Stacked Auto-encoder Based Feature Transfer Learning and Optimized LSSVM-PSO Classifier in Bearing Fault Diagnosis

Abstract

This paper proposes a new diagnosis technique for predicting the big data of roller bearing multi-level fault, which uses the deep learning method for the feature representation of the vibration signal and an optimized machine learning model. First, vibration feature extraction by stacked auto-encoders (VFE-SAE) with two layers in roller bearing fault signals is proposed. The unsupervised learning algorithm in VFE-SAE is used to reveal significant properties in the vibration data, such as nonlinear and non-stationary properties. The extracted features can provide good discriminability for fault diagnosis tasks. Second, a classifier model is optimized based on least squares support vector machine classification and particle swarm optimization (LSSVM-PSO). This model is used to perform supervised fine-tuning and classification; it is trained with the labelled features to identify the target data. Especially, using transfer learning, the performance of the bearing fault diagnosis technique can be fine-tuned. In other words, the features of the target vibration signal can be extracted by the learning of feature representation, which is dependent on the weight matrix of hidden layers of the VFE-SAE method. The experimental results (by analyzing the roller bearing vibration signals with multi-status fault) demonstrate that VFE-SAE based feature extraction in conjunction with the LSSVM-PSO classification is more accurate than other popular classifier models. The proposed VFE-SAE – LSSVM-PSO method can effectively diagnose bearing faults with 97.76 % accuracy, even when using 80 % of the target data.