Time series anomaly detection remains challenging due to non-stationary dynamics, multi-scale patterns, and scarce labeled anomalies. Existing deep models often struggle to learn reliable normal behavior, particularly for short and low-dimensional time series. We propose an unsupervised architecture that combines a fixed one-level Haar wavelet decomposition with a shared-parameter bidirectional cross-attention encoder and branch-specific Temporal Convolutional Network (TCN) decoders. The Haar transform decomposes each input window into low-frequency (trend) and high-frequency (transient) components, while cross-attention mutually conditions the two bands to expose inconsistent dynamics that may indicate anomalies. Experiments on six public benchmarks show that our method outperforms twelve representative baselines on five datasets, achieving F₁scores of 100% on UCR and 2D-Gesture, 99.39% on Power Demand, 99.09% on ECG, and 99.85% on SMD, while remaining competitive on MSL (95.09%). The source code and pretrained checkpoints are available at https://github.com/NNNguyenDuyyy/ACIIDS_2026_Anomaly_Timeseries_Detection.git.