An ensemble attention-based residual convolutional neural network for rolling bearing fault diagnosis

Abstract

Rolling bearings are essential components in mechanical systems, whose fault diagnosis is vital for operational efficiency. But in real industrial environments, the harsh conditions, including noise, missing data, and compound faults, severely limit the diagnostic performance of the algorithm. Thus, we propose an ensemble attention-based residual convolutional neural network (CNN) optimized by the vortex search algorithm. First, a new residual CNN with the improved residual structure, the separable convolution, and the global average pooling layer is designed to extract features from the vibration signals automatically. Second, a residual cooperative attention mechanism is presented. To guarantee the difference between the base models, different base models are constructed employing multiple convolutional kernels, activation functions, as well as attention mechanisms, respectively. And different training sets are allocated to each base model by Bootstrap. Third, a new exponential threshold decision fusion strategy is put forward to achieve ensemble learning. Eventually, the vortex search algorithm is employed to optimize the parameters of the decision fusion strategy. The noise, missing data, and compound fault datasets constructed separately using data from two rolling bearing experiments reveal that the proposed ensemble model can effectively overcome the limitations of individual models and achieve superior fault identification performance than existing methods under many types of severe conditions.