Abstract:

In transformer fault diagnosis accuracy, addressing the limitations of traditional neural networks such as insufficient interpretability and weak temporal feature extraction capabilities, this study proposes a novel diagnostic model,LKAN which integrates Long Short?Term Memory (LSTM) with Kolmogorov?Arnold Network (KAN). The model first employs LSTM to model time?series data from transformer operations, extracting hidden states as temporal features. These features are then fed into the KAN layer, where B?spline functions enable nonlinear mapping and function decomposition, thereby enhancing both the model's expressiveness and interpretability. Experimental results on real?world power transformer datasets demonstrate that the LKAN model achieves a diagnostic accuracy of 98.80%, outperforming LSTM, Convolutional Neural Network（CNN）, Gated Recurrent Unit（GRU）, and single KAN models.Meanwhile, it exhibits strong generalization ability and stability. The LKAN model effectively integrates the temporal modeling capability of LSTM and the interpretability advantage of KAN. It provides a technical path with high accuracy and strong interpretability for intelligent fault diagnosis of transformers, and has good engineering promotion value.

Key words: Transformer fault diagnosis, LKAN model, LSTM, KAN, Interpretable neural network

摘要：

针对传统神经网络在变压器故障诊断中存在可解释性不足、时序特征提取能力弱等问题，提出了一种融合长短期记忆网络（LSTM）与柯尔莫戈洛夫-阿诺德网络（Kolmogorov?Arnold Network，KAN）的新型诊断模型——LKAN。该模型首先利用LSTM对变压器运行时序数据进行建模，并从隐藏状态中提取关键时序特征；随后将特征输入KAN层，通过B?spline基函数实现非线性映射与函数分解，提升模型的表达能力与可解释性。在真实电力变压器数据集上的实验结果表明，LKAN模型的故障诊断准确率达到98.80%，优于LSTM、卷积神经网络（CNN）、门控循环单元（GRU）及单一KAN模型，同时展现出较强的泛化能力与稳定性。LKAN模型有效融合了LSTM的时序建模能力与KAN的可解释性优势，为变压器智能故障诊断提供了一种高精度、可解释性强的技术路径，具有良好的工程推广价值。

关键词: 变压器故障诊断, LKAN模型, LSTM, KAN, 可解释性神经网络