Multi-step prediction of automobile rear axle assembly torque based on adaptive time series data decomposition hybrid deep learning model

Abstract

The prediction of assembly torque of automobile rear axles is crucial for the timely detection of assembly abnormalities and early feedback control to reduce the number of production line shutdowns. However, due to the complex assembly conditions, the assembly torque change has non-stationary and abnormal mutation characteristics, and the conventional prediction model is difficult to accurately predict. Therefore, this article proposes an adaptive time series data decomposition hybrid deep learning assembly torque multi-step prediction framework based on the GWO-VMD-CNN-BiLSTM-Attention model. Firstly, the preprocessed original assembly torque sequence will be decomposed into multiple subsequences by an improved adaptive variational mode decomposition method based on grey wolf optimization (GWO). Then, the subsequence is subjected to 1D-CNN convolution to obtain local features. The BiLSTM layer is used to predict the feature data output by the CNN layer, and the Attention mechanism is added to weight the output data of the BiLSTM layer to extract the torque change characteristics in different time steps. Finally, iterative multi-step prediction is performed on each subsequence of assembly torque. Six-steps prediction experiments were carried out on the actual production data of the actual automobile rear axle assembly process. The highest accuracy of the proposed prediction method reaches 98.5%, 97.8%, 95.4%, 91.3%, 89.7%, and 86.5%, respectively. The experimental results show that the proposed model has higher prediction accuracy than other benchmark models. The model is coupled with the digital twin system and deployed, which realizes the real-time prediction of assembly torque and has a high industrial application value.