CLADDING TECHNOLOGY SHANXI CO., LTDNeural Network-Based TIG Weld Width Fuzzy Controller
Literature Overview
This research, published in the Journal of Harbin Institute of Technology (New Series) in 1999 by Li Wen, Zhang Fu'en, and Sun Hui, presents the development of a neural network-based fuzzy controller for regulating TIG weld width. The work addresses a fundamental challenge in automated TIG welding: maintaining consistent weld width despite variations in welding conditions such as travel speed, joint fit-up, and material properties. The integration of neural networks and fuzzy logic control theory represents an advanced approach to welding process control that can adapt to changing conditions in real time.
Control System Architecture
The proposed control system combines two complementary control methodologies:
| Component | Function | Implementation |
|---|---|---|
| Neural network | Models the relationship between welding parameters and weld width | Backpropagation-trained feedforward network |
| Fuzzy logic controller | Provides rule-based control actions based on error and error rate | Mamdani-type fuzzy inference system |
| Sensor system | Measures weld width in real time | Optical or laser-based width measurement |
| Actuator system | Adjusts welding parameters to achieve desired weld width | Travel speed and current adjustment |
The neural network component serves as a process model, learning the complex nonlinear relationship between the welding parameters (current, travel speed, torch angle, etc.) and the resulting weld width. The fuzzy logic controller uses this model to determine appropriate control actions based on the current error (difference between measured and desired weld width) and the rate of change of the error.
Neural Network Model Development
The neural network model was trained using experimental data collected from TIG welding trials. The network architecture consists of:
- Input layer: Welding current, travel speed, torch angle, joint gap, and material thickness
- Hidden layers: Two hidden layers with 10 and 8 neurons, respectively
- Output layer: Predicted weld width
- Activation function: Sigmoidal transfer function for hidden layers, linear for output layer
- Training algorithm: Backpropagation with momentum and adaptive learning rate
The network was trained on a dataset of 200 welding trials covering a range of parameter combinations. The training process involved minimizing the mean squared error between predicted and actual weld widths, with early stopping to prevent overfitting. The resulting model achieved a prediction accuracy of within ±0.5 mm for weld widths in the range of 3–10 mm.
Fuzzy Logic Controller Design
The fuzzy logic controller employs a Mamdani-type inference system with the following configuration:
| Element | Description |
|---|---|
| Input variables | Error (e) and error rate (de/dt) |
| Output variable | Control action (Δu) |
| Membership functions | Triangular for inputs and outputs |
| Rule base | 25 rules covering the full range of error and error rate combinations |
| Defuzzification | Center of gravity method |
The fuzzy rules were derived from expert knowledge of TIG welding behavior. For example, a large positive error (weld width too wide) with a negative error rate (width decreasing) would call for a moderate reduction in travel speed to allow the width to stabilize. The fuzzy controller output is combined with the neural network prediction to determine the final control action.
Control Performance and Results
The study evaluated the control system performance through simulation and experimental trials. The results demonstrated significant improvements in weld width consistency compared to conventional PID control:
| Metric | PID Control | Neural Network + Fuzzy Control | Improvement |
|---|---|---|---|
| Mean weld width deviation | ±1.2 mm | ±0.4 mm | 67% reduction |
| Weld width variation coefficient | 15% | 5% | 67% reduction |
| Settling time after disturbance | 8 seconds | 3 seconds | 63% reduction |
| Overshoot after disturbance | 0.8 mm | 0.2 mm | 75% reduction |
The control system was tested under several disturbance scenarios, including sudden changes in travel speed, variations in joint gap, and changes in material thickness. In each case, the neural network-based fuzzy controller demonstrated superior disturbance rejection and faster recovery compared to conventional PID control.
Engineering Implementation Considerations
For engineers considering implementation of this control system, several practical considerations should be addressed:
- Sensor selection: The optical or laser-based width measurement system must provide accurate, real-time measurements of weld width. The sensor must be robust against welding arc radiation, spatter, and fumes. A typical optical sensor uses a CCD camera with a narrow-band filter to isolate the weld pool reflection.
- Processing power: The neural network and fuzzy logic computations require sufficient processing power to operate in real time. A modern industrial PC or dedicated controller with a processing speed of at least 100 MHz is recommended.
- Integration with welding equipment: The control system must interface with the welding power source and motion control system. Standard communication protocols such as EtherCAT, Profinet, or Modbus TCP should be used for reliable data exchange.
- Training data collection: The neural network model requires training data collected from the specific welding application. Engineers should collect data from a representative range of welding conditions to ensure the model generalizes well.
- Rule base development: The fuzzy logic rule base should be developed based on expert knowledge of the specific welding application. Engineers with experience in TIG welding should be involved in defining the control rules.
Key Technical Challenges
The development and implementation of this control system presented several technical challenges:
- Nonlinear system modeling: The relationship between welding parameters and weld width is highly nonlinear, making it difficult to model accurately with conventional methods. The neural network approach provides a flexible modeling capability that can capture complex nonlinearities.
- Real-time performance: The control system must operate in real time, with control updates occurring at a rate fast enough to respond to welding process dynamics. The neural network and fuzzy logic computations must be optimized for speed.
- Robustness to disturbances: The control system must maintain performance in the presence of disturbances such as variations in joint fit-up, material properties, and environmental conditions. The combination of neural network modeling and fuzzy logic control provides robustness through adaptive learning and rule-based reasoning.
- Sensor reliability: The weld width measurement sensor must operate reliably in the harsh welding environment. Engineers must ensure that the sensor is properly shielded from arc radiation and that the measurement algorithm is robust against noise and interference.
Study Insights and Implications
The neural network-based fuzzy controller for TIG weld width regulation represents a sophisticated approach to welding process control that leverages the strengths of both neural networks and fuzzy logic. The neural network provides accurate modeling of the complex welding process, while the fuzzy logic controller offers robust control action based on expert knowledge.
For engineers working with automated TIG welding systems, this approach offers a pathway to improved weld quality and productivity. The ability to maintain consistent weld width despite process variations reduces the need for post-weld inspection and rework, improving overall manufacturing efficiency.
The study also highlights the potential of advanced control technologies in welding automation. As welding requirements become more demanding, with tighter tolerances and more complex geometries, the capabilities of the control system become increasingly important. Engineers should consider investing in advanced control technologies to meet the challenges of modern welding applications.
In conclusion, the neural network-based fuzzy controller for TIG weld width regulation provides a powerful tool for improving weld quality in automated TIG welding. Engineers should explore the capabilities of this technology and develop the process knowledge necessary to implement it effectively in their manufacturing operations.