Prediction of Welding Distortion Using Autoencoders with Thermal and Stress Data in Deep Learning Framework

Various welding defects remain crucial challenges in manufacturing; particularly welding distortion degrades the structurality of welded components.. Predicting welding distortion using thermal and stress data, this work propounds a deep learning system using SSAE. The study hinges on the GTAW tool and the mild steel workpieces were chosen because they are widely used in industries. The values for model training were gathered from finite element simulations and experimental tests for thermal and stress parameters when welding differently. SSAE model was completed with the help of TensorFlowadoop preprocessing and analysis was done using the numpy, pandas. Training and evaluation were also performed on NVIDIA CUDA-Enabled GPU for improved computation. Finally, the proposed model of the research is static, dynamic, and sound, with a prediction accuracy of 98.2%, mean absolute error of 0.016 mm and root mean squared of 0.019 mm. Comparison of the proposed SSAE model with ANNs and SVMs showed higher accuracy of the SSAE model with similar amounts of data and fewer computations. The performance of the model was further verified for various unseen circumstances of heat input and material properties where the overall predicted errors did not exceed 4%. Such results suggest that the developed framework is rather good at generalization to various welding regimes. With the help of deep learning approaches in combination with high-dimensional thermal and stress data, this study provides a sound approach for the predictive modeling of welding distortion with immediate application in the context of intelligent manufacturing systems.