Concept Drift Mitigation on Resource-Constrained IoT Devices via Self-Learning

Real-world deployments of Internet of Things (IoT) sensing systems equipped with artificial intelligence (AI) models generally experience a reduction in accuracy over time due to concept drift and data shift. To overcome this issue, it is suggested to periodically retrain the embedded AI model using new incoming data. However, this requires powerful processing hardware and labelled data, which are not generally available on IoT edge-based devices deployed in the real-world. In this study, we propose a method that benefits from a Deep Learning (DL) model for feature extraction and K-Means clustering for classification. The K-Means clustering technique can use the new incoming unlabeled data to update the model, thus helping the system to adapt to any changes that might occur in the new data. The proposed method is evaluated on two image datasets: the first is a public dataset with artificially added concept drift, and the second is a real-world dataset that suffers from concept drift issues. This is a lightweight model with only 610 KB of size and 608 KB of peak memory, which needs less than 0.6 s to perform and lower than 0.7 J to analyse each sample and update the model. Therefore, the method could easily be stored and executed on resource-constrained microcontroller-based (MCU-based) devices to deal with concept drift.