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

Area of Science:

• Robotics and Automation
• Machine Learning
• Industrial Safety

Background:

• Conventional safety methods for CNC press brakes struggle with open die regions and frequent human interaction.
• Existing systems lack sufficient active recognition capabilities for dynamic hazards.
• Sheet-metal bending poses significant risks due to tool motion and workpiece placement.

Purpose of the Study:

• To develop an improved ResNet50 model for accurate hazardous-region state recognition in CNC press brakes.
• To enhance the safety of human-machine interaction during sheet-metal bending operations.
• To provide a robust visual protection system for industrial machinery.

Main Methods:

• Defined a visual monitoring region based on tool motion and workpiece placement.
• Constructed an on-site image dataset for training and validation.
• Optimized the ResNet50 architecture by replacing convolutions, incorporating depthwise separable convolutions, efficient channel attention (ECA), and LeakyReLU activation.

Main Results:

• The developed ResNet50-Ours model achieved 99.26% accuracy with 10.50M parameters and 119.20 FPS.
• Accuracy improved by 3.24% points, parameters reduced by 55.36%, and frame rate increased by 48.50% compared to the original ResNet50.
• Successfully distinguished between safe states, hand intrusion, and foreign-object intrusion in test scenarios.

Conclusions:

• The optimized ResNet50 model offers superior performance for hazardous-region state recognition.
• The findings support the embedded deployment of advanced visual protection systems for CNC press brakes.
• This advancement contributes to improved safety and efficiency in sheet-metal bending operations.