Driving with Data: Modeling and Forecasting Vehicle Fleet Maintenance in Detroit

Josh Gardner, Danai Koutra, Jawad Mroueh, Victor Pang, Arya Farahi, Sam Krassenstein, Jared WebbarXiv 2017 · OpenReview

Abstract

The City of Detroit maintains an active fleet of over 2500 vehicles, spending an annual average of over $5 million on new vehicle purchases and over $7.7 million on maintaining this fleet. Understanding the existence of patterns and trends in this data could be useful to a variety of stakeholders, particularly as Detroit emerges from Chapter 9 bankruptcy, but the patterns in such data are often complex and multivariate and the city lacks dedicated resources for detailed analysis of this data. This work, a data collaboration between the Michigan Data Science Team and the City of Detroit's Operations and Infrastructure Group, seeks to address this unmet need by analyzing data from the City of Detroit's entire vehicle fleet from 2010-2017. We utilize tensor decomposition techniques to discover and visualize unique temporal patterns in vehicle maintenance; apply differential sequence mining to demonstrate the existence of common and statistically unique maintenance sequences by vehicle make and model; and, after showing these time-dependencies in the dataset, demonstrate an application of a predictive Long Short Term Memory (LSTM) neural network model to predict maintenance sequences. Our analysis shows both the complexities of municipal vehicle fleet data and useful techniques for mining and modeling such data.

Introduction: The Logistical Puzzle of Municipal Fleets

Managing a city's vehicle fleet is a massive, expensive logistical challenge. The City of Detroit maintains over 2,500 vehicles—ranging from police cruisers and fire trucks to specialized public works equipment like lawnmowers and salt trucks. Annually, maintenance costs exceed $7.7 million.

The primary difficulty in optimizing these operations lies in the complexity of the data. Maintenance records are inherently multivariate, involving disparate systems such as brakes, engines, and electronics, and temporal, with patterns shifting across seasons and as vehicles age. Most municipal governments lack the specialized data science resources required to analyze these complex patterns, often resulting in reactive maintenance strategies. This research bridges that gap by demonstrating how modern machine learning can transform "data exhaust" into actionable insights for urban infrastructure management.

Methodology

The researchers employed a three-tiered analytical approach to decode the fleet's maintenance patterns:

  1. Tensor Representation and PARAFAC Decomposition: The data was structured as a 3D "tensor" (Vehicle $\times$ System $\times$ Time). Using PARAFAC (Parallel Factors) decomposition, the team broke this complex data block into simpler components, allowing for the visualization of interactions between specific vehicle types, mechanical systems, and time periods.
  2. Differential Sequence Mining: To identify predictable "pathways" in vehicle wear, sequence mining was used to find common strings of repairs (e.g., Oil Change $\rightarrow$ Brake Repair $\rightarrow$ Tire Replacement). Statistical tests then determined if these sequences were unique to specific makes and models.
  3. LSTM Neural Networks: Maintenance history was treated as a sequential language. A Long Short-Term Memory (LSTM) network was trained to "read" a vehicle's past 20 repairs and predict the most likely subsequent maintenance task.

Dataset Overview

The study utilized a comprehensive dataset provided by the City of Detroit’s Operations and Infrastructure Group, covering the period from 2010 to 2017.

The City of Detroit's diverse vehicle fleet, including emergency services and public works vehicles, presents a complex maintenance challenge.

The data consisted of two primary tables:

  • Vehicles Table: 6,725 records detailing every vehicle owned by the city since 1944, including make, model, and purchase cost.
  • Maintenance Table: 229,540 individual job records ranging from routine inspections to full engine replacements.
  • Refined Subset: For predictive modeling, the authors focused on 1,087 vehicles purchased after 2010 to ensure data reliability, encompassing over 25,000 maintenance records.

Key Insights and Results

The analysis yielded several critical findings regarding the predictability of municipal maintenance:

  • Maintenance "Fingerprints": Different vehicle types exhibit statistically unique maintenance signatures. For instance, repair sequences for a Dodge Charger police cruiser are fundamentally distinct from those of a Smeal SST Pumper fire truck ($p < 0.0001$).
  • Seasonality vs. Aging: Tensor analysis successfully separated seasonal patterns (e.g., summer maintenance for lawnmowers) from age-related wear (e.g., garbage trucks requiring hydraulic repairs specifically in years 2–4 of service).
  • Forecasting Accuracy: The LSTM model achieved a test perplexity of 15.7. Compared to a baseline "random" model (perplexity of 260), the LSTM is significantly more effective at narrowing down the next likely maintenance need, proving that even "noisy" human-entered data contains strong underlying mechanical patterns.
  • Visual Discovery: PARAFAC decomposition identified "factors" mapping directly to real-world operations, such as specific engine and body repair cycles for 2014 Terrastar Horton ambulances.

Conclusion and Future Work

While this work proves that municipal maintenance is highly predictable, several avenues for improvement remain. The current model relies on the order of repairs; future iterations could incorporate mileage (odometer readings), garage location, and technician ID for more granular predictions. Additionally, addressing "data entry noise"—where different technicians use different codes for similar jobs—could further refine model accuracy.

Ultimately, these predictive tools can be translated into economic models to calculate the "Total Cost of Ownership," helping cities make data-driven decisions on when to retire aging vehicles versus continuing costly repairs.

Summary updated

Abstract

The City of Detroit maintains an active fleet of over 2500 vehicles, spending an annual average of over \$5 million on new vehicle purchases and over \$7.7 million on maintaining this fleet. Understanding the existence of patterns and trends in this data could be useful to a variety of stakeholders, particularly as Detroit emerges from Chapter 9 bankruptcy, but the patterns in such data are often complex and multivariate and the city lacks dedicated resources for detailed analysis of this data. This work, a data collaboration between the Michigan Data Science Team (http://midas.umich.edu/mdst) and the City of Detroit's Operations and Infrastructure Group, seeks to address this unmet need by analyzing data from the City of Detroit's entire vehicle fleet from 2010-2017. We utilize tensor decomposition techniques to discover and visualize unique temporal patterns in vehicle maintenance; apply differential sequence mining to demonstrate the existence of common and statistically unique maintenance sequences by vehicle make and model; and, after showing these time-dependencies in the dataset, demonstrate an application of a predictive Long Short Term Memory (LSTM) neural network model to predict maintenance sequences. Our analysis shows both the complexities of municipal vehicle fleet data and useful techniques for mining and modeling such data.
BibTeX
@misc{gardner2017driving,
  author = {Josh Gardner and Danai Koutra and Jawad Mroueh and Victor Pang and Arya Farahi and Sam Krassenstein and Jared Webb},
  title = {{Driving with Data: Modeling and Forecasting Vehicle Fleet Maintenance in Detroit}},
  year = {2017},
  eprint = {1710.06839},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1710.06839},
}
Figure