173 / 2025-06-11 15:59:29

Research on Sensor Fault Isolation Method for Transmission Systems Based on Convolutional Autoencoder

deep learning, transmission system, sensor, fault isolation

    Transmission systems are usually applied in industrial scenarios with high load, long time continuous operation or complex environmental disturbances, and the sensors are prone to stuck, constant deviation, constant gain and mutation faults, which lead to the distortion of monitoring data and threaten the reliability of the system. Aiming at the lack of isolation capability of traditional methods for high-dimensional nonlinear dynamic data and concurrent multi-type faults, this paper proposes a sensor fault isolation method that integrates convolutional autoencoder (CAE), attention mechanism and support vector machine (SVM). The method firstly utilizes the unsupervised learning ability of CAE to automatically extract the spatio-temporal features of the sensor time series data; secondly, it introduces the attention mechanism to dynamically strengthen the key fault features and inhibit the noise interference, which significantly improves the discriminative property of the features; finally, it realizes the accurate classification and isolation of faults by means of the powerful nonlinear classification ability of SVM. The experimental results show that the model has an average accuracy of 98% and F1-score of 0.96 in the typical fault detection of pressure, temperature and speed sensors, which provides a new technical solution for the isolation of transmission system sensor faults under complex working conditions.