Enhanced analog circuit fault diagnosis via continuous wavelet transform and dual-stream convolutional fusion

Abstract

Abstract Analog circuit fault diagnosis is crucial for ensuring the reliability and safety of electronic systems. To overcome the limitations of traditional methods, this study proposes a novel analog circuit fault diagnosis method based on Continuous Wavelet Transform (CWT) and Dual-Stream Convolutional Neural Network (DSCNN). The method uses CWT to convert raw fault waveform data into two-dimensional time–frequency images and employs a one-dimensional convolutional neural network (1D-CNN) to extract temporal features and a two-dimensional convolutional neural network (2D-CNN) to extract image features, achieving feature fusion. Additionally, the model incorporates a Convolutional Block Attention Module (CBAM), which includes channel and spatial attention modules, to enhance the model’s expressive power. Experiments on the Sallen–Key band-pass filter circuit, four-op-amp biquad high-pass filter circuit, and Tow-Thomas filter circuit validate the effectiveness of the proposed method, demonstrating excellent fault classification accuracy. Their classification accuracies reached 1.0000 ± 0.0000, 99.66% ± 0.0016, and 0.9771 ± 0.0023, respectively. Under various SNR conditions, our proposed model consistently maintains the highest classification accuracy with minimal impact from SNR variations. Furthermore, detailed practical experiments on the four-op-amp biquad high-pass filter circuit show that this model outperforms 1DCNN, CWT-CNN, and ISSA-SVM by 3.85%, 5.50%, and 6.39%, respectively, further proving the model’s superior feature extraction capability.