Electronic Commutator Defect Detection Using Deeplabv3

In this work, we tackle the task of defect detection on the electronic switch of a battery by implementing the DeepLabv3 model which is a deep learning algorithm and we make two main contributions to the industrial domain: The first one is the identification and fast detection of faults of different shapes and types, the second one is the accuracy in fault identification For this, we improved or modified some parameters of our DeepLabV3 model established by the following authors [1] such as adopting Network Backbone MobileNet instead of xception, changing the num_class to 4 because we have 4 types of faults, using a downsample_factor of 16, freezing the weight to 0. 9e-4 and then thawing the weight to continue to increase the accuracy 4e-5 and running the cycle 500 times for good computational training. To measure the efficiency and accuracy of our model, we compared our results with other models like DeepLab v2, ResNet-101, ASSP (Actrous spatial pyramid pooling), Deep Layer cascade, CentraleSupelec Deep G-CRF, LRR_4x_ResNet-CRF, PSPNet on datasets like PASCAL VOG2012 and it appears that DeepLabv3 model mainly based on hole convolution obtains one of the best results with 85. 7% MoU followed by PSPN and 85.4% also for MIoU. This Deep LabV3 model has improved the original Deep Lab V3 model by increasing the detection accuracy by almost 80% compared to the original model, reducing the processing time, and allowing the identification of one or multiple defects simultaneously.