WaveSpectro-XAF: Dual-branch deep learning with STFT spectrograms for automotive brushed DC motor noise classification

This paper presents WaveSpectro-XAF, a dual-branch deep learning model that fuses raw waveforms and STFT spectrograms through cross-attention for DC-motor noise classification. Using 4008 recordings across OK, NG High Noise, NG Vibration, and NG Abnormal classes, the model was evaluated under three STFT windowing schemes. WaveSpectro-XAF achieved the highest performance among all baselines, reaching 0.981 mean F1, 0.989 best F1, and above 99.9% AUC. Statistical tests (ANOVA, Kruskal–Wallis, post-hoc analyses) confirm significant robustness. Results demonstrate the model’s suitability for automated, reliable Industry 4.0 motor noise diagnostics.