Wind turbine generator migration learning fault diagnosis method and device based on CC-FOA

Abstract

The invention discloses a CC-FOA-based wind turbine generator migration learning fault diagnosis method and device, and relates to the field of wind turbine generator fault diagnosis, and the method comprises the steps: collecting vibration signal data of a plurality of wind turbine generators, dividing the vibration signal data into source domain data and target domain data, and carrying out the preprocessing; training a stacked sparse denoising auto-encoder network by using the preprocessed source domain data; optimizing the number of hidden layers of the stacked sparse denoising auto-encoder network and the number of neurons in each layer by using a chaos crisscross fruit fly optimization algorithm; inputting the preprocessed source domain data and target domain data into a stacked sparse denoising auto-encoder network, adding a joint probability maximum mean difference constraint to each hidden layer, and retraining the network based on the sum of errors; and applying the retrained stacked sparse denoising auto-encoder network to target domain data, extracting features and carrying out fault classification. According to the invention, the accuracy and reliability of fault identification can be effectively improved.

Description

Technical Field

[0001] The present application relates to the field of wind turbine fault diagnosis, and in particular to a wind turbine fault diagnosis method and device based on CC-FOA transfer learning. Background Art

[0002] Wind turbine multi-component fault identification has become a major challenge in the wind power industry, which is closely related to the safe operation, power generation efficiency and economic benefits of wind farms. As a complex electromechanical system, wind turbines are composed of multiple key components such as blades, gearboxes, generators, etc. These components are prone to various faults during long-term operation. Timely and accurate identification of these faults can not only prevent major accidents, but also optimize maintenance strategies, reduce downtime and improve power generation efficiency. Therefore, wind turbine multi-component fault identification is of great significance for ensuring the stable operation of wind farms, reducing maintenance costs and extending equipment life. At present, the main technologies for wind turbine multi-component fault identification include signal processing-based methods, model-based methods and data-driven methods. Signal processing methods mainly detect anomalies by analyzing sensor data such as vibration, sound, and temperature; model-based methods identify faults by establishing a physical or mathematical model of the wind turbine and comparing the differences between actual operating data and model prediction results; while data-driven methods use machine learning algorithms to learn fault characteristics from a large amount of historical data. The working principle of these methods is based on continuous monitoring and analysis of the operating status of wind turbines, and the occurrence and type of faults are determined by identifying abnormal patterns or deviations from normal operating conditions.

[0003] However, these existing technologies have some problems. First, due to the complex and changeable operating environment of wind turbines, a single signal processing or model method is often difficult to cope with various complex fault conditions. Second, traditional data-driven methods require a large amount of labeled data for training, and in practical applications, it is often difficult and costly to obtain sufficient labeled fault data. In addition, the difference in data distribution between different wind farms or different types of wind turbines also poses a challenge to the generalization ability of fault identification models. To solve the above problems, researchers have proposed a variety of improvement methods. Among them, methods based on deep learning have attracted widespread attention due to their powerful feature extraction capabilities. For example, convolutional neural networks (CNNs) or recurrent neural networks (RNNs) are used to automatically learn the time-frequency characteristics of vibration signals, thereby improving the accuracy of fault identification. Another method is to use ensemble learning to combine the advantages of multiple algorithms to improve the robustness of the model. In addition, some researchers have tried to use unsupervised learning or semi-supervised learning methods to reduce the dependence on labeled data. However, these methods still have some limitations. First, they usually require a large amount of labeled data, which is difficult to obtain in practical applications. Secondly, these methods often have difficulty in effectively handling the distribution differences between different data domains, resulting in poor performance of the model on new wind farms or new types of wind turbines. Finally, traditional methods may not be able to make full use of existing knowledge and need to train the model from scratch each time, which is not only time-consuming and labor-intensive, but may also lead to unstable model performance.

[0004] Therefore, there is an urgent need for a technical solution that can effectively utilize the knowledge of source domain data to assist the learning of target domain data and improve the accuracy and reliability of fault identification. Summary of the invention

[0005] In order to solve the deficiencies of the prior art, the embodiment of the present application provides a wind turbine fault diagnosis method and device based on CC-FOA (ChaoticSpanning Crossover Fruit Fly Optimization Algorithm). The present application solves the technical problems that the prior art model has low generalization ability between different fault types.

[0006] The embodiment of the present application provides a wind turbine group transfer learning fault diagnosis method based on CC-FOA, including: collecting vibration signal data of multiple wind turbines, dividing the vibration signal data into source domain data and target domain data; preprocessing the source domain data and the target domain data, including signal denoising and dimensionality transformation; using the preprocessed source domain data to train a stacked sparse denoising autoencoder network; using a chaotic cross-linked fruit fly optimization algorithm (CC-FOA) to optimize the number of hidden layers and the number of neurons in each layer of the stacked sparse denoising autoencoder network; inputting the preprocessed source domain data and target domain data into the optimized stacked sparse denoising autoencoder network, adding a joint probability maximum mean difference constraint to each hidden layer, and retraining the network based on the sum of errors; applying the retrained stacked sparse denoising autoencoder network to the target domain data, extracting features and performing fault classification.

[0007] In one possible implementation, the number of hidden layers and the number of neurons in each layer of the stacked sparse denoising autoencoder network are optimized using a chaotic crossover fruit fly optimization algorithm, including: setting the fruit fly population size, the maximum number of iterations, the error accuracy, the acceleration factor, and the vertical crossover probability; converting the initial population of fruit flies into the number of hidden layers and the number of neurons in each layer of the autoencoder for training, and calculating the reconstruction error as the fitness value of the fruit flies; sorting the fitness values ​​and individual positions, and recording the global optimal value and the corresponding optimal individual ; All fruit flies in the population are randomly combined without duplication, and horizontal crossover is performed to calculate the fitness value of the fruit flies, and the fruit flies with the best fitness in the current population are retained; each dimension of the fruit flies obtained by horizontal crossover is normalized, and then all dimensions are randomly paired without duplication, and a vertical crossover operation is performed. After denormalization, the fitness values ​​are compared, and the global optimal value is updated. and the corresponding optimal individual The algorithm continues to iterate until the maximum number of iterations is reached or the error accuracy reaches the preset threshold. The fruit fly with the best fitness is selected as the optimal solution and output. The optimal solution is the optimal number of hidden layers and the number of neurons in each layer of the stacked sparse denoising autoencoder.

[0008] In a possible implementation, the preprocessed source domain data and target domain data are input into the optimized stacked sparse denoising autoencoder network, a joint probability maximum mean difference constraint is added to each hidden layer, and the network is retrained based on the sum of errors, including: inputting the preprocessed source domain data and target domain data into the stacked sparse denoising autoencoder network; using the maximum mean difference to measure the distribution difference between the source domain and the target domain, and constructing a joint probability maximum mean difference constraint including marginal distribution and conditional distribution; embedding the joint probability maximum mean difference into the hidden layer of each sparse denoising autoencoder to migrate the model layer by layer; training the stacked sparse denoising autoencoder network by minimizing the loss function including reconstruction error and joint probability maximum mean difference to obtain the optimal parameter set.

[0009] In a possible implementation, the maximum mean difference is used to measure the distribution difference between the source domain and the target domain, and a joint probability maximum mean difference constraint including marginal distribution and conditional distribution is constructed, including: establishing a joint probability difference model between the source domain and the target domain, including a homogeneous transferability metric and a heterogeneous distinguishability metric to define the maximum mean difference of the discriminant joint probability; constructing a label matrix of the source domain data and a pseudo-label matrix of the target domain data, and calculating the first and second normalized matrices; generating an extended label matrix of the source domain and the target domain to calculate the third and fourth normalized matrices; constructing an optimization objective function of the maximum mean difference of the discriminant joint probability based on the first, second, third and fourth normalized matrices to solve the optimal linear mapping.

[0010] In a possible implementation, the maximum mean difference of the joint probability is embedded in the hidden layer of each sparse denoising autoencoder to migrate the model layer by layer, including: embedding the maximum mean difference of the joint probability into the hidden layer of each sparse denoising autoencoder, extracting features of the source domain data and the target domain data by minimizing the loss function of the first sparse denoising autoencoder to align the feature space; using the output of the first sparse denoising autoencoder as the input of the second sparse denoising autoencoder, reducing the distribution difference of the source domain data and the target domain data in the new feature space by minimizing the loss function of the second sparse denoising autoencoder; using the output of the second sparse denoising autoencoder as the input of the third sparse denoising autoencoder, and achieving full alignment of the source domain data and the target domain data in the common feature space by minimizing the loss function of the third sparse denoising autoencoder to migrate the model.

[0011] In a possible implementation, a joint probability difference model of the source domain and the target domain is established, including a homogeneous transferability metric and a heterogeneous distinguishability metric, to define the maximum mean difference of the discriminant joint probability, including: establishing a joint probability difference model of the source domain and the target domain including the homogeneous transferability metric and the heterogeneous distinguishability metric, wherein the homogeneous transferability metric uses the maximum mean difference MMD to calculate the transferability from the source domain to the target domain; based on an empirical formula, obtaining the product formula of the conditional expectation and prior probability of the source domain and the target domain, substituting it into the maximum mean difference MMD formula to obtain the final source domain to target domain transferability metric formula; establishing a heterogeneous distinguishability metric model, using the maximum mean difference MMD to calculate the distinguishability between different fault categories in the source domain and the target domain, and obtaining the final source domain and target domain heterogeneous distinguishability metric formula; substituting the obtained homogeneous transferability metric formula and heterogeneous distinguishability metric formula into the established joint probability difference model to obtain a complete definition of the maximum mean difference of the discriminant joint probability.

[0012] In a possible implementation, the transferability metric includes: , in, represents the transferability metric; C represents the total number of categories, that is, the number of fault types; c represents the current category index, ranging from 1 to C; Represents the total number of samples of source domain data; Represents the number of samples belonging to category c in the source domain; Represents the total number of samples of target domain data; Represents the number of samples belonging to category c in the target domain; Represents the transpose of the linear mapping matrix A, which is used to map data from the original feature space to the new feature space; represents the iith sample belonging to category c in the source domain; represents the jjth sample in the target domain belonging to category c.

[0013] In a possible implementation, the discriminative metric includes: , in, It represents a discriminative measure, which is used to measure the distribution difference between source and target domains in different categories. represents another category index, ranging from 1 to C, and ; Indicates that the target domain belongs to the category The number of samples; Indicates that the jjth item in the target domain belongs to the category of the sample.

[0014] An embodiment of the present application also provides a wind turbine group transfer learning fault diagnosis device based on CC-FOA, including: a processor, a memory, and a system bus; wherein the processor and the memory are connected via the system bus; the memory is used to store one or more programs, and the one or more programs include instructions, and when the instructions are executed by the processor, the processor executes the method described in the above embodiment.

[0015] In the method and device for wind turbine transfer learning fault diagnosis based on CC-FOA as provided above, the embodiment of the present application collects vibration signal data of the wind turbine, trains the stacked sparse denoising autoencoder network, and optimizes the number of hidden layers and the number of neurons in each layer of the stacked sparse denoising autoencoder network in combination with the chaotic crisscross fruit fly optimization algorithm, and uses the optimized stacked sparse denoising autoencoder network and the joint probability maximum mean difference constraint to achieve domain adaptation, which can effectively utilize the knowledge of the source domain data to assist the learning of the target domain data, thereby improving the accuracy and reliability of fault identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0017] Figure 1 A flow chart of a wind turbine group fault diagnosis method based on CC-FOA transfer learning provided in an embodiment of the present application; Figure 2 A schematic diagram of a maximum mean difference constraint method provided in an embodiment of the present application; Figure 3 A schematic diagram of a process of a layer-by-layer migration method provided in an embodiment of the present application; Figure 4 A schematic diagram of a chaotic cross-cutting fruit fly optimization algorithm flow provided in an embodiment of the present application; Figure 5 A schematic diagram of the overall process of a migration identification method provided in an embodiment of the present application; Figure 6 A schematic diagram of a migration identification model provided in an embodiment of the present application; Figure 7 A schematic diagram of migration recognition performance provided in an embodiment of the present application. DETAILED DESCRIPTION

[0018] Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that the relative arrangement of components and steps, numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present disclosure unless otherwise specifically stated.

[0019] Those skilled in the art can understand that the terms "first", "second" and the like in the embodiments of the present disclosure are only used to distinguish different steps, devices or modules, etc., and neither represent any specific technical meaning nor represent the necessary logical order between them. It should also be understood that in the embodiments of the present disclosure, "multiple" can refer to two or more, and "at least one" can refer to one, two or more. It should also be understood that for any component, data or structure mentioned in the embodiments of the present disclosure, in the absence of explicit limitation or contrary revelation given in the context, it can generally be understood as one or more. In addition, the term "and / or" in the present disclosure is only a description of the association relationship of the associated objects, indicating that there can be three relationships, for example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in the present disclosure generally indicates that the associated objects before and after are an "or" relationship. It should also be understood that the description of each embodiment in the present disclosure emphasizes the differences between the embodiments, and the same or similar parts can refer to each other. For the sake of brevity, they will not be repeated one by one.

[0020] At the same time, it should be understood that, for ease of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present disclosure and its application or use. The techniques, methods and devices known to ordinary technicians in the relevant fields may not be discussed in detail, but where appropriate, the techniques, methods and devices should be considered part of the specification. It should be noted that similar numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further discussed in subsequent drawings.

[0021] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0022] Figure 1A flow chart of a wind turbine transfer learning fault diagnosis method 100 based on CC-FOA provided in an embodiment of the present application. The method aims to solve various fault problems that may occur in the actual operation of wind turbines, collect vibration signal data of wind turbines, train stacked sparse denoising autoencoder networks, and optimize the number of hidden layers and the number of neurons in each layer of the stacked sparse denoising autoencoder network by combining the chaotic cross-linked fruit fly optimization algorithm (CC-FOA), and use transfer learning technology to process and analyze data to achieve fault diagnosis of multiple components of wind turbines.

[0023] like Figure 1 As shown, at step S101, vibration signal data of multiple wind turbines are collected, and the vibration signal data are divided into source domain data and target domain data. Specifically, the method first collects vibration signal data of multiple wind turbines, and divides the collected vibration signal data into source domain data and target domain data. Source domain data generally refers to historical data of known fault types, while target domain data refers to real-time data of faults to be identified or data of unknown fault types. By dividing the source domain data and the target domain data, a basis can be provided for subsequent transfer learning.

[0024] At step S102, the source domain data and the target domain data are preprocessed, including signal denoising and dimensional transformation. After obtaining the source domain data and the target domain data, the method preprocesses these data, including operations such as signal denoising and dimensional transformation. Signal denoising refers to removing noise interference in the vibration signal to improve the quality and reliability of the signal. This can be achieved through various signal processing techniques, such as wavelet transform, Fourier transform, etc. Dimension transformation refers to converting the original high-dimensional vibration signal data into a more compact and effective low-dimensional representation to reduce computational complexity and improve the efficiency of feature extraction. Common dimensional transformation methods include principal component analysis (PCA), independent component analysis (ICA), etc.

[0025] At step S103, the stacked sparse denoising autoencoder network is trained using the preprocessed source domain data. The stacked sparse denoising autoencoder is a deep learning model, which is composed of a stack of multiple sparse denoising autoencoders. The sparse denoising autoencoder is an unsupervised learning algorithm that can learn high-level feature representations of data and has denoising and sparsification properties. By introducing sparsity constraints in the hidden layer of each sparse denoising autoencoder, the model can learn a more compact and discriminative feature representation.

[0026] At step S104, the number of hidden layers and the number of neurons in each layer of the stacked sparse denoising autoencoder network are optimized using the chaotic cross-linked fruit fly optimization algorithm (CC-FOA). In order to optimize the hyperparameters of the model, the number of hidden layers and the number of neurons in each layer of the stacked sparse denoising autoencoder network are optimized in combination with the chaotic cross-linked fruit fly optimization algorithm. The chaotic cross-linked fruit fly optimization algorithm is an improved fruit fly optimization algorithm. By combining chaos theory and cross-linked strategies, the global search capability and convergence accuracy of the algorithm are improved. The execution process of the chaotic cross-linked fruit fly optimization algorithm includes initializing the population with chaos, calculating the fitness value of each individual, updating the position of the fruit fly based on the chaotic map, and optimizing the individual position using the cross-linked strategy. The loop is iterated until the stop condition is met, and the global optimal solution is finally output.

[0027] Figure 4 A schematic diagram of a chaotic cross-cutting fruit fly optimization algorithm flow provided in an embodiment of the present application. The method aims to solve the hyperparameter optimization problem of stacked sparse denoising autoencoders. Figure 4 As shown in the figure, we first initialize the parameters, set the fruit fly population size to 30, the maximum number of iterations to 200, and the probability of vertical and horizontal crossing to 0.6. Then we set the number of iterations to Set to 1. Convert the initial population of fruit flies into the number of hidden layers and neurons in each layer of the stacked sparse denoising autoencoder for training, and calculate the reconstruction error by the following formula, which is the fitness value of the fruit fly.

[0028] ; Mean Squared Error (MSE) is used to measure the difference between the model prediction value and the true value. In the invention, MSE is used to calculate the reconstruction error of the stacked sparse denoising autoencoder network, which is used as the fitness value of the fruit fly individual in the fruit fly optimization algorithm to evaluate the performance of the stacked sparse denoising autoencoder network. is the true value (target value), indicating the true label or true output of the t-th sample. In the present invention, Represents the true value (or original input data) of the wind turbine vibration signal; is the predicted value (model output), which represents the model predicted value of the t-th sample. In the invention, represents the reconstructed value of the input data by the stacked sparse denoising autoencoder network (i.e., the model's predicted output for the input data); K is the total number of samples, indicating the total number of samples in the data set; m is an index variable, indicating the error calculation for the mth sample, from 1 to N. By minimizing MSE, the number of hidden layers and the number of neurons in each layer of the autoencoder can be optimized, thereby improving the model's ability to reconstruct wind turbine vibration signals and the accuracy of fault diagnosis. The calculation of MSE is one of the core indicators of model training and optimization, which directly affects the performance and generalization ability of the model. Sort the fitness values ​​and individual positions, and record the global optimal value and the corresponding optimal individual and the individual optimal value ; When the maximum number of iterations is not reached, the fruit fly population enters the optimization process according to the crossover algorithm generated by the following formula; ; Among them: Chaos fruit flies In the cross-difference fruit fly population, for is the population size of the crossover algorithm The dimension of variables included for Drosophila; That is ; Randomly generate chaotic initial fruit flies ; u is the control parameter of the chaotic map, with a value of 2; s is the perturbation intensity (or mixing coefficient), which is used to control the influence of the chaotic map and the global optimal value on the position update of the fruit fly. s is a parameter between 0 and 1, which is used to balance the chaotic map ( ) and the global optimum ( ) contributes to the update of the fruit fly position. Preferably, the value of s is 0.4; For the The updated value of the position of the fruit fly, in the chaotic cross-cutting fruit fly optimization algorithm, It represents the position of the fruit fly after being updated by chaotic mapping and cross-border strategy; For the The position of a chaotic fruit fly represents the position of the fruit fly generated by chaotic mapping. It is the result of chaotic mapping and is used to introduce chaotic characteristics to enhance the global search ability of the algorithm. is the position of the current global optimal solution, indicating the optimal fruit fly position found by the algorithm during the iteration process, which is used to guide the fruit fly population to search in the direction of the global optimal solution and improve the convergence of the algorithm; For the In the iteration The chaotic variable value of the dimension is the input value of the chaotic map, which is used to generate the next generation of chaotic variables. .

[0029] according to Different chaotic mapping formulas are used to update the positive and negative cases of , to ensure the dynamics and diversity of chaotic variables.

[0030] Through chaotic mapping and cross-cutting strategies, the algorithm can more effectively explore the global optimal solution in the search space and avoid falling into the local optimum. In the invention, the above formula is used to optimize the number of hidden layers and the number of neurons in each layer of the stacked sparse denoising autoencoder network, thereby improving the performance of the wind turbine fault diagnosis model.

[0031] All fruit flies in the population are randomly combined in pairs without duplication, and horizontal crossover is performed according to the following formula to calculate the fitness value of the fruit flies, and the fruit flies with small fitness values ​​are retained; ; ; ; in: A random number or weight coefficient in [0,1] used to control and Contribution ratio in the mixing operation; A random number between -1 and 1, used to control the difference term or The effect in the mixed operation, the difference term reflects the difference between the two individuals on dimension d; M is the fruit fly size, that is, the total number of individuals; is the variable dimension, that is, the total number of dimensions of the problem; , Drosophila and No. dimension , They are and The first dimension offspring; i, j represent the index of the individual; N represents a set of natural numbers, N(1,M) represents a set of natural numbers from 1 to the total number of individuals M; and N(1,D) represents a set of natural numbers from 1 to the number of variable dimensions D.

[0032] Normalize each dimension of the fruit flies obtained by horizontal crossover, then randomly pair all dimensions without duplication, perform vertical crossover operation according to the following formula, compare fitness values ​​after denormalization, and update the fruit flies; ; ; in, Indicated in dimension The result of some mixing or crossover operation of individual i; Represents individual i in dimension The eigenvalue or function value on Represents individual i in dimension The characteristic value or function value on the two values ​​are the attributes or states of individual i in different dimensions. r is a random number or weight coefficient with a value range of [0,1], which is used to control and The contribution ratio in the mixing operation. When r is close to 1, The contribution of is greater; when r is close to 0, The contribution is greater.

[0033] The algorithm continues to iterate until the maximum number of iterations is reached or the error accuracy reaches a preset threshold. The final global optimal solution is the optimal number of hidden layers and the number of neurons in each layer of the stacked sparse denoising autoencoder.

[0034] At step S105, the preprocessed source domain data and target domain data are input into the optimized stacked sparse denoising autoencoder network, a joint probability maximum mean difference constraint is added to each hidden layer, and the network is retrained based on the sum of errors. Specifically, it includes: inputting the preprocessed source domain data and target domain data into the stacked sparse denoising autoencoder network; using the maximum mean difference to measure the distribution difference between the source domain and the target domain, constructing a joint probability maximum mean difference constraint including marginal distribution and conditional distribution (such as Figure 2 As shown); embed the joint probability maximum mean difference into the hidden layer of each sparse denoising autoencoder to migrate the model layer by layer (as shown Figure 3 As shown); the stacked sparse denoising autoencoder network is trained to obtain the optimal parameter set by minimizing the loss function including the reconstruction error and the maximum mean difference of the joint probability.

[0035] The maximum mean difference of joint probability is an indicator to measure the difference in distribution between the source domain and the target domain, which takes into account both marginal distribution and conditional distribution. By embedding this constraint into the hidden layer of each sparse denoising autoencoder, the layer-by-layer migration of the model can be achieved, so that the feature distribution of the source domain and the target domain gradually approaches, reducing the difference between the domains. This transfer learning method can effectively use the knowledge of the source domain to assist the learning of the target domain and improve the generalization ability and adaptability of the model.

[0036] After adding the joint probability maximum mean difference constraint, this method constructs a loss function based on the reconstruction error and the joint probability maximum mean difference, and retrains the stacked sparse denoising autoencoder network. The reconstruction error measures the model's ability to reconstruct the input data, while the joint probability maximum mean difference measures the distribution difference between the source domain and the target domain. By minimizing the loss function containing these two terms, the model can effectively reduce the difference between the source domain and the target domain while maintaining good reconstruction ability, thus achieving domain adaptation.

[0037] At step S106, the retrained stacked sparse denoising autoencoder network is applied to the target domain data to extract features and perform fault classification. Finally, the method applies the retrained stacked sparse denoising autoencoder network to the target domain data to extract features and perform fault classification. By inputting the target domain data into the trained network, its high-level feature representation can be obtained. These feature representations contain key information of the target domain data and can effectively distinguish different fault types. Based on the extracted features, various classification algorithms, such as support vector machines (SVMs), random forests, etc., can be used to realize fault identification of multiple components of wind turbines.

[0038] In summary, the wind turbine transfer learning fault diagnosis method based on CC-FOA provided in the embodiment of the present application collects vibration signal data of the wind turbine, trains the stacked sparse denoising autoencoder network, and optimizes the number of hidden layers and the number of neurons in each layer of the stacked sparse denoising autoencoder network in combination with the chaotic cross-linked fruit fly optimization algorithm, and uses the optimized stacked sparse denoising autoencoder network and the joint probability maximum mean difference constraint to achieve domain adaptation, which can effectively use the knowledge of the source domain data to assist the learning of the target domain data and improve the accuracy and reliability of fault identification. This method has important practical application value for the operation and maintenance and fault diagnosis of wind turbines, and can reduce the downtime of wind turbines, improve power generation efficiency, and reduce maintenance costs.

[0039] Figure 2 A flow chart of a maximum mean difference constraint method provided for an embodiment of the present application. The method aims to solve the feature migration and classification tasks in the domain adaptation problem. Domain adaptation refers to the process of migrating knowledge learned in one domain (source domain) to another related but different domain (target domain). In practical applications, there are usually differences in data distribution between the source domain and the target domain. This difference will lead to performance degradation when the model trained in the source domain is directly applied to the target domain. Therefore, the present method effectively reduces the distribution difference between the source domain and the target domain by establishing a discriminant joint probability maximum mean difference model while maintaining the distinguishability between different categories.

[0040] like Figure 2As shown, at step S201, a joint probability difference model of the source domain and the target domain is established, including a homogeneous transferability metric and a heterogeneous distinguishability metric, so as to define the maximum mean difference of the discriminant joint probability. Specifically, it includes: establishing a joint probability difference model of the source domain and the target domain including a homogeneous transferability metric and a heterogeneous distinguishability metric, wherein the homogeneous transferability metric uses the maximum mean difference MMD to calculate the transferability from the source domain to the target domain; based on the empirical formula, the product formula of the conditional expectation and the prior probability of the source domain and the target domain is obtained, and it is substituted into the maximum mean difference MMD formula to obtain the final transferability metric formula from the source domain to the target domain; establishing a heterogeneous distinguishability metric model, using the maximum mean difference MMD to calculate the distinguishability between different fault categories in the source domain and the target domain, and obtaining the final heterogeneous distinguishability metric formula of the source domain and the target domain; substituting the obtained homogeneous transferability metric formula and heterogeneous distinguishability metric formula into the established joint probability difference model to obtain a complete definition of the maximum mean difference of the discriminant joint probability.

[0041] Among them, the transferability metrics include: , in, represents the transferability metric, which is used to measure the distribution difference between the source domain and the target domain under the same category; C represents the total number of categories, that is, the number of fault types; c represents the current category index, which ranges from 1 to C; Represents the total number of samples of source domain data; Represents the number of samples belonging to category c in the source domain; Represents the total number of samples of target domain data; represents the number of samples belonging to category c in the target domain; A represents the linear mapping matrix, which is used to map the original feature space to the new feature space in order to reduce the distribution difference between the source domain and the target domain. Represents the transpose of the linear mapping matrix A, which is used to map data from the original feature space to the new feature space; represents the iith sample belonging to category c in the source domain; represents the jjth sample belonging to category c in the target domain; Represents the square of the L2 norm, which is used to calculate the square of the Euclidean distance between two vectors.

[0042] Among them, the discriminative measures include: , in, It represents a discriminative measure, which is used to measure the distribution difference between source and target domains in different categories. represents another category index, ranging from 1 to C, and ; Indicates that the target domain belongs to the category The number of samples; Indicates that the jjth item in the target domain belongs to the category of the sample.

[0043] At step S202, a label matrix of the source domain data and a pseudo label matrix of the target domain data are constructed, and the first and second normalized matrices are calculated. The label matrix is ​​a matrix used to represent data category information, in which each row corresponds to a sample and each column corresponds to a category. For the source domain data, the label matrix can be directly constructed using the real label information. However, for the target domain data, since there is usually no real label information, a pseudo label matrix needs to be constructed. The pseudo label can be obtained by predicting the target domain data through some initial classifier (for example, a classifier trained on the source domain).

[0044] The calculation of the normalization matrix is ​​to standardize the data so that the scale differences between different features will not have a negative impact on the subsequent optimization process. The first normalization matrix is ​​calculated based on the label matrix of the source domain data, while the second normalization matrix is ​​calculated based on the pseudo-label matrix of the target domain data. These two normalization matrices will play an important role in the subsequent optimization process.

[0045] At step S203, an extended label matrix of the source domain and the target domain is generated to calculate the third and fourth normalized matrices. The extended label matrix is ​​an extension of the original label matrix, which may contain more information, such as the similarity between samples or the relationship between categories. By generating an extended label matrix, more prior knowledge or structural information can be introduced, thereby improving the effect of domain adaptation. The third and fourth normalized matrices are calculated based on the extended label matrix. The introduction of these matrices can help better capture the intrinsic structure and category information of the data, thereby better maintaining the discriminability of the data during the optimization process.

[0046] At step S204, an optimization objective function of the maximum mean difference of the discriminant joint probability is constructed based on the first, second, third and fourth normalized matrices to solve the optimal linear mapping. This optimization objective function is the core of the entire method, which comprehensively considers the transferability of the same class and the distinguishability of different classes, and utilizes the label information (including pseudo labels) of the source domain and the target domain.

[0047] The process of constructing the optimization objective function may involve a weighted combination of multiple sub-objectives. For example, it may include minimizing the MMD distance between the source domain and the target domain (improving transferability), maximizing the MMD distance between different categories (improving distinguishability), minimizing the classification error on the source domain, etc. These sub-objectives are balanced by appropriate weight coefficients to form the final optimization goal.

[0048] Solving the optimal linear mapping is the ultimate goal of the entire method. The linear mapping can be regarded as a feature transformation that maps the data points in the original feature space to a new feature space. In this new feature space, the distribution difference between the source domain and the target domain is minimized, while the distinguishability between different categories is maximized. This optimal linear mapping can be solved by an optimization algorithm (such as gradient descent).

[0049] It is worth noting that although this method uses linear mapping as a means of feature transformation, it does not mean that it is only applicable to linearly separable problems. Through the kernel trick, this method can be extended to nonlinear situations to handle more complex domain adaptation problems.

[0050] In general, the maximum mean difference constraint method provided in the embodiment of the present application constructs an effective domain adaptation framework by cleverly combining the same transferability metric and the heterogeneous distinguishability metric. This method not only considers the distribution difference between the source domain and the target domain, but also makes full use of the category information, thereby achieving effective knowledge transfer while maintaining category discriminability.

[0051] Figure 3 A flow chart of a layer-by-layer migration method provided for an embodiment of the present application. The method aims to solve the problem of inconsistent data distribution between the source domain and the target domain in cross-domain learning. The layer-by-layer migration method refers to gradually reducing the distribution difference between the source domain data and the target domain data through multiple consecutive processing steps, and finally realizing the effective migration of the model. The core idea of ​​this method is to use a multi-layer sparse denoising autoencoder (SDAE) structure, combined with the joint maximum mean difference (JMMD) technology, to align the feature distribution of the source domain and the target domain layer by layer.

[0052] Specifically, the sparse denoising autoencoder is an unsupervised learning algorithm that discovers useful features in data by learning a compressed representation of the input data. "Sparse" means that in the hidden layer, the activation values ​​of most neurons are close to zero, and only a few neurons are activated, which helps to learn more robust and meaningful feature representations. "Denoising" means adding noise to the input data during training, and then requiring the network to reconstruct the original noise-free data. This method can improve the model's robustness to noise and learn more stable features.

[0053] Joint maximum mean difference is a method to measure the difference between two probability distributions. In cross-domain learning, JMMD can be used to measure the difference between the source domain and the target domain data distribution and achieve domain adaptation by minimizing this difference. By embedding JMMD into the hidden layer of SDAE, the consistency of the source domain and target domain data distribution can be considered while extracting features, thereby better aligning the feature space.

[0054] This method adopts a three-layer sparse denoising autoencoder structure, each layer undertakes different tasks, and works together to achieve the goal of gradual migration. This multi-layer structure design allows the model to gradually refine and optimize feature representation, thereby more effectively handling complex cross-domain learning problems.

[0055] like Figure 3 As shown, at step S301, the joint probability maximum mean difference is embedded in the hidden layer of each sparse denoising autoencoder, and features of the source domain data and the target domain data are extracted by minimizing the loss function of the first sparse denoising autoencoder to align the feature space.

[0056] Specifically, the joint probability maximum mean difference is embedded into the hidden layer of the first sparse denoising autoencoder. The purpose of this step is to consider the consistency of the source domain and target domain data distribution while performing feature extraction. By minimizing the loss function of the first sparse denoising autoencoder, preliminary feature extraction and alignment of the source domain data and the target domain data can be achieved. The loss function here may include reconstruction error, sparsity constraint and JMMD term, where the reconstruction error ensures that the autoencoder can effectively reconstruct the input data, the sparsity constraint prompts the hidden layer to learn sparse feature representation, and the JMMD term is used to reduce the difference in feature distribution between the source domain and the target domain.

[0057] In step S302, the output of the first sparse denoising autoencoder is used as the input of the second sparse denoising autoencoder, and the distribution difference between the source domain data and the target domain data in the new feature space is reduced by minimizing the loss function of the second sparse denoising autoencoder.

[0058] The method further reduces the distribution difference between the source domain and the target domain data in the new feature space. The specific operation is to use the output of the first sparse denoising autoencoder as the input of the second sparse denoising autoencoder. This cascade structure allows the model to perform further feature extraction and optimization based on the features learned in the first layer. By minimizing the loss function of the second sparse denoising autoencoder, the distribution difference between the source domain and the target domain data can be further reduced in the new feature space. The loss function of this step may be similar to the first step, but the weights of each item may be adjusted to pay more attention to reducing the difference between domains.

[0059] In step S303, the output of the second sparse denoising autoencoder is used as the input of the third sparse denoising autoencoder, and the source domain data and the target domain data are fully aligned in the common feature space by minimizing the loss function of the third sparse denoising autoencoder to perform model migration.

[0060] Specifically, the output of the second sparse denoising autoencoder is used as the input of the third sparse denoising autoencoder. By minimizing the loss function of the third sparse denoising autoencoder, sufficient alignment of the source and target domain data can be achieved in the final feature space. The loss function of this step may further increase the weight of domain alignment to ensure that the data distribution of the source and target domains is as consistent as possible in the final feature space.

[0061] It is worth noting that the loss function of each layer of sparse denoising autoencoder may contain multiple components, such as reconstruction error, sparsity constraint, regularization term, and domain alignment term (such as JMMD). The specific forms and weights of these components may be adjusted according to the specific application scenario and data characteristics. For example, the reconstruction error can be in the form of mean square error or cross entropy, the sparsity constraint can be implemented through KL divergence, the regularization term can use L1 or L2 norm, and the domain alignment term can use JMMD or other domain adaptation techniques.

[0062] In general, the layer-by-layer migration method provided in the embodiment of the present application achieves the gradual alignment of the data distribution of the source domain and the target domain by cleverly combining the sparse denoising autoencoder and the maximum mean difference of the joint probability. This method can not only effectively extract the feature representation of the data, but also consider the differences between domains in the process of feature extraction, thereby achieving more effective knowledge transfer. The layer-by-layer design of this method allows the model to gradually refine and optimize the feature representation, which is of great significance for dealing with complex cross-domain learning problems. Through this method, the knowledge of the source domain can be fully utilized to improve the performance of the model in the target domain when the target domain data is limited.

[0063] Figure 5A schematic diagram of the overall process of a migration identification method provided for an embodiment of the present application. First, the source domain data and the target domain data are destroyed with the same noise level. Then, the destroyed source domain data is used to train the AE with sparsity constraints added. The trained AE can better handle the data of the noise level and has higher robustness and analysis capabilities. Next, the destroyed source domain data and target domain data are input into the trained AE to obtain the hidden features of the source domain and the target domain, respectively. Then, DJPMMD is used to measure the difference between the hidden features of the source domain and the target domain, and the AE is trained again until DJPMMD converges. Finally, the noise level is reduced, the original source domain data and target domain data are destroyed again, and the AE is retrained until the noise level reaches the set value. The hidden layer input of the trained AE is used as the input of the next AE, and the above process is repeated until all AEs are trained, and multiple trained AEs are stacked to obtain the final TL-SSMDAE.

[0064] Multi-level noise training can capture more detailed information from different scales hidden in the original data, which is beneficial for fault classification. In addition, data corruption using multiple noise levels actually generates artificial data to expand the data size, which can alleviate the problem of small dataset size to some extent.

[0065] Figure 6 A schematic diagram of a migration identification model provided in an embodiment of the present application. Figure 6 The model shown collects vibration signals of different components of each wind turbine from the SCADA system. The vibration signals are preliminarily processed and appropriately transformed. The source domain vibration signal data is input into the SSMDAE to train the SSMDAE. At the same time, the source domain vibration signal and the target domain vibration signal data are simultaneously input into the trained SSMDAE, and DJPMMD is added to the hidden layer of each AE of the SSMDAE. The sum of the errors of DJPMMD and SSMDAE is used as the new error to train the SSMDAE. The retrained SSMDAE is used in the target domain, and the final output is classified.

[0066] It should be understood that the systems shown in the figure are exemplary and non-restrictive. This means that the system architecture involved is not limited to a specific form or design, but is presented as an example. In other words, the architecture shown in the figure can be regarded as a way of expression to clearly describe related concepts and relationships, and does not exclude other forms of architecture. Therefore, when interpreting the architecture in the picture, it should be understood that the model is flexible and diverse, and its purpose is to provide an exemplary description rather than a restrictive provision on a specific form.

[0067] Figure 7A schematic diagram of migration recognition performance provided by an embodiment of the present application. The generalization ability of the model is tested in the experiment, which is mainly divided into two parts: one is the generalization ability between data sets measured at different locations, which mainly analyzes the generalization ability of the model between different components. The other is the generalization ability between different fault data sets measured at the same location, which mainly analyzes the generalization ability of the model between different faults.

[0068] First, the generalization ability of the model between data sets collected at different locations is analyzed. The specific research content is shown in Table 1. The DE, FE, and BA data sets are divided into 70% training sets and 30% test sets. The model is trained with the training set of one domain, and the classification performance of the model is tested with the test set of another domain to measure the generalization ability of the model.

[0069] Table 1 Migration experiment information

[0070] To avoid contingency, all experiments were repeated 10 times, the maximum and minimum values ​​were removed and the average was taken. The classification results of SSMDAE without DJMMMD are shown in Figure 7 As shown, since the BA side does not collect normal data sets, experiments 7, 8 and 9 only use 9 types of fault data.

[0071] from Figure 7 As can be seen from (A) in the figure, when the source domain and the target domain are from the same dataset, that is, experiments 1, 5 and 9, the classification accuracy of these experiments is above 94%. However, when the source domain and the target domain are from different datasets, the classification accuracy does not exceed 60%, which indicates that the model has poor generalization ability between datasets obtained at different locations.

[0072] Figure 7 (B) in the figure is the classification accuracy of the proposed method in experiments 1-9. Figure 7 As can be seen in (B), the classification accuracy of the model between different domains has increased by more than 35%. Among them, after the model trained on BA is migrated to DE, it has a classification accuracy of 94.88%, which is comparable to BA itself. This shows that the proposed method architecture can significantly improve the generalization ability of the model between data sets collected at different locations and improve the robustness of the model.

[0073] Furthermore, an embodiment of the present application also provides a wind turbine group transfer learning fault diagnosis device based on CC-FOA, including: a processor, a memory, and a system bus; the processor and the memory are connected through the system bus; the memory is used to store one or more programs, and the one or more programs include instructions, and when the instructions are executed by the processor, the processor executes any of the above methods.

[0074] Furthermore, an embodiment of the present application also provides a computer program product, which, when executed on a terminal device, enables the terminal device to execute any of the above methods.

[0075] It can be known from the description of the above implementation mode that those skilled in the art can clearly understand that all or part of the steps in the above-mentioned embodiment method can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a storage medium such as ROM / RAM, a disk, an optical disk, etc., including several instructions for enabling a computer device (which can be a personal computer, a server, or a network communication device such as a media gateway, etc.) to execute the methods described in the various embodiments of the present application or certain parts of the embodiments.

[0076] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0077] It should also be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0078] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A wind turbine fault diagnosis method based on CC-FOA transfer learning, characterized in that: include: Collecting vibration signal data of multiple wind turbine generator sets, and dividing the vibration signal data into source domain data and target domain data; Preprocessing the source domain data and the target domain data, including signal noise reduction and dimensionality transformation; Use the preprocessed source domain data to train the stacked sparse denoising autoencoder network; CC-FOA is used to optimize the number of hidden layers and neurons in each layer of the stacked sparse denoising autoencoder network; Input the preprocessed source domain data and target domain data into an optimized stacked sparse denoising autoencoder network, add a joint probability maximum mean difference constraint to each hidden layer, and retrain the network based on the sum of errors; The retrained stacked sparse denoising autoencoder network is applied to the target domain data to extract features and perform fault classification.

2. The method according to claim 1, characterized in that in, CC-FOA is used to optimize the number of hidden layers and neurons in each layer of the stacked sparse denoising autoencoder network, including: Set the fruit fly population size, maximum number of iterations, error accuracy, acceleration factor, and vertical crossover probability; convert the initial population of fruit flies into the number of hidden layers and neurons in each layer of the autoencoder for training, and calculate the reconstruction error as the fitness value of the fruit flies; Sort the fitness values ​​and individual positions, and record the global optimal value and the corresponding optimal individual ; All fruit flies in the population are randomly combined without duplication, crossover is performed, the fitness values ​​of the fruit flies are calculated, and the fruit flies with the best fitness in the current population are retained; Each dimension of the fruit flies obtained by the horizontal crossover is normalized, and then all dimensions are randomly paired without duplication, and a vertical crossover operation is performed. After denormalization, the fitness values ​​are compared, and the global optimal value is updated. and the corresponding optimal individual ; The algorithm continues to iterate until the maximum number of iterations is reached or the error accuracy reaches a preset threshold. The fruit fly with the best fitness is selected as the optimal solution and output. The optimal solution is the optimal number of hidden layers and the number of neurons in each layer of the stacked sparse denoising autoencoder.

3. The method according to claim 1, characterized in that in, Inputting the preprocessed source domain data and target domain data into the optimized stacked sparse denoising autoencoder network, adding a joint probability maximum mean difference constraint to each hidden layer, and retraining the network based on the sum of errors, including: Input the preprocessed source domain data and target domain data into the stacked sparse denoising autoencoder network; The maximum mean difference is used to measure the distribution difference between the source domain and the target domain, and the maximum mean difference constraint of the joint probability including the marginal distribution and the conditional distribution is constructed; Embedding the joint probability maximum mean difference into the hidden layer of each sparse denoising autoencoder to transfer the model layer by layer; The stacked sparse denoising autoencoder network is trained to obtain the optimal parameter set by minimizing the loss function including the reconstruction error and the maximum mean difference of the joint probability.

4. The method according to claim 3, characterized in that in, The maximum mean difference is used to measure the distribution difference between the source domain and the target domain, and the joint probability maximum mean difference constraint containing marginal distribution and conditional distribution is constructed, including: Establish a joint probability difference model between the source domain and the target domain, including the same transferability measure and the different distinguishability measure to define the maximum mean difference of the discriminant joint probability; Constructing a label matrix of source domain data and a pseudo label matrix of target domain data, and calculating the first and second normalized matrices; Generate the extended label matrices of the source domain and the target domain to calculate the third and fourth normalization matrices; An optimization objective function for determining the maximum mean difference of the joint probability is constructed based on the first, second, third and fourth normalized matrices to solve the optimal linear mapping.

5. The method according to claim 3, characterized in that: in, The joint probability maximum mean difference is embedded into the hidden layer of each sparse denoising autoencoder to transfer the model layer by layer, including: The maximum mean difference of the joint probability is embedded into the hidden layer of each sparse denoising autoencoder, and the source domain data and the target domain data are feature extracted by minimizing the loss function of the first sparse denoising autoencoder to align the feature space; The output of the first sparse denoising autoencoder is used as the input of the second sparse denoising autoencoder, and the distribution difference between the source domain data and the target domain data in the new feature space is reduced by minimizing the loss function of the second sparse denoising autoencoder; The output of the second sparse denoising autoencoder is used as the input of the third sparse denoising autoencoder. By minimizing the loss function of the third sparse denoising autoencoder, the source domain data and the target domain data are fully aligned in the common feature space to perform model migration.

6. The method according to claim 4, characterized in that in, Establish a joint probability difference model between the source domain and the target domain, including the same transferability measure and the heterogeneous distinguishability measure to define the maximum mean difference of the discriminant joint probability, including: A joint probability difference model of source domain and target domain is established, which includes the transferability metric of the same class and the distinguishability metric of different classes. The transferability metric of the same class uses the maximum mean difference (MMD) to calculate the transferability from the source domain to the target domain. Based on the empirical formula, the product formula of the conditional expectation and prior probability of the source domain and the target domain is obtained, which is substituted into the maximum mean difference (MMD) formula to obtain the final transferability measurement formula from the source domain to the target domain. A heterogeneous distinguishability measurement model is established, and the maximum mean difference (MMD) is used to calculate the distinguishability between different fault categories in the source domain and the target domain, and the final source domain and target domain heterogeneous distinguishability measurement formula is obtained; Substitute the obtained transferability measurement formula of the same class and the discriminability measurement formula of the different class into the established joint probability difference model to obtain the complete definition of the maximum mean difference of the discriminant joint probability.

7. The method according to claim 6, characterized in that in, Transferability metrics, including: ， in, represents the transferability metric; C represents the total number of categories, that is, the number of fault types; c represents the current category index, ranging from 1 to C; Represents the total number of samples of source domain data; Represents the number of samples belonging to category c in the source domain; Represents the total number of samples of target domain data; Represents the number of samples belonging to category c in the target domain; Represents the transpose of the linear mapping matrix A, which is used to map data from the original feature space to the new feature space; represents the iith sample belonging to category c in the source domain; represents the jjth sample in the target domain belonging to category c.

8. The method according to claim 7, characterized in that in, Discrimination measures, including: ， in, represents a discriminative measure; represents another category index, ranging from 1 to C, and ; Indicates that the target domain belongs to the category The number of samples; Indicates that the jjth item in the target domain belongs to the category of the sample.

9. A wind turbine fault diagnosis device based on CC-FOA transfer learning, characterized in that: include: A processor, a memory, and a system bus; wherein the processor and the memory are connected via the system bus; The memory is used to store one or more programs, wherein the one or more programs include instructions, and when the instructions are executed by the processor, the processor executes the method according to any one of claims 1 to 8.

Images