This paper proposes a framework to assist in managing predictive maintenance by detecting progressive surface wear on spur gears through the analysis of digital images of gear teeth using computer vision (CV) techniques.
An experimental setup was constructed to capture images of gear teeth using endoscopic cameras. The images were selected, pre-processed, stored in a database and used in the experimental study of the proposed framework. Three CV techniques were explored within the framework for detecting wear in spur gears: (1) edge detection; (2) gray level co-occurrence matrix (GLCM) combined with machine learning (ML) algorithms and (3) deep learning with convolutional neural networks (CNN).
The results showed 85% accuracy using the edge detection algorithm. Among the ML algorithms, accuracy was above 60% for the support vector machine (SVM) and above 70% for K-nearest neighbors (KNN). Principal component analysis (PCA) indicated that as the distance between the principal components increased, it characterized the formation and progression of surface wear on the gear teeth. With the CNN, an accuracy of 99.999981% was achieved in the training loss rate, with a classification accuracy rate (CAR) of 91.6666%, an F1 score of 90.9090% and a recall of 83.3334% during the testing phase.
This framework is applicable to a variety of gear systems and industrial contexts requiring predictive maintenance, making it a highly scalable solution for industry professionals.
This paper proposes a novel framework that considers various CV techniques to detect and assess the level of wear on spur gear surfaces. Moreover, the results provide guidelines for selecting the most appropriate method for detecting wear in gear systems.
