Multiple faults diagnosis for an industrial robot fuse quality test bench using deep-learning

Abstract

Diagnosing faults occurring in industrial equipment and monitoring its status is crucial for maintaining the continuous operation of the equipment. The conventional diagnostic approaches, such as model-based methods, need to model the dynamics of the physical process; however, obtaining an accurate dynamic model of complex interconnected industrial systems poses a significant challenge. Modern diagnostic methods based on data-driven using artificial intelligence models rely heavily on sensor data to perform fault detection and diagnosis. Conventional machine learning methods have low detection accuracy and rely on domain knowledge to extract meaningful features from data acquired from the equipment. On the other hand, deep learning models can automatically extract useful features from fault data with high diagnostic performance. This paper proposes a two-dimensional Convolutional Neural Network (2DCNN) diagnostic model to effectively improve the diagnostic accuracy of predicting faults on a tabular dataset with multiple fault classes. A simple sliding window approach is proposed to effectively transform and reshape the features in the dataset as inputs to the proposed model architecture. The method's feasibility was evaluated using data from an industrial robotic fuse quality test bench. Experimental results show high diagnostic performance with an average accuracy of 99.98% compared to conventional methods commonly used for diagnostics. Moreover, validation of the diagnostic model on the experimental dataset using raw multi-sensor fusion from a robot manipulator platform was carried out. The results demonstrate the model's outstanding diagnostic performance, with an average testing accuracy of 99.74% for four fault classes.