Transformer Failure Diagnosis via Multi-Stage Transfer Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The limited availability of abnormal data for deep learning methods, such as CNNs, hinders the effective diagnosis of power transformer failures, as these methods require large datasets to function accurately.
Innovation Solution
A multi-stage transfer learning approach is employed, utilizing a finite element model to simulate transformer parameters, constructing a convolutional neural network for two-stage transfer learning, and performing data enhancement to leverage simulation and detection data sets, thereby improving the accuracy of failure diagnosis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If deep learning methods (CNN) are used for transformer failure diagnosis, then automatic identification and rapid processing of abnormal states are achieved, but the lack of valid abnormal data considerably limits its application
Solution Approach 1:
The patent creates virtual copies of transformer data through finite element modeling and simulation, generating synthetic abnormal data that replicates real failure scenarios. This allows the deep learning model to be trained on simulated data that mirrors actual equipment behavior, overcoming the scarcity of real abnormal data while maintaining diagnostic accuracy
Solution Approach 2:
The patent introduces simulation data as an intermediary between the limited real abnormal data and the deep learning model requirements. The simulation acts as a bridge, translating available real data into expanded training datasets through virtual experimentation, enabling the model to learn from both real and synthesized examples
2Quantity of substance
If simulation data is used for deep learning training, then a large quantity of training data is obtained, but the diagnosis accuracy may be affected by the difference between simulation and real data
Solution Approach 1:
The patent performs preliminary validation by training the deep learning model on simulated data first, then fine-tuning it with real abnormal data before deployment. This staged approach allows the model to learn general patterns from simulation while adapting to real-world variations, ensuring both data quantity and diagnostic precision
Solution Approach 2:
The patent implements a feedback mechanism where the model's performance on real data is continuously evaluated and used to adjust the simulation parameters and data generation processes. This closed-loop approach ensures that simulated data remains aligned with actual equipment behavior, maintaining high diagnosis accuracy
Data Source
AI summary
A transformer failure identification and location diagnosis method based on a multi-stage transfer learning theory is provided. Simulation is set up first, a winding parameter of a transformer to be tested is calculated, and a winding equivalent circuit is accordingly built. Different failures are configured for the equivalent circuit, and simulation is performed to obtain a large number of sample data sets. A sweep frequency response test is performed on the transformer to be tested, and detection data sets are obtained. Initial network training is performed on simulation data sets by using the transfer learning method, and the detection data sets are further trained accordingly. A failure support matrix obtained through diagnosis is finally fused. The multi-stage transfer learning theory is provided by the disclosure.