Abstract
The wind turbine drive system is one of the key components in converting wind energy into electrical energy. The life prediction of drive system is very important for the maintenance of wind turbine. With increasing capacity, the wind turbine system has become more complicated. Consequently, for the life prediction of drive system, it is necessary to consider the problems of multi-information fusion of big data, quantification of time-varying dynamic loads, and analysis of multiple-damage coupling. In order to solve the above challenges, the fatigue life analysis and evaluation method considering the interaction of coupled multiple damages are proposed in this study. The hierarchical Bayesian theory with fault physics technology is introduced to deal with the uncertainty of wind turbine drive system. Then, a time-varying performance analysis model is established based on the multiple-damage coupling competition failure mechanism. Moreover, the Internet of Things (IoT) technology is introduced and combined with the proposed model. Through the data collection by IoT, the time-stress curve of drive system can be obtained. A case study about the remaining fatigue life estimation of drive system is utilized to illustrate the effectiveness of the proposed method.
1. Introduction
Wind energy is a common clean and renewable energy. The effective utilization of wind energy can reduce the energy crisis and protect the environment. The early development of wind energy is mainly carried out on land. Engineers have successfully built many onshore wind turbines. Compared with offshore wind farms, onshore wind farms are more mature in technology and management [1]. Generally, the offshore wind farms are located on the ocean and have very little impact on humans. Therefore, offshore wind power will account for more proportion in the future wind power industry.
Uncertain environment has restricted the development of offshore wind turbines and threatened the wind turbine drive system security [2]. The design requirements for offshore wind farms are more difficult than those for onshore wind farms [3–6]. Risk assessment can improve the reliability of systems [7–11]. For wind turbine drive system, it is important to study the failure mechanism, fault diagnosis, and life prediction of key components. Accordingly, recent research involves multiple subjects like dynamics, material mechanics, fatigue, reliability, and signal analysis [12, 13].
Taking the wind turbine planetary gearbox as an example, the dynamic analysis mainly focuses on the characteristics of gears and bearings. The gear dynamics model mainly analyzes the gear mesh stiffness characteristics [14], the tooth surface contact force [15, 16], and the influence of assembly and machining errors [17]. The bearing dynamics model mainly analyzes the interaction force between the rolling element and the cage [18, 19] and the influence of the cage parameters on the bearing characteristics [20]. During the operation of the gearbox, the tooth surfaces and roots are subjected to alternating and impact loads, which leads to fatigue damage of the gears. Many uncertainties also cause the failure of the bearing. Thus, the dynamic model is proposed for failure mechanism analysis, fault diagnosis, and life prediction.
The research of gear failure mechanism starts from the tooth root crack [21], the tooth surface pitting [15], and the relationship between crack change and meshing stiffness change [14]. The research of bearing failure mechanism starts from the change of bearing rolling element clearance [22, 23], the number of pressure rolling elements [24], the lubrication characteristics [25] and the cage flexibility characteristics [26]. Recent research results include rolling bearing-rotor systems [27], gear transmission systems [28], gear-rotor-bearing system [29], and gearbox bearing system [30].
The main working mode of wind turbine planetary gearbox is that the load changes with the wind speed. Therefore, the fault diagnosis is the core technology in research process. Many methods have been proposed, such as using instantaneous power spectrum and genetic programming methods to identify fault feature parameters [31]; fault diagnosis of rotating machinery under variable operating conditions based on wavelet packet coefficients and entropy information indicators [32]; and fault information extraction based on the Vold–Kalman method [33].
In the process of fault diagnosis, it is necessary to analyze the internal and external load characteristics. However, the current research mainly focuses on the relationship between a certain fault state and system dynamics. These research works are short on developing dynamic fault response mechanism. This paper proposes the coupling fault diagnosis mechanism and the fatigue life prediction method based on wind farm complex conditions.
Many scholars have carried out research on product remaining life prediction methods, such as model-based methods, data-driven methods, and hybrid methods [34–37]. The product monitoring data have the characteristics of large quantity and variety. Consequently, it is necessary to achieve automatic real-time analysis of product monitoring data by the Internet of Things (IoT). Deep learning methods are also effective ways to solve this problem. However, most of them cannot solve the product remaining life prediction problems. Hence, the product remaining life prediction based on the Internet of Things has a good research prospect.
In this paper, a remaining life prediction method based on deep learning is proposed, and it can be operated under multiple failure models. This method can integrate the multisensor monitoring data with the dynamic working condition data into the fusion data. Through the usage of bidirectional neural network in the deep learning, the time correlation model of the fusion data can be obtained. Then, this method can achieve nonlinear regression analysis of the fusion data and product remaining life under multiple failure models. This paper takes the remaining life prediction of wind turbine drive system as a case to prove the effectiveness of the method.
2. The Construction of Internet of Things (IoT) in the Drive System of Wind Turbine
The working environment of wind turbine drive system is hard, so it is necessary to monitor working conditions and environmental information. Therefore, a sensor is used to collect information as data support for the system. The sensor can convert physical signals into electrical signals so as to satisfy the actual demand [38]. Its accuracy affects the performance of the system, so the selection of sensors should take the system requirements and environment factors into consideration.
The state monitoring and life prediction system of the wind turbine drive system based on the Internet of Things can be divided into three functional levels. (a) Perception layer: The perception layer includes transmission equipment, sensors, and wireless sensor nodes. Its function is to perceive the working conditions of the transmission system. (b) Transmission layer: The transmission layer uses the CAN communication protocol to achieve wired communication between devices and uses the ZigBee communication protocol to build a wireless sensor network. Its function is to deliver the working condition information of the transmission system. (c) Application layer: The application layer is composed of PC/104, operating system, software, and life prediction algorithm. Its function is to achieve acceptance, storage, display, and intelligent processing of the working condition information and manage the entire life cycle of the wind turbine.
Figure 1 shows the system hardware framework. The wireless sensor network collects speed and temperature information of the transmission system and delivers them to the PC/104 for processing. At the same time, the ARM processor collects fault signals, status signals, current signals, and vibration signals of the transmission system and delivers them to the PC/104 for processing. The PC/104 processes the collected information and displays the results and status monitoring information on the LCD screen. The PC/104 can share the information with the centralized control center through the Ethernet communication module. Then, the centralized control center can monitor the transmission system in real time.
3. Data-Driven Fatigue Life Prediction Based on IoT
A large amount of data has been accumulated during the sensor operation. After being processed, the data still contain unused information. The data-driven multivariate statistical algorithm can use these data to carry out process monitoring, fault monitoring and diagnosis, and quality and life prediction [39–41].
The product remaining life is the time interval between the current moment and the failure moment. Specifically, the product monitoring information at this moment is compared with the historical information to get the comparison conclusion and predict the product remaining life.
3.1. Data Collection
The degradation status information of products during operation is monitored by multiple sensors. Then, the data is defined as online measurement data. The monitoring data of training products are recorded until failure and saved to . represents the online monitoring data of the i-th training product. The data are monitored by p sensors. Each sensor employs M data points. L is the total number of training products. represents the degradation state of the i-th training product and needs to be detected offline. The monitoring data of texting products are recorded and saved to . represents the online real-time monitoring data of the j-th testing product. C is the total number of testing products. The information about the degradation status and remaining life of the testing product is unknown, and it needs to be further estimated based on .
3.2. Data Preprocessing
The data preprocessing can be divided into three sections, namely, feature extraction, data regularization processing, and sliding time window processing.
The feature extraction of the monitoring data can reduce the data size and increase work efficiency. The noisy and abnormal data points will affect prediction accuracy of the method. The feature extraction can reduce the above-mentioned adverse effects. First, the time domain features of the monitoring data can be extracted, such as mean, standard deviation, root mean square, skewness, and kurtosis. Second, the frequency domain features of the monitoring data can be extracted by the fast Fourier transform technique, such as average frequency, spectral skewness, spectral kurtosis, and spectral power. Third, the time-frequency domain features can be extracted by the empirical mode decomposition and wavelet change technique.
After the above steps, the monitoring data and can be converted into set and . and represent the initial features of the i-th training product and the j-th testing product . N represents the number of initial features. However, the extracted features have different data size, so the z-score regularization method is used to normalize data, as follows:where represents the d-th feature quantity for the k-th sample, represents the result after normalization, and and represent the mean and standard deviation of the d-th feature in set.
3.3. Life Prediction Method Based on Particle Filter Algorithm
The network structure based on the bidirectional long short-term memory (Bi-LSTM) model can be divided into two sections: (1) extracting the time-dependent features of normalized data by the Bi-LSTM network; (2) processing the time-dependent features and outputting the prediction result of product degradation indicators. A schematic diagram of the correlation between the two sections is shown in Figure 2.
More details can be found as follows: Step 1: the training set can be obtained by processing monitoring data. represents the data of the k-th training feature. represents the length of the data. represents the normalized feature data at time t. represents the product degradation state of . The value of is the data of the offline degradation state. is normalized and its time-dependent features are extracted by the Bi-LSTM network, as follows: where represents the hidden layer function of the Bi-LSTM network. The parameter set of the Bi-LSTM network is described by . All output features of the Bi-LSTM network are given in (2). However, only the output features of the last time point of the second layer are used for the fully connected layer; the calculation process is as follows: where and represent the output features of the forward and reverse Bi-LSTM networks at the last moment, and represents the sum operation point-by-point. Step 2: The output feature of the Bi-LSTM network is used in the fully connected layer to further extract the advanced features. The calculation process is as follows:where is the output features of the fully connected layer and is the parameter set of the fully connected layer, including the weight matrix and the bias vector . Then, is used to predict the degradation state of the product. The process can be described by the following equation:where is the prediction of the product degradation index and is the weight vector of the last linear regression layer.
Step 1 and Step 2 both use single-layer bidirectional long short-term memory neural network expression and fully connected neural network expression. In actual engineering, more layers can be stacked to form a deep structure. However, the deep neural network structure has a more complex model and needs more training data.
The calculation process of the model forward propagation has been given earlier. Then, the model loss function based on mean variance can be defined as follows:where represents the normalized value of the degradation indicators in the shutdown state, represents the prediction of the product degradation indicators , and is defined by (2)–(5).
For some complex products, the degradation state data is difficult to obtain online directly because of actual engineering limitations. Hence, online monitoring techniques are used as auxiliary means to obtain indirect data and complete the estimation of the degradation state. In order to define the problem of product degradation state estimation, firstly, the relationship between the degradation state and time is given as follows:where k represents the time point, represents the product state transition model, represents the internal degradation state of the product at time k-1, and represents an independent and identically distributed process noise. At any time point k, the online measurement data of the product can be obtained, and the nonlinear relationship between and is given as follows:where is the observation model and is an independent and identically distributed measurement noise.
The online measurement data is collected by multiple sensors at very high sampling frequency, and its signal to noise ratio is low, so it is difficult to obtain the display analytical model between the online monitoring data and the internal degradation state. Therefore, the online measurement data is processed earlier, including feature extraction, normalization, and serialization. Then, the online monitoring data can be converted into a set of normalized data . Finally, the model between the extracted feature and the internal degradation state is established based on the Bi-LSTM network and is given as follows:where represents the predicted output value of the product degradation index, represents the internal degradation state of the product in the shutdown state, and is the measurement noise and follows a normal distribution.
In the framework of Bayesian learning, the main task is to calculate the posterior probability distribution of the product degradation state when the real-time online measurement data is known [42–47]. is processed and transmitted to the Bi-LSTM network to obtain the product degradation index predicted output value . Thus, the posterior probability distribution is actually obtained. If is known at time k, the probability density function based on the Chapman–Kolmogorov equation is given as follows:
At time k + 1, the new online measurement data is processed and transmitted to the Bi-LSTM network to obtain the product degradation index prediction output . Then, the posterior probability distribution can be updated based on the Bayesian criterion as follows:where represents the likelihood function at time k + 1 and can be defined by (9), the probability density distribution can be calculated through the prediction, and represents the normalization factor. The calculation process is as follows:
Because of the nonlinear features, it is difficult to obtain exact solutions of above equations, so particle filtering is introduced to solve these problems. The structure of particle filtering is actually the Monte Carlo method (which refers to the probability of the event by the frequency of occurrence at a certain time) with a layer of importance sampling proposed in it. The basic idea of this method is to use a set of samples (or particles) to approximate the posterior probability distribution of the system, and then use this approximate representation to estimate the state of the nonlinear system. Using this idea, particle filtering can handle any form of probability in the filtering process. At the same time, is approximated by Bayesian filtering method based on Monte Carlo simulation. If is known at time k and can be described by discrete values and their weights , the calculation process is given as follows:where , represents the Dirac function.
Equation (10) can be rewritten as follows:where i represents the i-th particle, N represents the total number of particles, and can be calculated based on the state transition model.
The weight value of (9) can be updated by the importance sampling method, and the update process is given as follows:where represents the likelihood function of the predicted output .
In order to avoid the particle degradation during the operation of the particle filtering algorithm, the resampling method based on the inverse cumulative function is used to obtain particles with equal weights. The steps are as follows: Step 1. Construct the cumulative density function based on the likelihood function . Step 2. Sample a random value from the uniform distribution U (0, 1) and use it as the input of the cumulative density function in Step 1. Step 3. Take out the particle whose value is most similar to the output of CDF in Step 2 and consider it as the result after resampling. Step 4. Repeat the calculation process of Steps 1–3 N times and obtain N equal weight resampling particles. Then, use these particles to approximate the posterior probability density distribution .
4. Case Study
4.1. Doubly Fed Wind Turbine Drive System
The wind turbine drive system can transmit torque and change rotating speed. Its structure has a great influence on the layout of wind turbine. At present, the wind turbine has two types: doubly fed and direct-drive. The difference between them is the structure of the drive system. The drive system of the doubly fed wind turbine has a speed-increasing gearbox, while the drive system of the direct-drive wind turbine does not have one. The object of this paper is the doubly fed wind turbine drive system.
The wheel hub of the doubly fed wind turbine is connected to one side of the low-speed main shaft. The other side is connected to the input of the speed-increasing gearbox. The output of the gearbox (high-speed shaft) is connected to the generator rotor through the flexible coupling. The hydraulic brake disc is installed on the high-speed shaft. The entire drive system is installed on the main frame. The thrust and radial load generated by the wind wheel are transmitted to the main frame first and then to the tower.
Because of the speed-increasing gearbox, the generator rotor can reach a high rotating speed and the requirement for magnetic poles is low. The gearbox calls for a high speed-increasing ratio and use multistage planetary gears combined with fixed-shaft gears. The characteristics of the doubly fed wind turbine drive system are as follows: (1) The operating conditions change with the wind speed and lead to constant change of vibration signal amplitude and frequency. As a result, the signal contains a large number of uncertainties. (2) The drive system is located in the top of the tower. Strong wind will cause the body shake. Hence, the signal will contain noise interference. (3) The mechanical structure of wind turbine drive system is complex. It calls for high standard of related technologies. (4) Wind turbines are severely affected by environmental conditions and have high requirements for the reliability and anti-interference ability of the test system.
4.2. Dynamic Modeling of Wind Turbine Drive System
This paper uses the dynamic analysis process of the wind turbine gearbox drive system as shown in Figure 3.
First of all, the external and internal incentives of the gearbox are obtained through the monitoring of the Internet of Things technology. They are transmitted as dynamic incentives to the analysis model of gearbox system. Second, the analysis model of gearbox system processes the input dynamic incentives through centralized parameter method, finite element method, hybrid modeling method, and bong graph theory. After processing, the inherent features and dynamic responses are output, and they are used as the dynamic responses of the analysis system. Finally, the particle filter algorithm is used for condition monitoring and life prediction applications. The above is the main application process of the method proposed in this paper for the wind turbine drive system.
The external excitation of the gearbox mainly includes the dynamic input torque, the nontorsional loads, the generator feedback torque, the braking torque, and the impact load. The internal excitation includes time-varying stiffness excitation, time-varying error excitation, meshing shocks, and tooth backlash. This paper uses dynamic response analysis method, dynamic stability analysis method, and inherent characteristic analysis method to optimize the system structure. Based on the IoT technology, the dynamic model has been proposed to monitor the gearbox status and provide theoretical basis of fault diagnosis and life prediction.
4.3. Method Validation
The gearbox wear tests are recorded as A1, A2, and A3. Each test includes online monitoring data from six signal channels, and each test collected 600 sets of online monitoring data during transmission process. In order to describe the online monitoring data clearly, 50 equidistant data monitoring points are used for each online monitoring data group.
In order to reduce the data size and improve prediction efficiency, the method extracts the time domain feature first and then the frequency domain feature. The BL-STM model has been used to simplify the feature processing flow.
During the experiment, A1 is used as the test sample, and A2 and A3 are used as the training samples. The prediction result of the final transmission pair loss is shown in Table 1. According to the experiment results, the top five features with the highest scores can achieve the best prediction performance.
The influence of the feature fusion step on the final transmission pair loss prediction result is shown in Table 2. The optimal feature set used in Table 2 is composed of the top eight features with the highest scores. It can be inferred from Table 2 that the fusion feature with the highest scores can bring the best prediction performance.
Combined with the state transfer model of transmission pair wear and its observation model based on Bi-LSTM, the particle filter algorithm is used to predict the evolution of the transmission pair degradation index. When the index exceeds the failure threshold, the remaining useful life (RUL) of transmission pair can be obtained.
5. Conclusions
Based on the bidirectional long short-term memory neural network and particle filter technology, the remaining life prediction method for wind turbine drive system is proposed. Compared with the traditional method, it simplifies the feature processing flow by using deep learning technology. Through the engineering examples, the effectiveness of the method proposed in this paper has been proved. The research results in this paper will further promote the engineering application of deep learning technology to the product fatigue life prediction higher safety requirements.
Data Availability
All data used to support the findings of this study are included within the article.
Conflicts of Interest
The authors declare that they have no conflicts of interest regarding the publication of this paper.
Acknowledgments
This research was supported by the National Key Research Development Plan Funds under contract no. 2018YFB1702804.