Wind turbine generator performance dynamic evaluation method and system based on multi-source data fusion

Abstract

The invention discloses a wind turbine generator performance dynamic evaluation method and system based on multi-source data fusion, and the method comprises the steps: synchronizing multi-modal heterogeneous data through a quantum encryption algorithm and an edge gateway; constructing a space-time semantic graph network through a deep semantic analysis technology, and generating a precise space-time feature matrix; based on an attention mechanism, generating a physical enhancement feature vector; constructing cross-working-condition health index mapping by applying a machine learning algorithm, and quantifying a cross-working-condition comparable health index; a hierarchical incremental learning architecture is adopted, and a multi-objective optimization algorithm is utilized to generate a dynamic maintenance priority sequence; and building a digital twin platform, performing closed-loop verification on the maintenance priority sequence, and generating a unit maintenance scheme. The problem that multi-source data fusion of the wind turbine generator is difficult is solved, the health index quantification accuracy under the complex working condition is improved, the dynamic maintenance strategy is optimized, and closed-loop optimization of the maintenance scheme is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system monitoring, and particularly relates to a wind turbine performance dynamic evaluation method and system based on multi-source data fusion. BACKGROUND

[0002] As the core equipment of new energy power generation, the running state of the wind turbine directly affects the power generation efficiency and safety. In the current operation and maintenance of the wind farm, there are problems of heterogeneous multi-source monitoring data format and poor time and space synchronization, which leads to great difficulty in data fusion. The traditional evaluation method relies on a single data source or a static model, and it is difficult to capture the performance degradation law of the unit under complex working conditions, and the quantification accuracy of the health index is insufficient.

[0003] At the same time, the existing maintenance strategy is mostly based on fixed period or single index, and lacks dynamic priority adjustment mechanism, which is easy to cause excessive maintenance or maintenance lag. Although the digital twin technology has been applied, the combination degree with multi-source data fusion and dynamic evaluation is low, and it is difficult to realize the closed-loop optimization of the maintenance scheme. The present application solves the fusion difficulty problem caused by the heterogeneous multi-source data of the wind turbine and the time and space asynchronization, improves the quantification accuracy of the health index under complex working conditions, optimizes the dynamic maintenance strategy, and realizes the closed-loop optimization of the maintenance scheme. SUMMARY

[0004] The present application provides a wind turbine performance dynamic evaluation method based on multi-source data fusion, which comprises:

[0005] Synchronizing multi-modal heterogeneous data and time and space alignment by quantum encryption algorithm and edge gateway;

[0006] Constructing a time and space semantic graph network through deep semantic analysis technology to generate a precise time and space feature matrix;

[0007] Fusing the time and space feature matrix based on the attention mechanism to generate a physical enhanced feature vector rich in time and space information;

[0008] Using a machine learning algorithm to construct a cross-condition health index mapping based on the physical enhanced feature vector to quantify the cross-condition comparable health index;

[0009] Using a hierarchical incremental learning architecture combined with the cross-condition comparable health index to generate a dynamic maintenance priority sequence using a multi-objective optimization algorithm;

[0010] Building a digital twin platform to perform closed-loop verification on the maintenance priority sequence to generate a unit maintenance scheme.

[0011] The wind turbine performance dynamic evaluation method based on multi-source data fusion as described above, wherein the multi-modal heterogeneous data is synchronized and time and space alignment by quantum encryption algorithm and edge gateway, comprising:

[0012] A high-precision timestamp synchronization mechanism is deployed at the edge gateway to perform nanosecond-level time alignment on multi-modal heterogeneous data streams, and simultaneously map them to a unified space-time coordinate system.

[0013] A lightweight encryption algorithm with dynamic key management is used to perform real-time encryption transmission of data at the edge gateway layer.

[0014] The wind turbine performance dynamic evaluation method based on multi-source data fusion as described above, wherein a space-time semantic graph network is constructed through deep semantic analysis technology to generate a precise space-time feature matrix, including:

[0015] An initial space-time graph network is constructed according to the physical topology of the equipment and the semantic relationship of the sensors.

[0016] A graph neural network is applied to fuse physical constraints for deep space-time semantic analysis, extract high-dimensional space-time semantic features, and generate a precise space-time feature matrix.

[0017] The wind turbine performance dynamic evaluation method based on multi-source data fusion as described above, wherein a space-time feature matrix is fused based on an attention mechanism to generate a physically enhanced feature vector rich in space-time information, including:

[0018] Based on a multi-scale space-time attention mechanism, the feature importance of heterogeneous sensor modalities is adaptively weighted and fused.

[0019] The physical model of the equipment is used as a prior constraint and integrated into the feature transformation process to generate a global physically enhanced feature vector rich in space-time information.

[0020] The wind turbine performance dynamic evaluation method based on multi-source data fusion as described above, wherein a machine learning algorithm is used to construct a cross-condition health index mapping based on the physically enhanced feature vector, quantify the cross-condition comparable health index, including:

[0021] Based on the physically enhanced features and the operating condition parameters, an end-to-end operating condition adaptive cross-condition comparable health index mapping function is trained.

[0022] The mapping function is applied to calculate the cross-condition comparable health index in real time and verify its adaptability and accuracy.

[0023] The wind turbine performance dynamic evaluation method based on multi-source data fusion as described above, wherein a hierarchical incremental learning architecture is used in combination with the cross-condition comparable health index to generate a dynamic maintenance priority sequence using a multi-objective optimization algorithm, including:

[0024] A hierarchical incremental learning framework is constructed to continuously absorb the cross-condition comparable health index and operating condition parameters, and dynamically update the health evaluation parameters.

[0025] The cross-operation comparable health index is taken as input, a multi-objective optimization algorithm is applied in combination with the health evaluation parameters to generate a dynamic unit component maintenance priority sequence.

[0026] The wind turbine performance dynamic evaluation method based on multi-source data fusion as described above, wherein a digital twin platform is built to close-loop verify the maintenance priority sequence, generate a unit maintenance scheme, including:

[0027] The physical model and multi-modal heterogeneous data are integrated to build a high-fidelity digital twin, inject the maintenance priority sequence, and simulate the execution of the maintenance strategy.

[0028] The maintenance priority sequence is evaluated and optimized to generate a final unit maintenance scheme, and the verification result is fed back to the incremental learning.

[0029] The wind turbine performance dynamic evaluation system based on multi-source data fusion, wherein it includes:

[0030] The quantum encryption module is used to synchronize multi-modal heterogeneous data and space-time alignment using quantum encryption algorithms and edge gateways.

[0031] The data analysis module is used to construct a space-time semantic graph network through deep semantic analysis technology to generate a precise space-time feature matrix.

[0032] The matrix fusion module is used to fuse the space-time feature matrix based on the attention mechanism to generate a physical enhanced feature vector rich in space-time information.

[0033] The health mapping module is used to use machine learning algorithms to construct a cross-operation health index mapping based on the physical enhanced feature vector to quantify the cross-operation comparable health index; a hierarchical incremental learning architecture is used in combination with the cross-operation comparable health index to generate a dynamic maintenance priority sequence using a multi-objective optimization algorithm.

[0034] The simulation verification module is used to build a digital twin platform to close-loop verify the maintenance priority sequence to generate a unit maintenance scheme.

[0035] The beneficial effects realized by the present application are as follows:

[0036] The present scheme solves the problems of wind turbine data heterogeneity and space-time asynchronization through multi-source data fusion technology, improves data utilization; deep semantic analysis and attention mechanism enhance the quantification accuracy of health indicators, and cross-operation evaluation is more accurate; dynamic maintenance priority generation optimizes resource allocation, reduces excessive or lagging maintenance; digital twin closed-loop verification improves the reliability of the scheme, and comprehensively improves the operation efficiency and safety of the unit. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0038] Figure 1 is the flow chart of the wind turbine performance dynamic evaluation method based on multi-source data fusion provided by the embodiment one of the present application.

[0039] Figure 2 is the schematic diagram of the wind turbine performance dynamic evaluation system based on multi-source data fusion provided by the embodiment two of the present application. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0041] Embodiment one

[0042] As shown in Figure 1 , the embodiment one of the present application provides a wind turbine performance dynamic evaluation method based on multi-source data fusion, which comprises:

[0043] S110: synchronizing multi-modal heterogeneous data by quantum encryption algorithm and edge gateway and time-space alignment;

[0044] Based on the geometric topological relationship of sensor physical coordinates and the signal propagation delay characteristics, a unified space-time coordinate system is constructed in the edge gateway, and the multi-modal data is mapped to the coordinate system to complete the spatial alignment. By integrating the quantum key distribution module in the edge gateway, the end-to-end quantum key is generated to encrypt the real-time transmission of multi-modal heterogeneous data.

[0045] Synchronizing multi-modal heterogeneous data by quantum encryption algorithm and edge gateway and time-space alignment comprises the following sub-steps:

[0046] S111: deploying a high-precision timestamp synchronization mechanism in the edge gateway, performing nanosecond-level time alignment on the multi-modal heterogeneous data stream, and simultaneously mapping to a unified space-time coordinate system;

[0047] High-precision time synchronization architecture is deployed in the edge gateway to achieve nanosecond-level time alignment. Multi-modal heterogeneous data streams (at least including sensor real-time data, working condition parameters, physical metadata, etc.) enter the timestamp analysis module of the edge gateway, extract the 64-bit nanosecond-level timestamp in the data frame header, and convert it to a time value under the local time reference.

[0048] A hybrid clock synchronization algorithm is adopted: for high-frequency data streams with high sampling rate, a Kalman filter is used to predict time offset and compensate for network transmission delay; for low-frequency data streams, a moving window interpolation algorithm is used for time resampling, which is unified to the standard time grid.

[0049] The sensor physical coordinates and signal features are extracted from the time-aligned data stream to construct a multi-dimensional feature vector. A hierarchical coordinate conversion architecture is adopted: the bottom layer converts the device coordinate system to the workshop coordinate system through the sensor calibration database, the middle layer constructs a spatial relationship graph using a graph neural network, and the upper layer maps different device coordinate systems to the global coordinate system through a rigid transformation matrix. A space-time calibration algorithm is used to compensate for signal propagation delay and multipath effects by combining a physical propagation model and a Bayesian filter. Finally, the data mapped to the unified space-time coordinate system is stored as a space-time tensor.

[0050] S112: A lightweight encryption algorithm with dynamic key management is used to encrypt and transmit data in real time at the edge gateway layer.

[0051] A quantum encryption transmission architecture with dynamic key management is constructed at the edge gateway layer. The quantum key distribution (QKD) module built-in the edge gateway establishes a quantum channel with the quantum key distribution center based on the BB84 protocol, generates an initial symmetric key through single-photon polarization state encoding, and inputs the key parameters into the key management module for initialization after verifying the quantum state non-cloning.

[0052] Multi-modal heterogeneous raw data is uniformly converted by the data adaptation layer of the edge gateway to enter the encryption process: the key management module monitors the number of key uses or time intervals in real time, generates a new key and marks the old key as invalid when the trigger condition is met, ensuring the timeliness of the dynamic key; the encryption algorithm module calls the current valid key and uses a quantum-secure hybrid encryption mechanism (quantum key encryption session key + AES-256 algorithm encryption data payload) to encrypt each data block one by one, while embedding a 128-bit key ID, an 8-bit device identifier, and a 64-bit nanosecond-level timestamp in the data header to form an encrypted data frame. The encrypted data is sent in real time to the backend system through the edge transmission interface of the edge gateway, and the dynamic updating and packet-by-packet verification of the quantum key are performed during transmission.

[0053] S120: A space-time semantic graph network is constructed through deep semantic analysis technology to generate a precise space-time feature matrix;

[0054] The semantic graph network is constructed based on a space-time tensor to construct an initial graph structure. A graph neural network is applied to fuse physical constraints for deep space-time semantic analysis, to learn complex nonlinear space-time dependency between nodes, to mine hidden device state semantic information, to extract high-dimensional space-time semantic features, and to form a precise space-time feature matrix representing the global and local space-time state of the device.

[0055] The space-time semantic graph network is constructed by a deep semantic analysis technology to generate a precise space-time feature matrix, including the following sub-steps:

[0056] S121: An initial space-time graph network is constructed according to the device physical topology and sensor semantic relationship;

[0057] Based on the space-time tensor, an initial space-time graph network is constructed to extract sensor physical coordinates and signal features (time domain, frequency domain, and time-frequency domain feature vectors) from a unified space-time coordinate system, and a graph structure is constructed in combination with a device topology database. The node generation module classifies sensors by type, and each node contains a 128-dimensional feature vector, and a key component node additionally embeds physical metadata.

[0058] The edge construction module generates three types of connections based on space-time correlation: spatial connection, temporal connection, and physical dependency connection. The graph initialization algorithm generates an initial adjacency matrix by hierarchical clustering, and then performs sparse processing to retain the top 10 strongest connections of each node to construct a sparse graph structure.

[0059] The physical constraint injection module converts the device dynamics equation into a node attribute constraint condition, and embeds the physical constraint into the graph structure by the Lagrange multiplier method. The finally generated space-time graph network is stored as an adjacency matrix and a node feature matrix for deep semantic analysis.

[0060] S122: A graph neural network is applied to fuse physical constraints for deep space-time semantic analysis, to extract high-dimensional space-time semantic features, and to generate a precise space-time feature matrix;

[0061] Based on the initial space-time graph network, deep semantic analysis is performed, and the adjacency matrix and node feature matrix are input into the graph neural network. The graph neural network adopts a hybrid architecture: a 3-layer graph convolutional network at the bottom layer extracts spatial features, a 2-layer graph attention network at the middle layer captures nonlinear dependency, and a 1-layer graph autoencoder at the upper layer performs feature compression and reconstruction. The physical constraint is injected in two ways, including adding Laplacian smoothing constraint in the GCN layer to ensure that the features of spatially adjacent nodes are similar, and designing a physical perception attention mechanism in the GAT layer to convert the device dynamics equation into an attention weight prior.

[0062] The graph neural network training adopts a multi-task learning framework: the main task is node state prediction, and the auxiliary task is edge existence prediction. The optimizer adopts Adam, and the early stopping strategy monitors the validation set loss. To handle non-steady-state conditions, the system adopts an online learning mechanism: every 100 new data points are received, the GCN layer parameters are frozen, and the GAT layer is fine-tuned. The interaction information of node embedding is calculated to mine the spatio-temporal dependence relationship and identify the key spatio-temporal correlation path. The optimized graph embedding is stored as a spatio-temporal semantic tensor and output to the feature extraction module.

[0063] The node and edge embeddings in the optimized graph neural network are analyzed to extract high-dimensional spatio-temporal semantic features. Based on the output of the graph neural network, the embedding vector of each node and edge is used as a feature source. The embedding vector is processed using a feature extraction function to generate a high-dimensional feature vector.

[0064] The system organizes these feature vectors according to the physical structure of the device and the spatio-temporal relationship to form the rows and columns of the matrix. Each row represents a local spatio-temporal state of the device, and each column represents the feature dimension under different states. The node i feature vector is obtained through a physical semantic driven feature aggregation mechanism, and the spatio-temporal feature matrix is constructed to generate an accurate spatio-temporal feature matrix.

[0065] The core formula is:

[0066]

[0067] σ(·) represents the activation function, which is a nonlinear transformation of the calculation result inside the parentheses; W self is the weight matrix, which is used for linear transformation of the original feature H v,i of node i itself, learning the representation method of the node's own features; H v,i represents the original feature vector of node i, which is the initial expression of the node's own attributes, state, and other basic information. W self H v,i gets the node's own features based on its original features; j∈N(i) represents the neighbor node set of node i, and j is each neighbor j of node i; a ij is the attention weight between node i and neighbor node j, and the larger the value, the stronger the influence of the neighbor on node i; H v,j ||H v,i represents the concatenation of the original feature H v,j of neighbor node j and the original feature H v,i of node i; φ(·) calculates the physical interaction strength / feature correlation degree between neighbor node j and itself node i.

[0068] S130: Fuse the spatio-temporal feature matrix based on the attention mechanism to generate a physical enhanced feature vector rich in spatio-temporal information;

[0069] The spatio-temporal feature matrix is weighted and fused, a multi-scale attention mechanism is constructed, and the importance of different sensor modal features is adaptively evaluated. The device physical model is taken as a priori constraint and is integrated into the attention weight calculation or feature transformation process. The physical model parameters are taken as weight adjustment factors to dynamically adjust the weights of the feature vectors, and a global physical enhanced feature vector rich in spatio-temporal information is generated.

[0070] The spatio-temporal feature matrix is fused based on the attention mechanism to generate a physical enhanced feature vector rich in spatio-temporal information, including the following sub-steps:

[0071] S131: Based on the multi-scale spatio-temporal attention mechanism, the feature importance of heterogeneous sensor modalities is adaptively weighted and fused;

[0072] The input spatio-temporal feature matrix is decomposed into sub-matrices according to the modalities (dimension remains unchanged). The time attention layer uses a bidirectional gated recurrent unit to capture the temporal dependence, and the forward and backward outputs are spliced and then passed through an attention weight matrix to generate time attention coefficients. The spatial attention layer constructs a k-neighbor graph and calculates the spatial correlation between nodes through a graph attention network, and then generates spatial attention coefficients through softmax.

[0073] Weighted average pooling with different time windows is used to process the feature matrix in parallel to obtain pooling weights. The output is spliced through the channel and then compressed to the original dimension through a fully connected layer. The modal fusion layer weights and sums the spatio-temporal attention coefficients of each modality and the multi-scale features. Through adaptive weight learning, a meta-learning mechanism is designed, and through a task-independent meta-learning algorithm, the network is pre-trained on the support set task to output the fusion weights of each modality and generate a weighted feature matrix. The weighted feature matrix not only retains multi-scale spatio-temporal features, but also adaptively highlights key modality information.

[0074] S132: The device physical model is taken as a priori constraint and integrated into the feature transformation process to generate a global physical enhanced feature vector rich in spatio-temporal information.

[0075] A double-branch network is designed: the left branch receives the weighted feature matrix output by the multi-scale spatio-temporal attention mechanism and maps it to the physical state space through a 3-layer MLP; the right branch encodes the physical model parameters into a 64-dimensional vector and generates a time-series physical constraint through a physical simulator network.

[0076] The physical consistency score is calculated through the attention weight adjustment layer. The time steps with scores greater than the threshold have their attention weights enhanced, and the time steps with scores less than the threshold have a physical regularization term activated. A physical perception transformation matrix is designed in the feature transformation layer to keep the transformed features consistent with the physical model prediction.

[0077] The global physical enhanced feature vector generated by the fusion layer contains deep spatio-temporal semantic information and conforms to the physical law of the device, and is used for cross-condition comparable health index mapping.

[0078] S140: Using a machine learning algorithm to construct a cross-condition health index mapping based on the physical enhanced feature vector to quantify the cross-condition comparable health index;

[0079] Based on the physical enhanced feature vector, a cross-condition comparable health index mapping is constructed. The physical enhanced feature vector and the condition parameters (load, speed, etc.) are spliced into an input vector for data preprocessing, including data cleaning, normalization, and feature extraction. The mapping function uses kernel regression method to obtain the cross-condition comparable health index. A condition clustering mechanism is designed, and each cluster corresponds to an independent mapping parameter. Determine the cluster to which the current condition belongs, call the corresponding mapping function for calculation, and verify the cross-condition comparability through cross-validation.

[0080] Using a machine learning algorithm based on the physical enhanced feature vector, a cross-condition adaptive cross-condition comparable health index mapping is constructed to accurately quantify the cross-condition comparable health index, including the following sub-steps:

[0081] S141: Based on the physical enhanced feature and the condition parameter, an end-to-end cross-condition adaptive cross-condition comparable health index mapping function is trained;

[0082] The physical enhanced feature vector and the condition parameter are spliced into a joint input vector. A hierarchical kernel regression framework is used for adaptive mapping: a Gaussian mixture model clusterer is trained based on the condition data (the covariance type is a diagonal matrix, and the number of clusters is automatically determined by the Bayesian information criterion), a membership function of the condition cluster is defined, and the matching degree of the current condition and the typical condition mode (such as "no load / low load / medium load / high load / overload" of wind power) is quantified.

[0083] The membership function expression is:

[0084]

[0085] ω k (θ) represents the matching degree value of the current condition and the kth typical condition mode; π k represents the prior probability of the current cluster, which represents the proportion of data belonging to the kth condition cluster in the historical data; N(θ|μ k ,Σ k ) represents the probability density function form of the multivariate Gaussian distribution; wherein N(·) represents the probability density function symbol of the Gaussian distribution, which is used to calculate the probability density value of the current kth condition cluster under the given condition parameter θ, which follows the Gaussian distribution with mean μ k and covariance ∑ k ; d represents the total number of condition clusters; πj P (θ | Cj) represents the prior probability of the jth operating condition cluster; N (θ | μ j ,Σ j ) represents the Gaussian distribution probability density function corresponding to the jth operating condition cluster, where μ j is the mean vector of the jth cluster, and ∑ j is the covariance matrix of the jth cluster, and the Gaussian distribution probability density value of the current operating condition parameter θ belonging to the jth cluster is calculated.

[0086] An independent radial basis kernel regression model is trained for each operating condition cluster, and the core formula is as follows:

[0087]

[0088] H k (u) represents the output of the radial basis kernel regression model trained for the kth operating condition cluster, which is used to evaluate the equipment health state; S k represents the historical sample set corresponding to the kth operating condition cluster; i represents the ith historical sample. represents the weight coefficient of the kernel regression model corresponding to the historical sample i in the kth operating condition cluster. represents the radial basis function, which measures the similarity between the current input splicing vector u and the splicing vector corresponding to the historical sample i in the kth operating condition cluster; ||u-u i || 2 represents the distance between the current input vector u and the historical sample vector u i in the feature space; the smaller the distance, the more similar the two are; σ k represents the width parameter of the radial basis function corresponding to the historical sample i in the kth operating condition cluster; b k represents the bias term of the radial basis and regression model corresponding to the kth operating condition cluster.

[0089] S142: Apply the mapping function to calculate the cross-condition comparable health index in real time, and verify its adaptability and accuracy.

[0090] The membership degrees of each operating condition cluster are calculated in real time, and the regression models of the top two clusters with the highest membership degrees are called to perform weighted fusion. The radial basis kernel regression parameters of the corresponding cluster are loaded, and the weighted cross-condition comparable health index is calculated. The core formula is as follows:

[0091]

[0092] wherein, θ represents the key parameter of the current operating condition cluster core state; K represents the total number of operating condition clusters; ω k (θ) represents the membership degree of the kth operating condition cluster under the current operating condition; I k (θ) represents the physical rationality factor of the kth operating condition cluster, which is represented by a piecewise Gaussian function: when θ ∈ safe domain, denotes the health weight decay term within the safety domain, θ k,safe is the kth cluster optimal safety operating point, is the safety domain operating tolerance, the weight decay rate when the operating condition deviates from the safety operating point; when θ ∈ the risk domain, denotes the health weight penalty term within the risk domain, ∈ is the risk penalty coefficient, used to reduce the health contribution of the risk operating condition, is the risk domain operating tolerance; H k (u) denotes the health regression model output of the kth operating condition cluster.

[0093] S150: Adopting hierarchical incremental learning architecture combined with cross-condition comparable health index, using multi-objective optimization algorithm to generate dynamic maintenance priority sequence;

[0094] The edge end cleans the cross-condition comparable health index and operating condition data in real time, updates the component state; the cloud end fuses maintenance feedback, dynamically calibrates health evaluation parameters, and drives model evolution. Decision optimization layer: embedded industrial constraints, adaptive crossover and variation based on NSGA-III algorithm, hierarchical screening of elite solutions, output maintenance sequence every 5 minutes, and synchronous matching of resources and qualifications.

[0095] Adopting hierarchical incremental learning architecture combined with health adaptive index, using multi-objective optimization algorithm to generate dynamic maintenance priority sequence, including the following sub-steps:

[0096] S151: Build a hierarchical incremental learning framework, continuously absorb cross-condition comparable health index, operating condition, and dynamically update health evaluation parameters;

[0097] Receive cross-condition comparable health index and real-time operating condition parameters as input to the bottom module of the framework. Perform data preprocessing, including data cleaning, normalization, and feature extraction. Pass the processed data to the middle layer with a recurrent neural network structure to capture dynamic features in time series data. By learning the variation of device state over time, extract the feature vector related to the health state.

[0098] Based on the Bayesian update rule, dynamically update the health evaluation parameters combined with new input data, adjust the model weights to adapt to new operating conditions and maintenance states, record the historical health state of the device, operating condition changes, and the effect of maintenance measures.

[0099] S152: Take the cross-condition comparable health index as input, apply multi-objective optimization algorithm combined with health evaluation parameters to generate dynamic unit component maintenance priority sequence.

[0100] Based on the updated health assessment parameters and multi-modal heterogeneous data, load the unit holographic state matrix, and construct a multi-objective optimization model: maximize equipment availability, minimize maintenance cost, and minimize failure risk as the target, impose cross-condition comparable health index threshold, resource constraints, and other maintenance conditions.

[0101] Apply improved NSGA-III algorithm for solution: initialize 100 candidate maintenance sequences, encode as priority vector; perform adaptive crossover and mutation operation - crossover probability is dynamically adjusted according to cross-condition comparable health index threshold; adopt hierarchical reference point strategy to divide target space into 5 emergency levels, combined with demand for elite solution screening, eliminate redundant schemes, output unit component maintenance priority sequence.

[0102] S160: Build a digital twin platform to verify the maintenance priority sequence in a closed loop and generate a unit maintenance scheme;

[0103] Generate a joint digital twin through an adaptive weighted fusion layer. After receiving the maintenance priority sequence, run the physical inversion verification synchronously, output the maintenance scheme and digital twin evaluation report verified by physics, and form a decision-making closed loop.

[0104] Build a digital twin platform to verify the maintenance priority sequence in a closed loop and generate a unit maintenance scheme, including the following sub-steps:

[0105] S161: Integrate physical models and multi-modal heterogeneous data to build a high-fidelity digital twin, inject the maintenance priority sequence, and simulate the execution of maintenance strategies;

[0106] Integrate device physical models, cross-condition comparable health index mapping functions, and real-time multi-modal heterogeneous data streams to generate a joint digital twin through an adaptive weighted fusion layer. After receiving the maintenance priority sequence, start the discrete event simulation engine: based on historical work order statistics, establish a maintenance operation time model, combine the current resource state to simulate the whole process, and dynamically update the health state parameters of the virtual device.

[0107] S162: Evaluate and optimize the maintenance priority sequence to generate the final unit maintenance scheme, and feed back the verification results to machine learning.

[0108] Run physical inversion verification synchronously during simulation, input the updated cross-condition comparable health index, derive the theoretical vibration value, and compare it with the actual sensor data. Through the verified maintenance scheme, automatically generate executable work orders (including spare parts list, time window and personnel demand), and mark the conflict points (such as material fatigue coefficient deviation) and feedback to machine learning, trigger health assessment model parameter calibration. Finally, output the maintenance scheme and digital twin evaluation report verified by physics, forming a decision-making closed loop.

[0109] Embodiment two

[0110] As Figure 2 shown, embodiment two of the present application provides a wind turbine performance dynamic evaluation system based on multi-source data fusion, comprising:

[0111] Quantum encryption module 21: using quantum encryption algorithm and edge gateway to synchronize multi-modal heterogeneous data and time-space alignment;

[0112] Data analysis module 22: constructing a time-space semantic graph network through deep semantic analysis technology, and generating a precise time-space feature matrix;

[0113] Matrix fusion module 23: fusing the time-space feature matrix based on attention mechanism, and generating a physical enhanced feature vector rich in time-space information;

[0114] Health mapping module 24: using machine learning algorithm to construct a cross-condition health index mapping based on the physical enhanced feature vector, and quantifying a cross-condition comparable health index; adopting a hierarchical incremental learning architecture combined with the cross-condition comparable health index, and generating a dynamic maintenance priority sequence using a multi-objective optimization algorithm;

[0115] Simulation verification module 25: building a digital twin platform to perform closed-loop verification on the maintenance priority sequence, and generating a unit maintenance scheme.

[0116] Corresponding to the above-mentioned embodiments, the embodiment of the present application provides a computer storage medium, comprising: at least one memory and at least one processor;

[0117] The memory is used to store one or more program instructions;

[0118] The processor is used to run one or more program instructions to execute the wind turbine performance dynamic evaluation method and system based on multi-source data fusion.

[0119] Corresponding to the above-mentioned embodiments, the embodiment of the present application provides a computer readable storage medium, the computer storage medium contains one or more program instructions, and the one or more program instructions are used to execute the wind turbine performance dynamic evaluation method and system based on multi-source data fusion by the processor.

[0120] The embodiment disclosed by the present application provides a computer readable storage medium, and the computer readable storage medium stores computer program instructions, when the computer program instructions run on the computer, the computer executes the wind turbine performance dynamic evaluation method and system based on multi-source data fusion.

[0121] In the embodiments of the present application, the processor can be an integrated circuit chip with a processing capability of signals. The processor can be a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0122] The disclosed methods, steps and logic block diagrams in the embodiments of the present application can be implemented or executed by using general purpose processors, which can be microprocessors or the processors of any kind. The steps of the methods disclosed in the embodiments of the present application can be directly embodied as hardware code executed by the processor, or be executed by a combination of hardware and software modules in the processor. The software modules can be located in storage media such as random access memory (RAM), flash memory, read only memory (ROM), programmable read-only memory (PROM), electrically programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, or other mature storage media in the field. The processor reads information in the storage media and combines it with hardware to complete the steps of the above methods.

[0123] The storage media can be a memory, which can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories.

[0124] The non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically EPROM (EEPROM) or a flash memory.

[0125] The volatile memory can be Random Access Memory (RAM), which is used as external cache memory. By way of example, and not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous Dynamic Random Access Memory (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), and direct Rambus RAM (DRRAM). The below-described embodiments do not limit the scope of the application to any particular implementation.

[0126] The storage media described in the embodiments of the present application is intended to include, but not limited to, these and any other suitable types of memory.

[0127] Those skilled in the art should be aware that the functions described in the embodiments of the present application can be implemented in combination of hardware and software in one or more of the above examples. When the software is applied, the corresponding functions can be stored in a computer readable medium or transmitted as one or more instructions or codes on the computer readable medium. The computer readable medium includes computer storage medium and communication medium, wherein the communication medium includes any medium that facilitates the transfer of computer programs from one place to another. The storage medium can be any available medium that can be accessed by a general or special purpose computer.

[0128] The above detailed description sets forth the purposes, technical solutions and beneficial effects of the present application. It should be understood that the above detailed description is only a specific implementation of the present application and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made on the basis of the technical solutions of the present application should be included in the protection scope of the present application.

Claims

1. A wind turbine performance dynamic evaluation method based on multi-source data fusion, characterized in that, Comprising: synchronizing multi-modal heterogeneous data and spatio-temporal alignment using quantum encryption algorithm and edge gateway; constructing spatio-temporal semantic graph network through deep semantic analysis technology to generate precise spatio-temporal feature matrix; fusing spatio-temporal feature matrix based on attention mechanism to generate physical enhanced feature vector rich in spatio-temporal information; using machine learning algorithm to construct cross-condition health index mapping based on physical enhanced feature vector to quantify cross-condition comparable health index; adopting hierarchical incremental learning architecture combined with cross-condition comparable health index to generate dynamic maintenance priority sequence using multi-objective optimization algorithm; building digital twin platform to close-loop verify maintenance priority sequence to generate unit maintenance scheme.

2. The wind turbine performance dynamic assessment method based on multi-source data fusion according to claim 1, characterized in that, Synchronizing multi-modal heterogeneous data and spatio-temporal alignment using quantum encryption algorithm and edge gateway, comprising: deploying high-precision timestamp synchronization mechanism in edge gateway to perform nanosecond-level time alignment on multi-modal heterogeneous data stream, and simultaneously mapping to unified spatio-temporal coordinate system; adopting lightweight encryption algorithm with dynamic key management to perform real-time encryption transmission of data at edge gateway layer.

3. The wind turbine performance dynamic assessment method based on multi-source data fusion of claim 1, wherein, Constructing spatio-temporal semantic graph network through deep semantic analysis technology to generate precise spatio-temporal feature matrix, comprising: constructing initial spatio-temporal graph network according to device physical topology and sensor semantic relationship; applying graph neural network to fuse physical constraints for deep spatio-temporal semantic analysis, extracting high-dimensional spatio-temporal semantic features to generate precise spatio-temporal feature matrix.

4. The wind turbine performance dynamic assessment method based on multi-source data fusion of claim 1, wherein, Fusing spatio-temporal feature matrix based on attention mechanism to generate physical enhanced feature vector rich in spatio-temporal information, comprising: based on multi-scale spatio-temporal attention mechanism, adaptively weighting and fusing feature importance of heterogeneous sensor modalities; integrating device physical model as prior constraint into feature transformation process to generate global physical enhanced feature vector rich in spatio-temporal information.

5. The wind turbine performance dynamic assessment method based on multi-source data fusion of claim 1, wherein, Using machine learning algorithm to construct cross-condition health index mapping based on physical enhanced feature vector to quantify cross-condition comparable health index, comprising: based on physical enhanced features and condition parameters, training end-to-end condition-adaptive cross-condition comparable health index mapping function; applying mapping function to calculate cross-condition comparable health index in real time and verifying its adaptability and accuracy.

6. The wind turbine performance dynamic assessment method based on multi-source data fusion of claim 1, wherein, Adopting hierarchical incremental learning architecture combined with cross-condition comparable health index to generate dynamic maintenance priority sequence using multi-objective optimization algorithm, comprising: constructing hierarchical incremental learning framework to continuously absorb cross-condition comparable health index and condition parameters, dynamically updating health evaluation parameters; taking cross-condition comparable health index as input, applying multi-objective optimization algorithm combined with health evaluation parameters to generate dynamic unit component maintenance priority sequence.

7. The wind turbine generator performance dynamic assessment method based on multi-source data fusion according to claim 1, characterized in that, Building digital twin platform to close-loop verify maintenance priority sequence to generate unit maintenance scheme, comprising: integrating physical model and multi-modal heterogeneous data to construct high-fidelity digital twin, injecting maintenance priority sequence to simulate execution of maintenance strategy; evaluating and optimizing maintenance priority sequence to generate final unit maintenance scheme, and feeding back verification results to incremental learning.

8. A wind turbine performance dynamic assessment system based on multi-source data fusion, characterized in that, Comprising: quantum encryption module for synchronizing multi-modal heterogeneous data and spatio-temporal alignment using quantum encryption algorithm and edge gateway; The data analysis module is configured to construct a spatio-temporal semantic graph network through deep semantic analysis technology, and generate a precise spatio-temporal feature matrix; The matrix fusion module is configured to fuse the spatio-temporal feature matrix based on an attention mechanism, and generate a physical enhanced feature vector rich in spatio-temporal information; The health mapping module is configured to use a machine learning algorithm to construct a cross-condition health index mapping based on the physical enhanced feature vector, and quantify a cross-condition comparable health index; adopt a hierarchical incremental learning architecture combined with the cross-condition comparable health index, and use a multi-objective optimization algorithm to generate a dynamic maintenance priority sequence; The simulation verification module is configured to build a digital twin platform, and perform closed-loop verification on the maintenance priority sequence to generate a unit maintenance scheme.

Images