Novel Key-Value Pair Attention Neural Network Based on Trend and Period Features for Fault Diagnosis

In recent years, the task of preventing and resolving major safety risks of hazardous chemicals has still difficult, and fault diagnosis has become an important link to improve the safety of the petrochemical equipment. Traditional fault diagnosis methods mainly relied on static estimation of system parameters, which are measured and estimated during the operation of the equipment. However, industrial process data is dynamic and time-varying, so fault characteristics are easily submerged in dynamic time-varying information. Therefore, a novel key-value pair attention neural network based on trend and period (KATP) features is proposed to solve the time-varying problem of industrial process data. A dynamic variational mode decomposition (VMD) is proposed to solve the most suitable number of modes and penalty factors in the decomposition process. Furthermore, an adaptive key-value pair Attention module is used to dynamically extract the trend cycle characteristics of the signals decomposed by the VMD. Meanwhile, the key-value pair attention mechanism enables the adaptive key-value pair Attention module to dynamically adjust the trend and periodic features of the Attention according to the current input by creating key and value pairs, and fuse these features to generate the final representation. Moreover, the Softmax function implements the mapping from fused features to fault categories for fault diagnosis. Finally, the proposed method is applied in the fault diagnosis of the Tennessee Eastman (TE) dataset and the actual catalytic cracking unit reaction regeneration system. The experimental results show that compared with other fault diagnosis methods in terms of the accuracy, the fault diagnosis performance of the KATP is improved by 10.3% and 15.5%, respectively.