Determination of Machine Condition Using Neural Networks

John MacIntyre, Peter Smith, John TaitAAAI 1994

Abstract

This study explores the transition from manual, labor-intensive machinery maintenance to automated, "condition-based" monitoring at a major UK power station. By applying multi-layer perceptron (MLP) neural networks to vibration frequency data, researchers developed the Neural Bearing Analyzer (NBA), a system capable of identifying bearing defects with a 93% accuracy rate compared to human experts.

Introduction

In the early 1990s, industrial maintenance typically followed two paths: fixing machines only upon failure (breakdown maintenance) or adhering to rigid, often unnecessary, time-based schedules (planned maintenance). "Condition monitoring" offers a middle ground, where maintenance is performed only when physical indicators suggest wear or impending failure. However, manual analysis of these indicators is labor-intensive and requires significant expertise. This paper presents an automated approach using artificial intelligence to monitor Cooper rolling element bearings at the Blyth Power Station.

The Challenge of Condition Monitoring

The primary tool for monitoring machine health is vibration analysis. Sensors attached to machinery capture time-domain signals, which are then converted into the frequency domain via Fast Fourier Transform (FFT). The resulting frequency spectrum contains peaks that correspond to specific mechanical components and potential defects, such as cracked bearing races.

Vibration frequency spectrum from the Blyth Power Station showing characteristic peaks used for defect diagnosis.

As shown in the spectrum above, identifying these defects is not always straightforward. While some peaks are clear, others are obscured by background noise, low-frequency interference, or complex mechanical geometries. Manually analyzing these spectra across an entire power plant is a massive task, prompting the need for automated diagnostic tools.

Methodology: The Neural Bearing Analyzer

The researchers developed two distinct versions of the Neural Bearing Analyzer (NBA) using multi-layer perceptron (MLP) architectures trained via back-propagation:

  1. Feature-Based Model: This model utilized a limited set of pre-selected features from the frequency spectrum, such as specific peaks and known defect frequencies. It was designed to classify the severity of defects in individual bearing components.
  2. Raw Spectrum Model: This "brute force" approach used the entire 400-point frequency spectrum as input. The goal was to allow the neural network to identify complex patterns that might be too subtle or difficult for human experts to define manually.

The training data consisted of vibration signals from Cooper rolling element bearings at the 1,180MW Blyth Power Station. These signals were preprocessed using band-pass and envelope filters before FFT conversion.

Key Results and Insights

The NBA demonstrated significant success in replicating expert human judgment:

  • High Expert Agreement: The system achieved a 93% agreement rate with a professional consultant condition monitoring engineer on live test cases.
  • Handling Real-World Data: The neural networks proved robust enough to handle the "vague" and noisy data inherent in a working power plant environment, where ideal sensor placement is often impossible.
  • Convergence Challenges: The researchers noted that training the "Raw Spectrum Model" was significantly more difficult due to the high dimensionality of the 400-point input. However, once convergence was achieved, the model proved highly effective at extracting meaningful diagnostic features.

Conclusion and Future Work

The study successfully demonstrated that neural networks can automate complex diagnostic tasks in industrial settings. While the current system focuses on rolling element bearings, future work aims to address more complex machinery, such as coal mill gearboxes, which present higher background noise and more intricate gear-meshing frequencies. The team is also investigating alternative architectures, such as Self-Organizing Maps (SOMs), to improve the system's ability to categorize faults with less supervised oversight.

Summary updated

BibTeX
@inproceedings{macintyre1994determination,
  author = {John MacIntyre and Peter Smith and John Tait},
  title = {{Determination of Machine Condition Using Neural Networks}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}
Figure