Noisy Label Filtering for Nonintrusive Appliance Identification via Negative Learning

Nonintrusive load monitoring (NILM) can extract power consumption information for individual appliances by analyzing voltage and/or current measurements from a limited number of locations within the power distribution system. Cutting-edge techniques based on deep neural networks have shown promising performance for NILM tasks. Their success requires large-scale curated datasets with human annotations, which are expensive and time-consuming to obtain. Typically, these model trainings use positive learning (PL) to label input appliance data in a supervised manner, which will result in a noticeable degradation in performance when inaccurate or noisy labels exist. In this article, an indirect learning method called negative learning (NL) is implemented to address the noisy label issue, in which the model is trained using the information that “labels of other appliances are not labels of the specific appliance to be identified.” Via this, it enhances the feature representation and reduces the misidentification risk of the concerned appliances. Furthermore, filtering NL and cleaning PL are applied to selected high-confidence training data to achieve improved noisy data filtering. Experiments on publicly available datasets show that, with a semi-supervised learning (SSL) technique, the proposed method reaches superior performance for appliance identification when noisy data labels exist in the adopted dataset.