Autoencoder Fault Prediction for Long-Time-Constant Deterioration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing fault prediction methods using LSTM technology struggle with devices having long time constants, leading to large amounts of past information and difficulty in realizing fault prediction with a small number of parts, while methods relying solely on current information lack accuracy in estimating deterioration states.
Innovation Solution
A fault prediction device utilizing a plurality of autoencoders corresponding to different deterioration states, which determine the device's state based on state signals, and employs a combination of inference and state calculation to predict accurate deterioration states without requiring extensive past information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If LSTM technology is used to predict fault by storing past information, then prediction accuracy is improved, but the amount of information to be stored becomes very large when the time constant is long
Solution Approach 1:
The patent segments the prediction task by dividing the input information into two distinct parts: current state information and past information. This segmentation allows the system to process only relevant past information rather than storing all historical data, thereby reducing the quantity of information needed while maintaining prediction accuracy for devices with long time constants.
Solution Approach 2:
The patent extracts and utilizes only the necessary past information required for accurate prediction, rather than storing comprehensive historical data. By selectively extracting relevant temporal patterns and combining them with current state information, the system achieves accurate fault prediction without the burden of large-scale information storage.
2Quantity of substance
If only current information is used to estimate deterioration state, then the amount of information is reduced, but the accuracy of estimating deterioration state becomes insufficient
Solution Approach 1:
The patent segments the information input into current state information and selectively extracted past information. This segmentation enables the system to maintain low information requirements while improving estimation accuracy by incorporating only the most relevant temporal patterns, avoiding the need to process all historical data.
Solution Approach 2:
The patent applies local quality by giving different weights and processing methods to different types of information. Current state information is processed with high priority, while past information is selectively extracted and integrated only where it provides predictive value. This differentiated approach maximizes estimation accuracy while minimizing the total amount of information required.
3Reliability
If LSTM technology is used for fault prediction, then prediction capability is improved, but device complexity increases making it difficult to realize with a small number of parts
Solution Approach 1:
The patent segments the fault prediction function into modular components: a current state information processing unit and a past information processing unit. This segmentation enables the system to achieve LSTM-level prediction capability through simpler, more modular architecture that can be implemented with fewer integrated components, reducing overall device complexity.
Solution Approach 2:
The patent extracts the essential predictive functionality from the complex LSTM architecture, retaining only the core temporal pattern recognition capability while eliminating the need for large-scale information storage infrastructure. This extraction enables fault prediction capability to be realized with a minimal number of parts.
Data Source
AI summary
A fault prediction device capable of predicting an accurate deterioration state is provided. A fault prediction device for predicting fault of a target device whose deterioration state transitions with elapse of time includes autoencoders AED1 to AED4 respectively corresponding to deterioration states of the target device. The autoencoder AED2 corresponding to a first deterioration state determines whether the target device exists in the first deterioration state or not based on a state signal indicating a state of the target device. In a case where it is determined that the target device does not exist in the first deterioration state, the autoencoder AED3 corresponding to a second deterioration state determines whether the target device exists in the second deterioration state or not based on the state signal.