Abstract
Most existing laser welding process monitoring (LWPM) technologies focus on detecting post-process defects. However, in sheet metal laser welding applications such as welding of electronic consumer products during mass production, in-process defect detection is more important. In this article, a compact LWPM system using multi-sensor data fusion to detect in-process defects has been built. This system can collect the time series of plasma intensity, light intensity and temperature data for feature analysis. To verify the system’s effectiveness, a plasma-light-temperature dataset has been compiled, which consists of 5,836 samples of nine classes, including one positive class and eight negative classes of typical in-process defects. A multi-sensor data fusion network based on a convolution neural network for in-process defect detection, called IDDNet, has also been proposed. Experimental results have demonstrated that IDDNet can achieve better multi-classification results than the support vector machine, with an overall accuracy of 97.57%. In particular, considering this monitoring process as a binary classification problem, IDDNet can achieve a 99.42% accuracy. Moreover, IDDNet can reach an average speed of 0.79ms per sample on a single GTX 1080ti graphics card, which meets the real-time requirement for industrial production. The proposed LWPM system has been successfully verified in real applications of sheet metal laser welding.
Introduction
Laser welding is a cornerstone of modern electronics and automotive manufacturing, prized for its high precision and speed. However, the process is inherently volatile. Minor deviations—such as a tilted workpiece, surface impurities like glue or water, or a laser defocusing—can lead to significant structural failures.
Traditionally, manufacturers have relied on post-process monitoring, which inspects parts after production is complete. This approach risks "batch failure," where a machine misalignment at the start of a shift results in hundreds of defective parts before detection. While visual camera-based systems offer real-time monitoring, they are often computationally expensive, data-intensive, and susceptible to interference from the intense laser light. This research introduces a compact, high-speed solution designed for robust in-process defect detection.
System Design and Methodology
Hardware and Sensor Fusion
The proposed system utilizes a hardware-software co-design integrated into a Han’s Laser welding machine. Rather than relying on bulky cameras, the system integrates three lightweight sensors to capture critical process data:
- Photodiodes: Two sensors measure plasma intensity and reflected light intensity.
- Pyrometer: One sensor monitors the thermal radiation (temperature) of the weld pool.
IDDNet Architecture
To process the multi-sensor data, the researchers developed the In-Process Defect Detection Network (IDDNet), a 1D Convolutional Neural Network (CNN). Unlike traditional machine learning methods like Support Vector Machines (SVM) that require manual feature engineering, IDDNet automatically extracts complex patterns from raw time-series data. The architecture consists of three convolution blocks—incorporating Batch Normalization and ReLU activation—followed by Global Average Pooling, optimized for 128-point time-series signals.
Experimental Setup and Dataset
The system was evaluated using the Plasma-Light-Temperature Dataset (PLTD), which contains 5,836 samples. The dataset covers nine distinct classes: one "Qualified" class for perfect welds and eight specific defect types, including Defocus, White Glue, Missing Weld, Drift, Tilt, Repetition, and Water contamination. Each sample comprises three parallel time-series signals, each 128 points in length.
Results and Discussion
Classification Performance
IDDNet demonstrated superior performance compared to traditional methods. While SVMs achieved approximately 83% accuracy, IDDNet reached a multi-class accuracy of 97.57%. When simplified to a binary classification problem (detecting any defect vs. a qualified weld), the system achieved 99.42% accuracy.
Analysis revealed that while temperature is the most predictive single variable (~90% accuracy), the fusion of all three sensors is essential for reaching the 97%+ threshold. Plasma and light intensity provide critical context that thermal data alone lacks.
Real-Time Industrial Viability
For industrial applications, processing speed is as critical as accuracy. IDDNet processes a single sample in just 0.79ms on a standard GTX 1080ti GPU. Given that the welding process itself occurs on a millisecond scale, this performance allows for near-instantaneous defect flagging, preventing the production of substandard parts in high-speed lines.
Conclusion
The proposed LWPM system successfully demonstrates that multi-sensor data fusion combined with deep learning can provide a robust, real-time solution for laser welding quality control. Future work will focus on improving model robustness across varying power levels and speeds, as well as developing closed-loop systems that can automatically adjust laser parameters to prevent defects as they begin to form.
Summary updated