Limited Data Rolling Bearing Fault Diagnosis With Few-Shot Learning

Abstract

This paper focuses on bearing fault diagnosis with limited training data. A major challenge in fault diagnosis is the infeasibility of obtaining sufficient training samples for every fault type under all working conditions. Recently deep learning based fault diagnosis methods have achieved promising results. However, most of these methods require large amount of training data. In this study, we propose a deep neural network based few-shot learning approach for rolling bearing fault diagnosis with limited data. Our model is based on the siamese neural network, which learns by exploiting sample pairs of the same or different categories. Experimental results over the standard Case Western Reserve University (CWRU) bearing fault diagnosis benchmark dataset showed that our few-shot learning approach is more effective in fault diagnosis with limited data availability. When tested over different noise environments with minimal amount of training data, the performance of our few-shot learning model surpasses the one of the baseline with reasonable noise level. When evaluated over test sets with new fault types or new working conditions, few-shot models work better than the baseline trained with all fault types. All our models and datasets in this study are open sourced and can be downloaded from https://mekhub.cn/as/fault_diagnosis_with_few-shot_learning/. INDEX TERMS Deep learning, few-shot learning, bearing fault diagnosis, limited data.

A Novel Deep Learning Model for the Detection and Identification of Rolling Element-Bearing Faults

“…To validate the effectiveness of the novel dual-path recurrent neural network with wide first kernel and deep convolutional pathway proposed in this work for intelligent data-driven fault diagnosis applications, the benchmark bearing fault dataset from Case Western Reserve University (CWRU) Bearing Data Center [ 46 ] is used. The CWRU bearing fault dataset has been widely used in the literature for investigating both conventional fault diagnosis techniques [ 47 , 48 ] and data-driven fault diagnosis methods using machine learning [ 49 , 50 ] and deep learning [ 35 , 36 , 37 ]. The data used in this study were collected from both the drive end accelerometer (close to the fault location) and the fan end accelerometer (remote from the fault location) of the test apparatus shown in Figure 5 .…”

“…In particular, as shown in the confusion plots in Figure 9 , this domain shift can act to hinder classification when there are potential additional undiagnosed fault conditions. Following the approach in Zhang et al [ 37 ], this scenario considers the case where labeled fault data from multiple operating conditions is available. In this case, the model can be trained using data acquired under these multiple operating conditions to provide a more robust fault diagnosis system.…”

“…Zhuang et al [ 36 ] also proposed a 1D CNN capable of operating on raw vibration signals but used dilated convolutions and residual connections to improve robustness to noise and domain shifts at the cost of some complexity in tuning the model parameters. Zhang et al [ 37 ] adapted the WDCNN proposed in [ 35 ] to work with limited data—a common challenge in fault diagnosis where components are rarely allowed to run to failure.…”

One-Shot Learning for Partial Discharge Diagnosis Using Ultra-High-Frequency Sensor in Gas-Insulated Switchgear

“…In this paper, we propose a new one-shot learning model for fault diagnosis with a small number of PRPDs while using Siameses neural networks that have the advantage of reducing the parameters to train and avoiding the problem of over-fitting [ 29 ]. One-shot learning and few-shot learning can learn features when only a few labeled samples are provided [ 30 ]. The proposed method uses a Siamese structure consisting of two identical CNNs and a distance metric function [ 31 ].…”

“…The main contributions of this paper are summarized, as follows: One-shot learning is introduced for the first time to classify the PRPDs in a GIS. This method offers the advantages of a high classification accuracy while requiring a small amount of data compared with a linear SVM and CNN [ 30 ]. The proposed model uses pairs of samples of the same class or different classes during the training phase and recognizes the test sample with a single training sample for each class.…”