On the edge implementation of quantized ANN for interturn short-circuit diagnostics
The prepared software package contains a set of scripts for the conversion of neural networks for interturn short-circuit detection from float arithmetic to integer arithmetic suitable for implementation in tensor processing units TPU. A trained neural network and a representative dataset are needed to use the quantization process. The resulting neural network contains only integer types and can be easily implemented in low-cost microcontrollers without float processing units FPU or into TPU platforms. The size of the neural network is reduced to approximately one-quarter of the original size with minimal reduction in classification accuracy. The resulting quantized networks were tested on a test system for the analysis of interturn short-circuits in a multiphase PMS motor with a nerve platform with an additional Coral TPU accelerator.