← Platform Overview

AI & Analytics

Anomaly Detection

Multi-algorithm anomaly detection that catches equipment issues hours before traditional threshold alarms. Five algorithms running in consensus to minimize false positives.

SENSOR WAVEFORMVIB_X [mm/s]+3σ-3σTHRESHOLD: 4.5 mm/sCURRENT: 2.7 mm/sISOLATION FOREST??AN?NNANOMALY SCORE1.00.50.0THRESH0.15NORMALConfidence35%CROSS-SENSOR CORRELATIONVIBVIBTEMPTEMPFLOWFLOWPRESPRES1.000.650.140.490.731.000.270.440.480.071.000.020.030.330.251.00ALGORITHM CONSENSUSIsolation Forest0.12Autoencoder0.18SPC / CUSUM0.08One-Class SVM0.22Spectral Res.0.10CONSENSUS: NORMAL OPERATIONALERT TIMELINELAST 24 HOURS00:0002:0004:0006:0008:0010:0012:0014:0016:0018:0020:0022:0024:00WARNCRITWARNCRITWARNanomaly-engine@twinedge:~$ watch --sensors=4 --interval=1sVIB_X2.1 mm/sTEMP67.0 °CFLOW42.0 L/sPRES3.2 bar[00:00:00]ALL CLEAR -- 0/5 algorithms flaggedlatency: 22ms

Detection Algorithms

Five algorithms running in parallel, each with different strengths.

AlgorithmTypeStrengthLatency
Isolation ForestUnsupervisedCatches multi-dimensional outliers across correlated sensor groups<50ms
AutoencoderDeep LearningLearns complex normal operating patterns; detects subtle deviations<80ms
Statistical Process ControlStatisticalCUSUM and EWMA charts for gradual drift detection over time<5ms
One-Class SVMSemi-SupervisedEffective with limited training data; good for rare equipment types<30ms
Spectral ResidualFrequency DomainDetects periodic anomalies and unexpected frequency components<20ms

Detection Timeline

0hNormalAll sensors within baseline. Models continuously updating normal operating profile.
-6hEarly WarningAutoencoder reconstruction error rises 2.3 sigma. Cross-sensor correlation shifts detected.
-2hAnomaly ConfirmedIsolation Forest and SPC both flag. Multi-algorithm consensus triggers formal anomaly alert.
-1hAlert DispatchedAlert routed to operations team with root cause hypothesis, affected sensors, and recommended action.
0hThreshold AlarmTraditional alarm fires. But your team already diagnosed the issue 6 hours earlier.

Core Capabilities

Hours of Early Warning

Detect anomalies 2-8 hours before traditional threshold alarms fire. Enough lead time to plan a response.

Cross-Sensor Correlation

Analyze relationships between vibration, temperature, pressure, and flow simultaneously. Catch issues no single sensor reveals.

Adaptive Baselines

Models retrain continuously on recent data. Seasonal changes, load variations, and process shifts are learned automatically.

Consensus Scoring

Multiple algorithms must agree before an anomaly is confirmed. Reduces false positives by 85% compared to single-model approaches.

Detect Problems Before They Become Failures

Hours of early warning instead of seconds. Multi-algorithm consensus instead of single-threshold guessing.