SmartBuildSim

Anomaly Detection

smartbuildsim.models.anomaly labels anomalous sensor readings with IsolationForest, a teraz również wspiera porównania z LocalOutlierFactor przez smartbuildsim.evaluation.benchmark.run_anomaly_benchmark.

Configuration

AnomalyDetectionConfig fields:

• sensor: sensor name to monitor.
• contamination: expected proportion of anomalies (passed to IsolationForest).
• random_state: ensures deterministic model output via smartbuildsim.config.
• rolling_window: forwarded to FeatureConfig to control rolling statistics.

Access the derived FeatureConfig via the feature_config property when calling smartbuildsim.features.engineering.engineer_features directly.

Detection workflow

detect_anomalies(data, config) performs the following steps:

1. Filter rows for the configured sensor.
2. Engineer rolling mean, standard deviation, and optional derivative features.
3. Train an IsolationForest with the engineered features.
4. Append anomaly_score and is_anomaly columns to the returned dataframe.

CLI usage

smartbuildsim model anomalies examples/configs/default.yaml

The command writes outputs/anomalies.csv, which is then consumed by smartbuildsim viz plot for annotated visualisations.

Python example

from smartbuildsim.data.generator import DataGeneratorConfig, generate_dataset
from smartbuildsim.models.anomaly import (
    AnomalyDetectionConfig,
    detect_anomalies,
)
from smartbuildsim.scenarios.presets import get_scenario

scenario = get_scenario("office-small")
dataset = generate_dataset(
    scenario.building, DataGeneratorConfig(**scenario.data.dict())
)
config = AnomalyDetectionConfig(**scenario.anomaly.dict())
result = detect_anomalies(dataset, config)

flagged = result.data[result.data["is_anomaly"]]
print(flagged[["timestamp", "value", "anomaly_score"]].head())

See examples/scripts/run_example.py for the full pipeline that chains anomaly labelling with clustering and plotting. Porównawcze eksperymenty (krzyżowa walidacja, testy istotności i analiza skalowania) są dostępne w examples/scripts/run_benchmarks.py.

This site is open source. Improve this page.