Analog circuit fault diagnosis method using modified inception and effective spatial-channel attention

Abstract Circuit fault diagnosis based on deep learning is a crucial approach to mitigate risks and reduce losses. To extract the features of circuit compound faults and capture their correlations, a circuit fault diagnosis method using the Modified Inception (MInception) architecture and Effective Spatial-Channel Attention (ESCA) is proposed in this paper. First, an MInception feature extraction network is designed to extract features of voltage signals at different scales and depths. Second, an ESCA mechanism is constructed to enhance channel features and alleviate the gradient vanishing problem. Third, to adaptively focus on the most critical parts of input features with low computational cost, a Depthwise Pointwise Convolution Linear (DPCL) module is designed to assign higher weights to key channel features and positions. Finally, circuit fault diagnosis is accomplished through fully connected layers and a Softmax classifier. Given the wide application of Butterworth low-pass filters and the occurrence of compound faults (multiple faults happen simultaneously), Data acquisition with single faults and compound faults is performed on Butterworth low-pass filters. Experimental results demonstrate that the proposed model is effective in diagnosing the circuit faults, even for the compound faults. It can achieve an overall accuracy (OA) of 99.52%, which is 2.35% and 5.25% higher than those of the ASCAN method and the RNDCAM method respectively.