Research on fault diagnosis method of MS-CNN rolling bearing based on local central moment discrepancy

Highlights

• A multi-scale convolutional neural network (MS-CNN) model for transfer learning is proposed.

• A domain alignment method based on local central moment discrepancy in class subspace is proposed.

• Compared with the central moment discrepancy (CMD), the local center moment discrepancy (LCMD) performs better in transfer diagnosis.

Abstract

Transfer learning is an excellent approach to deal with the problem that the target domain label can not be adequately obtained when rolling bearing cross-condition fault detection. A transfer learning fault diagnosis method of multi-scale CNN rolling bearings based on local central moment discrepancy is presented in this research. The method maps bearing vibration data to a shared space by building a shared multi-scale feature extraction structure and fully connected layers. The source domain label and target domain pseudo-label are used to divide the category subspace in the shared space. And then the local central moment discrepancy is used to match source and target domain in the category subspace to realize fault knowledge transfer under different conditions. The experimental findings reveal that multi-scale CNN migration diagnosis based on local central moment discrepancy has superior accuracy and stability in diverse diagnostic tasks when compared to classic transfer learning approaches.

Introduction

In rotating machinery and equipment, a rolling bearing is an important supporting component [1], [2], [3]. The easiest way to prevent irreparable damage to rotating machinery caused by abnormal rolling bearing operating states is to detect the problem early and distinguish between the fault kinds [4], [5], [6]. Equipment failure detection and diagnosis is a very challenging problem [7], [8]. Researchers have proposed a variety of methods to solve the problem of equipment fault diagnosis, including time–frequency analysis [9], [10], fuzzy-based control [11], sparse signal decomposition [12], etc. In current years, the intelligent fault detection approach based on a data-driven has received considerable attention [13], [14]. The data-driven intelligent fault detection method could gain knowledge from the collected equipment operation data. Then applied the learned knowledge to perform an intelligent fault diagnosis of the equipment. Effectively converting the fault diagnosis problem, which previously required professional skills and extensive experience, into a classification problem [15]. Until present, researchers have presented a variety of intelligent defect diagnostic models based on deep learning (DL), such as convolutional neural networks (CNN) [16], Continuous Deep Belief Network (CDBN) [17], Residual building unit, Soft thresholding and Global context (RSG) [18]. The DL model can better adapt to diverse learning tasks and extract feature information from fault data in an adaptable approach.

The traditional DL approach offers high accuracy in fault detection. Unfortunately, to train a failure diagnostic model based on DL, a substantial amount of labeled data from target equipment is required. In an actual industrial environment, bearing equipment is frequently subjected to different working conditions. This makes obtaining a large number of bearing data sets with comprehensive fault categories challenging [19], [20]. When the rotational speed and load of the equipment change, for example, the characteristic distribution of the gathered data will depart from the prior state's characteristic distribution. The diagnostic system's detection performance will suffer as a result of these alterations [21], [22]. The abovementioned limitations prevent intelligent fault techniques from being used in real-world situations. The huge quantity of labeled data that may be taught in practice is usually derived from data collected during operations in a variety of settings. As a result, researchers have developed intelligent fault diagnoses based on transfer learning (TL) [23], [24].

The approach of TL has been successfully confirmed in fault diagnosis at this time [25], [26], [27]. The goal of TL is to improve the generalization task between source domain (D^s) and target domain (D^t) [28]. The model trains on data from the D^s try to learn certain common data features, and then builds a model with better generalization abilities to test data from the D^t [29]. A definition of transfer learning was put out by Pan et al. in their review [28]. TL's goal is to provide models the ability to apply information acquired from one activity to another. TL involves two key ideas, namely domain and task. Typically, we use D for the domain and T for the task. With the help of the knowledge in the given D^s, the learning of D^t is realized. Under specific situations, the domain adaption (DA) strategy can help the model extract transferable information in the D^s, according to the researchers [30], [31], [32], [33]. In the field of fault diagnostics, the Maximum Mean Discrepancy (MMD) has been frequently employed [34]. DA approaches based on MMD commonly use a kernel function to map the D^s and D^t data into a high-dimensional space and then minimize the distance between the two domains. Schwendemann presented a Layered MMD and CNN based bearing defect diagnostic system [35]. To accomplish the cross functioning condition diagnostics of bearings, Che et al. suggested a DBN network based on multi-kernel MMD (MK-MMD) [36]. To overcome the problem of data offset between D^s and D^t, Wu et al. presented an adversarial domain adaptive CNN structure based on MMD [37]. Zheng et al. proposed a locality-preserving projection-based transfer method for embedding data into subspaces and using MMD to reduce distributional differences between different datasets [38].

Rolling bearings usually work under complex conditions, such as changes in speed and load. When the working conditions change, the condition monitoring and fault diagnosis of the equipment will face more difficult challenges. Although the above intelligent defect detection approach based on transfer learning has a high detection ability, finding a suitable mapping kernel function when utilizing MMD for measurement learning is critical. The relevant kernel function must be updated when the data methods of the D^s and D^t are adjusted. Furthermore, measuring in a high-dimensional space will lengthen the model's computation time. To address the above-mentioned issues, this research offers a TL fault diagnostic technique based on the Central Moment Discrepancy (CMD) [39]: multi-scale CNN (MS-CNN) migration fault detection of rolling bearings based on the Local CMD (LCMD). The approach locates the same subdomain, in which samples with the same label but from distinct domains are found. In this domain, the difference in central moment across domains is therefore reduced. The following are the primary contributions of this paper.

• The approach based on MS-CNN-LCMD extracts information that can be transferred between D^s and D^t more effectively.

• The suggested technique makes greater use of transfer learning across multiple domains by avoiding the need for numerous kernel parameter selections.

• LCMD performs better in cross-domain diagnosis than classic CMD metric learning.

The rest of this manuscript is laid out as follows. Some fault diagnostic research on CNN and DA are reviewed in Section 2. The proposed approach is described in detail in Section 3. Section 4 delves into the details of the experiments and their outcomes and provides a discussion of the results. Finally, Section 5 is the conclusion.