HF
H.J. Frima
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
1 records found
1
Explainable Unsupervised Anomaly Detection for Predictive Maintenance
Using a Feed-Forward Neural Network
Master thesis
(2026)
-
H.J. Frima, M. Wisse, I.I. de Pater, Martijn Oerlemans, Konstantinos Pyrovolakis
Data-driven methods are increasingly used in aircraft maintenance to detect component degradation before operational disruptions occur. In aviation, however, predictive maintenance remains challenging due to the scarcity of labelled failure data, the high dimensionality of operational sensor data and the limited interpretability of many anomaly detection models. This thesis develops an unsupervised, SHapley Additive exPlanations (SHAP)-supported anomaly detection framework for detecting and localizing degradation in trim air valves (TAVs) within an aircraft air-conditioning system (ACS).
The proposed framework transforms high-frequency operational flight data into aggregated flight-level features and trains a feed-forward neural network (FFNN) to learn healthy TAV behaviour. Deviations between predicted and observed valve behaviour are converted into valve-level anomaly scores, enabling component-specific alarm generation instead of a single system-level anomaly indicator. SHAP-based explainability is applied to selected alarm events to assess whether the model predictions can be related to physically meaningful sensor-group contributions.
The framework is evaluated on real-world fleet data comprising approximately 34,000 flights from 31 wide-body aircraft and 25 reported TAV failure events. The results suggest that aggregated flight-level features can preserve degradation-relevant information while substantially reducing data volume. Under the selected decision-interval criteria, the simplified phase- and target-specific FFNN configuration achieved the most favourable operational trade-off, detecting 4 of the 25 reported failures while reducing false-positive alarm events from 39 to 16 compared with the operational baseline. In the detected failure cases, valve-level anomaly scores generally localized the alarms to the degrading valve. SHAP analysis provided useful diagnostic context for selected alarms, although the interpretability of these explanations was not consistent across all events. An exploratory generalization study on trim air pressure valve (TAPV) degradation further suggests that the framework can be applied to a related ACS component, although the most suitable target representation may be component-dependent.
Overall, the results indicate that unsupervised anomaly detection based on aggregated flight-level features can support predictive maintenance in aviation by combining scalable data representation, component-level localization and case-dependent diagnostic insight. However, the findings remain conditional on the limited number of failure events, the single-fleet evaluation, and the case-dependent nature of SHAP. Further validation on larger datasets, independent fleets and additional failure modes is required before broader operational applicability can be established. ...
The proposed framework transforms high-frequency operational flight data into aggregated flight-level features and trains a feed-forward neural network (FFNN) to learn healthy TAV behaviour. Deviations between predicted and observed valve behaviour are converted into valve-level anomaly scores, enabling component-specific alarm generation instead of a single system-level anomaly indicator. SHAP-based explainability is applied to selected alarm events to assess whether the model predictions can be related to physically meaningful sensor-group contributions.
The framework is evaluated on real-world fleet data comprising approximately 34,000 flights from 31 wide-body aircraft and 25 reported TAV failure events. The results suggest that aggregated flight-level features can preserve degradation-relevant information while substantially reducing data volume. Under the selected decision-interval criteria, the simplified phase- and target-specific FFNN configuration achieved the most favourable operational trade-off, detecting 4 of the 25 reported failures while reducing false-positive alarm events from 39 to 16 compared with the operational baseline. In the detected failure cases, valve-level anomaly scores generally localized the alarms to the degrading valve. SHAP analysis provided useful diagnostic context for selected alarms, although the interpretability of these explanations was not consistent across all events. An exploratory generalization study on trim air pressure valve (TAPV) degradation further suggests that the framework can be applied to a related ACS component, although the most suitable target representation may be component-dependent.
Overall, the results indicate that unsupervised anomaly detection based on aggregated flight-level features can support predictive maintenance in aviation by combining scalable data representation, component-level localization and case-dependent diagnostic insight. However, the findings remain conditional on the limited number of failure events, the single-fleet evaluation, and the case-dependent nature of SHAP. Further validation on larger datasets, independent fleets and additional failure modes is required before broader operational applicability can be established. ...
Data-driven methods are increasingly used in aircraft maintenance to detect component degradation before operational disruptions occur. In aviation, however, predictive maintenance remains challenging due to the scarcity of labelled failure data, the high dimensionality of operational sensor data and the limited interpretability of many anomaly detection models. This thesis develops an unsupervised, SHapley Additive exPlanations (SHAP)-supported anomaly detection framework for detecting and localizing degradation in trim air valves (TAVs) within an aircraft air-conditioning system (ACS).
The proposed framework transforms high-frequency operational flight data into aggregated flight-level features and trains a feed-forward neural network (FFNN) to learn healthy TAV behaviour. Deviations between predicted and observed valve behaviour are converted into valve-level anomaly scores, enabling component-specific alarm generation instead of a single system-level anomaly indicator. SHAP-based explainability is applied to selected alarm events to assess whether the model predictions can be related to physically meaningful sensor-group contributions.
The framework is evaluated on real-world fleet data comprising approximately 34,000 flights from 31 wide-body aircraft and 25 reported TAV failure events. The results suggest that aggregated flight-level features can preserve degradation-relevant information while substantially reducing data volume. Under the selected decision-interval criteria, the simplified phase- and target-specific FFNN configuration achieved the most favourable operational trade-off, detecting 4 of the 25 reported failures while reducing false-positive alarm events from 39 to 16 compared with the operational baseline. In the detected failure cases, valve-level anomaly scores generally localized the alarms to the degrading valve. SHAP analysis provided useful diagnostic context for selected alarms, although the interpretability of these explanations was not consistent across all events. An exploratory generalization study on trim air pressure valve (TAPV) degradation further suggests that the framework can be applied to a related ACS component, although the most suitable target representation may be component-dependent.
Overall, the results indicate that unsupervised anomaly detection based on aggregated flight-level features can support predictive maintenance in aviation by combining scalable data representation, component-level localization and case-dependent diagnostic insight. However, the findings remain conditional on the limited number of failure events, the single-fleet evaluation, and the case-dependent nature of SHAP. Further validation on larger datasets, independent fleets and additional failure modes is required before broader operational applicability can be established.
The proposed framework transforms high-frequency operational flight data into aggregated flight-level features and trains a feed-forward neural network (FFNN) to learn healthy TAV behaviour. Deviations between predicted and observed valve behaviour are converted into valve-level anomaly scores, enabling component-specific alarm generation instead of a single system-level anomaly indicator. SHAP-based explainability is applied to selected alarm events to assess whether the model predictions can be related to physically meaningful sensor-group contributions.
The framework is evaluated on real-world fleet data comprising approximately 34,000 flights from 31 wide-body aircraft and 25 reported TAV failure events. The results suggest that aggregated flight-level features can preserve degradation-relevant information while substantially reducing data volume. Under the selected decision-interval criteria, the simplified phase- and target-specific FFNN configuration achieved the most favourable operational trade-off, detecting 4 of the 25 reported failures while reducing false-positive alarm events from 39 to 16 compared with the operational baseline. In the detected failure cases, valve-level anomaly scores generally localized the alarms to the degrading valve. SHAP analysis provided useful diagnostic context for selected alarms, although the interpretability of these explanations was not consistent across all events. An exploratory generalization study on trim air pressure valve (TAPV) degradation further suggests that the framework can be applied to a related ACS component, although the most suitable target representation may be component-dependent.
Overall, the results indicate that unsupervised anomaly detection based on aggregated flight-level features can support predictive maintenance in aviation by combining scalable data representation, component-level localization and case-dependent diagnostic insight. However, the findings remain conditional on the limited number of failure events, the single-fleet evaluation, and the case-dependent nature of SHAP. Further validation on larger datasets, independent fleets and additional failure modes is required before broader operational applicability can be established.