ABNN: Adaptive-Gating Binary Neural Network With Dynamic Activation Quantization for Industrial Health Status...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Edge Computing

Background:

• Industrial equipment health prediction is crucial for safety and reliability.
• Deploying high-precision deep learning models at the industrial edge is challenging due to resource and real-time constraints.

Purpose of the Study:

• To propose an efficient adaptive-gating binary neural network (ABNN) for industrial edge scenarios.
• To overcome the limitations of deploying complex deep learning models in resource-constrained edge environments.

Main Methods:

• Developed a trend-aware encoder (TAE) for optimized input layer binarization.
• Introduced a learnable precision indicator (LPI) for adaptive inference precision.
• Designed an adaptive-gating convolution to enhance representational capabilities without increasing computational cost.
• Implemented a field-programmable gate array (FPGA) hardware accelerator.

Main Results:

• The proposed ABNN achieved approximately a 7% improvement in accuracy compared to the baseline model.
• The ABNN demonstrated a 45% gain in efficiency over the baseline model.
• The network effectively balances representational power and computational efficiency.

Conclusions:

• The ABNN offers an efficient solution for real-time industrial equipment health prediction at the edge.
• The proposed methods enable the deployment of accurate deep learning models in resource-limited environments.
• The ABNN framework, coupled with FPGA acceleration, shows significant promise for industrial edge applications.