A Novel Meta-Learning Network With Adaptive Input Length and Attention Mechanism for Bearing Fault Diagnosis Under Different Working Conditions

In recent years, rolling bearing fault diagnosis technology based on deep learning (DL) has provided a more intelligent and reliable method for the safe operation of mechanical systems due to its powerful feature learning ability. However, in real industrial scenarios, the acquisition of fault samples is limited and knotty, which makes it difficult for DL methods that require a large number of fault samples to be successful. To overcome the above problems, in this article, a novel meta-learning network with adaptive input (AI) and attention mechanism is proposed for rolling bearing fault diagnosis with small samples under different working conditions. First, inspired by the envelope demodulation signal processing method, an AI length selection strategy considering the different working conditions is proposed, which improves the disadvantages of Gram angle field (GAF) 2-D coding method with the traditional fixed input. Second, the residual structure and attention mechanism are introduced to make the network have stronger feature extraction and generalization performance and further improve the classification accuracy. Finally, the effectiveness of the method is verified on the Case Western Reserve University (CWRU) bearing fault datasets and the high-speed train axle box bearing fault datasets conducted by us. The results show that the proposed improved model-agnostic meta-learning (MAML) network is superior to the other four mainstream meta-learning methods under the same conditions, and satisfactory fault diagnosis results can be obtained on the bearing fault datasets under different working conditions.

I. Introduction

Rolling bearing is a kind of precise mechanical element running between shaft and shaft seat, which can transform sliding friction into rolling friction to reduce friction loss. Rotating machinery is very complex, usually working in high-temperature, high-pressure, and high-speed rotating environments. High-intensity motion and friction loss make rolling bearings vulnerable to damage [1], [2], [3]. Therefore, it is necessary to continuously monitor and diagnose the health status of rolling bearings. In the field of intelligent fault diagnosis, traditional diagnostic methods, such as artificial neural network (ANN) [4], support vector machine (SVM) [5], and extreme learning machine (ELM) [6], [7], need to use signal processing techniques to extract effective features from the original diagnostic signals. However, in actual cases, the fault vibration signal is usually nonstationary and has strong noise, which makes it difficult to design an effective feature extraction method manually. The bearing fault diagnosis method based on deep learning (DL) has made remarkable achievements with its automatic feature extraction ability and end-to-end approach in recent years [8], [9], [10]. Zhang et al. [11] designed an adaptive activation function with a slope and threshold of the tanh function (STAC-tanh) and combined the STAC-tanh with the deep residual network to achieve adaptive extraction of effective fault features. Shen et al. [12] proposed a deep multilabel learning framework called multilabel convolutional neural network (MLCNN), which can use missing label samples for network training by learning relevant features. Based on a simple spectrum matrix obtained by short-time Fourier transform (STFT), He and He [13] established an optimized DL structure, large memory storage retrieval (LAMSTAR) neural network, to realize bearing fault diagnosis.