半监督环境下风电机组群的智能化故障诊断与寿命预测

There are some disadvantages impeding the development of wind energy in China, e.g., high failure rates of wind turbines and disordered maintenance strategy. Usually, tens of even hundreds of wind turbines in service operate together under complicated and various operating conditions. Moreover, complex drive train and diverse failure modes make it difficult to obtain all samples under complete fault types, thus bringing a challenge to the intelligent fault diagnostic and life prognostic of wind turbines. In this project, a generative adversarial network based semi-supervised learning approach is studied, which can be applied to fault classification, diagnosis and location of concrete defects rapidly among numerous wind turbines. Through fusing the results from conditional generative adversarial network and joint probability of operating conditions, a health degradation index is designed to represent the degradation tendency of critical components of wind turbine, which can exclude the disturbance of various loads. For batch faults, a segmented degradation model is deduced from stochastic statistic, and the remaining useful life is estimated using multi-mode particle filter. For individual fault, a data driven model is constructed through analyzing the development law of incipient fault and a long-term degradation tendency is fitted, further the remaining useful life is computed accurately. This project is significant in theory and practice for reducing the cost of operation and maintenance, developing the high efficient maintenance system for wind farm and guaranteeing a healthy progress in wind power industry.

我国风电行业存在机组故障率高、维护决策无序的问题。风电机组以集群规模化运行、工况复杂多变，且其传动链结构复杂、失效形式多样导致完备类型的故障样本获取困难，进而给风电机组群的智能化故障诊断与寿命预测带来挑战。项目以生成对抗网络为基础，研究半监督诊断环境下风电机组群的智能故障分类方法，实现故障机组的快速筛查；融合条件生成对抗网络和多工况联合概率信息，设计能够排除工况干扰的风电机组健康退化指标；针对批次故障机组，构建分阶段的健康退化统计模型，进行基于多模态粒子滤波的剩余寿命估计；针对个体故障机组，依据自身退化规律构建数据驱动模型，获得长期性能退化趋势，实现风电机组精确的剩余寿命预测。项目的研究对于降低风电场的运营维护成本，形成合理高效的维修体制以促进风电行业健康发展具有重要的理论和现实意义。

我国风电行业存在机组故障率高、维护决策无序的问题。风电机组以集群规模化运行、工况复杂多变，且其传动链结构复杂、失效形式多样导致完备类型的故障样本获取困难，进而给风电机组群的智能化故障诊断与寿命预测带来挑战。项目针对大规模风电机组群的故障类别多样、全故障样本难以全面收集的问题，研究了基于深度神经网络、深度变分自编码、原型学习的风电机组半监督智能故障诊断方法，设计了SCADA的预处理方法，利用健康数据进行深度学习模型训练，兼顾模型正则化和泛化能力，实现多类故障的准确识别；针对风电机组故障诊断与预警不稳定的问题，研究了基于双判别器的生成对抗网络以解决故障样本不均衡的难点，开发了基于WGAN-GP算法的SCADA数据对抗训练，并利用距离测度量化风电机组传动链的异常状态，提出了稀疏字典学习与对抗变分自编码融合的故障预警算法，获得具有较高稳定性的风电传动系统故障预警结果；针对风电机组剩余寿命准确预测的难点，研究了基于单调性和平滑性的风电机组健康指标构建的约束条件，自适应选择时域、频域、时频域中影响约束条件的关键特征，并改进了粒子滤波和无迹卡尔曼粒子滤波方法，对风电机组轴承的批量故障和单一故障进行寿命预测，取得较好的预测精度。项目研究有助于提高风电机组群的智能化故障诊断与寿命预测能力，降低风电场的运维成本，对促进风电行业健康发展具有重要的理论和现实意义。