Overcoming Limited Fault Data: Intermittent Fault Detection in Analog Circuits via Improved GAN

Owing to the characteristics of intermittent faults (IFs) such as short duration and randomness in analog circuits, it is difficult to collect a sufficient amount of manually labeled fault data for training detection model. To this end, this paper explores the application of generative adversarial network (GAN) in the field of IF detection for analog circuits. The proposed model consists of two networks, an auto-encoder and a discriminator, which are trained by competing against each other while working together to grasp the underlying concepts in the target. In addition, a spatial Fourier convolution (SFC) block is proposed and introduced into the discriminator to enhance the detection performance of the model. The training process is divided into two stages, the first stage through the auto-encoder and the discriminator of the adversarial learning, so that the discriminator has the initial fault detection ability. The second stage trains the auto-encoder and the discriminator with a few fault samples to enhance the fault detection capability of the two networks working cooperatively. The proposed method is applied to two typical analog circuits, namely, a power amplifier circuit and a low-pass filter circuit. The results show that, guided by only a small number of fault samples, the proposed method has a great advantage over other state-of-the-art detection methods.