A lightweight evidence-guided acoustic diagnostic network for automotive component fault detection in maintenance scenarios

Non-contact acoustic diagnosis is attractive for automotive maintenance because abnormal sounds can be recorded without installing additional sensors on compact or inaccessible components. Practical service recordings, however, are affected by low-frequency dominance, transient impacts, device differences, propagation paths, and background noise. This study proposes a lightweight evidence-guided convolutional neural network (CNN)-graph convolutional network (GCN) for 12-class automotive component acoustic fault detection. Long-window short-time Fourier transform (STFT), short-window STFT, and log-Mel spectrograms are combined as a three-channel time-frequency representation to preserve low-frequency spectral details, impact-related temporal changes, and perceptual acoustic information. A dual-branch network is then used: the CNN branch extracts local time-frequency textures, while the GCN branch treats temporal frames as graph nodes and constructs a sample-adaptive Top-K graph to model non-local temporal relations. To reduce dependence on noise-related shortcuts, a frequency-band attribution-prior loss is introduced using class-specific fault-sensitive bands estimated from the training data. Under a source-group-level evaluation protocol, the proposed model achieves 0.9749 segment-level accuracy, 0.9749 macro- F 1, and 0.9976 macro-AUC on an independent test set. Ablation, noise robustness, and prototype deployment tests indicate the value of the proposed representation, local-global architecture, and evidence-guided loss for maintenance-oriented acoustic diagnosis.