Enhanced machine anomaly identification through a heterogeneous sensors fusion approach based on Dempster-Shafer theory of evidence

This study presents a heterogeneous sensor fusion approach that jointly exploits sensors that monitor machine conditions and those that assess product quality for enhanced anomaly identification. In industrial practice, these sensor categories are typically treated independently, and their joint use remains limited in current literature. The proposed method adopts a two-level probabilistic strategy for fault classification at the local level, Gaussian Process Classifiers are trained on features selected from each sensor; at the global level, the estimated class probabilities are fused using the Dempster-Shafer theory of evidence, with belief assignments incorporating model accuracy and confusion matrix statistics. The method was validated on an ultrasonic welding machine used for producing disposable hygiene products. Results show that the weighted Dempster-Shafer fusion approach achieves an average classification accuracy of 99.9% with a standard deviation of 0.1%, outperforming individual sensors and classical fusion methods. The technique also demonstrated robustness to noise, maintaining classification accuracy above 90% even when additive Gaussian noise with variance up to 100% of the original signal was applied to the test data.