Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Dual-Stream Physics-Augmented Unsupervised Architecture for Runtime Embedded Vehicle Health Monitoring

Enzo Nicolas Spotorno, Antonio Augusto Medeiros Frohlich

TL;DR

This work addresses the gap where unsupervised anomaly detectors fail to capture sustained mechanical work in vehicles by introducing a Dual-Stream Physics-Augmented architecture that combines an LSTM Autoencoder operating on 10 Hz sensor data with macroscopic physics proxies to estimate cumulative load. The health vector $H_t=[A_{ML}(t),W_{Phys}(t)]$ is fused via Score$(t)=\max(\text{Norm}(A_{ML}(t)),\text{Norm}(W_{Phys}(t)))$, enabling edge-accurate detection of both dynamic hazards and sustained drivetrain work. Validation on a RISC-V embedded platform shows low latency and energy overhead (~$6.35\mu$s,~$0.680\mu$J per inference), confirming feasibility for runtime monitoring on resource-constrained ECUs. The results demonstrate that macroscopic proxies correlate with cumulative wear and that the dual-stream approach fills the blind spot of purely data-driven methods, enabling robust, interpretable predictive maintenance using existing sensors. This approach holds significant practical impact for dynamic maintenance in mixed-use fleets by providing continuous, on-board health assessment without cloud dependence.

Abstract

Runtime quantification of vehicle operational intensity is essential for predictive maintenance and condition monitoring in commercial and heavy-duty fleets. Traditional metrics like mileage fail to capture mechanical burden, while unsupervised deep learning models detect statistical anomalies, typically transient surface shocks, but often conflate statistical stability with mechanical rest. We identify this as a critical blind spot: high-load steady states, such as hill climbing with heavy payloads, appear statistically normal yet impose significant drivetrain fatigue. To resolve this, we propose a Dual-Stream Architecture that fuses unsupervised learning for surface anomaly detection with macroscopic physics proxies for cumulative load estimation. This approach leverages low-frequency sensor data to generate a multi-dimensional health vector, distinguishing between dynamic hazards and sustained mechanical effort. Validated on a RISC-V embedded platform, the architecture demonstrates low computational overhead, enabling comprehensive, edge-based health monitoring on resource-constrained ECUs without the latency or bandwidth costs of cloud-based monitoring.

A Dual-Stream Physics-Augmented Unsupervised Architecture for Runtime Embedded Vehicle Health Monitoring

TL;DR

This work addresses the gap where unsupervised anomaly detectors fail to capture sustained mechanical work in vehicles by introducing a Dual-Stream Physics-Augmented architecture that combines an LSTM Autoencoder operating on 10 Hz sensor data with macroscopic physics proxies to estimate cumulative load. The health vector is fused via Score, enabling edge-accurate detection of both dynamic hazards and sustained drivetrain work. Validation on a RISC-V embedded platform shows low latency and energy overhead (~s,~J per inference), confirming feasibility for runtime monitoring on resource-constrained ECUs. The results demonstrate that macroscopic proxies correlate with cumulative wear and that the dual-stream approach fills the blind spot of purely data-driven methods, enabling robust, interpretable predictive maintenance using existing sensors. This approach holds significant practical impact for dynamic maintenance in mixed-use fleets by providing continuous, on-board health assessment without cloud dependence.

Abstract

Runtime quantification of vehicle operational intensity is essential for predictive maintenance and condition monitoring in commercial and heavy-duty fleets. Traditional metrics like mileage fail to capture mechanical burden, while unsupervised deep learning models detect statistical anomalies, typically transient surface shocks, but often conflate statistical stability with mechanical rest. We identify this as a critical blind spot: high-load steady states, such as hill climbing with heavy payloads, appear statistically normal yet impose significant drivetrain fatigue. To resolve this, we propose a Dual-Stream Architecture that fuses unsupervised learning for surface anomaly detection with macroscopic physics proxies for cumulative load estimation. This approach leverages low-frequency sensor data to generate a multi-dimensional health vector, distinguishing between dynamic hazards and sustained mechanical effort. Validated on a RISC-V embedded platform, the architecture demonstrates low computational overhead, enabling comprehensive, edge-based health monitoring on resource-constrained ECUs without the latency or bandwidth costs of cloud-based monitoring.
Paper Structure (10 sections, 4 equations, 3 figures, 2 tables)

This paper contains 10 sections, 4 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Proposed Dual-Stream Architecture
  4. Experimental Setup
  5. Experimental Results and Analysis
  6. Validation of Macroscopic Physics Proxies
  7. Analysis of Anomaly Detection Limitations
  8. Evaluation of the Dual-Stream Architecture
  9. Computational Feasibility Analysis
  10. Discussion and Conclusion

Figures (3)

  • Figure 1: Correlation analysis ($N=192,857$ windows): Stream A correlates with lateral stability but decouples from drivetrain work ($r \approx 0$).
  • Figure 2: Scenario-specific correlations: negligible drivetrain correlation in Ramp events vs. strong instability capture in Potholes.
  • Figure 3: Dual-stream fusion in ramp climb: (a) physics stream detects load, (b) ML stream misses it, (c) fused output correctly flags high intensity.