Rolling bearing fault diagnosis based on feature fusion with parallel convolutional neural network

摘要

Deep learning has seen increased application in the data-driven fault diagnosis of manufacturing system components such as rolling bearing. However, deep learning methods often require a large amount of training data. This is a major barrier in particular for bearing datasets whose sizes are generally limited due to the high costs of data acquisition especially for fault scenarios. When small datasets are employed, over-fitting may occur for a deep learning network with many parameters. To tackle this challenge, in this research, we propose a new methodology of parallel convolutional neural network (P-CNN) for bearing fault identification that is capable of feature fusion. Raw vibration signals in the time domain are divided into non-overlapping training data slices, and two different convolutional neural network (CNN) branches are built in parallel to extract features in the time domain and in the time-frequency domain, respectively. Subsequently, in the merged layer, the time-frequency features extracted by continuous wavelet transform (CWT) are fused together with the time-domain features as inputs to the final classifier, thereby enriching feature information and improving network performance. By incorporating empirical feature extraction such as CWT, this proposed method can effectively enable deep learning even with dataset size limitation in practical bearing diagnosis. The algorithm is validated through case studies on publicly accessible experimental rolling bearing datasets. A wide range of dataset sizes is tested with cross-validation, and influencing factors on network performance are discussed. Compared with existing methods, the proposed approach not only possesses higher accuracy but also exhibits better stability and robustness as training dataset sizes and load conditions vary. The concept of feature fusion through P-CNN can be extended to other fault diagnosis applications in manufacturing systems.