Anomaly Detection

Anomaly Detection

Train and use machine learning models to detect anomalous vibration patterns. Supports OneClassSVM and Local Outlier Factor (LOF) algorithms with PCA visualization for clustering analysis.

Prerequisite: The predictive-maintenance-mcp MCP server must be connected.

Workflow

Training a New Model

Step 1 — Prepare Normal Baseline Data

Gather a set of signals representing NORMAL machine operation. Call list_stored_signals() to check available signals, or load signals with load_signal(...).

Ask the user:

• Which signals represent normal/healthy operation?
• What model algorithm to use? (OneClassSVM recommended for small datasets, LOF for larger ones)

Step 2 — Extract Features

For each baseline signal, call extract_features_from_signal(signal_id=...) to verify data quality. Signals with very high kurtosis or ISO Zone C/D should be excluded from the normal baseline.

Step 3 — Train the Model

Call train_anomaly_model(signal_ids=[...], model_type="oneclass_svm").

Parameters:

• signal_ids: list of signal IDs for normal baseline
• model_type: "oneclass_svm" or "lof" (Local Outlier Factor)
• contamination: expected fraction of anomalies in training data (default 0.05, lower = stricter)

The model is saved to the models/ directory for future use.

Step 4 — Validate

Run predict_anomalies(signal_ids=[...]) on the training signals themselves. Most should be classified as normal (anomaly_score near 0). If too many are flagged, increase contamination or review the baseline data.

Predicting Anomalies on New Signals

Step 1 — Load Target Signals

Load the signals to analyze with load_signal(...).

Step 2 — Run Prediction

Call predict_anomalies(signal_ids=[...]).

The tool returns for each signal:

• anomaly_score: distance from normal (higher = more anomalous)
• is_anomaly: boolean classification
• confidence: model confidence

Step 3 — Interpret Results

Anomaly Score	Classification	Action
Low (near 0)	Normal	No action needed
Medium	Borderline	Monitor more frequently
High	Anomaly	Investigate with bearing-diagnosis or quick-screening

Step 4 — Visualize with PCA

Call generate_pca_visualization_report(signal_ids=[...]) to project all signals into 2D space. Normal signals cluster together; anomalies appear as outliers.

Important Notes

• Training requires multiple signals from the SAME machine under normal conditions — mixing machines produces unreliable models
• Models are deterministic when random seeds are fixed
• Retraining is needed when machine operating conditions change significantly
• The anomaly model supplements, but does NOT replace, physics-based diagnosis
• All ML processing happens locally — no data leaves the machine