STREAM-VAE: Dual-Path Routing for Slow and Fast Dynamics in Vehicle Telemetry Anomaly Detection

TL;DR

STREAM-VAE addresses the challenge of detecting anomalies in automotive telemetry that exhibit both slow drifts and fast spikes. It introduces a dual-path encoder that separately models drift and spike dynamics, coupled with a decoder that uses a per-feature mixture of experts and an event-residual path with soft-thresholding to isolate transients without inflating nominal tails. The approach yields improved robustness across operating modes and datasets, with favorable threshold calibration and real-time inference suitable for in-vehicle deployment and backend fleet analytics. The work demonstrates superior threshold-free ranking and competitive threshold-based metrics compared to strong baselines, and it identifies ablations that confirm the value of explicit time-scale separation and modular decoding. Practical impact includes more reliable, portable anomaly detection for intelligent vehicles and scalable fleet monitoring, though calibration remains per-entity and future work could focus on transferability across environments.

Abstract

Automotive telemetry data exhibits slow drifts and fast spikes, often within the same sequence, making reliable anomaly detection challenging. Standard reconstruction-based methods, including sequence variational autoencoders (VAEs), use a single latent process and therefore mix heterogeneous time scales, which can smooth out spikes or inflate variances and weaken anomaly separation. In this paper, we present STREAM-VAE, a variational autoencoder for anomaly detection in automotive telemetry time-series data. Our model uses a dual-path encoder to separate slow drift and fast spike signal dynamics, and a decoder that represents transient deviations separately from the normal operating pattern. STREAM-VAE is designed for deployment, producing stable anomaly scores across operating modes for both in-vehicle monitors and backend fleet analytics. Experiments on an automotive telemetry dataset and the public SMD benchmark show that explicitly separating drift and spike dynamics improves robustness compared to strong forecasting, attention, graph, and VAE baselines.

Figures (4)

• Figure 1: Illustration of the core detection challenge: fast spikes and slow drifts distort different aspects of the signal. STREAM-VAE maps clean segments to a compact latent manifold, while anomalous patterns should fall outside this region.
• Figure 2: Overview of the STREAM-VAE architecture with dual drift (top, purple) and spike (bottom, red) paths. Each block is annotated with the shape of its output tensor. The BI-LSTM Encoder (yellow) maps the input sequence $X \in \mathbb{R}^{T \times F}$ to encoder states $H_E \in \mathbb{R}^{T \times H_{\text{E}}}$ and latent states $Z \in \mathbb{R}^{T \times D}$. Slow EMA-based features define the drift path, while high-pass residual features define the spike path; both produce attention outputs in $\mathbb{R}^{T \times D}$. A gated fusion yields the latent context $Z_{\text{ctx}} \in \mathbb{R}^{T \times D}$, and the first difference $\Delta Z \in \mathbb{R}^{T \times D}$ carries transient information. The BI-LSTM Decoder maps these to hidden states $H_D \in \mathbb{R}^{T \times H_{\text{D}}}$, which drive per-feature mixture-of-experts (MoE) heads that output Gaussian reconstruction parameters $(\hat{x}, \hat{\sigma}^2) \in \mathbb{R}^{T \times F}$ used for anomaly scoring.
• Figure 3: Dual-path global attention (per head). Drift path (top). Multiple heads show bright vertical bands at consistent key indices, so many queries align to the same columns. This produces a few global reference anchors and a low-rank, anchor-based alignment that reflects slow and persistent context. Spike path (bottom). Heads display dense, oscillatory checkerboard patterns and narrow query-centric stripes, which indicate short-lag, high-frequency coupling typical of localized transients. The mean maps remain structured in both rows, which shows that heads specialize consistently and that the two paths split cleanly: The drift path captures global, low-frequency drift, and the spike path captures localized, high-frequency spikes.
• Figure 4: Component-wise contribution to reconstruction error by anomaly type.