Failure Mode Classification

Unlike binary anomaly detection, this task assigns a specific fault label to an input sample. A vibration signature is not just 'abnormal'; it is classified as inner race bearing wear, shaft misalignment, or cavitation. This requires a labeled dataset where each historical failure event is tagged with its root cause, enabling the model to learn the unique statistical fingerprints of distinct physical degradation mechanisms.

Raw sensor time-series data is rarely fed directly into a classifier. A preprocessing stage extracts discriminative features that separate failure modes:

• Time-domain: RMS, crest factor, kurtosis for impulsiveness.
• Frequency-domain: FFT magnitudes at specific fault frequencies (BPFO, BPFI).
• Time-frequency: Wavelet coefficients for non-stationary signals. Domain knowledge is critical here; the feature set must be engineered to amplify the differences between a gear tooth crack and a lubrication failure.

The choice of algorithm balances accuracy with interpretability:

• Support Vector Machines (SVM): Effective in high-dimensional feature spaces with clear margin separation.
• Random Forests: Provide feature importance scores, helping engineers understand which sensors are most diagnostic.
• 1D Convolutional Neural Networks (CNNs): Learn hierarchical features directly from raw time-series data, often eliminating manual feature engineering.
• Transformer Encoders: Apply self-attention to capture long-range dependencies in vibration sequences for complex, interacting faults.

Complex systems often use a hierarchical taxonomy. A first-level model might distinguish between electrical and mechanical faults. A second-level mechanical model then classifies the specific type: imbalance, looseness, or bearing defect. A third level might pinpoint the bearing defect to the inner race, outer race, or rolling element. This divide-and-conquer strategy improves accuracy on rare failure modes and aligns with a technician's diagnostic workflow.

A raw Softmax probability is often an overconfident estimate of true likelihood. Calibration techniques like Platt scaling or isotonic regression map model scores to true probabilities. A well-calibrated model ensures that a prediction of 'outer race spall with 90% confidence' is actually correct 90% of the time. This is essential for risk-based maintenance scheduling, where the cost of a false positive (unnecessary downtime) must be weighed against a false negative (catastrophic failure).

A closed-world classifier assumes all failure modes are known during training. In reality, previously unseen faults emerge. A robust system couples the classifier with a novelty detection module, often an autoencoder or a one-class SVM. If the reconstruction error of an input is high, the sample is flagged as an 'unknown failure mode' rather than being forcibly misclassified into a known category. This triggers a manual root cause analysis and subsequent model retraining.

The process of extracting discriminative statistical attributes from raw sensor data to enable accurate classification. Effective feature engineering transforms high-frequency vibration or acoustic signals into a compact, informative vector.

• Time-domain features: Root mean square (RMS), crest factor, kurtosis, and skewness capture signal energy and impulsiveness.
• Frequency-domain features: Spectral centroid, band power ratios, and envelope spectrum peaks isolate fault-specific frequencies.
• Time-frequency features: Wavelet packet energy and empirical mode decomposition coefficients capture non-stationary transients.

Domain expertise is critical—a bearing cage defect and a shaft misalignment may produce similar RMS values but distinct spectral signatures.

An algorithm that converts a time-domain vibration signal into its constituent frequency components, revealing the spectral fingerprint of specific failure modes. The FFT decomposes complex waveforms into a sum of sinusoids, each with a distinct amplitude and phase.

• Bearing faults manifest at characteristic frequencies: Ball Pass Frequency Inner Race (BPFI), Ball Pass Frequency Outer Race (BPFO), and Ball Spin Frequency (BSF).
• Gear mesh faults appear as sidebands around the gear mesh frequency, modulated by the shaft rotational speed.
• Unbalance produces a dominant peak at 1x running speed with low harmonic content.

Modern classification pipelines often feed FFT magnitude spectra directly into convolutional neural networks for automated pattern recognition.

A game-theoretic method for interpreting failure mode classification outputs by quantifying each sensor feature's contribution to a specific prediction. SHAP values decompose a model's output into additive feature attributions, ensuring local accuracy and consistency.

• Global interpretability: Aggregate SHAP values reveal which frequency bands or statistical moments most influence bearing fault vs. misalignment decisions.
• Local interpretability: For a single classification event, SHAP identifies the exact spectral peak that triggered the alert, enabling root cause validation.
• Model-agnostic: Works with gradient-boosted trees, support vector machines, and deep neural networks alike.

This explainability is essential for earning operator trust and meeting regulatory requirements in safety-critical industries.

A machine learning paradigm where a model trained on a source domain with abundant labeled failure data is adapted to a target domain with scarce examples. This is critical in manufacturing, where certain rare failure modes may have only a handful of historical occurrences.

• Pre-training: A convolutional neural network learns generalizable features from a large corpus of bearing vibration spectrograms.
• Fine-tuning: The final classification layers are retrained on a small, domain-specific dataset of gearbox or pump failures.
• Domain adaptation: Techniques like Maximum Mean Discrepancy (MMD) alignment minimize distribution shift between lab-generated training data and real-world operational data.

Transfer learning dramatically reduces the data acquisition burden for deploying classifiers on new equipment types.

The fundamental evaluation framework for multi-class failure mode classifiers, quantifying true positives, false positives, false negatives, and true negatives for each fault category. A well-constructed confusion matrix reveals systematic misclassification patterns.

• Precision: Of all instances predicted as 'inner race fault,' what fraction were correct? High precision minimizes unnecessary inspections.
• Recall: Of all actual 'inner race faults,' what fraction were detected? High recall ensures no critical failure goes undiagnosed.
• F1-score: The harmonic mean of precision and recall, providing a single metric robust to class imbalance.
• Macro vs. weighted averaging: Macro-averaging treats all failure modes equally; weighted averaging accounts for class frequency.

Misclassification cost is asymmetric—confusing a benign condition for a catastrophic failure wastes resources, but the reverse risks unplanned downtime.

The algorithmic combination of heterogeneous sensor streams—vibration, temperature, acoustic emission, and motor current—to create a unified feature representation for robust failure mode discrimination. No single sensor modality captures all fault signatures.

• Early fusion: Raw signals are concatenated before feature extraction, allowing the model to learn cross-modal correlations.
• Late fusion: Independent classifiers per modality are combined via voting or a meta-learner, providing modularity and interpretability.
• Attention-based fusion: Transformer architectures learn to dynamically weight modalities based on operational context.

For example, a shaft misalignment may produce subtle vibration changes but a pronounced thermal gradient, making temperature data critical for disambiguation.