ABSTRACT. Predicting the sequence pattern of the handwritten text images is a challenging problem due to the various writing style, insufficient training data, and even background and noise appearing in the text images. The architecture of the combination between convolutional neural network (CNN) and recurrent neural network (RNN), called CRNN architecture, is the most successful sequence learning method for handwritten text recognition systems. For handwritten text recognition in historical Thai document images, In this paper, we first trained nine different CRNN architectures with both training from scratch and transfer learning techniques to find out the most powerful technique. We discovered that the transfer learning technique does not significantly outperform scratch learning. Second, we examined training the CRNN model by applying the basic transformation data augmentation techniques; shifting, rotation, and shearing. Indeed, the data augmentation techniques provided more accurate performance than without applying data augmentation techniques. However, it did not show significant results. The original training strategy aims to find the global minima value and not always solve the overfitting problems. Third, we proposed a cyclical data augmentation strategy, called CycleAugment, to discover many local minima values and prevent overfitting. In each cycle, it rapidly decreased the training loss to reach the local minima. The CycleAugment strategy allowed the CRNN model to learn the input images with and without applying data augmentation techniques to learn from many input patterns. Hence, the CycleAugment strategy consistently achieves the best performance when compared with other strategies. Finally, we prevented image distortion by applying a simple technique to the short word images and achieved better performance on the historical Thai document image dataset.

ABSTRACT. Convolutional neural networks (CNNs) are now the state-of-the-art method for several types of image recognition. One challenging problem is vehicle image classification. However, applying only a single CNNs model is limited due to the weakness of each model. These can be solved using the ensemble method. Using the power of multiple CNNs together helps increase the final output accuracy but requires a lot of time-consuming. This paper introduced the new method called "Ensemble Multiple CNNs methods with Partial Training Set”. By combining the advantages of the ensemble to increase the accuracy and using the idea of "Partial Training Set" to decrease the time on training make it a good choice to compete with other single CNNs models.

ABSTRACT. Thai dragon fruit is an interesting fruit with beautiful colors and high nutritional value, which can be used for food and pharmaceutical. In Thailand, there are 7 species of dragon fruits. Its skin can be divided into red or yellow groups, but inside can be divided into white, red or pink groups. If farmers know the species of dragon fruits, they can export fruits at a good price. Most people are unable to distinguish species of the dragon fruits. This paper focuses on classifying species of Thai dragon fruits from digital images. Algorithm of DIPDEEP method consisted of three steps. The first step read an image converted its color into HSV, and calculate the ratio of yellow and red colors from H-channel. If the ratio of yellow is greater than, then it will be classified as the Israeli yellow species, and then exit. Otherwise, the RGB image of the dragon fruit will be segmented into a binary image with mean threshold of R channel. A dragon fruit can be found from its background, and then crop an image with size 100*100 pixels. Finally, deep learning was used to identify its species. The experiments were processed in a dataset with 9,754 dragon fruit images on a black background (laboratory), and 10,072 images of dragon fruits at outdoor environment (outdoor). The results showed that accuracy of classification between the red and yellow dragon fruit was 100% (laboratory) and 95.26% (outdoor), respectively. The red dragon fruit is classified its species with accuracy 98.80%.The DIPDEEP has smallest file size, and can save time by separating yellow bark out at first step.

ABSTRACT. Named Data Networking (NDN) is a new paradigm for the future Internet, aiming for efficient content delivery using in-network cache and information-centric communication. Security is built into NDN by embedding a public key signature in each data packet to enable verification of authenticity and integrity of contents. Access control is one of the most challenging issues in NDN. Several previous studies have proposed access control models over NDN. However, there are several drawbacks, particularly access revocation issues. We present a novel access control scheme to solve the problem by achieving immediate revocation. We have evaluated the performance of our scheme by comparing it to previous work. From the evaluation results, the proposed mechanism can provide an immediate revocation. We have also found that our access control scheme is suitable for NDN architecture that enables independent caching, and the computational burden for immediate revocation is less than previous proposals.

ABSTRACT. This paper presents the conversion of the Isarn Thai phonetic alphabet for the Nyaw language using basic rules of linguistics in which the characteristics of input sentences are written in the standard Thai language. The proposed framework consists of 2 main processes: the first process is used the string matching method to search for words from the Nyaw language dictionary, and the second process is applied the word segmentation method by comparing the longest word with a Thai dictionary. The first part results are divided into two cases, i.e., the case of the word found in the dictionary and the case of the word not found in the dictionary. For the first case, the words are identified to Nyaw phonetic alphabet. For the second case, the rest words are processed in the second process to re-segmentation using the longest matching algorithm with Thai dictionary. Finally, the word segmentation results will be converted into phonetic alphabets using the basic rules of the Isarn Thai language. To confirm the proposed method, we employed about 568 sentences in the experiments in which we found that the efficiency of the sentence conversion was up to 80.41%.

ABSTRACT. In this paper, we propose a method for detecting a human fall in colour video sequences. We combine a Support Vector Machine (SVM) with the Open-Pose method to classify parts of individual frames as human falling or not falling. We use Open-Pose to detect the 25 human body keypoints. We select two keypoints: neck and mid abdomen, to represent a human upper body. We subsequently draw a line on those keypoints to obtain the angle against the Y-axis. The angle is used as a feature for SVM to train. Finally, we used five differences dataset to test the proposed model. We demonstrate good results on a number of videos, comparing the detection and identification produced by this method with the manually extracted truth.

ABSTRACT. We introduce the novel technical results of a machine learning (ML) model for the detection of air leaks from fittings in a pneumatic pipe system using an accelerometer sensor system. The two main contributions are the following. The first is that we train the collected data generated from our experimental pneumatic pipe system based on the actual pneumatic pipe systems from one company production lines, then we obtain the model from convolutional neural networks deep learning ML technique. The model yields the best accuracy as 99.2%, compared to the accuracies of 89.23% and 89.56% of the decision tree and random forest ML techniques. The model can be one of the guidelines for the company to build the application to detect air leaks from fittings in a pneumatic pipe system using an accelerometer sensor system. The second is that, based on the model, we also discuss how to evaluate the accuracy of the model and the simple algorithm to build an application based on the model and the model's evaluation results. We believe the application can be applied to alert the engineer of this company about 'near-failure' or 'fail' situations of the fittings in the company's pneumatic pipe system. This can help the engineer to avoid the line's breakdown. This can also prevent costly maintenance and low production of this company mentioned above. This hopefully can benefit other companies with the same air leaks issues of this company. The compressed air leaks can cause half of the losses of manufacturing sectors. To the best of our knowledge, the contributions are not in the literature.

ABSTRACT. Fake news is a massive hassle in the global, especially on social media. Most people spend a lot of time-consuming social media every day which is very feasible for people to receive fake news as a social media user without realizing it. Primarily based on this hassle, we developed machine learning to detect fake news with the aid of the usage of various based algorithms such as Decision Tree, K-Nearest Neighbor, Naïve Bayes. The problem is based on algorithm for machine learning of which was selected one of those techniques used to classify the data by finding the model set. In addition, the performance of set describes classification the model and across the inconsistency solution for each iteration. In this study proposed a model which used the probability weighting of model into features extraction processing for data classification. The concept is the enhancing of probability weighting features which convergence exactly class label of classification. Our work also implemented based on traditional Count Vectorizer and TF-IDF Vectorizer sentiment analysis and combined probability weighting feature for fake news article. The experimental result of work shown that the best accuracy achieved by a proposed model using probability weighting feature and impact of classifiers models. In addition, the result of experimental information is represented by a data visualization which is categorized a data type according to the task taxonomy

ABSTRACT. This study presents a method of aspect-based sentiment analysis for customer reviews related to hotel. The considered hotel aspects are staff attentiveness, room cleanliness, value for money and convenience of location. The proposed method consists of three main processing stages. The first involves pre-processing customer reviews. This commences with breaking down each customer review into sentences, and then performing tokenization and bigram generation, stop-word removal, lowercase converting and lemmatization. The second stage assembles relevant sentences of each hotel aspect into relevant clusters of hotel aspects using BM25. Finally, the clusters are classified into positive and negative classes. Two algorithms as Support Vector Machines (SVM) with a linear and a RBF kernel, and Convolutional Neural Network (CNN) were applied to develop the classifier models. The classifier model based on SVM with a linear kernel returned better results than other models with an AUC score of 0.87. Therefore, this model was chosen for the classification stage. Four hotel experts were asked to provide initial keywords of each hotel aspect to generate other relevant keywords from a dataset by the word2vec algorithm. The corpus of keywords was called the Keywords of Hotel Aspect (KoHA) Corpus, and the keywords of each aspect were used as query to assemble relevant sentences of each hotel aspect into relevant clusters. The four experts also assisted with ground truths. The proposed method was evaluated using recall, precision and F1 with satisfactory results at 0.85, 0.87 and 0.86, respectively. Our proposed method showed the overview picture of customer feelings based on score, and also provided reasons why customers liked or disliked each aspect of the hotel. The best model from the proposed method was used to compare with the state-of-the-art. The results show that our method increases recall, precision, and F1 scores by 2.44%, 2.50% and 1.84% respectively.

ABSTRACT. When a software bug is found, the first thing a software development team should do is resolve the situation. Therefore, information concerning software bugs is essential. Previous studies concerning bug reports concentrated on binary classification (i.e. bug and non-bug). However, other information is also essential for software quality improvement or maintenance. Information designed to improve software products is called enhancement, while information concerning any other software engineering task is called task. After fixing all the bugs in the software, quality improvement and maintenance are still required, and bug report classification is not limited to only two classes. This study proposed multiclassification of bug reports into three classes as bug (or defect), enhancement and task. The dataset relevant to the Firefox opensource was downloaded from the Bugzilla system, and the pre-processing step was performed using tokenization, stop word removal, word inflection and lemmatization. We used unigram words along with CamelCase that indicated the specificity of problem domains in particular software. The second processing step represented pre-processed bug reports in a vector space model (VSM) and each term was weighted by tf-igm. The bug report vector for modeling the multi-classifiers involved eight supervised machine learning algorithms (i.e. Logistic Regression, Multinomial Naïve Bayes, Support Vector Machines, Random Forest, eXtreme Gradient Boosting, k-Nearest Neighbor, Decision Trees, Neural Networks. The best multi-classifier from the proposed method was chosen to compare with the baseline. The accuracy, F1, the area under curve (AUC) and the Matthews Correlation Coefficient (MCC) scores of the best multi-classifier from our proposed method were slightly better than the baseline, with improved accuracy, F1, AUC and MCC scores at 1.69%, 1.83%, 2.71% and 4.09%, respectively.

ABSTRACT. Coalition formation is an important area of research in multiagent systems. Among multiple solution concepts, Shapley value, the principle of distributing benefit among agents fairly, has been studied and extended widely in many domains. While advances in research on this area have been made extensively, it is doubtful that how much can people, particularly, the poor, benefit from them? This work presents algorithms for coalition formation of buyers in agriculture domain, where farmers use an application on their mobile phones to buy fertilizers (or other resources) in group and benefit from fair distribution of discount. We propose algorithms that $i$) helps farmers to form optimal coalitions of buyers, and $ii$) calculates fair distribution of discount rapidly in polynomial time. To our best knowledge, there is no other system that serves individual farmers with this advanced technique. In addition, other related parties can also benefit from this system.

ABSTRACT. DOAE of Thailand, maintaining databases of terabytes of real world data collected over several decades, is keenly courageous to develop an AI-based system and provide suggestions to Thai farmers, in order to successfully deliver reasonably acceptable, yet understandable, suggestions to millions of them in a conveniently affordable fashion. While modern techniques in AI research use standard data sets which have certain characteristics , normal distribution, etc., data collected from real world environment may not comply with such assumption. Here, we investigate whether Artificial Neural Network with Multi-layer Perceptron (MLP) and Random Forest (RF) models could effectively predict yield (product), cost, and price of crops in order to evaluate MLP model performance. Adjusting parameters such as learning rate and number of hidden nodes affected the accuracy of crop yield predictions. Smaller data sets required fewer hidden layers in model optimization. MLP models consistently produced more accurate yield predictions than regression models. MAPE (Mean Absolute Percentage Error) is used to measure the model of regression and MLP. MLP models proved that it produces accurate prediction. Similarly, RF is also deployed, particularly to provide suggestions when time constraints are tight. Although considered a less accurate method, compared to MLP, RF works just fine in most cases.

ABSTRACT. Predicting the optimal domestic power peak demand is very important for long-term electricity construction planning as the electricity cannot be stored permanently. If the prediction can give a yield close to the actual demand, the electricity suppliers can save their construction cost and provide their customer with benefits such as lower cost of electricity. However, the accurate prediction still needs to be improved. Therefore, this works presents the predicting problem using the modified artificial emotional neural network (AENN) based on the improved JAYA optimizes. This study also applies the extreme learning machine (ELM) to compute the expanded feature in the AENN. A real case study of Thailand power peak demand was considered, which was prepared using a rolling mechanism, to demonstrate the performance of the developed predicting model (AENN-ELM-IJAYA) when contrasted with the state of the art of AENN models, the artificial neural network with Levenberg-Marquardt (ANN-LM), AENN methods based on the winner-take-all approach, and the improved BEL-based on AENN model. The performance analysis demonstrated the proposed model provided performance and generalized stability improvement better than the comparative models.

ABSTRACT. The Twitter platform has been widely popular to reflect on feeling to various events personal lives. Their data can be analyzed using text mining to predict depression. The symptoms of depression consist of nine symptoms classification by DSM-5 criteria from American Psychiatric Association that is difficult to identify each symptom of depression. Then we propose the multi-objective optimization algorithms for depressive symptoms prediction modeling (MOADSP). This research aim includes: 1) find the appropriate number of features; 2) improve the weights for the prediction models based on the recall of the class for the Ensemble; and 3) compare performance model of the classification: single model, unweighted and weighted voting ensemble model for depressive disorder. The training data and deployment set came from Twitter. The features selection modeling is Information gains. The single classification techniques use the Naïve Bayes, Random Forest, and K-Nearest, and the vote ensemble models use the unweighted, TP weighted, and AVG TP weighted models. The best recall classifier is KNN (98.60%) from the loss of interest class from the single classification techniques, and the highest recall classifier is AVG TP weighted (98.43%) of the loss of interest class of the training model. The highest recall in the class depressive classifier is AVG TP weighted (80.00%). The result of the paired samples t-test process of the TP and AVG TP weighted in the depression is statistically significant (p-value < 0.05) for the deployment. This purpose is beneficial in predicting depressive disorder, where the number of terms is more than three terms.

ABSTRACT. One of the important strategy’s factors of petrochemicals plant maintenance is spare parts inventory policy that is significant on the efficiency, reliability, and productivity of the petrochemicals industry, unsuitable or not right spare part inventory policy will lead to a loss in long engineering machinery downtime due to the shortage of spare parts. To implement the spare parts inventory policy which is able to fulfill the future demand of the spare parts, the optimal calculation by various statistical theories and working processes are used to custom the spare parts inventory policy. However, to validate the custom spare parts inventory policy is the right or not right inventory policy, Simpy: discrete-event simulation library is used to mimic the actual spare part inventory system, has to be involved in the performance evaluation process of the custom spare parts inventory policy. The inventory simulation model consists of many events depending on the supply chain system but the crucial event which is the most complex for the simulation of spare part inventory is the demand event. This work applies the demand forecasting technique to the simulation by deep learning using a pre-built architecture model: Temporal Fusion Transformers (TFT) which the averaged MAE of the point predictions from a global model is 0.4874+/-6.7744 on the validation dataset and 0.6424+/-3.4963 on the test dataset, to predict a quantile forecast of the future demand which is able to handle the stochastic nature of the spare parts demand in the petrochemicals industry that makes the result from the simulation outcome is more accurate and close to the outcome from an actual inventory system. Then, the information from the analysis of the simulation outcome is used by the inventory management team to make a decision on the custom inventory policy before deploying it to the actual system.

ABSTRACT. This research aims to develop a new clustering algorithm called C-Algorithm that the document can classify to the previous domain or create the new domain and solve the K-means problem. This problem comes from the distance measurement of similarity from the new document to the centroid of each group. The new document will classify the group so that the relationship between groups and the new document is analogous or divergent. This experiment observes the proper group numbers using the Elbow method before starting the process. After this process, the Threshold value will be calculated from the centroid of the document in the group and percentile. The new document will compare with the Threshold and decision to set to the group or create the new document. This research compares the performance of the weight between the TF-IDF and BM25. These results show that the best performance comes from the BM25, Euclidean distance, and 80-85 percentile. The result of this research is more accurate than the conventional algorithm.

ABSTRACT. Price forecasting is one of the fundamental techniques used in most businesses to improve the competitiveness and decision-making level. Nonetheless, it is not simple to make a model that provides high accuracy price prediction, especially in modern enterprises with ever longer, and more complex supply chains across the globe. In the classical approach for predictive problem, researchers solved with time series forecasting, but no decent outcome has been developed so far. This work suggests a new way to tackle this problem in the modern complex business world to predict the price of plastic resin by using integration of textual information and numerical indicators input to deep learning models. Since the traditional methods which based on historical price itself is not sufficient, external data like economic indicators or textual information gathered from news articles, can help improve the performance of the models by catching the overall global economic sentiment. Word semantic is retrieved as vector representation from pre-trained word embeddings GloVe. In additional, deep learning models have gained great attention in the past decade after showing promising performance in various applications including Natural Language Processing (NLP), computer vision, and voice recognition. Hence, deep leaning models, Artificial Neural Network (ANN) and Recurrent Neural Network (RNN), are utilized in this research to deal with complex and fluctuated price of plastic resin. The models’ performances are validated with root mean square error metric. The outcomes are the robust models that show sufficient and satisfied result for plastic resin price prediction. This research has pointed out new models designs to handle time series input data with combination between textual and numerical data and contribute a new alternative strategy in petrochemical industries for more accurate price prediction which is the starting point for developing even more sophisticated and more accurate models in the future.

ABSTRACT. The Data Processing Impact Analysis (DPIA) is a process or methodology that should help an organization to ensure a reasoned and reliable use of personal data during the process. In the healthcare sector, medical information usually is very sensitive personal data that needs to be taken care of by assuring a high level of privacy. Additionally, the government of Thailand has recently announced a new personal data protection act (PDPA), which requires the organization to guarantee personal data protection by law. Furthermore, if the healthcare industry collects, stores, or processes patients’ data from the European Union (EU), the data processors are obligated to conduct a DPIA in accordance with the General Data Protection Regulation (GDPR). The DPIA should be implemented as soon as the new processing of personal data is designed, and it is a continuous improvement process. However, the DPIA process is still new for the healthcare industry in Thailand. Therefore, This study aims to propose a process model for conducting DPIA that is suitable to the Thai healthcare industry. As a result, the proposed process model was adapted from leading DPIA guidelines in EU countries and the PDPA of Thailand. The study’s findings should assist data controllers, and project managers in conducting an adequate DPIA, which should help increase public confidence in their organization’s information systems and comply with Thailand’s new personal data protection act.

ABSTRACT. Social media has grown in popularity over the past decade. Numerous services from a wide variety of industries have migrated to social media. The banking industry is likewise recognizing social media’s potential for client engagement. Initially, social media was being used as a platform for marketing communication. However, the beneficial aspects of social media include its ubiquitous nature, convenience, ease of use, and large user base that could support financial services. As a result, financial institutions begin to offer transactional banking services directly through social media platforms. However, the concept of using social media for financing purposes is relatively new, particularly among teenagers. As such, the main objective of this study is to determine the factors that influence youths in Northeast Thailand’s intention to use social media banking. The research model for this study was adapted from UTAUT by extending perceived convenience. The data of 159 participants were gathered through the online questionnaires. The hypotheses were empirically tested using partial least square structural equation models. The findings indicated that Performance Expectancy (β=0.202), Social Influence (β = 0.745), and Perceived Convenience (β = 0.561) significantly influenced the intention to use social media banking. The only component that is not relevant is the Facilitating Condition. Finally, the paper discussed and provided practical implications for stakeholders on social media banking to design a platform or marketing campaign that increases teenagers’ willingness to use social media banking.

ABSTRACT. SPOC (Small Private Online Course) is an online learning platform combining classroom and online lessons. For the development of the SPOC, it is worth looking at producing online learning as creating lessons for learners. Therefore, the learning behaviors are in-depth throughout the course. Not only about finishing each course, but also looking at the knowledge gained from studying. As Higher Education courses in Thailand focused on English language benchmarking, this research decided to choose the English Level 1 from August 2020 to June 2021 as a concentrated course containing the reading and grammar skills assessment. The study was conducted the Descriptive Analytics by Visual Analytics and K-Means Clustering, both student-level and lesson-level learning behavior. The result found four types of learners in Student-level learning behavior: 1) Intelligent 2) Weak-cognitive 3) Inattentive 4) Unenthusiastic. Moreover, the Predictive Analytics, for predicting learning quality through learning behaviors of each cluster by four machine learning models, were comparatively experimented: Generalized Linear Model, Decision Tree, Random Forest, and Gradient Boosted Trees. The Optimization method is used for tuning the optimum parameter of each method. For student-level behavior prediction, the Unenthusiastic's Decision Tree was 0.0449, and Lesson-level Weak-cognitive's Gradient Boosted Trees was 0.0371 relative error. Additionally, the factor importance in quality prediction was found that amount of quiz was the essential variable among all clusters. The result of this research is that the instructors can further develop content and teaching methods in the course to meet the needs of the learners truly.

ABSTRACT. Data Governance is defined as general Data Management practices and redundant data structures, and it sets direction and controls to ensure policy compliance rules and regulations. However, this is not enough for organizations with large amounts of data and complex organizational structures. Focusing on people and interactions, based on the agile concept, which entrusts a co-design and decision-making team, non-invasive in the original role to support the organization’s business in operations. The data can be embedded in the analysis group to sprint the data analysis cycles. This research proposes an operating model the Medical Agile Non-Invasive Data Governance (MANI-DG), which aims for the assignment of roles and responsibilities to working groups with specific expertise formally. The resulting streamlined workflow benefit from various solutions in planning strategies for further healthcare services. A healthcare organization in Thailand was selected as a case study. It emphasizes the roles and responsibilities throughout the organization and shows the more explicit implementation process of the prototype. The implementation and evaluation are divided into two levels: organizational level and operational level. The resulting RASCI matrix for the policy establishment process is evaluated by an in-depth interview at the organizational level. The operational level assessment makes the function of the concept clearly visible by a role-playing representation of data operations. A questionnaire of role-playing roles assessed it in terms of agility, connectivity, and redundancy, reducing roles and process responsibilities. The highest level of overall satisfaction is 4.73 regarding to 5-score rating scale. After comparative results with existing frameworks or studies, the researchers found that this prototype design is complete, coverage in both roles and responsibilities of each level according to the organizational structure. It can simplify streamline work processes and lead to analytics to connect with valuable, accurate, and transparent targeted outcomes for organizations.

ABSTRACT. Crowd simulation is an interesting area of research because it can be used as an important tool to make decision in various situation in real world problems.There are two main streams of technique in crowd simulation namely macroscopic and microscopic simulation, respectively. While the former is used to predict behavior of crowd based on mathematical models, the latter focuses on individual behavior in order to collectively demonstrate the behavior of crowd. Among many techniques, intelligent agent is a popular approach to simulate individual behavior at decision making level. However, another technique, Recripocal Velocity Obstacle (RVO), in particular, is also needed to properly simulate agent movement at physical level. In this research we combine both agent and RVO techniques. We propose a crowd simulation framework based on BDI (Belief-Desire-Intention), a popular psychological principle for decision making, using RVO as an underlying navigation mechanism for agents. The benefit of BDI is allowing for agent to plan and re-plan according to ever changing environmental status. RVO is also improved to allow for larger number of agents. The objectives are to observe the performance of our approach in terms of ability to carry out large number of agents at low cost as possible. We achieve satisfactory results as a large number of agents can be simulated smoothly.

ABSTRACT. This research aims to classify Thai defamatory messages or sentences on Facebook, where the learned text is derived from the comments under the pictures or under the articles of the person being mentioned. The research indicates whether the text is within the scope of defamation or non-defamation by using the text and special extracted features of the contents as the input data. We present the model creation to classify defamatory statements on Facebook, using three deep learning techniques: Long Shot-Term Memory (LSTM), Bidirectional Long-Short Term Memory (Bi-LSTM), and Convolutional Neural Networks (CNN). The results show that CNN using thai2fit for word embedding combines two key features inputs: Term Frequency of judges' vocabulary related to seven types of defamation (TF of judges’ vocabulary), and Part-of-Speech tag (POS-tag) gives the best result. We observe that CNN integrated with the presented features is more effective than LSTM and Bi-LSTM, which are set up the same input. This research focuses on the defamatory statements, which consist of unique legal characteristics. This result relates to the research conducted by Wenpeng Yin, who comparatively experimented the performance of CNN and RNN models with each domain text type. The study results show that the model's efficiency depends on the nature of the data. We set up a simple model for our research in which tuning the model's parameters can result in better efficiency.

ABSTRACT. Learning and understanding customer needs is one of the business strategies that will help build long-term customer relationships. This research has analyzed customer opinions compared to hotel features. To allow hotel businesses to use the information obtained to develop their business to meet guests' needs. This research proposed 1) Compilation of English comments from the website. www.tripAdvisor.com 12,644 reviews from 30 hotels 2) Word segmentation process consists of labeling the types of words. (Path-of-Speech Tagging) using the Penn Treebank Target was used to select verbs (Verb), adjectives (ADJ), and adverbs (ADV) to be processed. 3) The customer feedback analysis process is used to identify the feedback poles of each feature. 4) Extracting the hotel description, and 5) feature matching between hotel description and prediction result. It uses to check customer reviews and hotel strengths. The results showed that Test data, the number of comments from 600 messages, the overall efficiency was 0.78.

ABSTRACT. Multi-label classification (MLC) is a supervised classification method that allows a data instance with more than one class-labels (or targets). Solving MLC is still a challenging task. MLC can potentially generate complex decision boundaries as the method is a non-mutual exclusive classification. Recently, a number of techniques have been proposed to cope with this complexity of MLC problems, such as the Problem transform method (PTM), Adaptation method (AM), and Ensemble method (EM). These techniques can deliberately produce good results with certain datasets. However, they can outcome poor classification performance when the number of possible class-label is larger, even the dataset is well-presented (high density). This work aims to solve the MLC problems by performing a feature reconstruction process on the original data features. The proposed feature reconstruction method generates a set of compact features from the original data instances. AutoEncoder is deployed to learn and encode the features of the data (as the constructed feature steps) before they are classified by learning algorithms (or classifiers). We conducted the experiments using different multi-label classifiers based around on PTM, AM, and EM, on the set of the standard dataset. The results from the experiments demonstrate that the proposed feature reconstruction technique provides promising classification results, especially with high density data.

ABSTRACT. This research focuses to conduct an experiment to find the best classification model for Thai Timely and Timeless content classification which Timely and Timeless concept is an essential content marketing strategy. Timely is applied for the viral purpose, and timeless is adopted to gain brand awareness and marketing performance in the long term, is not just a temporal trend. Among six Timely-Timeless text classification models from two major state-of-the-art methods as CNN Neural Network and its variants: CNN Con1D, CNN Conv1D Skip-gram and Transformer and its variants: BERT, RoBERTa, and WangchanBERTa. The research results found the best one is WangchanBERTa, the large Pre-trained corpus exclusively for the Thai language. It has the highest validation accuracy at 93.06% while testing accuracy, getting more than 90% at 93.00%. Not only performing well for accuracy, but the validation loss of WangchanBERTa is also the lowest, at only 23.85%. Furthermore, this transformer-based text classification model is considered the best with performance metrics such as 93% for both Precision and Recall, while F1-score is highest at 92.00%. Although it is limited in the small dataset size as 600 articles with at least 250 words for each article, this dataset is separated into training 336, validation 144, and test dataset 120 articles. Therefore, this research can be a guideline for text classification on a small Thai language dataset. Our other contributions in this research are to present the way for pre-processing as cleansing Facebook dataset based on Timely and Timeless classification purpose and we also introduce the detail of model hyperparameters tuning which leads to high performance in our experiment.

ABSTRACT. Rural students such as those living in Northernmost regions of Thailand, tend to have lower learning performance compared to native Thai students due to the lack of access to educational resources, and learning technologies. To empower the knowledge, and 21st century skills of this user group, promoting active learning, and connecting subjects with students’ everyday lives is crucial. This paper presents a novel educational platform that integrates IoT (Internet of Things), and Chatbot-based learning system called ZiZy-BOT to facilitate learning for primary school students in Northern Thailand. This platform was designed to enable young learners to gain a more meaningful learning experience with an interactive agent and the real world information, anywhere anytime, and when they needed. The user context study was conducted, and observed to understand how 12 rural-students design, and interact with their paper-prototype. The results of the study contributed to a conceptual framework, and the key findings will lead to developing the platform, including augmented real-world information, conversation-based interface with friendly design, and a collaborative learning space. The developed platform will then be tested to examine the students’ learning outcome, and record their perceptions in a future study.

ABSTRACT. Pali Samas words cannot be found in any dictionary. They are created by placing Pali words which contain meaning after one another without changing their morphemes, or with changes in both morphemes and pronunciation. The changes in both aspects are generated from the link between the last syllable of the former word and the first syllable of the latter word into one same syllable, and the letters joined between each word are also changed. Thus, Pali Samas words need to be segmented back into their previously original words to obtain their meanings. Due to this reason, this research presents a novel approach to Pali Samas segmentation by using BiLSTM to predict the splitting locations as well as applies the rules obtained from Samas words segmentation to adjust their grammatical errors and achieve the correct meanings. For the dataset used in this research, the total of 2,757 Thai Pali Samas words from Dhammapada Atthakatha book are used to further created 4,478 Samas words by text augmentation. The result from Samas segmentation indicates that the prediction for the splitting locations is 99.20% weighted average of F1-score, which has 81.91% of the original words derived from the reverse segmentation based on the rules.

ABSTRACT. The foot strike pattern has a great impact on the knee joint of the runner during running and may be related to injury and performance. Currently, the most popular method for foot strike pattern detection is the strike index (SI), a measure that requires a force plate or force treadmill; therefore, foot strike pattern detection outside of a laboratory environment is difficult and costly. This research paper is aimed to optimize the Foot Strike Pattern Recognition by the novel feature extraction. The foot strike patterns are collected real-time by the novel and inexpensive designed smart wireless wearable sensor system installed equipped with the running shoes and pre-processing using the hybrid of CoP and PCA for feature extractions before classified the running pattern by machine learning. The comparison of six machine learning models including KNN, SVM (Linear kernel function), SVM (Polynomial kernel function), SVM (RBF kernel function), ANN and RF are also examined to find the best classifier. Experimental results show that the proposed optimized feature extraction that is used in SVM (RBF kernel function) provides higher recognition ability compared with that of an unoptimized model and helps in increasing the speed of recognition. The resulting system succeeds in classifying foot strike patterns correctly up to 98.68 %. The results are very promising, and it can apply to support runners for the foot strike pattern adjustment during running.

ABSTRACT. Nowadays the trend of environmental conservation has been interested. A lot of waste materials have been reused and recycled, or even reduced in order to preserve the environment. The objective of this research was to design and develop a low-cost reverse vending machine (RVM) for clear plastic (Polyethylene Terephthalate: PET) bottles and aluminium cans to support recycling process of solid waste in Khao Rup Chang Municipality, Songkhla Province, Thailand. Clear plastic bottles and aluminium cans were classified using a colour sensor (TCS230) and a proximity sensor. In the system, coupon slips could be printed out to show earned points for customers to exchange into money or rewards under the project. All of the mentioned processes were controlled by a microcontroller (Arduino Mega 2560). In terms of the test of its performance, it was found that the accuracy to sort each types of materials—clear plastic bottles and aluminium cans—and both materials at the same time was at 96.5, 100 and 95.25 percent, respectively.

ABSTRACT. Due to increasing urbanization at many cities, it is observed that there are an increasing number of people living in high density apartment/condominium complexes. Using Thailand as a case study, one of the common services that are provided at those complexes are public coin operated washing machines which allow tenants to do their laundry at a cheap and affordable price. This leads to a new set of challenges, as the quantity of washing machines are usually inadequate during peak hours, leading to long waiting queues for tenants. Due to that issue, the development of a smart public coin operated washing machine can help improve the quality of life for tenants. This research proposes a smart public coin operated washing machines platform by extending the washing machine using IoT applications. The proposed platform would be discussed including the infrastructure, the proposed application features, and on the vacancy state detection of the washing machine. With the proposed platform, tenants using public coin-operated washing machines can plan when to do their laundry off peak hours, improving their quality of life.

ABSTRACT. Face recognition is an important task in smart security for detecting a face or monitoring a person in a live video. The usage of face recognition is to verify the identity of an authentic user. However, there have been face spoofing methods that can trick a face recognition algorithm into wrongly verifying the identity of the person. In this paper, we proposed a new hybrid framework for spoofing face detection based on Convolutional Neural Network (CNN) and instance-based learning algorithm. In addition, a new dataset called FSA-CCTV is proposed, which contains face images from CCTV video clips. Our dataset consists of 2,045 spoofed face images and 1,015 genuine face images. The spoofed faces containing many types of spoofing attacks. The proposed anti-spoofing algorithm was run on this dataset. The performance of the proposed method was compared to several other anti-spoofing methods: CNN and RI-LBP, SLRNN, HSV+YCbCr, ResNet50, YCbCr+SVM and YCbCr+KNN. The experimental results show that our method achieved an Accuracy of 93.2%, Recall of 96.8%, Precision of 94%, F1-score of 94.8% and AUC of 93%. From the experimental results can conclude that the proposed algorithm outperforms the others on our dataset and provides a stable accuracy for each class.

ABSTRACT. The objective of this research is to comparatively study the analysis of depression potentiality from the videos recorded during PHQ-9 assessment and clinical psychologists’ expertise (psychometric) in major depression disorder (MDD) diagnosis by implementing experimental research from 3 subject: 1) subjects without depression disorder 2) subjects who have been suffering from depression without any treatment and 3) subjects who are undergoing depression treatment. The subjects were selected by using a purposive sampling which was consented according to Human Research Ethics Committee of Somdej Phra Phuttaloetla Hospital. In addition, the data was collected via experimentation at Psychiatry and Drug Addiction Subdivision in Hospital by using a research tool created from the application between facial expression recognition, user behavior tracking, and PHQ-9, in accordance with clinical manifestations and behavioral congruence. The relationship between PHQ-9 questions and the subjects’ emotions, which occurred while they are taking the questionnaires, are analyzed by using correlation statistics. After that, the results are compared by 3 clinical psychologists expertizing in depression diagnosis. Consequently, the relationship between the facial expressions and the PHQ-9 questionnaires illustrates interesting outcomes as follows: 1) Most of the emotions derived from the experiment are in the similar trends. To elaborate, the untreated subjects and the undergoing-treatment subjects are prone to Sad whereas the normal subjects tend to show Happy and barely have relationship with PHQ-9; 2) The occurring behaviors can be detected easier by the system than by naked-eye observation; 3) In terms of reaction time, the system is able to provide a more clearly detailed description of the reaction time than the time tracking observation method. For the future research, the controlled variables of the subjects and the experimental groups, and it should opt for a more unique facial expression recognition or train model to achieve the optimal accuracy in MDD prediction.

ABSTRACT. This article presents an image classification for quality grading of crown flowers using Neural Network and Transfer Learning. It consists of 1) image acquisition, 2) pre-process, and 3) data classification. Using data of 1,500 white crown flowers, divided into five classes as follows: 1) flowers with the base of the petals close together, 2) flower with mold, 4) flowers with more than one bad trait, 4) flowers with unequal petal length, and 5) complete flower. The experiment divided the data into two parts, including 80% of Training Data and 20% of Testing Data. This article experiments with four models, including Convolutional Neural Network (CNN), Xception, VGG16, and InceptionV3. The experiment results show that both Convolution Neural Networks and InceptionV3 Networks were 85%; moreover, Xception and VGG16 were 83% and 81%, respectively. Finally, the result of performance measurements, the CNN model, shows the highest average of all precision, recall, and F1-score equations. Accordingly, the CNN model is the deep learning of the characteristics of the image, which gets to more efficiently and accurately than other models. It can apply to classify the quality of crown flowers, which supports the gardeners in grading flowers more efficiently.

ABSTRACT. The 5G cellular networks are expected to provide the high data rates for indoor environments by utilizing the mmWave bands. Due to missing of general path loss models of these bands, various works have been carried out to find the most appropriate model that can provide accurate values of path loss over distance for different indoor scenarios. This paper focuses on some well-known path loss model, called mixed and open office models were introduced by 3GPP; dual-stripe model was presented by Alcatel-Lucent, picoChip Designs and Vodafone at 3GPP meeting; single and dual ABG models were designed by leading academic and industrial labs. Two network scenarios are utilized in simulation to analyze the effects of path loss models on the network performance. The simulation results indicate that the single-slop and dual-slop model can derive the similar performance when only a cell is deployed. However, in a 3-cell network system with intercell interference, the achieved performance in these systems are significantly different since dual-slop model can capture the Light-of-Sight (LoS) and Non LoS (NLoS) effects.

ABSTRACT. Laem Chabang is the one of Thailand’s most air-polluted cities because it is the largest seaport and nearby industrial factories in the country’s Eastern Economic Corridor. Therefore, a PM2.5 forecasting system is necessary so that the public can be notified of the air quality. This paper implemented a PM2.5 forecasting flow system with three main processes: data preprocessing, a PM2.5 forecasting model and its evaluation. We selected the hourly concentration of PM2.5 at Laem Chabang Municipal Stadium station as our dataset. Two forecasting models were investigated (T2V_CNN and T2V_LSTM) that applied Time2Vec to transform a sequence of PM2.5 records to Time2Vec vector representation and then forwarded them as input to CNN and LSTM models. These experiments evaluated model accuracy using the mean absolute error (MAE) and root mean square error (RMSE) and evaluated the processing time as an indicator of the speed of each model. The experimental results showed that T2V_CNN had the lowest MAE and RMSE values of the all compared techniques. The T2V_LSTM methods clearly reduced prediction error from the original LSTM. Accordingly, T2V_CNN model was further investigated and performed well in forecasting the concentration of PM2 in terms of accuracy and processing time.

ABSTRACT. In academic institutions, the grading approval process is an important part of the educational quality assurance process. In the example of Assumption University in Thailand, course coordinators can set the grading distribution for their respective courses, but the grading would be scrutinized by internal and external committees in the grading approval process. The grading approval process is a manually intensive process that is an error-prone and time-consuming especially when the committees provide suggestions in modifying the grading distribution. The research proposes a grading approval system that help streamlines the process. The key feature of the proposed system includes a rule-based grading distribution system that allows grading coordinators to quickly setup grading distributions that are based on patterns that are commonly suggested by the internal and external grading committees of the Department of Digital Business Management at Assumption University. Comparing the proposed system's grading distribution, the course coordinator's grading distribution, three standard distributions, and the bell curve distribution with the approved grading distribution of courses offered by the Department of Digital Business Management at Assumption university of the academic year of 2019 to 2020, the proposed system has performed well compared to other control grading distributions.

ABSTRACT. With the increasing population, the amount of waste has significantly increased. One of these waste streams contains a considerable amount of recyclable plastic. As a result, recyclable waste must be managed effectively to mitigate the impact of improper waste disposal. This article discusses the application of smart technology to waste management in social enterprises. This research study is divided into two major sections: Part 1 involves developing an automated waste sorting system utilizing image analysis techniques; Part 2 consists of the development of a collection of waste management applications for a social enterprise. A robotic automation uses deep learning techniques and digital image processing to separate recycled waste correctly. The smart bin was developed with a garbage sorting model that automatically uses an embedded microcontroller to separate four different types of garbage. On the basis of motion detection, a web camera connected to the Raspberry Pi captures individual images of the garbage. After capturing the image, it is sent to the classification model, which predicts the garbage class. Waste is transferred directly to the appropriate bin obtained from the results, using a motor-driven moving plate as a fully automated sorting machine. Additionally, a social enterprise housed within the university are used to assist with waste management through mobile applications. These activities encourage students and faculty to recycle waste and convert it to reward points redeemable for goods and services at the university. The data collected from the automated waste sorting system is stored in a cloud-based database system to analyze waste management policies at the university. The findings of this study can be used to improve waste management in social enterprises in other universities throughout the country.

ABSTRACT. Online learning is a primary strategy for learning during the COVID-19 pandemic, which takes advantage of the physical distancing and contactless among lecturers and learners. Technology-enhanced learning is widely utilized over the Internet, while the demand for video communications applications has increased as a crucial tool for education. Even the video communication application supports to create a virtual classroom environment instead of an onsite classroom, the learning behavior is invisible. In this paper, an online lecture support system for visualizing the learning behavior is proposed. The system can provide learning materials as similar as the traditional lecture. The system allows learners to access the learning materials and take their notes in class period and overtime. Moreover, the system can visualize the learning behavior regarding the number of visiting and re-visiting of each material. The spending time is calculated to illustrate the learners’ engagement in the course. The word cloud is generated by using note-taking of learners to visualize the recording information. Furthermore, the clustering of learning behavior was conducted for demonstrating the pattern of behavior in the online lectures. The system can inform the learning behavior to the lecturers, which describes how their learners learn in the virtual classroom.

ABSTRACT. A dragon fruit is nutritious having low sugar levels suitable for diabetics. Its shape is spherical with petals surrounding the bark. In Thailand, 7 species of dragon fruit are grown. Its appearance can be divided into red bark (6 species) and yellow bark (1 species). People can separate between yellow and red bark. The big problem is people cannot classify species of red bark. Since the red-skinned dragon fruit may have its fruit pulp one of three colors: white, red or pink. It is difficult to guess from the outside what color of the fruit pulp will be. Some people may be allergic or dislike dragon fruit in some species. This research is trying to find the best way to automatically classify the species of Thai red dragon fruit from its image. The CBML stands for content based and Machine Learning. This method use content based to extract key features (34 attributes),and use Machine Learning for giving optimization results with the Support Vector Machine (SVM), setting kernel for polynomial in 8 degree. The results showed that the CBML method was able to identify any species of dragon fruits in the red bark group with an accuracy of 98.47%.

ABSTRACT. At present, in Thailand there is a policy to develop the economic system and industry into the Thailand 4.0 era that will transform the traditional economy into a better economy driven by technology and creativity. In which the industry has an idea to bring innovation and technology. Artificial intelligence is also known as AI that has been written and developed to be intelligent. Ability to think, analyze, plan and make decisions. In addition, AI technology can be further developed for use in the manufacturing industry or in other fields. Infinitely, using Machine Learning techniques in the field of deep learning that has the ability to allow machines or devices to learn by teaching or learning by themselves without programming into use in detecting the produced workpieces and counting the number of workpieces instead of employees, which this technology can reduce the number of workers. At this point, the workload is reduced to allow for a change of duty to another part that lacks or reduces the cost of hiring employees by applying the principles of deep learning to help the manufacturing industry to make the production process accurate. The model used in the recognition is 97% accurate. The developed system is used to produce satisfactory results.

ABSTRACT. This paper presents machine learning techniques to detect the failure in hard disk drive manufacturing test process. The data is high dimensionality and highly imbalance. Feature selection technique with filter method and embedded method with light gradient boost are used to reduce the dimension of data. We apply three techniques: SMOTE, Different Cost and SMOTE with Different Cost to handle imbalance data. Several machine learning methods are compared. The XGBoost with SMOTE and XGBoost with Different Cost (XGB DC) give the best performance with 91% ROC AUC and 73% PRC AUC. The SVM algorithm shows good performance on ROC AUC while low performance on PRC AUC. The XGBoost algorithm shows good performance of both ROC AUC and PRC AUC.

ABSTRACT. Even before the current (2020/2021) pandemic began, Conversational User Interfaces (CUI) had been seen as a valuable way to ease the burden on medical staff in many countries. In times of restricted direct contact with people, the need for online or virtual tools to connect patients with physicians has become even more visible. In addition, these restrictions hamper the training conditions for prospective doctors. This paper describes the design and implementation of a CUI covering patients' minor complaints of the ear, nose, and throat (ENT), which can be correlated with infection by COVID-19. The purpose of this study is to provide pilot test results for an online anamnesis and diagnosis tool supporting the cooperative work of specialists and non-specialists at their workplaces. We have designed and created the Cooperative Online Anamnesis and Diagnosis System (COLDS) using (1) a knowledge-based system for the anamnesis mainly of complaints relating to ENT including the eyes, (2) a knowledge base of disorders regarding ENT and eyes, and (3) a user interface that assists patients as well as cooperative processes involving non-specialists and specialists. COLDS is part of a clinical decision support system. The system has been evaluated in a two-tier pilot test process set in a real-life environment: Tier 1 was concerned with the usability of the system, whereas Tier 2 involved medical specialists to evaluate the outcome and recommendations created by the system based on an adapted Objective Structured Clinical Examination (OSCE) framework.

ABSTRACT. Unstructured data is the most common type of data generated on a daily. It is critical that businesses discover ways to retain such data so that they can analyze and use it to make business choices. Document databases have emerged to the needs of applications. This database has a schemaless feature that allows for more flexibility in data management. Particularly, a document database does not define conceptual modeling. The conceptual data model is one of the most important phases in database development. In a relational database, the Object-Role Modeling (ORM) is used for data modeling purposes. However, when it comes to designing document databases, which model is the best? Therefore, these databases need their conceptual model to understand business users and easily in the implementation. This work provided a standard design strategy for document representation. ORM is utilized in this approach for standard conceptual modeling that is adopted through a document model. This paper defined a set of graphical symbols to document database concept such as collection and two kinds of relationships. Furthermore, the current challenges in document database modeling were highlighted throughout the discussion of this topic.

ABSTRACT. With the Thai government and education policy investigating ways to drive Thailand towards an innovative and creatively based society, STEM education has become increasingly an important learning method. Typically, STEM education implementation requires teachers to be expert in STEM areas to helps to bridge the gaps sometimes found in STEM-related fields. Moreover, STEM educators have increasingly sought to explore tools and techniques to enable creating suitable approaches to promote STEM engagement. Simultaneously, two-dimensional (2D) virtual environments have been applied in various disciplines and areas of study. This has led to the essential approach - how is a 2D virtual environment developed and how does it handle delivering and supporting collaborative STEM learning and communication between STEM students and STEM-related field teachers? The aim of this project is to collaborate with learning processes across different STEM learning fields. In this work we describe the development and deployment of a 2D virtual environment tool to bring diverse STEM teacher teams and STEM students to learning together. This project proposes an interactive multimedia STEM instructional module developed to provide a delivering and supporting collaborative STEM learning and communication between STEM students, and STEM-related field teachers. This module was built following the interaction design principle to create instructional material for learning based on 2D virtual environment in teacher team delivery. Unity was used to create the 2D virtual environment and containing a learning digital media, control learning digital media using C#. This project can support those who teach STEM and those who are considering collaborating with learning processes across different STEM learning fields.

ABSTRACT. Since The World Health Organization declared the 2019 coronavirus disease (COVID-19). In 2020, 77 countries have reported cases of COVID-19. Various organization have demonstrated the ability to reduce the COVID-19 virus transmission. In response, many software industries have transformed to an electronic work culture. It was a big adaptation for the workers. Scrum is an agile project management framework which is widely used in many software organizations. Use of online Scrum tools is increasing rapidly nowadays. For education, the Scrum approach was adopted for Software Engineering students to learn the current way of developing software projects. This research aims to compare the effectiveness of learning to develop software utilizing a Scrum framework by comparing between online and onsite groups of students. Gamification, which is the application of game elements in non-game context, was adopted to enable this learning. Game components are used to increase team motivation and change behavior. Six research questions were produced to compare the difference between learning Scrum utilizing gamification online and onsite. The first four questions compare the differences between online and onsite students’ opinions toward using the agile framework, based on four key values of the agile manifesto. The online group responded to change better than the onsite group, whereas there was no significant difference in the other three values. The fifth question examined negative Scrum activities. The result was that the Velocity chart was the most difficult technique to understand. The fun game aspects were evaluated in the last question. The results show that for both online and onsite participants, the “Leader board” was the students’ favorite fun game element.

ABSTRACT. Currently there are many user interface design styles using a variety of app icon designs. The earliest style was skeuomorph which imitates the design of real-world objects. Flat and material design have emerged more recently and all of them are still currently in use. This paper observed 15 basic app icons from five mobile phone brands representing different icon styles; skeuomorph, skeuominimalist, flat, material design, and broken line. Young and senior adult participants were recruited into this study to examine their satisfaction with app icon styles. A visual search task was employed to measure the designs’ efficiency and effectiveness. The results show that both age groups are most satisfied with skeuomorph and skeuominimalist. Seniors prefer flat rather than material design or broken line, and vice versa for young adults. In the visual search task, both age groups took longest for flat design and also made the most errors. In visual search, the skeuomorph icon design was the fastest for young adults, while material design was the fastest for senior adults. There were also some significant interactions between age, styles, and app icons.

ABSTRACT. An industry sector is important for the economic growth in Thailand. Among those industries, the electrical and mechanical parts manufacturers are also essential to drive the production process in the factory. Due to the foundation activity supporting, the Industrial part manufacturer becomes more competitive in the market proven by The Business report in 2019 that stated the lost customers and open status of quotation is increasing dramatically. In order to solve and further prevent these problems and gain more competitive advantage, the Data Mining technique would be necessary to descriptively understand and predict customer behavior which can improve the business strategy to be more effective, which the Return-of-Investment of simulated business scenario will prove. The data used in this paper is customer data between 2017 and 2019 in two entities: 1) customer characteristic data including Registered capital, Industry code, Business type, Business size value, and 2) customer transaction data including purchase history. The combination of descriptive segmentation and predictive modeling towards decision-making strategies that tend to increase the Return-of-Investment of the industries is challenging, and the main contribution is specified in electrical and mechanical parts manufacturing. The expected results should support the Sales and Marketing team in increasing sales value and new customers and maintaining existing customers by offering highly accurate strategy segmentation.

ABSTRACT. Data silos are a challenge in both public and private companies. As a result of the organization's division of work functions into numerous departments, each party has distinct responsibilities and creates duplicate software, applications, and data. There is relatively little information sharing or exchange between departments. Data discrepancies or mismatches happen as a result of this. Although the same data is stored in multiple silos (databases), the data architecture, names, and meanings differ. As a result, users are perplexed as to how to use the data. This research applied the data lake concept to solve the data silo problem. The scope of the research focuses on structured data silos. The objective of this research is to design a data lake architecture and its internal working framework by using Hive and Spark technologies to integrate data within a data lake and write functional testing programs in Java Spark. According to the result of testing based on a detailed developed framework, integrating data silos on data lakes can reduce data heterogeneity and data inconsistencies by 100%, and it was able to reduce the redundancy of the test data by 78.6% from the total of 13 separate data cases.

ABSTRACT. Packaging is one of the essential factors in the production and selling process. Knowing the leading indicators and predicting the demand of consumer goods that use a high volume of packaging can help packaging manufacturers plan their production to satisfy consumers' needs and plan production costs effectively. In this research, we present the economic factors that affect the demand for consumer goods in Thailand and predict the production for each product using machine learning approaches. We use model-based selection, f-regression, mutual information, and recursive feature elimination (RFE) for the feature selection. Moreover, for forecasting, we compare the performance of machine learning models that can describe algorithms inside included multiple linear regression, random forest regression, and gradient boosting. The results show that using random forest regression with model-based feature selection gives the best score for more than 60% of the total products.

ABSTRACT. Nowadays driving a car is necessary to travel to places. The benefits of a car to travel are comfortable and fast. Since the increasing number of cars are being driven every day, the number of on-site parking spaces is insufficient to park cars. Therefore, this work is created to help us know how many parking spaces are left inside the facility. The equipment used in the experiment was a camera and a laptop. The camera was installed at an electricity post near the dormitory. After the video of cars entering and exiting the road have been recorded, they were processed by YOLOv4 techniques. The images of the cars were analyzed. Which has been experimented with two forms: The first form is a separate type of car. The result is sedan car has the highest detection accuracy of 100%, while the van has the lowest detection accuracy of 66.7%. The second form was the accuracy of vehicle detection at times intervals. The result of the highest detection accuracy is 94.8% in the evening and the lowest detection accuracy is 68.4% at night. Finally, the number of cars entering and exiting has been counted automatically and stored in the database. Not only the number of cars entering, exiting, and remaining in a place have been shown on the web page in real-time, but also the statistics of the total number of cars that enter each day has been stored. This project will help to make it easier to see how many parking spaces are left to be able to park. Which our web page, the part that received the most satisfaction score is contact at an excellent level.

ABSTRACT. Coronavirus pandemic has changed the classrooms globally. Due to the pandemic, the online classroom plays a vital role in education systems instead of the traditional classroom. However, there are some difficulties that learners have to deal with which affect their satisfaction and engagement. This study aims to survey the critical aspects of instructors and student peers on the students’ engagement. In particular, interactions (instructor-student interaction and peer interaction), social presence, and communication channels that meet students’ needs to foster effective online learning were observed. The online questionnaires were used to capture student attitudes, experience, and broad insights into synchronous online learning. Results suggested that interaction with instructors and peers tended to positively impact students’ engagement (x̅ = 4.49, S.D = 0.82, and x̅ = 4.02, S.D = 1.12, respectively). Moreover, instructor presence tended to have a higher impact on students’ engagement than peer presence (x̅ = 4.00, S.D = 1.24 for the instructor presence, and x̅ = 3.36, S.D = 1.30 for the peer presence, respectively). Finally, it was found that the most favorite communication channels with instructors were chat applications (i.e. Line), e-mail and mobile phones, respectively. Additionally, the favorite communication channels with peers were chat applications (i.e. Line), social media (i.e. Facebook and Instagram), and mobile phones. These study findings should benefit the classroom design on choosing suitable activities and interactive tools to increase student engagement in an online learning environment.

ABSTRACT. The research on stop motion animation short film : the spread of covid-19 virus prevention, the objectives of the study were 1) how to prevent the spread of COVID-19 2) to realize the prevention of COVID-19 people have more self-protective behaviors; Especially wearing a hygienic mask every time when going to the community, regularly washing your hands with soap or alcohol gel and keeping a distance of at least two meters. Avoid bringing your hands to touch your eyes, nose and mouth. Moreover, the prohibit share personal items with others (such as handkerchiefs, glasses, towels) due to pathogens in the gastrointestinal tract and breathing can enter the body through contact with the secretions of an infected person. Should be eating hot cooked food. The awareness of the impact of the COVID-19 pandemic is of great importance, in order to study the knowledge of protection and prevent possibly spreading COVID-19. The media that can enhance learning and remembering in the form of digital online media will help to understand. It is part of stopping the spread of the virus and explaining the prevention of the spread of the epidemic. A samples of 50 people aged 16 to 21 years, using simple random sampling method. The mean and the standard deviation data were used for data analysis and the evaluation from 3 experts. A quality assessment and performance evaluation of our media were a good level. The findings indicated that, learned how to make a stop motion animation film with the interesting process. The equipment preparation various settings in production which will have many limitations, learning from the beginning process until the completed of the process. It will be the experience and can be accumulated for animated films jobs.

ABSTRACT. Augmented Reality (AR) technology is frequently applied in simulation, particularly in medical training. The AR application renders the 3D model on mobile devices that be resource constraint devices. Hence, the characteristics of the 3D model for the mobile devices are low polygon is to reduce the resource constraint in rendering time and realistic rendering in the animated model. The topology of the polygon is the key to these issues. This paper proposes the retopology of the polygon based on the papercraft technique. We compare our approach with the experience of the digital artist in two groups of samples. We compare the performance in terms of the number of polygons, man-hour, and the quality of the model. The result shows that our approach reduces the number of polygons and man-hour better than the control group whereas the quality of the model is not different significantly.

ABSTRACT. A decision tree is a common approach used for classifying unseen data into defined classes. The Information Gain is usually applied as splitting criteria in the node selection process for constructing the decision tree. However, the bias in selecting the multi-variation attribute is the major limitation of using this splitting condition, leading to unsatisfactory classification performance. To deal with this problem, a new decision tree algorithm called “Knowledge-based decision tree (KDT),” which exploits the knowledge in ontology to assist the decision tree construction is proposed. The novelty of the study is that an ontology is applied to determine the attribute importance values using the PageRank algorithm. These values are used to modify the Information Gain to obtain the appropriate attribute to be a node in the decision tree. Four different datasets, Soybean, Heart disease, Dengue fever, and COVID-19 dataset, are employed to evaluate the proposed approach. The experimental results show that the proposed method is superior to the other decision tree algorithms, such as the traditional ID3 and the Mutual Information Decision tree (MIDT), and also perform better than a non-decision tree algorithm, e.g., the k-Nearest Neighbors.

ABSTRACT. Presented in this article is a student-led project to retrofit a tissue culture temporary immersion system (TIS) for remote operation. This equipment, prevalent in biotechnology research and agriculture industry, is a great candidate for automation and IoT enablement as the systems are semi-automatically controlled and pose modest hurdles to additional automation. The proposed system is capable of being remotely regulated and real-time monitored via an online web application. This system also organizes the data using a cloud platform -- allowing researchers to access the records and images for further analysis. Currently deployed at a university laboratory, the IoT-based TIS setup adds remote monitoring and configuring capability to the existing system and provides a base for future expansion.

ABSTRACT. Typical non-cooperative game theory studies how agents will behave and what their payoffs will be within certain games, i.e. specific actions and corresponding payoffs. We are interested in wider spectrum of outcomes in games, i.e. how much payoffs can vary within a trend such that the agents' strategies remain unchanged. Furthermore, traditional game theory, particularly, strategic form game, considers merely actions and payoffs of agents. In real world, such information may not be known a priori. In non-cooperative environment, we choose Bakery Game Intara and Sombattheera (2018) as our testbed to explore how much payoffs, more precisely, resources of agents can very within certain trends such that agent's strategies do not change.

ABSTRACT. CORONAVIRUS 2019 or known as COVID-19, began to spread in December 2019. The first affected covid-19 person was found in Wuhan district, China, and then COVID-19 impact worldwide, including Thailand. This research uses the Data Mining technique by applying the Association Rule to understand customer behavior insight who buy orchid pots during the first coronavirus pandemic period in Thailand. This study applies to the database of customer purchase transactions who buy orchid pots. This research adopts the Market Basket Analysis concept like an FP-Growth model to get a solid understanding of groups of products customers typically buy. Finally, the Association rule generates seven rules of orchid pot types that customers purchase at the same basket. Each rule shows Confidence, Lift, and Conviction range from 0.833 – 0.857, 2.629 – 5.602, 4.098 – 5.929, respectively. And then, this study also deployed a Predictive model to get the relationship between label feature and other dependent variables by utilizing a Generalized linear model, Deep learning, Random Forest, and Gradient Boosted Tree. As a result of the Predictive model, Gradient Boosted Tree without Auto feature selection and feature extractions methods produce the lowest relative error at 15.2%. The association rule finds out an orchid pot that customers purchased one of the items in the group. The expected result of this study is that orchid entrepreneurs can adopt this outcome from Association Rule and Predictive Modeling Analytics applying when a difficult situation similar to the COVID-19 pandemic happens again.

ABSTRACT. การให้บริการระบบสารสนเทศสำหรับนักศึกษาเกี่ยวกับงานทะเบียนของมหาวิทยาลัย ยังคงเป็นปัญหาในเรื่องการอำนวยความสะดวกในการถาม-ตอบปัญหาต่าง ๆ เกี่ยวกับงานทะเบียน เช่น การลงทะเบียน การรักษาสภาพ การย้ายวิทยาเขต การย้ายคณะ การเพิ่ม ถอนรายวิชา การสำเร็จการศึกษา หรือ ปัญหาอื่น ๆ ที่นักศึกษาประสบอยู่ เนื่องจากมหาวิทยาลัยยังไม่มีระบบถาม-ตอบออนไลน์แบบอัตโนมัติ งานวิจัยนี้จึงมีวัตถุประสงค์เพื่อ สร้างต้นแบบระบบถามตอบออนไลน์ ด้วยเทคโนโลยีออนโทโลยี โดยมีวิธีการดำเนินการวิจัยประกอบด้วย 1) การรวบรวมข้อมูลจากคู่มือนักศึกษาโดยการวิเคราะห์ข้อความ และแบ่งหมวดหมู่ต่าง ๆ เพื่อสร้างเป็นตัวแทนคำตอบ โดยการตัดคำ สกัดคำหยุด หาตัวแทนของคำ 2) กระบวนการพัฒนาในการสร้างออนโทโลยี เพื่อระบุขอบเขตขององค์ความรู้ที่ครอบคลุมในเรื่องต่าง ๆ แต่ละหมวด 3) รับค่าคำถาม เป็นการระบุคำถามจากผู้ใช้ โดยนำคำถามไปประมวลผล และ แสดงคำตอบที่เป็นไปได้มากที่สุดออกมา 4) ประเมินประสิทธิภาพการทำงานของระบบด้วยค่า Precision, Recall และ F-measure แหล่งข้อมูลที่ใช้ในการวิจัย จะใช้กรณีศึกษา: งานทะเบียนมหาวิทยาลัยเทคโนโลยีราชมงคลศรีวิชัย ผลการทดลองเบื้องต้นแสดงให้เห็น การประเมินประสิทธิภาพระบบถาม-ตอบ ออนไลน์ ด้วยเทคโนโลยีออนโทโลยี สำหรับการถาม-ตอบเกี่ยวกับงานทะเบียน มีค่าความแม่นยำ (Precision) 90.91% ค่าความระลึก (Recall) 83.33% และ ค่าเฉลี่ย (F-measure) 86.96%

ABSTRACT. การพัฒนาซอฟต์แวร์เป็นการพัฒนาโครงการซอฟต์แวร์ในรูปแบบทีม ทักษะการทำงานเป็นทีมจึงมีความสำคัญและจำเป็นสำหรับนักพัฒนาซอฟต์แวร์ในอนาคต ซึ่งการทำงานร่วมกันเป็นทีมจะประสบความสำเร็จในโครงการพัฒนาซอฟต์แวร์ได้ขึ้นอยู่กับความสามารถทั้ง 2 ด้านของนักพัฒนาซอฟต์แวร์ คือ ทางด้านเทคนิค (Hard skills) และด้านบุคคล (Soft skills) โดยบุคลิกภาพเป็นส่วนหนึ่งของด้านบุคคล ถ้าสมาชิกในทีมเข้าใจถึงความเหมือนและแตกต่างของบุคลิกภาพของสมาชิกภายในทีมก็จะช่วยให้การทำงานร่วมกันมีประสิทธิภาพมากขึ้น ผู้วิจัยทำการศึกษาความเหมือนทางบุคลิกภาพของสมาชิกทีมมีผลต่อประสิทธิภาพการพัฒนาซอฟต์แวร์อย่างไร โดยเก็บรวบรวมข้อมูลบุคลิกภาพของนิสิตสาขาวิศวกรรมซอฟต์แวร์ชั้นปีที่ 3 ที่ทำงานร่วมกันเป็นทีมสกัม (Scrum) เป็นระยะเวลา 1 ปี จำนวน 147 คน แบ่งเป็นทีมๆ ละ 8 -10 คน จำนวน 16 ทีม วัดบุคลิกภาพด้วยแบบประเมินบุคลิกภาพ MBTI และวัดประสิทธิภาพของทีมจากคะแนนเฉลี่ยการพัฒนาซอฟต์แวร์ในแต่ละวงรอบทั้งหมด 4 วงรอบ ประกอบด้วยคะแนนประเมินกระบวนการในการพัฒนาซอฟต์แวร์ (Process) จำนวน 8 ประเด็นคะแนนเต็ม 40 คะแนน และคะแนนประเมินผลลัพธ์ของผลิตภัณฑ์ (Product) คะแนนเต็ม 10 คะแนน โดยผู้เชี่ยวชาญในการพัฒนาซอฟต์แวร์จำนวน 10 คน ผลการวิจัยพบว่า ความเหมือนทางบุคลิกภาพของบุคคลในทีมไม่มีผลต่อประสิทธิภาพการพัฒนาซอฟต์แวร์ทั้งการทำงานร่วมกันและคุณภาพของผลิตภัณฑ์ซอฟต์แวร์ ถึงแม้ว่าผลการวิเคราะห์จะแสดงค่าความสัมพันธ์กันในเชิงบวกแต่ก็ไม่มีนัยสำคัญทางสถิติที่ระดับ 0.05

ABSTRACT. งานวิจัยฉบับนี้เป็นการแก้ปัญหาในช่วงสถานการณ์ covid19 ที่เกี่ยวข้องกับการบริการให้คำปรึกษาที่มีอยู่ในมหาวิทยาลัย เนื่องด้วยนักศึกษาไม่สามารถเข้ามาในพื้นที่มหาวิทยาลัยจึงทำให้ไม่สามารถติดต่อสื่อสารโดยตรงกับเจ้าหน้าที่แต่ละแผนกที่เกี่ยวข้องกับการบริการนักศึกษา เช่น กิจการนักศึกษา งานทะเบียน เป็นต้น ปกติช่องทางที่ใช้สำหรับการติดต่อได้แก่โทรศัพท์และช่องทางสังคมออนไลน์เพสบุ๊ค ซึ่งทางโทรศัพท์ใช้ได้เฉพาะในช่วงเวลาราชการเท่านั้น ส่วนช่องทางสังคมออนไลน์เพสบุ๊คก็พบว่าไม่เป็นระบบและไม่สามารถติดตามผลของการแก้ปัญหาของนักศึกษาได้ ทางผู้วิจัยจึงได้ศึกษาและพัฒนาระบบโดยเลือกรูปแบบในการให้บริการแบบเว็บแอพพลิเคชั่น สำหรับเทคนิคการแลกเปลี่ยนข้อมูลใช้หลักการ REST ซึ่งทำให้การขยายและการรองรับผู้ใช้ได้จำนวนมากอย่างมีประสิทธิภาพ ในการวิจัยได้ใช้กรณีศึกษาการให้คำปรึกษางานทะเบียน จากการวิเคราะห์ระบบพบว่าการตั้งคำถามควรแบ่งออกเป็นกลุ่ม เพื่อให้ง่ายต่อการจำกัดขอบเขตของปัญหา สำหรับงานทะเบียนได้วิเคราะห์กลุ่มคำถามได้ทั้งสิ้น 12 กลุ่มได้แก่ การขอเปิดรายวิชา,การลงทะเบียน,การถอนรายวิชา เป็นต้น ซึ่งหลักการทำงานของระบบจะให้ผู้ขอคำปรึกษาจะต้องทำการเข้าสู่ระบบโดยใช้บัญชีรายชื่อของมหาวิทยาลัย เพื่อการแยกแยะบุคคลจากนั้นจึงทำการบันทึกเสียงผ่านทางเว็บบราวเซอร์(Web Browser) จากนั้นทำการอัพโหลดไฟล์เสียงเข้าสู่ระบบ เพื่อให้เจ้าหน้าที่ฟังและทำการตอบ โดยการตอบของเจ้าหน้าที่สามารถทำได้ทั้งการตอบด้วยเสียงหรือข้อความ นักศึกษายังสามารถติดตามสถานะการให้คำปรึกษาผ่านทางระบบได้ การประเมินความพึงพอใจของผู้ใช้ระบบได้ทำการประเมินแบ่งเป็น 7 ด้านระดับความพึงพอใจ 5 ระดับ พบว่าผลจากการประเมินมีค่าเฉลี่ยความพึงพอใจอยู่ที่ 4.07 และส่วนเบี่ยงเบนมาตรฐาน 0.75 ผลการสรุปการให้บริการระบบขอคำปรึกษาผ่านเว็บแอพพลิเคชั่นมีความพึงพอใจอยู่ที่ระดับดี

ABSTRACT. งานวิจัยนี้มีจุดประสงค์เพื่อสร้างแบบจำลองการฟุ้งกระจายสำหรับฝุ่นละอองอนุภาคขนาดเล็กไม่เกิน 2.5 ไมครอน (PM2.5) ที่เกิดจากการเผาเศษเหลือทิ้งทางการเกษตรในพื้นที่จังหวัดนครสวรรค์ ด้วยโมเดล NRC และ Gaussian Plume โดยใช้การทดสอบทั้งแนวดิ่ง และแนวระนาบ ในการทดสอบแนวดิ่ง จะทดสอบที่ระยะ 600 เมตร เหนือพื้นดินเหมือนกันทั้ง 2 โมเดล เนื่องจากเป็นระยะความสูงที่ให้ผลลัพธ์ดีที่สุด แต่ตัวแปรสำหรับการทดสอบในแนวระนาบนั้น จะใช้ค่าตัวแปรที่ไม่เหมือนกัน เพื่อให้เห็นความแตกต่างอย่างชัดเจนในแต่ละโมเดล ในส่วนโมเดล NRC สามารถใช้ทดสอบกับแรงลมเบาได้เพียงอย่างเดียวนั้น กำหนดให้ใช้ตัวแปรระยะทางทดสอบ 2 ระยะ คือ 100 และ 1,000 เมตร โดยผลที่ได้สำหรับระยะ 100 เมตร ค่าความเข้มข้นอยู่ที่ 0.0815 ไมโครกรัมต่อลูกบาศก์เมตร และเมื่อระยะทางห่างออกไปเป็น 1,000 เมตร อัตราความเข้มข้นลดลงเหลือเพียง 0.0793 ไมโครกรัมต่อลูกบาศก์เมตร สำหรับโมเดล Gaussian Plume สามารถใช้ทดสอบกับแรงลมได้หลายระดับ ประกอบด้วย เบา ปานกลาง และแรง ใช้ค่าตัวแปรทดสอบในแนวระนาบ 3 ระยะ ได้แก่ 5 10 และ 100 เมตร โดยผลที่ได้ สามารถจำแนกได้ดังนี้ ระดับลมเบา ค่าความเข้มข้นอยู่ที่ 0.1940 ไมโครกรัมต่อลูกบาศก์เมตร 0.1183 ไมโครกรัมต่อลูกบาศก์เมตร และ 0.0023 ไมโครกรัมต่อลูกบาศก์เมตร ระดับลมปานกลาง ค่าความเข้มข้นอยู่ที่ 3.1042x10-5 ไมโครกรัมต่อลูกบาศก์เมตร 0.0057 ไมโครกรัมต่อลูกบาศก์เมตร และ 0.0013 ไมโครกรัมต่อลูกบาศก์เมตร และระดับลมแรง ค่าความเข้มข้นอยู่ที่ 0.00002 ไมโครกรัมต่อลูกบาศก์เมตร 0.0030 ไมโครกรัมต่อลูกบาศก์เมตร และ 0.0007 ไมโครกรัมต่อลูกบาศก์เมตร

ABSTRACT. จากสถานการณ์การแพร่ระบาดของโรคติดเชื้อไวรัสโคโรนา 2019 (COVID-19) ที่มาพร้อมกับมาตรการการควบคุมการติดต่อของโรค ซึ่งนำออกมาบังคับใช้โดยรัฐบาล ได้ส่งผลกระทบต่อการดำเนินชีวิตประจำวันของประชาชนคนไทยเป็นอย่างมากเพราะต้องเจอกับภาวะกดดันต่างๆ ความเครียดและความวิตกกังวล โดยงานวิจัยชิ้นนี้ ผู้วิจัยได้ทำการเปรียบเทียบความวิตกกังวลที่เกิดจากโรคติดเชื้อไวรัสโคโรนา 2019 (COVID-19) ระหว่างสหราชอาณาจักรและประเทศไทยว่ามีความเหมือนหรือแตกต่างกันอย่างไร ในการวิเคราะห์ความรู้สึก ซึ่งเป็นการเก็บรวบรวมข้อมูลที่มาจากการพูดคุยสื่อสารผ่านเว็บบอร์ด และในการวิจัยครั้งนี้ผู้วิจัยใช้ข้อมูลจากเว็บ pantip.com รวบรวมกระทู้ที่เกี่ยวข้องกับโรคติดเชื้อไวรัสโคโรนา 2019 และมาตรการควบคุมโรคจากรัฐในช่วงเดือนมีนาคม 2563 ถึงพฤษภาคม 2563 โดยเป็นการใช้คำสำคัญหรือ Keyword ในการสกัดชื่อหัวข้อกระทู้ที่ใช้ได้ออกมา 10,746 กระทู้ และกระทู้ที่ถูกรวบรวมมาในขั้นแรกก็จะถูกนำมาตรวจสอบอีกครั้งว่าเนื้อหาในกระทู้สามารถนำมาใช้ในการวิเคราะห์ต่อไปได้หรือไม่ ซึ่งหัวข้อกระทู้ที่ถูกเลือกมาใช้จะถูกนำมาวิเคราะห์ด้วยการนับคำที่แสดงอารมณ์ต่างๆ โดยข้อมูลของไทยจะใช้แพลตฟอร์ม AI for Thai แพลตฟอร์มนี้ให้บริการปัญญาประดิษฐ์ (AI) ภายใต้แนวคิด “AI สัญชาติไทย” โดยสกัดเป็นตัวอิโมจิแสดงอารมณ์ก็จะถูกนำมาแปลงเป็นคำพูด และจากผลลัพท์ที่ได้พบว่าคนไทยตอบสนองทางด้านอารมณ์ต่อเรื่องดังกล่าวด้วยอารมณ์เศร้ามากที่สุด ซึ่งผลลัพธ์ท์ที่ได้จะถูกนำไปเปรียบเทียบกับข้อมูลของสหราชอาณาจักร โดยข้อมูลในส่วนของสหราชอาณาจักรนั้นเป็นรวบรวมกลุ่มตัวอย่างซึ่งเป็นผู้ที่อาศัยอยู่ในสหราชอาณาจักรจำนวน 2,500 คน โดยการให้ผู้ตอบแบบสอบถามเขียนข้อความสั้น และข้อความแบบยาวรวมเป็น 5,000 ตัวอย่าง ว่ามีความรู้สึกอย่างไรต่อสถานการณ์การแพร่ระบาดและมาตรการของรัฐในขณะนั้น โดยนำข้อความสกัดออกมาเป็นอารมณ์ผ่านเทคนิค LIWC และผลลัพท์ที่ได้พบว่าโดยส่วนใหญ่แล้วมีความรู้สึกกังวลมากกว่าความรู้สึกแบบอื่นซึ่งมีความแตกต่างจากของไทย อย่างไรก็ตามความเศร้าหรือความกังวลที่เกิดขึ้นส่วนใหญ่ล้วนมีผลต่อเรื่องงานและรายได้เหมือนกันทั้งสองประเทศ แต่ของไทยมีเรื่องหนี้สินเข้ามาเป็นอีกปัจจัย จึงทำให้เกิดความเศร้ามากกว่าแค่กังวลและจากการวิจัยในครั้งนี้ ผู้วิจัยเห็นว่าเป็นเรื่องที่สำคัญที่จะศึกษาและวิเคราะห์เกี่ยวกับอารมณ์ของประชาชน เพื่อให้ทราบถึงผลกระทบกับประชาชนได้รับทางด้านอารมณ์ ซึ่งอาจส่งผลกระทบต่อการดำเนินชีวิต เมื่อต้องเจอกับการระบาดของโรคร้ายแรงอื่นๆในอนาคต ทั้งนี้หน่วยงานหรือองค์กรที่มีบทบาทในการออกแบบนโยบาย สามารถนำข้อมูลดังกล่าวไปวางแผน ปรับใช้ในมาตรการเพื่อส่งผลกระทบกับสุขภาพจิตของประชาชนน้อยที่สุด

ABSTRACT. งานวิจัยนี้นำเสนอการพัฒนาต้นแบบเครื่องคัดแยกพุทราอัตโนมัติด้วยการประมวลผลภาพโดยใช้เทคนิคการเรียนรู้เชิงลึกมีวัตถุประสงค์เพื่อสร้างชุดอุปกรณ์วัดขนาดอัตโนมัติที่สามารถคัดแยกขนาดของพุทรา ด้วยวิธีการประมวลผลภาพโดยใช้เทคนิคการเรียนรู้เชิงลึก วิธีการดำเนินวิจัยแบ่งออกเป็น 5 ขั้นตอน ได้แก่ การเก็บรวบรวมข้อมูล การเตรียมข้อมูล การวัดประสิทธิภาพ การสร้างแบบจำลอง และการประเมินผล ผลการทดลองแสดงให้เห็นว่าการประมวลผลภาพโดยใช้เทคนิคการเรียนรู้เชิงลึก มีประสิทธิภาพในการจำแนกพุทรา มีความถูกต้องร้อยละ 80 และมีความรวดเร็วในการคัดแยกผลไม้กว่าการคัดแยกด้วยมนุษย์

ABSTRACT. เทคโนโลยีคลังข้อมูลมีประโยชน์ต่อการบริหารจัดการข้อมูลที่มีปริมาณมาก โดยมีวัตถุประสงค์เพื่อการวิเคราะห์ชุดข้อมูลที่มีในองค์กร เพื่อการกำหนดยุทธศาสตร์ และทิศทางขององค์กรตนเอง ซึ่งในบทความวิจัยนี้ได้นำเสนอกระบวนการการพัฒนาคลังข้อมูลตามวิธีการของ Ralph Kimball ตั้งแต่ การศึกษาแหล่งข้อมูล กระบวนการในการสกัด แปลง และถ่ายโอนข้อมูล (ETL) จนกระทั่งได้มาซึ่งคลังข้อมูลสรุปงบประมาณ ทั้งนี้เพื่อแก้ไขปัญหาในการเรียกดูรายงานสรุปงบประมาณของคณะวิทยาศาสตร์ มหาวิทยาลัยนเรศวร ซึ่งประสบปัญหาในการเรียกดูรายงานจากฐานข้อมูล OLTP ซึ่งเป็นฐานข้อมูลที่ใช้บริหารจัดการด้านการเบิกจ่ายงบประมาณที่มีจำนวนข้อมูลมาก โดยนำคลังข้อมูลซึ่งเป็นฐานข้อมูลแบบ OLAP มาใช้งานแทนในส่วนของการออกรายงานเพื่อวิเคราะห์และติดตามผลการดำเนินการเบิกจ่ายได้สะดวกยิ่งขึ้น ผ่านหน้าต่างระบบต้นแบบที่ผู้วิจัยได้พัฒนาขึ้น อีกประการหนึ่งในอนาคตยังสามารถนำเทคโนโลยีคลังข้อมูลไปใช้ในการวิเคราะห์กับงานด้านอื่นๆภายในคณะวิทยาศาตร์ฯได้เช่นกัน

ABSTRACT. การสูญเสียที่เกิดจากผู้ขับขี่รถจักรยานยนต์ที่ไม่สวมหมวกกันน็อกถือว่าเป็นปัญหาที่สำคัญเพื่อลดการสูญเสียและลดโอกาสที่จะได้รับการบาดเจ็บที่บริเวณศีรษะได้อย่างมีประสิทธิภาพ ดังนั้นผู้ขับขี่จึงต้องสวมหมวกกันน็อก และในปัจจุบันได้มีการพัฒนาระบบตรวจจับผู้ไม่สวมใส่หมวกกันน็อกด้วย Deep Learning แต่อย่างไรก็ตามประสิทธิภาพที่ได้ขึ้นอยู่กับจำนวนข้อมูลที่ใช้ในการฝึกสอน ดังนั้นในวิทยานิพนธ์นี้เราจึงทำการเพิ่มจำนวนข้อมูลที่ใช้ในการฝึกสอนด้วยวิธีการ Data augmentation โดยหลังจากการเพิ่มข้อมูลการฝึกสอนส่งผลให้ประสิทธิภาพของระบบตรวจจับผู้ไม่สวมใส่หมวกกันน็อกด้วย Deep Learning ที่อยู่ในช่วง 90-95 เพิ่มเป็น 99.3%