Anomaly Detection in Sensor Networks for Unmanned Systems: Machine Learning Methods and Simulation Analysis

Abstract

Sensor networks in unmanned systems are vulnerable to various anomalies including hardware malfunctions, environmental interference, cyber-attacks, and physical damage, necessitating robust anomaly detection mechanisms to maintain operational safety and mission success. This paper presents a comprehensive examination of machine learning methods for sensor anomaly detection and provides extensive simulation analysis across representative unmanned system scenarios. We systematically categorize anomalies by their characteristics: point anomalies representing isolated erroneous readings, contextual anomalies where values are abnormal given temporal or spatial context, and collective anomalies involving patterns across multiple sensors or time steps. The study examines supervised learning approaches including support vector machines with anomaly-specific feature engineering, random forests leveraging ensemble decision-making, and neural networks learning complex normal behavior patterns, analyzing their requirement for labeled anomaly data which is often scarce in practice. Unsupervised methods are comprehensively reviewed including statistical approaches (Gaussian mixture models, kernel density estimation), clustering-based techniques (DBSCAN, isolation forests), and autoencoder neural networks that learn compressed representations of normal sensor patterns with reconstruction errors indicating anomalies. Particular attention is devoted to time-series specific methods including recurrent neural networks (LSTM, GRU) capturing temporal dependencies, temporal convolutional networks offering computational efficiency, and transformer architectures with self-attention for long-range temporal relationships. We examine semi-supervised approaches including one-class SVM and deep SVDD that train exclusively on normal data, addressing the practical scenario where anomalies are rare and diverse. The paper develops a comprehensive simulation framework implementing fifteen prominent anomaly detection algorithms across multi-sensor unmanned system scenarios: UAV flight with IMU, GPS, barometer, and magnetometer sensors; autonomous ground vehicle with camera, LiDAR, wheel encoders, and ultrasonic sensors; and underwater vehicle with sonar, depth sensor, DVL, and compass. Simulation environments inject realistic anomalies including sensor bias drift, intermittent failures, noise spikes, calibration errors, and coordinated cyber-attacks affecting multiple sensors. Performance evaluation encompasses detection accuracy (precision, recall, F1-score), detection latency critical for real-time response, false alarm rates affecting operator trust, and computational requirements for embedded deployment. Comparative results reveal that deep learning approaches achieve superior detection accuracy (95%+ F1-scores) but require substantial training data and computational resources, while statistical methods offer interpretability and efficiency with moderate performance degradation. We investigate the impact of sensor correlation exploitation, demonstrating that multi-sensor fusion-based anomaly detection significantly outperforms independent sensor monitoring by identifying inconsistencies across redundant measurements. The paper examines adaptive anomaly detection that continuously updates models based on operational data, addressing concept drift where normal sensor patterns evolve over vehicle lifetime. Advanced techniques are analyzed including adversarial training for robustness to sophisticated attacks, attention mechanisms identifying which sensors contribute to anomaly decisions, and uncertainty quantification providing confidence estimates for detection decisions.