Contrastive Self-Supervised Learning in Transportation Monitoring Systems: A Practical Approach to Early Anomaly Detection

To detect defects and prevent accidents in transport systems, sensors are used that use models based on classical designs or threshold rules, which is fraught with false alarms or delays. Purpose: to propose and describe a practical approach to the use of contrastive self-supervised learning methods for early detection of anomalies in transport monitoring systems. Methods: contrastive learning methods with the InfoNCE loss function, lightweight convolutional encoders (1D-CNN/1D-ResNet), variational autoencoders (VAE) for reconstruction control, attention mechanisms for assessing sensor contributions, adaptive calibration of thresholds based on the exponential moving average (EMA), representation clustering for multiple normal modes, and testing scenarios on semi-real data are applied. Results: features of multisensory time series have been identified, limitations of field devices have been described, and ways to reduce false positives with limited markup have been identified. Algorithmic blocks for practical implementation are proposed: time series augmentation, adaptive threshold calibration, attention mechanisms for explainability and validation on semi-real data. A hybrid anomaly criterion based on contrastive and reconstruction scores has been developed. Practical significance: the introduction of a contrast unit with adaptive calibration provides a decrease in the time for detecting anomalies and a decrease in the frequency of false positives compared to basic VAE and threshold systems. Discussion: integration of contrast modules and adaptive calibration into existing transport monitoring systems using a three-level alarm system with per-channel explanation.