Machine Learning-Assisted, Process-Based Quality Control for Detecting Compromised Environmental Sensors

This study presents a machine learning-assisted data qualitycontrol methodology for environmental sensor data. This process-constrainedmethodology is shown to be more robust than existing state-of-the-artsin detecting faulty data. Machine learning (ML) techniquespromise to revolutionize environmentalresearch and management, but collecting the necessary volumes of high-qualitydata remains challenging. Environmental sensors are often deployedunder harsh conditions, requiring labor-intensive quality assuranceand control (QAQC) processes. The need for manual QAQC is a majorimpediment to the scalability of these sensor networks. Existing techniquesfor automated QAQC make strong assumptions about noise profiles inthe data they filter that do not necessarily hold for broadly deployedenvironmental sensors, however. Toward the goal of increasing thevolume of high-quality environmental data, we introduce an ML-assistedQAQC methodology that is robust to low signal-to-noise ratio data.Our approach embeds sensor measurements into a dynamical feature spaceand trains a binary classification algorithm (Support Vector Machine)to detect deviation from expected process dynamics, indicating whethera sensor has become compromised and requires maintenance. This strategyenables the automated detection of a wide variety of nonphysical signals.We apply the methodology to three novel data sets produced by 136low-cost environmental sensors (stream level, drinking water pH, anddrinking water electroconductivity), deployed by our group across250,000 km(2) in Michigan, USA. The proposed methodologyachieved accuracy scores of up to 0.97 and consistently outperformedstate-of-the-art anomaly detection techniques.