Fault Diagnosis Model via Vibration Signal Analysis With an Improved BKA-VMD and CNN-TELM Hybrid Framework

Abstract

Rolling bearings are fundamental components of contemporary machinery, yet their prolonged usage frequently leads to wear, performance deterioration, and potential faults. In scenarios characterized by limited sample sizes and complex, noisy environments, traditional diagnostic methods often encounter difficulties achieving satisfactory fault identification results. To address these challenges, this study introduces an innovative approach for rolling bearing fault diagnosis. Initially, the black-winged kite algorithm (BKA) is enhanced through the integration of a differential evolution strategy and an iterative search method, enabling the precise determination of optimal parameters for variational mode decomposition (VMD). Subsequently, a comprehensive index evaluation criterion is established to identify the optimal signal components, which are then subjected to a detailed analysis to extract diverse sensitive features, ultimately forming a hybrid feature set. To further improve the accuracy and efficiency of fault diagnosis, this study proposes an enhanced extreme learning machine model, termed twin extreme learning machine (TELM). Moreover, the TELM model is seamlessly integrated into the architecture of a convolutional neural network (CNN), specifically as a component of its output layer, resulting in a novel hybrid fault diagnosis model. Rigorous data validation performed on a rolling bearing testbed underscores that the proposed fault diagnosis model significantly surpasses conventional approaches, including SVM, KELM, ELM, LSTM, and softmax, in terms of accuracy, recall, and F1 score. Notably, the model maintains robust fault diagnosis capabilities even in environments with varying degrees of noise interference.