基于深度迁移学习网络的水电机组故障诊断方法

Fault diagnosis method for hydropower units based on deep transfer learning networks

摘要

【目的】针对水电机组实际运行中故障数据少、不同机组间故障信号不符合同分布假设而导致模型训练困难、现有诊断模型泛化能力差的问题，提出了一种结合迁移学习策略与CNN-BIGRU-attention网络(TCBA)的故障诊断方法。【方法】以转子试验台数据为源域，真实水电机组振动数据为目标域数据。首先，组合卷积神经网络(CNN)、双向门控单元(BIGRU)与注意力单元，构建CNN-BIGRU-attention诊断模型，然后利用源域数据对模型进行训练，训练完成后将模型的参数迁移至目标领域的故障诊断模型中，迁移过程中将低层网络冻结，并通过部分目标域数据对高层网络进行微调，最终得到适应目标设备的故障诊断模型。为验证所述方法的有效性，通过转子试验台数据集与真实水电机组故障数据的迁移试验对比所述模型与传统深度学习方法的识别准确率、训练速度以及所需样本量等指标。【结果】结果表明：与传统训练方法相比，所提方法能够显著提高模型的收敛速度并有效降低训练所需样本量。在小样本下对实际水电站故障样本数据故障状态的识别率可达99.02%,相比传统方法提升约3.00%。【结论】研究结果充分证明了该方法具有较好的故障状态识别能力，为解决数据有限情况下的水电机组故障识别问题提供了一种有效的解决途径。

Abstract

[Objective] Due to the limited fault data during actual operation of hydropower units and differences in fault signal distributions across different units that do not conform to the assumption of identical distribution, model training becomes difficult, and existing diagnostic models have poor generalization capabilities. To address these issues, a fault diagnosis method combining transfer learning strategies with a CNN-BIGRU-attention network(TCBA) is proposed. [Methods] Rotor test bench data was used as the source domain, and real vibration data from hydropower units served as the target domain data. First, a CNN-BIGRU-attention diagnostic model was constructed by combining Convolutional Neural Network(CNN), Bidirectional Gated Recurrent Unit(BIGRU), and attention units. The model was initially trained using source domain data, and its parameters were then transferred to the fault diagnosis model of the target domain. During the transfer process, the lower-layer network was frozen, and the upper-layer network was fine-tuned using partial target domain data, resulting in a fault diagnosis model adapted for the target equipment. To verify the effectiveness of the proposed method, a comparison was conducted between the proposed method and traditional deep learning method through transfer experiments using rotor test bench datasets and real hydropower unit fault data, evaluating indicators such as recognition accuracy, training speed, and sample size requirements. [Results] The result showed that, compared with traditional training method, the proposed method significantly improved the model's convergence speed and effectively reduced the sample size required for training. Under small sample conditions, the fault state recognition accuracy for actual hydropower station fault sample data reached 99.02%, which was about 3% higher than that of the traditional method. [Conclusion] This study demonstrates that the proposed method has strong fault state recognition capabilities, providing an effective solution for fault diagnosis of hydropower units under limited data conditions.

关键词

水电站 / 水电机组 / 振动信号 / 故障诊断 / 迁移学习 / 数据驱动

Key words

hydropower station / hydropower units / vibration signals / fault diagnosis / transfer learning / data-driven

基金资助

中国长江电力股份有限公司科研项目资助(2323020046)

中国长江电力股份有限公司科研项目资助(Z232302044)

国家自然科学基金重点项目(52339006)

江苏省创新支撑计划国际科技合作项目(BZ2023047)