Abstract
To ensure the reliability of sensors, it is very important to study Fault Diagnosis (FD) methods for sensors. This study puts forward a Convolutional Neural Network (CNN)-FD model with randomly discarding network units. This model reduces the huge computational burden caused by excessive parameters in the CNN through silent neural nodes to achieve efficient automatic extraction and analysis of fault features. Considering the signal data imbalance caused by small sample failures, this study used Generative Adversarial Networks (GANs) to achieve intelligent expansion of samples. The performance test results showed that the proposed model achieved the highest classification accuracy in binary classification tasks, with a size of 93.5%, which is 5.5% and 3.5% higher than the Densenet model and ResNet model, respectively. In multi-classification tasks, the model still realized the best classification accuracy, with a size of 89.9%, which is 8.8% higher than the Densenet model. The experimental results of fault detection denoted that the proposed model arrived the highest recognition accuracy in aging failure and sensitivity failure, with 96.8% and 92.6% respectively, while the recognition accuracy of shallow neural networks and deep confidence networks was lower than 90%. This indicated that the proposed algorithm can effectively perform feature extraction and fault pattern recognition.