Generative Learning for Simulation of Vehicle Faults

TL;DR

This work develops a practical, generative approach to predictive maintenance for US Army ground vehicles by forecasting faults and time-to-first-fault using a hybrid model that combines DeepAR with spatio-temporal attention (STAM) and a variational autoencoder (VAE). Fault covariates are generated with STAM from recent history and with a VAE from out-of-sample data, enabling state-conditioned simulations via latent-space clustering of vehicle attributes. The method achieves high fault-prediction performance (ROC-AUC up to 0.978) and a robust time-to-first-fault forecast ($r^{2} = 0.77$), while yielding interpretable insights through attention weights and latent-state analyses. This framework supports proactive maintenance and what-if fleet simulations, with public code and emphasis on practical deployability in DoD predictive maintenance programs.

Abstract

We develop a novel generative model to simulate vehicle health and forecast faults, conditioned on practical operational considerations. The model, trained on data from the US Army's Predictive Logistics program, aims to support predictive maintenance. It forecasts faults far enough in advance to execute a maintenance intervention before a breakdown occurs. The model incorporates real-world factors that affect vehicle health. It also allows us to understand the vehicle's condition by analyzing operating data, and characterizing each vehicle into discrete states. Importantly, the model predicts the time to first fault with high accuracy. We compare its performance to other models and demonstrate its successful training.

Figures (9)

• Figure 1: Description of Condition Based Maintenance dataset. Data from field vehicle operations, collected over several months from various vehicle types, are presented. This dataset includes time series of various fault signals, each with associated recording timestamps, as well as time series from multiple sensors also timestamped. Our analysis aims to utilize the sensor data as feature covariates to predict fault occurrences.
• Figure 2: Overview of generative model for conditional vehicle fault prediction on vehicle $k \in K$. At Step 1, a DeepAR model ($Q_\theta$) is trained using time series of fault target and sensor feature values from time point $0$ to $q$, and is subsequently evaluated at randomly selected time periods, with $t_0 \in S$. At Step 2, a STAM model ($g_{\theta}$) is trained over a period of $b=$ 3 hours and a VAE model ($v_\theta$) is trained on out of sample data. At Step 3, the DeepAR model ($Q_\theta$) generates representations of the fault indicator $\tilde{h}^k_{t_0:t_0+T}$ (resp. $\hat{h}^k_{t_0:t_0+T}$) of $T =$ 30 minutes using the sensor feature covariates generated from the STAM model (resp. the VAE model), where STAM utilizes a lookback length of 20 minutes. At Step 4, the classifier ($C$) and the regression model ($R$) use the predictions from both the STAM and VAE model to forecast if there is a fault and what time the first fault is (given a fault is predicted) from time point $t_0$ to $t_0 + T$.
• Figure 3: Illustration of mapping vehicle operational condition to generate future sensor readings.
• Figure 4: K-means cluster representation labeled on VAE latent space with metadata interpretation.
• Figure 5: Latent space fault visualization.