Online prediction of mechanical and electrical quality in ultrasonic metal welding using time series generation and deep learning
Honghuan Chen,Xin Dong,Yaguang Kong,Zhangping Chen,Song Zheng,Xiaoping Hu,Xiaodong Zhao
DOI: https://doi.org/10.1016/j.engfailanal.2024.108162
IF: 4
2024-03-02
Engineering Failure Analysis
Abstract:Ultrasonic metal welding (UMW) is a reliable solid-state joining technique used in manufacturing lithium batteries and wire terminals. Current research leverages deep learning methods to predict tensile strength of UMW joint. However, there is a gap in studying the prediction of contact resistance in UMW joints, which is a critical factor in the performance of lithium batteries and wire terminals. We propose a novel online method for predicting both the tensile strength and contact resistance in UMW. To enhance prediction accuracy, we use Variational Mode Decomposition (VMD) with Particle Swarm Optimization (PSO) to extract vibration signals and employ the Spectrum Customized Denoising Diffusion Probabilistic Model (SCDDPM) for time series data generation. Firstly, given the non-stationary nature of the raw vibration signal, VMD method is used for effective signal decomposition. To optimally separate effective vibration signals from interference, PSO is used to automatically determine the ideal decomposition number K and penalty factor α in VMD. Secondly, Considering the high-frequency and large data size of vibration signals, as well as the low-frequency of power and pressure signals, we develop a spectrum customized time series signal generation method named SCDDPM. This method overcomes the mode-collapse issue encountered with Generative Adversarial Networks and increases the number of abnormal samples. Lastly, a 17-layer MobileNetV2 network is used to extract features from vibration, power, and pressure. These extracted features are then concatenated for UMW quality prediction. Our proposed method has been evaluated in predicting two key quality aspects of UMW: contact resistance and tensile strength. It excels in predicting tensile strength with a RMSE of 10.11 and an R2 of 0.85, outperforming existing best methods by reducing RMSE by 30.75% and increasing R2 by 13.33%. Moreover, this is the first instance of successfully predicting contact resistance in UMW joints.
engineering, mechanical,materials science, characterization & testing