Logo of Acerta
Book a demo

Case Study

On-road engine failure predictions 100km in advance

Objectives

Warn drivers early about engine issues with real-time health and failure predictions—helping Nissan improve customer satisfaction and reduce unexpected breakdowns.

Challenge

Accurately predicting how fast engine parts degrade was difficult, especially when using either mechanical models or AI models alone. Limited real-world failure data and biases in test data added to the challenge.

Key Results

  • 90% accuracy predicting engine injector failures; 86% for air-fuel sensor.
  • Predicted engine injector failure 100+ km in advance.
  • Provided real-time engine health and failure warnings.

90%

Injector model accuracy

100km

Warning distance before failure

90%

Sensor model accuracy

Background

Nissan Motor Company partnered with Acerta to enhance predictive maintenance capabilities for critical engine components. As vehicle mileage increases, components deteriorate and may require significant repairs, but traditional approaches struggled to provide accurate predictions. Nissan sought a robust, real-time solution that could notify drivers of impending failures with enough lead time for preemptive repairs. Leveraging Acerta’s expertise in anomaly detection and machine learning, Nissan aimed to bridge the gap between mechanical data models and AI-based predictive analytics. This project was funded by the Ontario Vehicle Innovation Network (OVIN).

Problem

  • Complex Signal Relationships: Precise relationships between signals from various engine components were not fully understood, complicating failure prediction.
  • Limited On-Road Failure Data: Traditional AI models required large volumes of on-road vehicle data with labeled failures, which was difficult to obtain.
  • Dyno Data Bias: Simulated dyno test data lacked the diversity and unpredictability of real-world driving conditions, introducing biases into the models.
  • Difficulty Estimating Degradation Rate: Classical models couldn’t effectively predict the rate of component deterioration, limiting the accuracy of failure predictions.

Solution

Hybrid Data Approach

Acerta combined dyno test data and real-world on-road data collected from pre-production Nissan vehicles driven by employees. This hybrid strategy enabled the creation of more reliable and representative failure prediction models.

Targeted Modeling for Three Components
  1. Fuel Injector (INJ): Developed both classifier and degradation models to predict failure at least 100 km in advance.
  2. Air by Fuel Sensor (AFS): Built classifier models achieving 86% accuracy.
  3. Air Flow Meter (AFM): Utilized an autoencoder model to overcome dyno data bias, achieving perfect accuracy distinguishing normal from anomaly.
Remaining Distance Prediction Models
  • Parametric Approach: Used exponential degradation curves to predict injector failure based on anomaly scores.
  • Non-Parametric Approach: Estimated remaining useful life without assumptions, providing stable, reliable predictions based on a fleet of vehicles.
On-Road Demonstration

The system was installed in a Nissan Rogue and demonstrated in Waterloo, Ontario. Guests experienced real-time engine health monitoring, with simulations of injector failure to showcase the system’s predictive capabilities.

Results

  • Achieved 90% accuracy in predicting injector failures.
  • Predicted injector failures at least 100 km before breakdown.
  • Reached 86% accuracy in detecting air-fuel sensor anomalies.
  • Provided drivers with real-time engine health monitoring and early failure warnings.

We’re excited to continue our partnership

“Nissan recognizes the strength in Ontario’s thriving automotive ecosystem combined with expertise in AI and manufacturing. We worked with Acerta to accelerate the development of this new technology for our vehicles to bring future tangible benefits to our customers, and we’re excited to continue our partnership.”
Kazuhiro Doi
Corporate Vice President and Alliance Global VP Research Division

Case Studies

Proven Impact