Detection Model Retraining for Plant Anomaly Prediction Precision

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing techniques for detecting anomalous states in plants using unsupervised machine learning struggle to improve precision due to unlabeled pre-collected data, making it difficult to evaluate the learning model and determine its appropriateness.

Innovation Solution

A detection apparatus and method that collects measurement data, inputs it into a trained learning model, and generates a retrained model by using user-provided label values and collected data when the detection result differs from the user's determination, thereby improving the model's precision.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If unsupervised machine learning is performed on pre-collected normal data to generate a normal state model, then the model can be trained without labeled data, but the precision of the learning model to detect anomaly predictors is difficult to improve

Engineering Contradiction:
Improveprecision of learning modelVSAvoidlabeled data
Core Design Contradiction:
Measurement precisionVSLoss of information

Solution Approach 1:

The system implements feedback by having users check detection results and provide correction information. When a user determines that a detected anomaly predictor is incorrect or misses an actual anomaly, this feedback is used to generate new labeled training data, which is then used to retrain and improve the model's precision for future detections.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The system enables self-service by automatically generating labeled training data from user corrections without requiring manual labeling of all training data. The user-provided corrections are automatically converted into labeled datasets that the system uses to retrain its own model, allowing continuous self-improvement.

Inventive Principle:
Principle #25Self-service

2Productivity

If pre-collected data are not labelled, then data collection is simpler and faster, but an evaluation index of the learning model is unable to be calculated

Engineering Contradiction:
Improvedata collection efficiencyVSAvoidevaluation capability
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The system performs preliminary unsupervised learning on unlabeled data to generate initial detection results. These preliminary results serve as a foundation that users can then evaluate and correct, combining the efficiency of unlabeled data processing with the evaluation capability provided by subsequent user feedback.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

User feedback acts as an intermediary that bridges unlabeled data and model evaluation. The user's correction of detection results provides the labeling information needed to calculate evaluation indices, while maintaining the initial efficiency benefit of using unlabeled pre-collected data.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Ease of manufacture

If a learning model is trained without labeled data, then the training process is simpler, but it is difficult to determine whether the learning model is appropriate

Engineering Contradiction:
Improvemodel training simplicityVSAvoidmodel appropriateness
Core Design Contradiction:
Ease of manufactureVSReliability

Solution Approach 1:

User feedback on detection results provides a mechanism to assess model appropriateness. When users consistently correct certain types of detections or provide systematic corrections, this indicates the model may not be appropriate for certain scenarios, triggering retraining with the new labeled data to improve reliability.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The system performs self-validation through user feedback loops. By monitoring user corrections and automatically retraining when patterns of incorrect detections are identified, the system self-assesses and improves its appropriateness without requiring external validation processes.

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP4498198A1Detection apparatus, detection method, and detection program
Publication Date: 2025.01.29 YOKOGAWA ELECTRIC CORP
  • EP4498198A1 patent drawingFigure 1
  • EP4498198A1 patent drawingFigure 2
  • EP4498198A1 patent drawingFigure 3~4

AI summary

A server device (10) collects measurement data measured by a field device (20), inputs the measurement data collected, into a trained detection model that predicts a predetermined event in response to input of measurement data, obtains an output result from the trained detection model, executes, in a case where the output result from the trained detection model is different from a determination result by a worker W who has checked the field device (20) for which the predetermined event was predicted, retraining of the trained detection model by using a label value input by the worker W and the measurement data collected, and thereby generates a retrained model.