Weld joint penetration state sequential identification algorithm based on representation learning of weld images

The penetration state of weld joints determines the quality of the weld, but it is limited by space and is often difficult to measure directly. It is typically estimated using available weld process information. This study proposes a Weld Joint Penetration State Sequential Identification algorithm based on the representation learning of weld images (WJPSSI) and makes contributions in the following two aspects: (1) Features are extracted from weld pool images using a self-supervised representation learning method under label-free conditions, addressing the challenge of label acquisition in industrial production; (2) A few-shot timing learning method is established to determine the timing correlation in the evolution process of the penetration state. By establishing a dual-camera visual sensing system with Pulsed Gas Tungsten Arc Welding (GTAW-P) as the application focus, the synchronous collection of weld images and the penetration states of weld joints is completed, and a dataset of sequential melting time — backside weld pool width under various weld conditions is created. This study utilizes an autoencoder model (New-AutoEncoder, NAE) based on the channel attention mechanism for self-supervised representation and feature extraction of the weld pool image. Additionally, a self-attention mechanism is introduced to develop an improved backside weld pool width monitoring model using the Attention-Gated Recurrent Unit (AGRU) for timely prediction of backside weld pool width and penetration state determination. Through model structure optimization, parameter training, and performance verification, the time-series prediction accuracy achieved an RMSE of 0.29 mm across two groups, marking a significant improvement over other prediction methods. Experimental results show that the proposed method sets a new benchmark in the few-shot sequential identification of weld joints.