Track Anomaly Detection Using Log-Likelihood and MAD Thresholding

The patent segments the anomaly detection process into two distinct phases: an offline training phase where the neural network is trained on historical data, and an online inference phase where the trained model is applied to new tracks. This segmentation allows computationally intensive operations to be performed offline, reducing real-time resource consumption while maintaining detection accuracy.

The patent performs preliminary training of the neural network offline using historical track data before deployment. The trained model parameters are stored and reused during online operation, eliminating the need to perform heavy computational training during real-time anomaly detection, thus reducing online computational resource consumption.

The patent divides the detection process into offline training (time-consuming but performed once) and online inference (fast and repeated). This segmentation enables real-time processing during online operation while achieving high accuracy through the preliminary offline training phase.

The neural network is trained in advance offline, so that during online operation, only lightweight inference is required. This preliminary action eliminates the need for time-consuming training during real-time detection, enabling real-time processing while maintaining high detection accuracy.

The patent transforms the detection threshold from a variable that changes frequently with track duration into a fixed value derived from statistical parameters (mean and standard deviation of log-likelihood scores). This parameter transformation simplifies threshold management while maintaining detection accuracy across different track durations.

The system automatically computes the detection threshold using statistical properties of the trained model's output on validation data. This self-service approach eliminates the need for manual threshold adjustment and complex threshold management, reducing system complexity while adapting to different track characteristics.

The patent segments computational tasks between offline training (performed on powerful systems) and online inference (performed on embedded systems). This segmentation enables deployment on resource-constrained embedded devices by moving heavy computational burden to the offline phase.

The neural network is trained in advance on powerful systems, and only the trained model parameters are deployed to embedded systems. This preliminary action reduces the resource requirements of embedded systems during online operation, making anomaly detection feasible on devices with limited computational resources.