Design and implementation of automatic fault diagnosis system for wind turbine

Highlights

• A new method based on time-frequency domain combined with phase space reconstruction and singular value decomposition is proposed.

• The proposed method integrated with wavelet transform, phase space reconstruction and singular value decomposition can lead to fast and efficient results.

• Validity of the proposed method proposed is verified by pratical fault diagnosis application of wind turbine.

Abstract

Operation of wind turbines under fault state will directly affect the power output efficiency of wind farms. This paper proposes a new automatic fault diagnosis method for wind turbines. A fault diagnosis system framework is constructed and data of vibration status of wind turbines collected is processed and used for fault diagnosis. Firstly, wavelet coefficients are obtained using a discrete wavelet transform (DWT) for vibration acceleration signals collected from wind turbines. Then, the wavelet coefficients are sequentially subjected to phase space reconstruction (PSR) and singular value decomposition (SVD) to extract the fault features. Finally, an extreme learning machine (ELM) is used to classify the faults. Experimental results show that the proposed method is more effective and accurate than other fault diagnosis methods for wind turbines, such as support vector machine (SVM) and multiscale convolutional neural network (MSCNN).

Introduction

Energy is the material basis for the sustenance and development of human society. At present, coal, oil, natural gas, and nuclear energy constitute the world's main sources of energy. However, coal combustion causes considerable pollution, oil and natural gas supplies are being depleted, and the disposal costs of nuclear waste are much higher than construction costs. Thus, humankind is facing an energy crisis. Wind energy, as a type of clean energy, is becoming increasingly popular around the world [1], [2]. In recent years, China's wind power industry has witnessed rapid development. According to the statistics of the World Wind Energy Association (WWEA) published on February 25, 2019, the total installed capacity of wind power in the world reached 600 GW by the end of 2018, of which the new installed capacity was 53.9 GW. Furthermore, China, with a new installed of 25.9 GW, is the first country in the world whose installed capacity of wind power has exceeded 200 GW. China is the country with the largest installed capacity of wind power in the world. However, due to the short time for China to introduce wind power, the fault analysis technology for wind turbines is relatively backward, which has a lot of room for improvement. According to incomplete statistics, the average utilization rate of wind turbines in China is generally less than 95%, in which the high failure rate is a major factor, which leads to the maintenance cost becoming the main cost of wind farms. Therefore, reducing the maintenance costs is an important way to improve the operating efficiency of wind farms. In general, wind farms are located in harsh environments and face unsteady operating conditions. The signals detected by the sensors are affected by faults and various external factors. Moreover, the vibration signals of wind turbines are non-stationary and nonlinear. Wind turbine faults will not only affect electricity generation but also lead to serious safety issues, resulting in considerable economic losses. For the existing mature fault diagnosis method for wind turbines, it takes a lot of economic and manpower to better realize the fault diagnosis of the wind turbine. In recent years, a large number of researchers have attracted a lot of research on unstable and nonlinear signals.

Through steady development, fault diagnosis technology has emerged as an independent and comprehensive interdisciplinary information processing technology. It can be categorized into traditional diagnosis methods, mathematical diagnosis methods, and intelligent diagnosis methods. Traditional diagnosis methods include vibration detection technology, oil analysis technology, noise detection technology, and infrared temperature measurement technology [3]. Mathematical diagnosis methods include pattern recognition based on Bayesian decision criteria [4] and linear/non-linear discriminant functions, time series model diagnosis based on probability statistics, fault diagnosis based on distance criteria, the fuzzy diagnosis principle, gray system diagnosis, fault tree analysis, SVM [5], principal component analysis (PCA) [6], and wavelet analysis [7]. Intelligent diagnosis methods include fuzzy logic [8], expert systems, neural networks [9], and genetic algorithms. As the research of artificial intelligence becomes more and more mature, intelligent diagnostic methods enter the factory and gradually become the main direction of fault diagnosis. The SVM classifier can effectively solve the small sample and over-fitting problems, but the SVM needs to solve the support vector by means of quadratic programming, which is difficult to implement for large-scale training samples, and it is difficult for SVM to solve the multi-class problem. The main idea of PCA is to map n-dimensional features to k-dimension. This k-dimensional is a new orthogonal feature, also called principal component, which is a k-dimensional feature reconstructed on the basis of the original n-dimensional features. The significance of wavelet transform (WT) is that it can decompose the signal at different scales, and the selection of different scales can be determined according to different targets, which can realize the simultaneous representation of signals in the time domain and the frequency domain.

Fault signal processing and analysis methods have also witnessed considerable advancement and widespread use, from traditional analysis methods such as time domain waveform analysis, speed synchronization analysis, power spectrum analysis, detailed spectrum analysis, correlation analysis, coherence analysis, cepstrum analysis, and demodulation analysis, to new analysis methods such as wavelet analysis [10], Wigner–Ville technology [11], Hilbert demodulation, and other time-frequency analyses [12]. With the onset of the 21st century, researchers have increasingly focused on combinations of traditional or mathematical diagnostic methods with intelligent methods, such as the combination of a DWT and an ELM [14], [15]. Li et al. [15] proposed a fault detection and classification method for medium voltage DC shipborne power system combining wavelet transform multiresolution analysis technology with artificial neural network. The wavelet transform multiresolution analysis technique and Parseval theorem are used for different faults. The features are extracted and then the artificial neural network is used to automatically classify the fault types. Further, methods for fault diagnosis have been proposed on the basis of a combination of pattern recognition and artificial intelligence, fault feature extraction and separation based on empirical mode decomposition (EMD), and resonance demodulation.

Based on the theoretical analysis and the processing of the field data, this paper presents a new method of reasonable and higher accuracy for the signal of the non-stable state of the wind power generation, and can finally be used in the actual wind farm. On the basis of the fault diagnosis of wind power generation, a new hybrid diagnosis method is proposed in this paper. According to the unsteady and nonlinear characteristics of the acceleration vibration signal of wind power generation, combined with DWT, PSR and SVD, the fault feature extraction of wind turbine is realized. Finally, the classification is realized by using an ELM.

The remainder of this paper is organized as follows. A background of fault diagnosis and the whole fault diagnosis system construction of wind turbines are provided in Section 2. A sequential method consisting of a DWT, PSR, SVD, and an ELM is proposed in Section 3. Experimental results conducted on real wind turbine data are presented and comparison results with other methods are given in Section 4. Finally, Section 5 concludes the paper.