Analytical ferrography image recognition using convolutional neural network

From the viewpoint of preventive maintenance, the analytical ferrography method is widely used as a method to diagnose the wear condition of worn surfaces through wear particle analysis of lubricating oil, and production losses are prevented by replacing or renewing lubricating oil based on the analysis results. On the other hand, wear particle analysis by analytical ferrography requires determining in the wear classification and its cause by comprehensively judging the wear particle size, color, shape, and surface properties, so that the analysis needs specialist knowledge and often takes time. Therefore, attention was paid to convolutional neural networks (CNN), which are compatible with image analysis. Currently, studies evaluating the accuracy of wear particle image analysis models using CNN have been reported. Still, it is not clear what features of the wear particles are important in determining the wear mode from the wear particle image. Therefore, in this study, a CNN model that discriminates four types of wear modes was first created using the Neural Network Console, 100 single wear particles were analyzed, and it was shown that the CNN model in this study could distinguish wear modes. Next, visualizing the basis of AI's decision-making when identifying wear particles using Grad-CAM from the analysis results could demonstrate that AI would learn by focusing on the lightness and darkness of the color in the wear particles image analysis. Furthermore, concerning the lightness and darkness of the color, it was found that cutting wear particles and spherical wear particles were identified based on the black color in the image, flat wear particles were identified based on relatively light colors such as yellow, yellowish brown and light blue, and severe wear particles were identified based on relatively dark colors such as brown and dark blue.