Development of laser cleaning state classification model through the acquired acoustic signal using the empirical mode decomposition and one dimensional convolutional neural network

Abstract

Laser cleaning is an efficient, non-invasive method that utilizes high-energy laser beams to eliminate contaminants. However, variations in laser process parameters can lead to challenges such as inconsistent cleaning depth, thermal damage, and uneven surface treatment, ultimately compromising the quality of the cleaned surface. To address these issues, developing a predictive model for cleaning states is crucial to enhance online monitoring systems, enabling earlier detection of potential problems. This manuscript outlined the development of a classification model intended for predicting the states of laser cleaning by employing the EMD-1DCNN methodology. The primary objective of integrating Empirical Mode Decomposition (EMD) is to enhance the precision and reliability of the model generated from a one-dimensional Convolutional Neural Network (1D-CNN). The laser cleaning experiments were executed at velocities of 100 mm/s and 300 mm/s on corroded boron steel substrates. Acoustic signals within the frequency spectrum of 20 Hz to 10,000 Hz were systematically recorded during the entirety of the cleaning procedure.These signals were categorized into three phases, which were corrosion removal stage, low roughness formation, and engraving stage, which indicates the surface damage. The results show that the time-domain signal recorded a random non-linear pattern during the corrosion removal stage. The frequency was active at 6300 Hz for all laser cleaning conditions, but the peak amplitude decreased as the grooves started to form on the cleaned area. Instead, the peak at 1800 Hz was increased. However, the implementation of EMD revealed a significant trend that could separate corrosion removal and groove formation stage at another bandwidth, which was 20 Hz to 500 Hz. Moreover, the EMD-1D-CNN classification model achieved an average accuracy of 95.75% with a deviation of 1.99%, demonstrating enhanced performance compared to a model developed without EMD. This research highlights the importance of the classification model in predicting cleaning process stages, facilitating real-time monitoring and ensuring cleaning quality. The preprocessing methods employed not only enhanced model accuracy but also improved consistency, potentially reducing computational demands while fostering a stable model.