Semi-Supervised Anomaly Detection

Anomaly Detection

Stack
  • Python

Anomaly detection for improving heavy-truck failure detection by 70%

Semi-Supervised Anomaly Detection
GitHub

Overview

A predictive-maintenance system that flags likely failures of a critical engine component (“Component X”) across a fleet of heavy-duty trucks. Each truck trip is scored, where a higher anomaly score means the component is more likely to fail in the near future. The project was built under a deliberate constraint of using no deep learning, relying only on classical data-mining methods to keep the solution interpretable and effective on scarce, imbalanced data.

The Problem

The work uses the SCANIA Component X dataset from the IDA 2024 Industrial Challenge. The setup is semi-supervised: the training set holds 318 trips that are all normal, so the model must learn the normal operational profile and flag deviations from it. The validation (67) and test (167) sets mix normal and failing trips. Each trip is a multivariate sensor time series with irregular sampling and missing readings, joined to a static table of truck specifications.

Pre-processing

  • Time series regularization: linear interpolation onto a regular interval, since trips had irregular sampling rates and varying lengths.
  • Missing value handling: a binary missing-value indicator per feature, then a forward-fill followed by a back-fill, applied independently per trip so values never leak across trips.
  • Data integration: the processed sensor series joined to the static truck-specification table on trip_id.

Approach

The work progressed through three stages, all built on tree-based and distance/density-based detectors evaluated with Average Precision (AP) as the primary metric and AUC-ROC as a secondary one.

Isolation Forest baseline pipeline
The Isolation Forest baseline pipeline: merge and clean the data, process each variable type, score every trip timestep, then take the maximum anomaly score per trip as its trip-level signal.
  1. Isolation Forest baseline: minimal features scored a validation AP of 0.3126. Adding temporal features from the interpolated series (percentage change between steps, plus min/max/mean/std summaries and sequence metadata) lifted validation AP to 0.5324, about a 70% gain, confirming that temporal dynamics carry far more signal than raw counter values.
  2. Score-averaging ensemble: anomaly scores were averaged across complementary detectors (Isolation Forest, HBOS, KNN, LOF, ABOD, OCSVM, INNE), chosen because precision-oriented and ranking-oriented models offset each other’s weaknesses. The best six-model ensemble reached a validation AP of 0.5879 and AUC-ROC of 0.7642.
  3. Rich feature engineering: smoothing, first and second derivatives, histogram statistics (mean, standard deviation, skew, kurtosis, Shannon entropy), and short (10-step) plus long (100-step) rolling-window aggregates, reduced to 200 principal components via PCA, with each window scored individually.

Trip-level scores were produced by taking the maximum anomaly score across a trip’s windows, on the assumption that a single highly anomalous moment should flag the entire trip.

Final Model

The submitted model was a six-detector ensemble (Isolation Forest, HBOS, INNE, ABOD, OCSVM, LOF) with averaged anomaly scores, reaching a validation AP of 0.5803 and AUC-ROC of 0.7653. PCA and t-SNE visualizations were used throughout to check whether each new feature improved class separability.

Validation performance versus ensemble size
Validation AP (primary) and AUC-ROC (secondary) across ensemble sizes, with AP peaking at six models.

Key Technologies

  • Isolation Forest, HBOS, KNN, LOF, ABOD, OCSVM, INNE: classical outlier detectors combined into a score-averaging ensemble
  • Linear interpolation and rolling-window features: temporal feature engineering on irregular sensor series
  • PCA / t-SNE: dimensionality reduction and visual diagnostics

Results

Temporal feature engineering, not detector choice, drove the gains: features from the interpolated series alone lifted Isolation Forest AP about 70%, the single largest jump in the project. The final ensemble then cleared the All-Zeros (AP 0.2687) and Random-Guess (AP 0.2852) baselines by a wide margin, showing that a carefully composed set of classical detectors can surface rare engine failures from noisy, irregular sensor data without any deep learning.