Electric current classification with tiny machine learning for home appliances

Non-intrusive electric load monitoring methods analyze changes in voltage and current measured at the household's power plug connection point. This helps to disaggregate the total power consumption into individual appliances contributions. The idea is to use their unique fingerprint to identify and classify the active appliances. In this paper the mel power spectrogram of the current measurement was used as an input transformed data for both convolutional neural network and more traditional machine learning models. The primary goal was to achieve high accuracy classification. Moreover, the memory and computational complexity of these models were compared against the challenging constraints set by limited resource embedded systems and during the development. The work was concerned on developing machine learning models targeting software execution and deploying those models into a future intelligent meter. Microcontrollers were considered as the primary execution targets in order to evaluate the models' ground truth performance. As a result, the classification accuracy of the appliances measured on the WHITED dataset for the convolutional neural network was 98.6%. While the accuracy with the COOLL dataset achieved 95%. By carefully analyzing both convolutional neural network and off the shelf machine learning models, only the former achieved a tiny memory suitable for the microcontroller deployment.