Estimation of Al/Cu laser weld penetration in photodiode signals using deep neural network classification

In Al/Cu overlap laser welding joints for electric vehicle batteries, power and load are transmitted through the connection between the upper and lower sheets. As a thin sheet is placed on the upper part of the overlap joint, the laser processing parameters should be precisely adjusted to avoid weld defects such as incomplete penetration into the lower part or excessive upper sheet melting. In this work, a support vector machine and two deep neural networks (i.e., a fully connected neural network and a convolutional neural network) were used to classify penetration as unsatisfactory, transient, or good in photodiode signals. A bandpass filter was applied to the photodiode signals to transmit only the Cu emission wavelength. For modeling, 405 datapoints were collected, and 283, 61, and 61 datapoints were used for training, validation, and testing of the models, respectively. The machine learning models predicted the penetration mode every 50 ms, and the test results showed a high classification performance, exceeding 90% accuracy. The convolutional neural network was verified experimentally by gradually increasing the laser output power, thus demonstrating the feasibility and applicability of neural network classification to estimate Al/Cu laser weld penetration.