Wind turbine fault early warning method and system based on optimized deep belief network

Abstract

The application belongs to the technical field of wind turbine units, and specifically provides a wind turbine unit fault early warning method and system based on an optimized deep belief network, comprising: a data acquisition and monitoring control SCADA system based on a wind farm installation, which obtains SCADA data; wherein the SCADA data comprises wind farm surrounding environment parameters and wind turbine unit component performance parameters; a KNN-box chart combination algorithm is used to detect outliers of the SCADA data; a sparrow search algorithm SSA and a particle swarm algorithm PSO are combined to optimize parameters of a deep belief network DBN, and a wind turbine unit fault early warning model based on SSA-PSO-DBN is constructed; and based on the wind turbine unit fault early warning model, early warning of wind turbine unit faults is completed. The algorithm provided by the application has high accuracy in judging the running state of a wind turbine unit, can further reduce the risk of misjudgment, and can realize early warning of wind turbine unit faults.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wind turbine units, and particularly relates to a wind turbine unit fault early warning method and system based on an optimized deep belief network. BACKGROUND

[0002] In recent years, wind energy has received extensive attention from countries around the world. As a clean energy, wind energy is rising at an unprecedented speed in the global energy landscape, becoming a key force to drive energy transformation, address climate change and achieve sustainable development. Wind turbine units, as the core equipment for wind energy utilization, have complex and diverse working environments, containing many advantages conducive to power generation and facing many severe challenges. These factors jointly affect the operating efficiency, service life and safety of wind turbine units.

[0003] In the prior art, fault early warning algorithm research is based on SCADA data under the normal state of wind turbine units. A data reconstruction model based on a neural network is established by selecting specific variables. However, the parameters such as the number of nodes in the deep learning network are not deeply researched during the model establishment stage. However, neural network parameters such as the number of nodes in the layer and the learning rate will affect the accuracy of the model for fault early warning. Therefore, the parameters of the neural network need to be more accurately determined before modeling to improve the recognition accuracy of the model for the operating state of the unit.

[0004] Moreover, most current researches use a single algorithm to clean the SCADA data of wind turbine units, but many algorithms have their own defects, and sometimes a single algorithm cannot effectively clean the data. SUMMARY

[0005] To solve the problems in the prior art, the application provides a wind turbine unit fault early warning method and system based on an optimized deep belief network, which further reduces the risk of fault misjudgment and can realize early warning of wind turbine unit faults.

[0006] To achieve the above-mentioned purpose, the application provides the following solutions.

[0007] The wind turbine unit fault early warning method based on an optimized deep belief network comprises the following steps:

[0008] Based on the data acquisition and monitoring control SCADA system installed in the wind farm, SCADA data is acquired; wherein the SCADA data comprises wind farm surrounding environment parameters and wind turbine unit component performance parameters;

[0009] The KNN-box plot combination algorithm is used to detect abnormal values of the SCADA data, and cleaned SCADA data is obtained;

[0010] The joint sparrow search algorithm (SSA) and the particle swarm optimization (PSO) are used to perform parameter optimization on a deep belief network (DBN), and a wind turbine fault early warning model based on the SSA-PSO-DBN is constructed.

[0011] Based on the wind turbine fault early warning model, the wind turbine fault early warning is completed.

[0012] Preferably, the method for performing outlier detection on the SCADA data comprises:

[0013] A wind speed threshold is established, for a wind speed interval less than the wind speed threshold, a KNN algorithm is used to calculate the distance between a target sample and a training sample in the SCADA data, the training samples are sorted in ascending order of distance, k nearest neighbor training samples that meet a preset distance requirement are screened out, and the number of labels of the k nearest neighbor training samples is counted, and the label with the highest occurrence frequency is taken as the outlier detection result of the target sample.

[0014] For a wind speed interval greater than or equal to the wind speed threshold, a box plot algorithm is used to obtain the quartiles and interquartile ranges of the SCADA data, and the outliers of the SCADA data are identified.

[0015] Preferably, the method for constructing the wind turbine fault early warning model comprises:

[0016] The sparrow population and related parameters are initialized, each sparrow is classified, and DBN training is performed on the classified sparrows to obtain a first DBN training result; wherein the types of sparrow classification include discoverer, joiner and alarm.

[0017] The sparrow positions are updated based on the first DBN training result, and the sparrows are reclassified and then subjected to DBN training again until a first preset iteration number is reached, and the optimal solution of SSA optimization is obtained.

[0018] Particles are generated around the optimal solution of SSA optimization, DBN training is performed on each particle to obtain a second DBN training result, and the positions and velocities of the particles are updated based on the second DBN training result until a second preset iteration number is reached, and the optimal solution of PSO optimization is obtained.

[0019] The optimal solution of PSO optimization is input into a deep belief network (DBN) model, the SSA-PSO joint optimization is completed, and the wind turbine fault early warning model is obtained.

[0020] Preferably, in the SSA-PSO joint optimization process, the method for obtaining the optimal solution is:

[0021] The fitness value of each sparrow in the sparrow search algorithm (SSA) and the fitness value of each particle in the particle swarm optimization (PSO) are calculated respectively, and the optimal solution is found by taking the fitness value as a key index; wherein, the fitness function for calculating the fitness value is as follows:

[0022] ,

[0023] In the formula, represents the fitness value, S represents the data length, D represents the dimension, represents the reconstructed data, represents the original data, s and d represent the first data and the first dimension, respectively. s d

[0024] Preferably, the method for early warning of wind turbine faults based on the wind turbine fault early warning model comprises:

[0025] The target SCADA data is input into the wind turbine fault early warning model to generate reconstructed data, and the reconstruction error is calculated based on the reconstructed data;

[0026] When the reconstruction error exceeds the preset alarm threshold, it is determined that the wind turbine state is abnormal, otherwise it is determined that the wind turbine state is normal.

[0027] The application also provides a wind turbine fault early warning system based on an optimized deep belief network, which is used to implement the method, comprising:

[0028] A data acquisition module is configured to acquire SCADA data based on a data acquisition and monitoring control (SCADA) system installed in a wind farm; wherein, the SCADA data comprises environmental parameters around the wind farm and performance parameters of wind turbine components;

[0029] A data cleaning module is configured to detect outliers in the SCADA data by using a KNN-box plot combination algorithm, and obtain cleaned SCADA data;

[0030] A model construction module is configured to jointly use a sparrow search algorithm (SSA) and a particle swarm optimization (PSO) to optimize parameters of a deep belief network (DBN), and construct a wind turbine fault early warning model based on SSA-PSO-DBN;

[0031] A fault early warning module is configured to complete early warning of wind turbine faults based on the wind turbine fault early warning model.

[0032] Preferably, the data cleaning module comprises:

[0033] ​​The KNN cleaning unit is configured to set a wind speed threshold, and for a wind speed interval less than the wind speed threshold, calculate distances between a target sample and training samples in SCADA data by using a KNN algorithm, sort the training samples in ascending order of distance, filter out k nearest neighbor training samples satisfying a preset distance requirement, and count numbers of labels of the k nearest neighbor training samples, and take a label with a maximum number of occurrences as an anomaly value detection result of the target sample.

[0034] The box plot cleaning unit is configured to, for a wind speed interval greater than or equal to the wind speed threshold, obtain quartiles and interquartile ranges of SCADA data by using a box plot algorithm, and identify anomaly values of the SCADA data.

[0035] Preferably, the model construction module comprises:

[0036] The SSA initialization unit is configured to initialize sparrow populations and related parameters, classify each sparrow, and perform DBN training on the classified sparrows to obtain a first DBN training result; wherein the types of sparrow classification include discoverer, joiner and alarm.

[0037] The SSA optimization unit is configured to update sparrow positions based on the first DBN training result, reclassify the sparrows, and perform DBN training again until a first preset iteration number is reached to obtain an optimal solution of SSA optimization.

[0038] The PSO optimization unit is configured to generate particles with the optimal solution of SSA optimization as a center, perform DBN training on each particle to obtain a second DBN training result, and update positions and speeds of the particles based on the second DBN training result until a second preset iteration number is reached to obtain an optimal solution of PSO optimization.

[0039] The model construction unit is configured to input the optimal solution of PSO optimization into a deep belief network (DBN) model, complete SSA-PSO joint optimization, and obtain a wind turbine fault early warning model.

[0040] Compared with the prior art, the KNN-box chart combination algorithm of the application combines the advantages of two data cleaning algorithms, divides the wind turbine SCADA data according to wind speed, and divides the KNN algorithm and the box chart algorithm into respective bin intervals for data cleaning. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more clearly illustrate the technical solutions of the present application, the following briefly introduces the drawings needed in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0042] Figure 1 For the box chart algorithm of the embodiment of the present application Figure Four

[0043] Figure 2 For the KNN algorithm data cleaning result graph of the embodiment of the present application

[0044] Figure 3 For the box chart algorithm data cleaning result graph of the embodiment of the present application

[0045] Figure 4 For the RBM structure diagram of the embodiment of the present application

[0046] Figure 5 For the DBN structure diagram of the embodiment of the present application

[0047] Figure 6 For the SSA-PSO-DBN wind turbine fault early warning flow chart of the embodiment of the present application

[0048] Figure 7 For the SCADA data wind speed-power binning result graph of the embodiment of the present application

[0049] Figure 8 For the KNN-box chart combination algorithm data cleaning result graph of the embodiment of the present application

[0050] Figure 9 ​A comparison chart of reconstruction errors of two models of embodiments of the present application is shown in the figure;

[0051] Figure 10 A schematic diagram of reconstruction errors and alarm thresholds of a test data set under a normal state of embodiments of the present application is shown in the figure;

[0052] Figure 11 A schematic diagram of reconstruction errors and alarm thresholds of a test data set under an abnormal state of embodiments of the present application is shown in the figure. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0054] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] Embodiment one:

[0056] As shown in the figure, the wind turbine fault early warning method based on an optimized deep belief network comprises: Figure 6

[0057] S1: Based on the data acquisition and monitoring control (SCADA) system installed in the wind farm, SCADA data is obtained; wherein the SCADA data comprises environmental parameters around the wind farm and performance parameters of the wind turbine components.

[0058] S2: The KNN-box plot combination algorithm is used to detect the abnormal values of the SCADA data, and the cleaned SCADA data is obtained. Specifically, before using the SCADA data to model the fault early warning, the abnormal values in the data need to be detected and removed to improve the modeling accuracy and prevent the abnormal values from affecting the model performance. At present, most data detection methods are single processing of one algorithm, but this method often has the problem of over-cleaning or insufficient cleaning of part of the data. Specifically, although the KNN algorithm and the box plot algorithm can both clean the SCADA data, they have their own problems. For example, Figure 2 , Figure 3 ​As shown. When using the KNN algorithm for data cleaning, for data with wind speed greater than 0 m / s and less than 5 m / s, the cleaning is effective and there is no over-cleaning, but for data with wind speed greater than or equal to 5 m / s, the cleaning effect is poor, and many outliers are not cleaned. When using the box plot algorithm for data cleaning, for data with wind speed greater than 0 m / s and less than 5 m / s, there is a significant over-cleaning problem, and many normal data are identified as abnormal data, but for data with wind speed greater than or equal to 5 m / s, such problems do not occur, and the identification of outliers is more accurate. In summary, it can be seen that although both algorithms are effective for cleaning abnormal data, they both have obvious defects, and if used alone, it will definitely cause deviation in subsequent fault diagnosis model training.

[0059] Therefore, the present application proposes a KNN-box plot combined outlier detection algorithm to solve such problems.

[0060] Further embodiments are directed to a method for detecting outliers in SCADA data, comprising:

[0061] A wind speed threshold is established, with wind speed 0.5 m / s as a wind speed interval, and wind speed 5 m / s as the wind speed threshold.

[0062] For the wind speed interval less than the wind speed threshold 5 m / s, the KNN algorithm (k-nearest neighbor classification algorithm) needs to select a suitable distance measurement method (such as the commonly used Euclidean distance), which is used as a reference to calculate the distance between the target sample and the training sample in the SCADA data. The training samples are sorted in order of distance from small to large, and k nearest neighbor training samples that meet the preset distance requirement are selected, and the number of labels of the k nearest neighbor training samples is counted, and the label with the most occurrences is taken as the outlier detection result of the target sample.

[0063] For the wind speed interval greater than or equal to the wind speed threshold 5 m / s, the box plot algorithm is used to obtain the quartiles and interquartile ranges of the SCADA data, and the outliers of the SCADA data are identified. Specifically, as shown in the figure, the box plot criterion (Box Plot Criterion) is a statistical method for identifying outliers (outliers) in a data set based on quartiles and interquartile ranges of data distribution, where the median represents the average level of the data, the difference between the quartiles reflects the dispersion of the data, and the length of the box represents the concentration of the data. The core data in the box plot are: Figure 1

[0064] Lower quartile Q 1: The middle value of the smaller half of the data set, i.e. The number at the position, n is the length of the data set sequence.

[0065] Median​Q 2: the value of the middle position in the data set, i.e. the number at position n when n is odd, and when n is even.

[0066] when n is odd, and n when n is even, and the mean of the two numbers at position .

[0067] the upper quartile Q 3: the median of the larger half of the data set, i.e. the number at position .

[0068] the interquartile range IQR : the difference between Q 3 and Q 1 IQR = Q . - Q 1

[0069] the upper outlier boundary: expressed as Q 3 + 1.5 IQR .

[0070] the lower outlier boundary: expressed as Q 1 - 1.5 IQR .

[0071] Numbers in the data set that are outside the upper and lower outlier boundaries are considered outliers and should be removed.

[0072] S3: the SSA-PSO-DBN wind turbine fault early warning model is constructed by using the SSA-PSO combined algorithm to optimize the parameters of the DBN.

[0073] Deep Belief Network (DBN) is a commonly used machine learning algorithm at present, which simulates the multi-level algorithm of human brain processing information, is composed of multiple stacked Restricted Boltzmann Machines (RBM), and deeply mines important information and key features contained in the data. During model training, each RBM is first trained through unsupervised training, and then the parameters are fine-tuned through BP neural network in reverse, so that the model obtains the optimal solution. For the SSA-PSO-DBN-based wind turbine fault early warning model proposed in the embodiment, the SSA-PSO combined algorithm is used to optimize the parameters of the DBN, so as to obtain the optimal parameters, and train the DBN model under the optimal parameters, so as to realize more accurate early warning of faults.

[0074] DBN is a multi-level algorithm simulating human brain processing information, which is composed of multiple stacks of restricted Boltzmann machines, deeply mining important information and key features contained in data. Restricted Boltzmann machine contains hidden layer and visible layer structure, and there is no connection between the layers, and the layers are connected bidirectionally. Its network structure is as follows Figure 4 .

[0075] RBM is an energy-based model, when the model reaches the optimum, the energy is minimum. If the visible layer of RBM contains I nodes, and the hidden layer contains J nodes, the energy function of the RBM is formula (1).

[0076] , (1)

[0077] In the formula, represents the parameters of the RBM, represents the connection weight between the i th node in the visible layer and the j th node in the hidden layer; and represent the bias of the i th node in the visible layer and the bias of the j th node in the hidden layer. represents the i th node in the visible layer, represents the j th node in the hidden layer.

[0078] Because the energy function conforms to the Boltzmann distribution, the joint probability distribution of state v , h ) can be obtained from the above formula, as formula (2).

[0079] , (2)

[0080] In the formula, Z ( θ ) represents the normalization factor, that is, the sum of the energy of all states, as formula (3).

[0081] . (3)

[0082] Integrating the visible layer and the hidden layer can obtain the visible layer unit distribution P ( v | θ ) and the hidden layer unit distribution P ( h | θ ), as formula (4) and formula (5).​​

[0083] , (4)

[0084] , (5)

[0085] When the visible layer or hidden layer state is given, the activation function of a certain node in the hidden layer or visible layer, that is, the activation probability of a certain node 、 Can be expressed as formula (6), formula (7).

[0086] , (6)

[0087] , (7)

[0088] For a given sample, training the RBM is to update θ , through the update θ Make the model better fitting training samples, let the RBM can better learn the internal law of training samples.

[0089] The deep belief network DBN is a neural network model stacked by multiple RBMs, and the basic structure of the DBN model used in the application is composed of an input layer, an output layer and four RBMs, and the structure is as shown in Figure 5 The training process of the DBN model mainly includes two parts, network pre-training and backpropagation fine-tuning. The model training process firstly pre-trains the network, inputs the training data processed by the KNN-box chart algorithm into the DBN model, extracts features through the first layer of RBM, and then outputs to the next layer of RBM, and then extracts features layer by layer. After the feature extraction of the four layers of RBM, the initial parameters of the network are generated, and then the backpropagation neural network is used for backpropagation fine-tuning, so that the model parameters reach the optimum.

[0090] In view of the problem that the model hyperparameters in the DBN training are difficult to determine, the application proposes an SSA-PSO-DBN algorithm, which intelligently optimizes the DBN model by using the SSA-PSO algorithm, so as to improve the accuracy of the reconstruction error of the DBN model and improve the accuracy of the fault warning.

[0091] The particle swarm optimization (Particle Swarm Optimization, PSO) is a commonly used DBN optimization algorithm. The basic principle is to randomly generate a batch of particles with only position and speed in the search space. The particles update their positions by the position relationship between other particles and their own speed. The position and speed updating formula of the particles in the algorithm is as shown in formula (8), formula (9).

[0092] , (8)

[0093] (9)

[0094] In the formula v t sd Indicates the first s The particle in the first d Vidi t The speed of time; Representing the s The particle in the first d Vidi t Location at any given moment Represents the dimension. Represents the individual's historical optimal solution; This represents the historical optimal solution of the population; w The inertia weight is used to adjust search capabilities; c 1 and c 2 represents the learning factor and the population learning factor, respectively; r 1 and r 2 is a random number between 0 and 1.

[0095] When using the PSO algorithm for DBN hyperparameter optimization, particle updates depend on the population's historical best solution and the individual's historical best solution. Poor initialization can easily lead to a rapid contraction of the particle search range, trapping the model in local optima and affecting DBN model accuracy. To address this, this embodiment proposes the SSA-PSO algorithm for DBN model optimization. Leveraging the strong global search capability of the SSA algorithm, it improves the particle initialization position of the PSO algorithm, avoiding the local optima problem and thus enhancing the accuracy of the DBN model.

[0096] The Sparrow Search Algorithm (SSA) is a swarm intelligence-based algorithm that mimics the foraging and anti-predation behavior of sparrow populations. The algorithm divides the population into discoverers, joiners, and vigilants. The basic principle of SSA is that discoverers are responsible for finding food-rich areas and guiding the population's movement; joiners follow the discoverers, adjusting their own positions using the discoverers' location information; and when the population detects danger, some individuals quickly relocate—these relocating individuals are the vigilants. The position update formulas for discoverers, joiners, and vigilants are Equations (10), (11), and (12), respectively.

[0097] (10)

[0098] In the formula x s, t d Indicates in t At this moments the first sparrow in the d th position; α is a random number in the range of 0 to 1; itermax represents the maximum number of iterations; Q is a random number following a normal distribution; L is a 1 x D matrix with all elements being 1; R 2 represents the warning value; ST represents the safety value. When R 2 ST , it means that there is no predator around, the environment is safe, and the population can search widely. Otherwise, the population will gather in the safe area and search again.

[0099] , (11)

[0100] where x worst represents the current global worst position; x p represents the best position among the current discoverers; A + = A T ( AA T ) -1 where A is a 1 x D matrix, where all elements are randomly assigned as 1 or -1. When s>S / 2, it means that the s th joiner with poor fitness may have no food.

[0101] , (12)

[0102] where represents the global optimal position at time t. β is the step control coefficient, which is a random number following a normal distribution with a mean of 0 and a variance of 1; K is a random number in the range of [-1, 1]; represents the fitness of the s th sparrow; represents the current global worst fitness; represents the current global optimal fitness; ε represents a very small constant to prevent the denominator from being 0. When , it means that the warning is not in the center of the population and is vulnerable to attack, and needs to move towards the global optimal position. When When the alarm is given, it indicates that the alarm is in the center of the population, the population is aware of the danger and needs to move to other positions.

[0103] Further embodiments are directed to a method for constructing a wind turbine fault early warning model, comprising:

[0104] Initializing the sparrow population and related parameters, classifying each sparrow, and performing DBN training on the classified sparrows to obtain a first DBN training result; wherein the sparrow classification types include discoverer, joiner, and alarm.

[0105] Updating the sparrow position based on the first DBN training result, reclassifying the sparrows, and performing DBN training again until a first preset iteration number is reached to obtain an optimal solution of SSA optimization.

[0106] Generating particles centered on the optimal solution of SSA optimization, performing DBN training on each particle to obtain a second DBN training result, and updating the position and velocity of the particles based on the second DBN training result until a second preset iteration number is reached to obtain an optimal solution of PSO optimization;

[0107] Inputting the optimal solution of PSO optimization into the deep belief network DBN model to complete SSA-PSO joint optimization and obtain a wind turbine fault early warning model.

[0108] Further embodiments are directed to a method for constructing a wind turbine fault early warning model, comprising:

[0109] Calculate the fitness value of each sparrow in the sparrow search algorithm SSA and each particle in the particle swarm optimization algorithm PSO, and use the fitness value as a key indicator to find the optimal solution; wherein the fitness function for calculating the fitness value is as follows:

[0110] , (13)

[0111] In the formula, represents the fitness value, S represents the data length, D represents the dimension, represents the reconstructed data, represents the original data, s and d represent the first s data and the firstd dimension.

[0112] S4: Based on the wind turbine fault early warning model, complete the early warning of wind turbine faults.

[0113] A further implementation method for providing early warning of wind turbine faults based on a wind turbine fault early warning model includes:

[0114] Input the target SCADA data into the wind turbine fault early warning model to generate reconstructed data, and calculate the reconstructed error based on the reconstructed data.

[0115] If the reconstruction error exceeds the preset alarm threshold, the wind turbine is determined to be in an abnormal state; otherwise, it is determined to be in a normal state. In this embodiment, if the reconstruction error exceeds the preset alarm threshold for three consecutive points, the turbine is determined to be in an abnormal state; otherwise, it is in a normal state.

[0116] Specifically, in this embodiment, the reconstruction error is selected as the core criterion for fault warning. This is because when the wind turbine is in normal operation, its reconstruction error will fluctuate within a very small range; however, once the turbine fails, the reconstruction error will show significant anomalies. Based on this, this paper adopts a method of setting alarm limits using a fixed threshold combined with the 3-σ criterion. The implementation of this method requires first calculating the reconstruction error and its mean and root mean square error, as shown in equations (14), (15), and (16).

[0117] (14)

[0118] , (15)

[0119] (16)

[0120] In the formula R e ( s ) indicates the first s One reconstruction error; μ This represents the mean of the reconstruction error; σ The root mean square error represents the reconstruction error; N This represents the total number of reconstruction errors. According to 3- σ The alarm threshold limit set by the criteria is Equation (17).

[0121] , (17)

[0122] In the formula U This indicates the alarm threshold limit. To avoid misjudging the generator set status due to data fluctuations, this embodiment stipulates that when three consecutive points exceed the alarm threshold limit, the generator set status is determined to be abnormal.

[0123] The present embodiment provides an experimental verification process:

[0124] The present embodiment takes a wind farm B08 unit as the research object and the SCADA data source. The unit experienced a high temperature fault of the main shaft floating bearing on November 21, 2023, 15:00. Therefore, 13262 pieces of data between August 21, 2023, 0:00 and November 21, 2023, 15:00 before the unit failed are used for research, and the training data set, the normal state test data set and the abnormal state test data set are divided according to the ratio of 6:1:3. For the input variables of the model, after correlation analysis, 10 variables such as wind speed, active power of power quality module monitoring, engine speed, etc. are selected as the input variables of the model.

[0125] Before training SSA-PSO-DBN using wind turbine SCADA data, the collected SCADA data needs to be preprocessed and the outliers need to be cleaned. In this paper, the KNN-boxplot algorithm proposed in Chapter 2 is used to clean the SCADA data, and the data is binned according to wind speed every 0.5 m / s. The data before cleaning is shown in Figure 7 After binning the data, the KNN-boxplot combination algorithm is used to clean the data in different bins. The results are shown in Figure 8 From Figure 8 , it can be seen that the KNN-boxplot combination algorithm can effectively clean the outliers in the wind turbine SCADA data. The outliers are basically eliminated and the normal data are not mis-cleaned. The cleaning result is ideal.

[0126] After cleaning the wind turbine SCADA data, the parameters related to the wind turbine generator are selected and input into the SSA-PSO-DBN model proposed in this embodiment for model training. This embodiment selects wind speed, generator speed, active power of power quality module monitoring, generator winding temperature, etc. as input variables for model training and fault warning.

[0127] The parameter settings of the SSA-PSO-DBN model established in this embodiment are as follows: the DBN model sets the number of hidden layers to 4, and the optimization interval of the number of internal nodes is [1, 150]; the SSA algorithm sets the population size to 20, the maximum number of iterations to 30, the discoverer ratio to 30%, the joiner ratio to 50%, and the guard ratio to 20%; the PSO algorithm sets the population size to 30, the maximum number of iterations to 30, the inertia weight to 0.729, and the group learning factor and individual learning factor to 2.05.

[0128] The DBN model parameters after optimization using SSA-PSO are: the first hidden layer node number is 128, the second hidden layer node number is 73, the third hidden layer node number is 63, and the fourth hidden layer node number is 129. The RMSE and MAE evaluation indexes of the DBN model trained using the parameters obtained by SSA-PSO optimization are 0.0464 and 0.0296, respectively, which are very close to 0, indicating that the performance of the DBN model after optimization by SSA-PSO algorithm has been effectively improved. The two indexes show that the SSA-PSO-DBN model has strong generalization ability in processing high-dimensional and nonlinear data.

[0129] The 5305 data points in the normal state test set and the abnormal state test set are input into the SSA-PSO-DBN model and the PSO-DBN model, and the reconstruction error is output. The reconstruction error amplitude output by the two models is observed, as shown in Figure 9 .

[0130] As can be seen from Figure 9 , the reconstruction error amplitude of the SSA-PSO-DBN model is smaller than that of the PSO-DBN model, indicating that the SSA-PSO-DBN model has better nonlinear fitting ability and higher reconstruction accuracy, and can reduce the possibility of misjudgment and make more accurate early warning.

[0131] According to the selected wind turbine SCADA data and the normal state test data set, the data in this data set are all wind turbine normal state data with 1326 data points. The reconstruction error output by the model is shown in Figure 10 .

[0132] As can be seen from Figure 10 , the reconstruction error of the training set is basically below the alarm threshold limit, and most of them fluctuate below 0.05, and only a few exceed 0.15, and only one data point exceeds the alarm threshold limit.

[0133] The abnormal state test data set data is 30% of the selected SCADA data, with 3979 data points. This data set contains data within a period of time before failure, so the results of the data set can test whether the model can correctly warn of failure. The reconstruction error is shown in Figure 11 .

[0134] The alarm threshold is exceeded, which meets the wind turbine state abnormality determination standard. In the interval of 2500-3000, 2878-2880 consecutive three data points exceed the alarm threshold limit, which is the second time to meet the wind turbine state abnormality determination standard; 2957-2960 consecutive four data points exceed the alarm threshold limit, which is the third time to meet the wind turbine state abnormality determination standard.

[0135] According to the SCADA data, on October 30, 2023, the reconstruction error meets the wind turbine state abnormality determination standard, at this time, there are still 22 days to the fault occurrence, on November 14 and 15, 2023, the reconstruction error meets the wind turbine state abnormality determination standard twice, at this time, there are still 6 days to the fault occurrence. If the on-board inspection and maintenance are carried out after the first alarm, the loss caused by the fault shutdown can be completely avoided. Even if no fault signal is found at the first time, there is still sufficient time for on-board inspection and maintenance after the second and third alarms to avoid the occurrence of the fault.

[0136] In summary, through the above experimental verification, it can be proved that the SSA-PSO-DBN algorithm proposed in the present application has high accuracy for wind turbine operation state judgment, can further reduce the risk of misjudgment, and can realize early warning of wind turbine fault.

[0137] The KNN-box plot data cleaning algorithm and the wind turbine fault early warning algorithm based on SSA-PSO-DBN proposed in the present application. The KNN-box plot algorithm is combined for the advantages and disadvantages of the two data cleaning algorithms, the wind turbine SCADA data is binned according to the wind speed, and the data cleaning is carried out in the respective binning intervals of the KNN algorithm and the box plot algorithm. The SSA-PSO-DBN algorithm uses SSA-PSO to optimize the parameters of DBN, and uses the optimization results to train the DBN model. The algorithm improves the shortcomings of the single PSO algorithm in global optimization, avoids the problem that the PSO algorithm is easy to fall into local optimum, and realizes high-precision early warning of wind turbine fault. Through experimental verification, the KNN-box plot data cleaning algorithm can effectively clean the abnormal data, while maximizing the retention of normal data, and the SSA-PSO-DBN algorithm can detect the large fluctuation of the reconstruction error early, realize early warning of the wind turbine, and leave sufficient time for fault maintenance.

[0138] Example two:

[0139] The present application also provides a wind turbine fault early warning system based on an optimized deep belief network, which is used to realize the method of example one, comprising:

[0140] A data acquisition module is used to acquire SCADA data based on a data acquisition and monitoring control SCADA system installed in a wind farm, wherein the SCADA data includes wind farm surrounding environment parameters and wind turbine component performance parameters.

[0141] A data cleaning module is used to detect abnormal values of the SCADA data by using a KNN-box plot combined algorithm, and obtain cleaned SCADA data.

[0142] The model construction module is configured to jointly use a sparrow search algorithm (SSA) and a particle swarm optimization algorithm (PSO) to perform parameter optimization on a deep belief network (DBN) to construct a wind turbine fault early warning model based on SSA-PSO-DBN.

[0143] The fault early warning module is configured to perform early warning on the wind turbine fault based on the wind turbine fault early warning model.

[0144] In a further implementation, the data cleaning module includes:

[0145] The KNN cleaning unit is configured to set a wind speed threshold, for a wind speed interval less than the wind speed threshold, use a KNN algorithm to calculate distances between a target sample and training samples in SCADA data, sort the training samples in ascending order of distance, filter out k nearest neighbor training samples that meet a preset distance requirement, and count the number of labels of the k nearest neighbor training samples, and take the label with the highest occurrence frequency as an anomaly value detection result of the target sample.

[0146] The box plot cleaning unit is configured to, for a wind speed interval greater than or equal to the wind speed threshold, use a box plot algorithm to obtain quartiles and interquartile ranges of SCADA data, and identify anomalies in the SCADA data.

[0147] In a further implementation, the model construction module includes:

[0148] The SSA initialization unit is configured to initialize sparrow populations and related parameters, classify each sparrow, and perform DBN training on the classified sparrows to obtain a first DBN training result; wherein the types of sparrow classification include discoverer, joiner, and alarm.

[0149] The SSA optimization unit is configured to update the positions of the sparrows based on the first DBN training result, reclassify the sparrows, and perform DBN training again until a first preset number of iterations is reached to obtain an optimal solution of SSA optimization.

[0150] The PSO optimization unit is configured to generate particles with the optimal solution of SSA optimization as the center, perform DBN training on each particle to obtain a second DBN training result, and update the positions and velocities of the particles based on the second DBN training result until a second preset number of iterations is reached to obtain an optimal solution of PSO optimization.

[0151] The model construction unit is configured to input the optimal solution of PSO optimization into a deep belief network (DBN) model to perform SSA-PSO joint optimization and obtain a wind turbine fault early warning model.

[0152] The above described embodiments are only to illustrate the preferred modes of the present application, and are not intended to limit the scope of the present application. Any modification and improvement made by those skilled in the art to the technical solutions of the present application without departing from the design spirit of the present application shall fall within the protection scope of the present application as defined by the claims.

Claims

1. A wind turbine fault early warning method based on optimized deep belief networks, characterized in that, include: The data acquisition and monitoring control SCADA system installed in the wind farm acquires SCADA data, which includes environmental parameters around the wind farm and performance parameters of wind turbine components. The KNN-box plot combination algorithm is used to detect outliers in the SCADA data to obtain cleaned SCADA data. The parameters of the Deep Belief Network (DBN) are optimized by combining the Sparrow Search Algorithm (SSA) and the Particle Swarm Optimization Algorithm (PSO), and a wind turbine fault early warning model based on SSA-PSO-DBN is constructed. Based on the aforementioned wind turbine fault early warning model, early warning of wind turbine faults is achieved. The method for outlier detection of the SCADA data includes: A wind speed threshold is established. For wind speed ranges less than the threshold, the KNN algorithm is used to calculate the distance between the target sample and the training sample in the SCADA data. The training samples are sorted in ascending order of distance. The k nearest neighbor training samples that meet the preset distance requirements are selected. The number of labels for each of the k nearest neighbor training samples is counted. The label that appears most frequently is taken as the outlier detection result of the target sample. For wind speed ranges greater than or equal to the wind speed threshold, the box plot algorithm is used to obtain the interquartiles and interquartile ranges of the SCADA data and identify outliers in the SCADA data. Methods for constructing wind turbine fault early warning models include: Initialize the sparrow population and related parameters, classify each sparrow, and perform DBN training on the classified sparrows to obtain the first DBN training result; among them, the sparrow classification types include discoverer, joiner, and vigilant. The sparrow's position is updated based on the first DBN training result, and the sparrow is reclassified and DBN training is performed again until the first preset number of iterations is reached to obtain the optimal solution for SSA optimization. Particles are generated centered on the optimal solution of SSA optimization. DBN training is performed on each particle to obtain the second DBN training result. The position and velocity of the particles are updated according to the second DBN training result until the second preset number of iterations is reached to obtain the optimal solution of PSO optimization. The optimal solution obtained from PSO optimization is input into the Deep Belief Network (DBN) model to complete the joint optimization of SSA-PSO and obtain the wind turbine fault early warning model.

2. The method according to claim 1, characterized in that, The method for obtaining the optimal solution during the joint optimization process of SSA-PSO is as follows: The fitness value of each sparrow in the Sparrow Search Algorithm (SSA) and each particle in the Particle Swarm Optimization (PSO) algorithm is calculated separately, and the fitness value is used as a key indicator to find the optimal solution; the fitness function used to calculate the fitness value is as follows: ， In the formula, Represents fitness value, S Represents the length of the data. D Represents dimensions, This indicates data reconstruction. Represents the original data. s and d They represent the first s The data and the first d dimension.

3. The method according to claim 1, characterized in that, The method for issuing early warnings of wind turbine faults based on the aforementioned wind turbine fault early warning model includes: The target SCADA data is input into the wind turbine fault early warning model to generate reconstructed data, and the reconstructed error is calculated based on the reconstructed data. If the reconstruction error exceeds the preset alarm threshold, the wind turbine generator is determined to be in an abnormal state; otherwise, the wind turbine generator is determined to be in a normal state.

4. A wind turbine fault early warning system based on an optimized deep belief network, used to implement the method described in any one of claims 1-3, characterized in that, include: The data acquisition module is used to acquire SCADA data based on the data acquisition and monitoring control SCADA system installed in the wind farm; wherein, the SCADA data includes environmental parameters around the wind farm and performance parameters of wind turbine components; The data cleaning module is used to perform outlier detection on the SCADA data using the KNN-box plot combination algorithm to obtain cleaned SCADA data. The model building module is used to combine the Sparrow Search Algorithm (SSA) and the Particle Swarm Optimization Algorithm (PSO) to optimize the parameters of the Deep Belief Network (DBN) and build a wind turbine fault early warning model based on SSA-PSO-DBN. The fault early warning module is used to provide early warning of wind turbine faults based on the wind turbine fault early warning model.

5. The system according to claim 4, characterized in that, The data cleaning module includes: The KNN cleaning unit is used to establish a wind speed threshold. For wind speed ranges less than the threshold, the KNN algorithm is used to calculate the distance between the target sample and the training sample in the SCADA data. The training samples are sorted in ascending order of distance, and the k nearest neighbor training samples that meet the preset distance requirements are selected. The number of labels of each of the k nearest neighbor training samples is counted, and the label that appears most frequently is taken as the outlier detection result of the target sample. The box plot cleaning unit is used to obtain the interquartile range and interquartile range of SCADA data for wind speed ranges greater than or equal to the wind speed threshold, and to identify outliers in the SCADA data.

6. The system according to claim 4, characterized in that, The model building module includes: The SSA initialization unit is used to initialize the sparrow population and related parameters, classify each sparrow, and perform DBN training on the classified sparrows to obtain the first DBN training result; among them, the sparrow classification types include discoverers, joiners, and vigilants; The SSA optimization unit is used to update the sparrow's position based on the first DBN training result, and to reclassify the sparrow and perform DBN training again until the first preset number of iterations is reached to obtain the optimal solution of SSA optimization. The PSO optimization unit is used to generate particles centered on the optimal solution of SSA optimization, perform DBN training on each particle to obtain the second DBN training result, and update the position and velocity of the particles according to the second DBN training result until the second preset number of iterations is reached to obtain the optimal solution of PSO optimization. The model building unit is used to input the optimal solution of PSO optimization into the deep belief network (DBN) model to complete the joint optimization of SSA-PSO and obtain the wind turbine fault early warning model.

Images