Sensor Time-Series Correction Using GAN Fault Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting and correcting faulty sensor data in electrical motor systems, particularly in harsh operating conditions, struggle to differentiate between erroneous and normal data, leading to challenges in sensor fault detection and recovery, especially with abrupt and incipient failures.
Innovation Solution
A method involving a generative adversarial network (GAN) with an autoencoder structure, including convolutional and deconvolutional layers, and a classifier to predict corrected time series signals and identify error types, enabling the detection and recovery of erroneous data alongside missing values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional sensor fault detection methods are used, then missing or unavailable data can be handled using redundant sensors or estimators, but erroneous sensor data cannot be effectively differentiated from normal data
Solution Approach 1:
A GAN-based intermediary system is introduced between the sensor and the control system. The generator creates corrected sensor readings by learning the underlying data distribution, while the discriminator validates whether readings are erroneous or normal. This intermediary structure enables effective differentiation of erroneous data without requiring changes to the sensor hardware itself.
Solution Approach 2:
The patent replaces traditional mechanical/redundant sensor-based fault detection with a data-driven GAN model. Instead of using physical redundant sensors or classical estimation algorithms, the system uses deep learning models (generator and discriminator networks) to detect and correct erroneous sensor data, achieving higher precision in error differentiation.
2Measurement precision
If a GAN model is implemented to differentiate erroneous data, then measurement precision improves, but device complexity increases
Solution Approach 1:
The GAN model is segmented into two distinct functional components: the generator (which creates corrected readings) and the discriminator (which detects errors). This segmentation allows each component to be optimized independently and simplifies the overall training process, reducing the practical complexity despite the advanced functionality.
Solution Approach 2:
The GAN model serves multiple functions simultaneously: it detects erroneous sensor data, generates corrected readings, and provides uncertainty estimates. This multi-functionality reduces the need for separate systems for each task, thereby managing complexity while achieving comprehensive fault detection and correction capabilities.
3Reliability
If redundant sensors are used to handle missing data, then reliability improves, but device complexity and cost increase
Solution Approach 1:
Instead of using physical redundant sensors, the system creates virtual copies of sensor readings through the GAN generator. The generator learns the underlying data distribution from training data and generates synthetic sensor readings that replace missing or erroneous values, providing redundancy without additional physical hardware.
Solution Approach 2:
The patent replaces the mechanical approach of using redundant physical sensors with a software-based GAN model that generates synthetic sensor data. This substitution eliminates the need for additional hardware while maintaining data availability and system reliability.
Data Source
Figure 1~2
Figure 3~4
Figure 5~6
AI summary
The present document proposes a method for training a model able to predict a corrected time series signal and a classification of an error of said signal, from an input time series signal from at least one sensor, that may comprise an error or missing values, said model comprising a generative adversarial network or GAN comprising a generator and a discriminator each based on an autoencoder structure, and a classifier able to predict a class of the predicted error outputting the second head of the generator, said method comprising a training phase where the generator, the discriminator and the classifier are trained using a training set of data comprising input time series signals from at least one sensor, said signals comprising or not an error and/or missing values.