Converter Fault Prediction Using ML Temperature Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for predictive maintenance of power semiconductor devices struggle to accurately detect temperature trends, especially in dynamic operating conditions, making it difficult to identify gradual changes that may indicate faulty behavior or aging, leading to potential unexpected failures.
Innovation Solution
A method using a machine learning algorithm trained with historical operation point and temperature data to estimate device temperatures, allowing for the prediction of faulty behavior by comparing estimated and measured temperatures, and incorporating ambient temperature for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If simple temperature trending is used to monitor static conditions, then sudden rises can be detected as component faults or aging, but gradual temperature changes under dynamic operating conditions cannot be reliably detected
Solution Approach 1:
The patent applies dynamics by making the temperature monitoring system adaptive to changing operating conditions. The machine learning model is trained on historical data covering various operating points, enabling it to dynamically adjust temperature predictions based on current load, ambient temperature, and operational context. This resolves the contradiction by making the system versatile enough to handle dynamic conditions while maintaining precise detection of abnormal temperature trends.
Solution Approach 2:
The patent changes parameters by using multiple input variables (load, ambient temperature, historical temperature data) instead of relying solely on raw temperature measurements. The machine learning model processes these varying parameters to predict expected temperature under current operating conditions, allowing accurate detection of anomalies regardless of whether the system is in static or dynamic operation. This enables reliable gradual temperature change detection across all operating modes.
2Measurement precision
If machine learning algorithms are used to estimate device temperature, then accurate temperature prediction under dynamic conditions is achieved, but system complexity increases
Solution Approach 1:
The patent applies partial action by using a machine learning model with a practical complexity level - not overly simple to be inaccurate, but not excessively complex to be computationally burdensome. The model uses a selected set of relevant input parameters (load, ambient temperature, historical temperature) rather than all possible variables, achieving sufficient accuracy while maintaining reasonable computational requirements for implementation in power converter systems.
3Loss of information
If only measured temperature data is used for monitoring, then direct temperature information is available, but meaningful temperature trends and gradual changes are difficult to identify
Solution Approach 1:
The patent applies preliminary action by pre-training the machine learning model offline using historical temperature data and operating conditions. This preliminary training phase extracts and stores the relationships between operating parameters and temperature behavior. During actual operation, the pre-trained model quickly processes current inputs to predict expected temperature, enabling timely detection of deviations without requiring complex real-time analysis of raw temperature sequences. This resolves the contradiction by preserving temperature trend information while keeping online processing requirements manageable.
Data Source
Figure 1
Figure 2~3
Figure 4A~4B
AI summary
A method for predicting a faulty behaviour of an electrical converter (12) comprises: receiving an operation point indicator (I) of the electrical converter (12) indicative of an actual operation point of the electrical converter (12); receiving a measured device temperature (Td) of a power semiconductor device (18) of the electrical converter (12) indicative of an actual temperature of the power semiconductor device (18); inputting the operation point indicator (I) as input data into a machine learning algorithm (32) trained with historical data comprising operation point indicators and associated device temperatures; estimating an estimated device temperature (II) with the machine learning algorithm; and predicting the faulty behaviour (F) by comparing the estimated device temperature (II) with the measured device temperature (Td).