Method for Diagnosing Bearing Faults in Electromechanical Equipment Based on Improved Prototypical Networks

Abstract

Due to the complexity of electromechanical equipment and the difficulties in obtaining large-scale health monitoring datasets, as well as the long-tailed distribution of data, existing methods ignore certain characteristics of health monitoring data. In order to solve these problems, this paper proposes a method for the fault diagnosis of rolling bearings in electromechanical equipment based on an improved prototypical network—the weight prototypical networks (WPorNet). The main contributions of this paper are as follows: (1) the prototypical networks, which perform well on small-sample classification tasks, were improved by calculating the different levels of influence of support sample distributions in order to achieve the prototypical calculation. The change in sample influence was calculated using the Kullback–Leibler divergence of the sample distribution. The influence change in a specific sample can be measured by assessing how much the distribution changes in the absence of that sample; and (2) The Gramian Angular Field (GAF) algorithm was used to transform one-dimensional time series into two-dimensional vibration images, which greatly improved the application effect of the 2D convolutional neural network (CNN). Through experiments on MAFAULDA and CWRU bearing datasets, it was shown that this network effectively solves the shortcomings of a small number of valid samples and a long-tail distribution in health monitoring data, it enhances the dependency between the samples and the global data, it improves the model’s feature extraction ability, and it enhances the accuracy of model classification. Compared with the prototypical network, the improved network model increased the performance of the 2-way 10-shot, 2-way 20-shot, and 2-way 50-shot classification tasks by 5.23%, 5.74%, and 4.37%, respectively, and it increased the performance of the 4-way 10-shot, 4-way 20-shot, and 4-way 50-shot classification tasks by 12.02%, 10.47%, and 4.66%, respectively. Experimental results show that the improved prototypical network model has higher sample classification accuracy and stronger anti-interference ability compared with traditional small-sample classification models.

A Long-Tailed Image Classification Method Based on Enhanced Contrastive Visual Language

“…After extracting the features of each modality, use two transformation matrices W v ∈ R d v ×d and W l ∈ R d 1 ×d project the original visual and textual features into a shared embedding space, where v and u are d-dimensional normalized vectors, as shown in Equation (3).…”

“…However, this type of imbalance problem often makes the training of deep neural networks very difficult. Classification and recognition systems that directly use long-tailed distribution data for training often tend to lean towards the head class data, making them insensitive to tail class features during prediction and affecting the correct judgment of the system [ 3 ]. In traditional methods, a series of common methods to mitigate performance degradation caused by long-tailed distribution data are based on categories re-balancing strategy, including re-sampling training data and re-weighting to redesign the loss function [ 3 ].…”

“…Classification and recognition systems that directly use long-tailed distribution data for training often tend to lean towards the head class data, making them insensitive to tail class features during prediction and affecting the correct judgment of the system [ 3 ]. In traditional methods, a series of common methods to mitigate performance degradation caused by long-tailed distribution data are based on categories re-balancing strategy, including re-sampling training data and re-weighting to redesign the loss function [ 3 ]. These methods can effectively reduce the bias of the model to the head class in the training process, thus producing a more accurate classification decision boundary.…”

“…With the increase in the rarity of the category, the collection volume tends to show exponential growth, as shown in Figure 1. In practical applications, such as facial recognition, species classification, autonomous driving, medical diagnosis, drone detection, and other fields, there is a problem of long-tailed class imbalance [3]. For example, for autonomous driving, the data on normal driving will account for the vast majority, while there are very little data on actual abnormal situations/car accident risks.…”

“…After pre-training, model training is conducted separately for the head class and tail class, which can improve the performance of the tail class without sacrificing the performance of the head class. In practical applications, such as facial recognition, species classification, autonomous driving, medical diagnosis, drone detection, and other fields, there is a problem of longtailed class imbalance [3]. For example, for autonomous driving, the data on normal driving will account for the vast majority, while there are very little data on actual abnormal situations/car accident risks.…”