An improved ResNet50 method for hazardous-region state recognition in CNC press brakes

To address the open die region, frequent human-machine interaction, and insufficient active recognition capability of conventional protection methods during sheet-metal bending with CNC press brakes, this study develops an improved ResNet50 model for hazardous-region state recognition. A visual monitoring region is defined according to the overlap between the tool motion range and the workpiece placement region, and an on-site image dataset is constructed accordingly. In the baseline model selection stage, VGG16, GoogLeNet, MobileNetV3, and ResNet50 are compared, and ResNet50 is selected as the basis for subsequent optimization. In the model improvement stage, the initial 7 × 7 convolution is replaced with stacked 3 × 3 convolutions. Depthwise separable convolution is introduced into the Bottleneck blocks, and efficient channel attention (ECA) and LeakyReLU activation are further incorporated to construct the ResNet50-Ours model. The experimental results show that ResNet50-Ours achieves an accuracy of 99.26%, with 10.50 M parameters and a frame rate of 119.20 FPS. Compared with the original ResNet50, the proposed model improves accuracy by 3.24% points, reduces the number of parameters by 55.36%, and increases the frame rate by 48.50%. The test-scenario results further indicate that the method can distinguish typical operating conditions, including safe states, hand intrusion, and foreign-object intrusion. The results provide model support for the embedded deployment of visual protection systems for hazardous regions of CNC press brakes.