Fault Diagnosis for Hydraulic Servo System: A Stacked Denoising Autoencoder Method based on Self-Learning of Robustness Features

摘要

The fault diagnosis of hydraulic servo system attracts more attention in complex system prognostics and health management. As the precondition of most fault diagnosis methods, feature extraction could efficiently draw information from the initial data and supply more evidence to the following results. However, traditional time-frequency analysis largely depend on artificial selection and optimization, which are generally limited by the quality of input data and working environment. Thus this paper proposes a stacked deep learning based model to represent robust feature information in terms of the advantage of cognitive computing and pattern classification theory, which is shown to be suitable for certain applications with inevitable ambient noise and working condition fluctuations. To effectively realize feature reconstruction, a stacked denoising autoencoder (SDA) in which multiple encoders are established and trained is used. The employed deep neural network is trained layer by layer to extract high-level features, where the sparsity representation is applied to map the original inputs to better high-level features. Considering better robustness of the learnt features to avoid external interferences, the original input neural parameters of each autoencoder are denoised by randomly assigning some units to be zero. The modified denoising autoencoders are then stacked to initialize the deep hierarchical architecture instead of the original. High-level feature representations of the monitoring data samples are obtained based on unsupervised self-learning, and are set as the inputs of a top fault pattern classifier for final training, followed by a fine-tuning process. Validation data are collected to facilitate the comparison and evaluation of the fault diagnosis results of the SDA models, of which the denoising proportion is different for each. Experiments show an obvious advantage of the SDA model based on the self-learning of the robustness features for fault pattern classification.