Ensemble Autoencoder Training for Drift-Resilient Anomaly Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing PHM systems struggle with integrating real-world data from complex engineering assets, particularly in maritime vessels, due to system drift, confounding variables, and under-represented data sets, leading to inaccurate anomaly detection and failure prediction.

Innovation Solution

A method and system for training an autoencoder using a combination of clustering, KDtree algorithm, and bagging to handle under-represented data, incorporating Mahalanobis distance for anomaly scoring and decomposition, enabling real-time anomaly detection and failure prediction.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If traditional PHM algorithms are applied to real-world engineering assets, then the system can operate in complex environments, but the detection accuracy deteriorates due to system drift and data irregularities

Engineering Contradiction:
Improveadaptability to complex operating environmentsVSAvoidanomaly detection accuracy
Core Design Contradiction:
Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The system dynamically adapts to changing operating conditions by continuously updating the autoencoder model with new data from different operating envelopes. The model evolves over time to accommodate system drift and environmental changes, maintaining detection accuracy despite varying operational contexts.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The system changes key parameters including training data composition (incorporating under-represented failure modes), model architecture (using ensemble autoencoders), and evaluation metrics (Mahalanobis distance) to improve detection accuracy while maintaining adaptability to complex environments.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If more comprehensive training data is collected to cover all operating envelopes, then the model becomes more accurate, but the data collection time and complexity increase

Engineering Contradiction:
Improveanomaly detection accuracyVSAvoiddata collection time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system performs preliminary actions by proactively collecting and storing data from all possible operating envelopes during normal operation, even before failures occur. This preparatory data collection enables the model to be trained on comprehensive datasets that include rare failure modes, improving accuracy without extending the actual deployment time.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system creates synthetic copies of failure scenarios by using data augmentation techniques and transfer learning from similar systems. This allows the model to learn from multiple operating envelopes and failure modes without requiring physical access to every possible failure scenario, reducing data collection time while maintaining accuracy.

Inventive Principle:
Principle #26Copying

3Reliability

If the autoencoder model is trained on under-represented data sets, then the detection of rare failure modes improves, but the training data quality and consistency deteriorate

Engineering Contradiction:
Improvedetection of rare failure modesVSAvoiddata consistency
Core Design Contradiction:
ReliabilityVSStability of the object's composition

Solution Approach 1:

The system merges multiple data sources including under-represented failure modes, normal operating data, and synthetic generated data into a unified training dataset. This combination balances the representation of rare failure modes with overall data consistency, allowing the model to detect rare events while maintaining stable performance across all operating conditions.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The system introduces intermediary processing steps including data normalization, feature engineering, and anomaly filtering that mediate between raw inconsistent data and the training model. These intermediary layers standardize the data composition while preserving the distinctive characteristics of under-represented failure modes, improving both reliability and consistency.

Inventive Principle:
Principle #24Intermediary (Mediator)

4Measurement precision

If multiple cloned autoencoders are aggregated to handle different data sets, then the detection accuracy for diverse failure modes improves, but the system complexity increases

Engineering Contradiction:
Improveanomaly detection accuracyVSAvoidmodel architecture complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system segments the overall detection task into multiple specialized autoencoder models, each trained on specific operating envelopes or failure modes. This segmentation allows each model to specialize in detecting particular types of anomalies, improving overall accuracy while maintaining manageable complexity through modular architecture and independent training of each segment.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4402610B1System and method for training an autoencoder to detect anomalous system behaviour
Publication Date: 2025.08.06 BAE SYSTEMS PLC
  • EP4402610B1 patent drawingFigure 1
  • EP4402610B1 patent drawingFigure 2
  • EP4402610B1 patent drawingFigure 3a

AI summary

The invention relates to a system and method for detecting anomalous system behaviour. The system comprises a plurality of sensors and a trained autoencoder. The method of training comprises: obtaining training data and test data comprising multiple data records for at least one engineering asset which corresponds to the engineering asset whose behaviour is to be classified, wherein the data records comprise a plurality of sensor readings for the engineering asset; fitting the autoencoder to the obtained training data; running the test data through the encoder of the fitted autoencoder to obtain encodings of the test data; generating a plurality of data sets from the obtained encodings, wherein the generated plurality of data sets include under-represented data sets; cloning the fitted autoencoder to create a cloned autoencoder for each of the generated plurality of data sets; and aggregating the cloned autoencoders to form an over-arching autoencoder. The method further comprises calculating an error data set between the training data and data reconstructed by the over-arching auto encoder; obtaining, using the calculated error data set, estimated parameters for calculating an anomaly score for each data record, wherein the anomaly score is selected from a Mahalanobis distance and a squared Mahalanobis distance; and estimating, using the calculated error set, parameters for calculating a decomposition of the anomaly score to identify a contribution from each sensor reading to the anomaly score.