Ultrasonic Metal Welding Quality Prediction Model Based on MHSA–LSTM–AHP

Abstract

Ultrasonic Metal Welding (UMW) technology is widely used in industries such as electric vehicle manufacturing owing to its efficiency, low heat input, and ability to join dissimilar metals. However, the weld quality of UMW technology is susceptible to various process parameters; thus ensuring consistency is challenging. A quality prediction model that combines the Analytic Hierarchy Process (AHP), Long Short-term Memory (LSTM), and Multi-head Self-attention (MHSA) was proposed in this study to evaluate the joint quality of UMW technology accurately. The UMW process was analyzed in four stages based on current production operations, with characteristic information extracted from the process characteristic data sampled by the in-line inspection equipment. A hierarchical framework was developed by drawing on the principles of the AHP to elucidate the interrelationships between the real-time process data, the process characteristics, and the UMW joint quality. An MHSA-LSTM-AHP quality prediction model was established in this study by harnessing the advantages of LSTM and MHSA in learning temporal dependencies. A comparative analysis was conducted by using the Genetic Algorithm-optimized Backpropagation (GA-BP) Neural Network and LSTM-AHP models to evaluate the performance of the proposed model. Results indicate that the proposed model performs well in predicting tensile strength and contact resistance, with a mean error of 3.21% and 3.7%, respectively. This study can provide a satisfactory reference for the construction of an online quality monitoring system and the optimization of the UMW process.