Analog circuit fault diagnosis method using auto-encoder based feature extraction

Motivated by the development of artificial intelligence, data-driven fault diagnosis for analogue circuits has become popular. Data-driven methods commonly include two stages: the feature extraction stage and the fault isolation stage. Traditional feature extraction methods, mainly using signal processing, suffer from the problem of information loss and are highly dependent on prior knowledge. In order to deal with the limitations above, a semi-supervised representation learning method is proposed for fault diagnosis. An autoencoder-based structure is adopted to establish a mathematical encoder. The proposed encoder can directly capture the latent feature from the original signal data without traditional data processing. And then a two-layer neural network is applied to further fault classification task. Experiments on two classical filter circuits are performed to demonstrate the effectiveness and merits of the proposed fault diagnosis method for analogue circuits.