Single-event-latchup (SEL) in a semiconductor device is an undesirably induced high current state, typically rendering the affected device to be non-functional and compromising its operating lifetime. The lower-current SEL phenomenon – the micro-SEL – is often difficult to detect, particularly when the normal operating current of the protected device is variable and the magnitude of micro-SEL currents is different under different operating conditions. In Machine-Learning (ML), the said variable current inadvertently affects the multiple features of the input current profile required for micro-SEL detection, thereby severely reducing the detection accuracy. In this paper, we propose a data pre-processing module to improve the accuracy of the ML-based micro-SEL detection under the aforesaid current conditions. The proposed pre-processing module encompasses the following. Prior to classification by ML, the input current profile is processed by a data pre-processing module employing a proposed background subtraction algorithm and proposed adaptive normalization algorithm. By filtering the irrelevant base current and normalizing the micro-SEL current based on the base current value, the data pre-processing module provides improved accurate features of the input current profile and widens the difference between normal samples and micro-SEL samples in the feature space. Ultimately, the proposed module facilitates ML algorithms to generate a more accurate decision boundary. The outcome is a worthy ~13% accuracy improvement (from ~79% to ~92%) in the micro-SEL detection in a device operating with variable currents.