Electric Current Classification With Deeply Quantized Neural Network for Home Appliances

The development of Deeply Quantized Neural Networks is a relatively new field for the tiny machine learning community. It enables lower bit depth (integer and fractional) arithmetic for weights, biases, and activations at a point it could severely decrease performances if pushed to just one-bit signed representation. In this paper, it was investigated in the context of electric current classification of the home appliances. A neural network was developed to classify the appliances based on their electric current consumption. However, achieving that at lowest bit depth to save complexity, was challenging from an accuracy point of view. The Mel power spectrogram of the current signal was used as pre-processing to feed a deeply quantized neural network. The objective was to build a high accurate model to classify the appliances while respecting the limited memory and computation latency constraints of a cheap microcontroller accordingly to the model complexity. The classification accuracy achieved was 98.73% on the WHITED dataset, 97.04% on the COOLL dataset while it reached 94.73% on the first group of classes of PLAID dataset and 92.80% on the second group of classes. The model was characterized on two microcontrollers in term of execution time and power consumption. On STM32H7, it ran around 11 ms. It was 7 ms faster than a floating-point precision based convolutional neural network, estimated consuming 2.5 mA of current. On STM32L4 instead, the model was executed around 92 ms and required 0.89 mA of current.