Smart energy & NILM

Non-Intrusive Load Monitoring (NILM) infers what individual appliances consume from a single aggregate electricity signal, without installing a sensor on every device. This line of work started during my PhD and continued through several student supervisions and publications.

⚡ Event detection with Tukey’s Fences

The core contribution is a statistical event detector for aggregate current signals: a fast Fourier transform isolates the relevant frequency content, and Tukey’s fences flag the samples that fall outside the expected spread. Appliance switching events are then detected without training data, and with a high accuracy compared to threshold-based baselines.

Published as a preprint: Event Detection for Non-intrusive Load Monitoring using Tukey’s Fences (Kaddour, Lehsaini, Bouchachia).

🧠 Unsupervised disaggregation

Building on the detected events, I worked on convolutional deep embedded clustering: a convolutional autoencoder compresses each event window into a small latent representation, K-Means initializes the cluster centroids, and a DEC objective refines the assignment. The goal is to group switching events by appliance without labelled data, which is the main practical obstacle to deploying NILM in real homes.

📉 Anomaly detection in consumption data

Earlier work compared unsupervised outlier detection methods (Isolation Forest, One-Class SVM, K-Means) on real consumption traces to identify abnormal electricity usage, published in the International Journal of Software Science and Computational Intelligence.

🧑‍🎓 Related supervision

This topic fed three Master’s internships I supervised at the University of Tlemcen: daily activity detection in smart homes, deep clustering for non-intrusive device monitoring, and anomaly detection in electricity consumption.