Research on Improved Residual Network Classification Method for Defect Recognition of Thermal Battery

Thermal battery is an ideal power supply for military applications such as artillery and ship equipment. Due to the sheet-type process of the thermal battery, various installation error defects occur in the assembly of thermal battery. Aiming at the problems of low efficiency and low defect-recognition rate of thermal battery detection, a thermal battery defect detection model is proposed based on residual network. First, the squeeze-and-excitation networks (SENet) structure based on the attention mechanism is introduced into residual block of the residual neural network, the connection between the feature extraction channels is established, and the improved deep residual network I-ResNet50 is obtained; Second, in order to prevent overfitting, the defect images processed in the production line and the laboratory are data-enhanced and labeled. Transfer learning strategy is introduced into the recognition model I-ResNet50, and then the training set data samples are input into the recognition model I-ResNet50 for training, and the activation function LReLu and Dropout skills are introduced to improve the classification ability of the I-ResNet50 model; Finally, the recognition model I-ResNet50 is applied to the test set and validation set, and each defect of the thermal battery are output. Comparison experiments are tested under different migration strategies and different optimizers and learning rates, and comparison experiments with the five classic network structures of ResNet50, YOLOV3, MobileNetV2, VGG16, and YOLOV4 are also tested. The test data show that the recognition accuracy rates of qualified images and the three types of defective images (Qualified Assembly, Missing Current Plate, Wrong Number of Stacks, and Reverse Stack) can reach 99.64%, 98.17%, 99.11%, and 95.40%, respectively, the overall recognition accuracy rate can reach 98.10%. The test results illustrate the model can detect thermal battery defects more accurately and quickly, and has good defect diagnosis ability, which is nearly 5% higher than the traditional method, and a new solution for defect detection in practical industrial scenarios of thermal battery is provided.