A Practical Approach on Performance Assessment of Federated Learning Algorithms for Defect Detection in Industrial Applications

One of the most common problems to be solved in industrial environments is the detection of defects in the manufacturing quality control of different products. The design of such automatic systems, especially if they are based on image processing, often presents difficulties related to the availability of sufficient well- labeled data for initial training. Usually, it is not easy to have enough industrial samples with defects to have a sufficiently large and balanced dataset. This research work presents a new method of hybridisation of convolutional neural networks used for defect detection in industry by means of image processing, using Federated Learning (FL) techniques. This method is able to overcome the limitations and problems presented by this type of systems stated above. In this research, it is demonstrated with examples of defect detection in textures that it is possible to reach a defect detection effectiveness above 90% on average, in problems where no dataset is available, using federated learning algorithms with classifiers based on Convolutional Neural Networks previously trained in other problems of defect detection in other types of textures. The creation of such systems for error detection on previously untrained data using federated learning and obtaining this effectiveness represents an advance on the state of the art with respect to the existing approaches, and constitutes the main contribution of this research work.