BP Neural Network Modeling of TIG Welding Deposited Weld Width

Literature Overview and Research Background

The deposited weld width is a critical geometric parameter that directly influences the mechanical properties, fatigue performance, and service life of weldments. In TIG welding, the weld width is governed by a complex interplay of process parameters including welding current, travel speed, arc length, torch angle, and shielding gas composition. Traditional approaches to predicting weld width rely on empirical formulas or extensive experimental matrices, which are time-consuming and resource-intensive. This literature presents a backpropagation (BP) neural network model for predicting TIG welding deposited weld width, offering a data-driven alternative that can rapidly predict weld geometry from process parameters without the need for physical experimentation.

Core Technical Findings

The study develops and validates a BP neural network model for predicting TIG weld width, with the following key findings:

• Model architecture: A three-layer feedforward neural network with 5 input neurons (welding current, travel speed, arc length, torch angle, and gas flow rate), 10 hidden neurons, and 1 output neuron (weld width) achieves the best prediction accuracy.
• Training and validation: The model is trained on 120 experimental data points and validated on 30 independent test samples, achieving a mean absolute percentage error (MAPE) of 3.2% and a coefficient of determination (R²) of 0.96.
• Parameter sensitivity: The neural network analysis reveals that welding current is the most influential parameter on weld width, followed by travel speed, arc length, torch angle, and gas flow rate.
• Prediction range: The model is validated for welding currents of 80–200 A, travel speeds of 50–200 mm/min, and arc lengths of 2–6 mm, covering a wide range of practical TIG welding applications.

Neural Network Model Development

Model Architecture and Training

The BP neural network model is structured as follows:

The training process employs the Levenberg-Marquardt algorithm with a learning rate of 0.01, momentum of 0.9, and a maximum of 5000 iterations. The dataset is divided into 70% training, 15% validation, and 15% testing subsets to prevent overfitting and ensure generalization capability.

Input Parameters and Data Distribution

Prediction Accuracy Analysis

The model performance is evaluated across the parameter space:

Engineering Practice Implications

Application in Welding Procedure Development

The BP neural network model can be integrated into the welding procedure development process to:

1. Reduce experimental iterations: Predict optimal parameter combinations for target weld widths without extensive trial-and-error experimentation
2. Optimize process parameters: Identify parameter sets that achieve desired weld geometry while minimizing heat input and distortion
3. Support digital twin development: Provide real-time weld width prediction for in-process monitoring and control
4. Accelerate procedure qualification: Reduce the time and cost associated with welding procedure qualification (WPQ) testing

Quality Control Integration

The model can be incorporated into the quality control workflow as follows:

1. Pre-weld planning: Input target weld width and other geometric requirements to obtain recommended process parameters
2. In-process monitoring: Use real-time sensor data (current, travel speed, arc voltage) to predict weld width and adjust parameters as needed
3. Post-weld verification: Compare predicted and actual weld widths to identify systematic deviations and calibrate the model
4. Process optimization: Analyze model predictions across the parameter space to identify robust operating windows and avoid parameter combinations that produce unacceptable weld geometry

Key Reflections and Study Insights

The most significant engineering insight from this literature is the demonstration that a relatively simple BP neural network architecture can achieve high prediction accuracy for weld width across a wide parameter range. This finding validates the potential of data-driven approaches for welding process modeling and optimization, offering a practical alternative to traditional empirical formulas and extensive experimental matrices.

However, the study also highlights important limitations and considerations for practical implementation. The model's accuracy is bounded by the quality and coverage of the training data, and extrapolation beyond the validated parameter range is not reliable. Furthermore, the model does not capture the underlying physical mechanisms governing weld formation, which limits its ability to predict weld properties beyond geometry (such as mechanical properties, residual stresses, and microstructure). Engineers must therefore use the model as a complementary tool rather than a replacement for fundamental process understanding.

From a broader perspective, the findings underscore the growing importance of data-driven approaches in welding engineering. As sensor technology and data acquisition capabilities improve, the availability of high-quality training data will increase, enabling more sophisticated and accurate models. The BP neural network model presented in this literature represents a practical first step toward intelligent welding process optimization, and future work should explore more advanced architectures and multi-output models that can simultaneously predict weld geometry, mechanical properties, and quality metrics.