Wind turbine state recognition model training method, recognition method, equipment and medium

Abstract

The invention provides a wind turbine state recognition model training method, a recognition method, equipment and a medium, and relates to the technical field of data processing, the method expands the coverage range of parameter spaces by using a plurality of mutually orthogonal initialized parameter spaces, calculates loss in each parameter space, and improves the recognition efficiency of the wind turbine state recognition model. And normalizing a loss result to determine a combination coefficient of each parameter space so as to form a joint parameter space for meta-level updating, and updating the joint parameter space in an inner loop. In the outer circulation, each parameter space can be updated in sequence based on the performance of the updated gradient updating parameter space on the task query set. The parameter space used by the target recognition model has better generalization ability compared with any target parameter space, the target recognition model with a better recognition effect can be obtained only through a small amount of target domain training data, and the generalization ability of the model can be further improved by adopting samples of a plurality of working conditions of the wind turbine for training.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a wind turbine state recognition model training method, recognition method, equipment and medium. Background Art

[0002] As a clean and renewable energy source, the development and utilization of wind energy is of great significance for reducing greenhouse gas emissions and responding to climate change. In the context of the global energy structure transitioning to a low-carbon one, wind turbines are playing an increasingly important role in the field of new energy. In the daily management of wind turbines, state identification is crucial for operation and maintenance. It can assess the health status and power generation capacity of wind turbines, so as to formulate maintenance strategies and technical transformation methods in a timely manner, reduce downtime, prevent major failures, and save maintenance costs. However, the scarcity of wind turbine fault states in actual projects makes it very difficult to obtain large-scale labeled fault state data, and the limited data makes it impossible to train a robust model well. In addition, since the monitoring data may come from different operating conditions of the wind turbine, its operating conditions may involve the coupling of multiple factors, such as wind speed, temperature, altitude, etc., which brings additional challenges to state identification. Summary of the invention

[0003] The present application provides a wind turbine state recognition model training method, recognition method, equipment and medium, which can eliminate the problem of poor recognition effect caused by large data distribution differences and small number of samples in the wind turbine state recognition process due to the fact that the wind turbine parameter signals collected in actual projects come from different wind turbine operating conditions.

[0004] The present application provides a wind turbine state recognition model training method, comprising:

[0005] Obtain a recognition model to be trained, obtain source domain training samples for training the recognition model, obtain an initialized parameter space set, the parameter space set includes multiple parameter spaces of the recognition model, the source domain training samples include a support set and a query set, the source domain training samples include multiple wind turbine parameter signals, and the wind turbine status is marked as a label in the source domain training samples;

[0006] Execute an iterative training step, the iterative training step comprising: sampling multiple batches of target training samples in the source domain training samples based on a preset number of support set samples and a number of query set samples, each batch of target training samples corresponding to a different working condition of the wind turbine; iteratively training multiple parameter spaces of the recognition model based on the multiple batches of target training samples, determining the loss of each parameter space in the parameter space set in the support set in an inner loop of the iterative training, normalizing each loss to obtain a combination coefficient, combining each initialized parameter space into a joint parameter space based on the combination coefficient, and performing gradient update on the joint parameter space based on the support set to obtain a gradient updated parameter space; in an outer loop of the iterative training, determining the loss of the gradient updated parameter space in the query set to obtain a first loss, determining a second loss based on the current number of training rounds and the first loss, determining the sum of the second losses of each batch of target training samples as a third loss, updating the parameter space in the parameter space set based on the third loss in a polling manner, and obtaining the parameter space after the current iterative update;

[0007] When the current number of training rounds is equal to the preset number of training rounds, the training is stopped to obtain multiple target parameter spaces after the training is completed;

[0008] Obtain target domain training samples;

[0009] Based on multiple target parameter spaces and the target domain training samples, a target recognition model is determined.

[0010] In one embodiment of the present application, obtaining an initialized parameter space set includes:

[0011] Each parameter space in the parameter space set is initialized based on the Schmidt orthogonalization method to obtain multiple parameter spaces.

[0012] In one embodiment of the present application, each loss is normalized to obtain a combination coefficient, including:

[0013] The loss of each parameter space in the parameter space set is normalized based on a nonlinear activation function to obtain a plurality of space coefficients, and a set of the plurality of space coefficients is determined as a combination coefficient.

[0014] In one embodiment of the present application, the gradient updating of the joint parameter space based on the support set to obtain the gradient updated parameter space includes:

[0015] loading the joint parameter space into the recognition model to determine a fourth loss of the joint parameter space on a support set;

[0016] Based on the joint parameter space, a preset learning rate of the base learner, and a fourth loss, a gradient update is performed to obtain the gradient updated parameter space.

[0017] In one embodiment of the present application, determining the second loss based on the current number of training rounds and the first loss includes:

[0018] When the current number of training rounds is an integer multiple of a preset number of regularization rounds, the dot product of each initial parameter space is determined as a regularization loss, and the sum of the first loss and the regularization loss is determined as a second loss, and the preset number of regularization rounds is less than the preset number of training rounds;

[0019] When the current number of training rounds is not an integer multiple of a preset number of regularization rounds, the first loss is determined as the second loss.

[0020] In one embodiment of the present application, a target recognition model is determined based on multiple target parameter spaces and the target domain training samples, including:

[0021] Determining a loss value of each target parameter space based on the target domain training samples;

[0022] Normalize the loss values ​​of each target parameter space to obtain the target combination coefficient;

[0023] combining the target parameter spaces into a target joint parameter space based on the target combination coefficient;

[0024] Performing gradient update on the target joint parameter space based on the target domain training samples to obtain a target model parameter space;

[0025] The target model parameter space is loaded into the recognition model to obtain the target recognition model.

[0026] In one embodiment of the present application, updating the parameter space in the parameter space set based on the third loss in a polling manner includes:

[0027] In a polling manner, based on the parameter space in the parameter space set, the preset learning rate of the meta-learner, and the third loss, a gradient update is performed to obtain the parameter space after the current iterative update.

[0028] To achieve the above objectives and other related objectives, the present application also provides a method for identifying a wind turbine state, comprising:

[0029] Obtaining wind turbine parameters to be identified;

[0030] Inputting the wind turbine parameters to be identified into the target identification model to obtain the wind turbine state;

[0031] The target recognition model is a target recognition model trained by one or more of the aforementioned training methods.

[0032] To achieve the above objectives and other related objectives, the present application also provides an electronic device, the electronic device comprising:

[0033] one or more processors;

[0034] a memory for storing program code executable by the processor;

[0035] The processor is configured to execute the program code to implement one or more of the methods described above.

[0036] To achieve the above objectives and other related objectives, the present application also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor of a computer, the computer executes one or more of the methods described above.

[0037] As described above, the wind turbine state recognition model training method, recognition method, device and medium provided by the present application have the following beneficial effects:

[0038] A method for training a wind turbine state recognition model in the present application expands the coverage of the parameter space by utilizing multiple mutually orthogonal initialized parameter spaces, calculates the loss in each parameter space, and normalizes the loss result to determine the combination coefficient of each parameter space, thereby forming a joint parameter space for meta-level update, and the joint parameter space is updated in the inner loop. In the outer loop, each parameter space is updated in sequence based on the performance of the updated gradient update parameter space on the task query set. The target recognition model is determined using target domain training samples and multiple target parameter spaces. The parameter space used by the target recognition model has better generalization ability than any target parameter space. Only a small amount of target domain training data is needed to obtain a target recognition model with better recognition effect, and the use of samples of multiple working conditions of wind turbines for training can further improve the generalization ability of the model.

[0039] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The drawings herein are incorporated into the specification and constitute a part of the specification, showing embodiments consistent with the present application, and together with the specification, are used to explain the principles of the present application. Obviously, the drawings described below are only some embodiments of the present application, and for those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0041] Figure 1is a flow chart of a wind turbine state recognition model training method shown in an exemplary embodiment of the present application;

[0042] Figure 2 is a training logic diagram of an inner loop and an outer loop shown in an exemplary embodiment of the present application;

[0043] Figure 3 is a flow chart of a source domain training phase of a wind turbine state recognition model shown in an exemplary embodiment of the present application;

[0044] Figure 4 is a flow chart of a target domain training phase of a wind turbine state recognition model shown in an exemplary embodiment of the present application;

[0045] Figure 5 It is a flow chart of a method for identifying a wind turbine state shown in an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0046] The following will describe the implementation methods of the present application with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. The present application can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be understood that the preferred embodiments are only for illustrating the present application, not for limiting the scope of protection of the present application.

[0047] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present application, and thus the drawings only show components related to the present application rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed at will, and the component layout may also be more complicated.

[0048] In the following description, a large number of details are discussed to provide a more thorough explanation of the embodiments of the present application. However, it is obvious to those skilled in the art that the embodiments of the present application can be implemented without these specific details. In other embodiments, well-known structures and devices are shown in the form of block diagrams rather than in detail to avoid making the embodiments of the present application difficult to understand.

[0049] See also Figure 1 , Figure 1 FIG. 1 is a flow chart of a method for training a wind turbine state recognition model according to an exemplary embodiment of the present application. Figure 1 It can be seen that the wind turbine state recognition model training method may include:

[0050] Step S110, obtaining a recognition model to be trained, obtaining source domain training samples for training the recognition model, and obtaining an initialized parameter space set.

[0051] The parameter space set includes multiple parameter spaces of the recognition model, the source domain training samples include a support set and a query set, the source domain training samples include multiple wind turbine parameter signals, and the wind turbine states are marked as labels in the source domain training samples.

[0052] In one embodiment of the present application, the recognition model to be trained may be an untrained neural network model. When iteratively training the recognition model, multiple batches of target training samples may be sampled from the source domain training samples. A batch of target training samples represents a task, and each task comes from a different distribution domain. The tasks have both common characteristics and their own specific characteristics. The goal of meta-learning is to learn the common characteristics of training tasks from the source domain, that is, general knowledge. When faced with a specific task, only a small number of samples are needed to transfer general knowledge to the specific distribution domain of the task, forming specific knowledge adapted to the task.

[0053] Wind turbine state identification under various operating conditions can be considered a unique meta-learning task. Constrained by actual engineering conditions, multiple training tasks from different data distribution domains can be constructed based on limited data. The recognition model trained by meta-learning can be applied to wind turbine state identification tasks under different operating conditions, rather than being limited to state identification tasks in a single data distribution domain. In addition, when faced with a state identification task that has never been seen before, the recognition model only needs a small amount of data to adapt to achieve satisfactory performance.

[0054] In one possible implementation, obtaining an initialized parameter space set includes: initializing each parameter space in the parameter space set based on the Schmidt orthogonalization method to obtain multiple parameter spaces. Schmidt orthogonalization (Gram-Schmidt Orthogonalization) is a method used in mathematics and numerical analysis to construct a set of orthogonal vectors from a set of linearly independent vectors. This process can be applied to any inner product space, but it is most commonly used in Euclidean space (i.e., a real or complex vector space with a standard inner product). The vectors obtained by Schmidt orthogonalization are not only mutually orthogonal (i.e., the inner product between them is zero), but can also be further normalized so that each vector has a unit length, thereby forming a set of orthogonal bases. Multiple randomly initialized first parameter spaces can be first obtained, and the multiple first parameter spaces can be initialized using Schmidt orthogonalization to obtain multiple parameter spaces.

[0055] The support set can be used to update the parameter space of the recognition model so that the recognition model can quickly adapt to new tasks. After the recognition model is quickly adjusted according to the support set, it can be classified on the query set to evaluate the performance of the recognition model.

[0056] The wind turbine parameter signal may include the current, speed, vibration or load of the wind turbine. The wind turbine state may include a normal operating state or a fault state. The fault state may include a stator winding fault, a rotor winding fault, a demagnetization fault, etc. The characteristics of the wind turbine parameter signal are different under each fault of the wind turbine, so the wind turbine state can be obtained by identifying the wind turbine parameter signal. The source domain training samples may include the relevant parameters of the wind turbine running without load and the relevant parameters of the wind turbine running with load.

[0057] It should be noted that the execution subject of this application can be any electronic device with computing capabilities, such as a server, etc. The electronic device can be deployed locally or in the cloud, and this application does not limit it.

[0058] Step S120, executing an iterative training step.

[0059] Among them, the iterative training step includes: sampling multiple batches of target training samples in the source domain training samples based on the preset number of support set samples and the number of query set samples, and each batch of target training samples corresponds to different working conditions of the wind turbine; iteratively training multiple parameter spaces of the recognition model based on the multiple batches of target training samples, in the inner loop of the iterative training, determining the loss of each parameter space in the parameter space set in the support set, normalizing each loss to obtain a combination coefficient, combining each initialized parameter space into a joint parameter space based on the combination coefficient, and gradient updating the joint parameter space based on the support set to obtain a gradient updated parameter space; in the outer loop of the iterative training, determining the loss of the gradient updated parameter space in the query set to obtain a first loss, determining a second loss based on the current number of training rounds and the first loss, determining the sum of the second losses of each batch of target training samples as a third loss, and updating the parameter space in the parameter space set based on the third loss in a polling manner to obtain the parameter space after the current iterative update.

[0060] In one embodiment of the present application, each initialized parameter space can be loaded into the recognition model in turn, and the loss corresponding to each support set can be determined based on a preset loss function. The gradient updated parameter space can be loaded into the recognition model to determine the loss of the gradient updated parameter space in the query set.

[0061] It should be noted that when training and updating the parameter space, the parameter space can be loaded into the network architecture of the recognition model, and then the parameters can be updated on the task. Each batch of target training samples corresponds to different working conditions of the wind turbine. For example, batch one includes various wind turbine parameter signals of the wind turbine under light load and high speed working conditions, and batch two includes various wind turbine parameter signals of the wind turbine under high load and high speed working conditions.

[0062] The inner loop refers to the process of fast adaptation for a specific task. In this process, the model makes multiple gradient updates based on the support set of a task to adjust the parameter space to better adapt it to the task. The goal of the inner loop is to find a set of parameters that can make the model perform well on the task with a small number of training samples by minimizing the loss function on the support set. The outer loop refers to the process of meta-learning, in which the model can adjust its initial parameter space based on the average performance of all tasks. The goal is to find multiple sets of initialization parameters that allow the model to quickly adapt to any new task with a small number of updates in the inner loop.

[0063] In a possible implementation, normalizing each loss to obtain a combination coefficient includes: normalizing the loss of each parameter space in the parameter space set based on a nonlinear activation function to obtain multiple space coefficients, and determining the set of multiple space coefficients as the combination coefficient. The nonlinear activation function may be a Softmax function.

[0064] For example, the parameter space set to be initialized can be defined as N is the number of parameter spaces, θ (n) The process of normalizing the loss of each parameter space in the parameter space set based on the nonlinear activation function to obtain multiple space coefficients and determining the set of multiple space coefficients as the combination coefficient may include:

[0065]

[0066] in, is the nth spatial coefficient when using the target training sample of the i-th batch, The loss determined by using the target training samples of the ith batch when loading the nth parameter space for the recognition model. The combination coefficients corresponding to the target training samples of the ith batch can be represented as

[0067] In a possible implementation, the process of combining the initialized parameter spaces into a joint parameter space based on the combination coefficient may include:

[0068]

[0069] in, is the joint parameter space.

[0070] In one possible implementation, gradient updating is performed on the joint parameter space based on the support set to obtain a gradient updated parameter space, including: loading the joint parameter space into a recognition model to determine a fourth loss of the joint parameter space on the support set; and performing a gradient update based on the joint parameter space, a preset learning rate of a base learner, and the fourth loss to obtain a gradient updated parameter space.

[0071] Exemplarily, the process of obtaining the gradient update parameter space may include:

[0072]

[0073] in, The gradient update parameter space corresponding to the target training sample of the i-th batch, S i represents the support set of the target training samples in the i-th batch, is the fourth loss, and α is the learning rate of the preset base learner.

[0074] In a possible implementation, determining the second loss based on the current number of training rounds and the first loss includes: when the current number of training rounds is an integer multiple of a preset number of regularization rounds, determining the dot product of each initial parameter space as the regularization loss, determining the sum of the first loss and the regularization loss as the second loss, and the preset number of regularization rounds is less than the preset number of training rounds; when the current number of training rounds is not an integer multiple of the preset number of regularization rounds, determining the first loss as the second loss.

[0075] Exemplarily, the process of determining the sum of the first loss and the regularization loss as the second loss may include:

[0076]

[0077] Among them, Q i is the query set of target training samples in the i-th batch, The loss of the gradient update parameter space in the query set is also the first loss. is the regularization loss, For the second loss.

[0078] It should be noted that regularizing the loss can enhance the diversity of the parameter space and improve the generalization ability, which helps to expand the domain adaptation space.

[0079] In one possible implementation, the parameter space in the parameter space set is updated based on the third loss in a polling manner, including: performing a gradient update in a polling manner based on the parameter space in the parameter space set, a preset learning rate of the meta-learner, and the third loss to obtain the parameter space after the current iterative update.

[0080] Exemplarily, the sum of the second losses of each batch of target training samples is determined as the third loss, which can be recorded as for The third loss of the target training samples of the batch is the cumulative total loss. The parameter spaces in the initialized parameter space set are polled based on the cumulative total loss Update, each update formula is:

[0081]

[0082] Among them, θ (n) ' is the parameter space after the current iteration update, is the third loss, and β is the preset learning rate of the meta-learner.

[0083] Step S130, stopping the training when the current number of training rounds is equal to the preset number of training rounds, and obtaining a plurality of target parameter spaces after training.

[0084] In one embodiment of the present application, a preset number of training rounds may be set before iterative training, and training may be stopped when the current number of training rounds is equal to the preset number of training rounds, thereby obtaining a plurality of target parameter spaces after training.

[0085] Step S140, obtaining target domain training samples.

[0086] In one embodiment of the present application, the data distribution of the source domain and the target domain is often not exactly the same. Obtaining target domain training samples helps the recognition model understand and adapt to the difference in data distribution, ensuring that the recognition model is still effective in the new environment. The target domain training samples can be parameter signals corresponding to the wind turbine that needs to be state recognized, and the parameter signals can include parameter signals such as the current, speed, vibration, and load of the wind turbine.

[0087] Step S150, determining a target recognition model based on multiple target parameter spaces and target domain training samples.

[0088] In one embodiment of the present application, a target recognition model may be determined based on multiple target parameter spaces and target domain training samples.

[0089] In one embodiment, the process of determining the target recognition model in step S150 based on multiple target parameter spaces and the target domain training samples may include steps S151 to S155.

[0090] Step S151, determining the loss value of each target parameter space based on the target domain training samples.

[0091] In one embodiment of the present application, the loss value of each target parameter space on the target domain training sample support set can be determined to obtain a set of loss values ​​corresponding to all target parameter spaces:

[0092] Step S152, normalize the loss value of each target parameter space to obtain the target combination coefficient.

[0093] In one embodiment of the present application, the loss value of each target parameter space may be normalized by applying a Softmax nonlinear activation function to obtain a target combination coefficient.

[0094]

[0095] Among them, σ (n) is the nth spatial coefficient when using the target domain training sample, The loss value determined when the nth target parameter space is loaded for the recognition model. The target combination coefficient corresponding to the target domain training sample can be represented as

[0096] Step S153, combining the target parameter spaces into a target joint parameter space based on the target combination coefficient.

[0097] In one embodiment of the present application, each target parameter space may be combined into a joint parameter space based on a target combination coefficient.

[0098]

[0099] Among them, θ ＊ is the target joint parameter space, θ (n)＊ is the nth target parameter space.

[0100] Step S154, performing gradient update on the target joint parameter space based on the target domain training samples to obtain the target model parameter space.

[0101] In one embodiment of the present application, the target joint parameter space may be gradient updated based on the support set of the target domain training samples to obtain the target model parameter space.

[0102]

[0103] Among them, φ ★ is the target model parameter space, S represents the support set of the target domain training samples, is the loss of the target joint parameter space on the support set of the target domain training samples, and α is the learning rate of the preset base learner.

[0104] Step S155, loading the target model parameter space into the recognition model to obtain the target recognition model.

[0105] In one embodiment of the present application, the target model parameter space may be loaded into the recognition model to obtain a target recognition model.

[0106] For example, see Figure 2 , which is a training logic diagram of the inner loop and the outer loop shown in an exemplary embodiment of the present application. Among them, the solid line represents the inner loop, the dotted horizontal dashed line represents the outer loop, and the dotted dashed line represents the update of each parameter space to the joint parameter space. When faced with the task of wind turbine state recognition, the difference in data distribution between the source domain and the target domain usually poses a challenge. In traditional model-independent meta-learning methods, two-layer updates are only applied to a single parameter space, which may lead to insufficient coverage of the task space and cause the optimization to remain at a local optimum. The multi-parameter space exploration and expansion of the task space coverage provided in the embodiment of the present application are an effective solution to improve the generalization ability of the model.

[0107] The recognition model training method provided in the embodiment of the present application also follows the meta-learning task framework, allowing training in limited data and multi-source domain scenarios. By utilizing multiple mutually orthogonal initialized parameter spaces, the coverage of the task space is expanded. In a specific task, the embodiment of the present application calculates the loss in each model parameter space, and applies a nonlinear activation function to the loss result to determine the combined weights of each parameter space, thereby forming a joint parameter space for meta-level updates. The joint parameter space is updated in the inner loop. In the outer loop, each initial parameter space is updated in turn based on the performance of the updated gradient update parameter space on the task query set. During the update process, the scheme of the embodiment of the present application introduces the dot product of each parameter space as a regularization term to enhance the diversity of the parameter space, thereby helping to expand the domain adaptation space.

[0108] For example, see Figure 3 , which is a flow chart of the source domain training phase of the wind turbine state recognition model shown in an exemplary embodiment of the present application. c is the current training round number.

[0109] For example, see Figure 4 , which is a flow chart of a target domain training phase of a wind turbine state recognition model shown in an exemplary embodiment of the present application. A plurality of target parameter spaces constitute a target parameter space set.

[0110] Figure 5 FIG. 1 is a flow chart of a method for identifying a wind turbine state, shown in an exemplary embodiment of the present application. Figure 5 As shown, the exemplary wind turbine state identification method may include:

[0111] Step S510, obtaining parameters of the wind turbine to be identified.

[0112] In one embodiment of the present application, the wind turbine parameters to be identified may be obtained. The wind turbine parameters to be identified may include parameter signals such as current, rotation speed, vibration, load, etc. of the wind turbine to be identified.

[0113] Step S520: input the wind turbine parameters to be identified into the target identification model to obtain the state of the wind turbine.

[0114] Among them, the target recognition model is a target recognition model trained by the training method provided by any of the above embodiments.

[0115] It should be noted that the target recognition model can be used to identify the state of the wind turbine, and the target model parameter space has a better recognition effect than any target parameter space.

[0116] An embodiment of the present application also provides an electronic device, including: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by one or more processors, the electronic device implements the methods provided in the above embodiments.

[0117] Another aspect of the present application further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor of a computer, the computer executes the methods provided in the above embodiments. The computer-readable storage medium may be included in the electronic device described in the above embodiments, or may exist independently without being assembled into the electronic device.

[0118] Another aspect of the present application also provides a computer program product or a computer program, which includes a computer instruction stored in a computer-readable storage medium. A processor of a computer device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the method provided in each of the above embodiments.

[0119] In the embodiments of the present application, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance. "Including" and "comprising" mentioned throughout the specification and claims are open-ended terms and should be interpreted as "including but not limited to".

[0120] The above embodiments are merely illustrative of the principles and effects of the present application, and are not intended to limit the present application. Anyone familiar with the technology may modify or change the above embodiments without violating the spirit and scope of the present application. Therefore, all equivalent modifications or changes made by a person of ordinary skill in the art without departing from the spirit and technical ideas disclosed in the present application shall still be covered by the claims of the present application.

Claims

1. A wind turbine state recognition model training method, characterized in that: include: Obtain a recognition model to be trained, obtain source domain training samples for training the recognition model, obtain an initialized parameter space set, the parameter space set includes multiple parameter spaces of the recognition model, the source domain training samples include a support set and a query set, the source domain training samples include multiple wind turbine parameter signals, and the wind turbine status is marked as a label in the source domain training samples; Execute an iterative training step, the iterative training step comprising: sampling multiple batches of target training samples in the source domain training samples based on a preset number of support set samples and a number of query set samples, each batch of target training samples corresponding to a different working condition of the wind turbine; iteratively training multiple parameter spaces of the recognition model based on the multiple batches of target training samples, determining the loss of each parameter space in the parameter space set in the support set in an inner loop of the iterative training, normalizing each loss to obtain a combination coefficient, combining each initialized parameter space into a joint parameter space based on the combination coefficient, and performing gradient update on the joint parameter space based on the support set to obtain a gradient updated parameter space; in an outer loop of the iterative training, determining the loss of the gradient updated parameter space in the query set to obtain a first loss, determining a second loss based on the current number of training rounds and the first loss, determining the sum of the second losses of each batch of target training samples as a third loss, updating the parameter space in the parameter space set based on the third loss in a polling manner, and obtaining the parameter space after the current iterative update; When the current number of training rounds is equal to the preset number of training rounds, the training is stopped to obtain multiple target parameter spaces after the training is completed; Obtain target domain training samples; Based on multiple target parameter spaces and the target domain training samples, a target recognition model is determined.

2. The wind turbine state recognition model training method according to claim 1 is characterized in that: Get the initialized parameter space set, including: Each parameter space in the parameter space set is initialized based on the Schmidt orthogonalization method to obtain multiple parameter spaces.

3. The wind turbine state recognition model training method according to claim 1, characterized in that: Normalize each loss to get the combination coefficient, including: The loss of each parameter space in the parameter space set is normalized based on a nonlinear activation function to obtain a plurality of space coefficients, and a set of the plurality of space coefficients is determined as a combination coefficient.

4. The wind turbine state recognition model training method according to claim 1, characterized in that: Performing gradient updating on the joint parameter space based on the support set to obtain a gradient updated parameter space includes: loading the joint parameter space into the recognition model to determine a fourth loss of the joint parameter space on a support set; Based on the joint parameter space, a preset learning rate of the base learner, and a fourth loss, a gradient update is performed to obtain the gradient updated parameter space.

5. The wind turbine state recognition model training method according to claim 1, characterized in that: The second loss is determined based on the current training round number and the first loss, including: When the current number of training rounds is an integer multiple of a preset number of regularization rounds, the dot product of each initial parameter space is determined as a regularization loss, and the sum of the first loss and the regularization loss is determined as a second loss, and the preset number of regularization rounds is less than the preset number of training rounds; When the current number of training rounds is not an integer multiple of a preset number of regularization rounds, the first loss is determined as the second loss.

6. The wind turbine state recognition model training method according to claim 1, characterized in that: Determining a target recognition model based on multiple target parameter spaces and the target domain training samples includes: Determining a loss value of each target parameter space based on the target domain training samples; Normalize the loss values ​​of each target parameter space to obtain the target combination coefficient; combining the target parameter spaces into a target joint parameter space based on the target combination coefficient; Performing gradient update on the target joint parameter space based on the target domain training samples to obtain a target model parameter space; The target model parameter space is loaded into the recognition model to obtain the target recognition model.

7. The wind turbine state recognition model training method according to claim 1, characterized in that: The parameter spaces in the parameter space set are updated based on the third loss in a polling manner, including: In a polling manner, based on the parameter space in the parameter space set, the preset learning rate of the meta-learner, and the third loss, a gradient update is performed to obtain the parameter space after the current iterative update.

8. A method for identifying a wind turbine state, characterized in that: include: Obtaining parameters of the wind turbine to be identified; Inputting the wind turbine parameters to be identified into the target identification model to obtain the wind turbine state; Wherein, the target recognition model is a target recognition model trained by the training method described in any one of claims 1-7.

9. An electronic device, characterized in that: The electronic device comprises: one or more processors; a memory for storing program code executable by the processor; The processor is configured to execute the program code to implement the method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by a processor of a computer, the computer is caused to execute the method according to any one of claims 1 to 8.

Images