Accurate and efficient condition monitoring is the key to enhance the reliability and security of wind turbines. In recent years, an intelligent anomaly detection method based on deep learning networks has been receiving increasing attention. Since accurately labeled data are usually difficult to obtain in real industries, this paper proposes a novel Deep Small-World Neural Network (DSWNN) on the basis of unsupervised learning to detect the early failures of wind turbines. During network construction, a regular auto-encoder network with multiple restricted Boltzmann machines is first constructed and pre-trained by using unlabeled data of wind turbines. After that, the trained network is transformed into a DSWNN model by randomly add-edges method, where the network parameters are fine-tuned by using minimal amounts of labeled data. In order to guard against the changes and disturbances of wind speed and reduce false alarms, an adaptive threshold based on extreme value theory is presented as the criterion of anomaly judgment. The DSWNN model is excellent in depth mining data characteristics and accurate measurement error. Last, two failure cases of wind turbine anomaly detection are given to demonstrate its validity and accuracy of the proposed methodology contrasted with the deep belief network and deep neural network.

With the development of industrial robots and other mechanical equipment to a higher degree of automation, mechanical systems have become increasingly complex. This represents a huge challenge for condition monitoring. The separation of vibration source signals plays an important role in condition monitoring and fault diagnosis. The key to the separation method of the vibration source signal is prior knowledge, such as of the statistical features of the vibration source signal, the number of vibration sources, and so forth. However, effective prior knowledge is difficult to obtain in engineering applications. This study found that low rank is a common feature of rotating machinery vibration source signals. To address the problem of the difficulty obtaining the signal feature of a vibration source, the multi-low-rank constrained vibration source signal separation method was proposed. Its advantages and effectiveness have been verified through simulations and experimental tests. Compared with the blind source separation method of independent component analysis (BSS-ICA) and the ensemble empirical mode decomposition (EEMD) methods, it obtained better clustering results and higher signal-to-signal ratio (SSR) values.

Pitch system failures occur primarily because wind turbines typically work in dynamic and variable environments. Conventional monitoring strategies show limitations of continuously identifying faults in most cases, especially when rapidly changing winds occur. A novel selective-ensemble monitoring strategy is presented to diagnose the most pitch failures using Supervisory Control and Data Acquisition (SCADA) data. The proposed strategy consists of five steps. During the first step, the SCADA data are partitioned according to the turbine’s four working states. Correlation Information Entropy (CIE) and 10 indicators are used to select correlation signals and extract features of the partition data, respectively. During the second step, multiple Small-World Neural Networks (SWNNs) are established as the ensemble members. Regarding the third step, all the features are randomly sampled to train the SWNN members. The fourth step involves using an improved global correlation method to select appropriate ensemble members while in the fifth step, the selected members are fused to obtain the final classification result based on the weighted integration approach. Compared with the conventional methods, the proposed ensemble strategy shows an effective accuracy rate of over 93.8% within a short delay time.

The performance evaluation of wind power forecasting under commercially operating circumstances is critical to a wide range of decision-making situations, yet difficult because of its stochastic nature. This paper firstly introduces a novel TRSWA-BP neural network, of which learning process is based on an efficiency tabu, real-coded, small-world optimization algorithm (TRSWA). In order to deal with the strong volatility and stochastic behavior of the wind power sequence, three forecasting models of the TRSWA-BP are presented, which are combined with EMD (empirical mode decomposition), PSR (phase space reconstruction), and EMD-based PSR. The error sequences of the above methods are then proved to have non-Gaussian properties, and a novel criterion of normalized Renyi’s quadratic entropy (NRQE) is proposed, which can evaluate their dynamic predicted accuracy. Finally, illustrative predictions of the next 1, 4, 6, and 24 h time-scales are examined by historical wind power data, under different evaluations. From the results, we can observe that not only do the proposed models effectively revise the error due to the fluctuation and multi-fractal property of wind power, but also that the NRQE can reserve its feasible assessment upon the stochastic predicted error.

This article has been retracted by the publisher.