Rolling bearing fault diagnosis based on probabilistic mixture model and semi-supervised ladder network

Abstract

Fault diagnosis of rolling bearings is of great significance to ensure the production efficiency of rotating machinery as well as personal safety. In recent years, machine learning has shown great potential in signal feature extraction and pattern recognition, and it is superior to traditional fault diagnosis methods in dealing with big data. However, most of the current intelligent diagnostic methods are based on the ideal conditions that bearing data set and label information are sufficient, which are often not always available in engineering practice. In response to this problem, this paper proposes to use probabilistic mixture model (PMM) to approximate the data distribution of the bearing signal, and then use Markov Chain Monte Carlo (MCMC) algorithm to sample the probabilistic model to expand the fault data set. In addition, Semi-supervised Ladder Network (SSLN) can achieve the effect of supervised learning classifier with only a few labeled samples. Based on Case Western Reserve University (CWRU) Bearing Database, the recognition accuracy of the proposed PMM-SSLN model can reach 99.5%, and the experimental results show that this model is applicable to the case where both bearing data set and label information are insufficient.

Research on Mechanical Fault Diagnosis and Prediction Technology Based on Deep Learning

“…For the purpose of enhancing the detection and anticipation of mechanical malfunctions, this investigation has embraced a synergistic deep learning paradigm that integrates the Convolutional Neural Network (CNN) with the Long Short-Term Memory (LSTM) model. This integrated approach leverages the strengths of CNN in discerning critical features alongside the LSTM's capacity to effectively manage sequential data, culminating in a robust analysis of mechanical vibration signals, as illustrated in the literature [7]. This model consists of two parts: convolution layer and LSTM layer (Figure 1).…”

Metric Learning-Guided Semi-Supervised Path-Interaction Fault Diagnosis Method for Extremely Limited Labeled Samples under Variable Working Conditions

“…23, 6951 17 of 28 based on the CNN has higher accuracy and smaller fluctuation than that based on the autoencoder, and we believe that this phenomenon is related to the excellent generalization capability of the shared convolutional kernel. From the perspective of algorithm structure, the ladder-shaped encoder-decoder architecture is able to exploit an enormous quantity of unlabeled data which is usually ignored, and thus the actual distribution of each fault can be obtained.…”

“…Furthermore, to alleviate the limited labeled problem, generating data with the same distribution of labeled data is regarded as an intuitive solution [ 8 ]. Ding et al [ 23 ] utilized the probabilistic mixture model and the Markov Chain Monte Carlo algorithm to expand the fault dataset, which could provide large amounts of fake data. Tao et al [ 24 ] generated pseudo-cluster labels for labeled and unlabeled data by adopting density peak clustering strategies.…”

A unified in-time correction-based testability growth model and its application on test planning

“…However, the truth is not what we expected. 7 In the process of BITE design, especially for the new and unproven system, it is difficult to obtain abundant training data. This is because that the system works in various environments and load stresses, but the design condition is limited by the design time cost and the fault injection risk.…”