Robust fan fault early warning model training method based on double network collaborative optimization

Abstract

The application belongs to the technical field of fan fault early warning, and particularly relates to a robust fan fault early warning model training method based on double-network collaborative optimization, comprising: obtaining a to-be-tested sample of a fan system, constructing a preliminary fault early warning intelligent agent based on Markov attributes according to fault early warning task requirements based on the to-be-tested sample, the preliminary fault early warning intelligent agent comprising an ontology network and a target network; introducing a method based on adaptive action sampling to generate a decision sample; performing hierarchical progressive isomorphic filtering processing on state features corresponding to the decision sample to extract filtered features with multi-scale and multi-fine-granularity information; performing anti-noise modeling on the filtered features to obtain anti-noise features; introducing a feature weighting method of an adaptive weight vector to the anti-noise features to obtain enhanced features; using the enhanced features, the model parameters are updated and optimized in a way of collaborative optimization of the ontology network and the target network to obtain a robust fan fault early warning model; and the robustness of the model is improved.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine fault early warning technology, specifically involving a robust wind turbine fault early warning model training method based on dual-network collaborative optimization. Background Technology

[0002] The safety and reliability of wind turbine operation and maintenance are of paramount importance, directly impacting the safety of wind farm operations and the stability of power supply. However, wind turbines operate in harsh environments, continuously enduring complex and variable wind loads, climatic conditions, and electromagnetic interference. During long-term operation, critical components such as blades, main shafts, gearboxes, generators, and nacelle structures may experience various failure modes, including imbalance, fatigue damage, loosening, or performance degradation. Therefore, timely and effective fault warnings for wind turbine operation are crucial for ensuring the safe operation of wind power equipment.

[0003] In actual operating conditions, wind turbines operate in complex environments, and their monitoring signals are often accompanied by strong background noise and non-stationary disturbances. Early faults often manifest as weak amplitudes, scattered feature distributions, and are easily mixed with normal operating signals, making it difficult to effectively extract key fault information. Traditional fault early warning methods that rely on empirical rules, signal processing, or single feature extraction are sensitive to noise interference and struggle to balance early warning stability and accuracy under low signal-to-noise ratio conditions.

[0004] In recent years, data-driven intelligent fault early warning methods have been increasingly applied to the field of wind turbine condition monitoring, enabling automatic analysis of operational data and fault identification through the construction of learning models. However, some existing methods suffer from complex network structures, large parameter scales, and high requirements for computing resources and deployment environments. Furthermore, they still exhibit problems such as unstable feature representation and insufficient model generalization ability under conditions of strong noise or drastic changes in operating conditions. Fault early warning methods based on reinforcement learning provide a new technical path for the adaptive optimization of wind turbine fault early warning strategies by modeling the fault discrimination process as a sequential decision problem. However, existing methods still have shortcomings in terms of action sampling strategy design, feature noise resistance, and training process stability, making it difficult to achieve efficient and reliable fault early warning in complex noise environments.

[0005] Furthermore, during the model training phase, the lack of a stable objective constraint mechanism and adaptive parameter optimization strategy can easily lead to oscillations or performance degradation during model convergence. In the actual operation phase of the model, the lack of a continuous evaluation and dynamic update mechanism for the model's early warning performance will also limit the ability of the fault early warning method to adapt to changes in wind turbine operating conditions and new fault types.

[0006] Therefore, there is an urgent need for an intelligent wind turbine fault early warning method that can be used for the overall operation of wind turbines, operate stably under strong noise and complex operating conditions, and take into account both lightweight models and accurate fault early warning. Summary of the Invention

[0007] The purpose of this invention is to provide a robust wind turbine fault early warning model training method based on dual-network collaborative optimization, so as to solve the technical problems in the prior art that it cannot be applied to the overall operation of the wind turbine, can not operate stably under strong noise and complex operating conditions, and cannot balance the lightweight nature of the model with the accuracy of fault early warning.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, this invention provides a training method for a robust wind turbine fault early warning model based on dual-network collaborative optimization, characterized in that it includes: Acquire test samples of the wind turbine system, and construct a primary fault early warning agent based on Markov attributes according to the fault early warning task requirements based on the test samples. The primary fault early warning agent includes an ontology network and a target network. In the primary fault warning agent, an adaptive action sampling method is introduced to generate decision samples; the state features corresponding to the decision samples are subjected to hierarchical progressive isomorphic filtering to extract filtered features with multi-scale and multi-fine-grained information. Based on the filtered features, noise resistance modeling is performed to obtain noise resistance features; the noise resistance features are then subjected to an adaptive weight vector feature weighting method to enhance features related to the fault state and suppress redundant or interfering features, resulting in enhanced features; Using the enhanced features, and through the collaborative optimization of the ontology network and the target network, the parameters of the primary fault warning agent are updated based on the global training objective and loss function to obtain the updated fault warning agent; The parameters of the fault warning agent are updated using a gradient optimization method based on Taylor approximation and dynamic adaptive step size. The gradient optimization method based on Taylor approximation and dynamic adaptive step size includes: calculating the multivariate gradient vector of the objective function, determining the parameter update direction based on the first-order Taylor approximation, and dynamically adjusting the update step size through an adaptive factor that includes a preheating and decay mechanism to obtain a robust wind turbine fault warning model.

[0009] Preferably, the sample to be tested is constructed into a primary fault warning agent based on Markov properties according to the fault warning task requirements, including: Based on the requirements of fault early warning tasks, a task attribute mapping rule for reinforcement learning agents is established to match the agent's operating status, fault type discrimination results, and number of fault categories. A reward feedback mechanism is constructed based on the fault type identification results. The reward value when the fault type is correctly identified is set to a different value than the reward value when the fault type is incorrectly identified, thus forming an instant reward function based on diagnostic accuracy. The test samples are randomly shuffled, and the number of samples of each type of fault is evenly distributed. A training data sampling method for the agent with Markov properties is constructed, such that the state transition probability to the next state s' after performing action a in the current state s is given by... Furthermore, the state transition process is independent of the state at the previous moment; Based on the quadruple (s,a,r,s') consisting of the current state s, the action a, the reward value r, and the next state s', a training sample set with Markov decision process characteristics is established, and the quadruple data is stored and randomly sampled in combination with the experience replay mechanism to form an experience replay buffer. When the number of samples in the experience replay buffer reaches a preset condition, samples are randomly selected from the experience replay buffer and input into the ontology network for parameter update, thereby training a primary fault warning agent based on Markov properties.

[0010] Preferably, in the primary fault warning agent, an adaptive action sampling method is introduced to generate decision samples for training the fault warning agent, including: In the primary fault warning agent, an action sampling function is constructed based on an adaptive degradation factor to maximize the action value function. By combining a random sampling strategy with an action sampling function, an action sampling mechanism based on an adaptive degradation mechanism is constructed. The primary fault warning agent adaptively adjusts its training through an action sampling mechanism based on adaptive degradation, generating decision samples for training the fault warning agent; the action sampling mechanism based on adaptive degradation is as follows:

[0011] in, Indicates the initial state; Indicates the state The next action will be selected based on the current strategy; This represents the set of actions that the system can choose from; Indicates from the set of actions Randomly select an action; State-action pairs Action evaluation function; This represents the probability weights of using different action sampling methods at the current time step; Indicates the state The action that maximizes the evaluation function; The probability of the early warning agent performing greedy sampling is represented by the following formula:

[0012] in, This represents the adaptive degradation control factor, used to adjust the proportion of exploration and utilization of the strategy at different training stages.

[0013] Preferably, the step of performing hierarchical progressive isomorphic filtering on the state features corresponding to the decision samples to obtain filtered features includes: The decision samples are decoupled and grouped to obtain the original feature map. The original feature map is then divided into several feature subgroups according to the channel dimension. The progressive output features are determined based on the feature subgroups. Multiple identity mapping functions are used to connect the progressive output features across layers, constructing a multi-level feature representation pyramid structure to obtain the filtered features.

[0014] Preferably, the filtered features are combined with multi-scale and multi-fine-grained features to perform noise-resistant modeling on the original state features to obtain noise-resistant features. Specifically, the filtered features, basic general features, local correlation features and global abstract features are jointly retained and focused on for extraction to obtain noise-resistant features.

[0015] Preferably, the noise resistance feature incorporates an adaptive weight vector feature weighting method to enhance features related to the fault state and suppress redundant or interfering features, resulting in enhanced features, including: Extract the weight vectors corresponding to each channel of the filtered features to obtain the initial weight vectors that reflect the response intensity of different channel features; The initial weight vector is normalized using a feature weight vector extraction method based on the softmax function to obtain channel features; The channel features are weighted to enhance features related to fault states and suppress redundant or interfering features, resulting in enhanced features.

[0016] Preferably, the ontology network is used based on the current system state. With action The input is used to estimate the action value function in real time, and the corresponding action value is output. The model parameters of the ontology network are iteratively updated using the backpropagation algorithm. The target network is used for the next state. The stability of the action value is estimated, and its output action value is calculated. The model parameters of the target network According to the preset training step size period and the model parameters of the ontology network Perform synchronized updates.

[0017] In a second aspect, the present invention provides a robust wind turbine fault early warning model training system based on dual-network collaborative optimization, comprising: A primary fault warning agent construction unit is used to acquire the test samples of the wind turbine system and construct a primary fault warning agent based on Markov properties according to the fault warning task requirements. The primary fault warning agent includes an ontology network and a target network. The feature extraction unit is used to introduce an adaptive action sampling method in the primary fault warning agent to generate decision samples; and to perform hierarchical progressive isomorphic filtering on the state features corresponding to the decision samples to extract filtered features with multi-scale and multi-fine-grained information. The noise reduction enhancement unit is used to perform noise reduction modeling based on the filtered features to obtain noise reduction features; the noise reduction features introduce an adaptive weight vector feature weighting method to enhance features related to the fault state and suppress redundant or interfering features to obtain enhanced features; The first update and optimization unit is used to utilize the enhanced features and, through the collaborative optimization of the ontology network and the target network, update the parameters of the primary fault warning agent based on the global training objective and loss function to obtain the updated fault warning agent. The second update and optimization unit is used to update the parameters of the fault warning agent after the update using a gradient optimization method based on Taylor approximation and dynamic adaptive step size. The gradient optimization method based on Taylor approximation and dynamic adaptive step size includes: calculating the multivariate gradient vector of the objective function, determining the parameter update direction based on the first-order Taylor approximation, and dynamically adjusting the update step size through an adaptive factor that includes a preheating and decay mechanism to obtain a robust wind turbine fault warning model.

[0018] In a third aspect, the present invention provides an electronic device, characterized in that it includes a processor and a memory, wherein the processor is configured to execute a computer program stored in the memory to implement the robust wind turbine fault early warning model training method based on dual-network collaborative optimization as described in any one of the preceding claims.

[0019] In a fourth aspect, the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores at least one instruction, which, when executed by a processor, implements the robust wind turbine fault early warning model training method based on dual-network collaborative optimization as described above.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention models the fault early warning process of the wind turbine drive system as an intelligent agent decision-making process that satisfies Markov properties, enabling the fault early warning model to adaptively optimize the decision-making strategy during the interaction process, thereby improving the method's adaptability to complex working conditions. By introducing knowledge exploration based on adaptive action sampling and utilizing a balancing mechanism to dynamically adjust the action selection method, the model training stability and decision reliability are effectively improved, taking into account both exploration and convergence during the training process. A hierarchical progressive isomorphic filtering method combining multi-scale and multi-fine-grained features is adopted to suppress background noise and enhance the robustness of feature representation in strong noise environment. By using an adaptive weight vector feature weighting method to enhance weak fault features, the problem of fault features being easily interfered with by noise is improved. By constructing a global training objective based on dual-network collaborative training and combining dynamic parameter optimization and model evaluation and update mechanisms, the stability of the model training process and the continuous applicability of fault warning are improved. Attached Figure Description

[0021] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a flowchart of the system fault early warning method according to an embodiment of the present invention; Figure 3 This is a structural diagram of the hierarchical progressive isomorphic filtering method according to an embodiment of the present invention; Figure 4 This is a diagram illustrating the ontology network algorithm framework of an embodiment of the present invention. Figure 5 The following is a visualization result of the features in an embodiment of the present invention; wherein, (a) is an attenuation rate of 0.86; (b) is an attenuation rate of 0.88; (c) is an attenuation rate of 0.90; (d) is an attenuation rate of 0.92; (e) is an attenuation rate of 0.94; (f) is an attenuation rate of 0.96; and (g) is an attenuation rate of 0.98. Figure 6 The diagrams show the confusion matrix results of different methods in the embodiments of the present invention; (a) is the DCNN method; (b) is the Inception method; (c) is the RSNet method; (d) is the ResNet method; (e) is the present application. Figure 7 This is a system structure block diagram according to an embodiment of the present invention; Figure 8 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0023] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0024] See Figure 1 This application discloses a training method for a robust wind turbine fault early warning model based on dual-network collaborative optimization, characterized by comprising: S1: Obtain the test sample of the wind turbine system, and construct a primary fault early warning agent based on Markov properties according to the fault early warning task requirements. The primary fault early warning agent includes an ontology network and a target network. S2: In the primary fault warning agent, an adaptive action sampling method is introduced to generate decision samples; the state features corresponding to the decision samples are subjected to hierarchical progressive isomorphic filtering to extract filtered features with multi-scale and multi-fine-grained information. S3: Based on the filtered features, noise resistance modeling is performed to obtain noise resistance features; the noise resistance features are enhanced by introducing an adaptive weight vector feature weighting method to enhance features related to the fault state and suppress redundant or interfering features, thus obtaining enhanced features; S4: Using the enhanced features, the updated fault warning agent is obtained by updating the parameters of the primary fault warning agent based on the global training objective and loss function through the collaborative optimization of the ontology network and the target network. S5: The parameters of the fault warning agent are updated using a gradient optimization method based on Taylor approximation and dynamic adaptive step size. The gradient optimization method based on Taylor approximation and dynamic adaptive step size includes: calculating the multivariate gradient vector of the objective function, determining the parameter update direction based on the first-order Taylor approximation, and dynamically adjusting the update step size through an adaptive factor that includes a preheating and decay mechanism to obtain a robust wind turbine fault warning model.

[0025] This method uses a primary fault warning agent as its core, employing a replay sampling strategy to construct a state sequence satisfying Markov properties from the wind turbine drive system's operational monitoring data. It introduces an adaptive action sampling strategy to achieve a dynamic balance between knowledge exploration and strategy utilization during the agent's training process. The sampled state features undergo hierarchical progressive isomorphic filtering to enhance their robustness under strong background noise interference. Weak fault features are reinforced through an adaptive weight vector feature weighting method. Furthermore, a global training objective is constructed based on a dual-network collaborative training architecture, and a multivariate gradient dynamic adaptive descent optimization method under Taylor series approximation order control is used to stably optimize model parameters and improve model convergence performance. Finally, during model training and operation, multiple evaluation metrics are used to quantitatively assess the model's fault warning performance, and the model structure and parameters are dynamically updated based on changes in operating conditions or the addition of new fault types. This method can achieve stable and reliable fault warning for wind turbine drive systems in complex noise environments. To address the common challenges faced by drive shaft systems in actual industrial operation, such as strong background noise interference, weak fault characteristics, and limited computational resources, existing fault early warning methods still fall short in terms of feature extraction robustness, model generalization ability, and real-time decision stability. Therefore, this invention proposes a robust wind turbine fault early warning model training method based on dual-network collaborative optimization. This model can achieve highly reliable fault early warning of the drive shaft system's operating status under complex noise environments.

[0026] In some embodiments, see Figure 2 A robust wind turbine fault early warning model training method based on dual-network collaborative optimization includes: A primary fault early warning agent based on Markov properties is constructed. The operation monitoring data of the wind turbine drive system is represented as the agent state, the fault type is defined as the decision action, and a corresponding reward feedback mechanism is established based on the fault discrimination result, thereby forming an agent decision framework for fault early warning. During the training process of the intelligent agent, an adaptive action sampling strategy is introduced to dynamically adjust the action selection method according to the training stage, so as to balance the exploration of unknown states and the utilization of existing strategies, and generate decision samples for model training. The state features corresponding to the decision samples are subjected to hierarchical progressive isomorphic filtering, and the original state features are modeled for noise resistance by combining multi-scale and multi-fine-grained features, so as to improve the robustness of feature representation under strong background noise conditions. After constructing the noise resistance features, an adaptive weight vector feature weighting method is introduced to enhance the weak features related to the fault state and suppress redundant or interfering features. Based on a dual-network collaborative training architecture, the ontology network and the target network are used to predict and constrain the action evaluation results respectively. By constructing a global training objective and calculating the corresponding loss function, the model parameters are iteratively updated to improve the stability of the training process. During the model parameter update process, a multivariate gradient descent optimization method based on Taylor series approximation order control is introduced to achieve adaptive optimization of model parameters; During model training or operation, multiple evaluation indicators are used to quantitatively assess the performance of the fault early warning model. The model structure and parameters are dynamically and incrementally updated according to changes in operating conditions or the addition of new fault types to achieve continuous fault early warning for the wind turbine drive system.

[0027] In some embodiments, the fault warning agent modeling method based on Markov attributes specifically includes: Based on the requirements of fault diagnosis and fault early warning tasks, a task attribute mapping rule for reinforcement learning agents is established, and the state space, action space and reward function of the agents are redefined to match the fault early warning scenario of the drive shaft system. Here, state s is defined as a vibration signal or a feature representation composed of it that characterizes the operating state of the drive shaft system, and action a is defined as the fault type discrimination result. The action is represented in one-hot encoding form, where N is the preset number of fault categories. A reward feedback mechanism is constructed based on the fault identification results. The reward value for correct fault type identification is set to r=1, and the reward value for incorrect fault type identification is set to r= 1. To form an immediate reward function based on diagnostic accuracy; By randomly shuffling the training samples and evenly distributing the number of samples of each type of fault, a training data sampling strategy with Markov properties is constructed for the agent, such that the state transition probability to the next state s' after performing action a in the current state s is... Furthermore, the state transition process is independent of the state at the previous moment, thereby ensuring that the agent's decision-making process satisfies the Markov property. Based on the above-mentioned quadruple (s,a,r,s') consisting of "state-action-reward-next state", a training sample set with Markov decision process characteristics is established, and the quadruple data is stored and randomly sampled in combination with the experience replay mechanism to reduce the temporal correlation between samples and improve sample utilization efficiency. When the number of samples in the experience replay buffer reaches a preset condition, samples are randomly selected from the experience replay buffer and input into the ontology network for parameter updates, so as to realize the training of the fault warning agent based on Markov properties.

[0028] In some embodiments, the knowledge exploration based on adaptive action sampling—utilizing a balancing method—specifically includes: An action sampling function is constructed based on an adaptive degradation factor to maximize the action value function, so that the action selection probability changes dynamically as the agent training process progresses. By combining the random sampling strategy with the action sampling function, an action sampling strategy based on an adaptive degradation mechanism is constructed. In the early stage of training, the probability of exploring non-optimal actions is increased, and in the later stage of training, it gradually transitions to a deterministic sampling method guided by the maximum value of the action value function. By employing the action sampling strategy based on the adaptive degradation mechanism, the exploration and utilization ratio of the agent is adaptively adjusted in different training stages, enabling the agent to accelerate the policy convergence speed and improve the stability and accuracy of fault warning decisions while ensuring full exploration of the global state space.

[0029] The specific expression for this strategy is as follows:

[0030] in, Indicates the initial state; Indicates the state The next action will be selected based on the current strategy; This represents the set of actions that the system can choose from; Indicates from the set of actions Randomly select an action; State-action pairs Action evaluation function; This represents the probability weights of using different action sampling methods at the current time step; Indicates the state The action that maximizes the evaluation function; The probability of the early warning agent performing greedy sampling is represented by the following formula:

[0031] in, This represents the adaptive degradation control factor, used to adjust the proportion of exploration and utilization of the strategy at different training stages.

[0032] In some embodiments, see Figure 3 The hierarchical progressive isomorphic filtering method combining multi-scale and multi-fine-grained features specifically includes: First, the input feature map is decoupled and grouped. The original feature map is divided into several feature subgroups according to the channel dimension to obtain multi-scale feature representations with different receptive fields.

[0033] in, This represents the i-th input feature map. This represents the corresponding progressive output feature. This represents the i-th order nonlinear characteristic mapping function; Multiple identity mapping functions are used to connect the progressive feature pyramid across layers, constructing a multi-level feature representation pyramid structure. Basic general features, local correlation features, and global abstract features are jointly preserved and focused for extraction, which can alleviate the gradient vanishing and feature representation degradation problems in the training process of deep networks and improve the robustness of the model in strong noise environment.

[0034] In some embodiments, the feature weighting method for the adaptive weight vector of weak feature enhancement for noise resistance features specifically includes: The weight vectors corresponding to each channel are extracted based on the input feature map. By calculating the statistical features of each channel feature map, an initial weight vector reflecting the feature response intensity of different channels is obtained. The initial weight vector is normalized using a feature weight vector extraction method based on the softmax function, mapping the weights corresponding to each channel to a probability distribution. The normalization process is as follows:

[0035] in, This represents the initial weight value corresponding to the i-th channel. This represents the total number of channels in the feature map. This represents the normalized channel weight coefficients; Channel features are weighted to strengthen key channels and suppress irrelevant channels, thereby enhancing discriminative features and improving the model's robustness in noisy environments.

[0036] In some embodiments, the global training objective optimization method under the dual-network collaborative training strategy specifically includes: Within the dual-network collaborative training framework, an ontology network and a target network are constructed, wherein: The ontology network is used to base on the current system state. With action The input is used to estimate the action value function in real time, and the corresponding action value is output. The model parameters of the ontology network are iteratively updated using the backpropagation algorithm. See [link to documentation]. Figure 4 ; The target network is used for the next state. The stability of the action value is estimated, and its output action value is calculated. The model parameters of the target network According to the preset training step size period and the model parameters of the ontology network Perform synchronized updates; The dual-network collaborative training process includes: The ontology network is used to calculate the action value output in the current state. And based on this output, participate in the construction of the loss function; The target mapping quantity corresponding to the predicted output is constructed using the target network, which can be expressed as:

[0037] in, Indicates the amount of real-time feedback. This is the attenuation weighting coefficient. Represents a set of actions; Furthermore, the global training objective optimization method includes: Multi-class prediction logits are constructed based on the output of the ontology network, and the logits are normalized and mapped to obtain normalized action evaluation output. To enhance the distribution differences of different fault states in the feature space; By introducing a global optimization constraint based on logits distribution stretching, the output corresponding to the target fault category is maximized relative to the non-target category; Based on the deviation relationship between the normalized prediction output and the target mapping, a total loss function for dual-network collaborative optimization is constructed, which can be expressed as:

[0038] in, Indicates the number of samples used in training; By employing the aforementioned dual-network collaborative training mechanism and a global loss optimization strategy based on logits distribution differences, joint updates of model parameters are achieved, thereby improving the stability and robustness of fault warning in wind turbine drive systems under complex noise environments.

[0039] In some embodiments, the multivariate gradient dynamic adaptive descent optimization method based on Taylor series approximation order control specifically includes: Obtain the partial derivatives of the multivariate objective function, and construct the multivariate gradient vector of the objective function with respect to each parameter component in the high-dimensional parameter space, when the input variables are represented as... When the objective function g is expressed as a multivariate gradient vector, it is as follows:

[0040] Based on the aforementioned multivariate objective function, a first-order Taylor approximation control is applied to describe the variation of the objective function in the neighborhood of the current parameter point. Its first-order Taylor expansion is expressed as:

[0041] in, Represents the parameter perturbation vector; Based on the first-order Taylor approximation, the parameters are perturbed along the negative gradient direction, and the perturbation vector is set as follows. Substituting this into the first-order Taylor expansion, we obtain the equivalent expression for the objective function along the multivariate negative gradient direction:

[0042] Based on the Taylor approximation equation, a dynamic adaptive optimization and update rule for the original parameter vector under the preheating mechanism is constructed, and the parameter update rule is expressed as follows:

[0043] in, This is a dynamic adaptive factor in the parameter update process. During the parameter iteration update process, it adopts an adaptive adjustment and control of the update step size based on a preheating and exponential decay mechanism. The adjustment rule is as follows:

[0044] in, This represents the initial value of the dynamic factor, and t represents the current training time. The value represents the length of the preheating stage of the dynamic factor, and K represents the attenuation coefficient of the dynamic factor. In some embodiments, the method for jointly evaluating model performance and dynamically updating model structure using multiple evaluation metrics specifically includes: The model performance is jointly evaluated based on multiple metrics, including accuracy, precision, recall, and F1 score. The formula for calculating model accuracy is as follows:

[0045] The formula for calculating model accuracy is:

[0046] The formula for calculating model recall is:

[0047] The formula for calculating the F1 score of the model is:

[0048] Wherein, TP represents the number of samples that were correctly identified as the target fault category, FP represents the number of samples that were misclassified as the fault category, FN represents the number of samples that were incorrectly identified as other categories, and TN represents the number of samples that were not the fault category and were correctly identified. When a new fault type is detected, freeze the existing feature extractor. Create a new feature extractor And fine-tuning based on incremental data; During incremental training, only the new feature extractor is updated, enabling the model to quickly adapt to new fault types.

[0049] Example 1 This invention provides a training method for a robust wind turbine fault early warning model based on dual-network collaborative optimization, comprising the following steps: Step 1: Model a basic fault warning agent based on Markov properties; In step one, based on the requirements of fault diagnosis and fault early warning tasks, a task attribute mapping rule for the reinforcement learning agent is established, and the state space, action space, and reward function of the agent are redefined to match the fault early warning scenario of the drive shaft system. Here, state s is defined as a vibration signal or feature representation composed of it that characterizes the operating state of the drive shaft system, and action a is defined as the fault type discrimination result. The action is represented in one-hot encoding form, where N is the preset number of fault categories. A reward feedback mechanism is constructed based on the fault identification results. The reward value for correct fault type identification is set to r=1, and the reward value for incorrect fault type identification is set to r= 1. To form an immediate reward function based on diagnostic accuracy; By randomly shuffling the training samples and evenly distributing the number of samples of each type of fault, a training data sampling strategy with Markov properties is constructed for the agent, such that the state transition probability to the next state s' after performing action a in the current state s is... Furthermore, the state transition process is independent of the state at the previous moment, thereby ensuring that the agent's decision-making process satisfies the Markov property. Based on the above-mentioned quadruple (s,a,r,s') consisting of "state-action-reward-next state", a training sample set with Markov decision process characteristics is established, and the quadruple data is stored and randomly sampled in combination with the experience replay mechanism to reduce the temporal correlation between samples and improve sample utilization efficiency. When the number of samples in the experience replay buffer reaches a preset condition, samples are randomly selected from the experience replay buffer and input into the ontology network for parameter updates, so as to realize the training of a primary fault warning agent based on Markov properties.

[0050] By constructing a primary fault warning agent based on Markov properties, the operational monitoring data of the drive shaft system is represented as the agent's state. Through a Markov decision process, the agent selects appropriate actions based on historical data in each state to determine the fault type. This decision process adjusts the action selection based on feedback from the current system state to achieve efficient and accurate fault warning. In this way, the agent can perceive and adapt to changes in the operating environment in real time, automatically optimize decision-making strategies, and improve the accuracy and response speed of fault prediction.

[0051] Step 2: Knowledge exploration based on adaptive action sampling - utilizing a balancing method; In step two, an action sampling function is constructed based on the adaptive degradation factor under the condition of maximizing the action value function, so that the action selection probability changes dynamically as the agent training process progresses. By combining a random sampling strategy with the action sampling function, an action sampling strategy based on an adaptive degradation mechanism is constructed. In the early stages of training, this strategy increases the probability of exploring non-optimal actions, and in the later stages, it gradually transitions to a deterministic sampling method guided by the maximum value of the action value function. Through this action sampling strategy based on the adaptive degradation mechanism, the exploration and utilization ratio of the agent in different training stages is adaptively adjusted, enabling the agent to accelerate policy convergence and improve the stability and accuracy of fault warning decisions while ensuring sufficient exploration of the global state space. The specific expression of this strategy is as follows:

[0052] in, Indicates the initial state; Indicates the state The next action will be selected based on the current strategy; This represents the set of actions that the system can choose from; Indicates from the set of actions Randomly select an action; State-action pairs Action evaluation function; This represents the probability weights of using different action sampling methods at the current time step; Indicates the state The action that maximizes the evaluation function; The probability of the early warning agent performing greedy sampling is represented by the following formula:

[0053] in, This represents the adaptive degradation control factor, used to adjust the proportion of exploration and utilization of the strategy at different training stages.

[0054] During agent training, an adaptive action sampling strategy is employed. By adjusting the balance between exploration (unseen states) and utilization (learned strategies), the system's fault prediction capability is enhanced. As training progresses, the agent adaptively adjusts its action selection strategy according to different training stages. In the early stages, it relies more on exploration to discover unknown faults, while in the later stages, it gradually transitions to deterministic sampling based on learned strategies to better utilize existing knowledge for fault identification. This dynamic action sampling strategy optimizes data acquisition efficiency and accelerates the model's convergence process.

[0055] Step 3: Employ a hierarchical progressive isomorphic filtering method that combines multi-scale and multi-fine-grained features; In step three, the input feature map is first decoupled and grouped, and the original feature map is divided into several feature subgroups according to the channel dimension to obtain multi-scale feature representations with different receptive fields.

[0056] in, This represents the i-th input feature map. This represents the corresponding progressive output feature. This represents the i-th order nonlinear characteristic mapping function; Multiple identity mapping functions are used to connect the progressive feature pyramid across layers, constructing a multi-level feature representation pyramid structure. Basic general features, local correlation features, and global abstract features are jointly preserved and focused for extraction, which can alleviate the gradient vanishing and feature representation degradation problems in the training process of deep networks and improve the robustness of the model in strong noise environment.

[0057] To address the feature extraction problem under strong noise interference conditions, this method employs a hierarchical progressive isomorphic filtering technique. Through multi-scale and multi-fine-grained feature extraction, fault-related features are progressively extracted from the original signal while background noise is filtered out. This process improves robustness in strong noise environments by gradually increasing the receptive field layer by layer and fusing local and global information. This method not only reduces the impact of noise on system performance but also effectively preserves key signals and enhances the ability to extract fault features.

[0058] Step 4: Employ a feature weighting method for adaptive weight vectors designed for weak feature enhancement; In step four, weight vectors corresponding to each channel are extracted based on the input feature map. Initial weight vectors reflecting the feature response intensity of different channels are obtained by calculating the statistical features of each channel's feature map. These initial weight vectors are then normalized using a feature weight vector extraction method based on the softmax function, mapping the weights corresponding to each channel to a probability distribution. The normalization process is as follows:

[0059] in, This represents the initial weight value corresponding to the i-th channel. This represents the total number of channels in the feature map. This represents the normalized channel weight coefficients; Channel features are weighted to strengthen key channels and suppress irrelevant channels, thereby enhancing discriminative features and improving the model's robustness in noisy environments.

[0060] After noise-resistant feature extraction, an adaptive weighting mechanism is introduced to enhance weak fault signals and suppress redundant or irrelevant features based on the filtered features. In this process, the weighting vector is dynamically adjusted based on real-time data, enhancing features relevant to the fault state while suppressing noisy or irrelevant features. This mechanism improves the system's noise resistance while maintaining accuracy, ensuring accurate identification of fault signals in complex environments.

[0061] Step 5: Employ a global training objective optimization method based on a dual-network collaborative training strategy; In step five, under the dual-network collaborative training framework, an ontology network and a target network are constructed, wherein: The ontology network is used to base on the current system state. With action The input is used to estimate the action value function in real time, and the corresponding action value is output. The model parameters of the ontology network are iteratively updated using the backpropagation algorithm; the target network is used to update the model parameters for the next state. The stability of the action value is estimated, and its output action value is calculated. The model parameters of the target network According to the preset training step size period and the model parameters of the ontology network Perform synchronized updates; The dual-network collaborative training process includes: The ontology network is used to calculate the action value output in the current state. The output result is then used to construct the loss function; the target network is used to construct the target mapping quantity corresponding to the predicted output, which can be expressed as:

[0062] in, Indicates the amount of real-time feedback. This is the attenuation weighting coefficient. Represents a set of actions; Furthermore, the global training objective optimization method includes: Multi-class prediction logits are constructed based on the output of the ontology network, and the logits are normalized and mapped to obtain normalized action evaluation output. To enhance the distribution differences of different fault states in the feature space; by introducing a global optimization constraint based on logits distribution stretching, the output corresponding to the target fault category is maximized relative to the non-target category; based on the deviation relationship between the normalized prediction output and the target mapping, a total loss function for dual-network collaborative optimization is constructed, which can be expressed as:

[0063] in, Indicates the number of samples used in training; By employing the aforementioned dual-network collaborative training mechanism and a global loss optimization strategy based on logits distribution differences, joint updates of model parameters are achieved, thereby improving the stability and robustness of fault warning in wind turbine drive systems under complex noise environments.

[0064] A dual-network collaborative training architecture is adopted, using an ontology network and a target network separately for fault prediction decision-making. Through this collaborative training architecture, the agent can estimate action value through the ontology network and stably estimate the action value of the next state through the target network during training, thereby improving model stability. During training, the parameters of the target network are updated synchronously with the ontology network at regular intervals, avoiding overfitting and ensuring the model's robustness under noise interference. This dual-network collaborative constraint further enhances training stability and convergence accuracy.

[0065] Step 6: Employ a multivariate gradient dynamic adaptive descent optimization method based on Taylor series approximation order control; Step six, the Taylor series approximation order control and multivariate gradient dynamic adaptive descent optimization, includes the following steps: The first step is to obtain the partial gradients of the multivariate objective function and construct the multivariate gradient vectors of the objective function with respect to each parameter component in the high-dimensional parameter space, when the input variables are represented as... When the objective function g is expressed as a multivariate gradient vector, it is as follows:

[0066] The second step involves applying a first-order Taylor approximation control to the multivariate objective function to describe its variation within the neighborhood of the current parameter point. The first-order Taylor expansion is expressed as:

[0067] in, Represents the parameter perturbation vector; The third step, based on the first-order Taylor approximation, involves perturbing the parameters along the negative gradient direction, with the perturbation vector set as... Substituting this into the first-order Taylor expansion, we obtain the equivalent expression for the objective function along the multivariate negative gradient direction:

[0068] Fourth, based on the Taylor approximation equation, a dynamic adaptive optimization update rule for the original parameter vector under the preheating mechanism is constructed, and the parameter update rule is expressed as follows:

[0069] in, This is a dynamic adaptive factor in the parameter update process. During the parameter iteration update process, it adopts an adaptive adjustment and control of the update step size based on a preheating and exponential decay mechanism. The adjustment rule is as follows:

[0070] in, This represents the initial value of the dynamic factor, and t represents the current training time. K represents the length of the preheating stage of the dynamic factor, and K represents the decay coefficient of the dynamic factor.

[0071] Step 7: Employ a method for joint evaluation of model performance and dynamic self-incremental update of model structure using multiple evaluation indicators; In step seven, the model performance is jointly evaluated based on multiple metrics, including accuracy, precision, recall, and F1 score. The formula for calculating model accuracy is as follows:

[0072] The formula for calculating model accuracy is:

[0073] The formula for calculating model recall is:

[0074] The formula for calculating the F1 score of the model is:

[0075] Wherein, TP represents the number of samples that were correctly identified as the target fault category, FP represents the number of samples that were misclassified as the fault category, FN represents the number of samples that were incorrectly identified as other categories, and TN represents the number of samples that were not the fault category and were correctly identified. When a new fault type is detected, freeze the existing feature extractor. Create a new feature extractor It is fine-tuned based on incremental data; during incremental training, only the new feature extractor is updated, enabling the model to quickly adapt to new fault types.

[0076] During model training or operation, multiple evaluation metrics (such as accuracy, precision, recall, and F1 score) are used to jointly assess the system's fault warning performance. These metrics help quantify the model's performance under different operating conditions, thus comprehensively reflecting the system's reliability and accuracy. Through these evaluation results, the model can be dynamically updated based on changes in operating conditions or the addition of new fault types, ensuring the system can continuously adapt to new environments and task requirements, and improving the sustainability and applicability of fault warnings.

[0077] like Figure 5 As shown, the deep features extracted by the model are reduced to a two-dimensional space for visualization using t-SNE technology. The visualization results show that samples of different fault categories form well-defined clusters in the feature space, with high inter-class separation, which intuitively proves that the features extracted by the model have high discriminative power and robustness.

[0078] To verify the advanced nature and effectiveness of the proposed lightweight robust wind turbine drive system fault early warning method based on dual-network collaborative optimization in terms of structural design and training mechanism, this embodiment further selects a variety of typical neural network models as comparison objects and conducts comparative experimental analysis under the same experimental data and test conditions.

[0079] In the comparative experiments, all models used the same vibration signal dataset, sample partitioning method, and evaluation metrics to ensure consistency of experimental conditions. Experimental results show that under strong background noise interference, the comparative models exhibit varying degrees of performance degradation in feature representation stability and fault discrimination consistency, especially in the identification of weak fault features, where the misclassification rate increases significantly. In contrast, the method of this invention, through a hierarchical progressive isomorphic filtering and adaptive weight vector feature enhancement mechanism, effectively suppresses noise interference at the feature level, allowing fault-related features to maintain a clearer distribution structure in the deep feature space, such as... Figure 6 As shown.

[0080] This invention, through the specific embodiments described above, constructs and verifies a complete intelligent fault early warning solution for transmission systems in high-noise environments. This method, centered on improved deep reinforcement learning, integrates an adaptive exploration strategy, a lightweight, noise-resistant feature extraction network, and a feature-enhanced attention mechanism. In practical ship shafting fault diagnosis, it demonstrates significant advantages such as high accuracy, strong noise resistance, fast convergence, and ease of deployment, showing promising engineering application prospects and widespread value.

[0081] Example 2 like Figure 7 As shown, based on the same inventive concept as the above embodiments, the present invention also provides a robust wind turbine fault early warning model training system based on dual-network collaborative optimization, comprising: A primary fault warning agent construction unit is used to acquire the test samples of the wind turbine system and construct a primary fault warning agent based on Markov properties according to the fault warning task requirements. The primary fault warning agent includes an ontology network and a target network. The feature extraction unit is used to introduce an adaptive action sampling method in the primary fault warning agent to generate decision samples; and to perform hierarchical progressive isomorphic filtering on the state features corresponding to the decision samples to extract filtered features with multi-scale and multi-fine-grained information. The noise reduction enhancement unit is used to perform noise reduction modeling based on the filtered features to obtain noise reduction features; the noise reduction features introduce an adaptive weight vector feature weighting method to enhance features related to the fault state and suppress redundant or interfering features to obtain enhanced features; The first update and optimization unit is used to utilize the enhanced features and, through the collaborative optimization of the ontology network and the target network, update the parameters of the primary fault warning agent based on the global training objective and loss function to obtain the updated fault warning agent. The second update and optimization unit is used to update the parameters of the fault warning agent after the update using a gradient optimization method based on Taylor approximation and dynamic adaptive step size. The gradient optimization method based on Taylor approximation and dynamic adaptive step size includes: calculating the multivariate gradient vector of the objective function, determining the parameter update direction based on the first-order Taylor approximation, and dynamically adjusting the update step size through an adaptive factor that includes a preheating and decay mechanism to obtain a robust wind turbine fault warning model.

[0082] Example 3 like Figure 8 As shown, the present invention also provides an electronic device 100 for implementing a robust wind turbine fault early warning model training method based on dual-network collaborative optimization; The electronic device 100 includes a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on at least one processor 102, and at least one communication bus 104.

[0083] The memory 101 can be used to store the computer program 103. The processor 102 implements the steps of Embodiment 1 by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101.

[0084] The memory 101 may primarily include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created based on the use of the electronic device 100 (such as audio data), etc. In addition, the memory 101 may include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other non-volatile solid-state storage device.

[0085] At least one processor 102 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Processor 102 may be a microprocessor or any conventional processor. Processor 102 is the control center of electronic device 100, connecting various parts of electronic device 100 via various interfaces and lines.

[0086] The memory 101 in the electronic device 100 stores multiple instructions to implement a robust wind turbine fault early warning model training method based on dual-network collaborative optimization, and the processor 102 can execute multiple instructions to achieve the following: Acquire the test samples of the wind turbine system, and construct a primary fault early warning agent based on Markov properties according to the fault early warning task requirements. The primary fault early warning agent includes an ontology network and a target network. In the primary fault warning agent, an adaptive action sampling method is introduced to generate decision samples; the state features corresponding to the decision samples are subjected to hierarchical progressive isomorphic filtering to extract filtered features with multi-scale and multi-fine-grained information. Based on the filtered features, noise resistance modeling is performed to obtain noise resistance features; the noise resistance features are then subjected to an adaptive weight vector feature weighting method to enhance features related to the fault state and suppress redundant or interfering features, resulting in enhanced features; Using the enhanced features, and through the collaborative optimization of the ontology network and the target network, the parameters of the primary fault warning agent are updated based on the global training objective and loss function to obtain the updated fault warning agent; The parameters of the fault warning agent are updated using a gradient optimization method based on Taylor approximation and dynamic adaptive step size. The gradient optimization method based on Taylor approximation and dynamic adaptive step size includes: calculating the multivariate gradient vector of the objective function, determining the parameter update direction based on the first-order Taylor approximation, and dynamically adjusting the update step size through an adaptive factor that includes a preheating and decay mechanism to obtain a robust wind turbine fault warning model.

[0087] Example 4 If the modules / units integrated in the electronic device 100 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, and read-only memory (ROM).

[0088] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0089] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0090] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0091] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0092] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A robust wind turbine fault early warning model training method based on dual-network collaborative optimization, characterized in that, include: Acquire the test samples of the wind turbine system, and construct a primary fault early warning agent based on Markov properties according to the fault early warning task requirements. The primary fault early warning agent includes an ontology network and a target network. In the primary fault warning agent, an adaptive action sampling method is introduced to generate decision samples; Hierarchical progressive isomorphic filtering is applied to the state features corresponding to the decision samples to extract filtered features with multi-scale and multi-fine-grained information; Based on the filtered features, noise resistance modeling is performed to obtain noise resistance features; The noise-resistant feature introduces an adaptive weight vector feature weighting method to enhance features related to the fault state and suppress redundant or interfering features, resulting in enhanced features; Using the enhanced features, and through the collaborative optimization of the ontology network and the target network, the parameters of the primary fault warning agent are updated based on the global training objective and loss function to obtain the updated fault warning agent; The parameters of the fault warning agent are updated using a gradient optimization method based on Taylor approximation and dynamic adaptive step size. The gradient optimization method based on Taylor approximation and dynamic adaptive step size includes: calculating the multivariate gradient vector of the objective function, determining the parameter update direction based on the first-order Taylor approximation, and dynamically adjusting the update step size through an adaptive factor that includes a preheating and decay mechanism to obtain a robust wind turbine fault warning model.

2. The training method for a robust wind turbine fault early warning model based on dual-network collaborative optimization according to claim 1, characterized in that, The test sample is constructed based on Markov properties to form a primary fault warning agent according to the fault warning task requirements, including: Based on the requirements of fault early warning tasks, a task attribute mapping rule for reinforcement learning agents is established to match the agent's operating status, fault type discrimination results, and number of fault categories. A reward feedback mechanism is constructed based on the fault type identification results. The reward value when the fault type is correctly identified is set to a different value than the reward value when the fault type is incorrectly identified, thus forming an instant reward function based on diagnostic accuracy. The test samples are randomly shuffled, and the number of samples of each type of fault is evenly distributed. A training data sampling method for the agent with Markov properties is constructed, such that the state transition probability to the next state s' after performing action a in the current state s is given by... Furthermore, the state transition process is independent of the state at the previous moment; Based on the quadruple (s,a,r,s') consisting of the current state s, the action a, the reward value r, and the next state s', a training sample set with Markov decision process characteristics is established, and the quadruple data is stored and randomly sampled in combination with the experience replay mechanism to form an experience replay buffer. When the number of samples in the experience replay buffer reaches a preset condition, samples are randomly selected from the experience replay buffer and input into the ontology network for parameter update, thereby training a primary fault warning agent based on Markov properties.

3. The training method for a robust wind turbine fault early warning model based on dual-network collaborative optimization according to claim 1, characterized in that, In the initial fault warning agent, an adaptive action sampling method is introduced to generate decision samples for training the fault warning agent, including: In the primary fault warning agent, an action sampling function is constructed based on an adaptive degradation factor to maximize the action value function. By combining a random sampling strategy with an action sampling function, an action sampling mechanism based on an adaptive degradation mechanism is constructed. The primary fault warning agent adaptively adjusts its training through an action sampling mechanism based on adaptive degradation, generating decision samples for training the fault warning agent; the action sampling mechanism based on adaptive degradation is as follows: in, Indicates the initial state; Indicates the state The next action will be selected based on the current strategy; This represents the set of actions that the system can choose from; Indicates from the set of actions Randomly select an action; State-action pairs Action evaluation function; This represents the probability weights of using different action sampling methods at the current time step; Indicates the state The action that maximizes the evaluation function; The probability of the early warning agent performing greedy sampling is represented by the following formula: in, This represents the adaptive degradation control factor, used to adjust the proportion of exploration and utilization of the strategy at different training stages.

4. The training method for a robust wind turbine fault early warning model based on dual-network collaborative optimization according to claim 1, characterized in that, The hierarchical progressive isomorphic filtering process performed on the state features corresponding to the decision samples to obtain filtered features includes: The decision samples are decoupled and grouped to obtain the original feature map. The original feature map is then divided into several feature subgroups according to the channel dimension. The progressive output features are determined based on the feature subgroups. Multiple identity mapping functions are used to connect the progressive output features across layers, constructing a multi-level feature representation pyramid structure to obtain the filtered features.

5. The training method for a robust wind turbine fault early warning model based on dual-network collaborative optimization according to claim 1, characterized in that, The filtered features are combined with multi-scale and multi-fine-grained features to perform noise-resistant modeling on the original state features, resulting in noise-resistant features. Specifically, the filtered features, basic general features, local correlation features, and global abstract features are jointly retained and focused on for extraction to obtain noise-resistant features.

6. The training method for a robust wind turbine fault early warning model based on dual-network collaborative optimization according to claim 1, characterized in that, The noise-resistant features employ an adaptive weight vector feature weighting method to enhance features related to the fault state and suppress redundant or interfering features, resulting in enhanced features, including: Extract the weight vectors corresponding to each channel of the noise reduction feature to obtain the initial weight vectors that reflect the response intensity of different channel features; The initial weight vector is normalized using a feature weight vector extraction method based on the softmax function to obtain channel features; The channel features are weighted to enhance features related to fault states and suppress redundant or interfering features, resulting in enhanced features.

7. The training method for a robust wind turbine fault early warning model based on dual-network collaborative optimization according to claim 1, characterized in that, The ontology network is used to base on the current system state. With action The input is used to estimate the action value function in real time, and the corresponding action value is output. The model parameters of the ontology network are iteratively updated using the backpropagation algorithm. The target network is used for the next state. The stability of the action value is estimated, and its output action value is calculated. The model parameters of the target network According to the preset training step size period and the model parameters of the ontology network Perform synchronized updates.

8. A robust wind turbine fault early warning model training system based on dual-network collaborative optimization, characterized in that, include: A primary fault warning agent construction unit is used to acquire the test samples of the wind turbine system and construct a primary fault warning agent based on Markov properties according to the fault warning task requirements. The primary fault warning agent includes an ontology network and a target network. The feature extraction unit is used to introduce an adaptive action sampling method to generate decision samples in the primary fault warning agent; Hierarchical progressive isomorphic filtering is applied to the state features corresponding to the decision samples to extract filtered features with multi-scale and multi-fine-grained information; The noise reduction enhancement unit is used to perform noise reduction modeling based on the filtered features to obtain noise reduction features; The noise-resistant feature introduces an adaptive weight vector feature weighting method to enhance features related to the fault state and suppress redundant or interfering features, resulting in enhanced features; The first update and optimization unit is used to utilize the enhanced features and, through the collaborative optimization of the ontology network and the target network, update the parameters of the primary fault warning agent based on the global training objective and loss function to obtain the updated fault warning agent. The second update and optimization unit is used to update the parameters of the fault warning agent after the update using a gradient optimization method based on Taylor approximation and dynamic adaptive step size. The gradient optimization method based on Taylor approximation and dynamic adaptive step size includes: calculating the multivariate gradient vector of the objective function, determining the parameter update direction based on the first-order Taylor approximation, and dynamically adjusting the update step size through an adaptive factor that includes a preheating and decay mechanism to obtain a robust wind turbine fault warning model.

9. An electronic device, characterized in that, It includes a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the robust wind turbine fault early warning model training method based on dual-network collaborative optimization as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, which, when executed by a processor, implements the robust wind turbine fault early warning model training method based on dual-network collaborative optimization as described in any one of claims 1 to 7.

Images