A wind turbine fault diagnosis method based on SSA-BLS and related equipment

Abstract

The application discloses a wind turbine fault diagnosis method based on SSA-BLS and related equipment, and relates to the technical field of wind power generation and intelligent operation and maintenance. The method comprises the following steps: acquiring operation data collected by a wind turbine SCADA system as an original input matrix; after performing missing value elimination, repeated value elimination and cluster outlier elimination on the original input matrix, an input matrix after cleaning is obtained; dimensionality reduction is performed on the input matrix after cleaning as input of a principal component analysis method, and a low-dimensional feature matrix is obtained; the low-dimensional feature matrix is taken as an input sample set, and a test sample subset divided from the input sample set is input into a trained width learning system network; the width learning system of incremental learning is perfectly coupled with the high cluster intelligence sparrow search algorithm, the best super parameter combination of the BLS is adaptively locked through the SSA algorithm under the premise of maintaining the extremely fast training characteristics of the width learning, and the training underfitting or overfitting phenomenon caused by artificial blind parameter adjustment of a traditional algorithm is avoided.

Description

Technical Field

[0001] This invention relates to a wind turbine fault diagnosis method and related equipment based on SSA-BLS, belonging to the field of wind power generation and intelligent operation and maintenance technology. Background Technology

[0002] With the rapid growth of global demand for clean energy, wind power installed capacity continues to climb. However, to obtain high-quality and stable wind energy resources, wind farms are mostly located in remote areas or offshore, facing harsh operating environments such as high wind speeds, high humidity, large alternating loads, and strong corrosion. Under these conditions of overlapping overload and extreme weather, wind turbines are highly susceptible to fatigue damage and operational failures in mechanical and electrical components. Once a major failure leads to a shutdown, due to inconvenient transportation and dispersed turbine layout, maintenance personnel often find it difficult to respond in a timely manner. This not only results in long-term wind curtailment losses but also drives up operation and maintenance costs, severely restricting the economic benefits of wind farms.

[0003] Currently, newly installed wind turbines are generally equipped with Supervisory Control and Data Acquisition (SCADA) systems, accumulating massive amounts of turbine operating status parameters, such as temperature, power, speed, and voltage. Data-driven machine learning algorithms have become a core focus of current research on wind turbine fault diagnosis. However, existing technologies still have the following significant shortcomings in practical applications:

[0004] 1) Traditional deep learning algorithms, such as Deep Belief Networks (DBN) and Deep Convolutional Neural Networks, have complex network structures and large parameter scales. When processing massive amounts of SCADA data, the training time is too long, making it difficult to meet the needs of real-time status monitoring and agile operation and maintenance.

[0005] 2) Conventional shallow machine learning models, such as backpropagation neural networks and support vector machines (SVMs), have limited generalization capabilities and are prone to getting trapped in local optima when dealing with large-sample, multi-fault classification tasks, resulting in insufficient diagnostic accuracy. Furthermore, the parameter configurations of these models, such as the number of feature points and regularization parameters, largely rely on expert experience or manual grid search, lacking adaptive optimization mechanisms. Summary of the Invention

[0006] The purpose of this invention is to provide a wind turbine fault diagnosis method and related equipment based on SSA-BLS. By using the sparrow search algorithm to globally optimize the topology and parameters to be optimized in the width learning system, the accuracy and generalization performance of multi-fault category diagnosis are significantly improved while maintaining the second-level training speed of the width learning system.

[0007] To achieve the above objectives, the present invention is implemented using the following technical solution.

[0008] On the one hand, this invention provides a wind turbine fault diagnosis method based on SSA-BLS, including:

[0009] The operating data collected by the wind turbine SCADA system is used as the raw input matrix;

[0010] After performing missing value removal, duplicate value removal, and cluster outlier removal on the original input matrix, a cleaned input matrix is ​​obtained;

[0011] The cleaned input matrix is ​​used as the input for principal component analysis to reduce dimensionality and obtain low-dimensional eigenmotes.

[0012] Using the low-dimensional feature matrix as the input sample set, the test sample subset divided from the input sample set is input into the trained width learning system network. The width learning system network calculates the classification output matrix according to the pre-calculated output layer weight matrix, determines the fault category of each test sample according to the classification output matrix, and outputs the diagnosis result.

[0013] The trained width learning system network is obtained in the following way:

[0014] A wide learning system network is constructed, wherein the wide learning system network is a state matrix formed by horizontally concatenating a feature node matrix and an enhancement node matrix, and the output layer weight matrix is ​​calculated using the state matrix.

[0015] Using the input sample set as the basis for calculating the fitness function, the sparrow search algorithm is used to globally optimize the parameters to be optimized in the wide learning system network, and the optimal parameter combination is output.

[0016] Configure the topology of the width learning system network with the optimal parameter combination;

[0017] The training sample subset divided from the input sample set is used as the input of the width learning system network, and the output layer weight matrix is ​​solved to obtain the trained width learning system network.

[0018] Optionally, the parameters to be optimized include the number of feature points, the number of augmentation points, the number of feature node neurons, the number of augmentation node neurons, and the regularization coefficient.

[0019] The parameters to be optimized are defined as five specific variables: the number of feature points, the number of augmentation points, the number of neurons in feature nodes, the number of neurons in augmentation nodes, and the regularization coefficient. This ensures that the five-dimensional optimization space of the Sparrow Search algorithm is precisely mapped to the BLS network structure, guaranteeing that the optimization results directly affect the core topological dimension of the diagnostic network and improving the targeting of the optimization.

[0020] Optionally, the outlier removal in clustering is performed by using the K-means clustering algorithm to divide the running data into clusters, identifying edge noise points that deviate from the distribution range of the normal operating power curve, and removing them.

[0021] The K-means clustering algorithm is used to identify and remove edge noise in SCADA operation data. This can effectively separate abnormal data caused by non-fault factors such as power rationing and sudden wind speed changes, and prevent these interfering samples from participating in subsequent model training. This improves the purity of the training sample set and provides data quality assurance for the accuracy of the diagnostic model.

[0022] Optionally, the width learning system network is constructed as follows: a feature node matrix is ​​generated from the input sample set; an enhancement node matrix is ​​generated from the feature node matrix; and the feature node matrix and the enhancement node matrix are horizontally concatenated to form a state matrix, wherein the number of rows of the state matrix is ​​the number of samples in the training sample subset, and the number of columns of the state matrix is ​​the sum of the number of columns of the feature node matrix and the enhancement node matrix.

[0023] Optionally, the output layer weight matrix of the width learning system is calculated using the ridge regression algorithm for pseudo-inverse calculation, and the calculation formula is as follows:

[0024] In the formula, Let I be the output layer weight matrix, λ be the regularization coefficient, I be the identity matrix, A be the state matrix formed by horizontally concatenating the feature node matrix and the booster node matrix, and Y be the desired output matrix. It is the transpose of the state matrix.

[0025] The output layer weight matrix is ​​obtained by performing pseudo-inverse calculation using the ridge regression algorithm. A regularization coefficient is introduced to regularize the condition number of the state matrix, which effectively suppresses the overfitting problem caused by pseudo-inverse calculation due to the non-square or ill-conditioned state matrix, and improves the generalization performance of the model on the test set.

[0026] Optionally, in the sparrow search algorithm, the sparrow population is divided into producers and beggars according to a preset ratio, and scouts are randomly assigned; the producers determine the foraging direction and update their own spatial position according to the fitness value; the beggars follow the producers with the highest fitness value to update their own spatial position; the scouts trigger anti-predation warnings during the population spatial position update process and randomly update their own spatial position.

[0027] The sparrow population is divided into producers, beggars, and scouts. Producers are responsible for global search and exploration of new areas, beggars are responsible for local refinement and development in high-quality areas, and scouts are responsible for triggering anti-predation warnings to escape local extremes. The collaboration of the three enables the population to maintain a dynamic balance between global exploration and local development, thereby improving the convergence speed and solution quality of the optimization process.

[0028] Optionally, the sparrow search algorithm uses the classification accuracy on the verification sample set as the fitness value and aims to maximize the classification accuracy as the convergence objective; in response to the number of iterations reaching a preset maximum value or the fitness value no longer increasing, it is determined that the convergence condition is met.

[0029] Using the classification accuracy of the validation set as the fitness value and maximizing this value as the convergence objective, the optimization process is directly aligned with the actual evaluation index of the diagnostic task, eliminating the deviation between the proxy objective and the actual performance; the explicit iteration termination condition makes the optimization process operable and controllable.

[0030] Secondly, the present invention provides a wind turbine fault diagnosis system based on SSA-BLS, comprising:

[0031] The data acquisition interface is used to connect to the wind turbine's SCADA system and acquire the unit's operating data.

[0032] The data preprocessing module, whose input end is connected to the output end of the data acquisition interface, is used to perform missing value removal, duplicate value removal, and cluster outlier removal on the original input matrix to obtain a cleaned input matrix; and uses the cleaned input matrix as input for principal component analysis to perform dimensionality reduction to obtain a low-dimensional feature matrix.

[0033] The diagnostic module, whose input is connected to the output of the data preprocessing module, is used to take the low-dimensional feature matrix as the input sample set, input the test sample subset divided from the input sample set into the trained width learning system network, the width learning system network calculates the classification output matrix according to the pre-calculated output layer weight matrix, determines the fault category of each test sample according to the classification output matrix and outputs the diagnostic result.

[0034] A network construction module is used to construct a wide learning system network, wherein the wide learning system network is composed of a state matrix formed by horizontally concatenating a feature node matrix and an enhancement node matrix, and the output layer weight matrix is ​​calculated using the state matrix.

[0035] The parameter optimization module has its input terminals connected to the output terminals of the data preprocessing module and the network construction module, respectively. It is used to perform global optimization of the parameters to be optimized in the wide learning system network using the sparrow search algorithm, based on the input sample set as the basis for calculating the fitness function, and output the optimal parameter combination.

[0036] The network construction and training module has its input terminals connected to the output terminals of the data preprocessing module and the parameter optimization module, respectively. It is used to configure the topology of the width learning system network with the optimal parameter combination; it uses a subset of training samples divided from the input sample set as the input of the width learning system network, solves the output layer weight matrix, obtains the trained width learning system network, and outputs it to the diagnostic module.

[0037] The diagnostic display terminal has its input end connected to the output end of the diagnostic module, and is used to receive and display the diagnostic results.

[0038] By organically combining a data acquisition interface, a data preprocessing module, a parameter optimization module, a network construction module, a diagnostic module, and a diagnostic display terminal, a fully automated closed-loop chain is formed from the access of raw SCADA data to the output of fault diagnosis results. The data flow relationships between each module correspond one-to-one with the aforementioned method steps, ensuring that when the system is deployed in the actual operation and maintenance environment of a wind farm, it can complete the entire process of diagnosis from data acquisition to alarm output with extremely low end-to-end latency, without the need for manual intervention in hyperparameter configuration or model tuning, significantly improving the automation level and fault response speed of wind farm centralized control operation and maintenance.

[0039] Optionally, the network construction module includes a feature node generation unit, an enhancement node generation unit, and a state matrix concatenation unit; the feature node generation unit is used to randomly map the low-dimensional feature data into a feature node matrix, the enhancement node generation unit is used to nonlinearly transform the feature node matrix into an enhancement node matrix, and the state matrix concatenation unit is used to horizontally concatenate the feature node matrix and the enhancement node matrix into the state matrix.

[0040] The network construction module is further refined into a feature node generation unit, an enhancement node generation unit, and a state matrix splicing unit. The three units perform three independent functions: feature mapping, enhancement transformation, and lateral splicing, respectively. This achieves modular decoupling of the BLS network construction process at the system architecture level, which facilitates subsequent system maintenance and upgrades.

[0041] Thirdly, the present invention provides a wind turbine fault diagnosis device, comprising:

[0042] Memory, used to store computer programs;

[0043] A processor is configured to execute the computer program to implement the steps of the SSA-BLS-based wind turbine fault diagnosis method described in any of the preceding claims.

[0044] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: This invention uses a wide learning system as the core diagnostic network. It generates a low-dimensional feature matrix through random mapping, followed by a nonlinear transformation to generate a topological structure, replacing the traditional deep learning training paradigm that relies on multi-layer backpropagation iterative optimization. This structure allows the output layer weight matrix to be solved in one step through pseudo-inverse calculation, fundamentally eliminating the iterative overhead caused by gradient chain propagation. Furthermore, this invention introduces a sparrow search algorithm to globally optimize the number of feature points, augmentation points, number of neurons, and regularization coefficients of the wide learning system, overcoming the model performance instability problem caused by traditional BLS relying on manual experience or grid search to set hyperparameters. Therefore, this invention compresses model training time to the second level or even less while ensuring the global optimality of hyperparameter combinations, thus simultaneously achieving high real-time performance and high accuracy in fault diagnosis. Attached Figure Description

[0045] Figure 1 The diagram shown is a flowchart of a wind turbine fault diagnosis method based on SSA-BLS in some embodiments.

[0046] Figure 2 The diagram shown is a schematic of the unit status classification results of a wind turbine fault diagnosis method based on SSA-BLS in some embodiments.

[0047] Figure 3 The diagram shown is a schematic of the confusion matrix of a diagnostic model for a wind turbine fault diagnosis method based on SSA-BLS in some embodiments.

[0048] Figure 4 The diagram shows a comparison of actual and predicted classifications before and after the SSA-BLS optimization model of a wind turbine fault diagnosis method based on SSA-BLS in some embodiments. Detailed Implementation

[0049] It should be noted that this invention can be deployed on computing devices such as servers, edge computing nodes, or industrial control computers at the base of wind turbine towers in the wind farm's central control center. It can access the wind farm's SCADA system network via wired or wireless communication to obtain real-time operation monitoring data of each wind turbine unit.

[0050] The term "operational data" refers to the numerical values ​​of various monitored variables recorded by the wind turbine's SCADA system, including but not limited to parameters such as wind speed, blade speed, power, bearing temperature of various components, grid-side voltage, and excitation current. The term "parameters to be optimized" refers to the hyperparameters that need to be pre-set during the construction of the wide-range learning system; their values ​​affect the scale and generalization performance of the network structure. The term "optimal parameter combination" refers to the set of hyperparameters that maximizes the fitness function value obtained through iterative optimization using the sparrow search algorithm.

[0051] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0052] Example 1

[0053] like Figures 1-4 As shown in the figure, this embodiment introduces a wind turbine fault diagnosis method based on SSA-BLS, including:

[0054] The operating data collected by the wind turbine SCADA system is used as the raw input matrix. The rows of this raw input matrix correspond to different sampling times, and the columns correspond to the monitoring variables recorded by the SCADA system. The monitoring signals collected by the SCADA system contain at least 37 variables, which involve key physical parameters of the generator, converter, and pitch system.

[0055] After performing missing value removal, duplicate value removal, and cluster outlier removal on the original input matrix, a cleaned input matrix is ​​obtained. Missing value removal refers to deleting data rows containing null or invalid values ​​from the original input matrix. Duplicate value removal refers to deleting identical data rows in the original input matrix caused by duplicate records or data retransmissions in the system. Cluster outlier removal refers to using cluster analysis methods on the preprocessed data to identify and delete outliers that deviate from the edges of normal clusters.

[0056] The cleaned input matrix is ​​used as input for principal component analysis (PCA) to reduce dimensionality, resulting in a low-dimensional feature matrix. Specifically, the covariance matrix among the monitoring variables in the cleaned input matrix is ​​calculated, and the eigenvalues ​​and eigenvectors of this covariance matrix are solved. The principal components are arranged in descending order of eigenvalues, and the variance contribution rate of each principal component is calculated. Based on the cumulative variance contribution rate, the top few principal components are selected, and the cleaned input matrix is ​​projected onto the selected principal component directions to obtain the dimensionality-reduced low-dimensional feature matrix. When using PCA for feature dimensionality reduction, the original 37-dimensional or 26-dimensional SCADA feature variables are reduced to 9-dimensional principal component eigenvectors, retaining more than 95.8% of the information in the original dataset.

[0057] Using the low-dimensional feature matrix as the input sample set, a subset of test samples from the input sample set is input into the trained width learning system network. The width learning system network calculates a classification output matrix based on the pre-calculated output layer weight matrix, determines the fault category of each test sample based on the classification output matrix, and outputs a diagnostic result. Each row of the classification output matrix corresponds to a test sample, each column corresponds to a fault category, and each element value represents the confidence level of the test sample belonging to the corresponding category. For each test sample, the fault category corresponding to the largest element value in that row of the classification output matrix is ​​selected as the diagnostic result output for that sample.

[0058] The trained width learning system network is obtained in the following way:

[0059] A wide learning system network is constructed, wherein the wide learning system network is a state matrix formed by horizontally concatenating a feature node matrix and an enhancement node matrix, and the output layer weight matrix is ​​calculated using the state matrix.

[0060] Using the input sample set as the basis for calculating the fitness function, the sparrow search algorithm is employed to globally optimize the parameters of the width learning system network, outputting the optimal parameter combination. Specifically, a sparrow population is initialized, with each sparrow's spatial position vector containing a set of parameters to be optimized. The input sample set is divided into a training subset and a validation subset. A width learning system is constructed using the parameter values ​​corresponding to each sparrow. The network is trained using the training subset, and the classification accuracy is calculated using the validation subset. This classification accuracy is used as the fitness value of the sparrow. Based on the fitness values ​​of each sparrow in the population, the spatial position of the population is iteratively updated according to the position update rules of the sparrow search algorithm until the convergence condition is met. The spatial position vector of the sparrow with the largest fitness value at convergence is output as the optimal parameter combination.

[0061] Configure the topology of the width learning system network with the optimal parameter combination; determine the feature node size, augmentation node size and number of neurons in each layer of the width learning system according to the values ​​of the number of feature points, number of augmentation points, number of feature node neurons, number of augmentation node neurons and regularization coefficient in the optimal parameter combination.

[0062] The training sample subset, derived from the input sample set, is used as the input to the width learning system network. The output layer weight matrix is ​​then solved to obtain the trained width learning system network. The input sample set is divided into a training sample subset and a test sample subset according to a preset ratio. The training sample subset is transformed by the feature mapping layer and enhancement layer of the width learning system to form a state matrix. Based on this state matrix and the expected output matrix corresponding to the training sample subset, the output layer weight matrix is ​​solved using a pseudo-inverse algorithm to complete the network training.

[0063] When the SCADA system collects new real-time operating data, it processes the data according to the same data cleaning and principal component analysis dimensionality reduction process described above, and then inputs it into the trained width learning system network to output the current unit status classification and fault diagnosis results in real time.

[0064] In summary, the wind turbine fault diagnosis method based on SSA-BLS in this embodiment ensures training efficiency through the incremental learning characteristics of the width learning system, while ensuring diagnostic accuracy through the global optimization capability of the sparrow search algorithm, thus solving the technical problem of difficulty in balancing training speed and diagnostic accuracy in the prior art.

[0065] In some embodiments, the parameters to be optimized include the number of feature points, the number of augmentation points, the number of feature node neurons, the number of augmentation node neurons, and the regularization coefficient. It can be understood that the target parameters for optimization are the number of feature points n, the number of augmentation points m, the number of feature node neurons p, the number of augmentation node neurons q, and the regularization coefficient λ of the width learning system (BLS).

[0066] In some embodiments, the outlier removal in clustering involves: using a K-means clustering algorithm to cluster the operational data, identifying and removing edge noise points that deviate from the normal operating power curve distribution range. This can be understood as follows: A two-dimensional scatter space is constructed using two variables in the operational data: active power and wind speed. The operational data, cleaned of missing and duplicate values, is used as input to the K-means clustering algorithm. The preset number of clusters is K. The K-means clustering algorithm iteratively optimizes the sum of the squares of the distances from each sample point to its cluster center, dividing the data into K clusters. Under normal operating conditions, the active power of the wind turbine follows a specific power curve distribution with wind speed. Most data points are concentrated near this power curve. However, abnormal data points generated due to power curtailment, sudden wind speed changes, or sensor malfunctions show a significant deviation between the active power and wind speed relationship from the normal power curve. In the two-dimensional scatter space, these appear as edge noise points far from the centers of normal clusters. By setting a distance threshold, sample points exceeding this threshold are identified as outliers and removed.

[0067] In some embodiments, constructing the width learning system network involves: generating a feature node matrix from the input sample set; generating an augmentation node matrix from the feature node matrix; and horizontally concatenating the feature node matrix and the augmentation node matrix to form a state matrix, wherein the number of rows in the state matrix is ​​equal to the number of samples in the training sample subset, and the number of columns in the state matrix is ​​the sum of the number of columns in the feature node matrix and the augmentation node matrix. Optionally, the output layer weight matrix of the width learning system is calculated using a ridge regression algorithm for pseudo-inverse, with the following formula:

[0068] In the formula, Let I be the output layer weight matrix, λ be the regularization coefficient, I be the identity matrix, A be the state matrix formed by horizontally concatenating the feature node matrix and the booster node matrix, and Y be the desired output matrix. It is the transpose of the state matrix.

[0069] It can be understood that, assuming the input sample set is X, based on the number of feature points n and the number of neurons in the feature nodes in the optimal parameter combination, the input sample set X is processed through n sets of random mapping functions to generate n sets of feature nodes. Each set of feature nodes contains p neurons. The feature node matrix is ​​obtained by horizontally concatenating the feature nodes of each set. Based on the number of augmentation points m and the number of augmentation node neurons q in the optimal parameter combination, the feature node matrix is... m sets of enhancement nodes are generated by mapping m sets of nonlinear activation functions. Each set of enhancement nodes contains q neurons. The enhancement node matrix is ​​obtained by horizontally concatenating the enhancement nodes from each set. The feature node matrix and enhanced node matrix The components are horizontally joined and assembled into the system's state matrix. Based on this, according to the state matrix A, the expected output matrix Y corresponding to the training sample subset, and the regularization coefficient λ, the pseudo-inverse of the ridge regression algorithm is used to calculate the output weight matrix. The calculation formula is as follows:

[0070]

[0071] Where λ is the pseudo-inverse control regularization coefficient to prevent model overfitting.

[0072] In some embodiments, the sparrow search algorithm divides the sparrow population into producers and beggars according to a preset ratio, and randomly assigns scouts. Producers determine their foraging direction and update their spatial position based on their fitness values. Beggars follow the producer with the highest fitness value to update their spatial position. Scouts trigger anti-predation warnings during the population's spatial position update process and randomly update their spatial position. Optionally, the sparrow search algorithm uses the classification accuracy on the verification sample set as the fitness value and aims to maximize the classification accuracy as the convergence objective. Convergence is determined when the number of iterations reaches a preset maximum value or the fitness value no longer increases. It can be understood that the sparrow search algorithm divides the sparrow population into producers and beggars according to a preset ratio and randomly assigns scouts. Individuals with higher fitness values ​​act as producers, responsible for exploring potential optimal areas and providing foraging directions in a large search space; the remaining individuals act as beggars, following the producer with the highest fitness value to update their spatial position, and performing localized refined searches near high-quality areas. Meanwhile, a certain proportion of individuals are randomly selected from the population to act as scouts. During the population's spatial location update process, the scouts trigger anti-predation warnings and randomly update their own spatial locations to prevent the population as a whole from falling into a local optimum.

[0073] In the optimization process, the Sparrow Search Algorithm (SSA) represents the total population position of sparrows as follows:

[0074] Where M is the total number of virtual sparrows and N is the dimension of the optimization parameters; by iteratively updating the spatial positions of producers and beggars, the optimization is performed to maximize the classification accuracy as the fitness function.

[0075] Specifically, the sparrow search algorithm uses the classification accuracy on the validation sample set as the fitness value and aims to maximize this accuracy as the convergence objective. In each iteration, a width learning system is constructed using the parameter values ​​corresponding to the spatial position vectors of each sparrow individual. After training the network with a subset of training samples, the validation sample set is input into the network to calculate the classification accuracy, which is the fitness value of that sparrow individual. Based on the fitness values ​​of each sparrow individual, the population collaboratively updates the spatial positions using the producer position update formula, the beggar position update formula, and the scout position update formula. When the number of iterations reaches a preset maximum value or the fitness value no longer increases, the convergence condition is met, the iteration stops, and the position vector of the sparrow individual with the highest fitness value is output as the optimal parameter combination.

[0076] The implementation scheme of the present invention will be described in detail below with reference to specific experimental data:

[0077] Step 1: Acquisition and Cleaning Preprocessing of SCADA Operation Data

[0078] The data in this embodiment comes from a plain wind farm, with a sampling period covering December 4, 2019 to June 30, 2020. It covers SCADA records of 33 units, totaling 1,730,080 raw data entries. Each sample contains 37 raw monitoring signal variables (including wind speed, blade speed, power, bearing temperature of various components, grid-side voltage, excitation current, etc.).

[0079] 1. Data cleaning: Remove missing and duplicate values ​​from SCADA data.

[0080] 2. Outlier Removal: For outliers caused by power curtailment or sudden wind speed changes, k-means clustering analysis is used to divide the distribution space of the scatter plot, identify edge noise points that deviate significantly from the normal power generation curve, and remove them as a whole.

[0081] 3. Sample Balance Processing: To avoid classification bias, 1,500 samples were strictly selected for each of the five typical fault types, resulting in a total of 7,500 standard balanced datasets.

[0082] Step 2: Dimensionality reduction and extraction of fault feature parameters using PCA

[0083] Directly inputting the original 37-dimensional feature variables would lead to computational redundancy in the model. In this embodiment, principal component analysis (PCA) is used to perform linear transformation and dimensionality reduction on the cleaned feature matrix.

[0084] Table 1 Principal component distribution:

[0085]

[0086] Table 2: Influence of Original Features on Principal Component Weights

[0087]

[0088] Based on the analysis results in Tables 1 and 2, the variance contributions of each principal component are as follows: the variance contribution percentage of principal component PC1 is 38.13%, and the features with the highest influence weight on PC1 include eigenvalues ​​F2, F3, F23, F22, F21, F14, F16, F15, F1, F20, and F19; the variance contribution percentage of principal component PC2 is 25.4%, and the features with high influence weight are F2, F20, F23, F21, F22, F3, F19, and F1.

[0089] By comprehensively considering the load matrices of the first five principal components (PC1~PC5), a total of 16 core feature variables were ultimately selected and retained: F1, F2, F3, F6, F8, F9, F10, F14, F15, F16, F19, F20, F21, F22, F23, and F25. These 16 streamlined features cover more than 95.8% of the information expressed in the original dataset, effectively removing mutually coupled noise and greatly accelerating the convergence of subsequent networks.

[0090] Step 3: Construction of BLS Classification Network

[0091] A breadth-learning system is constructed, with the reduced 16-dimensional feature vector input to the network's input layer. The steps for solving the BLS matrix are as follows:

[0092] 1. Let the input sample input matrix be X. The input layer is obtained by performing linear mapping and activation function transformation on the original data matrix. Feature nodes assemble into the first Group feature nodes:

[0093]

[0094] in, For the first Group feature node matrix, For random mapping functions, and The first The feature node matrix is ​​obtained by horizontally concatenating the weight matrices and bias vectors randomly generated from each group of feature nodes. .

[0095] 2. Transform the feature mapping matrix Introducing a nonlinear activation function to enhance the topology, generating the j-th group of enhanced nodes:

[0096]

[0097] in, For containing The first neuron Group enhancement node matrix, It is a non-linear activation function. and The first The weight matrix and bias vector of the augmentation nodes are randomly generated. The augmentation nodes of the augmentation layer H are composed of feature nodes. The feature nodes are obtained by performing linear transformations and activation functions. Then, the augmentation nodes can be calculated. : The output layer of the width learning system network model. It can be represented as:

[0098]

[0099] in, The weights between the hidden layer (composed of feature nodes and augmentation nodes) and the output layer are denoted as . For feature nodes and enhanced nodes The hidden layer matrix is ​​formed by the joint construction of the hidden layers. to output layer The weights between Then the output layer This can be further expressed as:

[0100]

[0101] Since the feature nodes and enhancement nodes that make up A can be obtained by calculating the input data, and the data of the output layer is also known, the training of the BLS network model can be transformed into solving the problem of weights.

[0102] 3. Convert the feature node matrix and enhanced node matrix The components are horizontally joined and assembled into the system's state matrix. .

[0103] 4. Based on the state matrix The expected output matrix corresponding to the training sample subset and regularization coefficient The output weight matrix is ​​calculated using the ridge regression pseudo-inverse algorithm. The calculation formula is:

[0104]

[0105] in, This is the output layer weight matrix of the width learning system network. The regularization coefficient is . It is the identity matrix. The state matrix, It is the transpose of the state matrix. The expected output matrix corresponding to the training sample subset, for The pseudo-inverse matrix. The initial parameters W in the width learning system network. fi B fi W hj B hj The weights W are randomly generated and fine-tuned using a sparse autoencoder, and remain unchanged during training. The entire network model consists of only the weights W. m Training and learning are required.

[0106] Step 4: Introduce the SSA algorithm for multi-parameter collaborative global optimization.

[0107] To eliminate the blindness inherent in traditional grid search, this invention injects the Sparrow Search Algorithm (SSA) before model training, setting the population size (virtual sparrow number M=30), the optimization parameter space dimension N=5, and the optimization vector as follows: Based on the sparrow population division model in the paper, the population is divided into producers who provide foraging directions and beggars who follow the producers to intercept energy, with a fixed ratio. Wanderers are randomly assigned to act as scouts, triggering anti-predation warning mechanisms. During optimization, 40% of the 7500 fault dataset sets are designated as the training baseline, and 60% as the validation set. The classification accuracy on the validation set is used as the fitness function to drive the sparrow population to iteratively update its position coordinates in space until the fitness function curve converges to the global maximum value, at which point the optimal hyperparameter combination is obtained.

[0108] In the sample set of 7,500 wind farm SCADA fault balancing samples, the diagnostic targets cover three different unit states:

[0109] Status 1: Pitch system – toothed band with low risk of fracture;

[0110] Status 2: Pitch system – toothed band with high risk of fracture;

[0111] Status 3: Pitch system - Normal operation.

[0112] The BLS and SSA-BLS diagnostic models were used to classify the states of the three units involved in the industrial competition model. The training and classification results of the two models were compared, and the classification results are as follows: Figure 2 As shown.

[0113] Overall, both diagnostic models performed relatively accurately in classifying states 1 and 3, with an accuracy rate of 98% for both. The accuracy rate for state 3 was slightly higher, with the BLS model achieving 99.81% accuracy and the SSA-BLS model achieving an even higher accuracy rate of 99.91%. For state 2, both models performed poorly, with the BLS model achieving only 75.95% accuracy, while the SSA-BLS model achieved a slightly higher accuracy rate of 85.95%. A confusion matrix was used to further analyze the classification performance of the two algorithms. Figure 3 As shown, Figure 3 In the diagram, a is a schematic diagram of the confusion matrix of the BLS model, and b is a schematic diagram of the confusion matrix of the SSA-BLS model.

[0114] For the classification of state 1, the BLS model misclassified 114 data points as state 2, while the SSA-BLS model only identified 68 of them as state 2. For the classification of state 2, the BLS model identified 1,803 of the 7,500 state 2 data points as state 3, while the SSA-BLS model misclassified 1,054 data points. However, the classification of state 3 was almost perfect, with the vast majority of state 3 data points being correctly classified.

[0115] Four performance evaluation metrics for the classification models were calculated, and the time spent by BLS and SSA-BLS during training and testing was recorded. The specific metric calculation results are shown in Table 3.

[0116] Table 3 Performance Indicators of Fault Diagnosis Model

[0117]

[0118] The calculation results show that after optimizing the BLS model with the SSA algorithm, there are significant improvements in precision, recall, and F1 score. The BLS model has low classification accuracy for state 3, indicating a poor ability to correctly identify that the unit is in state 3. In contrast, the SSA-BLS model has a 7.08% improvement in classification accuracy for state 3. The BLS model has a low recall rate for state 2, only 75.95%, which suggests that when the unit is in state 2, it is prone to misclassifying the unit as belonging to other states. The SSA-BLS model has a higher recall rate for state 2, with a 10% improvement.

[0119] Because BLS uses an incremental learning algorithm, the initial training phase, which involves building the network model and its parameters, takes a relatively long time. However, subsequent testing with data or adding new features takes significantly less time than the initial training. Furthermore, since it's not a deep learning architecture, the overall training time is faster. For BLS, training on 25,200 datasets took 19.2307 seconds, while testing took 2.2128 seconds. The SSA-BLS algorithm, optimized by SSA, further reduces the training time to just 17.1098 seconds, and the testing time to 2.0601 seconds. Therefore, considering both classification performance metrics and training time, the SSA-BLS model outperforms the BLS model in both classification performance and training time.

[0120] For the aforementioned carefully selected 7,500 wind farm SCADA fault balance sample sets, the diagnostic targets cover 5 high-risk fault labels:

[0121] Fault 1 (S1): Generator – toothed band with low risk of breakage;

[0122] Fault 2 (S2): Inverter - air-cooled fault;

[0123] Fault 3 (S3): Converter-feed fault, line overvoltage;

[0124] Fault 4 (S4): Generator-excitation fault, DC link overvoltage;

[0125] Fault 5 (S5): Converter - power supply failure, line undervoltage.

[0126] Blind test confusion matrix analysis was performed using 600 test samples. BLS and SSA-BLS diagnostic models were used to train SCADA data from actual wind farms and diagnose five fault states of the turbines. The training results were compared, and the diagnostic results are as follows: Figure 4 As shown. Combined with Figure 4 It can be seen that, Figure 4 Table 4 shows the actual and predicted classifications of the BLS model on the test set (a) and the SSA-BLS model on the test set (b). Both BLS and SSA-BLS achieved 100% classification accuracy for the first three fault types. However, BLS misclassified faults 4 and 5, misclassifying a small number of states belonging to fault 5 as fault 4 and a larger number of states belonging to fault 4 as fault 5. The BLS model, after optimization with SSA, achieved higher classification accuracy, with only a smaller number of misclassified samples. Therefore, SSA-BLS demonstrates a significant improvement in classification performance, as detailed in Table 4.

[0127] Table 4 Performance Indicators of Fault Diagnosis Model:

[0128]

[0129] As shown in Table 4, due to the limited number of training and testing samples, the overall classification performance of the model is good, and most states can be accurately classified. The BLS model has a classification accuracy of 98.61% for fault 4 and 94.87% for fault 5. The accuracy and recall of the other three faults are both 100%. The training time of the model is short, at 2.7883s, and the testing time is even shorter, at only 0.5313s. The SSA-BLS has an accuracy of 99.5% and 99.7% for fault 4 and fault 5, respectively, with a training time of only 1.9295s and a testing time of only 0.2767s.

[0130] In the highly challenging task of diagnosing mixed generator excitation faults and converter power supply faults, the conventional BLS model, due to the blindness of unsupervised mapping, misdiagnosed 32 samples belonging to fault 5 as fault 4 and 8 samples belonging to fault 4 as fault 5. This invention, by introducing swarm intelligence SSA optimization, directly increases the global diagnostic accuracy from 98.67% to 99.83%. In all 3000 blind test samples, only 5 samples showed slight errors, and the entire diagnostic network training process was completed within 2.2155 seconds. Compared to the shortcomings of traditional BP neural networks or support vector machines, which are prone to getting trapped in local optima and training lag, the real-time performance and high accuracy of this invention are extremely prominent, making it highly suitable for deployment in the fully automated centralized control and maintenance systems of modern wind farms.

[0131] Example 2

[0132] Based on the same inventive concept as Embodiment 1, this embodiment introduces a wind turbine fault diagnosis system based on SSA-BLS, including:

[0133] The data acquisition interface is used to connect to the wind turbine SCADA system and acquire the unit's operating data. The data acquisition interface supports standard industrial communication protocols and can establish a data connection with the SCADA system server via Ethernet, serial port or wireless communication to achieve real-time acquisition of operating data.

[0134] The data preprocessing module, whose input end is connected to the output end of the data acquisition interface, is used to perform missing value removal, duplicate value removal, and cluster outlier removal on the original input matrix to obtain a cleaned input matrix; and uses the cleaned input matrix as input for principal component analysis to perform dimensionality reduction to obtain a low-dimensional feature matrix; the low-dimensional feature data output by the data preprocessing module serves as input for the subsequent parameter optimization module, and as a source of training and testing samples for the network construction and diagnosis module.

[0135] The diagnostic module, whose input is connected to the output of the data preprocessing module, is used to take the low-dimensional feature matrix as the input sample set, input the test sample subset divided from the input sample set into the trained width learning system network, the width learning system network calculates the classification output matrix according to the pre-calculated output layer weight matrix, determines the fault category of each test sample according to the classification output matrix and outputs the diagnostic result.

[0136] A network construction module is used to construct a wide learning system network, wherein the wide learning system network is composed of a state matrix formed by horizontally concatenating a feature node matrix and an enhancement node matrix, and the output layer weight matrix is ​​calculated using the state matrix.

[0137] The parameter optimization module has its input terminals connected to the output terminals of the data preprocessing module and the network construction module, respectively. It is used to perform global optimization of the parameters to be optimized in the wide learning system network using the sparrow search algorithm, based on the input sample set as the basis for calculating the fitness function, and output the optimal parameter combination.

[0138] The network construction and training module has its input terminals connected to the output terminals of the data preprocessing module and the parameter optimization module, respectively. It is used to configure the topology of the width learning system network with the optimal parameter combination; it uses a subset of training samples divided from the input sample set as the input of the width learning system network, solves the output layer weight matrix, obtains the trained width learning system network, and outputs it to the diagnostic module.

[0139] The diagnostic display terminal, whose input is connected to the output of the diagnostic module, is used to receive and display the diagnostic results. The diagnostic display terminal uses a graphical user interface to display the current status and fault alarm information of each wind turbine in the form of tables, trend graphs, or dashboards, and triggers audible and visual alarms when a fault is detected.

[0140] The specific functional implementation of each module is as described in the relevant content of the method in Embodiment 1, and will not be repeated here. It should be noted that the network construction module includes a feature node generation unit, an enhancement node generation unit, and a state matrix concatenation unit. The feature node generation unit is used to randomly map the low-dimensional feature data into a feature node matrix. The enhancement node generation unit is used to nonlinearly transform the feature node matrix into an enhancement node matrix. The state matrix concatenation unit is used to horizontally concatenate the feature node matrix and the enhancement node matrix into the state matrix.

[0141] Example 3

[0142] Based on the same inventive concept as other embodiments, this embodiment introduces a wind turbine fault diagnosis device, comprising:

[0143] Memory, used to store computer programs;

[0144] A processor is used to execute the computer program to implement the steps of the wind turbine fault diagnosis method based on SSA-BLS as described in any of Embodiment 1.

[0145] It is understandable that the device can be deployed on a server in the central control room of the wind farm's booster station, or in an industrial control cabinet at the bottom of each wind turbine tower. When deployed on the central control room server, the device communicates with the SCADA system of each wind turbine in the entire field through a fiber optic ring network to centrally perform status monitoring and fault diagnosis of all units in the field. When deployed at the bottom of a single wind turbine tower, the device only performs independent diagnostic tasks for that wind turbine, and the diagnostic results are sent to the central control center through the communication network.

[0146] In summary, this invention combines the incremental learning characteristics of a wide learning system with the swarm intelligence optimization capability of a sparrow search algorithm. After the SCADA operating data is cleaned and dimensionality reduced, the sparrow search algorithm is used to globally optimize the topological parameters and regularization coefficients of the wide learning system, constructing a diagnostic network with the optimal parameter combination and performing fault classification. In the above embodiments, the horizontal expansion of the state matrix of the wide learning system replaces the multi-layer backpropagation iteration of traditional deep networks, eliminating the gradient chain propagation overhead during training and enabling the model to complete training in seconds. The sparrow search algorithm, through the collaborative division of labor among producers, beggars, and scouts, performs a global parallel search in the parameter space, overcoming the defects of underfitting or overfitting caused by manual experience-based hyperparameter tuning. The synergy of both allows the diagnostic network to maintain extremely low training costs while fully capturing the deep features of SCADA data. Verified by publicly available data, this invention achieves an accuracy rate of 99.83% in diagnosing five typical faults, including generator toothed belt breakage, converter air cooling fault, converter feeder overvoltage, generator excitation overvoltage, and converter power supply undervoltage. Compared with conventional width learning systems without hyperparameter optimization, the diagnostic accuracy is significantly improved. The training time is short and the diagnostic response is rapid. It can be deployed in the actual operation and maintenance environment of wind farms, meeting the engineering application requirements for real-time status monitoring and agile alarms.

[0147] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0148] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0149] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0150] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0151] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A wind turbine fault diagnosis method based on SSA-BLS, characterized in that, include: The operating data collected by the wind turbine SCADA system is used as the raw input matrix; After performing missing value removal, duplicate value removal, and cluster outlier removal on the original input matrix, a cleaned input matrix is ​​obtained; The cleaned input matrix is ​​used as the input for principal component analysis to reduce dimensionality and obtain low-dimensional eigenmotes. Using the low-dimensional feature matrix as the input sample set, the test sample subset divided from the input sample set is input into the trained width learning system network. The width learning system network calculates the classification output matrix according to the pre-calculated output layer weight matrix, determines the fault category of each test sample according to the classification output matrix, and outputs the diagnosis result. The trained width learning system network is obtained in the following way: A wide learning system network is constructed, wherein the wide learning system network is a state matrix formed by horizontally concatenating a feature node matrix and an enhancement node matrix, and the output layer weight matrix is ​​calculated using the state matrix. Using the input sample set as the basis for calculating the fitness function, the sparrow search algorithm is used to globally optimize the parameters to be optimized in the wide learning system network, and the optimal parameter combination is output. Configure the topology of the width learning system network with the optimal parameter combination; The training sample subset divided from the input sample set is used as the input of the width learning system network, and the output layer weight matrix is ​​solved to obtain the trained width learning system network.

2. The wind turbine fault diagnosis method based on SSA-BLS according to claim 1, characterized in that, The parameters to be optimized include the number of feature points, the number of augmentation points, the number of feature node neurons, the number of augmentation node neurons, and the regularization coefficient.

3. The wind turbine fault diagnosis method based on SSA-BLS according to claim 1, characterized in that, The outlier removal process involves using the K-means clustering algorithm to cluster the operational data, identifying edge noise points that deviate from the normal operating power curve distribution range, and removing them.

4. The wind turbine fault diagnosis method based on SSA-BLS according to claim 1, characterized in that, The width learning system network is constructed as follows: a feature node matrix is ​​generated from the input sample set; an enhancement node matrix is ​​generated from the feature node matrix; the feature node matrix and the enhancement node matrix are horizontally concatenated to form a state matrix, wherein the number of rows of the state matrix is ​​the number of samples in the training sample subset, and the number of columns of the state matrix is ​​the sum of the number of columns of the feature node matrix and the enhancement node matrix.

5. The wind turbine fault diagnosis method based on SSA-BLS according to claim 4, characterized in that, The output layer weight matrix of the width learning system is calculated using a pseudo-inverse algorithm based on ridge regression. The calculation formula is as follows: , In the formula, Let I be the output layer weight matrix, λ be the regularization coefficient, I be the identity matrix, A be the state matrix formed by horizontally concatenating the feature node matrix and the booster node matrix, and Y be the desired output matrix. It is the transpose of the state matrix.

6. The wind turbine fault diagnosis method based on SSA-BLS according to claim 1, characterized in that, In the sparrow search algorithm, the sparrow population is divided into producers and beggars according to a preset ratio, and scouts are randomly assigned; the producers determine the foraging direction and update their spatial position according to the fitness value. The beggars update their spatial location by following the producer with the highest fitness value; the scouts trigger anti-predation warnings and randomly update their spatial location during the population spatial location update process.

7. The wind turbine fault diagnosis method based on SSA-BLS according to claim 6, characterized in that, The Sparrow Search algorithm uses the classification accuracy on the verification sample set as the fitness value and aims to maximize the classification accuracy as the convergence objective; it determines that the convergence condition is met when the number of iterations reaches a preset maximum value or the fitness value no longer increases.

8. A wind turbine fault diagnosis system based on SSA-BLS, characterized in that, include: The data acquisition interface is used to connect to the wind turbine's SCADA system and acquire the unit's operating data. The data preprocessing module, whose input end is connected to the output end of the data acquisition interface, is used to perform missing value removal, duplicate value removal, and cluster outlier removal on the original input matrix to obtain a cleaned input matrix; and uses the cleaned input matrix as input for principal component analysis to perform dimensionality reduction to obtain a low-dimensional feature matrix. The diagnostic module, whose input is connected to the output of the data preprocessing module, is used to take the low-dimensional feature matrix as the input sample set, input the test sample subset divided from the input sample set into the trained width learning system network, the width learning system network calculates the classification output matrix according to the pre-calculated output layer weight matrix, determines the fault category of each test sample according to the classification output matrix and outputs the diagnostic result. A network construction module is used to construct a wide learning system network, wherein the wide learning system network is composed of a state matrix formed by horizontally concatenating a feature node matrix and an enhancement node matrix, and the output layer weight matrix is ​​calculated using the state matrix. The parameter optimization module has its input terminals connected to the output terminals of the data preprocessing module and the network construction module, respectively. It is used to perform global optimization of the parameters to be optimized in the wide learning system network using the sparrow search algorithm, based on the input sample set as the basis for calculating the fitness function, and output the optimal parameter combination. The network construction and training module has its input terminals connected to the output terminals of the data preprocessing module and the parameter optimization module, respectively. It is used to configure the topology of the width learning system network with the optimal parameter combination; it uses a subset of training samples divided from the input sample set as the input of the width learning system network, solves the output layer weight matrix, obtains the trained width learning system network, and outputs it to the diagnostic module. The diagnostic display terminal has its input end connected to the output end of the diagnostic module, and is used to receive and display the diagnostic results.

9. The wind turbine fault diagnosis system based on SSA-BLS according to claim 8, characterized in that, The network construction module includes a feature node generation unit, an enhancement node generation unit, and a state matrix concatenation unit. The feature node generation unit is used to randomly map the low-dimensional feature data into a feature node matrix. The enhancement node generation unit is used to nonlinearly transform the feature node matrix into an enhancement node matrix. The state matrix concatenation unit is used to horizontally concatenate the feature node matrix and the enhancement node matrix into the state matrix.

10. A wind turbine fault diagnosis device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the wind turbine fault diagnosis method based on SSA-BLS as described in any one of claims 1 to 7.

Images