Vehicle Power Electronics Prognostics Using Contextual Autoencoders
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Power electronic (PE) modules in vehicles experience degradation and early failure due to high power densities and temperature conditions, making it difficult to predict imminent failures without advanced data analysis.
Innovation Solution
The use of autoencoders trained with datasets including contextual information to generate clusters representing operating conditions, allowing for prognostics of PE devices by determining the minimum distance of encoded features from these clusters, thereby predicting current and future operating conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If basic sensed data (current, voltage, temperature) is used to detect anomalies, then anomaly detection capability is improved, but the ability to predict imminent failure of specific PE modules deteriorates
Solution Approach 1:
The patent transitions from analyzing basic sensed data in the original feature space to using autoencoders that map data into a compressed latent feature space. This dimensional transformation enables the system to capture complex patterns and relationships that are not apparent in the original data, thereby improving failure prediction capability while maintaining anomaly detection accuracy.
Solution Approach 2:
The autoencoder acts as an intermediary between the raw sensed data and the failure prediction logic. By introducing this intermediate processing layer with contextual information integration, the system can bridge the gap between basic anomaly detection and sophisticated imminent failure prediction, resolving the contradiction between the two capabilities.
2Reliability
If autoencoders are trained with contextual information to improve prognostics accuracy, then failure prediction capability is improved, but system complexity increases
Solution Approach 1:
The patent applies preliminary action by pre-training autoencoders with contextual information from healthy PE devices before deployment. This offline training phase captures operational patterns and relationships, allowing the system to achieve high prognostics accuracy during operation without requiring complex real-time processing. The contextual information is integrated in advance, reducing the computational burden during actual failure prediction.
3Measurement precision
If K-means clustering is applied to encoded features to generate condition clusters, then operating condition classification is improved, but processing time increases
Solution Approach 1:
The patent extracts the most relevant features through the autoencoder's latent space representation before applying K-means clustering. By taking out only the essential compressed features that capture the critical variations in operating conditions, the system achieves accurate condition classification while reducing the dimensionality and computational complexity of the clustering process, thereby minimizing processing time.
Data Source
AI summary
A method, system, and non-transitory computer readable medium describing an autoencoder that creates a reduced feature space from healthy power electronics devices for training. Devices under test are then encoded and compared to the encoded features of the healthy devices to determine health of the other devices. Contextual information is used to build multiple models that compare power electronics devices from similarly operated vehicles with one another.