Fixed vibration thresholds set years ago degrade over time as equipment changes, generating false alerts and missing real faults. Isolation Forest offers a better approach: trained only on healthy sensor readings, it produces a continuous anomaly score reflecting how statistically unusual a reading is relative to the current fleet baseline. Combined with an XGBoost fault classifier, the two scores create a four-quadrant decision framework — familiar faults, unfamiliar faults, novel anomalies, and normal operation — each mapping to a different operational response. The post walks through training the Isolation Forest with scikit-learn, persisting the model, writing scores to CrateDB alongside fault probabilities, and querying the four quadrants in SQL for Grafana dashboards or maintenance reports. The key insight is that threshold alerting, anomaly detection, and fault classification are complementary layers, not alternatives.
