Diagnostic and Monitoring Technology Based on Extreme Learning Machine

Ref. [84] proposed a data-driven method based on support vector data description (SVDD) and extreme learning machine (ELM) algorithm to achieve effective monitoring of the fan’s bad state. First get the wind turbine data from the sensor. Use SVDD classification algorithm to separate unhealthy data. Then, based on this balance data set, an effective classifier is constructed by ELM to monitor the unhealthy state. The specific implementation process is shown in Figure 7.30.

Use three characteristic indicators (e.g., Gini index, information value and Cramer’s V) to reflect the influence of variables on pattern recognition. The

The framework of monitoring not-runnable status of wind turbines

FIGURE 7.30 The framework of monitoring not-runnable status of wind turbines.

proposed method mainly includes two stages to realize the monitoring of the fan status. The first stage is to separate health data from unhealthy data. The second stage is to classify different unhealthy states, namely the various nonoperational states of the wind turbine. By comparing with the six models combining support vector machine (SVM), ELM. principal component analysis (PCA) and SVDD, this method improves the accuracy of the classification of unhealthy data and inoperable states. However, this article only considers two specific nonoperational states (WTs are stopped due to vibration in tactical air control (TAC84); WTs are stopped by remote control due to excessive power generation and reduction capacity), the use of this method has certain limitations.

In reference [85], a data-based method was proposed to estimate the health status of wind turbines, so as to improve the active power output of wind turbines. The degree of disability is estimated using an ELM algorithm combined with Bonferroni intervals. The data collected from the measurement is selected as the input of ELM model to predict the output signal of physical process. And the actual output signal obtained from the wind turbine is compared with the predicted output signal of the corresponding input signal. The Bonferroni interval method, which has more accurate confidence intervals, is applied to estimate the deviation level of each component in this study.

Flowchart of proposed method

FIGURE 7.31 Flowchart of proposed method.

Next, using analytic hierarchy process (AHP), various factors such as the degree of failure, maintenance cost, and maintenance time were comprehensively examined to estimate the health condition of the turbine structure. The procedure to evaluate the health condition of the turbine using AHP is as follows: (1) Define the evaluation criteria and build a hierarchical structure; (2) compose the pairwise criterion elements; (3) establish the comparison matrix and the alternative pairwise comparison matrix; and (4) calculate the relative weights of each criterion to estimate the health of the wind turbines.

In the end, in order to reduce the fatigue load of the failed unit and improve the operating efficiency of the normal unit, power distribution control is performed according to the operating state of the unit. Therefore, the process of the power distribution scheme is defined as follows (shown in Figure 7.31): for a healthy wind turbine, they are set in normal mode, which is the maximum power under variable wind speed. The failed wind turbine is operated in the power-reduced mode according to its health condition. Therefore, the output demand P* is defined as follows:

where /V, and N2 represent the number of the healthy and faulty wind turbines respectively; Ры is the power output of the /th healthy turbine operating in the baseline mode, whereas Pf) is the power output of the /th faulty turbine working in a power- reduced mode. Cfj is the power distribution coefficient of the /th faulty turbine. Power was distributed according to the power distribution coefficient of the failed turbine considering the health condition.

Diagnostic and Monitoring Technology Based on Intelligent Algorithm

To solve the problem that the abnormal load change is not always the fault of the component itself, a WTG state assessment method is proposed. First, according to the structure and working principle of wind turbine generator unit (WTGU), WTG evaluation index system was established, and the mathematical model of each condition index was trained by using the monitoring data under normal conditions. Second, the residual and deterioration degree of each index were given, and the evaluation standard cloud was established. Then, during the period to be evaluated, one needed to select the monitoring data, standardize the data, and perform a conditional evaluation. The flow chart of condition assessment of WTG is shown in Figure 7.32.

First, the mapping relationship between input and output physical quantities is established by using the normal operation data of the pre-processed unit. Then, the prediction model of evaluation index is established to calculate the deviation between the predicted value and the measured value. The least square support vector machine (LSSVM) starts with the loss function of machine learning, optimizes the objective function with two norms, and replaces the inequality constraint in SVM with the equality constraint. And then normalization is done to get the degree of degradation. Since the sensitivity of each index to the evaluation result is different, the weight is used to represent the sensitivity degree of this index. A set of fused data is obtained through the weight fusion degradation degree, and the fused data is used to obtain the digital characteristics of the conditional cloud model to be evaluated. The advantage of this method is that it can reduce the downtime while maintaining the generator state.

Reference [87] presents an online operation optimization strategy by optimizing the operation parameters. The operation parameters related to active power are extracted by loss analysis. Then the dimension of the array of operation parameters is reduced by factor analysis, and finally the reference value of operation parameters is obtained by cluster analysis. The specific implementation process is shown in Figure 7.33. К-mean clustering algorithm is selected to obtain the reference value of operation parameters. According to the wind speed, the WT operation is divided into different working conditions. According to the clustering results, the cluster center with the maximum active power is taken as the reference value of the operation parameters. The advantage of this model is that it can accurately determine the effective values of wind speed and power during operation, and its calculation accuracy is high, avoiding the influence of high-dimensional data. However, this paper does not give the operating parameters and reasons that lead to the decrease of operating efficiency.

Because wind turbines undergo a variety of state changes during operation, including normal operation of the turbine, idling, maintenance/repair mode, failure mode, weather downtime, etc. Therefore, state-based maintenance tools are needed to predict failure patterns in the system. In reference [88], a method for

Flow chart of condition assessment of wind turbine generator

FIGURE 7.32 Flow chart of condition assessment of wind turbine generator.

predicting the state of wind turbines is proposed with three typical steps, as shown in Figure 7.34.

Step I: Abstraction of turbine state. The possible states of the turbine are classified according to the different stages of prediction. In the first stage, the turbine can be divided into four states: normal turbine, failure, weather shutdown and maintenance shutdown. Where the weather shutdown category corresponds to a turbine shutdown due to adverse weather conditions, and any other shutdown is considered a maintenance shutdown. The second stage is classified by replacing the faults in stage I with the actual fault types, which are pitch overrun 0°, pitch thyristor 2 fault, axle 1 fault pitch

Flow chart of model construction

FIGURE 7.33 Flow chart of model construction.

controller and pulse sensor motor defect. The third stage also added practical failures such as out of control yaw, brush wear warning, blade Angle unreliability and advanced response generator.

Step 2: Learn Strategy. Data mining algorithm is used to establish wind turbine failure prediction model, through the use of five data algorithm, NN. SVM, random forest algorithm (RFA), boosting tree algorithm (BTA) and general chi-square automatic interaction detector (CHAID) to training, two-thirds of the primitive data types of the remaining one third of the original data for testing. The geometric mean (gmean) of the output category is used as the evaluation index, and the expression is:

where ясс, is the accuracy of class i, n is the total number of output classes.

Framework of the proposed approach

FIGURE 7.34 Framework of the proposed approach.

Step 3: State prediction. The primary goal of the first phase is to predict any type of failure. The second stage predicts specific failures. Replace the type of fault with the output type in phase I with the actual fault type. The third stage predicts invisible failures in different wind turbines.

Other Diagnostic and Monitoring Methods

An overall monitoring system based on field programmable gate array CPU has been proposed [89]. The system can not only provide status and subsynchronous control interactions (SSCI) monitoring functions at the same time, but also record the data required for analysis after the event. A monitoring system based on the overall internet of things (IoT) was developed for wind turbines. The system used a field programmable gate array (FPGA)-CPU hybrid controller to continuously complete data acquisition, data processing, and data recording to improve system performance. In this monitoring system, as long as the predefined detection mechanism is triggered on the FPGA target, the vibration and voltage signals were recorded on the CPU target. The recorded data can be sent from the FPGA target to the CPU target and stored to the network-attached storage (NAS). Through the monitoring and detection of SSCI events and the status of wind turbines, this method can provide detailed information for the system operators to avoid unnecessary coordination between the two independent monitoring systems.

Reference [90] proposed a method for tracking the power signal using an adaptive filter based on continuous wavelet transform (CWT) to monitor the state of the wind turbine. The technology can track the energy in the power signal at the specified fault-related frequency band rather than at all frequencies of the monitoring signal.

and can display the results graphically. The energy A in the fault-related frequency band at each time interval can be calculated by using Equations (7.43) and (7.44). These calculations were repeated until the curve of the energy change in the fault- related frequency band was obtained, and then the change in the operating condition of the WT can be evaluated.

where CWT,oca, is the matrix of wavelet coefficients, a is the wavelet scale and a = ft),/ft), ft) represents frequency, and Г is the time interval.

This method has been proven to detect two types of faults, including electrical and mechanical faults. The advantage of this technique is to reduce the calculation time of feature extraction, low cost and strong versatility.

In reference [91], an overview on grid-friendly wind turbines (WTs) and relevant technologies for control and monitoring was presented. In this paper, the aerodynamic conversion system, generator and front-end speed regulation (FESR) system module composed of FESR are established, which is similar to double-fed induction generator (DFIG) system. Thus, a friendly grid connection is achieved, with better stability and low-voltage ride-through (LVRT) capability. As shown in Figure 7.35, a “virtual generator” is introduced to achieve the torques of the low-speed shaft virtually. Then, a Pi-based virtual controller can be developed, and the design interface for FESR is achieved in Bladed. As FESR contains two transmission modules, mechanical strength and fatigue tests should be performed.

Modeling and control scheme for FESR WT in reference [91]

FIGURE 7.35 Modeling and control scheme for FESR WT in reference [91].

Suggested torque control scheme for load reduction in reference [91]

FIGURE 7.36 Suggested torque control scheme for load reduction in reference [91].

The improvement methods for load reduction of the system are as follows: (1) IPC should solve various influences caused by coupling nonlinearity and uncertainty of pitch blade, so as to realize accurate system model of IPC; (2) Proper torque control is also helpful to reduce load. One suggested approach is showm in Figure 7.36. The schematic curve of generator torque-speed depicts the four generator operation areas, as AB, BC, CD, and DE. The torque control command is compensated wfith the additional torque derived from the band-pass filter and gain block, so that the compound command is achieved.

Similarly, grid-friendly WTs capabilities can be enhanced in pow'er electronics technologies, such as novel configurations or converter topologies. From the point of view of the monitoring scheme based on SCADA, a comprehensive scheme consisting of “data preprocess,” “regular data process,” and “fault data process” is proposed to get a comprehensive monitoring mechanism for the system.

< Prev   CONTENTS   Source   Next >