Feature

Anomaly Detection

An LSTM autoencoder continuously monitors 13 sensor-derived features and raises precision alerts the moment equipment behaviour deviates from normal — typically in under 2 seconds.

87.3%
F1 Score
91.2%
Precision
83.7%
Recall
<2s
Detection latency
Start Free Trial →

How it works

01

Feature engineering

13 engineered features including temp_diff, power, and thermal ratio are computed from raw 5-sensor input.

02

Z-score normalisation

Each feature is normalised using training-set statistics before inference — preventing scale bias in LSTM reconstruction.

03

LSTM reconstruction

A (None, 10, 13) autoencoder reconstructs the input window. Elevated reconstruction error signals anomalous behaviour.

04

Ensemble scoring

0.6×LSTM_norm + 0.4×RF_prob = ensemble score. Threshold >0.5 triggers an alert with severity classification.

Input features (13)

FeatureSourceDescription
air_temperatureRaw sensorAmbient air temperature (K)
process_temperatureRaw sensorProcess temperature (K)
rotational_speedRaw sensorSpindle RPM
torqueRaw sensorTorque (Nm)
tool_wearRaw sensorCumulative tool wear (min)
temp_diffEngineeredprocess_temp − air_temp
powerEngineeredtorque × rotational_speed × constant
thermalEngineeredtemp_diff / process_temp
type_HEncodedMachine type high quality
type_LEncodedMachine type low quality
type_MEncodedMachine type medium quality
rpm_torque_ratioEngineeredrotational_speed / torque
wear_temp_interactEngineeredtool_wear × temp_diff

Severity classification

Critical
Score ≥ 0.8
High
Score 0.6 – 0.8
Medium
Score 0.4 – 0.6
Normal
Score < 0.4