A multi-scale frequency domain information-driven method for ultrasonic welding quality prediction

Ultrasonic welding is extensively used in industries such as automotive and electronics, where the tensile strength of the welding joints serves as a critical indicator of manufacturing quality, yet is subject to fluctuations caused by multiple uncertain factors during production. Moreover, the complex physical interactions of the process produce time series features rich in high-frequency ultrasonic information, which is critical for achieving accurate predictions of quality performance. This work develops a deep learning framework that accurately predicts joint tensile strength by fusing multi-source signals. Specifically, a high-fidelity acquisition system was established to synchronously capture high-frequency vibration, pressure, and displacement signals during welding. The proposed Gated Multi-scale Fast Fourier Transform Network (GMS-FFTNet) operates directly in the frequency domain, integrating a global Fast Fourier Transform (FFT) to capture long-range dependencies and a local Short-Time Fourier Transform (STFT) to resolve transient details. It is enhanced by an adaptive spectral filtering mechanism that dynamically emphasizes critical frequency bands, overcoming the limitations of conventional time-domain models that often struggle with spectral information loss. To synergize the insights from different sensors, a dedicated gating network effectively fuses the multi-source signal features. The model accurately predicts tensile strength, achieving a coefficient of determination of 82.17%. Furthermore, the visualization of attention weights serves to correlate the model’s inference with the manufacturing process of the ultrasonic welding. This method not only presents a robust solution for reliable quality monitoring of ultrasonic welding but also offers a generalizable framework for other dynamic manufacturing processes with high-frequency information.