Locomotive Fault Diagnostics
Client: Iran Railways
Year: 2015
Type: Industrial research
Problem
General Motors (GM) locomotives used by Iran Railways rely on DC traction motors that degrade over time. Detecting faults early (bearing wear, armature imbalance, commutator defects) prevents costly unplanned downtime and improves safety. Traditional inspection requires taking the locomotive out of service.
What Was Novel
I combined Discrete Wavelet Transform (DWT) for vibration signal decomposition with a Learning Vector Quantization (LVQ) neural network for fault classification. The DWT decomposes the raw vibration signal into frequency bands that correspond to specific fault signatures (e.g., bearing defects produce characteristic high-frequency patterns). The LVQ network then classifies the decomposed features into fault categories.
The novelty was in the feature extraction pipeline: rather than using generic statistical features (RMS, kurtosis), I selected DWT coefficients at specific decomposition levels that align with the physical fault frequencies of the GM traction motor. This domain-informed feature selection is what pushed accuracy to 90% with a relatively simple classifier.
Results
- 90% accuracy in early fault detection from vibration data
- Real-time classification capability
- Reduced unplanned downtime for Iran Railways fleet