结合Diffusion-Based WGANGP的变压器油纸绝缘老化状态拉曼光谱检测方法

doi:10.3964/j.issn.1000-0593(2025)08-2164-10

摘要
参考文献
相关文章 (15)

全文: PDF (10917 KB)
输出: BibTeX | EndNote (RIS)

摘要：提出了一种结合拉曼光谱与扩散模型改进的Wasserstein生成对抗网络（WGANGP）方法，用于提高变压器油纸绝缘老化状态的检测精度。凭借拉曼光谱技术无接触、不损耗的优势，可通过其分析油浸式电力变压器内部油纸绝缘材料的老化产物来评估变压器的老化程度。结合深度学习分类模型简化了拉曼光谱数据预处理过程，但此类模型对训练数据的数量和质量有较高要求，由于热加速老化实验周期长，导致可用于训练的有效拉曼光谱数据集相对稀少，限制了分类模型性能。为了解决这一难题，本研究引入了一种新的数据增强方法，即基于扩散模型的WGANGP（Diffusion-Based WGANGP），该方法通过将去噪扩散概率模型的前向加噪过程与WGANGP相结合，向WGANGP中引入实例化的噪声，去除了传统WGANGP的生成器结构中的复杂向上采样过程，简化了数据增强模型结构，有利于模型参数优化。相比于传统GAN及其变体，这种方法不仅保持了变压器油纸绝缘老化样本拉曼光谱数据集的原始特征峰特征与老化程度相关的基线漂移趋势，且与原始数据集特征保持近似的空间分布，生成的数据集信噪比（SNR）为24.84 dB，相比于原始数据集提高了32.11%；同时，也提升了生成样本的多样性，提高了基于深度学习的老化诊断模型的泛化能力、定量分析能力和鲁棒性。实验结果表明，采用Diffusion-Based WGANGP数据增强模型所生成的拉曼光谱数据集，在多个分类模型上的表现均优于其他数据增强方法，特别是在与ResNet-SVM分类模型结合时，在Accuracy（准确性，0.997 4）、F1 score（F1分数，0.996 9）、Recall（召回率，0.996 0）和Precision（精确度，0.998 0）四个评价指标上均表现出优势，这表明改进后的数据增强模型能够有效解决变压器老化绝缘油样本稀缺的问题，同时提高了分类模型对变压器老化状态的定量诊断能力。

关键词：变压器；油纸绝缘；拉曼光谱；Diffusion-Based WGANGP；故障诊断

Abstract：In this paper, a Wasserstein Generative Adversarial Network (WGANGP) method combining Raman spectroscopy with diffusion model improvement is proposed for improving the detection accuracy of the aging state of transformer oil-paper insulation. Raman spectroscopy, with its advantages of no contact and no loss, is used to assess the aging degree of a transformer by analysing the aging products of the oil-paper insulation material inside an oil-immersed power transformer. Combining deep learning classification models simplifies the Raman spectroscopy data preprocessing process, but such models have high requirements on the quantity and quality of training data. The long cycle of thermally accelerated aging experiments results in a relatively scarce set of valid Raman spectral data available for training, limiting the performance of the classification model. To address this challenge, a new data augmentation method, Diffusion-Based WGANGP, is introduced in this study. By combining the forward noise addition process of the denoising diffusion probabilistic model with WGANGP, the method introduces instantiated noise into WGANGP, removes the complex up-sampling process in the generator structure of the traditional WGANGP, simplifies the data augmentation model structure, and facilitates the optimization of model parameters. Compared with the traditional GAN and its variants, this method not only maintains the baseline drift trend of the original eigenpeak features associated with the aging degree in the Raman spectral dataset of the aging samples of the transformer oil-paper insulation, but also maintains an approximate spatial distribution of the features of the original dataset,, and the generated dataset with a Signal-to-Noise Ratio of 24.84 dB, which is improved by 32.11% compared to the original dataset; and at the same time, it also improves the diversity of generated samples, and enhances the generalisation ability, quantitative analysis ability and robustness of the aging diagnosis model based on deep learning. The experimental results show that the Raman spectral dataset generated using the Diffusion-Based WGANGP data augmentation model outperforms other data augmentation methods on several classification models, especially when combined with the ResNet-SVM classification model, in terms of Accuracy (0.997 4), F1 score (0.996 9), Recall (0.996 0) and Precision (0.998 0), which indicates that the improved data augmentation model can effectively solve the problem of the scarcity of samples of transformer aged insulating oil, and at the same time improves the classification model's ability to diagnose the aging state of the transformer quantitatively.

Key words：Transformer; Oil-paper insulation; Raman spectroscopy; Diffusion-Based WGANGP; Fault diagnosis

收稿日期: 2024-10-18 修订日期: 2025-03-18

中图分类号:

O433.4

基金资助: 国家自然科学基金项目(62275035)，重庆市教育委员会科学技术研究项目(KJZD-K202101103)，重庆理工大学研究生教育高质量发展项目(gzlcx20233150), 重庆市研究生科研创新项目（CYS25745）资助

通讯作者: 张知先 E-mail: zhangzhixian@cqut.edu.cn

作者简介: 陈新岗，1968年生，重庆理工大学电气与电子工程学院教授 e-mail: chenxingang@cqut.edu.cn

引用本文:

陈新岗，敖怡，张知先，马志鹏，张文轩，万福，况露，罗博文. 结合Diffusion-Based WGANGP的变压器油纸绝缘老化状态拉曼光谱检测方法[J]. 光谱学与光谱分析, 2025, 45(08): 2164-2173.
CHEN Xin-gang, AO Yi, ZHANG Zhi-xian, MA Zhi-peng, ZHANG Wen-xuan, WAN Fu, KUANG Lu, LUO Bo-wen. Raman Spectroscopic Detection of the Aging State of Oil-Paper Insulation in Combined Diffusion-Based WGANGP Transformers. SPECTROSCOPY AND SPECTRAL ANALYSIS, 2025, 45(08): 2164-2173.

链接本文:

https://www.gpxygpfx.com/CN/10.3964/j.issn.1000-0593(2025)08-2164-10 或 https://www.gpxygpfx.com/CN/Y2025/V45/I08/2164

[1] Zhang E Z, Liu J, Zhang C H, et al. Energies, 2023, 16(3): 1396.
[2] Zhang Z X, Chen W G, Wu K J, et al. IEEE Transactions on Instrumentation and Measurement, 2024, 73: 9517711.
[3] Soni R, Mehta B. Electric Power Systems Research, 2023, 214(B): 108954.
[4] Gouda O E, El Dein A Z. Electric Power Components and Systems, 2019, 47(4-5): 406.
[5] Feng D W, Chen G, Yan X Y, et al. IEEE Transactions on Dielectrics and Electrical Insulation, 2023, 30(6): 2758.
[6] Zheng H B, Yang E C, Wu S Y, et al. IEEE Transactions on Dielectrics and Electrical Insulation, 2022, 29(4): 1226.
[7] Kostina Y V, Rusakova O Y, Mikhalitsyn L A, et al. Petroleum Chemistry，2023, 63(1): 1.
[8] CHEN Xin-gang, CHEN Shu-ting, YANG Ding-kun, et al(陈新岗, 陈姝婷, 杨定坤, 等). Spectroscopy and Spectral Analysis(光谱学与光谱分析), 2021, 41(5): 1463.
[9] CHEN Xin-gang, FAN Yi-jie, MA Zhi-peng, et al(陈新岗, 范益杰, 马志鹏, 等). Laser & Optoelectronics Progress(激光与光电子学进展), 2023, 60(21): 2130002.
[10] Yang D K, Chen W G, Shi H Y, et al. High Voltage, 2021, 6(1): 51.
[11] Chen W G, Zou J X, Wan F, et al. AIP Advances, 2018, 8(3): 035204.
[12] Wu Y J, He Y Z, Wang Y L. Sensors, 2023, 23(16): 7153.
[13] Tang C H, Thu Q X, Wu W J, et al. Mathematical Problems in Engineering, 2020, 2020: 1245924.
[14] ZOU Jing-xin, CHEN Wei-gen, WAN Fu, et al(邹经鑫，陈伟根，万福，等). Transactions of China Electrotechnical Society(电工技术学报), 2018, 33(5): 1133.
[15] Nayak A A, Venugopala P S, Ashwini B, Archives of Computational Methods in Engineering, 2024, 31(8): 4739.
[16] Arjovsky M, Bottou L. 5 th International Conference on Learning Representations, 2017.
[17] Gulrajani I, Ahmed F, Arjovsky M, et al. Improved Training of Wassersetin GANs. Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, 2017, USA(2017): 5767. doi: 10.48550/arXiv.1704.00028
[18] Shang Z W, Zhang J, Li W X, et al. Neural Processing Letter, 2023, 55(5): 6377.
[19] Yang R, Shu R, Nakayama H, IEICE Transactions on Information and Systems, 2022, E105 D(9): 1537.
[20] Ho J, Jain A, Abbeel P. Denoising Diffusion Probabilistic Models[C]. Advances in Neural Information Processing Systems 33 (NIPS 2020). Curran Associates, 2020, 6840. doi: 10.48550/arXiv.2006.11239.
[21] Song J M, Meng C L, Ermon S. Denoising Diffusion Implicit Models[C]. International Conference on Learning Representations (ICLR 2021), ICLR 2021, doi: 10.48550/arXiv.2010.02502.
[22] Kong Z, Ping W. On Fast Sampling of Diffusion Probabilistic Models[C]. International Conference on Learning Representations ( ICLR 2021), ICLR 2021, doi: 10.48550/arXiv.2106.00132.
[23] Jalbert J, Gilbert R, Denos Y, et al. IEEE Transactions on Power Delivery, 2012, 27(2): 514.
[24] DL/T 984-2018, Guide for the Diagnosis of Insulation Aging in Oil-Immersed Power Transformer(油浸式变压器绝缘老化判断导则)[S], 2018.