Description
The APS (Air Processing System) is critical for the operation of Scania trucks. APS failures can lead to significant industrial costs and production downtime. Historical sensor data enables the analysis of system behavior and early detection of anomalies before they escalate into critical failures.
In this study, a Random Forest model was trained to detect APS system failures, with special attention given to the highly imbalanced nature of the dataset by applying class weighting to the rare positive instances. Model performance was evaluated using standard metrics including precision, recall, F1-score, and confusion matrices, while the associated industrial cost was calculated based on the impact of false positives and false negatives. To further optimize failure detection and reduce total costs, XGBoost was optionally employed. Additionally, feature importance analysis was conducted to identify the most critical sensors influencing APS failure predictions, with the top features visualized through horizontal bar charts to provide interpretable insights for industrial decision-making.
Figure 1 illustrates the confusion matrix of the Random Forest model applied to APS failure detection under a highly imbalanced class distribution. The model correctly classifies the majority of non-failure cases, as reflected by the large number of true negatives and the very low number of false positives. This indicates a strong capability to avoid unnecessary maintenance actions. However, a non-negligible number of APS failures are misclassified as normal operations (false negatives), which represent critical cases from an industrial perspective due to their high associated cost. This result emphasizes the importance of cost-sensitive learning and motivates the use of alternative models, such as XGBoost, to further reduce false negatives and optimize industrial cost. In the confusion matrix, the value 15,607 corresponds to true negatives, that is, trucks that do not have an APS system failure and are correctly identified as such by the model, demonstrating its ability to avoid unnecessary maintenance interventions.
Figure 2 presents the top 15 most influential sensor variables used by the Random Forest model to predict APS failures. The results indicate that the prediction is driven by a limited subset of sensors, with aa_000 being the most dominant feature, followed by ci_000, ck_000, and dn_000. This suggests that APS failures are strongly associated with specific operational measurements rather than uniformly across all sensors. The concentration of importance among these variables highlights their potential relevance for targeted monitoring and preventive maintenance strategies, as focusing on key sensors could improve fault detection efficiency while reducing system complexity.
Conclusion
The experimental results confirm that the proposed machine learning approach is effective for detecting failures of the Air Pressure System (APS) in heavy-duty trucks. The Random Forest model achieved a high overall accuracy (≈99%) and a strong precision for the negative class, indicating reliable identification of non-failure cases. However, the recall for APS failures remained moderate (≈58–61%), highlighting the intrinsic difficulty of detecting rare failure events in highly imbalanced industrial datasets.
The integration of class weighting significantly reduced the number of false negatives, which are associated with the highest industrial cost. Using the defined cost function, the Random Forest model resulted in a total industrial cost of approximately 74,000–78,000 units, demonstrating a meaningful improvement over non–cost-aware baselines. Furthermore, the XGBoost model substantially outperformed Random Forest in cost optimization, reducing the total industrial cost to approximately 29,850 units, primarily by further decreasing missed APS failures.
Feature importance analysis revealed that a limited set of sensor variables (e.g., aa_000, ci_000, ck_000, dn_000) consistently contributed most to the predictive performance. This suggests that APS degradation can be detected through specific operational patterns captured by onboard sensors. Overall, these results validate the relevance of cost-sensitive learning for industrial reliability studies and demonstrate the practical value of data-driven predictive maintenance in intelligent transportation systems.
