An Anomaly Detection Method Based on Continuous Wavelet Transform and Deep Convolutional Generative Adversarial Network that Only Requires Normal Sample Training

This paper proposes a new solution to the problem of anomaly detection of key components of new-type major equipment in the industrial field, especially the anomaly detection of rolling bearings. Due to the scarcity of abnormal samples of new-type major equipment, traditional fault diagnosis methods based on deep learning are difficult to train effectively, limiting their application in the new-type equipment’s fault diagnosis. Although some methods based on generative adversarial networks (GANs) do not rely on training based on abnormal samples in anomaly detection, determining the anomaly threshold is still challenging, and most methods still require a small number of abnormal samples to determine the threshold. To overcome this challenge, this paper proposes a method combining continuous wavelet transform (CWT) and deep convolutional generative adversarial network (DCGAN) to train the model and determine the threshold without abnormal samples. In this paper, the normal vibration signal of the bearing is firstly transformed by continuous wavelet transform to extract the time-frequency feature map, and the size is compressed to fit the deep learning model. Subsequently, the feature map is input into the DCGAN based on variational autoencoder (VAE) for training. The reconstructed features generated by the trained decoder and its statistical characteristics (mean and 3 times the standard deviation) determine the anomaly judgment threshold. In addition, the CBAM attention mechanism is introduced to enhance the network’s attention to low-frequency features and improve the feature extraction and reconstruction performance. Experimental results show that this method can effectively train the model and perform anomaly detection without abnormal samples, with precision, recall, and F1 values all reaching over 99%. This achievement provides an effective technical means for anomaly detection of key components of newtype major equipment.