Transformer winding deformation diagnosis method based on dynamic time warping and multilayer perceptron

Guobin Wang, Kang Wang, Xiaolin Xu, Guangyu Shi, Yu Lu, Shengwen ShuFrontiers in Energy Research 2025 · DOI

Abstract

To address the issues of strong subjectivity and difficulty in feature extraction that are inherent to traditional frequency response analysis methods used for diagnosing transformer winding deformation, an intelligent diagnostic method is proposed based on Dynamic Time Warping (DTW) and a Multilayer Perceptron (MLP). First, the frequency response curve is normalized and segmented into multiple frequency bands to extract physically meaningful features. Subsequently, the Dynamic Time Warping algorithm is employed to perform nonlinear curve alignment and difference quantification processes, thereby enhancing robustness against frequency-axis misalignment and measurement noise. Finally, the extracted features are fed into a Multilayer Perceptron (MLP) model, which utilizes multilayer nonlinear mappings to automatically identify the deformation levels of the windings. Validation based on field measurement data indicates that the proposed method achieves significant improvements in diagnostic accuracy, balance, and robustness when compared with traditional correlation coefficient methods and other machine learning models. This approach enables high-precision automated diagnosis of transformer winding deformation, offering a physically interpretable reference for condition monitoring as well as intelligent operation and maintenance of power equipment.

Introduction

Power transformers are the heart of the electrical grid, but their internal windings are subject to massive mechanical and thermal stresses during short circuits. Over time, these windings can bulge, loosen, or displace—often leading to catastrophic failure.

The industry standard for detecting these shifts is Frequency Response Analysis (FRA). Think of it as an EKG for a transformer: you inject a signal and measure the output. If the internal geometry changes, the "signature" of the frequency response shifts. However, traditional FRA is notoriously subjective. Engineers often have to manually interpret subtle "peak shifts" or "curve trends," a process prone to human error and sensitive to measurement noise or instrument misalignment. This paper aims to replace that "expert intuition" with a high-precision, automated pipeline.

Methods

The authors propose a three-stage intelligent diagnostic framework designed to handle the complexities of real-world field data.

Overall framework flowchart of the proposed method.

Frequency-Specific Segmentation

Instead of looking at the whole spectrum at once, the FRA curve (1 kHz to 1,000 kHz) is divided into three functional bands:

  • Low-frequency (LF): Reflects inductive characteristics (overall winding displacement).
  • Medium-frequency (MF): Influenced by both inductance and capacitance (winding bulging).
  • High-frequency (HF): Dominated by capacitive elements (local short circuits).

DTW-Based Feature Extraction

The "secret sauce" of this method is Dynamic Time Warping (DTW). Unlike standard Euclidean distance, which compares points at the exact same frequency, DTW allows for nonlinear alignment. It "stretches" the curves to find the optimal match, making the system highly robust against frequency-axis misalignment and measurement noise. The "distance" calculated by DTW becomes a physically meaningful feature representing morphological change.

DTW sequence alignment and distance matrix illustrating the elastic matching process.

MLP Classification

These DTW-derived features are fed into a Multilayer Perceptron (MLP). The MLP is chosen because it is lightweight enough for industrial edge deployment while being powerful enough to map complex, nonlinear relationships between curve shifts and specific deformation levels.

Datasets and Validation

The study uses real-world field test data from the Fujian Provincial Power Grid Company in China, covering transformers at 110 kV, 220 kV, and 500 kV levels. The dataset includes normal states and three levels of deformation (mild, significant, and severe). Because severe faults are rare in the field, the authors used Generative Adversarial Networks (GANs) to augment the minority-class samples in the training set, ensuring the model didn't develop a bias toward "normal" readings.

Key Results

The proposed DTW-MLP method was validated against an independent test set of field data:

  • Overall Accuracy: 96.00%, significantly higher than the traditional correlation coefficient method (80.00%).
  • Safety Reliability: The model achieved 100% recall for severe deformations. In a power grid context, this is critical—it means the system never missed a high-risk fault.
  • Early Detection: It achieved 100% precision for minor deformations, meaning it didn't produce "false alarms" for healthy transformers, though it did miss some very subtle early-stage shifts (80% recall for mild cases).

Conclusion

The core advantage of this approach is the use of elastic matching (DTW) to handle "feature drift." In real-world testing, two identical transformers might show slightly shifted peaks due to different lead lengths or grounding. DTW ignores these trivial offsets to focus on the actual shape deformation. By segmenting the frequency bands, the model also provides physical interpretability, helping engineers identify where and what the physical change likely is. Future work will focus on expanding the dataset and exploring advanced learning paradigms like few-shot learning to adapt to new transformer designs.

Summary updated

Abstract

To address the issues of strong subjectivity and difficulty in feature extraction that are inherent to traditional frequency response analysis methods used for diagnosing transformer winding deformation, an intelligent diagnostic method is proposed based on Dynamic Time Warping (DTW) and a Multilayer Perceptron (MLP). First, the frequency response curve is normalized and segmented into multiple frequency bands to extract physically meaningful features. Subsequently, the Dynamic Time Warping algorithm is employed to perform nonlinear curve alignment and difference quantification processes, thereby enhancing robustness against frequency-axis misalignment and measurement noise. Finally, the extracted features are fed into a Multilayer Perceptron (MLP) model, which utilizes multilayer nonlinear mappings to automatically identify the deformation levels of the windings. Validation based on field measurement data indicates that the proposed method achieves significant improvements in diagnostic accuracy, balance, and robustness when compared with traditional correlation coefficient methods and other machine learning models. This approach enables high-precision automated diagnosis of transformer winding deformation, offering a physically interpretable reference for condition monitoring as well as intelligent operation and maintenance of power equipment.
BibTeX
@inproceedings{wang2025transformer,
  author = {Guobin Wang and Kang Wang and Xiaolin Xu and Guangyu Shi and Yu Lu and Shengwen Shu},
  title = {{Transformer winding deformation diagnosis method based on dynamic time warping and multilayer perceptron}},
  booktitle = {Frontiers in Energy Research},
  year = {2025},
  doi = {10.3389/fenrg.2025.1731410},
}
Figure