Behavior-Induced False Positives in Vehicle Telemetry Anomaly Detection: An Empirical Study

Unsupervised anomaly detection is a promising approach for vehicle health monitoring, but its deployment in human-driven telemetry is often limited by nominal false positives. A key difficulty is that many telemetry channels are influenced not only by vehicle dynamics, but also by partially observed driver behavior and operating context, making some signals systematically less predictable than others. This paper presents a systematic empirical analysis of how such variability affects false-positive behavior in vehicle telemetry anomaly detection. Using multiple public vehicle telemetry datasets and representative detector families, we analyze the feature-level structure of nominal residual error and its contribution to false-positive detections. The results show that residual error is not distributed uniformly across the telemetry space, but is instead concentrated in a small subset of driver-controlled channels. This concentration persists across model classes, becomes more pronounced in falsely flagged nominal windows, and is not eliminated by increasing training-driver diversity. Feature-level analysis further shows that steering-, braking-, and accelerator-related variables dominate the residual hierarchy, whereas more tightly regulated vehicle-state channels remain comparatively stable. The findings identify a structured and recurrent false-positive failure mode in human-driven telemetry. It is further suggested that reliable deployment will require channel-aware or context-conditioned scoring strategies, rather than increased model complexity alone.