Autoencoder Anomaly Detection for Engineering Assets

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current Prognostics and Health Management (PHM) systems face challenges in real-world applications due to complex engineering assets with manifold operating environments and long-term operations, leading to data biases and 'system drift' from hardware and software changes, maintenance activities, and limited failure data, resulting in inaccurate anomaly detection and failure prediction.

Innovation Solution

A method and system for training an autoencoder using real-time sensor data from engineering assets, which involves fitting an autoencoder to training data, generating under-represented data sets, cloning the model, aggregating cloned autoencoders, calculating error data sets, and estimating anomaly scores using Mahalanobis distance or squared Mahalanobis distance to identify anomalous behavior and its contributing sensor readings.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If current PHM systems are applied to complex engineering assets with manifold operating environments, then they can provide health management capabilities, but they produce unacceptable false results due to system drift and data biases

Engineering Contradiction:
Improveanomaly detection accuracyVSAvoidoperating environment variability
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The patent segments the training process into multiple iterations, each handling a subset of the data. The data is divided into training sets and test sets across multiple iterations, allowing the model to progressively adapt to different operating conditions without being overwhelmed by the full complexity of all environments simultaneously.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary data processing and segmentation before the main training process. Training data is pre-processed to identify and separate different operating conditions, and the model is prepared with initial configurations that enable it to handle system drift and data biases before actual anomaly detection begins.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If statistical methods are used to detect anomalies, then they can identify clearly labelled data sets, but they fail to handle sparsity of failure examples and unlabelled data sets from maintenance activities

Engineering Contradiction:
Improveanomaly score calculationVSAvoidavailability of labelled failure data
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent implements self-service through unsupervised learning where the model learns normal operating patterns from unlabeled data without requiring explicit failure labels. The autoencoder structure enables the system to serve itself by automatically identifying anomalies through reconstruction error rather than relying on externally provided labeled failure data.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent changes the approach from using labeled failure data to using reconstruction error metrics. By transforming the problem from classification (requiring labels) to anomaly detection based on reconstruction quality, the system can work with unlabelled data and maintain measurement precision without needing abundant labeled failure examples.

Inventive Principle:
Principle #35Parameter changes

3Adaptability or versatility

If multiple data sets are generated from encodings to address under-representation, then the model can handle diverse operating conditions, but the training process becomes computationally intensive requiring model cloning and aggregation

Engineering Contradiction:
Improvecoverage of operating envelopeVSAvoidtraining process complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent uses copying by creating multiple cloned autoencoders, each trained on different subsets of the data. These clones are then aggregated to form a more robust model that handles diverse operating conditions. The copying approach allows parallel training of multiple specialized models that can be combined for comprehensive coverage.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent merges multiple cloned autoencoders into a single aggregated model. By combining the strengths of multiple specialized models, each trained on different data subsets, the system achieves comprehensive coverage of the operating envelope while managing computational complexity through distributed training and aggregation.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP4152210A1System and method for training an autoencoder to detect anomalous system behaviour
Publication Date: 2023.03.22 BAE SYSTEMS PLC
  • EP4152210A1 patent drawingFigure 1
  • EP4152210A1 patent drawingFigure 2
  • EP4152210A1 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..