Anomaly Detection and Segmentation in Measurement Signals on Edge Devices Using Artificial Neural Networks

Area of Science:

• Environmental Science
• Computer Science
• Machine Learning

Background:

• Anomalies in soil signal time series can compromise data integrity.
• Deploying advanced anomaly detection on resource-constrained edge devices (MCUs) presents significant computational and power challenges.

Purpose of the Study:

• To investigate and compare three artificial neural network (ANN) approaches for in situ anomaly detection and cleaning in soil sensor data.
• To evaluate the trade-off between prediction accuracy and computational complexity for ANNs running on Microcontroller Units (MCUs).

Main Methods:

• Developed and tested three ANN models: Neural-Primed Heuristic (NPH), Autoencoder-Based (AEB), and U-Net-Based (UNB).
• Trained models on cloud infrastructure using labeled data, transforming anomaly detection into an anomaly segmentation task.
• Evaluated models for accuracy and computational complexity on constrained end devices.

Main Results:

• All three models demonstrated varying levels of prediction accuracy and computational complexity.
• The U-Net-Based (UNB) approach achieved good results but demands substantial training data and exhibits high time complexity, leading to increased power consumption.
• Neural-Primed Heuristic (NPH) and Autoencoder-Based (AEB) models offer a more balanced compromise between accuracy and resource efficiency.

Conclusions:

• While UNB shows potential, its resource demands limit its practicality for low-power IoT applications.
• NPH and AEB are recommended as viable alternatives for developing efficient in situ data cleaning solutions in IoT measurement systems.
• The study highlights the importance of considering computational constraints when deploying machine learning models on edge devices.