Neural Network Application to Diagnostics and Control of Vehicle Control Systems

NeurIPS 1990

TL;DR

This paper explores the application of neural networks (NNs) for both diagnosing faults in vehicle powertrain systems and controlling active suspension systems. The authors demonstrate that NNs can effectively handle complex, real-time diagnostic tasks that traditional methods struggle with, and can also develop effective controllers for non-linear systems without requiring explicit plant models.

What Problem Does This Paper Address?

Diagnosing faults in complex, real-time vehicle control systems is a significant challenge. Traditional methods often struggle with the sheer volume and complexity of data generated by these systems, especially when aiming for real-time, continuous diagnostics. Furthermore, developing effective controllers for non-linear dynamic systems like active suspensions is also difficult, often requiring accurate plant models that are hard to obtain. This paper addresses these challenges by investigating the use of neural networks as a more robust and adaptable solution for both fault diagnosis and control in automotive applications.

Methods

The paper primarily utilizes backpropagation neural networks for both diagnostic and control tasks.

For diagnostics, the approach involves:

  • Model-based diagnosis: Comparing observed system behavior to predicted behavior from a model to detect and identify faults. Neural networks are proposed as a way to learn these transforms, avoiding tedious model development.
  • Data-driven classification: Training neural networks on collected vehicle data to classify system states as either normal or faulty.
  • Iterative refinement: For misfire detection, an iterative process was used to remove misclassified data points from the training set to improve accuracy.
  • Feature extraction: Exploring techniques like principal component analysis to reduce the input vector size for computational efficiency on embedded microcontrollers.

For control, the methods include:

  • Emulating existing controllers: Training a NN to replicate the performance of a conventional controller for an active suspension system.
  • Model-independent control: Developing controllers for non-linear plants without explicit use of a plant model, relying instead on a cost function that compares observed performance to desired performance.

Datasets

The paper discusses the use of several datasets:

  • Powertrain fault diagnosis:
    • An initial dataset from a vehicle operated in a narrow range of speed and load for diagnosing common faults.
  • A subsequent dataset with a more restricted set of faults but drawn from a broader range of operating conditions, reflecting the need for real-time diagnostics.
  • "60 Pin Data": Data obtained directly from the engine.
  • "DCL Data": Data obtained through the control microprocessor via a multiplexed link.
  • Engine misfire detection:
    • Data collected from normal and misfiring engines under a modest range of dynamic conditions.
  • Data collected under a very wide range of conditions (potholed roads, severe accelerations/braking) to test performance limits.
  • Active suspension control:
    • Simulated data from a non-linear active suspension system.
  • Training data equivalent to less than 100 feet of bumpy road was sufficient for some control tasks.

Key Insights and Ideas

  • NNs excel where traditional methods falter: Neural networks proved capable of handling complex diagnostic tasks that had eluded automated solutions by other means, particularly in real-time scenarios.
  • Data quality and quantity are paramount: The performance of NN classifiers is highly dependent on the quality and representativeness of the training data. The problem implicitly directs the choice of classifier paradigm based on these factors.
  • On-board diagnostics are feasible: Shifting diagnostics to the control microprocessor enables continuous, real-time monitoring, but requires highly compact and efficient NN algorithms.
  • Model-independent control is achievable: Neural networks can learn to control non-linear systems effectively even without an accurate plant model, significantly reducing development effort and time.
  • Understanding NN solutions is an ongoing challenge: While NNs provide effective solutions, the underlying mechanisms are not fully understood, which is crucial for their widespread deployment.

Key Results

  • Diagnostic Accuracy: Neural networks were able to classify faults in powertrain systems and detect engine misfires with satisfactory accuracy, even in complex scenarios. For instance, in the misfire detection study, confusion matrices showed low misclassification rates for both limited and extended operation conditions.
  • Control Performance: A backpropagation network was trained to provide performance equivalent to a conventional controller for an active suspension system. A small network (2 hidden nodes) achieved satisfactory performance with less than 1000 training vectors.
  • Computational Efficiency: The paper highlights the need for compact NN architectures for on-board implementation, and suggests that techniques like principal component analysis can help reduce computational load.

Limitations and Future Work

  • Computational Constraints: The limited computational resources of current vehicle microcontrollers pose a significant challenge for deploying complex NN diagnostic algorithms. Future work involves reducing the input vector size and developing more compact NN forms.
  • Training Data Acquisition: Assembling large, accurate training datasets for supervised learning is a considerable task, especially for real-time diagnostics. The paper suggests that pre-trained classifiers might be the initial deployment strategy.
  • Understanding NN Solutions: The authors acknowledge that they are "just beginning to understand how these solutions arise." Further research into the internal workings of NNs is needed to fully determine their role in deployed systems.
  • Continuous Adaptation: While continuous adaptation of classifiers is desirable, it requires more computational capability and a supervised learning environment, which is not always feasible in automotive applications.

Summary updated

BibTeX
@inproceedings{1990neural,
  title = {{Neural Network Application to Diagnostics and Control of Vehicle Control Systems}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1990},
}
Figure