| Summary: | The increased demand for predictive monitoring in Laser Beam Welding (LBW) process for quality management has led to the development of systems utilising various signals with the combination of acoustic and optical signals being particularly cost-effective and less complex. However, further study is needed to explore the combination of both approaches, as the effect of signal noise and selectable features on the classification model remains unclear, with limited findings supporting the effectiveness of this combination. This project aims to develop a classification model to estimate the weld penetration conditions of LBW through the analysis of signals using the proposed method for denoising and feature extraction. To achieve the aims of this research, an experiment was conducted using Boron Steel (22MnB5) with two different thicknesses, 1.2 mm and 1.8 mm. The experiment involved manipulating laser power (280 – 320 W) and weld speed (1.0 – 1.5 mm/s) to obtain several penetration conditions, namely, full penetration, half penetration, incomplete penetration and overheat penetration. The acoustic signal and optical spectrum were acquired during the process. The study found that the acoustic average amplitude and optical spectrum intensity are linearly proportional to the penetration depth. Due to process instability, a significant amount of spatters appeared, exhibiting non-stationary random noise in both signals, which led to a high Misclassification Rate (MCR) for both signal features. To overcome the issue, EMDbased Denoising with Z-score threshold and Spectrum Averaging was implemented. The result showed that out of the 15 features analysed, 12 showed improvement after denoising, with the highest improvement recorded to be 68.17%. The feature selection analysis identified Intensity and Log Different Absolute Standard Deviation (LDASD) as the most significant features in the original signal, while Intensity and Absolute Value of Summation of Square Root (AVSSR) were identified as the most significant features in the denoised signals. When comparing the original model to the denoised version, the average accuracy improved from 97.4% to 99.8% in the training set and from 95.5% to 99.0% in the test set. Additionally, the precision improved by 53.3% in the training set and 11.17% for the test set, indicating greater consistency and reliability in the denoised model. This research successfully enhanced the denoising method for acoustic and optical signals which improved the effectiveness of signal features for prediction models with notable feature enhancements such as Intensity (68.17%), RMS (31.51%), and AVSSR (26.98%). By developing classification models using both original and denoised signals, the research achieved performance gains with the denoised model reaching 99.8% accuracy in the training set and significant improvements in precision which underscore the critical role of signal preprocessing in optimizing machine learning models for the LBW process. This research is expected to advance the LBW process by providing a predictive monitoring system that enhances quality and efficiency with potential benefits for defect detection across various industries, ultimately leading to cost savings and improved product quality.
|
|---|