A two-step machine learning approach for predictive maintenance and anomaly detection in environmental sensor systems

Area of Science:

• Environmental Science
• Data Science
• Machine Learning

Background:

• Environmental sensor systems are crucial for infrastructure and environmental quality monitoring.
• Sensor faults and anomalies can compromise the reliability of these systems.
• Existing methods struggle with unlabeled sensor telemetry data.

Purpose of the Study:

• To develop a novel methodology for anomaly detection and sensor failure prediction.
• To address the challenge of unlabeled data in environmental sensor telemetry.
• To enhance the reliability of environmental monitoring systems through predictive maintenance.

Main Methods:

• A hybrid approach combining unsupervised (Isolation Forest) and supervised machine learning models.
• Unsupervised learning used to generate labels for unlabeled sensor data.
• Supervised models (Random Forest, Neural Network, AdaBoost) trained on labeled data for anomaly prediction.

Main Results:

• The proposed framework achieved high accuracy in anomaly detection and sensor failure prediction.
• Random Forest: 99.93%, Neural Network (MLP Classifier): 99.05%, AdaBoost: 98.04%.
• Demonstrated the effectiveness of using Isolation Forest for labeling unlabeled IoT sensor data.

Conclusions:

• The methodology successfully transforms raw, unlabeled IoT sensor data into actionable insights.
• Provides a scalable and robust real-time solution for anomaly detection and sensor fault prediction.
• Advances intelligent infrastructure management and enhances the reliability of environmental monitoring.