Driving with Data in the Motor City: Mining and Modeling Vehicle Fleet Maintenance Data

Josh Gardner, Jawad Mroueh, Natalia Jenuwine, Noah Weaverdyck, Samuel Krassenstein, Arya Farahi, Danai KoutraarXiv 2020 · arXiv

Abstract

Managing a municipal vehicle fleet is a complex logistical challenge involving thousands of vehicles and millions in annual expenditures. This research, conducted in partnership with the City of Detroit, introduces PRISM, a novel algorithm that integrates tensor decomposition with Bayesian sequence mining to uncover latent repair patterns. By modeling maintenance records as a multidimensional tensor, the study demonstrates that fleet needs are highly predictable. Results show that LSTM networks can accurately forecast future maintenance events, while ARIMA models provide robust financial projections for budgeting, transforming messy administrative logs into actionable operational intelligence.

Introduction

Municipalities often face a "data rich, information poor" dilemma. The City of Detroit maintains a diverse fleet of over 2,500 vehicles—from emergency response units to groundskeeping equipment—spending upwards of $12 million annually on maintenance. While years of logs existed, the multidimensional and sequential nature of the data made it difficult to extract strategic insights. This study addresses three core questions: how to interpret complex multivariate patterns, how to predict specific next-step maintenance needs, and how to forecast the resulting financial burden.

Methodology: Tensor Decomposition and PRISM

The researchers employed a multi-stage approach to bridge the gap between raw logs and interpretable sequences.

PARAFAC Decomposition

To capture the relationships between vehicles, mechanical systems, and time, the data was modeled as a 3D tensor ($X \in \mathbb{R}^{V \times S \times T}$). Using PARAFAC (Parallel Factors) decomposition, this "cube" was broken down into $R$ latent factors. Each factor represents a grouping of vehicles that experience similar repairs at similar times.

Tensor decomposition of fleet maintenance data into latent factors.

PRISM (PARAFAC-Informed Sequence Mining)

The primary contribution of this work is PRISM. While tensor decomposition identifies groupings, it does not inherently capture the temporal order of events. PRISM uses a Bayesian Gaussian Mixture Model (BGMM) to identify observations belonging to a specific factor and then applies a Bayesian Difference in Proportions Test (BDPT) to extract sequences of repairs that are uniquely characteristic of that group.

For predictive modeling, the authors utilized Long Short-Term Memory (LSTM) networks to learn dependencies in maintenance sequences and ARIMA models for high-level cost forecasting.

Data and Implementation

The study utilized a comprehensive dataset from Detroit’s Operations and Infrastructure Group, spanning records since 2010:

  • Vehicles: Metadata for 1,087 active vehicles (make, model, year).
  • Maintenance: Over 25,000 records including labor hours, parts costs, and system categories (e.g., "Brakes," "Lighting").

The researchers have released their implementation on GitHub to facilitate replication by other municipalities.

Key Insights and Results

The analysis revealed that maintenance events are not random but follow predictable "sentences."

Interpretable Patterns

PRISM successfully identified "specialist signatures"—patterns corresponding to specific technician teams or seasonal needs. For instance, it highlighted the surge in lawnmower repairs during summer months and identified that most vehicles require a major tire and tube overhaul in their second year of service.

PRISM results showing factor weights across vehicles, systems, and time, alongside characteristic repair sequences.

Performance Metrics

  • Maintenance Prediction: The LSTM model achieved a perplexity score of 15.7, significantly better than the frequency-based baseline of 260.
  • Cost Forecasting: ARIMA models predicted monthly costs at the department level with an RMSE between $38 and $49 per vehicle, providing a reliable tool for fiscal planning.
  • Scalability: The models remained effective even with data from fewer than 200 vehicles, suggesting viability for smaller fleet operations.

Conclusion

This work demonstrates the power of combining unsupervised pattern discovery with supervised learning in the public sector. By moving beyond isolated analysis of repairs, PRISM allows fleet managers to anticipate needs before they become failures. Future work will focus on anomaly detection to identify "lemon" vehicles and optimizing technician scheduling across city garages to further improve operational efficiency.

Summary updated

Abstract

The City of Detroit maintains an active fleet of over 2500 vehicles, spending an annual average of over \$5 million on purchases and over \$7.7 million on maintenance. Modeling patterns and trends in this data is of particular importance to a variety of stakeholders, particularly as Detroit emerges from Chapter 9 bankruptcy, but the structure in such data is complex, and the city lacks dedicated resources for in-depth analysis. The City of Detroit's Operations and Infrastructure Group and the University of Michigan initiated a collaboration which seeks to address this unmet need by analyzing data from the City of Detroit's vehicle fleet. This work presents a case study and provides the first data-driven benchmark, demonstrating a suite of methods to aid in data understanding and prediction for large vehicle maintenance datasets. We present analyses to address three key questions raised by the stakeholders, related to discovering multivariate maintenance patterns over time; predicting maintenance; and predicting vehicle- and fleet-level costs. We present a novel algorithm, PRISM, for automating multivariate sequential data analyses using tensor decomposition. This work is a first of its kind that presents both methodologies and insights to guide future civic data research.
BibTeX
@misc{gardner2020driving,
  author = {Josh Gardner and Jawad Mroueh and Natalia Jenuwine and Noah Weaverdyck and Samuel Krassenstein and Arya Farahi and Danai Koutra},
  title = {{Driving with Data in the Motor City: Mining and Modeling Vehicle Fleet Maintenance Data}},
  year = {2020},
  eprint = {2002.10010},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2002.10010},
}
Figure