Fan gear box rupture early warning method and device based on acoustic emission positioning technology

Abstract

The application relates to the technical field of power generation equipment fault diagnosis, and discloses a fan gear box rupture early warning method and device based on acoustic emission positioning technology, which comprises the following steps: obtaining internal structure parameters of a fan gear box, constructing an acoustic wave propagation model based on the internal structure parameters of the fan gear box; performing acoustic wave propagation simulation at multiple acoustic emission source positions by using the acoustic wave propagation model to obtain acoustic wave propagation simulation data; obtaining acoustic emission real-time signals, performing path separation on the acoustic emission real-time signals by using a graph convolution network model to obtain classification probability data of multipath signal paths; and performing fault signal detection by using a reinforcement learning agent model based on the acoustic wave propagation simulation data and the classification probability data of the multipath signal paths to obtain fan gear box rupture early warning information. The application effectively reduces the false positive rate, greatly improves the accuracy and reliability of the early warning system, and provides more reliable protection for fan gear box rupture early warning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power generation equipment fault diagnosis, and particularly relates to a fan gear box rupture early warning method and device based on acoustic emission positioning technology. BACKGROUND

[0002] With the rapid development of the wind power industry, the reliability and stability of the fan as the core power generation equipment are crucial. The fan gear box as a key component of the fan is long-term in complex and variable operating conditions, and is prone to various faults. The internal rupture of the gear box is a relatively serious fault form, which may cause the fan to shut down and even cause safety accidents. Therefore, it is of great significance to accurately and timely warn the internal rupture of the fan gear box.

[0003] However, the related fan gear box rupture early warning method has problems such as multi-path interference, inaccurate signal processing, and high false positive rate, which cannot guarantee the accuracy and reliability of the fan gear box rupture early warning result. SUMMARY

[0004] Therefore, the present application provides a fan gear box rupture early warning method and device based on acoustic emission positioning technology to solve the problem that the accuracy and reliability of the fan gear box rupture early warning result cannot be guaranteed.

[0005] In a first aspect, the present application provides a fan gear box rupture early warning method based on acoustic emission positioning technology, which comprises:

[0006] Obtaining the internal structure parameters of the fan gear box, and constructing an acoustic wave propagation model based on the internal structure parameters of the fan gear box;

[0007] Using the acoustic wave propagation model to simulate acoustic wave propagation at multiple acoustic emission source positions to obtain acoustic wave propagation simulation data;

[0008] Obtaining acoustic emission real-time signals, and using a graph convolution network model to separate the acoustic emission real-time signals to obtain classification probability data of multi-path signal paths;

[0009] Based on the acoustic wave propagation simulation data and the classification probability data of the multi-path signal paths, using a reinforcement learning agent model to detect fault signals to obtain fan gear box rupture early warning information.

[0010] The fan gear box rupture early warning method based on the acoustic emission positioning technology provided in the embodiment can construct an acoustic wave propagation model based on internal structure parameters of the fan gear box, and then simulate acoustic wave propagation at multiple acoustic emission source positions by using the acoustic wave propagation model to obtain acoustic wave propagation simulation data, so that the propagation path and characteristics of the acoustic emission signal under multipath effect can be more accurately predicted, useful information and interference components in the signal can be more accurately identified, and the reliability of signal analysis is improved. Secondly, the path separation of the acoustic emission real-time signal is performed by using the graph convolution network model, which provides purer and more accurate signals for acoustic emission signal processing and fault diagnosis, and helps to improve the accuracy of fault diagnosis. Finally, based on the acoustic wave propagation simulation data and the classification probability data of the multipath signal path, the fault signal detection is performed by using the reinforcement learning agent model, so that the system can automatically optimize the fault detection strategy according to the actual working condition, effectively reduces the false positive rate, and greatly improves the accuracy and reliability of the early warning system, thereby providing more reliable protection for the fan gear box rupture early warning.

[0011] In an optional implementation, the internal structure parameters of the fan gear box are acquired, and the acoustic wave propagation model is constructed based on the internal structure parameters of the fan gear box, including:

[0012] The three-dimensional parameters of each component inside the fan gear box are acquired, and a three-dimensional physical model of the fan gear box is constructed based on the three-dimensional parameters of each component inside the fan gear box.

[0013] The acoustic parameters of each component inside the fan gear box are acquired, and the acoustic parameters of each component inside the fan gear box are given to the three-dimensional physical model of the fan gear box.

[0014] The three-dimensional physical model of the fan gear box to which the acoustic parameters of each component inside the fan gear box are given is subjected to finite element analysis to obtain the acoustic wave propagation model.

[0015] The fan gear box rupture early warning method based on the acoustic emission positioning technology provided in the embodiment can construct an acoustic wave propagation model based on internal structure parameters of the fan gear box, and then simulate acoustic wave propagation at multiple acoustic emission source positions by using the acoustic wave propagation model to obtain acoustic wave propagation simulation data, so that the propagation path and characteristics of the acoustic emission signal under multipath effect can be more accurately predicted, useful information and interference components in the signal can be more accurately identified, and the reliability of signal analysis is improved. Secondly, the path separation of the acoustic emission real-time signal is performed by using the graph convolution network model, which provides purer and more accurate signals for acoustic emission signal processing and fault diagnosis, and helps to improve the accuracy of fault diagnosis. Finally, based on the acoustic wave propagation simulation data and the classification probability data of the multipath signal path, the fault signal detection is performed by using the reinforcement learning agent model, so that the system can automatically optimize the fault detection strategy according to the actual working condition, effectively reduces the false positive rate, and greatly improves the accuracy and reliability of the early warning system, thereby providing more reliable protection for the fan gear box rupture early warning.

[0016] In an optional implementation, the path separation of the acoustic emission real-time signal is performed by using the graph convolution network model to obtain the classification probability data of the multipath signal path, including:

[0017] The acoustic emission real-time signal is subjected to signal preprocessing, and the acoustic emission real-time signal after the signal preprocessing is subjected to feature extraction to obtain signal propagation characteristics.

[0018] Based on the preprocessed acoustic emission real-time signal, the classification probability data of multipath signal paths are determined using a graph convolutional network model.

[0019] The wind turbine gearbox rupture early warning method based on acoustic emission localization technology provided in this embodiment uses a graph convolutional network model to separate the real-time acoustic emission signal path, reducing the contamination of features (such as amplitude and phase) by noise. The separated real signal path features are closer to the actual fault signal, making the "fault-noise" boundary clearer when adjusting the threshold.

[0020] In one optional implementation, based on acoustic wave propagation simulation data and classification probability data of multipath signal paths, a reinforcement learning surrogate model is used to detect fault signals and obtain early warning information for wind turbine gearbox rupture, including:

[0021] The location optimization was performed on the sound wave propagation simulation data to determine the multipath interference level index;

[0022] Signal separation accuracy is determined based on classification probability data of multipath signal paths;

[0023] A reinforcement learning agent model was used to determine threshold adjustment strategies under various wind turbine operating conditions.

[0024] Obtain the current wind turbine operating parameters, and based on the current wind turbine operating parameters, multipath interference level index and signal path separation accuracy, determine the fault trigger threshold using threshold adjustment strategies under various wind turbine operating conditions;

[0025] The signal propagation characteristics are compared with the fault triggering threshold. If the signal propagation characteristics exceed the fault triggering threshold, a wind turbine gearbox rupture warning message is generated.

[0026] The wind turbine gearbox rupture early warning method based on acoustic emission localization technology provided in this embodiment can make timely threshold adjustment decisions as the operating conditions change, ensuring that fault signals can be accurately detected under various circumstances. This not only improves the real-time performance of fault detection and can issue early warnings at the first moment of a fault occurrence, but also ensures the accuracy of detection, preventing the omission of real faults due to improper threshold settings, thus providing a more reliable guarantee for the safe operation of the equipment.

[0027] In one optional implementation, a multipath interference level index is determined based on acoustic wave propagation simulation data, including:

[0028] Based on acoustic wave propagation simulation data, the location vector of the acoustic emission source is determined using a positioning algorithm.

[0029] Based on the position vector of the acoustic emission source, the propagation parameters are simulated to obtain the basic simulation parameters of sound wave propagation inside the gearbox.

[0030] The actual measurement parameters of sound wave propagation inside the gearbox are obtained. The simulated parameters of sound wave propagation inside the gearbox are compared with the actual measurement parameters of sound wave propagation inside the gearbox to obtain the multipath interference level index.

[0031] The wind turbine gearbox rupture early warning method based on acoustic emission localization technology provided in this embodiment provides real-time environmental feedback for reinforcement learning through the quantitative results of localization accuracy (i.e., multipath interference level index), enabling the system to dynamically optimize the detection strategy according to the reliability of sound wave propagation, thereby improving the accuracy and robustness of rupture early warning.

[0032] In one alternative implementation, a reinforcement learning agent model is used to determine threshold adjustment strategies under various wind turbine operating conditions, including:

[0033] Define the state, action, and reward function; where the state consists of historical false alarm data, wind turbine operating parameters, multipath interference level indicators, and signal path separation accuracy corresponding to various wind turbine operating conditions; the action is the adjustment operation of the fault trigger threshold.

[0034] Get the current state, select the current action based on the current state, adjust the fault trigger threshold based on the current action, and update the current state to get the updated state;

[0035] Calculate the current reward value based on the updated state;

[0036] Calculate the current Q value based on the current state, current action, updated state, and current reward value;

[0037] The current Q value is compared with the target Q value. The threshold adjustment strategy is updated iteratively based on the comparison result until the current Q value matches the target Q value, thus obtaining the threshold adjustment strategy under various wind turbine operating conditions.

[0038] The wind turbine gearbox rupture early warning method based on acoustic emission localization technology provided in this embodiment uses a reinforcement learning agent model to dynamically adjust the fault trigger threshold based on historical false alarm data and current operating conditions. During wind turbine operation, the characteristics of acoustic emission signals vary greatly under different operating conditions, and a fixed threshold is prone to false alarms. Therefore, through a dynamic threshold adjustment mechanism, it can automatically adapt to changes in operating conditions. In situations where transient noise is likely to occur, such as during wind turbine startup, the threshold is automatically increased to avoid false alarms. Furthermore, the reinforcement learning agent continuously learns and adjusts the threshold in real time, ensuring that the system always maintains the best fault detection performance.

[0039] Secondly, the present invention provides a wind turbine gearbox rupture early warning device based on acoustic emission positioning technology, the device comprising:

[0040] The module is used to obtain the internal structural parameters of the wind turbine gearbox and to build a sound wave propagation model based on the internal structural parameters of the wind turbine gearbox.

[0041] The simulation module is used to simulate sound wave propagation at multiple sound emission source locations using a sound wave propagation model, and obtain sound wave propagation simulation data.

[0042] The separation module is used to acquire real-time acoustic emission signals and use a graph convolutional network model to perform path separation on the real-time acoustic emission signals to obtain classification probability data of multipath signal paths;

[0043] The detection module is used to detect fault signals based on acoustic wave propagation simulation data and classification probability data of multipath signal paths, and to obtain early warning information of wind turbine gearbox rupture by using a reinforcement learning agent model.

[0044] Thirdly, the present invention provides a computer device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the wind turbine gearbox rupture early warning method based on acoustic emission positioning technology described in the first aspect or any corresponding embodiment.

[0045] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the wind turbine gearbox rupture early warning method based on acoustic emission positioning technology described in the first aspect or any corresponding embodiment.

[0046] Fifthly, the present invention provides a computer program product, including computer instructions, which are used to cause a computer to execute the wind turbine gearbox rupture early warning method based on acoustic emission positioning technology described in the first aspect or any corresponding embodiment above. Attached Figure Description

[0047] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0048] Figure 1 This is a flowchart illustrating a wind turbine gearbox rupture early warning method based on acoustic emission positioning technology according to an embodiment of the present invention.

[0049] Figure 2This is a flowchart illustrating another wind turbine gearbox rupture early warning method based on acoustic emission positioning technology according to an embodiment of the present invention;

[0050] Figure 3 This is a flowchart illustrating another wind turbine gearbox rupture early warning method based on acoustic emission positioning technology according to an embodiment of the present invention;

[0051] Figure 4 This is a flowchart illustrating another wind turbine gearbox rupture early warning method based on acoustic emission positioning technology according to an embodiment of the present invention;

[0052] Figure 5 This is a structural block diagram of a wind turbine gearbox rupture early warning device based on acoustic emission positioning technology according to an embodiment of the present invention;

[0053] Figure 6 This is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Currently, acoustic emission localization technology, as an effective fault monitoring method, has been applied to some extent in the early warning of wind turbine gearbox faults. However, in practical applications, this technology faces many challenges:

[0056] 1) Multipath interference problem: The internal structure of the wind turbine gearbox is complex, with a large number of partitions, bearing seats and other structural components. When the sound waves generated by the acoustic emission source propagate inside the gearbox, these structures will cause the sound waves to be reflected and refracted, forming a multipath effect. The multipath effect makes the received acoustic emission signal complex and distorted, which seriously interferes with the accurate judgment of the location of the acoustic emission source.

[0057] For example, the superposition of reflected and direct waves may cause anomalies in the time and amplitude of the signals received by the sensor, resulting in large errors in positioning algorithms based on time difference or amplitude. In some complex gearboxes, the multipath effect may cause the positioning error to reach 20%-30% of the actual distance, which greatly affects the accuracy and reliability of acoustic emission positioning technology.

[0058] 2) Positioning error due to complex structure: The wind turbine gearbox contains numerous components, and the shapes, sizes, and material properties of these components vary considerably. This complex structure further exacerbates the difficulty of acoustic emission positioning. Related acoustic emission positioning algorithms are usually based on simple geometric models and propagation assumptions, which cannot accurately adapt to the complex sound wave propagation characteristics inside the gearbox. For example, in some narrow spaces or irregularly shaped areas inside the gearbox, sound wave propagation may be significantly affected by boundary conditions, resulting in a large deviation between the acoustic emission source position calculated by traditional algorithms and the actual position. In addition, the influence of different materials on the absorption, scattering, and propagation speed of sound waves also makes the propagation law of acoustic emission signals more complex, increasing the difficulty of positioning.

[0059] 3) Inappropriate sensor layout: The layout of acoustic emission sensors has a critical impact on positioning accuracy and fault monitoring effectiveness. In practical applications, the layout of acoustic emission sensors in many wind turbine gearboxes often lacks scientific planning. Either the sensors are too sparsely distributed, failing to fully cover potential fault areas inside the gearbox, resulting in some fault-generated acoustic emission signals not being effectively detected; or the sensor layout is too concentrated, causing redundant monitoring in some areas while insufficient monitoring in others. An inappropriate sensor layout not only wastes monitoring resources but also reduces the accuracy and timeliness of fault warnings. For example, in some large wind turbine gearboxes, due to an inappropriate sensor layout, early fault signals of some key components may be missed, delaying the timing of fault handling.

[0060] 4) High false alarm rate: Acoustic emission monitoring systems typically use a single trigger condition, such as relying solely on signal amplitude or frequency to determine fault occurrence. However, wind turbines operate in complex environments with various noise interferences. These interference signals may trigger the monitoring system, leading to false alarms. For example, when a wind turbine starts, stops, or is subjected to external impacts, it may generate brief high-amplitude or specific-frequency signals, but these signals are not caused by internal gearbox faults. According to actual operational data, the false alarm rate of acoustic emission monitoring systems with a single trigger condition can reach 30%-40%. This not only increases the workload of maintenance personnel but may also lead to the overlooking of genuine faults, reducing the reliability of the early warning system.

[0061] In summary, the wind turbine gearbox fault early warning method based on acoustic emission positioning technology has significant shortcomings in terms of multipath interference suppression, positioning accuracy of complex structures, sensor layout optimization, and reduction of false alarm rate, and urgently needs improvement.

[0062] In the application of acoustic emission localization technology for wind turbine gearbox rupture early warning, to overcome problems such as multipath interference, inaccurate signal processing, and high false alarm rate, this invention provides a wind turbine gearbox rupture early warning method based on acoustic emission localization technology. Improvements are made in three key aspects: multipath interference suppression, physical constraint-based deep learning signal path separation, and reinforcement learning-based dynamic threshold adjustment mechanism. Specifically, the method includes:

[0063] 1) Multipath interference suppression: The internal structure of the wind turbine gearbox is complex, and multipath interference during sound wave propagation severely affects the accuracy of acoustic emission positioning; to address this, an accurate sound wave propagation model is established:

[0064] a. Model Construction Method: Using professional modeling tools, a detailed 3D model of the internal structure of the wind turbine gearbox is carried out. By accurately depicting the shape, size, and relative position of various structural components such as partitions, bearing seats, gears, and shafts inside the gearbox, a 3D model that closely matches the actual situation is constructed. On this basis, advanced technologies such as finite element analysis are used to accurately simulate the propagation process of sound waves inside the gearbox. This simulation process comprehensively covers the reflection and refraction phenomena of sound waves, as well as their interaction with different structural components.

[0065] b. Parameter Assignment and Significance: Taking into full account the differences in acoustic characteristics of the materials of various components in the gearbox, accurate sound wave propagation parameters, such as sound velocity and attenuation coefficient, are assigned to each component in the model. Different materials have different effects on sound wave propagation. Accurate parameter settings can more realistically reflect the propagation law of sound waves in the actual structure. In this way, the propagation path and characteristics of acoustic emission signals under multipath effects can be predicted more accurately, providing a solid and reliable foundation for subsequent positioning algorithms and effectively improving the accuracy and reliability of positioning.

[0066] 2) Physically Constrained Deep Learning Signal Path Separation: In related signal processing methods, the separation of multipath signals is limited by simple geometric models. When processing reflection and refraction signals in complex structures such as wind turbine gearboxes, it is often difficult to achieve ideal results. Therefore, this invention innovatively proposes a physically constrained deep learning signal path separation method:

[0067] a. Introduction of a three-dimensional physical model: In the signal processing flow, a three-dimensional physical model of the gearbox structure established through finite element simulation is introduced. This model provides a detailed depiction of the internal structure of the gearbox, fully presenting the spatial position and geometry of each component, as well as the interrelationships between them, providing an accurate physical basis for signal propagation simulation.

[0068] b. Combining Graph Convolutional Network (GCN): A graph convolutional network is used to separate multipath signals. Specifically, a 3D model of the enclosure is used to generate virtual training data containing the characteristics of sound wave propagation under different paths. This data covers various characteristics of sound waves after reflection and refraction between different structural components, and is used to train the GCN model. In this way, the GCN model can learn the unique characteristics of reflected / refracted signals, significantly enhancing the model's ability to identify reflected / refracted signals, thereby achieving more accurate separation of multipath signals and improving the overall accuracy of acoustic emission signal processing.

[0069] c. Innovation: Existing algorithms primarily rely on simple geometric models to process signals. When faced with the complex structure of wind turbine gearboxes, they cannot comprehensively and accurately describe the sound wave propagation process, resulting in poor signal processing performance. This invention innovatively combines physical simulation with data-driven approaches, using prior knowledge provided by physical models to guide the training of deep learning models. This breaks the dependence of existing algorithms on simple geometric models. This innovative method can more effectively address the challenges of multipath signal processing in complex structures, significantly improving the accuracy and reliability of signal processing, and providing a more advanced technical means for acoustic emission signal processing in wind turbine gearboxes.

[0070] 3) Dynamic threshold adjustment mechanism based on reinforcement learning: While fault triggering thresholds are typically fixed, wind turbine operating conditions are complex and variable, leading to changes in acoustic emission signal characteristics. Fixed thresholds can easily cause false alarms, affecting the accuracy of the early warning system. This invention proposes a dynamic threshold adjustment mechanism based on reinforcement learning to address this issue.

[0071] a. Design reinforcement learning agent: Design a reinforcement learning agent based on Deep Q Network (DQN). This agent has powerful learning and decision-making capabilities and can dynamically adjust the fault triggering threshold based on historical false alarm data and current operating conditions (such as wind speed, load, etc.).

[0072] b. Dynamic adjustment strategy: In practical application scenarios, transient noise exists during the wind turbine startup phase. At this time, the acoustic emission signal characteristics are significantly different from those during normal operation. The reinforcement learning agent automatically increases the fault trigger threshold based on its perception of the current operating conditions, effectively avoiding false alarms caused by transient noise. As the wind turbine operating conditions continue to change, the agent continues to learn and adjusts the threshold in real time to ensure that the system always maintains the best fault detection performance.

[0073] c. Innovation: Fixed threshold methods cannot adapt to the complex and ever-changing operating conditions of wind turbines, resulting in a high false alarm rate. This invention utilizes reinforcement learning technology to achieve intelligent dynamic adjustment of the threshold, enabling the system to automatically optimize the fault detection strategy based on actual operating conditions. This innovative mechanism effectively reduces the false alarm rate, significantly improves the accuracy and reliability of the early warning system, and provides a more reliable guarantee for early warning of wind turbine gearbox rupture.

[0074] This invention provides a wind turbine gearbox rupture early warning method based on acoustic emission positioning technology. It should be noted that the execution subject of this method can be a wind turbine gearbox rupture early warning device based on acoustic emission positioning technology. This device can be implemented as part or all of an electronic device through software, hardware, or a combination of both. The electronic device can be a server or a terminal. In this embodiment, the server can be a single server or a server cluster composed of multiple servers. The terminal can be a smartphone, personal computer, tablet computer, wearable device, or other intelligent hardware device such as a smart robot. The following method embodiments all use an electronic device as the execution subject for illustration.

[0075] According to an embodiment of the present invention, a wind turbine gearbox rupture early warning method based on acoustic emission positioning technology is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0076] This embodiment provides a wind turbine gearbox rupture early warning method based on acoustic emission positioning technology, which can be used in the aforementioned electronic equipment. Figure 1 This is a flowchart of a wind turbine gearbox rupture early warning method based on acoustic emission positioning technology according to an embodiment of the present invention, such as... Figure 1 As shown, the process includes the following steps:

[0077] Step S101: Obtain the internal structural parameters of the wind turbine gearbox and construct a sound wave propagation model based on the internal structural parameters of the wind turbine gearbox.

[0078] Step S102: Use the sound wave propagation model to simulate sound wave propagation at multiple sound emission source locations to obtain sound wave propagation simulation data.

[0079] Step S103: Obtain the real-time acoustic emission signal, and use a graph convolutional network model to perform path separation on the real-time acoustic emission signal to obtain the classification probability data of the multipath signal path.

[0080] Specifically, before using a graph convolutional network model to perform path separation on real-time acoustic emission signals, the graph convolutional network model needs to be trained. The steps include:

[0081] 1) Three-dimensional physical model establishment and data generation:

[0082] Finite element simulation modeling: Using finite element analysis software, such as ANSYS or ABAQUS, an accurate finite element model is established based on the three-dimensional model of the gearbox. The finite element model is meshed reasonably to ensure that the propagation of sound waves can be accurately simulated without excessive computation. According to the actual material parameters, the corresponding physical properties such as elastic modulus and density are assigned to each component in the model to simulate the propagation characteristics of sound waves in different materials.

[0083] Virtual training data generation: Multiple virtual acoustic emission source locations are set in the finite element model to simulate acoustic emission events at different locations. Through finite element simulation, the signal characteristics of sound waves propagating inside the gearbox to each sensor location are calculated, including information such as sound pressure, phase, and propagation time. The above data is collected to generate a virtual training dataset containing sound wave propagation characteristics under different paths. The data in the virtual training dataset is preprocessed, such as normalization, to ensure data consistency and usability.

[0084] 2) Graph Convolutional Network (GCN) Model Construction and Training:

[0085] Model building: Based on deep learning frameworks such as TensorFlow or PyTorch, construct a graph convolutional network model; define the nodes and edges of the graph, where nodes can represent sensor locations or specific locations inside a gearbox, and edges represent sound wave propagation paths or connections between nodes; set appropriate graph convolutional layers, pooling layers, and fully connected layers to build a network model that can effectively process graph-structured data.

[0086] Model Training: The generated virtual training dataset is input into the GCN model for training. Appropriate training parameters are set, such as learning rate, number of iterations, and loss function. During training, the model learns the characteristics of sound wave propagation in the virtual training data, gradually mastering the characteristics of reflected / refracted signals. The validation set is used to monitor the training process, preventing overfitting and adjusting model parameters to optimize performance. The following loss function is used to measure the difference between the model's predictions and the true labels:

[0087]

[0088] in, represents the loss function value, used to measure the difference between the prediction results and the true labels of the physical constraint-based deep learning model in the signal path separation task. Minimizing this loss function optimizes the model parameters, enabling the model to better separate multipath signals; P is the number of samples in the virtual training dataset, collected through multiple simulations of different fault scenarios and normal operating conditions in the wind turbine gearbox rupture early warning research based on acoustic emission localization technology, and used to train the deep learning model. The number of samples affects the model's training effect and generalization ability; y p For the true label of sample p, y p ∈{0,1}, where 0 represents a noise signal path and 1 represents a real acoustic emission signal path. This is determined based on manual annotation of the acoustic emission signal or known fault simulation conditions, and is used to supervise model training, enabling the model to learn to distinguish between real and noise signal paths. σ(x) is the Sigmoid function, which maps the model output to the [0,1] interval. In neural networks, it is often used to convert the result of a linear transformation into a probabilistic form to determine the authenticity of the signal path, making the model output more consistent with the actual probability distribution, thus facilitating classification tasks. K is the number of graph convolutional layers in the GCN model, determined through multiple experiments comparing the impact of different numbers of layers on model performance. K determines the model's complexity and feature extraction capability; a suitable number of layers can achieve a good balance between computational complexity and model performance, effectively extracting features related to the acoustic emission signal path. W k W is the weight matrix of the k-th graph convolutional layer, which is continuously adjusted during model training to transform node features, learn the relationships and feature representations between different nodes, and optimize the model's ability to extract and classify signal features; k (m, n) represent matrix W k The element in the m-th row and n-th column participates in the specific transformation calculation of the node features. Its value is adjusted during training based on the feedback of the loss function to optimize model performance. W represents the weight matrix of the k-th layer. k The mean of the elements (m,n) during training is used for weight regularization to prevent overfitting. This mean is obtained by statistically calculating the elements of the weight matrix after each iteration, keeping the weight values ​​within a reasonable range and preventing the model from over-relying on certain features. ReLU(x) is the modified linear unit function, introducing non-linearity so that the model can learn more complex signal feature relationships. When dealing with combined features of different frequency components in multipath signals, the ReLU function can highlight useful features and suppress useless or negative features, enhancing the model's expressive power. pqX represents the connection weight between node p and node q in the adjacency matrix of the graph structure, reflecting the correlation between nodes. This weight is determined based on the spatial relationship of the nodes in the gearbox structure and the possible paths of sound wave propagation. For example, nodes that are close together and have unobstructed sound wave propagation have larger connection weights, helping the model capture the relationship information between nodes and better process graph structure data; q The feature vector of node q contains features related to the sound wave propagation path extracted from the virtual training data, such as propagation time, amplitude, and phase. It provides input features for the model, enabling the model to learn the feature representations of signals from different paths. and These are the feature vectors of sample p obtained based on simulation and actual measurement, respectively. It includes simulated sound wave propagation path features, which are compared with feature vectors obtained from actual measurements to evaluate the accuracy of the model simulation and assist in model training. The data is obtained through actual measurements and contains real sound wave propagation path related features, which serve as supervisory information for model training to guide the model in learning accurate signal features. λ1, λ2, λ3, and λ4 are coefficients that weigh different terms and were determined through multiple cross-validation experiments. They are used to balance the contributions of weight regularization, differences between simulated and measured features, correlation, and energy differences in the loss function, so that the model's performance in different aspects can be reasonably optimized. Kullback-Leibler divergence is used to measure the simulated eigenvectors. With actual measured feature vector The differences between the simulated and measured features are assessed by calculating the KL divergence, which helps the model adjust its parameters to reduce the differences. The correlation coefficient is used to measure the performance of simulated feature vectors. With actual measured feature vector linear correlation, and They are respectively and The mean, Used to evaluate the degree of linearity between simulated features and actual features, providing a reference for model optimization; To simulate the energy of eigenvectors, The energy of the actual measured eigenvector is ∈; ∈ is a very small positive number, such as ∈=1e-6, used to avoid the case where the denominator is zero, and to ensure the mathematical rationality of the formula when calculating the energy difference ratio.

[0089] Among them, KL divergence The calculation formula is:

[0090]

[0091] Where I represents the dimension of the feature vector. Assuming that in this application scenario, the feature vector dimension I = 10, it covers key features such as amplitude and phase in different frequency bands.

[0092] Correlation coefficient The calculation formula is:

[0093]

[0094] The formulas for calculating the energy of eigenvectors in both simulated and actual measurements are as follows:

[0095]

[0096] Where J represents the feature dimension used to calculate energy. Assuming J=5, key feature dimensions related to energy are selected for calculation. By calculating the energy of the simulated feature vector, the accuracy of the model simulation is evaluated from the energy perspective, which helps the model training. The energy of the actual measured feature vector corresponds to the energy of the simulated feature vector. By comparing the energy difference between the two, the model can better learn the features of the real signal.

[0097] Furthermore, the calculation of the loss function aims to measure the difference between the model's predictions and the actual situation by integrating multiple aspects, in order to optimize the performance of the physically constrained deep learning model in the acoustic emission signal path separation task. First, the cross-entropy loss term (containing the sigmoid function) is used to measure the accuracy of the model in classifying the signal path, i.e., determining whether the signal path is a real acoustic emission signal or noise. Then, the weight regularization term (based on the difference between the elements of the weight matrix and the mean) is used to prevent the model from overfitting, giving the model better generalization ability. Next, the KL divergence term and the correlation coefficient term are introduced to evaluate the consistency between simulated features and actual features from the perspectives of distributional difference and linear correlation, respectively, to further optimize the model's learning of features. Finally, the energy difference term measures the difference between simulated and actual features from the perspective of energy, comprehensively improving the model's ability to fit signal features.

[0098] Furthermore, The construction idea of ​​the calculation formula is as follows: The sigmoid function has the property of mapping any real number to the interval [0, 1], which is consistent with the range of probability values. In deep learning, it is often used to convert the linear output of the neural network into a probabilistic form in order to classify the signal path and determine the probability that it belongs to the real acoustic emission signal path or the noise signal path. Converting the linear output of the model into a probability value that can be used for classification solves the problem of how to combine the neural network output with the actual signal path classification task, so that the model can effectively distinguish between the real acoustic emission signal path and the noise signal path.

[0099] Furthermore, the construction idea of ​​the ReLU(x) = max(0, x) calculation formula is as follows: This function can effectively introduce nonlinear transformation by setting all negative inputs to 0 and retaining positive inputs. When processing complex nonlinear data such as acoustic emission signals, the ReLU function (Linear rectification function) can highlight useful features and suppress useless or negative features, enabling the model to learn more complex feature relationships. This solves the problem of how neural networks can learn more complex feature relationships when processing complex nonlinear data such as acoustic emission signals, improves the expressive power of the model, and enables it to better handle the complex features of multipath signals.

[0100] Furthermore, The calculation formula is constructed as follows: KL divergence is used to measure the difference between two probability distributions; simulated feature vectors are used... and actual measured feature vector Treating them as two distributions, the degree of difference between them is measured by calculating the weighted sum of the logarithmic ratios of their corresponding elements. This solves the problem of how to quantify the distributional differences between simulated features and actual measured features, helps the model understand the degree of deviation between its simulated features and real features, and thus adjusts the parameters to make the simulated features closer to the real features, thereby improving the model's learning accuracy of acoustic emission signal features.

[0101] Furthermore, The construction idea of ​​the calculation formula is as follows: the correlation coefficient measures the linear correlation between two variables (simulated feature vector and actual measured feature vector) by calculating the ratio of the product of their covariance and standard deviation. In the calculation process, the feature vector is first centered (subtracting the mean), then the sum of the products is calculated, and then normalized by dividing by the product of the standard deviations. This solves the problem of how to evaluate the strength of the linear relationship between simulated features and actual measured features, enabling the model to optimize the learning of acoustic emission signal features from the perspective of linear correlation, improve the model's ability to fit the real signal features, and thus improve the accuracy of multipath signal separation.

[0102] Furthermore, in the research on early warning of wind turbine gearbox rupture based on acoustic emission localization technology, multipath signals interfere with each other, making it difficult to accurately separate signals from different paths. The calculation of the loss function value, by comprehensively measuring the loss from multiple aspects, helps the deep learning model learn accurate signal path characteristics, improves the accuracy of multipath signal separation, and thus provides support for accurately identifying the propagation path of the acoustic emission source, which helps to more accurately determine whether the gearbox has ruptures or other faults.

[0103] Furthermore, the range of the above loss function is a non-negative real number. The smaller the value, the better the GCN model learns the signal features of different paths, and the more accurately it can identify real and noisy signal paths. By comprehensively considering multiple aspects such as weight regularization, differences between simulated and measured features, correlation, and energy differences, the signal path separation effect of deep learning based on physical constraints is comprehensively improved. The larger the value, the worse the model's learning effect and the lower the accuracy of signal path identification.

[0104] Furthermore, during training, the model continuously adjusts the weight matrix W by minimizing this loss function. k The parameters enable the model to learn signal features from different paths better and better; the training process is monitored using a validation set to prevent overfitting; if the value of the loss function starts to rise on the validation set, it indicates that the model may be overfitting, and at this time it is necessary to adjust the training parameters or use regularization and other methods to optimize the model performance.

[0105] Furthermore, before practical application, a large amount of data is needed to train the physical constraint-based deep learning model. Specific steps include:

[0106] 1) Data preparation: Collect a large amount of acoustic emission signal data of wind turbine gearbox under different operating conditions (normal, minor fault, severe fault), generate training dataset, determine the number of samples P, and preprocess the data, including normalization, to ensure data consistency and usability. At the same time, determine parameters such as the dimension I of the feature vector and the feature dimension J used to calculate energy according to the actual situation.

[0107] 2) Model parameter initialization: Set the number of graph convolutional layers K, and initialize the weight matrix W. k The weight regularization coefficients λ1, λ2, λ3, and λ4 are determined through multiple cross-validation experiments to balance the contributions of different terms in the loss function.

[0108] 3) Training process: Input the samples from the training dataset into the model sequentially, and calculate the loss function value based on the current model parameters. The weight matrix W is adjusted based on the loss function value using the backpropagation algorithm. k The parameters are adjusted to gradually decrease the loss function value. During training, the loss function value on the validation set is monitored in real time. If the loss on the validation set starts to rise, it indicates that the model may be overfitting and training parameters (such as the learning rate) need to be adjusted or regularization methods are used to optimize the model.

[0109] Step S104: Based on the acoustic wave propagation simulation data and the classification probability data of multipath signal paths, a reinforcement learning agent model is used to detect fault signals and obtain early warning information of wind turbine gearbox rupture.

[0110] Specifically, the original acoustic emission signal is processed by the acoustic wave propagation model and the GCN model to separate the real signal path from the noise path, thereby improving signal purity and positioning accuracy. The preprocessed signal features are input into the reinforcement learning agent model, which determines whether the signal is a fault signal based on the current threshold: if the signal features (such as amplitude and phase) exceed the dynamically adjusted threshold and meet other state conditions (such as low false alarm rate and high signal path separation accuracy), the system triggers a rupture warning; if the threshold is not exceeded, it is regarded as normal operation or noise interference, and no warning is triggered.

[0111] The wind turbine gearbox rupture early warning method based on acoustic emission localization technology provided in this embodiment constructs an acoustic wave propagation model based on the internal structural parameters of the wind turbine gearbox. This model is then used to simulate acoustic wave propagation at multiple acoustic emission source locations, obtaining acoustic wave propagation simulation data. This data allows for more accurate prediction of the propagation path and characteristics of acoustic emission signals under multipath effects, and more precise identification of useful information and interference components in the signal, improving the reliability of signal analysis. Secondly, a graph convolutional network model is used to separate the real-time acoustic emission signal path, providing a cleaner and more accurate signal for acoustic emission signal processing and fault diagnosis, thus improving the accuracy of fault diagnosis. Finally, based on the acoustic wave propagation simulation data and the classification probability data of multipath signal paths, a reinforcement learning surrogate model is used for fault signal detection. This enables the system to automatically optimize the fault detection strategy according to actual operating conditions, effectively reducing the false alarm rate and significantly improving the accuracy and reliability of the early warning system, providing a more reliable guarantee for wind turbine gearbox rupture early warning.

[0112] This embodiment provides a wind turbine gearbox rupture early warning method based on acoustic emission positioning technology, which can be used in the aforementioned electronic equipment. Figure 2 This is a flowchart of a wind turbine gearbox rupture early warning method based on acoustic emission positioning technology according to an embodiment of the present invention, such as... Figure 2 As shown, the process includes the following steps:

[0113] Step S201: Obtain the internal structural parameters of the wind turbine gearbox and construct a sound wave propagation model based on the internal structural parameters of the wind turbine gearbox.

[0114] Specifically, in the complex environment of wind turbine gearboxes, multipath interference severely affects the accuracy of acoustic emission positioning. To effectively suppress multipath interference, it is necessary to first accurately simulate the sound wave propagation process and then optimize the positioning algorithm based on the simulation results.

[0115] The above step S201 includes:

[0116] Step S2011: Obtain the three-dimensional parameters of each component inside the wind turbine gearbox, and construct a three-dimensional physical model of the wind turbine gearbox based on the three-dimensional parameters of each component inside the wind turbine gearbox.

[0117] Specifically, professional 3D modeling software, such as SolidWorks and ANSYS Design Modeler, is used to accurately create a 3D model of the internal structure of the wind turbine gearbox based on its design drawings and actual dimensions. During the modeling process, each structural component is depicted in detail, including but not limited to the shape, size, and relative position of gears, shafts, bearing housings, and partitions, ensuring that the model closely matches the actual gearbox structure.

[0118] Step S2012: Obtain the acoustic parameters of each component inside the wind turbine gearbox and assign the acoustic parameters of each component inside the wind turbine gearbox to the three-dimensional physical model of the wind turbine gearbox.

[0119] Specifically, by consulting relevant material handbooks, acoustic databases, or conducting experimental tests, the acoustic parameters of the materials used in each component of the gearbox, such as sound velocity, density, and attenuation coefficient, are obtained. For components made of different materials, corresponding accurate parameters are assigned; for example, for metal gears, the corresponding sound velocity and attenuation coefficient are determined based on their specific alloy composition; for seals made of plastic or rubber, appropriate acoustic parameters are also set according to their material properties.

[0120] Step S2013: Perform finite element analysis on the three-dimensional physical model of the wind turbine gearbox after assigning acoustic parameters to each component inside the gearbox to obtain the sound wave propagation model.

[0121] Specifically, the completed 3D model is imported into finite element analysis software, such as ANSYS or COMSOL Multiphysics. An acoustic analysis module is defined within the software, and appropriate boundary conditions are set, such as assuming the gearbox housing as a hard acoustic boundary to simulate the reflection of sound waves at the boundary. Simultaneously, an excitation source is set to simulate the sound waves generated by the acoustic emission source. By adjusting analysis parameters, such as mesh generation accuracy and time step, the accuracy and reliability of the simulation results are ensured.

[0122] Step S202: Use the sound wave propagation model to simulate sound wave propagation at multiple sound emission source locations to obtain sound wave propagation simulation data.

[0123] Specifically, finite element analysis software was used to simulate the propagation of sound waves inside the gearbox. The simulation results were analyzed to observe the reflection and refraction paths of the sound waves, as well as the sound pressure distribution and phase changes at different locations. By simulating sound emission sources at different locations multiple times, comprehensive sound wave propagation data was obtained, providing detailed propagation model data support for subsequent localization algorithms.

[0124] Step S203: Acquire the real-time acoustic emission signal, and use a graph convolutional network model to perform path separation on the real-time acoustic emission signal to obtain the classification probability data of the multipath signal path. For details, please refer to [link to relevant documentation]. Figure 1Step S103 of the illustrated embodiment will not be described again here.

[0125] Step S204: Based on the acoustic wave propagation simulation data and the classification probability data of multipath signal paths, a reinforcement learning surrogate model is used to detect fault signals and obtain early warning information for wind turbine gearbox rupture. For details, please refer to [link to relevant documentation]. Figure 1 Step S104 of the illustrated embodiment will not be described again here.

[0126] This embodiment provides a wind turbine gearbox rupture early warning method based on acoustic emission localization technology. By performing detailed three-dimensional modeling of the complex internal structure of the wind turbine gearbox and assigning accurate acoustic wave propagation parameters to each component, the method uses finite element analysis to accurately simulate the acoustic wave propagation process. This enables more precise prediction of the propagation path and characteristics of acoustic emission signals under multipath effects. Based on the acoustic wave propagation model, the subsequent localization algorithm can obtain a more reliable data foundation, thereby effectively reducing the localization error caused by multipath interference and significantly improving the localization accuracy of the acoustic emission source.

[0127] This embodiment provides a wind turbine gearbox rupture early warning method based on acoustic emission positioning technology, which can be used in the aforementioned electronic equipment. Figure 3 This is a flowchart of a wind turbine gearbox rupture early warning method based on acoustic emission positioning technology according to an embodiment of the present invention, such as... Figure 3 As shown, the process includes the following steps:

[0128] Step S301: Obtain the internal structural parameters of the wind turbine gearbox, and construct a sound wave propagation model based on these parameters. For details, please refer to [link to relevant documentation]. Figure 2 Step S201 of the illustrated embodiment will not be described again here.

[0129] Step S302: Simulate sound wave propagation at multiple sound emission source locations using a sound wave propagation model to obtain sound wave propagation simulation data. For details, please refer to [link to relevant documentation]. Figure 2 Step S202 of the illustrated embodiment will not be described again here.

[0130] Step S303: Obtain the real-time acoustic emission signal, and use a graph convolutional network model to perform path separation on the real-time acoustic emission signal to obtain the classification probability data of the multipath signal path.

[0131] Specifically, step S303 includes:

[0132] Step S3031: Perform signal preprocessing on the real-time acoustic emission signal, and extract features from the preprocessed real-time acoustic emission signal to obtain signal propagation features.

[0133] Specifically, multiple sensors are strategically arranged around the wind turbine gearbox to ensure the acquisition of acoustic emission signals from different directions and locations. When an acoustic emission event occurs, each sensor synchronously acquires signals, recording initial data such as the time and intensity of the signal received by each sensor. The acquired raw signals are preprocessed, including noise removal, for example, by using filtering algorithms such as low-pass filters to remove high-frequency noise and high-pass filters to remove low-frequency noise, in order to improve signal quality. At the same time, the signals are normalized to unify the amplitude of signals acquired by different sensors to the same scale range, which facilitates subsequent analysis.

[0134] Furthermore, features related to the propagation path can be extracted from the preprocessed signal. For example, based on the time difference of arrival of the signal, the time it takes for the signal to arrive at different sensors will differ depending on the path. By accurately measuring and calculating this time difference, the propagation direction of the signal can be preliminarily inferred. The phase change of the signal can also be analyzed. Since sound waves may undergo different reflections and refractions during propagation along different paths, the phase changes. Extracting phase features helps to distinguish different paths. The frequency component features of the signal can also be extracted. Different paths may have different effects on the frequency of the signal. Some frequency components may be enhanced or weakened on a specific path.

[0135] Step S3032: Based on the preprocessed acoustic emission real-time signal, the classification probability data of the multipath signal path is determined using a graph convolutional network model.

[0136] Specifically, after the GCN model is trained, it is applied to perform signal path separation during the real-time operation of the wind turbine gearbox. The specific steps include:

[0137] 1) Real-time signal preprocessing: Acquire acoustic emission signals in real time, perform preprocessing operations such as filtering and denoising to remove obvious noise interference and make the signals input into the model purer.

[0138] 2) Signal path separation and analysis: The preprocessed signal is input into the trained model. The model separates the multipath signal based on the learned features. By analyzing the signal components of different paths output by the model and combining the information provided by the accurate propagation model, the possible propagation paths of the acoustic emission source are determined. For example, if the energy, phase and other features of a certain path signal output by the model match the known fault signal features, and the path is consistent with the possible propagation paths of the fault in the accurate propagation model, it can be further determined that the acoustic emission signal corresponding to the path may be related to the fault.

[0139] Furthermore, the GCN model separates the input multipath signal based on the learned characteristics of the reflected / refracted signal. The model outputs the signal components of different paths. By analyzing these signal components, the possible propagation paths of the acoustic emission source can be determined. Combined with the information provided by the acoustic wave propagation model, the judgment of the signal path can be further optimized, and the accuracy of multipath signal separation can be improved.

[0140] Furthermore, using the established three-dimensional physical model, based on the geometry of the gearbox's internal structure, component positions, and material acoustic properties, the propagation of sound waves along different possible paths is simulated. The simulated signal characteristics, such as propagation time, phase change, and frequency characteristics, are compared with the actual acquired and extracted signal characteristics. Based on the feature comparison results, matching algorithms, such as correlation analysis, are used to find the simulated propagation path that best matches the actual signal characteristics. This matching analysis is performed on the signal received by each sensor, and then the results from multiple sensors are combined to determine the possible propagation paths of the acoustic emission source. If the matching results from multiple sensors all point to certain similar paths, then these paths are more likely to be the actual propagation paths of the acoustic emission source. After determining the possible propagation paths, the probability of these paths can be further ranked based on information such as signal strength, providing a more accurate basis for subsequent localization algorithms.

[0141] Furthermore, the final output layer of the GCN model is activated by the Sigmoid function, which maps the linear transformation result of the model to the [0,1] interval and outputs the probability value of each sample (signal path) belonging to the real acoustic emission signal path. If the output value is close to 1, it means that the model believes that the path is likely to be a real signal path; if it is close to 0, it is judged as a noise or interference path.

[0142] Furthermore, in the acoustic emission signals of the gearbox, the signals generated by real faults usually have specific propagation paths (such as direct propagation through the gear meshing surface), and their characteristics have a high degree of matching with the regular reflection / refraction patterns simulated in the virtual training data, with the model output probability close to 1; the propagation paths of noise signals (such as environmental vibration and electrical interference) are irregular, and their characteristics match the noise patterns in the training data, with the model output probability close to 0; through the above methods, the GCN model transforms the propagation laws simulated by the physical model (such as finite element analysis results) into calculable probability judgments, providing reliable signal preprocessing results for subsequent acoustic emission source localization and fault early warning.

[0143] Furthermore, the physically constrained deep learning method, after separating multipath signals through a graph convolutional network (GCN), outputs the following key information: signal path classification result: the probability of distinguishing between real acoustic emission signal paths and noise paths (output via the sigmoid function); separation accuracy (A). sep，t): Reflects the model's ability to separate multipath signals (e.g., the proportion of correctly classified paths).

[0144] Step S304: Based on the acoustic wave propagation simulation data and the classification probability data of multipath signal paths, a reinforcement learning surrogate model is used to detect fault signals and obtain early warning information for wind turbine gearbox rupture. For details, please refer to [link to relevant documentation]. Figure 2 Step S204 of the illustrated embodiment will not be described again here.

[0145] This embodiment provides a wind turbine gearbox rupture early warning method based on acoustic emission localization technology. It uses a graph convolutional network model to separate the real-time acoustic emission signal path, reducing the contamination of features (such as amplitude and phase) by noise. The separated real signal path features are closer to the actual fault signal, making the "fault-noise" boundary clearer when adjusting the threshold.

[0146] This embodiment provides a wind turbine gearbox rupture early warning method based on acoustic emission positioning technology, which can be used in the aforementioned electronic equipment. Figure 4 This is a flowchart of a wind turbine gearbox rupture early warning method based on acoustic emission positioning technology according to an embodiment of the present invention, such as... Figure 4 As shown, the process includes the following steps:

[0147] Step S401: Obtain the internal structural parameters of the wind turbine gearbox, and construct a sound wave propagation model based on these parameters. For details, please refer to [link to relevant documentation]. Figure 3 Step S301 of the illustrated embodiment will not be described again here.

[0148] Step S402: Simulate sound wave propagation at multiple sound emission source locations using a sound wave propagation model to obtain sound wave propagation simulation data. For details, please refer to [link to relevant documentation]. Figure 3 Step S302 of the illustrated embodiment will not be described again here.

[0149] Step S403: Acquire the real-time acoustic emission signal, and use a graph convolutional network model to perform path separation on the real-time acoustic emission signal to obtain the classification probability data of the multipath signal path. For details, please refer to [link to relevant documentation]. Figure 3 Step S303 of the illustrated embodiment will not be described again here.

[0150] Step S404: Based on the acoustic wave propagation simulation data and the classification probability data of multipath signal paths, a reinforcement learning agent model is used to detect fault signals and obtain early warning information of wind turbine gearbox rupture.

[0151] Specifically, step S404 includes:

[0152] Step S4041: Optimize the location of the sound wave propagation simulation data and determine the multipath interference level index.

[0153] In some optional implementations, step S4041 above includes:

[0154] Step a1: Based on the acoustic wave propagation simulation data, determine the location vector of the acoustic emission source using a positioning algorithm.

[0155] Specifically, after obtaining the sound wave propagation simulation data, the following formula is used to optimize the positioning accuracy:

[0156]

[0157] in, For the estimated acoustic emission source position vector The estimated position of the acoustic emission source in three-dimensional space is used to determine the location of potential ruptures or other faults inside the gearbox; r is the variable vector of the acoustic emission source position (r = (x, y, z)), used to find the variable that minimizes the objective function of the acoustic emission source position, and the optimal acoustic emission source position estimate is searched by changing its value; N is the number of sensors, representing the number of sensors actually arranged around the wind turbine gearbox to receive acoustic emission signals. These sensors are used to obtain information such as the arrival time and amplitude of the acoustic emission signals, providing data support for locating the acoustic emission source. The time it takes for the sound wave to reach the i-th sensor is obtained from the simulation based on the sound wave propagation model. The actual measured time of the sound wave reaching the i-th sensor is obtained through actual sensor acquisition and is used to compare with the simulated time to evaluate the accuracy of the simulation and determine the location of the sound emission source. λ represents the standard deviation of the time measurement of the i-th sensor, reflecting the uncertainty of the arrival time of the sound wave measured by that sensor. It is used to measure the fluctuation of the time measurement value and is used in the formula to normalize the time difference, making the time errors of different sensors comparable. λ is a coefficient that balances the effects of time and amplitude. By adjusting this coefficient, the relative importance of time difference and amplitude difference in the objective function can be changed to adapt to different measurement environments and needs, and optimize the estimation of the acoustic emission source location. M is the number of signal amplitude sampling points, used to quantify the number of samples of the acoustic emission signal amplitude. By analyzing the amplitude of multiple sampling points, the amplitude characteristics of the acoustic emission signal can be described more comprehensively. The amplitude of the acoustic emission signal at the j-th sampling point is obtained based on the acoustic wave propagation model simulation. The amplitude of the acoustic emission signal at the j-th sampling point is obtained through actual measurement and is used to compare with the simulated amplitude to help determine the location of the acoustic emission source. denoted as , where is the standard deviation of the amplitude measurement at the j-th sampling point, reflecting the uncertainty of the amplitude measurement at that sampling point. It is used to normalize amplitude differences and enhance the comparability of amplitude errors at different sampling points. μ is the coefficient for balancing the phase influence, used to balance the contribution of the phase difference to the objective function, enabling the comprehensive consideration of time, amplitude, and phase information when locating acoustic emission sources, thereby improving positioning accuracy. L is the number of signal phase sampling points, used to quantify the number of phase samples of the acoustic emission signal. Analysis of multiple phase sampling points provides a more comprehensive description of the phase characteristics of the acoustic emission signal. The phase of the acoustic emission signal at the l-th sampling point is obtained based on the acoustic wave propagation model simulation. The phase of the acoustic emission signal at the l-th sampling point is obtained through actual measurement and is used to compare with the simulated phase to help locate the acoustic emission source. The standard deviation of the phase measurement at the l-th sampling point reflects the uncertainty of the phase measurement and is used to normalize the phase difference, making the phase errors of different sampling points comparable.

[0158] in, The calculation formula is used to accurately simulate the time required for sound waves to propagate to each sensor within the complex structure of a wind turbine gearbox. It considers the influence of different material regions on the sound velocity, thus more accurately reflecting the actual propagation situation. The time it takes for the sound wave to reach the i-th sensor is simulated based on the sound wave propagation model. The calculation formula is as follows:

[0159]

[0160] Among them, s i (r) represents the propagation path from the acoustic emission source r to the i-th sensor, and v(s) represents the sound speed along the path s. By assigning corresponding sound speed values ​​to different material regions, the influence of material properties on sound speed is reflected.

[0161] Furthermore, The calculation process comprehensively considers the influence of factors such as attenuation, distance, and frequency on the amplitude during signal propagation, accurately simulates amplitude changes, and obtains the acoustic emission signal amplitude at the j-th sampling point based on the acoustic wave propagation model. The calculation formula is as follows:

[0162]

[0163] Where A0 is the initial acoustic emission signal amplitude, α is the attenuation coefficient, and l j (r) represents the propagation distance from the acoustic emission source r to the j-th sampling point, ω represents the angular frequency of the acoustic emission signal, and Δt j (r) represents the propagation time difference from the acoustic emission source r to the j-th sampling point.

[0164] Furthermore, To accurately simulate the phase change of acoustic emission signals during propagation, the phase of the acoustic emission signal at the l-th sampling point is simulated based on the acoustic wave propagation model. The calculation formula is as follows:

[0165]

[0166] in, Let r be the propagation time from the acoustic emission source r to the l-th sampling point, as simulated based on the acoustic wave propagation model. This is the initial phase.

[0167] Furthermore, the core idea of ​​the formula for calculating the acoustic emission source location vector is to determine the most probable location of the acoustic emission source by minimizing the differences between the simulated time, amplitude, and phase values ​​and the actual measured values. Specifically, the first term in the formula measures the difference between the simulated time and the actual measured time, the second term measures the difference between the simulated amplitude and the actual measured amplitude, and the third term measures the difference between the simulated phase and the actual measured phase. Taking into account the time of sound wave propagation to the sensor, signal amplitude, and phase information, the propagation of the above parameters in the complex structure inside the gearbox is simulated using a sound wave propagation model and compared with the actual measured values. By adjusting the weighting coefficients λ and μ of the time, amplitude, and phase difference terms, the importance of these three differences in positioning can be flexibly balanced according to the actual situation. For example, in some cases, time measurement may be more accurate, in which case λ can be appropriately increased to make the time difference play a more critical role in positioning.

[0168] Furthermore, in the study of wind turbine gearbox rupture early warning based on acoustic emission positioning technology, multipath interference can lead to complex acoustic emission signal propagation paths, making it difficult for the positioning method to accurately determine the location of the acoustic emission source. The above formula (6) effectively suppresses the influence of multipath interference on positioning by accurately simulating sound wave propagation and comprehensively comparing multiple parameters, thereby improving the positioning accuracy of the acoustic emission source and more accurately determining the location where the gearbox may rupture, providing a basis for timely early warning and maintenance.

[0169] Furthermore, The calculation formula is constructed based on the fact that the sound wave propagates through a complex path inside the gearbox and passes through different material regions, causing the sound speed to change. This function calculates the propagation path s from the sound emission source r to the i-th sensor. i (r) is integrated and combined with the sound speed v(s) at each point on the path to accurately calculate the sound wave propagation time to simulate the actual propagation situation; it accurately simulates the propagation time of sound waves in complex structures, solves the problem of the difficulty in accurately determining the propagation time under multi-path interference, and provides accurate time parameters for the location of acoustic emission sources.

[0170] Furthermore, The calculation formula is constructed based on the following idea: During the propagation of an acoustic emission signal, its amplitude is affected by factors such as propagation distance, attenuation, and frequency. This function takes an initial amplitude A0 and considers the propagation distance l. j (r) caused attenuation (by (represented by), and frequency ω and propagation time difference Δt j (r) Modulation of amplitude (by...) (Indicated) It accurately simulates the amplitude changes of signals; it accurately simulates the amplitude changes of acoustic emission signals during propagation, solves the problem of difficulty in accurately simulating signal amplitude under multipath interference, and provides accurate amplitude parameters for locating acoustic emission sources.

[0171] Furthermore, The calculation formula is based on the following idea: the phase of the acoustic emission signal is related to the propagation time and the initial phase. This function is calculated by multiplying the angular frequency ω of the acoustic emission signal by the propagation time. In addition to the initial phase It accurately simulates signal phase changes; precisely simulates the phase changes of acoustic emission signals during propagation, solves the problem of difficulty in accurately simulating signal phase under multipath interference, and provides accurate phase parameters for locating acoustic emission sources.

[0172] Step a2: Based on the position vector of the acoustic emission source, the propagation parameters are simulated to obtain the basic simulation parameters of sound wave propagation inside the gearbox.

[0173] Specifically, the obtained acoustic emission source position vector is a fundamental parameter for simulating the propagation of sound waves inside the gearbox. The fundamental parameters for simulating sound wave propagation inside the gearbox include: simulated propagation time. based on Using a gearbox structural model, calculate the theoretical time for the sound wave to reach the i-th sensor; simulate the signal amplitude. based on Amplitude attenuation after analog signal propagation; Phase of analog signal based on Phase change after analog signal propagation.

[0174] Step a3: Obtain the actual measurement parameters of sound wave propagation inside the gearbox. Compare the simulated parameters of sound wave propagation inside the gearbox with the actual measurement parameters of sound wave propagation inside the gearbox to obtain the multipath interference level index.

[0175] Specifically, the above-mentioned simulated basic parameters are compared with the actual measured values. By comparison, a multipath interference level index can be obtained. The formula for calculating the multipath interference level index is as follows:

[0176]

[0177] In the above formula, Error mp，t It is a comprehensive deviation index between simulated and measured values ​​in multipath interference suppression, used to assess the impact of multipath interference on the simulation of acoustic emission signal propagation; t represents the time when the sound wave arrives at the i-th sensor, based on simulation and actual measurement, respectively; Let be the standard deviation of the time measurement of the i-th sensor at time t; are the acoustic emission signal amplitudes at the j-th sampling point at time t, based on simulation and actual measurement, respectively; Let be the standard deviation of the amplitude measurement at the j-th sampling point at time t; These are the phases of the acoustic emission signal at the l-th sampling point at time t, based on simulation and actual measurement, respectively. Let N be the standard deviation of the phase measurement at the l-th sampling point at time t; N, M, and L are the number of sensors, the number of amplitude sampling points, and the number of phase sampling points, respectively. mp，t The smaller the value, the better the multipath interference suppression effect and the greater the positive contribution to the reward value.

[0178] Furthermore, considering the differences between simulated and measured values ​​of the time, signal amplitude, and phase of sound wave propagation to the sensor under multipath interference, a comprehensive deviation index is obtained by calculating the deviations of time, amplitude, and phase separately and averaging them according to the number of sensors and sampling points. This comprehensively measures the degree of influence of multipath interference on the simulation of acoustic emission signal propagation. Accurately assessing the impact of multipath interference on the simulation of acoustic emission signal propagation provides a quantitative index for the reward function that reflects the effect of multipath interference suppression. This helps the reinforcement learning agent understand the impact of the current multipath interference situation on fault detection, thereby making more reasonable threshold adjustment decisions.

[0179] Furthermore, the multipath interference index, as the state input of the reinforcement learning agent model, is used to characterize the reliability of the current acoustic emission signal propagation. For example, if Error mp，t A large value indicates severe multipath interference and low reliability of the localization results. The fault detection threshold needs to be increased to avoid false alarms. If the error value is large... mp，t A smaller value indicates high positioning accuracy and reliable signal characteristics, and the threshold can be appropriately reduced to improve fault detection sensitivity. The multipath interference index is embedded in the reward function to guide threshold optimization, thereby driving dynamic threshold adjustment decisions.

[0180] Step S4042: Determine the signal separation accuracy based on the classification probability data of the multipath signal path.

[0181] Step S4043: Use a reinforcement learning agent model to determine threshold adjustment strategies under various wind turbine operating conditions.

[0182] Specifically, the system uses a deep Q-network (DQN) agent to analyze the wind turbine's operating conditions (such as wind speed, load, and temperature), multipath interference level (Error_{mp,t}), and signal path separation accuracy (A_{sep,t}) in real time, and dynamically adjusts the fault trigger threshold. The threshold is a key parameter for determining whether the acoustic emission signal is a fault signal. When the acoustic emission signal characteristics (such as amplitude and frequency) collected by the sensor exceed the threshold, the system determines that there may be a fault such as gearbox breakage, and then triggers an early warning.

[0183] In some optional implementations, step S4043 above includes:

[0184] Step b1 defines the state, action, and reward function; where the state consists of historical false alarm data, wind turbine operating parameters, multipath interference level indicators, and signal path separation accuracy corresponding to various wind turbine operating conditions; the action is the adjustment operation of the fault trigger threshold.

[0185] Specifically, the state space of the DQN (Deep Q Network, a deep reinforcement learning) surrogate model is determined. The state includes historical false alarm data and current operating condition information. Historical false alarm data includes information such as the number of false alarms, the time of false alarm, and the signal characteristics at the time of false alarm over a period of time. Current operating condition information includes wind turbine operating parameters such as wind speed, load, and temperature, as well as multipath interference level indicators and signal path separation accuracy. The above information is encoded to form a state vector that the DQN surrogate model can understand.

[0186] Furthermore, the action space of the DQN agent model is defined, where an action represents an adjustment operation on the fault trigger threshold. For example, a series of discrete threshold adjustment steps are set, such as increasing or decreasing by 5dB, 10dB, etc. The DQN agent model adjusts the fault trigger threshold by selecting different actions.

[0187] Furthermore, the reward function uses the number of actual fault detections as a positive reward and the number of false alarms as a negative reward, guiding the reinforcement learning agent model to optimize the threshold to balance detection sensitivity and reliability.

[0188] Furthermore, the reward function guides the DQN agent model to learn the optimal threshold adjustment strategy. The design principle of the reward function is as follows: a positive reward is given when the agent's threshold adjustment operation reduces false positives and does not miss any real faults; a negative reward is given when false positives or missed faults occur. For example, if in a certain state, after the agent adjusts the threshold, the number of false positives decreases and all real faults are successfully detected, a large positive reward is given; if false positives or missed faults still exist, a corresponding negative penalty is given. The reward value r... t Calculate using the following formula:

[0189]

[0190] Where, r t For state s t Take action a t The reward value obtained is determined by comprehensively considering multiple factors to guide the reinforcement learning agent to learn the optimal threshold adjustment strategy. The higher the reward value, the more beneficial the action is to optimizing fault detection performance. The number of actual faults detected at time t is obtained by comparing and statistically analyzing them with actual fault records. It is an important indicator for measuring the accuracy of fault detection. The more actual faults detected, the greater the positive contribution to the reward value. βt represents the number of false alarms at time t, determined by comparison with the actual situation. Fewer false alarms contribute more positively to the reward value; reducing false alarms is one of the important goals of optimizing the fault early warning system. β1 and β2 are weighting coefficients used to measure the impact of actual fault detection and false alarms on the reward value, respectively. mp,t β3 represents the comprehensive deviation index between simulated and measured values ​​in multipath interference suppression (i.e., the multipath interference degree index); β3 represents the contribution weight of multipath interference suppression effect to the reward value; A sep,t β4 represents the accuracy of signal path separation based on physical constraints in deep learning. It is calculated by comparing the processed signal path classification results at the current time step with the true labels, and the proportion of correctly classified paths is used as the accuracy. A higher accuracy indicates better signal path separation, positively contributing to the reward value and helping to improve the accuracy of fault detection. β4 represents the impact of signal path separation on the reward value. (Noise) int，t β5 represents the cabin interior noise level; β5 represents the effect of noise on the reward value; State f，t and State gb，t These are the operating status indicators for the wind turbine and gearbox, respectively; β6 and β7 represent the weights of the wind turbine and gearbox operating status on the reward value.

[0191] Among them, the cabin interior noise index is Noise int，t The calculation formula is:

[0192]

[0193] In the above formula, N(t, f) is the interference noise signal at frequency f at time t, S(t, f) is the acoustic emission signal at frequency f at time t, and f1 and f2 are the frequency ranges of interest. Assuming f1 = 100Hz and f2 = 1000Hz, the values ​​of f1 and f2 are determined based on the main frequency range of the acoustic emission signal from the wind turbine gearbox. By calculating the ratio of interference noise energy to total signal energy within this frequency range, the influence of interference noise inside the nacelle on the acoustic emission signal is evaluated. The smaller this index value, the greater its contribution to the reward value, which means that the interference noise has less impact on the acoustic emission signal, thus improving the accuracy of fault detection.

[0194] Noise level inside the cabin int，t In the calculation process, the influence of internal engine room noise on acoustic emission signals is quantified by calculating the ratio of interference noise energy to total signal energy within a specific frequency range. f1 and f2, corresponding to the main frequency range of acoustic emission signals from the wind turbine gearbox, are selected so that the index can specifically reflect the level of interference noise that has a real impact on fault detection. The accuracy of the evaluation of the interference degree of internal engine room noise on acoustic emission signals provides a quantitative index for the reward function that reflects the impact of interference noise, helping the reinforcement learning agent to consider interference noise factors when adjusting the threshold and improving the accuracy of fault detection in complex noise environments.

[0195] Where, State f，t and State gb，t The calculation formula is:

[0196]

[0197] In the above formula, State f，t For wind turbine operating status indicators; F i (t) represents the i-th characteristic parameter of the wind turbine's operating state at time t, such as wind speed, rotational speed, etc. This is the normal mean of the characteristic parameter; I f The number of characteristic parameters of the wind turbine's operating state; assuming I f =3, corresponding to wind speed, rotational speed, and power, respectively; by calculating the relative deviation of the current characteristic parameters from the normal average, the operating status of the wind turbine is evaluated. The closer the operating status is to normal, the greater its contribution to the reward value, which helps to reasonably adjust the fault trigger threshold under different operating conditions of the wind turbine; State gb，t For gearbox operating status indicators; GB j (t) represents the j-th characteristic parameter of the gearbox operating state at time t, such as the wear degree of the gears, oil temperature, etc. This is the normal mean of the characteristic parameter; I gb The number of characteristic parameters of the gearbox operating state; assuming I gb=4, corresponding to gear wear, oil temperature, oil quality index and vibration amplitude respectively. By calculating the relative deviation of the current characteristic parameters from the normal average, the operating status of the gearbox is measured. The more stable the operating status, the greater the positive impact on the reward value, which helps to reasonably adjust the fault trigger threshold under different operating conditions of the gearbox.

[0198] State f，t In the calculation process, key characteristic parameters of the wind turbine's operating status (such as wind speed, rotational speed, and power) are selected. The operating status of the wind turbine is evaluated by calculating the relative deviation between the current characteristic parameters and the normal average. This method can intuitively reflect the degree of deviation between the wind turbine's operating status and the normal state, providing quantitative information on the wind turbine's operating status for the reinforcement learning agent. In other words, it quantifies the wind turbine's operating status, enabling the reinforcement learning agent to adjust the fault trigger threshold according to different wind turbine operating states. This solves the problem of how to reasonably adjust the threshold to ensure the accuracy of fault detection when the wind turbine's operating status changes.

[0199] State gb，t In the calculation process, key characteristic parameters of the gearbox operating status (such as gear wear, oil temperature, oil quality index, and vibration amplitude) are selected. By calculating the relative deviation between the current characteristic parameters and the normal average, the operating status of the gearbox is measured, providing quantitative information on the gearbox operating status for the reinforcement learning agent. That is, quantifying the gearbox operating status, the reinforcement learning agent can adjust the fault trigger threshold according to different gearbox operating states, solving the problem of how to reasonably adjust the threshold to ensure the accuracy of fault detection when the gearbox operating status changes.

[0200] Furthermore, by defining the above reward function, the threshold adjustment operation made by the agent model can reduce false alarms, accurately detect real faults, improve the multipath interference suppression effect, increase the signal path separation accuracy, reduce the impact of internal interference noise in the nacelle, and adapt to the operating status of the wind turbine and gearbox. When these conditions are met, the agent will obtain a higher reward value, thereby guiding the agent to learn the optimal threshold adjustment strategy.

[0201] Furthermore, considering multiple key aspects of wind turbine gearbox fault early warning, factors such as actual fault detection, false alarm control, multipath interference suppression, signal path separation, interference and noise impact, and the operating status of the wind turbine and gearbox themselves are quantified as components of the reward value. The importance of each factor is balanced by weighting coefficients β1-β7, so that the reward function can comprehensively and reasonably reflect the impact of different actions on the performance of the fault early warning system, guiding the reinforcement learning agent to learn the optimal threshold adjustment strategy. This solves the problem of how to comprehensively evaluate the impact of different threshold adjustment actions on the performance of the fault early warning system under complex and variable operating conditions, providing a clear learning direction for the reinforcement learning agent, enabling it to dynamically adjust the threshold according to different operating conditions and improve the accuracy of fault detection.

[0202] Step b2: Obtain the current state, select the current action based on the current state, adjust the fault trigger threshold based on the current action, and update the current state to obtain the updated state.

[0203] Step b3: Calculate the current reward value based on the updated state.

[0204] Specifically, during the operation of the wind turbine, the DQN agent model continuously observes the current state, selects an action (adjusting the threshold), executes the action, observes the new state and the reward obtained, and stores the state, action, reward and new state of each step in the experience replay buffer. When the data stored in the buffer reaches a certain amount, a batch of data is randomly sampled from the buffer for training.

[0205] Step b4: Calculate the current Q value based on the current state, current action, updated state, and current reward value.

[0206] Step b5: Compare the current Q value with the target Q value. Based on the comparison result, the threshold adjustment strategy will be updated iteratively until the current Q value matches the target Q value, thus obtaining the threshold adjustment strategy under various wind turbine operating conditions.

[0207] Specifically, using the sampled data, the error between the target Q value and the current Q value is calculated. The neural network parameters of the DQN proxy model are updated through the backpropagation algorithm, so that the current Q value gradually approaches the target Q value. This process is repeated continuously, and as training progresses, the DQN proxy model gradually learns the optimal threshold adjustment strategy.

[0208] Furthermore, the Q-value update formula is:

[0209]

[0210] Where Q(s) t a t ) is in state s t Take action a t The Q-value, in the study of wind turbine gearbox rupture early warning based on acoustic emission localization technology, comprehensively reflects the expectation of long-term cumulative reward for taking a specific action in the current state, and is an important indicator for reinforcement learning agents to evaluate the value of actions; t This represents the state of the deep Q-network at time t, including historical false alarm data, current operating condition information, and the multipath interference level indicator Error. mp,t Signal path separation accuracy A sep,t Noise level inside the cabin int,t Wind turbine operating status indicators f,t Gearbox operating status indicators gb，tThis comprehensively reflects various relevant information of the system at time t, providing a basis for the agent to make action decisions; a t For state s t The action to be taken, namely the adjustment of the fault trigger threshold, is assumed to have an action space of a. t ∈{Δθ1, Δθ2, ..., Δθ5}, corresponding to threshold increases of 5%, increases of 10%, decreases of 5%, decreases of 10%, and remaining unchanged, respectively. Different actions are selected to dynamically adjust the fault trigger threshold. α is the learning rate, controlling the step size of each update. Experiments show that setting α = 0.1 affects the speed and stability of model learning. A suitable learning rate allows the model to effectively adjust the Q-value during training and gradually learn the optimal policy. γ is the discount factor, determining the importance of future rewards. Its value ranges from 0 to 1, allowing the reinforcement learning agent to consider not only current rewards but also potential future rewards when making decisions. A larger γ value indicates a greater emphasis on future rewards, making the agent more inclined to choose actions that bring high long-term returns. A smaller γ value makes the agent more focused on current rewards. In this study, a suitable γ value helps the agent find the optimal threshold adjustment strategy under different operating conditions to maximize long-term cumulative rewards. For state s t+1 The maximum Q value among all possible actions.

[0211] Among them, the Q-value update is based on the idea of ​​temporal difference learning, that is, Q(s t a t ) represents the current state s t Take action a t Value estimate, r t It is to perform action a t The reward obtained immediately afterward reflects the impact of the action on system performance at that moment. This refers to the state s from the next state. t+1 Starting from the optimal action, the estimated future reward is multiplied by a discount factor γ to reflect the present value of the future reward. This is achieved by calculating the difference between the current estimate and the target value, which includes both the current reward and the estimated future optimal reward. ), and multiply by the learning rate α to update Q(s) t a tAs time goes by and experience accumulates, the agent can continuously adjust its value assessment of each state-action pair, gradually approaching the true value under the optimal strategy. In the dynamic threshold adjustment mechanism based on reinforcement learning, in order for the agent to learn the optimal threshold adjustment strategy under different operating conditions, an effective method is needed to update the value assessment of each state-action pair. The Q-value update process provides an iterative update mechanism, which continuously uses the current reward and the expectation of future rewards to correct the current value estimate, enabling the agent to learn the optimal threshold adjustment action sequence in complex environments (i.e., various operating conditions of wind turbine gearboxes). This solves the problem of how to enable the reinforcement learning agent model to learn an effective strategy in a dynamic environment to optimize fault detection performance.

[0212] Furthermore, by continuously repeating the process of experience replay and network update, as training progresses, the DQN agent model gradually learns the optimal threshold adjustment strategy, enabling it to dynamically adjust the fault trigger threshold according to different operating conditions in order to maximize long-term cumulative rewards.

[0213] Furthermore, through iterative learning, the reinforcement learning agent model gradually masters the optimal threshold strategy under different operating conditions, ensuring that early warnings are triggered in a timely manner when real fault signals are detected, while suppressing false alarms caused by noise.

[0214] Step S4044: Obtain the current wind turbine operating parameters. Based on the current wind turbine operating parameters, multipath interference level index and signal path separation accuracy, determine the fault trigger threshold using threshold adjustment strategies under various wind turbine operating conditions.

[0215] Specifically, the operating conditions of the wind turbine are monitored in real time, including changes in parameters such as wind speed, load, and temperature. The current operating condition information is fed back to the DQN agent model in a timely manner as a basis for the agent to adjust the threshold.

[0216] Furthermore, the DQN agent model selects an action based on the learned threshold adjustment strategy according to the current state, that is, whether to adjust the fault trigger threshold and the adjustment range. For example, during the wind turbine startup phase, when a rapid change in wind speed and a large fluctuation in the amplitude of the acoustic emission signal are detected, the agent automatically increases the threshold according to the learned strategy to avoid false alarms caused by transient noise. As the wind turbine's operating conditions stabilize, the agent adjusts the threshold according to the real-time state to ensure that the system always maintains the best fault detection performance.

[0217] Step S4045: Compare the signal propagation characteristics with the fault triggering threshold. If the signal propagation characteristics exceed the fault triggering threshold, generate a wind turbine gearbox rupture warning message.

[0218] This embodiment provides a wind turbine gearbox rupture early warning method based on acoustic emission localization technology. Based on the principle of reinforcement learning, it defines a state, action, and reward function, allowing a reinforcement learning agent model (deep Q-network) to learn how to adjust the fault trigger threshold under different operating conditions to maximize long-term cumulative rewards. The state includes various information related to the operation of the wind turbine gearbox, the action is the specific adjustment operation of the fault trigger threshold, and the reward function comprehensively considers multiple factors such as the accuracy of fault detection (number of actual fault detections, number of false alarms), multipath interference suppression effect, signal path separation accuracy, internal interference noise in the nacelle, and the operating status of the wind turbine and gearbox. The Q-value update formula continuously adjusts the agent's value assessment of each state-action pair based on the current reward and the estimate of future rewards, enabling the agent to gradually learn the optimal strategy. In the research on wind turbine gearbox rupture early warning based on acoustic emission localization technology, the fixed threshold fault detection method is difficult to adapt to the complex and ever-changing operating conditions of wind turbines, and is prone to false alarms or missed alarms. Therefore, by dynamically adjusting the threshold through reinforcement learning, the fault detection system can automatically optimize the fault trigger threshold according to real-time operating information, effectively solving the limitations of the fixed threshold detection method, improving the accuracy and real-time performance of fault detection, and providing more accurate early warning of wind turbine gearbox rupture.

[0219] The following specific embodiments illustrate the detailed steps of a wind turbine gearbox rupture early warning method based on acoustic emission positioning technology.

[0220] Example 1:

[0221] The specific steps of the wind turbine gearbox rupture early warning method based on acoustic emission localization technology include:

[0222] Step 1: In the complex environment of the wind turbine gearbox, multipath interference severely affects the accuracy of acoustic emission localization. To effectively suppress multipath interference, the core idea is to first accurately simulate the sound wave propagation process, and then optimize the localization algorithm based on the simulation results. The specific steps are as follows:

[0223] I. Steps for establishing a sound wave propagation model:

[0224] 1) 3D modeling:

[0225] Using 3D modeling software (such as SolidWorks, ANSYS Design Modeler, etc.), the internal structure of the wind turbine gearbox is accurately replicated based on the detailed design drawings and actual measured dimensions. This step is crucial because only by accurately reproducing the shape, size, and relative position of the gears, shafts, bearing seats, partitions, and other components inside the gearbox can an accurate physical model foundation be provided for subsequent sound wave propagation simulation.

[0226] Furthermore, during the modeling process, every detail needs to be carefully handled, such as the tooth profile parameters of the gears, the diameter and length of the shaft, the thickness and position of the partition, etc., to ensure that the model closely matches the actual structure. The above steps are like building an accurate virtual gearbox, creating a realistic "stage" for the simulation of sound wave propagation.

[0227] 2) Determination of material parameters:

[0228] By consulting relevant material handbooks and acoustic databases, or by conducting specialized experimental tests, key parameters such as sound velocity, density, and attenuation coefficient of the materials used in each component of the gearbox can be obtained.

[0229] Furthermore, since different materials have significantly different effects on sound wave propagation, for example, the acoustic characteristics of metal parts and plastic parts are completely different, it is crucial to assign accurate acoustic parameters to the specific material of each part to ensure accurate simulation. This step is like setting unique "performance attributes" for each "actor" in the virtual gearbox, so that it can exhibit realistic characteristics in the "performance" of sound wave propagation.

[0230] 3) Finite element analysis settings:

[0231] The completed 3D model (i.e., the 3D physical model of the wind turbine gearbox after assigning acoustic parameters to each component inside the gearbox) is imported into professional finite element analysis software (such as ANSYS or COMSOL Multiphysics). In the software, boundary conditions are set using the acoustic analysis module. For example, assuming that the gearbox shell is an acoustic hard boundary, the reflection of sound waves at the boundary is simulated, which is closer to the actual reflection characteristics of sound waves on the gearbox wall.

[0232] Furthermore, an excitation source is set up to simulate the sound waves generated by the acoustic emission source: by adjusting the analysis parameters, such as performing reasonable mesh division of the model, so that the mesh density can accurately capture the details of sound wave propagation without causing excessive computation; selecting an appropriate time step to accurately simulate the propagation process of sound waves over time, etc., to ensure the accuracy and reliability of the simulation results; this step is like setting rules and rhythm for the "performance" in the virtual gearbox, so that the sound wave propagation simulation can proceed in an orderly manner.

[0233] 4) Simulation calculation and result analysis:

[0234] Finite element analysis software is used to simulate the propagation process of sound waves inside the gearbox. During the simulation, the propagation position, reflection and refraction paths of the sound waves at different times are calculated, as well as the detailed characteristics such as sound pressure distribution and phase change at different positions.

[0235] Furthermore, by simulating acoustic emission sources at different locations multiple times, comprehensive acoustic wave propagation data is obtained. This data will provide detailed propagation model data support for subsequent positioning algorithms. For example, acoustic wave propagation data includes information such as the time, amplitude, and phase of the acoustic wave as it travels from an acoustic emission source at a specific location, through a series of reflections and refractions, to reach various sensors. In-depth analysis of the simulation results is like repeatedly studying a "performance video" to extract useful information, laying the foundation for accurately determining the location of the acoustic emission source.

[0236] II. Optimization of the localization algorithm based on simulation results:

[0237] After obtaining the sound wave propagation data simulated by the sound wave propagation model, it is applied to the positioning algorithm. By accurately simulating the sound wave propagation process and optimizing the positioning algorithm based on the simulation results, multipath interference is effectively suppressed and the positioning accuracy of the sound emission source is improved.

[0238] Step 2: Accurate separation of multipath signals is a key challenge when processing acoustic emission signals from wind turbine gearboxes. Physically constrained deep learning methods offer an effective solution. The basic idea is to integrate the physical structure information of the gearbox into the deep learning model to improve the accuracy of signal path separation. The specific implementation steps are as follows:

[0239] I. Establishment of 3D Physical Model and Data Generation:

[0240] 1) Finite element simulation modeling:

[0241] First, an accurate finite element model is constructed based on the 3D model of the gearbox using finite element analysis software (such as ANSYS or ABAQUS). During the modeling process, reasonable meshing of the model is crucial. If the mesh is too coarse, it may not be able to accurately capture the details of sound wave propagation, while if the mesh is too fine, it will lead to a sharp increase in the amount of computation. Therefore, it is necessary to select an appropriate mesh density according to the structural characteristics of the gearbox and the computational resources.

[0242] Simultaneously, based on actual material parameters, each component in the model is assigned corresponding physical properties such as elastic modulus and density. This step is to accurately simulate the propagation characteristics of sound waves in different materials, as different materials significantly affect the propagation speed and attenuation of sound waves. For example, metal components and plastic components have very different effects on sound wave propagation; accurately assigning physical properties makes the simulation more closely resemble reality.

[0243] 2) Virtual training data generation:

[0244] In the established finite element model, multiple virtual acoustic emission source locations are set to simulate acoustic emission events at different locations. Through finite element simulation, the signal characteristics of sound waves propagating inside the gearbox to each sensor location are calculated, including information such as sound pressure, phase, and propagation time. This information will form the basis of the virtual training data.

[0245] To make the virtual training data more representative, it is necessary to consider a variety of different operating conditions and fault scenarios; for example, to simulate acoustic emission signals under normal operating conditions, as well as acoustic emission signals under fault conditions such as gear wear and bearing failure; after collecting the above data, a virtual training dataset containing the characteristics of sound wave propagation under different paths is generated.

[0246] 3) Preprocess the generated data, such as normalization, to ensure data consistency and usability. Normalization can make data with different features have the same scale, which helps to improve the training efficiency and accuracy of deep learning models. For example, for sound pressure data and propagation time data, normalization maps them to the same numerical range.

[0247] II. Construction and Training of Graph Convolutional Network (GCN) Models:

[0248] 1) Model building:

[0249] Based on deep learning frameworks (such as TensorFlow or PyTorch), a graph convolutional network model is constructed. During the construction process, the nodes and edges of the graph are first defined. Nodes can represent sensor locations or specific locations inside the gearbox, and edges represent sound wave propagation paths or connection relationships between nodes. The above definition of the graph structure can effectively integrate the physical structure information of the gearbox into the model.

[0250] By setting appropriate graph convolutional layers, pooling layers, and fully connected layers, a network model that can effectively process graph-structured data is constructed. Graph convolutional layers are used to extract features from graph-structured data, pooling layers are used to reduce data dimensionality, and fully connected layers are used to integrate and classify the extracted features. By reasonably designing the parameters and connection methods of the above layers, the model can learn the features of multipath signals.

[0251] 2) Model training:

[0252] The generated virtual training dataset is input into the GCN model for training. Before training, appropriate training parameters need to be set, such as learning rate, number of iterations, and loss function. The learning rate determines the step size of the model when updating parameters each time. An excessively large learning rate may cause the model to fail to converge, while an excessively small learning rate will make the training process extremely slow. The number of iterations determines how many times the model learns from the training data and needs to be adjusted according to the actual situation to avoid overfitting or underfitting.

[0253] The following loss function is used to measure the difference between the model's predictions and the true labels:

[0254]

[0255] Where P is the number of samples in the virtual training dataset. Assuming that in an actual wind turbine gearbox monitoring scenario, P = 1000 samples are collected for training by simulating different fault scenarios and normal operating conditions multiple times; y p For the true label of sample p, y p ∈{0, 1}, where 0 represents the noise signal path and 1 represents the real acoustic emission signal path, y p It is determined based on manual annotation of acoustic emission signals or known fault simulation conditions; W is the sigmoid function, which maps the model output to the interval [0, 1]. K is the number of graph convolutional layers in the GCN model. After multiple experiments comparing the impact of different numbers of layers on model performance, it was determined that K=5 layers can achieve a good balance between computational complexity and model performance. k W is the weight matrix of the k-th graph convolutional layer; k (m,n) represents matrix W k The element in the m-th row and n-th column, W represents the weight matrix of the k-th layer. k The mean of the elements (m, n) during training is used for weight regularization; ReLU(x) = max(0, x) is the modified linear unit function, introducing non-linear characteristics; A pq X represents the connection weight between node p and node q in the adjacency matrix of a graph structure, reflecting the correlation between nodes; q Let be the feature vector of node q, containing features related to the sound wave propagation path extracted from the virtual training data. and λ1, λ2, λ3, and λ4 are the feature vectors of sample p obtained based on simulation and actual measurement, respectively; λ1, λ2, λ3, and λ4 are the coefficients that weigh different terms. After multiple cross-validation experiments, λ1 = 0.01, λ2 = 0.1, λ3 = 0.05, and λ4 = 0.2 were determined. Kullback-Leibler divergence is used to measure the simulated eigenvectors. With actual measured feature vector The difference between them is I, which is the dimension of the feature vector involved in its calculation. Assume I = 10. The correlation coefficient measures the simulated feature vector. With actual measured feature vector linear correlation, and They are respectively and The mean; and Let J represent the energy of the simulated and actual measured eigenvectors, respectively. Here, J represents the eigendimensional dimension used to calculate the energy. Assume J = 5, and ∈ is a very small positive number, such as ∈ = 1e-6, to avoid the case where the denominator is zero.

[0256] During training, the model continuously adjusts the weight matrix W by minimizing this loss function. k The parameters enable the model to learn signal features from different paths better and better; the training process is monitored using a validation set to prevent overfitting; if the value of the loss function starts to rise on the validation set, it indicates that the model may be overfitting, and at this time it is necessary to adjust the training parameters or use regularization and other methods to optimize the model performance.

[0257] III. Multipath Signal Separation and Processing:

[0258] 1) Real-time signal input: In practical applications, the acoustic emission signal acquired in real time is first preprocessed. Since the actual acquired signal often contains various noises, it is necessary to remove obvious noise interference through filtering, denoising and other operations to make the signal input into the model purer; for example, a bandpass filter can be used to filter out noise in other frequency bands according to the frequency range of the acoustic emission signal.

[0259] 2) Signal path separation: The preprocessed signal is input into the trained GCN model. Based on the characteristics of reflected / refracted signals learned during training, the GCN model separates the input multipath signal. The model outputs the signal components of different paths. By analyzing these signal components, the possible propagation paths of the acoustic emission source can be determined. Combined with the information provided by the previously established acoustic wave propagation model, the judgment of the signal path can be further optimized. For example, the acoustic wave propagation model can provide prior information such as the propagation time and amplitude of the sound wave at different locations, helping the GCN model to more accurately identify the real acoustic emission signal path, thereby improving the accuracy of multipath signal separation.

[0260] The physical constraint-based deep learning signal path separation method can effectively utilize the physical structure information of the gearbox, improve the accuracy of multipath signal separation, and provide a more reliable signal foundation for subsequent fault diagnosis.

[0261] Step 3: Fixed threshold fault detection methods are difficult to adapt to the complex and ever-changing operating conditions of wind turbines, easily leading to false alarms or missed alarms. The dynamic threshold adjustment mechanism based on reinforcement learning aims to improve the accuracy of fault detection by dynamically adjusting the fault trigger threshold according to different operating conditions through intelligent learning. Its implementation process mainly includes three key steps: reinforcement learning agent design, training, and dynamic threshold adjustment, as detailed below:

[0262] I. Design of the Deep Q-Network (DQN) Proxy Model:

[0263] State Definition: First, the state space of the DQN agent model is defined. The state contains rich information, including not only historical false alarm data but also current operating condition information. Historical false alarm data records the false alarm situation over a period of time, such as the number of false alarms in the past n time steps. This helps the agent model understand the system's past performance, thus allowing for better strategy adjustments; current operating condition information includes wind speed v. t Load L t Temperature T t Wind turbine operating parameters and multipath interference level indicators (Error) mp，t Signal path separation accuracy A sep，t Indicators related to acoustic emission signal processing are used; this information is encoded to form a state vector that the DQN surrogate model can understand. The above state definitions comprehensively reflect the system's history and current status, providing a sufficient basis for the agent model to make reasonable decisions.

[0264] Action Definition: Define the action space of the DQN agent model. An action represents an adjustment operation on the fault trigger threshold. Considering flexibility and operability in practical applications, a series of discrete threshold adjustment step sizes are set, for example... Where Δθ i For different threshold adjustment amounts, such as increasing the threshold by 5%, increasing by 10%, decreasing by 5%, decreasing by 10%, or keeping it unchanged, the proxy model dynamically adjusts the fault trigger threshold by selecting different actions to adapt to different operating conditions.

[0265] Reward function design: Setting a reasonable reward function is the key to guiding the DQN agent model to learn the optimal strategy. The design principle of the reward function is to encourage the agent to make decisions that can reduce false alarms and not miss real faults.

[0266] II. DQN Agent Training:

[0267] Experience replay: During wind turbine operation, the DQN agent continuously monitors the current status s t Choose an action a based on the current strategy. t After this action is performed, the environment will return a new state s. t+1 and the corresponding reward value r t s, the state at each step t Action a t Rewards r t and new state s t+1The data is stored in the experience replay buffer. The purpose of the experience replay buffer is to break the correlation between data and improve the stability and efficiency of learning. When the data stored in the buffer reaches a certain amount, a batch of data is randomly sampled from the buffer for training. This random sampling method can prevent the agent from over-relying on a certain part of the data during the learning process, thereby better generalizing to various working conditions.

[0268] Network Update: Using sampled data, calculate the error between the target Q-value and the current Q-value. The target Q-value represents the estimated long-term cumulative reward obtained by taking the optimal action in the current state, while the current Q-value is the estimated reward obtained by taking the action under the current policy. Through backpropagation, adjust the neural network parameters of the DQN agent so that the current Q-value gradually approaches the target Q-value.

[0269] By repeatedly replaying experiences and updating the network, the DQN agent model gradually learns the optimal threshold adjustment strategy as training progresses. It can dynamically adjust the fault trigger threshold according to different working conditions to maximize long-term cumulative rewards.

[0270] III. Dynamic Threshold Adjustment:

[0271] Real-time operating condition monitoring: The system monitors the operating conditions of the wind turbine in real time, including changes in parameters such as wind speed, load, and temperature, as well as indicators related to acoustic emission signal processing, such as the degree of multipath interference and the accuracy of signal path separation. This information is fed back to the DQN proxy model in real time as a basis for the proxy to adjust the threshold. For example, when the wind speed suddenly increases, it may introduce more environmental noise, affecting the detection of acoustic emission signals. At this time, the proxy needs to make corresponding threshold adjustment decisions based on this change in operating conditions.

[0272] Threshold Adjustment Decision: The DQN agent model selects an action based on the current state (including historical false alarm data and real-time operating condition information) and its learned strategy, namely, whether to adjust the fault trigger threshold and the adjustment range. For example, during the wind turbine startup phase, due to transient noise, the acoustic emission signal characteristics differ from those during normal operation, and multipath interference may be severe, potentially resulting in lower signal path separation accuracy. In this case, the DQN agent model automatically increases the threshold according to the learned strategy based on the current state information to avoid false alarms caused by transient noise. As the wind turbine's operating conditions change, the agent continuously learns and adjusts the threshold in real time, ensuring the system always maintains optimal fault detection performance, effectively reducing the false alarm rate and improving the real-time performance and accuracy of fault detection.

[0273] Through the above implementation of the dynamic threshold adjustment mechanism based on reinforcement learning, the fault detection system can intelligently adjust the fault trigger threshold according to the actual operating conditions of the wind turbine, adapt to the complex and ever-changing working environment, and improve the reliability and stability of the entire wind turbine gearbox rupture early warning system.

[0274] Example 2:

[0275] In practical applications, the multipath interference suppression formula provides a more accurate signal propagation model and initial data for signal path separation based on physical constraints in deep learning. Accurate signal path separation, in turn, provides reliable signal features and state information for dynamic threshold adjustment based on reinforcement learning. Specific applications in the early warning process of wind turbine gearbox rupture include:

[0276] I. Applications of multipath interference suppression formulas in the early warning process of wind turbine gearbox rupture include:

[0277] 1) Application under normal operating conditions: When the wind turbine gearbox is running normally, the acoustic emission signal is relatively stable, and although multipath interference exists, it is mild. At this time, the multipath interference suppression formula is as follows:

[0278]

[0279] The steps of multipath interference suppression include:

[0280] Determine the initial parameters: First, determine the number of sensors N, the number of amplitude sampling points M, and the number of phase sampling points L based on the structure of the wind turbine gearbox and the sensor layout. Simultaneously, determine the standard deviation of the time measurements of each sensor based on historical data or previous tests. Standard deviation of amplitude measurement Standard deviation of phase measurement For the balance coefficients λ and μ, the default empirical values ​​can be used first, and then fine-tuned according to the actual situation in subsequent runs.

[0281] Real-time monitoring and calculation: Real-time acquisition of the actual time it takes for sound waves to reach each sensor. Actual amplitude and actual phase The propagation of sound waves is simulated in real time using an accurate propagation model to obtain the simulation time. Simulated amplitude and simulated phase Substituting these values ​​into the formula, and by continuously adjusting the acoustic emission source position variable vector r, we find the value that minimizes the objective function. This is the estimated location of the acoustic emission source. This location information can be used as a reference for the characteristics of the acoustic emission signal during normal operation and for comparison in subsequent abnormal situations.

[0282] 2) Application under minor fault conditions: When a minor fault occurs in the wind turbine gearbox, such as slight gear wear or loose parts, the acoustic emission signal will change, and multipath interference may increase. The steps for suppressing multipath interference include:

[0283] Parameter adjustment: Because faults can cause changes in signal characteristics, the standard deviation of time, amplitude, and phase measurements may vary. It may be necessary to reassess and adjust; at the same time, adjust the balance coefficients λ and μ appropriately according to the type of fault and the emphasis on its impact on the signal; for example, if the fault mainly affects the signal amplitude, then increase λ appropriately.

[0284] Fault location and analysis: Following the real-time monitoring and calculation steps under normal operating conditions, the adjusted parameters are substituted into the calculation to obtain the estimated location of the acoustic emission source. By comparing this location with the reference location under normal operating conditions, if the deviation exceeds a certain threshold (this threshold needs to be determined based on the specific structure of the wind turbine gearbox and historical data), the location of the fault can be preliminarily determined. Further analysis of the possible causes and severity of the fault can be conducted by combining this with changes in the characteristics of the acoustic emission signal, such as increased amplitude and abnormal phase.

[0285] 3) Applications under severe fault conditions: In severe fault conditions, such as severely damaged gears or broken shafts, the acoustic emission signal will change significantly, and multipath interference will be more complex. The steps for multipath interference suppression include:

[0286] Enhanced model and parameter optimization: At this point, the original accurate propagation model may need to be further refined or modified to more accurately simulate complex sound wave propagation. At the same time, the adjustment range of each parameter needs to be larger, the standard deviation needs to be re-measured accurately, and the balance coefficients λ and μ need to be significantly optimized according to the degree of impact of the fault on different signal characteristics.

[0287] Fault severity assessment: Estimation of acoustic emission source location obtained through formula calculation If the position deviates greatly from the normal operating condition, and the signal characteristics (such as amplitude far exceeding the normal range and severe phase disorder) also indicate a serious fault, the severity of the fault can be quickly determined based on the above information and the pre-set fault severity assessment criteria (based on historical fault data and expert experience), providing a basis for timely maintenance measures.

[0288] II. Applications of reinforcement learning-based dynamic threshold adjustment in wind turbine gearbox rupture early warning processes include:

[0289] 1) Wind turbine start-up phase: During the wind turbine start-up phase, operating conditions change significantly, acoustic emission signal characteristics are unstable, and multipath interference and noise have a significant impact. The specific steps of dynamic threshold adjustment based on reinforcement learning include:

[0290] State initialization: Determine the initial state s0 of the deep Q network agent, including acquiring operating condition information such as wind speed, load, and temperature at startup, as well as the multipath interference level index Error based on previous simulations or empirical estimates. mp，0 Signal path separation accuracy A sep，0 Noise level inside the cabin int，0 Wind turbine operating status indicators f，0 Gearbox operating status indicators gb，0 Meanwhile, initialize historical false alarm data.

[0291] Action selection and threshold adjustment: Based on the current state s0, the agent selects an action a0 according to the current strategy (such as a random strategy or a strategy based on previous experience), which adjusts the fault trigger threshold. For example, due to the high noise during the startup phase, the threshold may be appropriately increased to avoid false alarms.

[0292] Reward Calculation and Q-Value Update: After executing action a0, obtain the new state s1, and calculate the reward value r0 according to the reward function:

[0293]

[0294] Then, update the Q value using the Q-value update formula:

[0295]

[0296] As the startup process continues, the above steps are repeated, and the agent gradually learns the threshold adjustment strategy suitable for the wind turbine startup phase.

[0297] 2) Stable operation phase of the wind turbine: When the wind turbine enters the stable operation phase, the operating conditions are relatively stable, but the threshold still needs to be dynamically adjusted based on real-time monitoring information. The specific steps of dynamic threshold adjustment based on reinforcement learning include:

[0298] Real-time status updates: Real-time monitoring of operating conditions such as wind speed, load, and temperature, as well as multipath interference level indicators (Error). mp，t Signal path separation accuracy A sep，t Noise level inside the cabin int，t Wind turbine operating status indicators f，t Gearbox operating status indicators gb，t Wait, update the proxy status s t At the same time, historical false alarm data is recorded.

[0299] Policy-based action selection: The agent selects actions based on the current state s. t Based on the learned strategies, select an action a.t For example, if the multipath interference level indicator Error mp，t Increase, and the signal path separation accuracy A sep，t By lowering the threshold, the agent may choose to appropriately reduce the threshold to improve the sensitivity of fault detection.

[0300] Reward Calculation and Q-Value Update: Performing Action a t Then, obtain the new state s. t+1 Calculate the reward value r t The Q-value is then updated, a process similar to the startup phase. By continuously updating the Q-value, the agent continuously optimizes the threshold adjustment strategy to adapt to subtle changes in operating conditions that may occur during the stable operation phase.

[0301] 3) Fault Occurrence Stage: When a fault occurs in the wind turbine gearbox, the acoustic emission signal characteristics, multipath interference, noise, and the operating status of the wind turbine and gearbox will all change significantly. The specific steps of dynamic threshold adjustment based on reinforcement learning include:

[0302] State Changes and Action Adjustments: A failure will cause the agent's state s to change. t Significant changes have occurred, such as in the multipath interference index Error. mp，t Significantly increased signal path separation accuracy A sep，t Decrease, State indicator for fan and gearbox operation f，t State gb，t Deviations from the normal range, etc., the agent selects action a based on the learned policy according to these state changes. t The threshold may be adjusted significantly to ensure that faults can be detected accurately.

[0303] Fault Confirmation and Reward Feedback: If a real fault is successfully detected by adjusting the threshold (i.e., (Adding a positive reward to the corresponding term in the reward function reinforces the agent's decision.) Simultaneously, based on the results of fault handling, such as the wind turbine and gearbox returning to normal operation after repair, the agent's state and Q-value are further updated, enabling the agent to learn the optimal threshold adjustment strategy during fault occurrence and handling, thus better preparing it to handle similar fault situations in the future.

[0304] 4) Re-operation Phase After Fault Repair: This is a transitional phase after the wind turbine gearbox fault is repaired and the system is put back into operation. The operating conditions gradually recover from the post-fault state to a normal, stable state. The specific steps of the reinforcement learning-based dynamic threshold adjustment include:

[0305] State Reset and Initialization: The agent's state is partially reset and reinitialized. First, historical false alarm data is appropriately cleaned up or adjusted, as the system status has changed after the fault repair. Based on the actual situation after the fault repair, the wind turbine operating status indicator (State) is reassessed and redefined. f，t Gearbox operating status indicators gb，t Simultaneously, based on the potential impact on the acoustic emission signal propagation path during fault repair, the multipath interference level index Error is re-estimated. mp,t And signal path separation accuracy A sep,t This determines the initial state s of the agent. t .

[0306] Gradual adjustment and adaptation: The agent adjusts according to the current state s t Select action a t At this point, the action may tend to adjust the threshold to a value close to the normal operating state, but it will be relatively conservative to avoid false alarms or missed alarms due to the system not being fully stable. As the running time goes by, the changes of various status indicators are monitored in real time, such as multipath interference gradually returning to normal levels and signal path separation accuracy gradually improving. Based on these changes, the agent continuously adjusts its actions so that the threshold gradually adapts to the new stable operating state.

[0307] Learning and Optimization: In this process, the reward value r is calculated using a reward function. t The Q-value is updated using the Q-value update formula. For example, if the system can accurately detect a small number of potential problems in the new operating state after adjusting the threshold (confirmed by subsequent detection methods), the reward function will give a positive reward, prompting the agent to further optimize the threshold adjustment strategy to adapt to the operating state after fault repair more quickly and prepare for future stable operation.

[0308] It is important to note the following points during the above application process:

[0309] 1) Continuous data updates and maintenance: The wind turbine operating environment and gearbox status will change over time, so it is necessary to continuously collect new data to update the model and parameters; for example, periodically re-evaluate the propagation model parameters in the multipath interference suppression formula, as well as the training data of the physical constraint-based deep learning model, to ensure that the model can adapt to new operating conditions and failure modes.

[0310] 2) Dynamic adjustment of parameters: The parameters in each formula are not fixed and need to be dynamically adjusted according to the actual operation. For example, in the dynamic threshold adjustment based on reinforcement learning, the weight coefficients may need to be adjusted according to different seasons and different operating stages (such as the initial stage of new equipment use and the long-term operation) in order to balance the impact of various factors on the reward value.

[0311] System Integration and Optimization: When integrating the algorithms involved in these three formulas into a complete monitoring system, attention must be paid to the interfaces and data interactions between the various parts. For example, how to accurately transmit the signal path separation results to the dynamic threshold adjustment module, and how the dynamic threshold adjustment feedbacks the parameter settings affecting multipath interference suppression. By continuously optimizing system integration, the performance and reliability of the entire wind turbine gearbox rupture early warning system based on acoustic emission positioning technology can be improved.

[0312] By discussing the practical application of the above methods in detail under different circumstances, comprehensive technical guidance can be provided to technical personnel in related fields, helping them to better utilize the above technologies to achieve effective early warning of wind turbine gearbox rupture.

[0313] Example 3:

[0314] The reinforcement learning-based dynamic threshold adjustment mechanism and the physical constraint-based deep learning signal path separation method logically form a complete chain of "signal preprocessing → feature optimization → decision execution". The two achieve overall optimization of wind turbine gearbox rupture early warning through data interaction and target collaboration. The specific relationship is as follows:

[0315] I. Data Interaction: Signal path separation provides key input features for threshold adjustment:

[0316] 1) Signal path separation output as a state parameter for threshold adjustment: After separating multipath signals through a graph convolutional network (GCN), the physical constraint-based deep learning method outputs the following key information: Signal path classification result: the probability of distinguishing between real acoustic emission signal paths and noise paths (output via the Sigmoid function); Separation accuracy (A sep,t ): Reflects the model's ability to separate multipath signals (e.g., the proportion of correctly classified paths); the above information is directly used as the state input (s) of the reinforcement learning agent. t (), used to evaluate the reliability of current signal processing.

[0317] For example: If A sep，t A higher percentage (e.g., >90%) indicates high signal purity, allowing for more sensitive threshold adjustment to detect early faults; if A sep，t A low threshold (e.g., <60%) indicates that multipath interference has not been effectively suppressed, and the threshold needs to be increased to avoid false alarms.

[0318] 2) The difference between simulated and measured characteristics provides environmental feedback for threshold adjustment: During signal path separation, the characteristics simulated by the physical model (X) sim，p ) and actual measured characteristics (X) meas,pDifferences in parameters such as KL divergence, correlation coefficient, and energy difference are integrated into a multipath interference index (Error). mp,t This metric, as part of the reinforcement learning state, is used to quantify the accuracy of the signal propagation model: Error mp,t The smaller the value, the more consistent the simulated and measured characteristics are, the higher the positioning accuracy, and the more the threshold adjustment can rely on signal characteristics; Error mp，t The larger the threshold, the greater the model bias, requiring threshold adjustment to compensate for the uncertainty.

[0319] II. Collaborative Goal: Jointly Optimize the Sensitivity and Reliability of Fault Detection:

[0320] 1) Signal path separation lays the physical foundation for threshold adjustment: Related threshold adjustment methods suffer from signal feature distortion due to multipath interference, making it difficult to set a fixed threshold. Physically constrained deep learning methods improve signal quality in the following ways:

[0321] Suppressing reflection / refraction interference: Using a 3D model of the gearbox and a GCN network to separate the real signal path, reducing noise contamination of features (such as amplitude and phase).

[0322] Enhanced feature distinguishability: The separated real signal path features are closer to the actual fault signal, making the "fault-noise" boundary clearer when adjusting the threshold.

[0323] For example, in gear wear faults, the separated signal may show a jump in the amplitude of a specific frequency component. In this case, the dynamic threshold can be precisely adjusted for this feature to avoid being falsely triggered by similar features in the noise.

[0324] 2) The value of dynamic threshold adjustment in amplifying signal separation: Even if signal path separation improves feature purity, the threshold still needs to be dynamically determined based on real-time operating conditions. Reinforcement learning optimizes this collaboratively through the following methods:

[0325] Adaptive operating conditions: Under operating conditions such as fan startup and sudden load changes, the signal characteristic distribution may deviate from the normal state (e.g., increased transient noise). Dynamic threshold adjustment can temporarily relax or tighten the threshold according to the current operating conditions (e.g., wind speed, load) and signal separation effect, avoiding the "one-size-fits-all" defect of fixed thresholds.

[0326] False alarm-false negative balance: Signal separation accuracy (A) sep，t ) as the positive term of the reward function (β4·A) sep，t This guides the agent to lower the threshold to increase sensitivity when the separation effect is good, and to raise the threshold to reduce false alarms when the separation effect is poor, forming a closed loop of "high-quality signal → active detection, low-quality signal → conservative decision-making".

[0327] III. Mathematical Connections: Parameter Coupling in Key Formulas

[0328] 1) Reward function integrates signal separation results: The reward function of reinforcement learning directly includes the signal path separation accuracy (A). sep，t ) and simulation-measured feature difference (Error) mp，t ):

[0329]

[0330] β3 and β4 are weighting coefficients, indicating the accuracy of signal separation (A sep，t ) and model reliability (Error) mp，t This directly affects the agent's reward feedback, driving it to prioritize threshold strategies that match the signal separation results.

[0331] 2) Q-value update depends on signal processing quality: When a reinforcement learning agent optimizes its threshold adjustment strategy using the Q-value update formula, the state transition (S... t →S t+1 The core basis is the comparison result between signal characteristics and threshold.

[0332] For example: if, after signal path separation, a certain feature (such as amplitude) is determined to be a "fault signal" by a threshold, and it is actually a real fault ( If the threshold is increased, the corresponding Q value will receive a positive update, strengthening the priority of the threshold adjustment action; if the separated signal characteristics are misjudged as faults (due to residual interference), the number of false alarms will be used to determine the fault. The negative reward inhibits the action, forcing the agent to adjust the threshold.

[0333] IV. Summary: The closed loop from "signal purification" to "intelligent decision-making" is shown in Table 1 below:

[0334] Table 1:

[0335] Link Signal path separation based on physical constraints Dynamic threshold adjustment based on reinforcement learning Core objective Separate real signals from noise, improve feature purity Dynamically determine fault threshold according to signal characteristics and working conditions Key outputs Signal path classification, separation accuracy, feature difference index Dynamic threshold, fault warning decision Logical association Provide reliable input features for threshold adjustment Convert signal characteristics into specific warning actions Synergistic value Reduce "noise interference error" of threshold adjustment Amplify "fault feature enhancement effect" of signal separation

[0336] Both technologies constrain the signal processing process through physical models and combine data-driven optimization decision-making strategies to form a complete technical chain of "first purifying the signal, then making intelligent decisions." Together, they solve the core problems of "difficulty in suppressing interference and difficulty in adapting thresholds" in traditional acoustic emission early warning systems, ultimately achieving high precision and high reliability in wind turbine gearbox rupture early warning.

[0337] The above embodiments have the following beneficial effects:

[0338] 1) Multipath interference suppression improves positioning accuracy: By establishing an accurate propagation model, a detailed 3D model of the complex internal structure of the wind turbine gearbox is created, and accurate acoustic wave propagation parameters are assigned to each component. Finite element analysis is used to accurately simulate the acoustic wave propagation process, enabling more precise prediction of the propagation path and characteristics of acoustic emission signals under multipath effects. Based on this model, subsequent positioning algorithms can obtain a more reliable data foundation, effectively reducing positioning errors caused by multipath interference and significantly improving the positioning accuracy of the acoustic emission source. In practical applications, positioning errors can be reduced by 30%-50%, more accurately determining the location of faults inside the gearbox. This provides strong support for timely maintenance and troubleshooting; it also enhances the reliability of signal analysis: an accurate propagation model helps to better understand the propagation law of acoustic emission signals inside the gearbox. When analyzing the collected acoustic emission signals, this model can more accurately identify useful information and interference components in the signal, improving the reliability of signal analysis. This is of great significance for accurately judging the operating status of the gearbox and discovering potential faults in advance. For example, when analyzing the frequency, amplitude and other characteristics of the signal, it can eliminate false features caused by multipath interference, making the analysis results more realistically reflect the fault situation inside the gearbox.

[0339] 2) Improve the accuracy of multipath signal separation by deep learning based on physical constraints: Introduce a three-dimensional physical model of the gearbox structure and combine it with a graph convolutional network to separate multipath signals. By generating virtual training data using the CAD model of the gearbox, the GCN model learns the characteristics of reflected / refracted signals, significantly enhancing its ability to identify and separate multipath signals in complex structures. Compared to signal processing methods based on simple geometric models, it can separate multipath signals more accurately, improving the accuracy of multipath signal separation by 20%-30%. This provides cleaner and more accurate signals for subsequent acoustic emission signal processing and fault diagnosis, contributing to improved fault diagnosis accuracy. It breaks through the limitations of traditional algorithms by innovatively combining physical simulation with data-driven approaches, overcoming the dependence of traditional algorithms on simple geometric models. This combination fully utilizes the prior knowledge provided by the physical model to guide the training of the deep learning model, enabling the model to better adapt to the complex internal structure of the wind turbine gearbox. Even in extremely complex acoustic wave propagation scenarios, it can effectively process multipath signals, improving the adaptability and robustness of the entire acoustic emission signal processing system. For example, when there are minor changes in the internal structure of the gearbox or the presence of irregular components, traditional algorithms may fail, while this method can still accurately separate multipath signals, ensuring the normal operation of the fault early warning system.

[0340] 3) Reinforcement Learning-Based Dynamic Threshold Adjustment Mechanism to Reduce False Alarm Rate: A reinforcement learning agent based on a deep Q-network is designed to dynamically adjust the fault trigger threshold based on historical false alarm data and current operating conditions. During wind turbine operation, the acoustic emission signal characteristics vary significantly under different operating conditions. Traditional fixed thresholds are prone to false alarms. The dynamic threshold adjustment mechanism can automatically adapt to changes in operating conditions. In situations prone to transient noise, such as wind turbine startup, the threshold is automatically increased to avoid false alarms. Practical applications show that this mechanism can reduce the false alarm rate by 40%-50%, greatly reducing the waste of manpower and resources caused by false alarms and improving the reliability of the early warning system; it also improves fault detection. Real-time performance and accuracy: The reinforcement learning agent continuously learns and adjusts thresholds in real time, ensuring the system always maintains optimal fault detection performance. As operating conditions change, the agent can make timely threshold adjustment decisions, ensuring accurate detection of fault signals under various circumstances. This not only improves the real-time performance of fault detection, enabling early warnings at the first sign of a fault, but also guarantees detection accuracy, preventing the omission of real faults due to improper threshold settings. For example, when the fan load suddenly changes, the agent can quickly adjust the threshold according to the current operating conditions, accurately detecting fault signals that may be caused by load changes, providing a more reliable guarantee for the safe operation of the equipment.

[0341] This embodiment also provides a wind turbine gearbox rupture early warning device based on acoustic emission positioning technology. This device is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0342] This embodiment provides a wind turbine gearbox rupture early warning device based on acoustic emission positioning technology, such as... Figure 5 As shown, it includes:

[0343] Module 501 is used to obtain the internal structural parameters of the wind turbine gearbox and to construct a sound wave propagation model based on the internal structural parameters of the wind turbine gearbox.

[0344] Simulation module 502 is used to simulate sound wave propagation at multiple sound emission source locations using a sound wave propagation model to obtain sound wave propagation simulation data;

[0345] The separation module 503 is used to acquire real-time acoustic emission signals and use a graph convolutional network model to perform path separation on the real-time acoustic emission signals to obtain classification probability data of multipath signal paths.

[0346] The detection module 504 is used to detect fault signals based on acoustic wave propagation simulation data and classification probability data of multipath signal paths, and to obtain early warning information of wind turbine gearbox rupture by using a reinforcement learning agent model.

[0347] The further functional descriptions of each of the above modules are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0348] In this embodiment, a wind turbine gearbox rupture early warning device based on acoustic emission positioning technology is presented in the form of a functional unit. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0349] This invention also provides a computer device having the above-described features. Figure 5 The image shows a wind turbine gearbox rupture early warning device based on acoustic emission positioning technology.

[0350] Please see Figure 6 , Figure 6 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of the present invention, such as... Figure 6 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 6 Take a processor 10 as an example.

[0351] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0352] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0353] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0354] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0355] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0356] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0357] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0358] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and all such modifications and variations fall within the scope defined by the appended claims.

Claims

1. A method for early warning of wind turbine gearbox rupture based on acoustic emission localization technology, characterized in that, The method includes: Obtain the internal structural parameters of the wind turbine gearbox, and construct a sound wave propagation model based on the internal structural parameters of the wind turbine gearbox; The sound wave propagation model was used to simulate sound wave propagation at multiple sound emission source locations to obtain sound wave propagation simulation data. Acquire real-time acoustic emission signals, and use a graph convolutional network model to perform path separation on the real-time acoustic emission signals to obtain classification probability data of multipath signal paths; Based on the acoustic wave propagation simulation data and the classification probability data of the multipath signal path, a reinforcement learning surrogate model is used to detect fault signals and obtain early warning information of wind turbine gearbox rupture. The classification probability data based on the acoustic wave propagation simulation data and the multipath signal path is used to perform fault signal detection using a reinforcement learning surrogate model to obtain wind turbine gearbox rupture early warning information, including: The sound wave propagation simulation data is used for positioning optimization to determine the multipath interference level index; The signal path separation accuracy is determined based on the classification probability data of the multipath signal paths. The reinforcement learning agent model is used to determine threshold adjustment strategies under various wind turbine operating conditions; Obtain the current wind turbine operating parameters, and based on the current wind turbine operating parameters, the multipath interference level index, and the signal path separation accuracy, determine the fault trigger threshold using the threshold adjustment strategy under various wind turbine operating conditions.

2. The method according to claim 1, characterized in that, The process of obtaining the internal structural parameters of the wind turbine gearbox and constructing a sound wave propagation model based on these parameters includes: Obtain the three-dimensional parameters of each component inside the wind turbine gearbox, and construct a three-dimensional physical model of the wind turbine gearbox based on the three-dimensional parameters of each component inside the wind turbine gearbox; Acoustic parameters of each component inside the wind turbine gearbox are obtained, and the three-dimensional physical model of the wind turbine gearbox is assigned the acoustic parameters of each component inside the wind turbine gearbox. Finite element analysis was performed on the three-dimensional physical model of the wind turbine gearbox after assigning acoustic parameters to each component inside the gearbox, and the sound wave propagation model was obtained.

3. The method according to claim 2, characterized in that, The method of using a graph convolutional network model to perform path separation on the real-time acoustic emission signal to obtain classification probability data of multipath signal paths includes: The real-time acoustic emission signal is preprocessed, and the preprocessed real-time acoustic emission signal is used to extract features to obtain signal propagation features; Based on the preprocessed acoustic emission real-time signal, the classification probability data of the multipath signal path is determined using the graph convolutional network model.

4. The method according to claim 3, characterized in that, The method for detecting fault signals using a reinforcement learning surrogate model based on the acoustic wave propagation simulation data and the classification probability data of the multipath signal paths to obtain wind turbine gearbox rupture early warning information also includes: The signal propagation characteristics are compared with the fault triggering threshold. If the signal propagation characteristics exceed the fault triggering threshold, a wind turbine gearbox rupture warning message is generated.

5. The method according to claim 1, characterized in that, The step of optimizing the location of the simulated sound wave propagation data and determining the multipath interference level index includes: Based on the sound wave propagation simulation data, the location vector of the sound emission source is determined using a positioning algorithm; Based on the acoustic emission source position vector, the propagation parameters are simulated to obtain the basic simulation parameters for sound wave propagation inside the gearbox. The actual measurement parameters of sound wave propagation inside the gearbox are obtained, and the simulated basic parameters of sound wave propagation inside the gearbox are compared with the actual measurement parameters of sound wave propagation inside the gearbox to obtain the multipath interference level index.

6. The method according to claim 1, characterized in that, The step of using the reinforcement learning agent model to determine threshold adjustment strategies under various wind turbine operating conditions includes: Define a state, an action, and a reward function; wherein, the state is historical false alarm data corresponding to various wind turbine operating conditions, the wind turbine operating parameters, the multipath interference level index, and the signal path separation accuracy; the action is the adjustment operation of the fault trigger threshold; Obtain the current state, select the current action based on the current state, adjust the fault trigger threshold based on the current action, and update the current state to obtain the updated state; Calculate the current reward value based on the updated state; Calculate the current Q value based on the current state, the current action, the updated state, and the current reward value; The current Q value is compared with the target Q value, and the threshold adjustment strategy is updated iteratively based on the comparison result until the current Q value matches the target Q value, thus obtaining the threshold adjustment strategy under the various wind turbine operating conditions.

7. A wind turbine gearbox rupture early warning device based on acoustic emission positioning technology, characterized in that, The device includes: The module is used to obtain the internal structural parameters of the wind turbine gearbox and to construct a sound wave propagation model based on the internal structural parameters of the wind turbine gearbox. The simulation module is used to simulate sound wave propagation at multiple sound emission source locations using the sound wave propagation model, and obtain sound wave propagation simulation data. The separation module is used to acquire real-time acoustic emission signals and perform path separation on the real-time acoustic emission signals using a graph convolutional network model to obtain classification probability data of multipath signal paths. The detection module is used to detect fault signals based on the acoustic wave propagation simulation data and the classification probability data of the multipath signal path, using a reinforcement learning agent model, and to obtain early warning information of wind turbine gearbox rupture. The detection module is specifically used to optimize the location of the acoustic wave propagation simulation data and determine the multipath interference level index; determine the signal path separation accuracy based on the classification probability data of the multipath signal path; determine the threshold adjustment strategy under various wind turbine operating conditions using the reinforcement learning surrogate model; obtain the current wind turbine operating parameters, and determine the fault trigger threshold based on the current wind turbine operating parameters, the multipath interference level index, and the signal path separation accuracy using the threshold adjustment strategy under various wind turbine operating conditions.

8. A computer device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the wind turbine gearbox rupture early warning method based on acoustic emission positioning technology as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the wind turbine gearbox rupture early warning method based on acoustic emission positioning technology as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, The method includes computer instructions for causing a computer to execute the wind turbine gearbox rupture early warning method based on acoustic emission positioning technology as described in any one of claims 1 to 6.

Images