Bearing Fault Diagnosis With Incomplete Training Data: Fault Data With Partial Diameters

Existing data-driven bearing fault diagnosis studies are based on strong assumptions: complete fault samples are required. The number of fault data can be more or less, but the data of each fault class must be available. However, such a condition is difficult to meet in the industry. Therefore, this paper addresses an open issue: bearing fault diagnosis with incomplete training data. In other words, only partial fault data are available in the training process. This issue is more in line with the industrial situation, and the issue is worthy of in-depth research. In response to this issue, the Cepstrum-Scale-Distance based Framework (CSD-Framework) is proposed, including C-stage, S-stage, and D-stage. The three stages realized vibration signal transformation, multi-scale adaptive adjustment, and multi-metric distance matching, respectively. This is a general framework, suitable for analyzing vibration signals, and is convenient to be combined with advanced AI algorithms. On this basis, the Multi-Metric-Adaptive-Clustering (MMA-Clustering) algorithm and the Multi-Metric-Weight-Classify (MMW-Classify) algorithm are proposed to form the D-stage of CSD. The proposed method has three advantages: 1) generic; 2) scalable; 3) good ability to classify unseen data. Experimental results showed that the performance of CSD was better than a variety of existing AI algorithms, as well as Ceps-AI methods based on cepstrum and AI algorithms.
Note to Practitioners
—Diagnose bearing faults by using incomplete fault data (only partial fault diameters). Existing approaches require a complete dataset, that is, each fault (diameter) needs to be available, and to achieve accurate classification of high-similar data through strong learning ability, but it is helpless for unseen fault diameters. This paper proposes a fault diagnosis framework based on cepstrum, which focuses on highlighting key fault features with cepstrum analysis technique, and realizes fault diagnosis based on incomplete data by constructing a multi-scale and multi-metric adaptive distance matching method. The framework is validated on a public dataset and a real-world dataset, where only partial fault data (fault diameter) is required for training.