Leveraging machine learning for predicting and monitoring clogging in laser cladding processes: An exploration of neural sensors

Abstract

This study addresses the development of smart neural sensors to predict the powder mass flow and track clogging in real time during laser cladding. The challenges posed by powder granulometry and challenging environmental conditions that can lead to delivery failures are considered. An extensive experimental setup was conducted that included manipulation of key factors, such as laser power, travel speed, Z-step, N-layers, nozzle-to-substrate distance, and two types of process patterns. The mass flow rate of the powder was used as an independent variable to evaluate the predictive ability of the neural sensor with respect to the mass flow rate. Several models were trained and evaluated with different datasets and images of the cladding equipment. The model that integrated all data and images showed the best accuracy and precision also showed a strong predictive power for real-time estimation of the powder mass flow rate. Considering two practical rules—an error detection time of no more than one second and a confidence interval of less than 1.8 g/min—two strategies were proposed to meet these criteria. The first recommends the use of the comprehensive “all-features” model, while the second proposes a simplified model (with Z-step, N-slices, and the external camera as inputs) for efficient real-time error detection. The study provides an understanding of powder clogging prediction in laser cladding and suggests strategies for leaders in the field. Future research should validate these results and test these models in different environments to predict complex cladding properties and support the development of stand-alone laser cladding systems.