Industrial Anomaly Detection With Iterative Model Retraining

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing anomaly detection models in industrial environments face challenges in effectively detecting abnormal datapoints, especially in large unlabeled datasets, requiring significant training data and relying heavily on expert validation, and are not efficient in continuous learning from unlabeled and labeled datapoints.

Innovation Solution

A predictive maintenance system that applies anomaly detection models to subsets of operation data, iteratively classifies datapoints as normal or abnormal, updates the training dataset, and re-trains the models based on a threshold time, enabling continuous learning and robust anomaly detection even with limited labeled data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If existing anomaly detection models are trained with a large amount of pre-labelled data, then detection accuracy is improved, but data preparation complexity and time consumption increase

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

Solution Approach 1:

The system performs preliminary action by pre-training anomaly detection models with available labeled normal data before actual deployment. This preliminary training establishes a baseline detection capability that can operate with limited labeled data, reducing the need for extensive pre-labeling in production environments.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system implements feedback mechanisms where detection results are continuously fed back to update and retrain the models. This allows the system to improve detection accuracy over time using unlabeled data through iterative learning, reducing dependency on large amounts of pre-labelled training data.

Inventive Principle:
Principle #23Feedback

2Ease of manufacture

If anomaly detection models are trained with limited labeled data, then data preparation effort is reduced, but detection reliability deteriorates

Engineering Contradiction:
Improvemodel training easeVSAvoidanomaly detection reliability
Core Design Contradiction:
Ease of manufactureVSReliability

Solution Approach 1:

The system ensures continuity of useful action by implementing continuous learning and model updates using streaming data. The models are continuously retrained with newly accumulated labeled data from detection results, maintaining and improving reliability over time without requiring extensive initial labeled datasets.

Inventive Principle:
Principle #20Continuity of useful action

Solution Approach 2:

The system applies dynamics by making the model training process adaptive and flexible. The training regime dynamically adjusts based on available labeled data, using semi-supervised and self-supervised learning techniques that can operate effectively with varying amounts of labeled data, thus maintaining reliability while easing data preparation requirements.

Inventive Principle:
Principle #15Dynamics

3Measurement precision

If expert validation is used to verify abnormal datapoint detection, then detection accuracy is improved, but operational complexity increases

Engineering Contradiction:
Improveabnormal datapoint detection accuracyVSAvoidsystem operational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system implements self-service by enabling automatic validation and verification of detection results through the model itself. The anomaly detection models perform self-verification by comparing new detections against learned patterns, reducing or eliminating the need for expert validation while maintaining detection accuracy.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system uses feedback loops where detection results are automatically validated and used to refine model performance. This automated feedback mechanism replaces manual expert validation, reducing operational complexity while maintaining or improving detection accuracy through continuous self-improvement.

Inventive Principle:
Principle #23Feedback

4Quantity of substance

If traditional anomaly detection models are used with large unlabeled datasets, then data utilization is improved, but learning efficiency deteriorates

Engineering Contradiction:
Improvedata utilizationVSAvoidlearning efficiency
Core Design Contradiction:
Quantity of substanceVSProductivity

Solution Approach 1:

The system applies parameter changes by employing different learning strategies for different data types. Semi-supervised learning techniques are used to leverage unlabeled data efficiently, while self-supervised learning approaches enable the model to learn from raw data without extensive labeling, thus improving both data utilization and learning efficiency simultaneously.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4150421B1System, device and method of detecting abnormal datapoints
Publication Date: 2024.10.16 SIEMENS AG
  • EP4150421B1 patent drawingFigure 1
  • EP4150421B1 patent drawingFigure 2
  • EP4150421B1 patent drawingFigure 3

AI summary

System, Device and Method of detecting at least one abnormal datapoint in operation data (U) associated with an industrial environment (610) is disclosed. The method comprising iteratively applying one or more anomaly detection models (fi) to at least one subset (S) of the operation data (U), wherein the anomaly detection models (fi) are trained based on a training dataset (L) consisting of datapoints labeled as normal; classifying subset-datapoints in the subset (S) as one of normal datapoints (N) and abnormal datapoints (A) using the anomaly detection models (fi); updating the training dataset at least with the normal datapoints; retraining the anomaly detection models (fi) with the updated training dataset after expiration of a threshold time, wherein the threshold time is based on the number of updates to the training dataset; and detecting the at least one abnormal datapoint in the operation data (U) using the anomaly detection models (f'i).