Data-Driven Feature Extraction for Analog Circuit Fault Diagnosis Using 1-D Convolutional Neural Network

Abstract

The present study applies the one-dimensional convolutional neural network (1D-CNN) to propose an intelligent approach of the feature extraction for the analog circuit diagnosis. The raw signals based on various soft faults from the output terminal of the circuit under test (CUT) are collected with appropriate data acquisition system to implement a data-driven fault diagnosis. The data-driven diagnosis process is typically encapsulated in two distinct blocks, including the feature extraction and the classification. In this study, the designed 1D-CNN model efficiently combines the aforementioned two phases into a single diagnosis body with fast learning rate and accurate classification. The main advantages of the 1D-CNN are: 1) it can be directly established to the raw signal with proper training so that it is more applicable in real applications; 2) its compact architecture and configuration has reasonable applicability in complex analog circuits; 3) convolutional kernels guarantee that the hierarchical features can be extracted from raw data with better anti-interference performance. Moreover, since the method can extract high-level features of raw signals, it resolves the necessity to employ other per-processing methods for the hand-crafted feature transformation. The performance of the proposed 1D-CNN model is evaluated through three benchmark circuits on the SIMULINK platform. Obtained results are compared with other intelligent fault diagnosis methods. The experimental results show that the 1D-CNN can be utilized effectively as the feature exactor and faults classifier for analog circuits.