Intelligent and Integrated Condition Monitoring of Distributed Generation Systems

分布式发电将在未来的电力系统中发挥重要作用，因为这种发电技术可以在靠近最终用户的地点利用广泛的可再生能源来提供电力。在过去几十年中，在DG系统的能力、规模和位置方面取得了相当大的进步，例如从陆上到近海。DG系统部署面临的最关键挑战之一具体涉及可用性和可靠性，以便在无人看管的情况下维持能源发电并最大限度延长能源系统的使用寿命。因此，这对创新的状态监测系统和解决方案提出了更高的要求，以应对这一领域出现的新挑战。首次拨款计划申请中提出的研究代表了一项努力，旨在探索对优化故障检测和诊断的状态监测技术具有普遍重要性的关键问题。这项研究面向的是以风力涡轮机为DG来源的DG系统，因为这一特定的应用提出了许多现实的挑战。首先，由于风源的间歇性和电网系统的波动，测量信号会表现出很强的非平稳特性。其次，小幅度的信号可能指示重大故障的开始，这通常是常规方法无法检测到的，特别是在恶劣的环境中。第三，需要处理和传输大量数据，特别是对于连续的在线监测。例如，如果我们假设一个典型的2兆瓦风力发电机组需要250个点来监测一个风力发电机组的大多数子系统，这将在5分钟的采样率下产生一个1 GW风电场每天3600万个数据。此外，一个需要迫切关注的关键问题是传感器系统的健康问题，这就要求监测技术应该评估当一些传感器读取数据不正确时发生了什么。为了满足这种多样化的要求，我们计划使用和应用加窗变换，这是一种以提取测量数据中非平稳成分的能力而闻名的技术。通过对窗口形状的优化选择，可以实现自动加窗小波变换，以适应不同的传感器数据，从而更好地定位、提取和关联特征。尽管早期的故障信号通常幅度小、持续时间短，但它本质上具有与大故障信号相同的特征，如规律性。如果我们能设计一种合适的算法来匹配信号的局部规律性或奇异性，那么任何潜在的故障、异常和紊乱都可以被检测出来，而不管它们的大小和持续时间。该项目还致力于设计一种混合神经模糊方法来进行最优传感器数据融合。使用这种人工智能方法可以通过系统地结合先验信息来最好地关联传感器数据和预测未知数。最小化传感器的数量，同时保持足够的数量来评估系统的条件，不仅可以最大限度地降低传感器系统的复杂性，而且还可以减少数据存储需求。项目的最后部分特别涉及到实用方面，在模块化的嵌入式系统上对所提出的算法进行实时验证，以实现在线监控目的。本项目中拟议的状态监测系统将在一个硬件模块中容纳所有监测技术，该模块可以很容易地适应其他应用。该项目将为实际应用提供更好的传感技术和改进的算法，通过提高我们对如何设计它们的理解，以帮助与现有和未来DG系统的资产维护和管理有关的决策过程。

Distributed electricity generation (DG) will play a significant role in future electric power system, as this type of power generation technology can provide electric power by utilising a wide range of renewable energy sources at a site close to end users. Considerable advances have been achieved during past decades in the capacity, scale and location of DG systems, e.g. from onshore to offshore. One of the most critical challenges for the deployment of DG systems relates specifically to availability and reliability in order to sustain energy generation and maximise a long service life of the energy systems unattended. This has, therefore, placed higher demand on predictive maintenance from innovative condition monitoring systems and solutions to tackle new arising challenges in this area.The research proposed in this first grant scheme application represents an effort to explore key issues of generic importance to condition monitoring techniques optimised for fault detection and diagnosis. The research is oriented towards DG systems with wind turbines being the DG sources as this particular application presents a number of realistic challenges. Firstly, measurement signals would exhibit strong non-stationary behaviour due to the intermittent nature of wind sources and fluctuations of grid system. Secondly, the signals of small magnitude may indicate a start of a significant failure, which are normally undetected by conventional methods particularly in a harsh environment. Thirdly, large volume of data needs to be processed and transmitted especially for continuous online monitoring. For example, if we assume that 250 points are required for a typical 2 MW wind turbine to monitor most subsystems of a turbine, this will give rise to 36 million data per day for a 1 GW wind farm under a sampling rate of 5 minutes. Furthermore, a critical issue needing urgent attention will be the health problems of the sensor system, which requires that the monitoring techniques should be assessing what is happening when some of the sensors read data incorrectly.In order to meet such diversified requirements, we plan to use and apply windowed transform, a technique well known for its ability to extract nonstationary components in the measurement data. By the optimal selection of a window shape, automatic windowed wavelet transforms can be achieved to accommodate different sensor data for better feature localisation, extraction and correlation. Although an incipient fault signal is usually of low magnitude and short duration, it would essentially carry the same features as the large ones, such as the regularity. If we can design a suitable algorithm to match the local regularity or singularity of a signal, any incipient faults, abnormalities and disorders can be detected irrespective of their magnitude and time duration. The project is also concerned with designing a hybrid neuro-fuzzy method for optimal sensor data fusion. The use of this artificial intelligence method can best correlate sensor data and predict the unknowns by systematic incorporation of priori information. Minimising the number of sensors whilst still maintaining a sufficient number to assess the system's conditions can not only minimise the complexity of sensor systems but it can also reduce data storage requirements. The final part of the project relates specially to the practical aspect, where the proposed algorithms are validated in real time for online monitoring purposes on a modular embedded system. The proposed condition monitoring system in this project would accommodate all monitoring techniques within one hardware module, which can be readily adapted to other applications. The project will provide better sensing techniques and improved algorithms towards real applications by improving our understanding of how to engineer them in order to aid the decision making process with respect to asset maintenance and management of existing and future DG systems.