Enhancing Robustness and Energy Efficiency of IoT-Coupled Machine Learning Applications

Abstract

Machine learning (ML) applications continue to revolutionize many domains. In recent years, there has been considerable research interest in building novel ML applications for a variety of Internet of Things (IoT) domains, such as precision agriculture, smart cities, and smart manufacturing. However, IoT characteristics pose several fundamental challenges to designing and implementing effective ML applications at the edge, specifically, inference-time data incompleteness and quality induced by sensor failures and the energy management of such applications. This dissertation comprises four components aimed at addressing the aforementioned challenges to enhance IoT ML-based systems’ robustness, energy efficiency, and reliability against sensor failures at inference time. First, to tackle the sudden and concurrent missing data of sensors at the inference time of the ML application, it introduces an ensemble of a few sub-models, each trained with distinct subsets of sensors chosen based on correlation to make ML robust against missing data. This ensemble-based approach maintains the accuracy of the IoTMLapplication during the occurrence of missing data at inference time via performing predictions using sub-model(s) that are built using the correlated sensors to the faulty sensors, excluding the faulty sensors. The second component involves an extensive study to identify energy bottlenecks for operating our ensemble-based approach, SECOE, on a resource-constrained edge device and find the number of ensemble models that further maintain the ML application’s performance while optimizing energy consumption. The third component introduces an energy-aware ensemble of models with enhanced energy management capabilities for handling data incompleteness during inference time to fit resource-constrained IoT devices. Finally, our research offers an innovative autoencoder model for correcting inaccurate/faulty sensor readings caused by the malfunctioning of multiple sensors simultaneously at inference time. This model treats different faulty readings (e.g., drift and bias) as one type via a correlation-based masking mechanism. The experimental results have demonstrated that our methods have effectively alleviated the impact of IoT sensor failures on the performance of theMLapplication during the inference time while optimizing its energy consumption, thereby decisively outperforming existing methods and significantly improving the overall reliability of ML coupled IoT applications.