Abnormality Determination Using Two-Stage Auto-Encoders
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing abnormality detection systems in large-scale power plants face challenges in accurately determining slight changes and stepwise changes in sensor data, leading to erroneous predictions and failure to detect abnormalities due to correlation issues between sensors and spurious correlations in machine learning models.
Innovation Solution
The use of a two-stage machine learning approach with first and second prediction models, where the first model trains on feature amounts of normal state data and the second model trains on slight fluctuations, improving prediction accuracy by distinguishing between main components and slight fluctuations in plant data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single machine learning model is used to predict sensor values, then the model complexity is low, but the prediction accuracy deteriorates due to inability to distinguish between main components and slight fluctuations
Solution Approach 1:
The patent divides the prediction task into two separate models: a first prediction model that predicts main sensor values and a second prediction model that predicts slight fluctuations. This segmentation allows each model to specialize in specific aspects of the data, improving overall prediction accuracy while managing complexity through functional division.
Solution Approach 2:
The patent introduces a new dimension of prediction by adding the second model that specifically targets slight fluctuations as a separate predictive layer. This dimensional addition to the prediction space enables capturing subtle variations that a single model would miss, thereby improving measurement precision.
2Reliability
If sensor correlation models are used to detect abnormalities, then the ability to detect abnormal changes is improved, but false positives increase due to spurious correlations between sensors
Solution Approach 1:
The patent segments the prediction task into main value prediction and fluctuation prediction, which helps distinguish genuine abnormalities from spurious correlations. By separating these functions, the system reduces false positives while maintaining reliable abnormality detection.
Solution Approach 2:
The patent applies different prediction strategies to different aspects of sensor data: the first model handles main sensor values while the second model specifically handles slight fluctuations. This local quality differentiation improves detection accuracy by treating different data characteristics with appropriate methods.
3Measurement precision
If regression models are used to predict sensor values, then the model simplicity is maintained, but the ability to detect slight changes deteriorates
Solution Approach 1:
The patent segments the detection task into main value prediction (first model) and slight fluctuation detection (second model). This segmentation enables the system to detect slight changes with high precision while keeping each individual model relatively simple in structure.
Solution Approach 2:
The patent adds a new predictive dimension by introducing the second model that specifically targets slight fluctuations. This dimensional addition to the prediction framework enables detection of subtle changes that traditional regression models would overlook.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
According to one embodiment, a processing circuit classifies a time-series data corresponding to process amounts generated in a target facility into groups. For each of groups, the processing circuit applies time-series data included in the group to a first auto-encoder, which differs depending upon each group, and outputs time-series data. The processing circuit applies input difference data, which are based on output time-series data on the process amounts and the input time-series data, to a single second auto-encoder, and outputs difference data. The processing circuit determines an abnormality of the target facility, based on the comparison between addition data which are based on the output difference data and the output time-series data, and the input time-series data.