Neural Networks Micro Memory Control Strategy for Mechanical Faults Edge Recognition

The edge recognition of mechanical faults using deep learning models requires the deployment and operation of neural networks, which consume a large amount of memory. However, edge devices have limited memory. To address the problem, a neural network micromemory control strategy is proposed. This strategy addresses the memory resource constraints through process quantization and byte indexing while maintaining the model structure and accuracy. First, the neural network parameters with Gaussian and uniform distributions are tested and corrected to achieve process quantization. Then, a parameter memory control is implemented to achieve low-memory deployment and runtime. Subsequently, a bytes indexing is proposed for parameter location reading, enabling low-memory process inversed quantization. Finally, temporary variables is computed through micro memory overlay. This memory control method using process quantization and bytes indexing not only allows for ultra-low memory deployment and operation of neural networks but also improves test accuracy. It achieves over 97% test accuracy and an edge computation speed of over 200 KB/s with a maximum of 3.89 KB of runtime memory.