AI-Powered Predictive Maintenance for Industrial Motors

The Problem

Industrial vibration data contains rich fault signatures, but those signatures can be difficult to separate without careful signal processing and modeling. The challenge was to build a fault-classification pipeline that combines vibration-signal analysis, physically meaningful features, and data-driven learning to distinguish multiple rotating-machinery fault conditions.

The Approach

• Vibration-signal preprocessing and normalization.
• FFT-based feature extraction and rotational harmonics.
• Statistical features including RMS, kurtosis, skewness, and crest factor.
• CNN-based representation learning integrated with physics-informed constraints in a hybrid diagnosis model.
• Evaluation across normal operation, imbalance, misalignment, and bearing-fault conditions.

Technical Decisions

• Spectral and statistical features were used to preserve physically meaningful information from vibration signals.
• The modeling pipeline combines data-driven learning with engineering knowledge instead of treating raw vibration measurements as arbitrary input.
• Physics-informed constraints were incorporated into the hybrid diagnosis framework alongside classification and data-matching objectives.
• The MAFAULDA dataset provides a controlled benchmark for rotating-machinery fault classification.

Results

The final classification pipeline achieved 98% overall accuracy on the MAFAULDA rotating-machinery fault dataset, demonstrating strong separation across the evaluated operating and fault conditions.

Why It Matters

The project demonstrates how signal processing, domain-informed feature engineering, deep learning, and physics-informed modeling can be combined to identify mechanical fault conditions from vibration data. This type of pipeline can form the diagnostic layer of broader condition-monitoring and predictive-maintenance systems.