Optimizing Self-Supervised Autoencoders for IoT Sensor Data Using Memory-Augmented Adaptive Simulated Annealing

Abstract

The emergence of the Internet of Things (IoT) has given rise to unprecedented volumes of unlabeled sensor data due to its constantly evolving nature. This phenomenon has created an acute need for learning models capable of intelligently use such data devoid of expensive manual labeling. Self-Supervised Learning (SSL) has recently arisen as an efficient paradigm to meet this challenge by directly deriving informative representations from unlabeled data. Nevertheless, the practical application of SSL models is hindered by their over-dependence on hyperparameter tuning, which remains a central challenge. In addressing this challenge, we introduce an MA-ASA (Memory-Augmented Adaptive Simulated Annealing) framework for hyperparameter tuning that has been optimized to this specificity. This framework is built on the concepts of temperature-adaptive mechanisms, short-term memory buffers to mitigate repetitive evaluations, and stagnation-initiated restarts, offering greater efficiency in dealing with hyperparameter space complexities. The proposed method employs a two-step framework consisting of denoising convolutional autoencoder (dCae) SSL pretraining on unlabeled data and logistic regression downstream classification on sparse labeled data. The approach was tested on UCI HAR, MHEALTH, and Environmental Sensor Telemetry datasets achieving classification accuracies of 96.00%, 97.33%, and 99.25%, respectively. The results showcase that the MA-ASA-based optimization strategy improves the performance and generalization of SSL models relative to other IoT domains, all while retaining computational efficiency and scalability.