Abstract
A steering actuator is the core component of an autonomous vehicle. Because of noise interference, the extraction of useful fault features for an accurate fault diagnosis of steering actuator becomes an arduous task. This article proposes an adaptive denoising residual network (AD-ResNet) for fault diagnosis of steering actuator with reduced noise interference and improved accuracy. First, a high-complexity model based on long short-term memory (LSTM) is built to estimate redundant signals, and then, fault features are generated by comparing redundant signals with the measured sensor signals. Second, an attention unit is designed to determine the unique set of thresholds for each fault feature. Finally, hard thresholding is embedded in a deep network as a nonlinear transform layer to reduce noise. Simulation experiments were performed at different signal-to-noise ratios (SNRs), and the results confirmed that the proposed method could adaptively denoise the high-noise signals. The proposed method was then validated against a fault dataset that was built using road experiments with an autonomous vehicle, and it achieved an accuracy, a sensitivity, a specificity, and an F1-score of 91.66%, 91.66%, 97.22%, and 0.9159, respectively, averaged over all fault states. The proposed method outperformed the comparison methods in the experiments.
Funding Information
  • Natural Science and Technology Special Projects (2019-1496)
  • Key-Area Research and Development Program of Guangdong Province (2020B0909050004)

This publication has 33 references indexed in Scilit: