Wind turbine health assessment and maintenance method and device based on graph attention network

Abstract

The application discloses a wind turbine health assessment and maintenance method and device based on a graph attention network, relates to the technical field of wind power generation intelligent operation and maintenance, and comprises the following steps: constructing a component relationship graph by taking each key component as a graph node and a coupling relationship between the key components as a base edge, generating a dynamic adjacency matrix according to a load transmission coefficient determined based on operation data, an abnormal propagation coefficient determined according to a degree to which vibration data deviates from a health benchmark vibration feature vector of the key components and the component relationship graph; inputting the operation data, the vibration data and the dynamic adjacency matrix into the graph attention network to output a health index; predicting the remaining service life of each key component based on a historical change sequence of the health index; and determining a maintenance action of each key component according to the health index and the remaining service life. The application can accurately assess the health status of each key component in a multi-component coupling scene, quantitatively predict the remaining service life of each key component, and output an action for guiding actual maintenance operation.

Description

Technical Field

[0001] This application relates to the field of intelligent operation and maintenance technology for wind power generation, and in particular to a method and apparatus for health assessment and maintenance of wind turbine units based on graph attention networks. Background Technology

[0002] As in-service wind turbines gradually enter their later stages of operation, frequent failures such as gearbox breakage, main shaft bearing wear, and generator insulation aging are occurring, significantly increasing the proportion of operation and maintenance costs in the cost per kilowatt-hour. The proportion of wind turbines with longer operating lives is rising year by year, leading to increasing maintenance pressure due to declining equipment reliability. Upgrading the operation and maintenance model from "reactive maintenance" and "periodic inspection" to "predictive maintenance" has become a core issue for cost reduction and efficiency improvement in the industry.

[0003] Existing technologies face the following bottlenecks: Firstly, current health assessment methods primarily rely on slowly varying signals such as temperature and power from SCADA systems. Abnormal characteristics inherent in vibration signals, such as kurtosis spikes and sideband energy increases, are not effectively utilized, resulting in output component health indices that fail to fully reflect the actual operating status of components. Secondly, while some graph neural network-based methods attempt to model relationships between components, the graph topology is pre-fixed and does not change with operating conditions. Under conditions of high winds or heavy loads, the coupling between components in the drivetrain significantly increases, and the fixed graph structure cannot accurately reflect the true degradation coupling strength. Furthermore, existing methods only output health assessment results or health indices without providing maintenance decision suggestions. Maintenance plans still require manual formulation based on experience, resulting in low levels of intelligence.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This application provides a method and apparatus for health assessment and maintenance of wind turbines based on graph attention networks, in order to solve the technical problems in the prior art where the fixed graph structure cannot respond to load changes, the output health index cannot fully reflect the actual operating status of components, and health assessment and maintenance decisions are disconnected.

[0006] In a first aspect, embodiments of this application provide a method for health assessment and maintenance of wind turbine generators based on graph attention networks, including: Obtain operating data and vibration data of key components in the wind turbine, wherein the operating data includes at least wind speed and active power; A component relationship graph is constructed using each key component as a graph node and the coupling relationship between key components as a base edge. The load transfer coefficient is determined based on the wind speed and the active power. The anomaly propagation coefficient is determined based on the degree to which the vibration data of the target key component deviates from its healthy reference vibration characteristic vector. Based on the load transfer coefficient, the anomaly propagation coefficient and the component relationship graph, a dynamic adjacency matrix representing the intensity of degradation influence between key components is generated. The operating data, the vibration data, and the dynamic adjacency matrix are input into a graph attention network. The graph attention network is used to perform weighted fusion of the operating data and the vibration data based on the dynamic adjacency matrix and output the health index of each key component. Based on the historical change sequence of the health index at multiple consecutive historical moments, the remaining service life of each key component is predicted. Based on the health index and the remaining service life, maintenance actions for each key component are determined, and these maintenance actions are used to guide the maintenance operations of the wind turbine.

[0007] In some embodiments, the key components include at least blades, main shaft bearings, gearboxes, generators, and converters; the coupling relationship between the key components includes at least physical transmission connections; the target key component is a transmitting component for abnormal propagation, used to propagate vibration abnormality information to other key components coupled with it, and the other key components are receiving components for abnormal propagation.

[0008] In some embodiments, determining the load transfer coefficient based on the wind speed and the active power includes: The load transfer factor is determined using the following formula:

[0009] Among them, F 载荷 (t) is the load transfer coefficient at time t; The load modulation intensity coefficient; The function is the Sigmoid function; v(t) is the wind speed at time t; v rated P is the rated wind speed; P(t) is the active power at time t; P rated This refers to the rated active power.

[0010] In some embodiments, the vibration data includes at least a vibration feature vector; correspondingly, the degree to which the vibration data based on the target critical component deviates from its healthy baseline vibration feature vector determines the anomaly propagation coefficient, including: The anomaly propagation coefficient is determined using the following formula:

[0011] in, Let be the intensity of the vibration anomaly information propagated from the transmitting component q to the receiving component p at time t, and let be the anomaly propagation coefficient at time t; exp is an exponential function with the natural constant as its base; K q (t) is the vibration characteristic vector of the transmitting component q at time t; The health reference vibration characteristic vector of the transmitting component q; For bandwidth parameters; is the vector norm between the vibration characteristic vector and the health reference vibration characteristic vector, used to characterize the degree to which the vibration characteristic vector of the transmitting component q deviates from its health reference vibration characteristic vector.

[0012] In some embodiments, generating a dynamic adjacency matrix representing the intensity of degradation impact among critical components based on the load transfer coefficient, the anomaly propagation coefficient, and the component relationship diagram includes: A load transfer matrix is ​​constructed based on the load transfer coefficient, and a vibration anomaly propagation matrix is ​​constructed based on the anomaly propagation coefficient. Based on the weights of each base edge in the component relationship diagram, construct the physical connection base matrix; The dynamic adjacency matrix is ​​generated by performing element-wise multiplication on the physical connection basis matrix, the load transfer matrix, and the anomaly propagation matrix.

[0013] In some embodiments, the step of performing element-wise multiplication of the physical connection basis matrix, the load transfer matrix, and the anomaly propagation matrix to generate the dynamic adjacency matrix includes: The dynamic adjacency matrix is ​​generated according to the following formula:

[0014] Where W(t) is the dynamic adjacency matrix at time t; W base For the physical connection basis matrix; For element-wise multiplication; f 载荷 (t) is the load transfer matrix at time t; W vib (t) is the vibration anomaly propagation matrix at time t.

[0015] In some embodiments, the graph attention network is used to perform weighted fusion of the operational data and the vibration data based on the dynamic adjacency matrix, and output a health index for each key component, including: The operational and vibration data of each key component are stitched together and linearly projected to obtain the initial nodal features of each key component. For each base edge in the component relationship graph, the two key components corresponding to the base edge are each other's neighbor key components. Based on the initial node features of each key component, the initial node features of its neighbor key components, and the weight of the corresponding base edge in the dynamic adjacency matrix, the attention score of each key component to its neighbor key components is calculated. The attention scores are normalized to obtain the normalized attention weights of each key component to its neighboring key components. According to the normalized attention weights, the initial node features of the neighboring key components of each key component are weighted and aggregated to obtain aggregated features. The aggregated features are then residually connected with the transformed features of each key component after residual transformation to obtain the fused features of each key component. The graph attention network contains at least two graph attention layers. The fused features of each key component are used as the input of the next graph attention layer. The operations of attention score calculation, normalization processing, weighted aggregation and residual connection are repeatedly performed until the graph attention calculation of the preset number of layers is completed, and the fused features of each key component in the last layer are obtained. The fusion features of each key component in the last layer are mapped to the range of health index values ​​through a fully connected layer and an activation function, and the health index of each key component is output.

[0016] In some embodiments, calculating the attention score of each critical component to its neighboring critical components includes: For each key component and its neighboring key components, the initial node features of the key component and the initial node features of its neighboring key components are linearly transformed to obtain the query features of the key component and the key features of its neighboring key components. The query features, key features, and physical bias components are added together, and after passing through a nonlinear activation function, an inner product operation is performed with the attention vector to obtain the attention score of the key component to its neighboring key components; wherein, the physical bias component is determined according to the weight of the connection relationship between the key component and its neighboring key components in the dynamic adjacency matrix.

[0017] In some embodiments, the step of adding the query features, the key features, and the physical bias components, applying the result to a nonlinear activation function, and then performing an inner product operation with the attention vector to obtain the attention score of the key component to its neighboring key components includes: Calculate the attention score using the following formula:

[0018] in, e ij Let i be the attention score of critical component i to its neighboring critical component j; W is the transpose of the attention vector; LeakyReLU is the non-linear activation function; query To query the transformation matrix, used for the initial node features of key component i. Perform a linear transformation to obtain the query features of key component i; W key Here is the key transformation matrix, used to initialize the node features of neighboring key component j. Perform a linear transformation to obtain the key features of the neighboring key component j; For query features; Key features; w ij (t) represents the weight of the base edge between key component i and its neighbor key component j in the dynamic adjacency matrix at time t; This is the physical bias component.

[0019] In some embodiments, the step of weighted aggregation of the initial node features of the neighboring key components of each key component according to the normalized attention weight to obtain aggregated features, and residual connection of the aggregated features with the transformed features of each key component after residual transformation to obtain the fused features of each key component, includes: Weighted aggregation and residual join are performed according to the following formula:

[0020] in, The fused features of key component i after processing by a graph attention network layer; The initial node features of key component i; The initial node features of the neighboring key component j; W represents the normalized attention weight of critical component i to its neighboring critical component j. res W is the residual transformation matrix; v Value transformation matrix; For transformation features; is the aggregation feature; N(i) is the set of neighboring key components of key component i; ELU is the activation function.

[0021] In some embodiments, mapping the fusion features of each key component at the last layer to the health index value range via a fully connected layer and an activation function, and outputting the health index of each key component, includes: Output the health index of each key component according to the following formula:

[0022] Among them, HI i (t) represents the health index of key component i at time t, and the health index ranges from [0,1]. The fusion features of key component i in the output of the Lth layer of the graph attention network; W HI b is the weight matrix of the fully connected layer; HI is the bias term for the fully connected layer; Sigmoid is the activation function.

[0023] In some embodiments, the method further includes: Obtain the current kurtosis value of each key component, which is extracted from the vibration data; When the current kurtosis value is greater than the healthy baseline kurtosis value, the vibration correction amount is determined based on the deviation between the current kurtosis value and the healthy baseline kurtosis value; The vibration correction amount is subtracted from the health index to obtain the corrected health index.

[0024] In some embodiments, determining the vibration correction amount based on the deviation between the current kurtosis value and the health baseline kurtosis value includes: The vibration correction amount is determined using the following formula:

[0025] in, This represents the vibration correction amount for the critical component i at time t. This is the maximum correction amplitude control parameter, used to limit the upper limit of the vibration correction amount; This is the kurtosis scaling factor, used to control how quickly the kurtosis reaches saturation; Let i be the current kurtosis value of the critical component i at time t; denoted as the health baseline kurtosis value of critical component i; tanh is the hyperbolic tangent function; max is the function for maximizing the value.

[0026] In some embodiments, the method further includes: The fusion features of each key component at multiple consecutive historical moments are obtained to form a historical sequence of fusion features; The operating conditions of each key component at multiple consecutive historical moments are obtained to form a historical operating condition sequence; The vibration feature vectors of each key component at multiple consecutive historical moments are obtained to form a historical sequence of vibration feature vectors; Accordingly, the prediction of the remaining service life of each key component includes: Based on the historical change sequence of the health index, the historical sequence of fused features, the historical operating condition sequence, and the historical sequence of vibration feature vectors, the remaining service life of each key component is predicted.

[0027] In some embodiments, predicting the remaining service life of each key component based on the historical change sequence of the health index, the historical sequence of fused features, the historical operating condition sequence, and the historical sequence of vibration feature vectors includes: The historical change sequence of the health index, the historical sequence of fused features, the historical operating condition sequence, and the historical sequence of vibration feature vectors are input into the gated loop unit encoder, and the input information is compressed into a context vector by the gated loop unit encoder. Using the context vector as the initial state, the gated loop unit decoder combines the predicted working condition information to generate the average predicted health index for each moment within a future preset time window. The predicted health index average is compared with a preset failure threshold at each time step, and the time when the predicted health index average first falls below the preset failure threshold is determined as the remaining service life.

[0028] In some embodiments, the step of inputting the historical change sequence of the health index, the historical sequence of fused features, the historical operating condition sequence, and the historical sequence of vibration feature vectors into a gated cyclic encoder, and compressing the input information into a context vector by the gated cyclic encoder, includes: The input information is compressed into a context vector using a gated cyclic encoder according to the following formula:

[0029] Among them, c i The context vector of key component i; GRU enc Encoder for gated loop unit; The historical change sequence of the health index of key component i from time tT to time t; The fusion feature history sequence of key component i from time tT to time t; This is the historical operating condition sequence of key component i from time tT to time t. This represents the historical sequence of vibration characteristic vectors of the key component i from time tT to time t.

[0030] In some embodiments, the step of generating the average predicted health index at each time point within a future preset time window by using the gated recurrent unit decoder with the context vector as the initial state and combining it with the predicted working condition information includes: The predicted mean health index is generated using the gated recurrent unit decoder according to the following formula:

[0031] in, Let i be the mean of the predicted health index of the key component i at a future time t+τ; Let GRU be the standard deviation of the predicted health index of critical component i at future time t+τ. dec For gated loop unit decoders; c i is the context vector of the key component i; This refers to the predicted operating conditions at a future time t+τ.

[0032] In some embodiments, comparing the predicted health index mean with a preset failure threshold time-by-time, and determining the moment when the predicted health index mean first falls below the preset failure threshold as the remaining service life includes: The remaining useful life is determined using the following formula:

[0033] Among them, RUL i The remaining service life of critical component i; Let i be the mean of the predicted health index of the key component i at a future time t+τ; The preset failure threshold for critical component i; T represents the time step forward from the current moment; max is the maximum number of prediction steps; min is the function to take the minimum value.

[0034] In some embodiments, the method further includes: Obtain real-time operating conditions, vibration alarm status, maintenance history information, spare parts inventory information, and weather window information; Accordingly, determining the maintenance actions for each key component based on its health index and remaining service life includes: The health index, remaining service life, real-time operating conditions, vibration alarm status, maintenance history, spare parts inventory, and weather window information of each key component are input as state vectors into the reinforcement learning network, which outputs maintenance actions. The maintenance actions include at least continued operation, planned maintenance, and emergency maintenance. The reinforcement learning network is optimized based on a reward function during training, which is determined according to a reliability penalty term, an operation penalty term, an over-maintenance penalty term, and a vibration alarm benefit term.

[0035] In some embodiments, the reward function is determined based on a reliability penalty term, an operational penalty term, an over-maintenance penalty term, and a vibration alarm benefit term, including: The reward function is determined using the following formula:

[0036] Wherein, R(s) t ,a t ) represents the state vector s t Perform maintenance action at Total reward; R rel (s t ,a t ) represents the state vector s t Perform maintenance action a t Reliability penalty item; R run (s t ,a t ) represents the state vector s t Perform maintenance action a t Runtime penalty items; R pen (s t ,a t ) represents the state vector s t Perform maintenance action a t Over-maintenance penalty; R vib (s t ,a t ) represents the state vector s t Perform maintenance action a t Vibration alarm revenue items; The reliability penalty term is determined according to the following formula:

[0037] in, The failure severity weight for critical component i; HI i (t) represents the health index of key component i at time t; The preset failure threshold for critical component i; For indicator functions, This indicates that the health index is set to 1 when it is below the preset failure threshold, and 0 otherwise; K represents the total number of critical components. The runtime penalty item is determined according to the following formula:

[0038] in, To maintain action a t Corresponding maintenance costs; To maintain action a t Corresponding losses due to power outages; The over-maintenance penalty is determined according to the following formula:

[0039] in, Penalty coefficient for over-maintenance; RUL represents the average remaining service life of critical component i. thr To preset the maintenance plan's lifespan threshold; Indicates when maintenance action at The value is 1 for emergency maintenance and 0 otherwise; K is the total number of critical components. The vibration alarm revenue item is determined according to the following formula: Among them, w vib K represents the vibration alarm benefit coefficient. i (t) represents the kurtosis value of the critical component i; K thr E is the preset kurtosis threshold. ratio,i (t) represents the high-frequency energy ratio; E thr The preset high-frequency energy ratio threshold is used; Indicates when maintenance action a t The value is set to 1 when the kurtosis value of the critical component is greater than the preset kurtosis threshold or the high-frequency energy ratio is greater than the preset high-frequency energy ratio threshold for planned maintenance; otherwise, the value is set to 0.

[0040] In some embodiments, the method further includes: Real-time monitoring of vibration characteristic vectors of key components; When the vibration characteristic vector of any critical component exceeds the preset vibration alarm threshold, the maintenance action of the critical component will be forcibly identified as planned maintenance.

[0041] In some embodiments, the method further includes: Collect equipment status information before maintenance, types of actions performed, equipment status information after maintenance, actual safe operating time of the equipment after maintenance, and vibration feature vectors before and after maintenance to construct experience samples; The experience samples are stored in the experience playback pool; At preset time intervals, experience samples are sampled from the experience replay pool for incremental fine-tuning to update the network parameters of the graph attention network and / or the network parameters of the reinforcement learning network. The updated parameters are used for subsequent health assessment, remaining service life prediction, and determination of maintenance actions.

[0042] Secondly, embodiments of this application provide a wind turbine health assessment and maintenance device based on graph attention networks, comprising: The acquisition module is used to acquire the operating data and vibration data of each key component in the wind turbine, wherein the operating data includes at least wind speed and active power; The component relationship graph construction module is used to construct a component relationship graph with each key component as a graph node and the coupling relationship between key components as a base edge. It determines the load transfer coefficient based on the wind speed and the active power, determines the anomaly propagation coefficient based on the degree to which the vibration data of the target key component deviates from its healthy reference vibration characteristic vector, and generates a dynamic adjacency matrix representing the intensity of degradation influence between key components based on the load transfer coefficient, the anomaly propagation coefficient and the component relationship graph. The health index output module is used to input the operation data, the vibration data and the dynamic adjacency matrix into the graph attention network. The graph attention network is used to perform weighted fusion of the operation data and the vibration data based on the dynamic adjacency matrix and output the health index of each key component. The remaining service life prediction module is used to predict the remaining service life of each key component based on the historical change sequence of the health index at multiple consecutive historical moments. The maintenance decision module is used to determine the maintenance actions for each key component based on the health index and the remaining service life. These maintenance actions are used to guide the maintenance operations of the wind turbine.

[0043] Thirdly, embodiments of this application also provide an electronic device, including a memory, a processor, and a computer program / instructions stored in the memory, wherein the processor executes the computing program / instructions to implement the steps of the above-described method for wind turbine health assessment and maintenance based on graph attention networks.

[0044] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the steps of the above-described method for wind turbine health assessment and maintenance based on graph attention networks.

[0045] Fifthly, embodiments of this application also provide a computer program product, wherein when the computer program / instructions are executed by a processor, they implement the steps of the above-described method for wind turbine health assessment and maintenance based on graph attention networks.

[0046] This application provides a method and apparatus for health assessment and maintenance of wind turbines based on graph attention networks. By acquiring operational and vibration data of key components in the wind turbine, including at least wind speed and active power, it achieves comprehensive perception of multi-source heterogeneous data of the wind turbine, providing a data foundation for subsequent health assessment. A component relationship graph is constructed using each key component as a graph node and the coupling relationship between key components as base edges. The load transfer coefficient is determined based on the wind speed and active power, and the anomaly propagation coefficient is determined based on the degree to which the vibration data of the target key component deviates from its health baseline vibration characteristic vector. Based on the load transfer coefficient, the anomaly propagation coefficient, and the component relationship graph, a dynamic adjacency matrix representing the intensity of degradation influence between key components is generated. This enables the graph structure to dynamically adjust with changes in wind speed, active power, and vibration data, effectively solving the problem in existing technologies where the fixed graph structure cannot respond to load changes. This allows the degradation coupling strength between components to change in real time with actual operational data and abnormal vibration states. By inputting the operational data, vibration data, and dynamic adjacency matrix into a graph attention network (GAN), the GAN performs weighted fusion of the operational and vibration data based on the dynamic adjacency matrix, outputting a health index for each key component. This achieves deep fusion of operational data (slowly changing signals) and vibration data (fastly changing signals / vibration signals), enabling the health index of each key component to comprehensively reflect its own state and the degradation effects of other coupled key components. This effectively solves the problem in existing technologies where the output health index cannot fully reflect the actual operating state of components. By predicting the remaining service life of each key component based on the historical change sequence of the health index over multiple consecutive historical moments, a quantitative prediction of the future health status of key components is achieved, providing time-dimensional information for maintenance decisions. By determining the maintenance actions for each key component based on the health index and the remaining service life, these maintenance actions guide the maintenance operations of the wind turbine, enabling the joint use of health assessment results and remaining service life prediction results in maintenance decisions. This effectively solves the problem of the separation between health assessment and maintenance decision-making in existing technologies. The above solution effectively solves the technical problems in the existing technology, such as the fixed structure of the graph cannot respond to load changes, the output health index cannot fully reflect the actual operating status of the components, and the separation between health assessment and maintenance decision-making. It can accurately assess the health status of each key component in multi-component coupled scenarios, quantitatively predict the remaining service life of each key component, and output actions to guide actual maintenance operations. Attached Figure Description

[0047] To more clearly illustrate the embodiments of this specification, the accompanying drawings used in the embodiments will be briefly introduced below. The drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart illustrating a wind turbine health assessment and maintenance method based on graph attention networks, provided in one embodiment of this specification. Figure 2 This is a schematic diagram illustrating an embodiment of a wind turbine health assessment and maintenance method based on graph attention networks provided in this specification, applied in a scenario example. Figure 3 This is a schematic diagram illustrating an embodiment of a wind turbine health assessment and maintenance method based on graph attention networks provided in this specification, applied in a scenario example. Figure 4 This is a schematic diagram illustrating an embodiment of a wind turbine health assessment and maintenance method based on graph attention networks provided in this specification, applied in a scenario example. Figure 5 This is a schematic diagram illustrating an embodiment of a wind turbine health assessment and maintenance method based on graph attention networks provided in this specification, applied in a scenario example. Figure 6 This is a schematic diagram illustrating an embodiment of a wind turbine health assessment and maintenance method based on graph attention networks provided in this specification, applied in a scenario example. Figure 7 This is a schematic diagram illustrating an embodiment of a wind turbine health assessment and maintenance method based on graph attention networks provided in this specification, applied in a scenario example. Figure 8 This is a schematic diagram illustrating an embodiment of a wind turbine health assessment and maintenance method based on graph attention networks provided in this specification, applied in a scenario example. Figure 9 This is a schematic diagram illustrating an embodiment of a wind turbine health assessment and maintenance method based on graph attention networks provided in this specification, applied in a scenario example. Figure 10 This is a schematic diagram of the structural composition of a wind turbine health assessment and maintenance device based on graph attention networks, provided in one embodiment of this specification. Figure 11 This is a schematic diagram of the structural composition of an electronic device provided in one embodiment of this specification. Detailed Implementation

[0049] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0050] Wind turbines are complex electromechanical systems composed of blades, main shaft bearings, gearboxes, generators, converters, and other components. These components are interconnected through physical transmission chains, exhibiting clear coupling relationships and fault propagation paths. As wind turbines age, faults such as gearbox breakage, main shaft bearing wear, and generator insulation aging occur frequently, significantly increasing the proportion of operation and maintenance costs in the cost per kilowatt-hour. Upgrading the operation and maintenance model from "reactive maintenance" and "periodic overhaul" to "predictive maintenance" has become a core issue for cost reduction and efficiency improvement in the industry.

[0051] Existing methods for assessing the health of wind turbines can be mainly categorized as follows: (1) Health assessment method based on independent modeling of single components. This type of method establishes health assessment models for key components such as gearbox, main shaft bearing, and generator, and judges their health status by monitoring the operating parameters of each component. This method is simple to implement, but ignores the physical coupling relationship and fault propagation effect between the components of the wind turbine. Abnormal vibration of the gearbox will be transmitted to the generator shaft end through the coupling, accelerating the wear of the generator bearing. The generator torque fluctuation will also have a reaction effect on the gearbox output end, forming a two-way coupling. Single component modeling cannot accurately depict this degradation propagation process that "affects the whole body when one part is affected".

[0052] (2) Health assessment methods based on static graph structures. Although a few methods based on graph neural networks attempt to model the relationships between components, the graph topology is fixed in advance and does not change with the operating conditions. Under strong wind / heavy load conditions, the components of the transmission chain are subjected to huge torques, and the coupling effect is significantly enhanced. Under light wind / light load conditions, the components tend to operate independently. Fixed graph structures cannot accurately reflect the true degradation coupling strength in the above two scenarios.

[0053] (3) Health assessment methods based on single data types. Existing methods mainly rely on slowly changing signals such as temperature and power in SCADA systems for health assessment. However, faults such as early pitting of gearbox bearings and microcracks on gear tooth surfaces are already manifested in vibration signals as kurtosis spikes and sideband energy increases before the temperature rises. The abnormal features such as kurtosis spikes and sideband energy increases contained in vibration signals are not effectively utilized, resulting in the output component health index failing to fully reflect the actual operating status of the components, thus leading to insufficient early fault detection capabilities. If vibration information can be effectively utilized, the fault warning window can be extended from the hourly level to the daily level. At the same time, existing methods only output health assessment results or health indices or directly output alarm signals, without providing maintenance decision suggestions. Maintenance plans still need to be formulated manually based on experience, resulting in low intelligence.

[0054] In other words, the existing technology has at least the following technical problems: (1) Single-component modeling cannot capture degradation propagation. There is a clear physical coupling relationship between the components of the wind turbine drivetrain—gearbox degradation will affect the generator input shaft load, and abnormal vibration of the generator will react on the gearbox output. Existing single-component modeling methods cannot capture this two-way coupling effect, resulting in a deviation between the health assessment results and the actual situation.

[0055] (2) Fixed graph structure cannot respond to load changes. Although existing graph neural network methods attempt to model the relationships between components, the graph topology is fixed in advance. Under high wind speed and high load conditions, the coupling between components is enhanced, but the edge weights in the graph remain unchanged, resulting in insufficient information propagation and decreased accuracy of health assessment under high load conditions.

[0056] (3) Relying on a single data mode, the health index cannot fully reflect the actual operating status of the component, resulting in insufficient early fault detection capability. Existing methods mainly rely on slowly changing signals such as temperature, and the early fault characteristics such as kurtosis surge and sideband energy rise contained in vibration signals are not effectively utilized, resulting in the health index being unable to fully reflect the actual operating status of the component.

[0057] (4) The health assessment and maintenance decision-making are disconnected. The health assessment results are not extended to the prediction of remaining life expectancy and maintenance decisions. The maintenance plan still needs to be formulated manually based on experience, and the level of intelligence is low.

[0058] (5) Insufficient accuracy in remaining life prediction. Existing remaining life prediction methods are mostly based on a single data source or single component modeling, and do not make full use of the historical change sequence of health index obtained based on multi-component coupled assessment, resulting in insufficient accuracy in remaining life prediction.

[0059] To address the aforementioned issues, this application provides a method and apparatus for wind turbine health assessment and maintenance based on graph attention networks. This method constructs a dynamic adjacency matrix, enabling the graph structure to dynamically adjust with changes in wind speed, active power, and vibration data, thus solving the problem of a fixed graph structure failing to respond to load changes. By using a graph attention network to weightedly fuse operational and vibration data, the health index of each key component comprehensively reflects its own state and the degradation effects of coupled components, solving the problems of single-component modeling failing to capture degradation propagation and the health index failing to fully reflect the actual operating state of components. By predicting the remaining service life based on the historical change sequence of the health index, time-dimensional information is provided for maintenance decisions. By determining maintenance actions based on the health index and remaining service life, the problem of the disconnect between health assessment and maintenance decision-making is solved. This application can accurately assess the health status of each key component in multi-component coupled scenarios, quantitatively predict the remaining service life of each key component, and output actions to guide actual maintenance operations.

[0060] It should be noted that the information and data related to users involved in the embodiments of this application are all information and data authorized by the user or fully authorized by the relevant parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with relevant laws, regulations, and standards, and necessary confidentiality measures have been taken. They do not violate public order and good morals, and corresponding operation entry points are provided for users or relevant parties to choose to authorize or refuse.

[0061] It should also be noted that the application embodiments may mention certain software, components, models and other existing solutions in the industry. These should be considered as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, they do not mean that the applicant has used or necessarily used the solution.

[0062] See Figure 1 As shown in the embodiments of this specification, a method for health assessment and maintenance of wind turbines based on graph attention networks is provided. In specific implementation, this method may include: S101: Obtain the operating data and vibration data of each key component in the wind turbine, wherein the operating data includes at least wind speed and active power; S102: Construct a component relationship graph with each key component as a graph node and the coupling relationship between key components as a base edge; determine the load transfer coefficient based on the wind speed and the active power; determine the anomaly propagation coefficient based on the degree to which the vibration data of the target key component deviates from its healthy reference vibration characteristic vector; and generate a dynamic adjacency matrix representing the intensity of degradation influence between key components based on the load transfer coefficient, the anomaly propagation coefficient, and the component relationship graph. S103: Input the running data, the vibration data, and the dynamic adjacency matrix into the graph attention network. The graph attention network is used to perform weighted fusion of the running data and the vibration data based on the dynamic adjacency matrix and output the health index of each key component. S104: Based on the historical change sequence of the health index at multiple consecutive historical moments, predict the remaining service life of each key component; S105: Based on the health index and the remaining service life, determine the maintenance actions for each key component, which are used to guide the maintenance operations of the wind turbine.

[0063] Specifically, the aforementioned operational data (or slow-changing signals) and vibration data (or fast-changing signals) can be collectively referred to as raw wind farm data. Operational data can be collected by the SCADA system, reflecting long-term degradation trends (increased temperature, decreased power), while vibration data can be collected by vibration sensors at various key components, reflecting instantaneous impact events (sudden increase in kurtosis, increase in sideband energy).

[0064] The aforementioned key components specifically refer to the core components of a wind turbine that significantly impact the safety and reliability of the unit's operation. The coupling relationships between these key components specifically refer to the physical transmission connections and other interactions between components. The aforementioned component relationship diagram specifically refers to a graph structure constructed with key components as nodes and the coupling relationships between key components as base edges, used to characterize the physical connections and degradation propagation paths between components. The degradation propagation path can be understood as the path through which the degradation state of a key component (such as increased gearbox vibration) transmits its influence to other coupled key components (such as the generator) via coupling relationships. For example, abnormal vibration of the gearbox is transmitted to the generator shaft end through the coupling, accelerating generator bearing wear; in this case, the physical connection between the gearbox and the generator is the degradation propagation path.

[0065] The aforementioned load transfer coefficient specifically refers to a coefficient dynamically determined based on wind speed and active power, used to characterize the modulation effect of operating load on the intensity of degradation effects between components. When both wind speed and active power are below their rated values, the load transfer coefficient approaches zero, and the intensity of degradation effects between components weakens. When wind speed and active power approach or exceed their rated values, the load transfer coefficient approaches a preset upper limit value, and the intensity of degradation effects between components increases. The intensity of degradation effects can be understood as the degree to which the degradation state of one critical component affects another critical component, i.e., the magnitude of the degradation effects transmitted along the degradation propagation path. For example, when the vibration of the gearbox intensifies abnormally, its degradation effect on the generator bearing increases, accelerating the wear of the generator bearing. The magnitude of the degradation effect intensity is characterized by the element values ​​in the dynamic adjacency matrix, which are dynamically determined jointly by the load transfer coefficient and the abnormal propagation coefficient. When wind speed and active power are high, or the degree of vibration abnormality deviation is large, the intensity of degradation effects increases. In this application, degradation refers to the process by which the health of critical components gradually declines over time, such as the cumulative wear of gearbox bearings leading to increased vibration, or the aging of generator insulation materials leading to decreased insulation performance.

[0066] The aforementioned anomaly propagation coefficient specifically refers to a coefficient dynamically determined based on the degree to which the vibration data (such as vibration characteristic vector) of the target critical component deviates from its healthy baseline vibration characteristic vector. It is used to characterize the intensity of vibration anomaly information propagating along the coupling relationship between components. The further the vibration characteristic vector of the target critical component deviates from its healthy baseline vibration characteristic vector, the greater the intensity of the anomaly information propagated by that component to other critical components coupled with it.

[0067] The aforementioned key components can specifically refer to the components whose vibration characteristic vector deviates from its healthy baseline vibration characteristic vector at the current moment, serving as the transmitting end components for abnormal propagation.

[0068] The aforementioned dynamic adjacency matrix can specifically refer to a matrix generated based on the load transfer coefficient, anomaly propagation coefficient, and component relationship diagram. Each element in the matrix represents the intensity of degradation influence between two corresponding key components.

[0069] The graph attention network described above can be understood as a neural network model that can propagate information and aggregate features based on the topology of a graph.

[0070] The aforementioned remaining useful life can also be simply referred to as remaining lifespan. Predicting remaining useful lifespan is equivalent to predicting remaining lifespan. The aforementioned health index specifically refers to the numerical value that characterizes the health status of key components, and its value range can be [0,1].

[0071] In practice, this application first acquires two types of heterogeneous data—operational data and vibration data—of key components in the wind turbine through a SCADA system and vibration sensors, respectively, providing a data foundation for the subsequent output of the health index.

[0072] Next, a component relationship graph is constructed using each key component as a graph node and the coupling relationships between key components as base edges. After the graph structure is determined, the load transfer coefficient is dynamically calculated based on the current wind speed and active power. Simultaneously, the anomaly propagation coefficient is dynamically calculated based on the degree to which the vibration characteristic vector of the target key component deviates from its healthy baseline vibration characteristic vector. The load transfer coefficient and the anomaly propagation coefficient are applied to each base edge in the component relationship graph to generate a dynamic adjacency matrix. This matrix is ​​dynamically adjusted according to changes in wind speed, active power, and vibration data, allowing the intensity of degradation effects between components to change in real time with the actual operating conditions.

[0073] Then, the operational data, vibration data, and dynamic adjacency matrix are input into the graph attention network. The graph attention network performs weighted fusion of the operational and vibration data of each key component based on the dynamic adjacency matrix, so that the health index of each key component not only reflects its own operational status, but also comprehensively reflects the degradation effects of other key components coupled with it, and outputs the health index of each key component.

[0074] Subsequently, based on the historical change sequence of the health index of each key component at multiple consecutive historical moments, the remaining service life of each key component is predicted.

[0075] Finally, based on the health index and remaining service life of each key component, the maintenance actions for each key component are determined to guide the actual maintenance operations of the wind turbine.

[0076] Based on the above embodiments, by constructing a dynamic adjacency matrix and utilizing a graph attention network for weighted fusion of multimodal data, the graph structure is dynamically adjusted according to the operating status, enabling the health index to comprehensively reflect the coupling degradation effects of multiple components. Simultaneously, through remaining life prediction and maintenance action determination, effective integration of health assessment and maintenance decision-making is achieved.

[0077] In some embodiments, the aforementioned key components may include at least blades, main shaft bearings, gearboxes, generators, and converters; the coupling relationship between the aforementioned key components may include at least physical transmission connections; the aforementioned target key component is a transmitting end component for abnormal propagation, used to propagate vibration abnormality information to other key components coupled with it, and the other key components are receiving end components for abnormal propagation.

[0078] Specifically, the aforementioned key components can include at least blades, main shaft bearings, gearboxes, generators, and converters. Blades are the core component of a wind turbine for capturing wind energy, and their health directly affects the unit's power generation efficiency and load distribution. Main shaft bearings connect the blades to the gearbox and bear the complex loads transmitted by the blades. The gearbox is the core component of the transmission chain, and wear on its internal gears and bearings directly affects transmission efficiency. The generator converts mechanical energy into electrical energy, and aging of its winding insulation leads to a decrease in power generation efficiency. The converter is responsible for power conversion and grid connection control, and degradation of its power modules affects the quality of power output.

[0079] The coupling relationships between the aforementioned key components specifically refer to the physical transmission connections and other interactive relationships between components. In wind turbine generators, the blades, main shaft bearings, gearbox, generator, and converter are sequentially connected by physical transmission chains, forming a clear coupling relationship. For example, abnormal vibrations in the gearbox can be transmitted to the generator shaft end through the coupling, accelerating the wear of the generator bearings. In this case, the gearbox is the transmitting component of the abnormal vibration, and the generator is the receiving component. Conversely, torque fluctuations in the generator can also have a reaction effect on the gearbox output end, forming a bidirectional coupling.

[0080] The aforementioned target critical component acts as the transmitting end component for anomaly propagation, used to transmit vibration anomaly information to other critical components coupled with it. These other critical components specifically refer to critical components connected to the target critical component through coupling, acting as receiving end components for anomaly propagation. For example, when the vibration characteristic vector of the gearbox deviates from its healthy reference vibration characteristic vector, the gearbox is the transmitting end component, and the generator coupled to the gearbox is the receiving end component. The vibration anomaly information of the gearbox propagates to the generator along the physical transmission connection, accelerating the generator's degradation process.

[0081] In some embodiments, the operating data and vibration data acquired in S101 above may specifically include: The above operating data may include at least the following parameters: wind speed v(t)(m / s), active power P(t)(MW), pitch angle β(t)(°), generator speed ω(t)(rpm), gearbox oil temperature Toil(t)(°), generator winding temperature Tgen(t)(°), main shaft bearing temperature Tmain(t)(°), and unit operating status Status(t).

[0082] The vibration data mentioned above includes at least a vibration characteristic vector. Vibration sensors are installed in at least three of the following components: gearbox input shaft bearing housing, gearbox output shaft bearing housing, generator drive end bearing housing, generator non-drive end bearing housing, and main shaft bearing housing. For example, vibration sensors can be installed in the gearbox input shaft bearing housing (V1, measurement direction: radial + axial, monitoring object: low-speed gearbox), gearbox output shaft bearing housing (V3, measurement direction: radial + axial, monitoring object: high-speed gearbox), and generator drive end bearing housing (V4, measurement direction: radial, monitoring object: front generator bearing) to form a minimum configuration of 3 measuring points. Based on this, additional gearbox intermediate shaft bearing housing (V2, measurement direction: radial, monitoring object: medium-speed gearbox), generator non-drive end bearing housing (V5, measurement direction: radial, monitoring object: rear generator bearing), and main shaft bearing housing (V6, measurement direction: radial + axial, monitoring object: main shaft bearing) can be added to form a recommended configuration of 6 measuring points, covering all critical components.

[0083] After obtaining the above-mentioned operational and vibration data, preprocessing operations can be performed on the two types of data to provide standardized inputs for subsequent dynamic adjacency matrix generation and health index output.

[0084] Specifically, the following preprocessing can be performed on the runtime data: (1) Time alignment and resampling: Resample to the model input time resolution Δt (e.g., 10 minutes). Use linear interpolation to fill missing values; if consecutive missing values ​​exceed the threshold (e.g., 3 consecutive points), discard the time window.

[0085] (2) Outlier removal: Remove values ​​that are obviously out of bounds (such as wind speed > 40 m / s, power > 1.2 times the rated power, temperature exceeding the sensor range, etc.).

[0086] (3) Data standardization: All numerical features are standardized using Z-score:

[0087] Where x represents the original data; μ represents the statistical mean of the training set; σ represents the statistical standard deviation of the training set; x norm This is the standardized data.

[0088] For vibration data, the following preprocessing can be performed: (1) Raw signal acquisition: Vibration sensor sampling frequency f s ≥25.6kHz, sampling duration T s =5~10 seconds / time, the data collection cycle is synchronized with the running data (e.g., 10 minutes / time).

[0089] (2) Temporal feature extraction: Extract temporal features from the original vibration signal, which may include at least the peak value X. peak Root mean square X rms , kurtosis X kurtosis and waveform factor X shape Among them, the peak value reflects the instantaneous impact energy; the root mean square (RMS) reflects the total vibration energy; the kurtosis is sensitive to the impact pulse, with a healthy bearing having a kurtosis ≈ 3 and a faulty bearing having a kurtosis > 5; the waveform factor is the RMS X. rms With peak X peak The ratio is used to reflect changes in waveform shape.

[0090] (3) Frequency domain feature extraction: After performing FFT transformation on the original vibration signal, the frequency domain features are extracted, which can include at least the rotational frequency amplitude A. 1X Meshing sideband energy A mesh Bearing characteristic frequency amplitude A bearing High-frequency energy ratio E ratio Among them, the rotational frequency amplitude is the amplitude of FFT at 1× rotational speed frequency, used to reflect rotor imbalance; the meshing sideband energy is the sideband energy of FFT at gear meshing frequency ± n× rotational speed frequency, used to reflect gear wear / pitting; the bearing characteristic frequency amplitude is the amplitude of FFT at bearing fault characteristic frequency, used to reflect early bearing faults; the high-frequency energy ratio is the ratio of energy in the 2kHz~10kHz frequency band to energy in the full frequency band, which is sensitive to early microcracks.

[0091] (4) Eigenvalue standardization: Z-score standardization is performed on each vibration feature to eliminate the influence of different dimensions. Through the above preprocessing, the original vibration waveform (approximately 25.6k×10s=256k points / time) can be compressed into 8 eigenvalues / measurement points, with a data compression ratio of over 3000:1, significantly reducing storage and transmission costs.

[0092] For critical components without vibration sensors (such as converters), their vibration characteristic vectors are filled with a preset zero vector. These critical components only participate in the information aggregation of the graph attention network as information receivers and do not actively propagate vibration anomaly information. The health status of critical blade components is indirectly characterized by operating data such as pitch angle fluctuations, power losses, and main shaft load changes, without directly relying on vibration sensors.

[0093] The preprocessed standardized running data is denoted as x. SCADA (t)∈R Ds Its dimension is Ds=8. The standardized vibration data is denoted as x. vib (t)∈R Dv Its dimension is Dv=8×N sensors , where N sensorsTo determine the number of measurement points, a minimum of 3 measurement points are configured, hence the typical value Dv=24. Standardized operational data and standardized vibration data are concatenated and used as input for the subsequent graph attention network. The subsequent graph attention network can then perform weighted fusion of the standardized operational and vibration data based on a dynamic adjacency matrix to accurately output the health index of each key component.

[0094] Based on the above embodiments, by performing time alignment, outlier removal, and standardization on the running data, and by extracting and standardizing the vibration data in the time and frequency domains, outliers and dimensional differences in the data can be eliminated, and the original vibration waveform can be compressed into a low-dimensional feature vector, which greatly reduces the cost of data storage and transmission, and provides a standardized multimodal data foundation for the subsequent generation of dynamic adjacency matrices and accurate output of health indices.

[0095] In some embodiments, the aforementioned key components can be represented as a graph node set V={v1,v2,v3,v4,v5}={blades, main shaft bearings, gearbox, generator, converter}, i.e., a default configuration of 5 key components (doubly fed induction generator), covering the key components with the highest failure rate and maintenance cost in the wind turbine drivetrain. For direct-drive turbines, the gearbox node can be deleted, in which case the number of key components K=4.

[0096] The specific construction method of the above component relationship graph can be as follows: each key component is a graph node, and the coupling relationship between key components is a base edge. The weight of the base edge can be a physical transmission connection (1 for connection, 0 for no connection). This component relationship graph is used to characterize the physical connections and degradation propagation paths between components, providing the basic topological structure for the subsequent generation of the dynamic adjacency matrix. In the dynamic adjacency matrix W(t), each element represents the degradation influence intensity (degradation propagation intensity) between two adjacent key components, which are two key components connected by a base edge. The larger the element value, the more significant the impact of degradation of one key component on the other. The weight of the base edge can also be dynamically modulated according to the load transfer coefficient, i.e.:

[0097] Where v(t) is the wind speed at time t; P(t) is the active power at time t; The load transfer weight of critical component i to critical component j, which is a neighbor of critical component i, is used as the weight of the modulated base edge. The weights of the base edges are used to characterize whether there is a physical transmission connection between two key components (1 for connection, 0 for no connection). represents the dynamic scaling factor. Here, i is the key component number (receiving component / affected key component, also denoted by p), and j is the neighboring key component number (transmitting component / affected key component, also denoted by q). The weights of the modulated base edges can be used for the subsequent generation of the dynamic adjacency matrix.

[0098] Based on the above embodiments, by dynamically modulating the base edge weights according to real-time wind speed and active power, and using the modulated base edge weights to generate the dynamic adjacency matrix, the accuracy of dynamic adjacency matrix generation can be further improved.

[0099] In some embodiments, determining the load transfer coefficient based on the wind speed and active power in S102 above may, in specific implementation, include: The load transfer factor is determined using the following formula:

[0100] Among them, F 载荷 (t) is the load transfer coefficient at time t; The load modulation intensity coefficient; The function is the Sigmoid function; v(t) is the wind speed at time t; v rated P is the rated wind speed; P(t) is the active power at time t; P rated This refers to the rated active power.

[0101] Specifically, the process of determining the load transfer coefficient reflects the dynamic modulation effect of the load on the coupling strength between key components. When the real-time wind speed v(t) is lower than the rated wind speed v... rated And the active power P(t) is lower than the rated active power P rated When the Sigmoid function outputs a smaller value, the load transfer coefficient F 载荷 When the sigmoid function output approaches zero, it indicates that the key components of the transmission chain tend to operate independently, and the degradation effect between key components is relatively weak. When the real-time wind speed v(t) and active power P(t) approach or exceed their corresponding rated values, the sigmoid function output approaches 1, and the load transfer coefficient F... 载荷 (t) approaches the load modulation intensity coefficient, indicating that the components of the transmission chain are subjected to a large torque, the coupling effect between key components is significantly enhanced, and the abnormality in the gearbox will be quickly transmitted to the generator, characterizing the tightly coupled state of the transmission chain. By multiplying the two Sigmoid functions and comprehensively considering the combined effects of wind speed and active power on the load transmission coefficient, the changes in the intensity of degradation effects between components under different operating conditions can be accurately reflected. Among them, the load modulation intensity coefficient... To control the maximum scaling factor of the load on the coupling strength between components (typically 0.3–0.8), it is determined by the validation set. The Sigmoid function is used to map the load level to the (0,1) interval. F 载荷 (t)∈[0,γ] represents the independent contribution of the load condition to the coupling strength between key components.

[0102] Based on the above embodiments, by using the above method to determine the load transfer coefficient, the wind speed deviation and active power deviation are mapped to the (0,1) interval by the Sigmoid function and then multiplied. This can comprehensively consider the nonlinear influence of wind speed and active power on the degradation intensity between components, making the calculation of the load transfer coefficient physically interpretable and smoothly reflecting the changing trend of the coupling strength between components under different load levels. This provides an accurate load transfer coefficient for the subsequent generation of the dynamic adjacency matrix.

[0103] In some embodiments, the vibration data may at least include a vibration feature vector; correspondingly, the degree to which the vibration data based on the target critical component deviates from its healthy baseline vibration feature vector in S102 determines the anomaly propagation coefficient, which, in specific implementations, may include: The anomaly propagation coefficient is determined using the following formula:

[0104] in, Let be the intensity of the vibration anomaly information propagated from the transmitting component q to the receiving component p at time t, and let be the anomaly propagation coefficient at time t; exp is an exponential function with the natural constant as its base; K q (t) is the vibration characteristic vector of the transmitting component q at time t; The health reference vibration characteristic vector of the transmitting component q; For bandwidth parameters; is the vector norm between the vibration characteristic vector and the health reference vibration characteristic vector, used to characterize the degree to which the vibration characteristic vector of the transmitting component q deviates from its health reference vibration characteristic vector.

[0105] Specifically, the aforementioned health benchmark vibration characteristic vector Specifically, a preset statistical quantile, such as the 90th quantile, can be taken from the vibration characteristics of the transmitting component q under normal operating conditions. The bandwidth parameters mentioned above... The sensitivity of the vibration anomaly propagation coefficient to the degree of deviation is used to control its value; the smaller the coefficient, the more significant the amplification effect of the vibration deviation on the propagation intensity. The physical meaning of the aforementioned anomaly propagation coefficient is: when the transmitting component q is at time t, the vibration characteristic vector K... q (t) deviates from its healthy baseline vibration characteristic vector The farther away the transmitting component q is, the stronger the intensity of the vibration anomaly information propagated from it to the receiving component p coupled with it, thus accurately characterizing the causal direction of fault propagation. For example, when the vibration characteristic vector of the gearbox (such as kurtosis, root mean square, high-frequency energy ratio, etc.) deviates further from its healthy reference vibration characteristic vector, the stronger the intensity of the abnormal information propagated from the gearbox to the generator, accelerating the degradation process of the generator.

[0106] For critical components (such as converters) without installed vibration sensors, their vibration characteristic vector K q (t) and the health baseline vibration characteristic vector All are defined as zero vectors and do not participate in the calculation of vibration anomaly propagation coefficients. This key component only participates in the information aggregation of the graph attention network as an information receiving end.

[0107] Based on the above embodiments, by determining the above-mentioned abnormal propagation coefficient, the intensity of the vibration abnormality information propagated from the transmitting end component to the receiving end component can be dynamically adjusted according to the degree to which the vibration characteristic vector of the transmitting end component deviates from its health reference vibration characteristic vector. This enables the corresponding elements in the dynamic adjacency matrix to respond to the vibration abnormality changes in a timely manner, accurately transmit the degradation information of the abnormal component to its coupled adjacent components in the graph attention network, and make the output health index more accurately reflect the coupling degradation relationship between components.

[0108] In some embodiments, see Figure 2 As shown, in S102 above, generating a dynamic adjacency matrix representing the intensity of degradation influence between key components based on the load transfer coefficient, the anomaly propagation coefficient, and the component relationship diagram can, in specific implementation, include: S201: Construct a load transfer matrix based on the load transfer coefficient, and construct a vibration anomaly propagation matrix based on the anomaly propagation coefficient; S202: Construct a physical connection basis matrix based on the weights of each base edge in the component relationship diagram; S203: Perform element-wise multiplication on the physical connection basis matrix, the load transfer matrix, and the anomaly propagation matrix to generate the dynamic adjacency matrix.

[0109] In some embodiments, the element-wise multiplication operation of the physical connection basis matrix, the load transfer matrix, and the anomaly propagation matrix in S203 above to generate the dynamic adjacency matrix may, in specific implementations, include: The dynamic adjacency matrix is ​​generated according to the following formula:

[0110] Where W(t) is the dynamic adjacency matrix at time t; W base For the physical connection basis matrix; For element-wise multiplication (Hadamard product); f 载荷 (t) is the load transfer matrix at time t; W vib (t) is the vibration anomaly propagation matrix at time t.

[0111] Specifically, the aforementioned load transfer matrix f 载荷 Each element in (t) is a load transfer coefficient corresponding to the base edge, representing the modulation effect of the operating load on the intensity of degradation between components. The aforementioned vibration anomaly propagation matrix W... vib Each element in (t) represents the anomaly propagation coefficient of the corresponding base edge, characterizing the intensity of vibration anomaly information propagating along the coupling relationship between components. The aforementioned physical connection base matrix W base Each element in the diagram represents the weight of each base edge in the component relationship diagram, used to characterize whether there is a physical transmission connection between two key components (1 for connection, 0 for no connection). For key components without vibration sensors (such as converters), the anomalous propagation coefficient is zero when acting as a transmitter.

[0112] The dynamic adjacency matrix is ​​generated using a joint influence form 1-(1-a)(1-b), where a represents an element in the load transfer matrix and b represents an element in the vibration anomaly propagation matrix. This form ensures that: when both the load transfer matrix and the vibration anomaly propagation matrix are in a minimum state or under light load conditions and without vibration anomalies (f 载荷 (t)=0, W vib When (t)=0, the corresponding elements in the dynamic adjacency matrix approach zero. When only the load transfer matrix or only the vibration anomaly propagation matrix increases, the corresponding elements monotonically increase. When both increase simultaneously, the corresponding elements further increase but do not exceed a preset physical upper limit, such as not exceeding 1. The physical meaning of this form is clear: load transfer and vibration anomaly propagation are two independent degradation propagation mechanisms. The existence of either mechanism will enhance the degradation effect intensity between components, but the superposition of the two will not exceed the preset physical upper limit of complete influence. For example, when the wind speed and active power are high (larger load transfer coefficient) and the degree of vibration anomaly deviation is large (larger anomaly propagation coefficient), the corresponding elements in the dynamic adjacency matrix further increase, and the degradation effect intensity between components reaches its maximum.

[0113] The elements in the dynamic adjacency matrix W(t) are then used in the graph attention network to calculate the attention scores of each key component to its neighboring key components, thereby determining the information propagation path and information propagation intensity between key components.

[0114] Based on the above embodiments, this application constructs a component relationship graph using key components as graph nodes and the coupling relationships between key components as base edges. It introduces load transfer coefficients and vibration anomaly propagation coefficients as dual driving mechanisms to generate a dynamic adjacency matrix, enabling real-time dynamic adjustment of the degradation impact intensity between components according to the actual operating state. Compared to existing methods with fixed graph structures or independent modeling of single components, this application proposes for the first time a component relationship graph based on the physical connection of the transmission chain and driven by real-time load and vibration anomaly propagation. When the vibration characteristics of a key component exhibit abnormalities (such as a sudden increase in kurtosis or an increase in sideband energy), vibration anomaly information is transmitted to adjacent key components through physical connection edges, allowing fault information to propagate rapidly along the transmission chain. The elements in the dynamic adjacency matrix are updated in real-time with changes in load and vibration anomalies, providing a reliable graph structure foundation for accurately calculating the attention scores of each key component to its neighboring key components in the subsequent graph attention network. This allows the output health index to more accurately reflect the coupling degradation relationship between components.

[0115] In some embodiments, see Figure 3 As shown, the graph attention network in S103 above is used to perform weighted fusion of the operating data and the vibration data based on the dynamic adjacency matrix, and output the health index of each key component. In specific implementation, it may include: S301: The operating data and vibration data of each key component are stitched together and linearly projected to obtain the initial nodal features of each key component; S302: For each base edge in the component relationship graph, the two key components corresponding to the base edge are each other's neighbor key components. Based on the initial node features of each key component, the initial node features of its neighbor key components, and the weight of the corresponding base edge in the dynamic adjacency matrix, the attention score of each key component to its neighbor key components is calculated. S303: Normalize the attention scores to obtain the normalized attention weights of each key component to its neighboring key components; S304: According to the normalized attention weights, the initial node features of the neighboring key components of each key component are weighted and aggregated to obtain aggregated features. The aggregated features are then residually connected with the transformed features of each key component after residual transformation to obtain the fusion features of each key component. The graph attention network contains at least two graph attention layers. The fusion features of each key component are used as the input of the next graph attention layer. The operations of attention score calculation, normalization processing, weighted aggregation and residual connection are repeatedly performed until the graph attention calculation of the preset number of layers is completed, and the fusion features of each key component in the last layer are obtained. S305: The fusion features of each key component in the last layer are mapped to the range of health index values ​​through a fully connected layer and activation function, and the health index of each key component is output.

[0116] Specifically, the aforementioned initial node features refer to the feature vector obtained by linearly projecting the concatenated operational and vibration data of each key component, which serves as the input features for the corresponding nodes in the graph attention network. The operational data reflects long-term degradation trends, while the vibration data reflects instantaneous impact events. By concatenating the data, the node features simultaneously include both slowly changing and rapidly changing signals, achieving the fusion and perception of multimodal information.

[0117] The aforementioned neighboring critical components can specifically refer to other critical components connected to the current critical component through the base edge, i.e., the neighboring nodes of the current critical component.

[0118] The aforementioned attention score specifically refers to the degree of attention the current critical component pays to its neighboring critical components, used to determine the weight of the features of the neighboring critical components in the feature aggregation of the current critical component. The aforementioned fused feature specifically refers to the updated feature, after processing by the graph attention layer, which integrates its own features and the features of its neighboring critical components.

[0119] In practice, the operational and vibration data of each key component are first concatenated and linearly projected to obtain the initial node features of each key component. These initial node features contain multimodal fusion information from the operational and vibration data. Next, for each base edge in the component relationship graph, the two key components corresponding to each base edge are designated as neighboring key components. Based on the initial node features of each key component, the initial node features of its neighboring key components, and the weights of the corresponding base edges in the dynamic adjacency matrix, the attention score of each key component to its neighboring key components is calculated. The weights of the corresponding base edges in the dynamic adjacency matrix determine the intensity of the degradation influence between two key components, affecting the magnitude of the attention score. Then, the attention scores are normalized to obtain normalized attention weights. Next, the initial node features of the neighboring key components are weighted and aggregated according to the normalized attention weights. The aggregated result is then residually concatenated with its own features after residual transformation to obtain the fused features. The graph attention network contains at least two graph attention layers. The fused features are used as input to the next layer, and the above operations are repeated until the preset number of layers is calculated, resulting in the fused features of the final layer. Finally, the fusion features of the last layer are mapped to the range of health index values ​​through a fully connected layer and an activation function, and the health index of each key component is output.

[0120] Based on the above embodiments, by stitching together and projecting operational and vibration data into initial node features, the initial node features simultaneously contain both slowly changing and rapidly changing signals, achieving multimodal information fusion perception. By using the weights of the corresponding basis edges in the dynamic adjacency matrix to participate in the attention score calculation, physical coupling information directly influences the information propagation process. Through the stacking of multiple graph attention layers, the fused features of key components not only reflect their own operational status but also comprehensively reflect the degradation effects of other key components coupled with them, thus providing a feature foundation for outputting an accurate health index.

[0121] In some embodiments, the step of stitching together the operating data and vibration data of each key component in S301 above, and obtaining the initial node features of each key component through linear projection, may, in specific implementations, include: The preprocessed standardized operational data and standardized vibration data are stitched together according to the following formula:

[0122] Where, x i SCADA (t)∈R Ds The standardized operational data features of key component i at time t have a dimension of Ds=8; x i vib (t)∈R Dv The data represents the standardized vibration data of the critical component i at time t, with dimensions Dv = 8 × N. sensors , where N sensors To determine the number of measurement points, a minimum of 3 measurement points are required. The two types of standardized features are concatenated to obtain the concatenated multimodal feature vector x. i (t)∈R Dx Its dimension is Dx=Ds+Dv=8+24=32. For critical components without vibration sensors (such as converters), their vibration data features are filled with a preset zero vector. At this time, the splicing features of the critical component only contain operational data features.

[0123] Then, the concatenated multimodal fusion feature vector x i (t) is mapped to the hidden layer space via linear projection to obtain the initial node features of each key component. :

[0124] Among them, W in b is a learnable projection matrix; in This is a bias term.

[0125] The initial node features mentioned above can then be used as a basis for further analysis. The aforementioned dynamic adjacency matrix W(t) serves as the input to the graph attention network (GATv2). The output of the graph attention network can include the health index HI of each key component. i (t) may also include the overall health index of the whole machine (to be explained later, and will not be elaborated here), and the fusion features of the key component i output by the Lth layer of the graph attention network. .

[0126] Based on the above embodiments, by stitching together the running data and vibration data and then linearly projecting them, the initial node features of each key component simultaneously contain multimodal fusion information of slow-changing signals (running data) and fast-changing signals (vibration data), providing a node feature basis for fusion perception for subsequent information propagation in the graph attention network.

[0127] In some embodiments, see Figure 4 As shown, the calculation of the attention score of each key component to its neighboring key components in S302 above can, in specific implementation, include: S401: For each key component and its neighboring key components, perform linear transformations on the initial node features of the key component and the initial node features of its neighboring key components to obtain the query features of the key component and the key features of its neighboring key components. S402: Add the query features, the key features, and the physical bias component, pass them through a nonlinear activation function, and then perform an inner product operation with the attention vector to obtain the attention score of the key component to its neighboring key components; wherein, the physical bias component is determined according to the weight of the connection relationship between the key component and its neighboring key components in the dynamic adjacency matrix.

[0128] Specifically, in step S401 above, for each key component and its neighboring key components, the initial node features of the key component are linearly transformed using a query transformation matrix to obtain the query features of the key component. These query features represent the state of the key component as the current node (receiver). Simultaneously, the initial node features of the neighboring key components are linearly transformed using a key transformation matrix to obtain the key features of the neighboring key components. These key features represent the state of the neighboring key components as neighboring nodes (senders). Both the query transformation matrix and the key transformation matrix are learnable parameter matrices.

[0129] In step S402 above, the query features, key features, and physical bias components are added together, processed by a nonlinear activation function, and then multiplied by the transpose of the attention vector to obtain the attention score of the key component to its neighboring key components. The physical bias component is determined based on the weights of the base edges between the key component and its neighboring key components in the dynamic adjacency matrix. When the base edge weight is zero, the physical bias component is zero and does not contribute to the attention score, effectively physically prohibiting information propagation from that edge. When the base edge weight is 1, the physical bias component is approximately 0.693, making a significant positive contribution.

[0130] The attention score mentioned above specifically refers to the degree of attention the current key component pays to its neighboring key components, and is used to determine the weight of the features of the neighboring key components in the feature aggregation of the current key component. The higher the attention score, the higher the degree of attention the current key component pays to its neighboring key components, and the greater the weight of the features of the neighboring key components in the feature aggregation of the current key component.

[0131] Based on the above embodiments, by incorporating the weights of the corresponding base edges in the dynamic adjacency matrix into the attention calculation in the form of additive bias, the intensity of degradation influence between components is directly applied to the attention score, avoiding the gradient saturation problem and enabling the attention score to more accurately reflect the degradation influence relationship between key components.

[0132] In some embodiments, the addition of the query features, the key features, and the physical bias component in S402 above, followed by a nonlinear activation function and then an inner product operation with the attention vector to obtain the attention score of the key component to its neighboring key components, may, in specific implementations, include: Calculate the attention score using the following formula:

[0133] in, e ij Let i be the attention score of critical component i to its neighboring critical component j; W is the transpose of the attention vector; LeakyReLU is the non-linear activation function; query To query the transformation matrix, used for the initial node features of key component i. Perform a linear transformation to obtain the query features of key component i; W key Here is the key transformation matrix, used to initialize the node features of neighboring key component j. Perform a linear transformation to obtain the key features of the neighboring key component j; For query features; Key features; w ij (t) represents the weight of the base edge between key component i and its neighbor key component j in the dynamic adjacency matrix at time t; This is the physical bias component.

[0134] Specifically, the aforementioned physical bias components The introduction method has clear physical interpretability and mathematical advantages. When the basis edge weights w ij When (t)=0 (lightly loaded decoupling or no physical connection), the physical bias component is log(1)=0, contributing no attention score, which is equivalent to physically prohibiting information propagation on that edge. When the base edge weight w ij When (t)=1 (strong coupling), the physical bias component is log(2)≈0.693, making a significant positive contribution. Mathematically, during backpropagation... The gradient remains within the range of [0.5,1], avoiding both gradient vanishing and gradient explosion when the weights are too small. By incorporating the physical weights into the attention score as additive biases instead of direct multiplication, the physically coupled information can stably influence the information propagation process.

[0135] Based on the above embodiments, by incorporating the weights of the corresponding base edges in the dynamic adjacency matrix into the attention calculation in the form of logarithmic bias, the contribution of the degradation influence intensity between components to the attention score increases monotonically with the increase of the base edge weights, and the gradient saturation problem is avoided. This ensures that the degradation influence relationship between components makes a stable contribution to the attention score during the training process, thereby more accurately reflecting the degradation influence relationship between key components in subsequent feature aggregation.

[0136] In some embodiments, the normalization process of the attention scores in S303 above, to obtain the normalized attention weights of each key component to its neighboring key components, may include: The attention score was normalized using the following formula:

[0137] in, Here is the normalized attention weight of critical component i to its neighboring critical component j; e ij Let be the attention score of critical component i to its neighboring critical component j; exp( ) is an exponential function with the natural constant as the base; N(i) is the set of neighboring key components of key component i, that is, the set of other key components connected to key component i through the base edge; The attention scores of all neighboring key components of key component i are exponentially summed, and this sum is used as the normalized denominator. This formula means that the attention scores are converted into a probability distribution (summing to 1), with edges having higher values ​​receiving higher weights during aggregation. Normalizing the attention scores provides a quantifiable basis for subsequent feature weighting and aggregation.

[0138] In some embodiments, the step S304 above, which involves weighted aggregation of the initial node features of the neighboring key components of each key component according to the normalized attention weights to obtain aggregated features, and residual connection of the aggregated features with the transformed features of each key component after residual transformation to obtain fused features of each key component, may, in specific implementations, include: Weighted aggregation and residual join are performed according to the following formula:

[0139] in, The fused features of key component i after processing by a graph attention network layer; The initial node features of key component i; The initial node features of the neighboring key component j; W represents the normalized attention weight of critical component i to its neighboring critical component j. res W is the residual transformation matrix; v Value transformation matrix; For transformation features; is the aggregation feature; N(i) is the set of neighboring key components of key component i; ELU is the activation function.

[0140] Specifically, in the aforementioned weighted aggregation and residual join process, the value transformation matrix W is first used... v Initial node features of neighboring key component j Perform a linear transformation to obtain the value transformation characteristics of the neighboring key component j. Then, according to the normalized attention weights... The aggregated features are obtained by weighted summation of the value transformation features of each neighboring key component. This aggregated feature integrates the degradation impact information of all neighboring key components on key component i, and normalizes the attention weights. The larger the value, the greater the proportion of the value transformation feature of the corresponding neighboring key component j in the aggregated features. Simultaneously, through the residual transformation matrix W... res Initial node features of the key component i itself Perform a linear transformation to obtain the transformation characteristics of the key component i. This transformation feature is used to preserve the information of the key component i itself, ensuring that the original features of the key component are not lost even if the graph structure is sparse or attention is distracted in the early stages of training. This combines aggregated features with the transformation feature. After addition, it is activated by the ELU function. The process involves processing key component i to obtain its fused features after passing through a graph attention network layer. Among them, the residual transformation matrix Wres It can be initialized to a preset scaling factor of the identity matrix (such as 0.1x) to ensure that the model mainly uses its own features in the early stages of training and gradually learns to trust neighbor information.

[0141] The fused features of the aforementioned key component i after processing by a graph attention network layer By integrating the state information of key component i itself and the degradation impact information of its neighboring key components, effective fusion of multi-component coupling features is achieved. The aforementioned graph attention network contains at least two graph attention layers. The fused feature of key component i after processing by one layer of the graph attention network is used as the input to the next layer. The operations of attention score calculation, normalization, weighted aggregation, and residual connection are repeatedly performed until a predetermined number of graph attention layers are completed, resulting in the fused feature of key component i in the last layer (e.g., layer L). By stacking multiple layers, information can be propagated to multiple hop neighbors, capturing indirect coupling effects (such as blade anomalies → main shaft load changes → gearbox input torque fluctuations → generator output fluctuations). The default number of layers is 2.

[0142] Based on the above embodiments, the weighted aggregation and residual connection methods enable the fusion characteristics of each key component to comprehensively reflect its own state and the degradation impact of its neighboring key components. Residual connections ensure that the characteristics of key components are not lost during information propagation, and multi-layer stacking enables information to propagate to multi-hop neighbors, capturing indirect coupling effects, thereby more accurately reflecting the coupling degradation relationship of multiple components in the subsequent health index output.

[0143] In some embodiments, the step S305 above, which maps the fusion features of each key component in the last layer to the health index value range via a fully connected layer and an activation function, and outputs the health index of each key component, may include, in specific implementation: Output the health index of each key component according to the following formula:

[0144] Among them, HI i (t) represents the health index of key component i at time t, and the health index ranges from [0,1]. The fusion features of key component i in the output of the Lth layer of the graph attention network; W HI b is the weight matrix of the fully connected layer; HI is the bias term for the fully connected layer; Sigmoid is the activation function.

[0145] Specifically, in the above-mentioned health index output process, the fusion features of the key component i output by the Lth layer of the graph attention network are passed through the weight matrix W of the fully connected layer. HI and bias term b HIA linear transformation is performed to obtain the linear mapping result of key component i. Then, the linear mapping result is compressed to the [0,1] interval using the Sigmoid function to obtain the health index HI of key component i at time t. i (t). The Sigmoid function maps the output of the fully connected layer to the interval [0,1], where 1 represents a fully healthy layer and 0 represents a fully ineffective layer.

[0146] The above health index HI i (t) is used to characterize the health status of critical component i at time t, providing input for subsequent remaining service life prediction and maintenance decisions. Health Index HI i The closer (t) is to 1, the better the health status of critical component i; the closer it is to 0, the worse the health status of critical component i.

[0147] Based on the above embodiments, by mapping the fusion features of each key component in the last layer to the range of health index values ​​through a fully connected layer and activation function, it is possible to output numerical values ​​representing the degree of health status of each key component, providing accurate quantitative basis for health status for subsequent prediction of remaining service life and maintenance decisions.

[0148] In some embodiments, see Figure 5 As shown, after outputting the health indices of each key component as described above, in specific implementation, it may also include: S501: Obtain the current kurtosis value of each key component, the current kurtosis value being extracted from the vibration data; S502: When the current kurtosis value is greater than the healthy baseline kurtosis value, determine the vibration correction amount based on the deviation between the current kurtosis value and the healthy baseline kurtosis value; S503: Subtract the vibration correction amount from the health index to obtain the corrected health index.

[0149] Specifically, the aforementioned current kurtosis value can refer to the kurtosis feature value extracted from the vibration data of the key component at the current moment, used to reflect the impact pulse component in the vibration signal. The aforementioned health baseline kurtosis value can refer to the preset statistical quantile (e.g., the 90th percentile) of the kurtosis value of the key component under normal operating conditions, with a typical value of 2.8~3.2. The aforementioned vibration correction amount can refer to the correction value determined based on the deviation between the current kurtosis value and the health baseline kurtosis value, used to correct the health index.

[0150] In step S501 above, the current kurtosis value is extracted from the vibration data of each key component. This current kurtosis value is used to characterize the vibration impact level of the key component at the current moment. When a key component experiences an early failure, the kurtosis value will increase significantly, so the kurtosis value can serve as a sensitive indicator of vibration anomalies.

[0151] In step S502 above, when the current kurtosis value is greater than the healthy baseline kurtosis value, the vibration correction amount is determined based on the deviation between the current kurtosis value and the healthy baseline kurtosis value. The vibration correction amount increases with the increase of the kurtosis deviation and is limited by a preset maximum correction amplitude. When the current kurtosis value is less than or equal to the healthy baseline kurtosis value, the vibration correction amount is zero, and no correction is made to the health index.

[0152] In step S503 above, the vibration correction amount is subtracted from the health index to obtain the corrected health index. This correction enables the health index to respond to sudden faults characterized by vibration anomalies, achieving rapid early warning of minor faults.

[0153] Based on the above embodiments, by obtaining the current kurtosis value of each key component, and determining the vibration correction amount based on the deviation when the current kurtosis value is greater than the health baseline kurtosis value, the vibration correction amount is deducted from the health index, so that the corrected health index can respond in a timely manner to the sudden fault represented by vibration abnormality, and realize the rapid early warning of early minor faults.

[0154] In some embodiments, determining the vibration correction amount based on the deviation between the current kurtosis value and the health baseline kurtosis value in S502 above may, in specific implementation, include: The vibration correction amount is determined using the following formula:

[0155] in, This represents the vibration correction amount for the critical component i at time t. This is the maximum correction amplitude control parameter, used to limit the upper limit of the vibration correction amount; This is the kurtosis scaling factor, used to control how quickly the kurtosis reaches saturation; Let i be the current kurtosis value of the critical component i at time t; denoted as the health baseline kurtosis value of critical component i; tanh is the hyperbolic tangent function; max is the function for maximizing the value.

[0156] Specifically, the process for determining the vibration correction amount is as follows: First, the deviation between the current kurtosis value and the healthy baseline kurtosis value is calculated using the maximum value function `max`. When the current kurtosis value is less than or equal to the healthy baseline kurtosis value, the deviation is zero, and the vibration correction amount is zero. When the current kurtosis value is greater than the healthy baseline kurtosis value, the deviation is positive. Then, the deviation is multiplied by the kurtosis deviation scaling factor. After passing through the hyperbolic tangent function tanh( Map to the (-1, 1) interval, then multiply by the maximum correction magnitude control parameter. The vibration correction amount of the key component i at time t is obtained. .

[0157] The maximum correction amplitude control parameter (typically 0.1) limits the maximum downward adjustment of the health index due to vibration anomalies. The kurtosis deviation scaling factor (typically 0.5) controls how quickly the correction amount reaches saturation after the kurtosis deviates from the baseline value. A larger kurtosis deviation scaling factor results in a more sensitive correction response, while a smaller factor provides stronger resistance to high-frequency noise interference. Through a bounded nonlinear mapping of the hyperbolic tangent function, the vibration correction amount monotonically increases with the increase of kurtosis deviation, and is limited by the preset maximum correction amplitude, avoiding excessive correction amount that could lead to abnormal fluctuations in the health index.

[0158] Based on the above embodiments, by using the above method for determining the vibration correction amount, the vibration correction amount can be dynamically determined according to the deviation between the current kurtosis value and the health baseline kurtosis value. Moreover, the vibration correction amount is limited by the preset maximum correction range, so as to avoid the health index from fluctuating abnormally due to excessive correction amount, and to achieve a sensitive response of the health index to vibration anomalies.

[0159] In some embodiments, subtracting the vibration correction amount from the health index in S503 above to obtain the corrected health index may, in specific implementation, include: The corrected health index is obtained using the following formula:

[0160] in, HI is the corrected health index of key component i at time t. i (t) represents the health index of key component i at time t; This represents the vibration correction amount for the critical component i at time t.

[0161] Specifically, the vibration correction described above enables the health index to simultaneously reflect both long-term degradation trends and instantaneous shock anomalies. Temperature increases are slow-changing signals with high thermal inertia, reflecting long-term degradation trends. Increased kurtosis, on the other hand, is a fast-changing signal, sensitive to shocks and responding instantaneously, thus promptly reflecting sudden anomalies. By subtracting the vibration correction from the health index, the corrected health index can capture both long-term degradation trends (such as temperature changes) and respond to sudden anomalies (such as sudden increases in kurtosis).

[0162] For example: Gearbox health baseline kurtosis The measured kurtosis suddenly increased to K. i If (t) = 5.5, then the kurtosis is 5.5 - 3.0 = 2.5. The kurtosis scaling factor is 0.5, so tanh(0.5 × 2.5) = tanh(1.25) ≈ 0.85. If the health index HI... i (t)=0.75, maximum correction amplitude control parameter β vib =0.1, then the vibration correction amount The corrected health index of key component i at time t The health score was downgraded from "mild deterioration" to "significantly abnormal," triggering an earlier warning.

[0163] Based on the above embodiments, by deducting the vibration correction amount from the health index, the corrected health index can simultaneously reflect the long-term degradation trend and the sudden failure characterized by vibration anomalies, thereby providing a more accurate quantitative basis for the prediction of the remaining service life and maintenance decisions.

[0164] In some embodiments, after outputting the health index of each key component as described above, the implementation may further include: The overall health index of the machine is determined using the following formula:

[0165] Among them, HI turbine (t) represents the overall health index of the wind turbine at time t; HI i (t) represents the health index of key component i at time t; Let be the weight coefficient of key component i, and satisfy . K represents the total number of critical components.

[0166] Specifically, the aforementioned overall health index characterizes the health level of the wind turbine as a whole. It weights and sums the health indices of each key component to provide maintenance personnel with a rapid assessment of the overall health status of the unit. The weight coefficients for each key component are determined based on its failure cost, with higher weights assigned to components with higher failure costs. For example, the gearbox's weight coefficient can be set to 0.30, and the generator's weight coefficient can be set to 0.25, as their failure costs are the highest. The weight coefficients for other key components are allocated accordingly based on their failure costs.

[0167] HI turbine The value of (t) ranges from [0,1]. The closer it is to 1, the better the overall health of the wind turbine; the closer it is to 0, the worse the overall health of the wind turbine. This comprehensive health index can be used as an additional output of the health assessment module, providing a reference for the entire turbine at the turbine level for subsequent prediction of remaining service life and maintenance decisions.

[0168] Based on the above embodiments, by weighted summing of the health indices of each key component, a comprehensive health index representing the overall health level of the wind turbine is obtained, providing maintenance personnel with a rapid assessment of the overall health status of the unit and a reference basis for subsequent maintenance decisions at the whole unit level.

[0169] In some embodiments, the graph attention network can be pre-trained using temporal contrastive learning before outputting the health index of each key component: Construct the contrastive learning loss function according to the following formula:

[0170] Among them, L contrast For contrastive learning loss function; z i The health representation of key component i can specifically refer to the fused features of key component i output by the Lth layer of the graph attention network. ;z i + For positive sample health characterization of key component i; z j For negative sample health characterization of key component j; sim( ) is the cosine similarity function; τ cl is the temperature hyperparameter (typical value 0.1); S(i) is the negative sample set of key component i, used for comparative learning with positive samples.

[0171] Specifically, the aforementioned positive sample z i + Specifically, this can refer to the health characteristics of the same critical component (such as a gearbox) under similar operating conditions (same wind speed and active power range) but at different times. The construction of positive sample pairs utilizes the time-stationary prior of health status—healthy components should exhibit similar health characteristics under similar load conditions. If the health status of a critical component does not change, then its health characteristics should remain consistent under the same operating conditions. This prior does not rely on any fault labels and can be pre-trained on massive amounts of unlabeled operational and vibration data.

[0172] The above negative sample z j Specifically, this can refer to the health representation of sample pairs from different key components, different operating conditions, or known fault periods and healthy periods. Through comparative learning of positive and negative sample pairs, the graph attention network learns feature representations that can distinguish between healthy and abnormal representations.

[0173] The training process is as follows: The remaining service life prediction module and maintenance decision module are frozen, and only the graph attention network is trained using all unlabeled operational and vibration data for pre-training. After pre-training, the encoder of the graph attention network can effectively distinguish between healthy and abnormal characteristics. Then, the graph attention network is fine-tuned end-to-end using a small amount of labeled fault data, addressing the industry pain point of scarce fault samples in wind turbines.

[0174] Based on the above embodiments, the graph attention network is pre-trained using temporal contrastive learning and a large amount of unlabeled running data and vibration data. This enables the encoder of the graph attention network to effectively distinguish between healthy and abnormal representations. Then, a small amount of labeled fault data is used for end-to-end fine-tuning, which solves the problem of insufficient generalization ability of the model due to the scarcity of fault samples. This allows the output health index to maintain high accuracy even under the condition of limited fault samples.

[0175] In some embodiments, see Figure 6 As shown, before predicting the remaining service life of each key component in the above-mentioned S104, in specific implementation, it may also include: S601: Obtain the fusion features of each key component at multiple consecutive historical moments to form a fusion feature history sequence; S602: Obtain the operating conditions of each key component at multiple consecutive historical moments to form a historical operating condition sequence; S603: Obtain the vibration feature vectors of each key component at multiple consecutive historical moments to form a historical sequence of vibration feature vectors; Accordingly, the prediction of the remaining service life of each key component in S104 above can, in specific implementation, include: S604: Based on the historical change sequence of the health index, the historical sequence of fused features, the historical operating condition sequence, and the historical sequence of the vibration feature vector, predict the remaining service life of each key component.

[0176] Specifically, the aforementioned historical change sequence of health indices can refer to a sequence formed by arranging the health indices of various key components in chronological order across multiple consecutive historical moments, such as... Where T is the length of the historical time window, which can be 48 (8 hours). The aforementioned fusion feature historical sequence specifically refers to the sequence formed by arranging the fusion features of each key component at multiple consecutive historical moments in chronological order, such as... The aforementioned historical operating condition sequence specifically refers to a sequence formed by arranging the operating conditions of key components at multiple consecutive historical moments in chronological order. These operating conditions can at least include wind speed and active power, such as... The aforementioned historical sequence of vibration characteristic vectors can specifically refer to the sequence formed by arranging the vibration characteristic vectors of each key component at multiple consecutive historical moments in chronological order, such as... .

[0177] The historical change sequences of the above health indices, the historical sequences of fused features, the historical sequences of historical operating conditions, and the historical sequences of vibration feature vectors can be input together into the GRU encoder-decoder, or together into the remaining service life prediction model. The structure of the remaining service life prediction model can include the GRU encoder-decoder. It outputs the remaining service life of each key component, and can also output the 95% confidence interval of the remaining service life of each key component, the predicted degradation trajectory of each key component, etc., which will be explained separately later and will not be elaborated upon here.

[0178] Since the health index indicates "how healthy it is now," but maintenance personnel need to know "how long it can last," the GRU encoder-decoder can be used to read the health index decline trajectory over the past 8 hours to predict the health index change curve for the next 48 hours. The moment when it first falls below the failure threshold is the remaining useful life (RUL). Unlike single-component, single-data-source RUL prediction, this application simultaneously inputs the historical change sequences of the health index, the historical sequence of fused features, the historical sequence of historical operating conditions, and the historical sequence of vibration feature vectors for all the aforementioned key components into the GRU. Vibration features (especially high-frequency energy ratio and sideband amplitude) are direct measures of component damage accumulation. By incorporating them into the RUL prediction network, the GRU can simultaneously utilize "slow degradation trends" (temperature, power) and "fast degradation signals" (vibration shock) for more accurate remaining useful life prediction.

[0179] Based on the above embodiments, by acquiring the fusion feature history sequence, historical operating condition sequence, and vibration feature vector history sequence of each key component, and using them together with the historical change sequence of the health index as input for remaining service life prediction, a multimodal data foundation can be provided for remaining service life prediction. This allows the prediction results to comprehensively reflect the multiple influences of health status change trends, operating condition changes, and vibration status changes, thereby improving the accuracy of remaining service life prediction.

[0180] In some embodiments, see Figure 7 As shown, the prediction of the remaining service life of each key component based on the historical change sequence of the health index, the historical sequence of fused features, the historical operating condition sequence, and the historical sequence of vibration feature vectors in S604 above can, in specific implementation, include: S701: Input the historical change sequence of the health index, the historical sequence of fused features, the historical operating condition sequence, and the historical sequence of vibration feature vectors into the gated loop unit encoder, and compress the input information into a context vector through the gated loop unit encoder; S702: Using the context vector as the initial state, the gated loop unit decoder generates the average predicted health index for each moment within a future preset time window, combined with the predicted working condition information. S703: The average predicted health index is compared with a preset failure threshold at each time step, and the time when the average predicted health index first falls below the preset failure threshold is determined as the remaining service life.

[0181] Specifically, the gated cyclic encoder can refer to a neural network module that compresses a multimodal input sequence into a context vector. Its inputs include a historical sequence of health index changes, a historical sequence of fused features, a historical sequence of operating conditions, and a historical sequence of vibration feature vectors. The output is a context vector. This context vector specifically refers to the compressed representation output by the encoder, containing all historical degradation information of the critical components. The gated cyclic decoder can refer to a neural network module that uses the context vector as its initial state and, combined with predicted operating condition information, autoregressively generates the average predicted health index for the future. The average predicted health index specifically refers to the predicted health index value of the critical components at future moments, output by the decoder. The preset failure threshold specifically refers to the critical health index value used to determine the failure of the critical components.

[0182] In practice, the historical change sequence of the health index, the historical sequence of fused features, the historical operating condition sequence, and the historical sequence of vibration feature vectors are input into the encoder of the gated loop unit. The encoder compresses these four types of information into a context vector. Then, using the context vector as the initial state, the decoder of the gated loop unit combines the predicted operating condition information to generate the predicted average health index for each moment within a preset future time window. Finally, the predicted average health index is compared with a preset failure threshold moment by moment, and the moment when the predicted average health index first falls below the preset failure threshold is determined as the remaining service life.

[0183] Based on the above embodiments, the historical sequence of health index changes, the historical sequence of fused features, the historical sequence of historical operating conditions, and the historical sequence of vibration feature vectors are compressed into a context vector by the gated loop unit encoder. Then, the average value of the future predicted health index is generated by the gated loop unit decoder and compared with the preset failure threshold to determine the remaining service life. This realizes the quantitative prediction of the future health status of key components and provides time dimension information for maintenance decisions.

[0184] In some embodiments, the input of the historical change sequence of the health index, the historical sequence of fused features, the historical operating condition sequence, and the historical sequence of vibration feature vectors into the gated loop unit encoder in step S701 above, and the compression of the input information into a context vector by the gated loop unit encoder, may include the following in specific implementations: The input information is compressed into a context vector using a gated cyclic encoder according to the following formula:

[0185] Among them, c iThe context vector of key component i; GRU enc Encoder for gated loop unit; The historical change sequence of the health index of key component i from time tT to time t; The fusion feature history sequence of key component i from time tT to time t; This is the historical operating condition sequence of key component i from time tT to time t. This represents the historical sequence of vibration characteristic vectors of the key component i from time tT to time t.

[0186] Specifically, tT represents the start time of the historical time window, t is the current time, and T is the length of the historical time window, which can be 48 (corresponding to 8 hours). The above four types of information are encoded using a gated cyclic encoder, compressing the multimodal historical information of the key component i over the past T time points into a context vector c. i This context vector contains all historical degradation information of the critical component i, used for subsequent remaining service life prediction. The inclusion of the historical sequence of vibration feature vectors enables the gated cyclic encoder to capture early damage accumulation information that cannot be reflected by operational data (such as slowly varying signals like temperature and power). Vibration features (especially high-frequency energy ratio and sideband amplitude) are direct measures of component damage accumulation; their inclusion allows the gated cyclic encoder to simultaneously utilize both "slow degradation trends" (such as temperature and power) and "fast degradation signals" (such as vibration and shock) for more accurate remaining service life prediction.

[0187] Based on the above embodiments, the historical sequence of health index changes, the historical sequence of fused features, the historical sequence of historical operating conditions, and the historical sequence of vibration feature vectors are compressed into a context vector by the gated cyclic unit encoder. This context vector comprehensively contains multiple information such as the health status change trend of key components, changes in operating conditions, and changes in vibration status, providing a comprehensive historical degradation information basis for subsequent remaining life prediction.

[0188] In some embodiments, the process in S702 above, which uses the gated recurrent unit decoder with the context vector as the initial state and combines it with the predicted working condition information to generate the average predicted health index at each time point within a future preset time window, may include, in specific implementation: The predicted mean health index is generated using the gated recurrent unit decoder according to the following formula:

[0189] in, Let i be the mean of the predicted health index of the key component i at a future time t+τ; Let GRU be the standard deviation of the predicted health index of critical component i at future time t+τ.dec For gated loop unit decoders; c i is the context vector of the key component i; This refers to the predicted operating conditions at a future time t+τ.

[0190] Specifically, t+τ above represents the τth time step after the current time t, where τ is the prediction time step. The gated recurrent unit decoder uses the context vector c i Given the initial state, combined with the predicted working conditions at future time t+τ The mean of the predicted health index of key component i in the future time t+τ is generated autoregressively. and the standard deviation of the predicted health index The standard deviation of the predicted health index is used to quantify the uncertainty of the prediction.

[0191] Based on the above embodiments, the gated loop unit decoder uses the context vector as the initial state and combines it with future predicted working condition information to generate the mean and standard deviation of the predicted health index at each time within a preset future time window. This achieves quantitative prediction of the future health status of key components. At the same time, the prediction uncertainty is quantified by the prediction standard deviation, providing more comprehensive time dimension information for maintenance decisions.

[0192] In some embodiments, the step S703 above, which compares the average predicted health index with a preset failure threshold time-by-time and determines the moment when the average predicted health index first falls below the preset failure threshold as the remaining service life, may include, in specific implementation: The remaining useful life is determined using the following formula:

[0193] Among them, RUL i The remaining service life of critical component i; Let i be the mean of the predicted health index of the key component i at a future time t+τ; The preset failure threshold for critical component i; T represents the time step forward from the current moment; max Maximum prediction steps (typical value T) max =288, corresponding to 48 hours); min is the function to take the minimum value.

[0194] Specifically, the process for determining the remaining lifespan is as follows: starting from the current time t, the predicted average health index is calculated hour by hour. Compared with the preset failure threshold Compare and find the first one that satisfies the condition. The time step τ is the remaining service life RUL of the critical component i. iAmong them, the preset failure threshold Used to determine whether critical component i has failed. When the average predicted health index is lower than the preset failure threshold for the first time, it indicates that critical component i has reached a failure state at that moment.

[0195] Based on the above embodiments, by comparing the predicted health index average with the preset failure threshold time by time, the moment when the predicted health index average first falls below the preset failure threshold is determined as the remaining service life, thereby realizing the quantitative prediction of the remaining service life of key components and providing clear time dimension information for maintenance decisions.

[0196] In some embodiments, the aforementioned preset failure threshold It can be determined in the following ways: Collect a set of health index calculation values ​​for critical component i at a preset time (e.g., 24 hours) before all historical failures. ; Take the set of health index calculation values The lower quartile is used as the preset failure threshold:

[0197] Among them, Q 25% ( The quantile represents the lower quartile, or 25th percentile. The physical meaning of taking the 25th percentile is that 75% of historical faults were determined to be below the preset failure threshold before they occurred, which can cover most fault scenarios.

[0198] When the number of historical failure samples is insufficient, a conservative estimation method of subtracting a preset multiple of the standard deviation from the mean is used to determine the preset failure threshold.

[0199] in, The mean of the critical health index for failure; This represents the standard deviation of the critical health index for failure. When the number of historical failure samples is less than the preset number (e.g., 10 times), the reliability of the above statistical estimate is insufficient. In this case, a transfer learning strategy (borrowing historical failure statistical thresholds from other wind farms of the same turbine model) or a physical prior based on the bearing life model is used to initialize the threshold, and it is dynamically corrected in subsequent experience playback. If none of the above external data is available, the default preset failure threshold is 0.25.

[0200] The aforementioned method for determining the preset failure threshold abandons arbitrary manual setting and is based on statistical analysis of historical fault data from wind farms, making the preset failure threshold more objective and accurate. The failure threshold is ultimately locked within a physically reasonable range [0.1, 0.4], within which the risks of missed and false alarms can be balanced. In this embodiment, the preset failure threshold for the gearbox can be 0.25, the preset failure threshold for the generator can be 0.28, and the preset failure threshold for the blades can be 0.20.

[0201] In some embodiments, the step S703 above, which compares the average predicted health index with a preset failure threshold time-by-time and determines the moment when the average predicted health index first falls below the preset failure threshold as before the remaining service life, may further include: The gated cyclic unit decoder is subjected to multiple random forward propagations to obtain multiple remaining lifetime predictions. Based on these multiple remaining lifetime predictions, the point estimate and confidence interval of the remaining lifetime are determined.

[0202] Specifically, during inference, Dropout is enabled, and M random forward propagations are performed (e.g., M=50) to obtain M predicted remaining useful lives. The median is taken as the point estimate of the remaining useful life, and a preset multiple of the standard deviation (e.g., 1.96 times the standard deviation) is taken as the 95% confidence interval.

[0203]

[0204] in, Estimate the remaining useful life of critical component i; The remaining lifetime prediction is obtained from the m-th random forward propagation; M is the number of random forward propagations; median( ) is the median function; σ RUL The standard deviation of the predicted remaining useful life; std( ) is the standard deviation function.

[0205] The confidence intervals mentioned above are used to quantify the uncertainty of remaining useful life prediction: the narrower the confidence interval, the more certain the model is about the prediction result; the wider the confidence interval, the more cautious the operations and maintenance personnel should be in making decisions. The MC-Dropout method is used to quantify prediction uncertainty, avoiding giving operations and maintenance personnel an overly absolute and misleading figure.

[0206] In some embodiments, the training process of the gated recurrent unit encoder and the gated recurrent unit decoder described above can be performed according to the following formula:

[0207] Among them, L RULThe training loss is denoted by N, where N is the number of samples in the batch; and T is the training loss of the remaining useful life prediction model. out To predict the window length; This represents the true health index of critical component i at the corresponding moment. The mean of the predicted health index of key component i at the corresponding time. Let be the standard deviation of the predicted health index of critical component i at the corresponding time.

[0208] Specifically, the aforementioned training loss not only penalizes the deviation between the predicted mean and the true value, but also... This penalizes the bias in the uncertainty estimation, preventing the model from reducing the first term's loss by infinitely increasing the variance. Among these, Used to quantify and predict uncertainty.

[0209] Based on the above embodiments, the training loss function not only penalizes the deviation between the predicted mean and the true health index, but also penalizes the deviation of uncertainty estimation, preventing the model from reducing the loss value by infinitely increasing the variance. This enables the model after training convergence to output a reasonable prediction standard deviation while predicting the health index mean, providing a basis for subsequent MC-Dropout uncertainty quantification.

[0210] In some embodiments, prior to S105 above, the following may be included in the specific implementation: Obtain real-time operating conditions, vibration alarm status, maintenance history information, spare parts inventory information, and weather window information; Accordingly, the maintenance actions for each key component determined based on its health index and remaining service life in step S105 above can, in practice, include: The health index, remaining service life, real-time operating conditions, vibration alarm status, maintenance history, spare parts inventory, and weather window information of each key component are input as state vectors into the reinforcement learning network, which outputs maintenance actions. The maintenance actions include at least continued operation, planned maintenance, and emergency maintenance. The reinforcement learning network is optimized based on a reward function during training, which is determined according to a reliability penalty term, an operation penalty term, an over-maintenance penalty term, and a vibration alarm benefit term.

[0211] Specifically, the aforementioned state vector can refer to the input vector of the reinforcement learning network, which includes the health index of each key component, remaining service life, real-time operating conditions, vibration alarm status, maintenance history information, spare parts inventory information, and weather window information. The aforementioned maintenance actions include at least continued operation, planned maintenance, and emergency maintenance, used to guide the maintenance operations of the wind turbine.

[0212] The reinforcement learning network described above can be understood as a decision-making model trained based on reinforcement learning algorithms. Its input is a state vector, and its output is a maintenance action. During training, the reinforcement learning network optimizes based on a reward function, which evaluates the quality of maintenance actions and guides the network to learn the optimal maintenance strategy.

[0213] In practice, the health index, remaining service life, real-time operating conditions, vibration alarm status, maintenance history, spare parts inventory, and weather window information of each key component are input as state vectors into the reinforcement learning network. Based on the current state vector, the reinforcement learning network outputs maintenance actions, which can include at least continue operation, planned maintenance, and emergency maintenance. It can also output maintenance priority ranking and suggested maintenance time windows to further guide the maintenance operations of the wind turbine.

[0214] The aforementioned reinforcement learning network is optimized based on a reward function during training. The goal of the reinforcement learning network is to maximize long-term cumulative rewards. During training, the network outputs maintenance actions based on the current state vector, and the environment returns corresponding reward values ​​based on these actions. The network then updates its policy parameters based on these reward values, gradually learning the optimal maintenance strategy. The reward function is designed to guide the network in balancing multiple objectives, including reliability, economy, and vibration alarm benefits. For example, when the health index falls below a preset failure threshold, the reward function imposes a negative penalty, guiding the network to take maintenance actions before critical components fail. When critical components still have a long remaining service life but emergency maintenance is chosen, the reward function applies additional penalties to prevent over-maintenance.

[0215] Based on the above embodiments, by inputting the health index, remaining service life, real-time operating conditions, vibration alarm status, maintenance history information, spare parts inventory information, and weather window information of each key component as state vectors into the reinforcement learning network, and outputting maintenance actions, the health assessment results and remaining service life prediction results are used together for maintenance decisions, solving the problem of the separation between health assessment and maintenance decisions, and making maintenance decisions more intelligent.

[0216] In some embodiments, the reward function described above is determined based on a reliability penalty term, an operational penalty term, an over-maintenance penalty term, and a vibration alarm benefit term. In specific implementations, it may include: The reward function is determined using the following formula:

[0217] Wherein, R(s) t ,a t ) represents the state vector s t Perform maintenance action a t Total reward; R rel(s t ,a t ) represents the state vector s t Perform maintenance action a t Reliability penalty item; R run (s t ,a t ) represents the state vector s t Perform maintenance action a t Runtime penalty items; R pen (s t ,a t ) represents the state vector s t Perform maintenance action a t Over-maintenance penalty; R vib (s t ,a t ) represents the state vector s t Perform maintenance action a t Vibration alarm revenue items; The reliability penalty term is determined according to the following formula:

[0218] in, The failure severity weight for critical component i; HI i (t) represents the health index of key component i at time t; The preset failure threshold for critical component i; For indicator functions, This indicates that the health index is set to 1 when it is below the preset failure threshold, and 0 otherwise; K represents the total number of critical components. The runtime penalty item is determined according to the following formula:

[0219] in, To maintain action a t Corresponding maintenance costs; To maintain action a t Corresponding losses due to power outages; The over-maintenance penalty is determined according to the following formula:

[0220] in, Penalty coefficient for over-maintenance; RUL represents the average remaining service life of critical component i. thr To preset the maintenance plan's lifespan threshold; Indicates when maintenance action a tThe value is 1 for emergency maintenance and 0 otherwise; K is the total number of critical components. The vibration alarm revenue item is determined according to the following formula: Among them, w vib K represents the vibration alarm benefit coefficient. i (t) represents the kurtosis value of the critical component i; K thr E is the preset kurtosis threshold. ratio,i (t) represents the high-frequency energy ratio; E thr The preset high-frequency energy ratio threshold is used; Indicates when maintenance action a t The value is set to 1 when the kurtosis value of the critical component is greater than the preset kurtosis threshold or the high-frequency energy ratio is greater than the preset high-frequency energy ratio threshold for planned maintenance; otherwise, the value is set to 0.

[0221] Specifically, the aforementioned reliability penalty term guides the reinforcement learning network to take maintenance actions before critical components fail, preventing failures caused by delayed maintenance. When the health index is below a preset failure threshold, the reliability penalty term is negative, and the closer the health index is to zero (i.e., the greater the difference from the preset failure threshold), the heavier the penalty. This penalty term guides the reinforcement learning network to take timely maintenance actions when critical components are close to failure, thus preventing failures.

[0222] The aforementioned penalty term is used to guide the reinforcement learning network to consider the continuity of unit operation during maintenance actions, thereby reducing power generation losses caused by downtime. This penalty term guides the reinforcement learning network to select maintenance actions with shorter downtime as much as possible, thus minimizing power generation losses.

[0223] The aforementioned over-maintenance penalty is designed to prevent reinforcement learning networks from taking emergency maintenance actions when critical components still have a long remaining lifespan, thus avoiding over-repair. When the remaining lifespan of a critical component exceeds the preset maintenance plan lifespan threshold and emergency maintenance is performed, the over-maintenance penalty is negative, and the larger the remaining lifespan exceeds the preset maintenance plan lifespan threshold, the heavier the penalty.

[0224] The aforementioned vibration alarm reward is used to guide the reinforcement learning network to proactively take planned maintenance when vibration is abnormal, thus achieving a closed loop of early warning and timely maintenance. When the vibration feature vector of a critical component exceeds the preset vibration alarm threshold and planned maintenance is executed, the vibration alarm reward is positive, providing an additional positive reward.

[0225] Based on the above embodiments, the design of the above reward function guides the reinforcement learning network to learn the optimal maintenance strategy, balancing multiple objectives such as reliability, economy, and vibration alarm benefits, so that the output maintenance actions can guide the actual maintenance operation of wind turbine units and improve the intelligence level of maintenance decision-making.

[0226] In some embodiments, prior to S105 above, the following may be included in the specific implementation: Real-time monitoring of vibration characteristic vectors of key components; When the vibration characteristic vector of any critical component exceeds the preset vibration alarm threshold, the maintenance action of the critical component will be forcibly identified as planned maintenance.

[0227] Specifically, the aforementioned preset vibration alarm threshold can refer to a critical value used to determine whether a key component has experienced abnormal vibration, and may include a preset kurtosis threshold and a preset high-frequency energy ratio threshold. The aforementioned vibration feature vector may include at least a kurtosis value and a high-frequency energy ratio.

[0228] In practice, the vibration characteristic vectors of each key component are monitored in real time. When the kurtosis value of any key component exceeds a preset kurtosis threshold (e.g., 5.0) or the high-frequency energy ratio exceeds a preset high-frequency energy ratio threshold (e.g., 0.3), the maintenance action for that key component is forcibly defined as planned maintenance.

[0229] Among them, K i (t) represents the kurtosis value of the critical component i at time t; K thr Set a preset kurtosis threshold (e.g., 5.0); E ratio,i (t) represents the high-frequency energy ratio of key component i at time t; E thr To preset the high-frequency energy ratio threshold (e.g., 0.3); a i =1 indicates that the maintenance action of critical component i is determined as planned maintenance; OR is a logical OR operation.

[0230] The aforementioned vibration alarm forced triggering logic ensures that even if the health index has not fallen below the preset failure threshold, abnormal vibration can trigger early inspection and maintenance, thus achieving a closed loop of early warning and timely maintenance.

[0231] The aforementioned vibration alarm forced triggering logic serves as a supplementary mechanism to the reward function. It ensures that even if the reinforcement learning network has not yet learned to proactively implement planned maintenance during vibration anomalies, the forced triggering logic guarantees timely repairs in such scenarios. This logic, together with the vibration alarm reward in the reward function, provides a double guarantee: the reward function guides the reinforcement learning network to proactively learn planned maintenance strategies during vibration anomalies, while the forced triggering logic acts as a fallback mechanism to ensure that planned maintenance is inevitably triggered in vibration anomaly scenarios. Through this dual guarantee mechanism, this application can effectively trigger planned maintenance when vibration anomalies occur, preventing delays in repair due to the health index not yet falling below the failure threshold, thereby reducing the risk of failure.

[0232] In some embodiments, the reinforcement learning network described above can be trained using a proximal policy optimization algorithm. The reinforcement learning network may include a policy network and a value network: Policy Network (Actor Network): Input layer (dimension Ds) → Fully connected 256+ReLU+BN → Fully connected 128+ReLU → Fully connected 64+ReLU → Output layer (dimension Da=3×K, Softmax activation), outputting the probability distribution of each action. .

[0233] Value Network (Critic Network): Input layer (dimension Ds) → Fully connected 256+ReLU+BN → Fully connected 128+ReLU → Fully connected 64+ReLU → Output layer (dimension 1, linear activation), output state value .

[0234] Specifically, the policy network described above is used to determine the current state vector s. t Output the probability distribution of each maintenance action; the value network is used to estimate the current state vector s. t The state value. Training hyperparameters may include: the optimizer is Adam, and the learning rate lr of the policy network. actor =3×10 -4 The learning rate lr of the value network critic =1×10 -3 Batch size is 64; Dropout (scale 0.1) is used between hidden layers to prevent overfitting.

[0235] In some embodiments, the training process of the reinforcement learning network described above can be performed according to the following steps: (1) Calculate the advantage function using generalized advantage estimation:

[0236] in, The value of the advantage function at time t is used to measure the advantage in state s. t Take action a t How much better than the average; γ is the discount factor, with a value of 0.95, used to balance the importance of current rewards and future rewards; λ is the generalized advantage estimation parameter, with a value of 0.95, used to balance bias and variance; l is the cumulative index, indicating the l-th step from the current time t; δ t+l The time series difference error at time t+l; δ t Let r be the time series difference error at time t; t V(s) is the reward value at time t. t ) represents state s t Value estimation; V(s) t+1 ) represents the state vector s at the next time step. t+1Value estimate.

[0237] (2) Update the step size by limiting the strategy through the pruning objective function:

[0238] in, Let θ be the shearing objective function; θ be the parameters of the policy network; E t [ [This represents the expectation at time t;] The probability ratio between the old and new strategies; For the current policy network in state vector s t Output maintenance action a t The probability of clip ( ) is the shearing function, which will... Limited to [1- ,1+ Within the interval; This is the shearing parameter, with a value of 0.2, used to limit the step size of policy updates.

[0239] (3) Calculate the value loss:

[0240] in, For the loss of value; These are the parameters of the value network; The state vector s output by the value network t Value estimation; E represents the estimated cumulative reward at time t. t [ [ ] represents the expectation at time t.

[0241] (4) Calculate the total loss:

[0242] in, Total loss; These are the parameters of the policy network; These are the parameters of the value network; The objective function for shearing; c1 is the value loss coefficient, with a value of 0.5; c2 is the entropy regularization coefficient, with a value of 0.01. E is an entropy regularization term used to encourage exploration of the policy network. t [ [ ] represents the expectation at time t.

[0243] In some embodiments, see Figure 8 As shown, after S105 above, in specific implementations, it may also include: S801: Collect equipment status information before maintenance, the type of action performed, equipment status information after maintenance, actual safe operating time of the equipment after maintenance, and vibration feature vectors before and after maintenance to construct experience samples; S802: Store the experience sample into the experience playback pool; S803: At preset time intervals, experience samples are sampled from the experience replay pool for incremental fine-tuning to update the network parameters of the graph attention network and / or the network parameters of the reinforcement learning network. The updated parameters are used for subsequent health assessment, remaining service life prediction and maintenance action determination.

[0244] Specifically, the aforementioned experience samples refer to samples encapsulated after each maintenance, including information such as the state before and after maintenance, the type of action performed, and the actual effect data after maintenance. These samples are used for continuous learning and updating of the model. Experience samples can be represented as:

[0245] Among them, s pre The complete status information before maintenance (which may include the health index of each key component, remaining service life, etc.); 'a' is the type of action to be performed (which may include continue operation, planned maintenance, or emergency maintenance); 's' post The equipment status information after maintenance; R represents the actual reward information for this maintenance (which may include actual maintenance costs and power generation losses due to downtime); ΔRUL actual The actual and safe operating time of the equipment after maintenance serves as a feedback label between the model and the physical world, used to verify and correct the accuracy of the remaining service life prediction; K pre and K post The vibration characteristic vectors before and after maintenance (which may include kurtosis values, high-frequency energy ratios, etc.) are used to evaluate the actual contribution of vibration signals to maintenance decisions.

[0246] The aforementioned experience replay pool specifically refers to a buffer used to store experience samples, with a capacity of M=2000. Every preset time interval (e.g., 30 days), a batch of experience samples is sampled from the experience replay pool for incremental fine-tuning to update the network parameters of the graph attention network and / or the reinforcement learning network. The updated parameters are used for subsequent health assessments, remaining lifespan predictions, and determination of maintenance actions.

[0247] During incremental fine-tuning, an elastic weight consolidation mechanism is employed to prevent catastrophic forgetting:

[0248] Among them, L EWC To consolidate the total loss for elastic weights; L task Loss due to new tasks; λ EWCTo reinforce the regularization coefficients for elastic weights; F i The i-th diagonal element of the Fisher information matrix is ​​used to measure the parameter θ. i The importance of the old task; The optimal parameter values ​​for the old task; θ i This is the current parameter value.

[0249] Through the aforementioned elastic weight consolidation mechanism, a secondary penalty is applied to important parameters to prevent them from deviating significantly from their optimal values ​​in the old tasks, thus enabling the absorption of new knowledge without forgetting past experiences. This forms a complete closed loop of evaluation, prediction, decision-making, execution, feedback, and re-evaluation, allowing the model to continuously evolve.

[0250] Based on the above embodiments, experience samples are constructed by collecting state information before and after maintenance, the types of actions performed, the actual safe operating time of the equipment after maintenance, and vibration feature vectors before and after maintenance. These samples are stored in an experience playback pool. Experience samples are sampled from the experience playback pool at preset time intervals for incremental fine-tuning to update the network parameters of the graph attention network and / or the network parameters of the reinforcement learning network. This enables the model to continuously learn and evolve from operation and maintenance practices, forming a complete closed loop of evaluation, prediction, decision-making, execution, feedback, and re-evaluation, thereby improving the accuracy of subsequent health assessments, remaining service life predictions, and maintenance action determination.

[0251] Based on the above embodiments, this application can achieve the following beneficial effects: (1) Dynamic multi-component coupling graph modeling: By constructing a component relationship graph and dynamically adjusting the base edge weights based on wind speed, active power and vibration data, the graph structure changes in real time with the actual operating state, accurately depicting the degradation propagation relationship between components.

[0252] (2) Multimodal fusion perception of operation data and vibration data: The operation data and vibration data are weighted and fused through graph attention network so that the health index of each key component can comprehensively reflect its own status and the degradation impact of other key components coupled with it. It can capture long-term degradation trends and respond to instantaneous impact anomalies.

[0253] (3) Temporal comparison learning pre-training: The graph attention network is pre-trained using a large amount of unlabeled data, and then fine-tuned using a small amount of labeled fault data, which solves the problem of insufficient generalization ability of the model due to the scarcity of fault samples.

[0254] (4) Multimodal remaining life interval prediction: By fusing the historical change sequence of health index, the historical sequence of fusion features, the historical operating condition sequence and the historical sequence of vibration feature vector through the gated cyclic unit encoder-decoder, the point estimate and confidence interval of the remaining life are output to quantify the prediction uncertainty.

[0255] (5) PPO multi-objective maintenance decision and continuous learning closed loop: By outputting maintenance actions through reinforcement learning network, multiple objectives such as reliability, economy and vibration alarm benefits are balanced; through the construction of experience samples and elastic weight consolidation mechanism, the model can continuously learn and evolve from operation and maintenance practice.

[0256] (6) Low data dependency: It can be achieved with only running data, 3 vibration measurement points and maintenance / fault records, without the need for additional hardware, thus reducing deployment costs.

[0257] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. For details, please refer to the foregoing descriptions of the relevant processing embodiments; they will not be elaborated upon here.

[0258] In a specific scenario example, see Figure 9 As shown, the overall process of the wind turbine health assessment and maintenance method based on graph attention networks provided in this application may include: Data acquisition and preprocessing stage: Operating data such as wind speed, power, and temperature are collected through the SCADA system, and vibration data of key components (3-6 measuring points × 8 features) are collected through vibration sensors. Historical maintenance / fault records are also acquired. The above data undergoes data cleaning, alignment, and standardization to output standardized multimodal features.

[0259] Multi-component degradation coupling graph construction stage: Define component nodes (K=5, i.e., blades, main shaft bearings, gearboxes, generators, and converters), and construct the transmission chain topology with physical base edges. Calculate the load influence factor and vibration anomaly propagation coefficient, and generate a dynamic adjacency matrix through joint influence forms to achieve dynamic adjustment of the degradation influence intensity between components.

[0260] The GATv2 multimodal health assessment stage uses standardized multimodal features as node inputs, avoids gradient saturation through log(1+w) addition attention, corrects kurtosis abrupt signals through vibration correction factors, combines time-series contrastive learning pre-training (time-stationary prior), and outputs the health index of each component and the overall health index of the whole machine.

[0261] Multimodal remaining life prediction stage: Historical health index sequences, multimodal node features, operating condition sequences, and vibration feature sequences are input into a GRU encoder and compressed into a context vector; future degradation trajectories are generated through autoregression using a GRU decoder. MC-Dropout is used for 50 samplings, combined with the failure threshold quantile method (Q... 25% Historical fault data), output RUL point estimates and 95% confidence intervals.

[0262] The PPO multi-objective decision-making and continuous learning phase constructs a state space using health indices, remaining lifespan, and vibration alarm status. An Actor-Critic network (Actor: 256-128-64, Critic: 256-128-64) outputs maintenance commands (0-Continue Operation / 1-Planned Maintenance / 2-Emergency Maintenance). Simultaneously, a forced vibration alarm trigger logic (kurtosis > 5.0 or high-frequency energy ratio > 0.3) ensures early maintenance in case of vibration anomalies. The reward function integrates four objectives: reliability, economy, over-maintenance penalty, and vibration alarm benefit.

[0263] Maintenance and Feedback Phase: Maintenance operations are performed and their effects recorded; the experience playback pool includes before-and-after comparisons of vibration characteristics. At preset time intervals, the model parameters are continuously fine-tuned through EWC learning. The final output includes a health score sheet, remaining life prediction, and maintenance work order recommendations.

[0264] The following is an illustration using a specific example: (1) Wind field overview and data A wind farm in operation has 25 2MW doubly-fed induction generator units. The transmission chain structure is as follows: blades → main shaft bearings → gearbox (two-stage planetary gearbox + one-stage parallel shaft gearbox) → generator → converter. Two years of SCADA data were collected (10 minutes / time) + 3 vibration measurement points (gearbox input shaft, output shaft, and generator drive end, 8 feature values ​​were extracted every 10 minutes).

[0265] (2) Model configuration

[0266] K = 5 component nodes (blade, main shaft bearing, gearbox, generator, converter), d′ = 64, GATv2 layer number L = 2, historical window T = 48 (8 hours), prediction window T out =96 (16 hours). Failure threshold HI thr Set the kurtosis threshold K to 0.25 for the gearbox, 0.28 for the generator, and 0.20 for the blades. thr =5.0. Training strategy: 50 rounds of comparative learning pre-training → 200 rounds of end-to-end joint training.

[0267] (3) Vibration data sources

[0268] The vibration sensor is an IEPE accelerometer with a sampling frequency of 25.6kHz and a sampling time of 10 seconds. The original waveform is not transmitted; only eight characteristic values ​​(peak value, RMS, kurtosis, waveform factor, rotational frequency amplitude, meshing sideband energy, bearing characteristic frequency amplitude, and high-frequency energy ratio) are uploaded to the model. The compression ratio of a single data transmission exceeds 3000:1.

[0269] (4) Health assessment performance

[0270] After adding vibration data, the early warning lead time increased from 18.6 hours to 35.0 hours, and the MSE decreased by 25%.

[0271] (5) Remaining lifetime prediction performance

[0272] Although this specification provides the following examples or appendices Figure 10 The method or apparatus structure shown may include more or fewer combined operational steps or module units based on conventional or non-inventive methods. In steps or structures where there is no logically necessary causal relationship, the execution order of these steps or the module structure of the apparatus is not limited to the execution order or module structure shown in the embodiments or drawings of this specification. When the method or module structure is applied in actual devices, servers, or terminal products, it can be executed sequentially or in parallel according to the method or module structure shown in the embodiments or drawings (e.g., in a parallel processor or multi-threaded processing environment, or even a distributed processing or server cluster implementation environment). Based on the above-described wind turbine health assessment and maintenance method based on graph attention networks, this specification also proposes an embodiment of a wind turbine health assessment and maintenance device based on graph attention networks. Figure 10 As shown, the device may specifically include the following modules: The acquisition module 1001 can be used to acquire the operating data and vibration data of each key component in the wind turbine, wherein the operating data includes at least wind speed and active power. The component relationship graph construction module 1002 can be used to construct a component relationship graph with each key component as a graph node and the coupling relationship between key components as a base edge. It can determine the load transfer coefficient based on the wind speed and the active power, determine the anomaly propagation coefficient based on the degree to which the vibration data of the target key component deviates from its healthy reference vibration characteristic vector, and generate a dynamic adjacency matrix representing the intensity of degradation influence between key components based on the load transfer coefficient, the anomaly propagation coefficient and the component relationship graph. The health index output module 1003 can be used to input the operation data, the vibration data and the dynamic adjacency matrix into the graph attention network. The graph attention network is used to perform weighted fusion of the operation data and the vibration data based on the dynamic adjacency matrix and output the health index of each key component. The remaining service life prediction module 1004 can be used to predict the remaining service life of each key component based on the historical change sequence of the health index at multiple consecutive historical moments. The maintenance decision module 1005 can be used to determine the maintenance actions of each key component based on the health index and the remaining service life, and the maintenance actions are used to guide the maintenance operations of the wind turbine.

[0273] The descriptions and functions of the above modules can be found in the section on "A Method for Health Assessment and Maintenance of Wind Turbine Units Based on Graph Attention Networks," and will not be repeated here.

[0274] This specification also provides an electronic device based on the above-described method for wind turbine health assessment and maintenance using a graph attention network. The device includes a processor and a memory for storing executable programs / instructions. Specifically, the processor can execute the following steps according to the program / instructions: acquiring operating data and vibration data of each key component in the wind turbine, wherein the operating data includes at least wind speed and active power; constructing a component relationship graph using each key component as a graph node and the coupling relationship between key components as a base edge; determining the load transfer coefficient based on the wind speed and the active power; and determining the abnormal propagation coefficient based on the degree to which the vibration data of the target key component deviates from its health baseline vibration characteristic vector. Based on the load transfer coefficient, the anomaly propagation coefficient, and the component relationship diagram, a dynamic adjacency matrix representing the intensity of degradation impact among key components is generated. The operating data, vibration data, and the dynamic adjacency matrix are input into a graph attention network, which performs weighted fusion of the operating data and vibration data based on the dynamic adjacency matrix to output a health index for each key component. Based on the historical change sequence of the health index over multiple consecutive historical moments, the remaining service life of each key component is predicted. Based on the health index and the remaining service life, maintenance actions for each key component are determined, and these maintenance actions guide the maintenance operations of the wind turbine.

[0275] To execute the above instructions more accurately, please refer to... Figure 11 As shown in the embodiments of this specification, another specific electronic device is also provided, wherein the electronic device includes a network communication port 1101, a processor 1102 and a memory 1103, and the above structures are connected by internal cables so that the various structures can perform specific data interaction.

[0276] Specifically, the network communication port 1101 can be used to acquire the operating data and vibration data of each key component in the wind turbine, and the operating data includes at least wind speed and active power. The processor 1102 can be specifically used to construct a component relationship graph with each key component as a graph node and the coupling relationship between key components as a base edge; determine the load transfer coefficient based on the wind speed and the active power; determine the anomaly propagation coefficient based on the degree to which the vibration data of the target key component deviates from its healthy baseline vibration characteristic vector; generate a dynamic adjacency matrix representing the intensity of degradation influence between key components based on the load transfer coefficient, the anomaly propagation coefficient, and the component relationship graph; input the operating data, the vibration data, and the dynamic adjacency matrix into a graph attention network, which is used to perform weighted fusion of the operating data and the vibration data based on the dynamic adjacency matrix to output the health index of each key component; predict the remaining service life of each key component based on the historical change sequence of the health index at multiple consecutive historical moments; and determine the maintenance actions of each key component based on the health index and the remaining service life, which are used to guide the maintenance operations of the wind turbine.

[0277] The memory 1103 can be used to store the corresponding instruction program.

[0278] In this embodiment, the network communication port 1101 can be a virtual port bound to different communication protocols, thereby enabling the sending or receiving of different data. For example, the network communication port can be a port responsible for web data communication, a port responsible for FTP data communication, or a port responsible for email data communication. Furthermore, the network communication port can also be a physical communication interface or communication chip. For example, it can be a wireless mobile network communication chip, such as GSM or CDMA; it can also be a Wi-Fi chip; or it can be a Bluetooth chip.

[0279] In this embodiment, the processor 1102 can be implemented in any suitable manner. For example, the processor can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers, etc. This specification is not limiting.

[0280] In this embodiment, the memory 1103 may include multiple layers. In a digital system, anything that can store binary data can be a memory. In an integrated circuit, a circuit with storage function but no physical form is also called a memory, such as RAM, FIFO, etc. In a system, a storage device with a physical form is also called a memory, such as a memory stick, TF card, etc.

[0281] This specification also provides a computer storage medium based on the above-described method for wind turbine health assessment and maintenance using a graph attention network. The computer storage medium stores a computer program / instruction that, when executed, performs the following: acquiring operational and vibration data of key components in the wind turbine, the operational data including at least wind speed and active power; constructing a component relationship graph using each key component as a graph node and the coupling relationship between key components as a base edge; determining the load transfer coefficient based on the wind speed and active power; determining the anomaly propagation coefficient based on the degree to which the vibration data of the target key component deviates from its health baseline vibration characteristic vector; and, according to... The load transfer coefficient, the anomaly propagation coefficient, and the component relationship diagram are used to generate a dynamic adjacency matrix representing the intensity of degradation effects among key components. The operating data, the vibration data, and the dynamic adjacency matrix are input into a graph attention network, which performs weighted fusion of the operating data and the vibration data based on the dynamic adjacency matrix to output a health index for each key component. Based on the historical change sequence of the health index over multiple consecutive historical moments, the remaining service life of each key component is predicted. Based on the health index and the remaining service life, maintenance actions for each key component are determined, and these maintenance actions guide the maintenance operations of the wind turbine.

[0282] In this embodiment, the storage medium includes, but is not limited to, Random Access Memory (RAM), Read-Only Memory (ROM), Cache, Hard Disk Drive (HDD), or Memory Card. The memory can be used to store computer program instructions. The network communication unit can be an interface configured according to standards specified in the communication protocol for network connection communication.

[0283] In this embodiment, the specific functions and effects implemented by the program instructions stored in the computer storage medium can be explained in comparison with other implementation methods, and will not be repeated here.

[0284] This specification also provides a computer program product based on the above-described method for wind turbine health assessment and maintenance using a graph attention network. The product includes a non-transitory computer-readable storage medium storing computer programs / instructions. These computer programs / instructions are operable to cause the computer to perform the following steps: acquiring operational and vibration data of key components in the wind turbine, wherein the operational data includes at least wind speed and active power; constructing a component relationship graph using each key component as a graph node and the coupling relationships between key components as base edges; determining the load transfer coefficient based on the wind speed and the active power; and determining abnormal transmission based on the degree to which the vibration data of the target key component deviates from its health baseline vibration characteristic vector. Based on the load transfer coefficient, the anomaly propagation coefficient, and the component relationship diagram, a dynamic adjacency matrix representing the intensity of degradation impact among key components is generated. The operating data, vibration data, and the dynamic adjacency matrix are input into a graph attention network, which performs weighted fusion of the operating data and vibration data based on the dynamic adjacency matrix to output a health index for each key component. Based on the historical change sequence of the health index over multiple consecutive historical moments, the remaining service life of each key component is predicted. Based on the health index and the remaining service life, maintenance actions for each key component are determined, and these maintenance actions guide the maintenance operations of the wind turbine.

[0285] While this specification provides the steps of operation for the methods described in the embodiments or flowcharts, more or fewer steps may be included based on conventional or non-inventive means. The order of steps listed in the embodiments is merely one possible order of execution among many steps and does not represent the only possible order. In actual device or client product execution, the methods shown in the embodiments or drawings may be executed sequentially or in parallel (e.g., in a parallel processor or multi-threaded processing environment, or even a distributed data processing environment). The terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, product, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, product, or apparatus. Without further limitations, the presence of other identical or equivalent elements in a process, method, product, or apparatus that includes said elements is not excluded. The terms "first," "second," etc., are used to denote names and do not indicate any particular order.

[0286] Those skilled in the art will also know that, besides implementing the controller using purely computer-readable program code, the same functions can be achieved by logically programming the method steps, making the controller function as logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers (PLCs), and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the devices within it used to implement various functions can also be considered structures within that hardware component. Alternatively, the devices used to implement various functions can be considered as both software modules implementing the method and structures within a hardware component.

[0287] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, classes, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0288] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this specification can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions of this specification can essentially be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, mobile terminal, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments of this specification.

[0289] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. This specification can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0290] Although this specification has been described by way of examples, those skilled in the art will recognize that many variations of this specification are possible without departing from its spirit.

Claims

1. A method for health assessment and maintenance of wind turbine units based on graph attention networks, characterized in that, include: Obtain operating data and vibration data of key components in the wind turbine, wherein the operating data includes at least wind speed and active power; A component relationship graph is constructed using each key component as a graph node and the coupling relationship between key components as a base edge. The load transfer coefficient is determined based on the wind speed and the active power. The anomaly propagation coefficient is determined based on the degree to which the vibration data of the target key component deviates from its healthy reference vibration characteristic vector. Based on the load transfer coefficient, the anomaly propagation coefficient and the component relationship graph, a dynamic adjacency matrix representing the intensity of degradation influence between key components is generated. The operating data, the vibration data, and the dynamic adjacency matrix are input into a graph attention network. The graph attention network is used to perform weighted fusion of the operating data and the vibration data based on the dynamic adjacency matrix and output the health index of each key component. Based on the historical change sequence of the health index at multiple consecutive historical moments, the remaining service life of each key component is predicted. Based on the health index and the remaining service life, maintenance actions for each key component are determined, and these maintenance actions are used to guide the maintenance operations of the wind turbine.

2. The method according to claim 1, characterized in that, The key components include at least blades, main shaft bearings, gearboxes, generators, and converters; the coupling relationships between the key components include at least physical transmission connections; the target key component is a transmitting end component for abnormal propagation, used to propagate vibration abnormality information to other key components coupled with it, and the other key components are receiving end components for abnormal propagation.

3. The method according to claim 1, characterized in that, The determination of the load transfer coefficient based on the wind speed and the active power includes: The load transfer factor is determined using the following formula: Among them, F 载荷 (t) is the load transfer coefficient at time t; The load modulation intensity coefficient; The function is the Sigmoid function; v(t) is the wind speed at time t; v rated P is the rated wind speed; P(t) is the active power at time t; P rated This refers to the rated active power.

4. The method according to claim 2, characterized in that, The vibration data includes at least a vibration feature vector; correspondingly, the degree to which the vibration data based on the target key component deviates from its healthy baseline vibration feature vector determines the anomaly propagation coefficient, including: The anomaly propagation coefficient is determined using the following formula: in, Let be the intensity of the vibration anomaly information propagated from the transmitting component q to the receiving component p at time t, and let be the anomaly propagation coefficient at time t; exp is an exponential function with the natural constant as its base; K q (t) is the vibration characteristic vector of the transmitting component q at time t; The health reference vibration characteristic vector of the transmitting component q; For bandwidth parameters; is the vector norm between the vibration characteristic vector and the health reference vibration characteristic vector, used to characterize the degree to which the vibration characteristic vector of the transmitting component q deviates from its health reference vibration characteristic vector.

5. The method according to claim 1, characterized in that, The step of generating a dynamic adjacency matrix representing the intensity of degradation impact among key components based on the load transfer coefficient, the anomaly propagation coefficient, and the component relationship diagram includes: A load transfer matrix is ​​constructed based on the load transfer coefficient, and a vibration anomaly propagation matrix is ​​constructed based on the anomaly propagation coefficient. Based on the weights of each base edge in the component relationship diagram, construct the physical connection base matrix; The dynamic adjacency matrix is ​​generated by performing element-wise multiplication on the physical connection basis matrix, the load transfer matrix, and the anomaly propagation matrix.

6. The method according to claim 5, characterized in that, The step of performing element-wise multiplication of the physical connection basis matrix, the load transfer matrix, and the anomaly propagation matrix to generate the dynamic adjacency matrix includes: The dynamic adjacency matrix is ​​generated according to the following formula: Where W(t) is the dynamic adjacency matrix at time t; W base For the physical connection basis matrix; For element-wise multiplication; f 载荷 (t) is the load transfer matrix at time t; W vib (t) is the vibration anomaly propagation matrix at time t.

7. The method according to claim 1, characterized in that, The graph attention network is used to weight and fuse the operational data and the vibration data based on the dynamic adjacency matrix, and output the health index of each key component, including: The operational and vibration data of each key component are stitched together and linearly projected to obtain the initial nodal features of each key component. For each base edge in the component relationship graph, the two key components corresponding to the base edge are each other's neighbor key components. Based on the initial node features of each key component, the initial node features of its neighbor key components, and the weight of the corresponding base edge in the dynamic adjacency matrix, the attention score of each key component to its neighbor key components is calculated. The attention scores are normalized to obtain the normalized attention weights of each key component to its neighboring key components. According to the normalized attention weights, the initial node features of the neighboring key components of each key component are weighted and aggregated to obtain aggregated features. The aggregated features are then residually connected with the transformed features of each key component after residual transformation to obtain the fused features of each key component. The graph attention network contains at least two graph attention layers. The fused features of each key component are used as the input of the next graph attention layer. The operations of attention score calculation, normalization processing, weighted aggregation and residual connection are repeatedly performed until the graph attention calculation of the preset number of layers is completed, and the fused features of each key component in the last layer are obtained. The fusion features of each key component in the last layer are mapped to the range of health index values ​​through a fully connected layer and an activation function, and the health index of each key component is output.

8. The method according to claim 7, characterized in that, The calculation of the attention score of each critical component to its neighboring critical components includes: For each key component and its neighboring key components, the initial node features of the key component and the initial node features of its neighboring key components are linearly transformed to obtain the query features of the key component and the key features of its neighboring key components. The query features, key features, and physical bias components are added together, and after passing through a nonlinear activation function, an inner product operation is performed with the attention vector to obtain the attention score of the key component to its neighboring key components; wherein, the physical bias component is determined according to the weight of the connection relationship between the key component and its neighboring key components in the dynamic adjacency matrix.

9. The method according to claim 8, characterized in that, The step of adding the query features, the key features, and the physical bias component, applying the result to a non-linear activation function, and then performing an inner product operation with the attention vector to obtain the attention score of the key component to its neighboring key components includes: Calculate the attention score using the following formula: in, e ij Let i be the attention score of critical component i to its neighboring critical component j; W is the transpose of the attention vector; LeakyReLU is the non-linear activation function; query To query the transformation matrix, used for the initial node features of key component i. Perform a linear transformation to obtain the query features of key component i; W key Here is the key transformation matrix, used to initialize the node features of neighboring key component j. Perform a linear transformation to obtain the key features of the neighboring key component j; For query features; Key features; w ij (t) represents the weight of the base edge between key component i and its neighbor key component j in the dynamic adjacency matrix at time t; This is the physical bias component.

10. The method according to claim 7, characterized in that, The process involves weighted aggregation of the initial node features of neighboring key components according to the normalized attention weights to obtain aggregated features, and residual connection of the aggregated features with the transformed features of each key component after residual transformation to obtain the fused features of each key component, including: Weighted aggregation and residual join are performed according to the following formula: in, The fused features of key component i after processing by a graph attention network layer; The initial node features of key component i; The initial node features of the neighboring key component j; W represents the normalized attention weight of critical component i to its neighboring critical component j. res W is the residual transformation matrix; v Value transformation matrix; For transformation features; is the aggregation feature; N(i) is the set of neighboring key components of key component i; ELU is the activation function.

11. The method according to claim 7, characterized in that, The process of mapping the fused features of each key component at the last layer to the health index value range via a fully connected layer and activation function, and outputting the health index of each key component, includes: Output the health index of each key component according to the following formula: Among them, HI i (t) represents the health index of key component i at time t, and the health index ranges from [0,1]. The fusion features of key component i in the output of the Lth layer of the graph attention network; W HI b is the weight matrix of the fully connected layer; HI is the bias term for the fully connected layer; Sigmoid is the activation function.

12. The method according to claim 11, characterized in that, The method further includes: Obtain the current kurtosis value of each key component, which is extracted from the vibration data; When the current kurtosis value is greater than the healthy baseline kurtosis value, the vibration correction amount is determined based on the deviation between the current kurtosis value and the healthy baseline kurtosis value; The vibration correction amount is subtracted from the health index to obtain the corrected health index.

13. The method according to claim 12, characterized in that, The step of determining the vibration correction amount based on the deviation between the current kurtosis value and the healthy baseline kurtosis value includes: The vibration correction amount is determined using the following formula: in, This represents the vibration correction amount for the critical component i at time t. This is the maximum correction amplitude control parameter, used to limit the upper limit of the vibration correction amount; This is the kurtosis scaling factor, used to control how quickly the kurtosis reaches saturation; Let i be the current kurtosis value of the critical component i at time t; denoted as the health baseline kurtosis value of critical component i; tanh is the hyperbolic tangent function; max is the function for maximizing the value.

14. The method according to claim 1, characterized in that, The method further includes: The fusion features of each key component at multiple consecutive historical moments are obtained to form a historical sequence of fusion features; The operating conditions of each key component at multiple consecutive historical moments are obtained to form a historical operating condition sequence; The vibration feature vectors of each key component at multiple consecutive historical moments are obtained to form a historical sequence of vibration feature vectors; Accordingly, the prediction of the remaining service life of each key component includes: Based on the historical change sequence of the health index, the historical sequence of fused features, the historical operating condition sequence, and the historical sequence of vibration feature vectors, the remaining service life of each key component is predicted.

15. The method according to claim 14, characterized in that, The prediction of the remaining service life of each key component based on the historical change sequence of the health index, the historical sequence of fused features, the historical operating condition sequence, and the historical sequence of vibration feature vectors includes: The historical change sequence of the health index, the historical sequence of fused features, the historical operating condition sequence, and the historical sequence of vibration feature vectors are input into the gated loop unit encoder, and the input information is compressed into a context vector by the gated loop unit encoder. Using the context vector as the initial state, the gated loop unit decoder combines the predicted working condition information to generate the average predicted health index for each moment within a future preset time window. The predicted health index average is compared with a preset failure threshold at each time step, and the time when the predicted health index average first falls below the preset failure threshold is determined as the remaining service life.

16. The method according to claim 15, characterized in that, The process involves inputting the historical change sequence of the health index, the historical sequence of fused features, the historical operating condition sequence, and the historical sequence of vibration feature vectors into a gated cyclic encoder. The gated cyclic encoder then compresses the input information into a context vector, including: The input information is compressed into a context vector using a gated cyclic encoder according to the following formula: Among them, c i The context vector of key component i; GRU enc Encoder for gated loop unit; The historical change sequence of the health index of key component i from time tT to time t; The fusion feature history sequence of key component i from time tT to time t; This is the historical operating condition sequence of key component i from time tT to time t. This represents the historical sequence of vibration characteristic vectors of the key component i from time tT to time t.

17. The method according to claim 15, characterized in that, The step of generating the average predicted health index for each moment within a future preset time window by using the context vector as the initial state through a gated recurrent unit decoder and combining it with predicted working condition information includes: The predicted mean health index is generated using the gated recurrent unit decoder according to the following formula: in, Let i be the mean of the predicted health index of the key component i at a future time t+τ; Let GRU be the standard deviation of the predicted health index of critical component i at future time t+τ. dec For gated loop unit decoders; c i is the context vector of the key component i; This refers to the predicted operating conditions at a future time t+τ.

18. The method according to claim 15, characterized in that, The step of comparing the predicted average health index with a preset failure threshold time-by-time, and determining the moment when the predicted average health index first falls below the preset failure threshold as the remaining service life, includes: The remaining useful life is determined using the following formula: Among them, RUL i The remaining service life of critical component i; Let i be the mean of the predicted health index of the key component i at a future time t+τ; The preset failure threshold for critical component i; T represents the time step forward from the current moment; max is the maximum number of prediction steps; min is the function to take the minimum value.

19. The method according to claim 1, characterized in that, The method further includes: Obtain real-time operating conditions, vibration alarm status, maintenance history information, spare parts inventory information, and weather window information; Accordingly, determining the maintenance actions for each key component based on its health index and remaining service life includes: The health index, remaining service life, real-time operating conditions, vibration alarm status, maintenance history, spare parts inventory, and weather window information of each key component are input as state vectors into the reinforcement learning network, which outputs maintenance actions. The maintenance actions include at least continued operation, planned maintenance, and emergency maintenance. The reinforcement learning network is optimized based on a reward function during training, which is determined according to a reliability penalty term, an operation penalty term, an over-maintenance penalty term, and a vibration alarm benefit term.

20. The method according to claim 19, characterized in that, The reward function is determined based on reliability penalty, operational penalty, over-maintenance penalty, and vibration alarm benefit, including: The reward function is determined using the following formula: Wherein, R(s) t ,a t ) represents the state vector s t Perform maintenance action a t Total reward; R rel (s t ,a t ) represents the state vector s t Perform maintenance action a t Reliability penalty item; R run (s t ,a t ) represents the state vector s t Perform maintenance action a t Runtime penalty items; R pen (s t ,a t ) represents the state vector s t Perform maintenance action a t Over-maintenance penalty; R vib (s t ,a t ) represents the state vector s t Perform maintenance action a t Vibration alarm revenue items; The reliability penalty term is determined according to the following formula: in, The failure severity weight for critical component i; HI i (t) represents the health index of key component i at time t; The preset failure threshold for critical component i; For indicator functions, This indicates that the health index is set to 1 when it is below the preset failure threshold, and 0 otherwise; K represents the total number of critical components. The runtime penalty item is determined according to the following formula: in, To maintain action a t Corresponding maintenance costs; To maintain action a t Corresponding losses due to power outages; The over-maintenance penalty is determined according to the following formula: in, Penalty coefficient for over-maintenance; RUL represents the average remaining service life of critical component i. thr To preset the maintenance plan's lifespan threshold; Indicates when maintenance action a t The value is 1 for emergency maintenance and 0 otherwise; K is the total number of critical components. The vibration alarm revenue item is determined according to the following formula: Among them, w vib K represents the vibration alarm benefit coefficient. i (t) represents the kurtosis value of the critical component i; K thr E is the preset kurtosis threshold. ratio,i (t) represents the high-frequency energy ratio; E thr The preset high-frequency energy ratio threshold is used; Indicates when maintenance action a t The value is set to 1 when the kurtosis value of the critical component is greater than the preset kurtosis threshold or the high-frequency energy ratio is greater than the preset high-frequency energy ratio threshold for planned maintenance; otherwise, the value is set to 0.

21. The method according to claim 19, characterized in that, The method further includes: Real-time monitoring of vibration characteristic vectors of key components; When the vibration characteristic vector of any critical component exceeds the preset vibration alarm threshold, the maintenance action of the critical component will be forcibly identified as planned maintenance.

22. The method according to claim 1, characterized in that, The method further includes: Collect equipment status information before maintenance, types of actions performed, equipment status information after maintenance, actual safe operating time of the equipment after maintenance, and vibration feature vectors before and after maintenance to construct experience samples; The experience samples are stored in the experience playback pool; At preset time intervals, experience samples are sampled from the experience replay pool for incremental fine-tuning to update the network parameters of the graph attention network and / or the network parameters of the reinforcement learning network. The updated parameters are used for subsequent health assessment, remaining service life prediction, and determination of maintenance actions.

23. A wind turbine health assessment and maintenance device based on graph attention networks, characterized in that, include: The acquisition module is used to acquire the operating data and vibration data of each key component in the wind turbine, wherein the operating data includes at least wind speed and active power; The component relationship graph construction module is used to construct a component relationship graph with each key component as a graph node and the coupling relationship between key components as a base edge. It determines the load transfer coefficient based on the wind speed and the active power, determines the anomaly propagation coefficient based on the degree to which the vibration data of the target key component deviates from its healthy reference vibration characteristic vector, and generates a dynamic adjacency matrix representing the intensity of degradation influence between key components based on the load transfer coefficient, the anomaly propagation coefficient and the component relationship graph. The health index output module is used to input the operation data, the vibration data and the dynamic adjacency matrix into the graph attention network. The graph attention network is used to perform weighted fusion of the operation data and the vibration data based on the dynamic adjacency matrix and output the health index of each key component. The remaining service life prediction module is used to predict the remaining service life of each key component based on the historical change sequence of the health index at multiple consecutive historical moments. The maintenance decision module is used to determine the maintenance actions for each key component based on the health index and the remaining service life. These maintenance actions are used to guide the maintenance operations of the wind turbine.

24. An electronic device, characterized in that, include: A memory and a processor, the processor and the memory being communicatively connected to each other, the memory storing computer instructions, the processor executing the computer instructions to implement the steps of the method according to any one of claims 1 to 22.

25. A computer storage medium, characterized in that, The computer storage medium stores computer program instructions, which, when executed, implement the steps of the method according to any one of claims 1 to 22.

26. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1 to 22.

Images