基于改进深度学习策略的水电机组多维度故障诊断方法研究

A multi-dimensional fault diagnosis method based on improved deep learning methods is put forward in this project in view of the problems existed in the fault diagnosis of hydropower units such as over-relying on the labeled samples and manually extracting the features uncertainly. Due to the excellent performance of deep learning method in dealing with the massive data without labels and extracting the features adaptively, it is used to explore new research ideas and technical approaches for fault diagnosis of hydropower units. The main contents are as follows. The data cleaning, de-noising, and compressing methods are proposed to improve the data quality based on compressed sensing technique. A new structure of the loss function regularization term is proposed through analyzing the learning rule of the model. A global optimization strategy for model hyper-parameters is put forward based on the symbiotic organisms search algorithm. Thus the computational speed and generalization performance of the model can be increased effectively. According to the multi-source heterogeneous characteristics of condition monitoring data, a multi-dimensional fault diagnosis model is established to locate the faults layer-by-layer based on the fusion of deep learning and uncertain reasoning methods. The mapping relationship between fault features and mechanisms are illustrated through the comparative analysis of feature characteristics under multiple operating conditions and fault modes. The research results have important significance for developing the deep learning theory, improving the accuracy of fault diagnosis and revealing the failure mechanism.

针对现有水电机组故障诊断方法过于依赖有标签样本，人工提取故障特征存在不确定性的问题，本项目提出基于改进深度学习策略的水电机组多维度故障诊断方法，旨在利用深度学习处理海量无标签数据和自适应特征提取的优异性能，探索水电机组故障诊断新的研究思路和技术途径。主要内容包括：基于压缩感知技术提出状态监测数据的清洗、降噪和压缩方法，有效提升数据质量；研究模型内部学习机制，设计损失函数正则化项的新型结构，结合共生生物搜索算法，提出超参数全局寻优策略，从而增强深度学习模型计算速度与泛化性能；考虑状态监测数据多源异构特点，融合深度学习与不确定推理方法，建立多维度故障诊断模型，实现水电机组的逐层故障定位，对比分析多工况、多故障模式下的特征组成，挖掘故障特征与故障机理间的映射关系。研究成果对完善深度学习方法，提高故障诊断准确性和揭示故障机理有重要意义。

水电机组故障机理复杂耦合，传统诊断方法过于依赖有标签样本和人工提取故障特征，使得水电机组故障诊断难度较大。为提升水电机组故障诊断智能化程度，释放水电大数据发展红利，本项目聚焦深度学习模型在水电故障诊断中的应用，以水电机组状态监测信号的自适应特征提取为突破点，提出针对多维信号的多模型水电机组故障诊断方法。首先针对水电机组状态监测系统海量数据分布不均问题，基于自编码器和重构误差的在线辨识方法建立了状态监测数据异常识别模型；研究基于堆叠自编码网络的水电机组故障诊断方法，提出了基于蛙跳粒子群算法和动态调节学习率的深度学习参数自适应寻优算法，有效避免了算法过拟合、局部极值解等问题；针对水电机组状态监测数据的多源异构特点，提出了改进卷积神经网络和改进深度残差网络结构，并以轴心轨迹图像和轴系多源振动信号为研究对象，建立水电机组轴系故障诊断模型；基于长短期记忆神经网络和自回归滑动平均模型融合方法，构建了水电机组状态监测参数多时间尺度预测与趋势报警模型。研究成果将系统完善水电机组故障诊断理论，并为提高我国水电机组运维智能化水平提供重要理论依据和技术参考。