Method and device for management decision of unmanned wind farm based on multi-source data fusion

Abstract

The application relates to the technical field of power system intelligent operation and maintenance, in particular to a multi-source data fusion unmanned wind farm management decision method and device. The multi-source data fusion unmanned wind farm management decision method relates to the technical field of wind farm intelligent operation and maintenance. The method comprises the following steps: acquiring and preprocessing multi-modal data covering high frequency, medium frequency, low frequency and environment; utilizing a dynamic twin combining physical mechanism and deep neural network to perform residual error analysis; then performing feature fusion and causal inference through a situational awareness attention mechanism and a causal knowledge graph; and finally combining a historical case library to generate multi-dimensional decision suggestions. The application can realize accurate prediction, root cause tracing and scientific decision of wind farm equipment faults, thereby improving the intelligent operation and maintenance level and economic benefits of the wind farm.

Description

Technical Field

[0001] This application relates to the field of intelligent operation and maintenance technology for wind farms, specifically to a management decision-making method and device for unmanned wind farms based on multi-source data fusion. Background Technology

[0002] Wind farms, as a crucial component of renewable energy, play a key role in the global energy structure transformation. With the continuous growth of installed wind power capacity, wind farm operation and maintenance management faces increasingly complex challenges. Modern wind farms generally employ sensor networks to collect equipment operating data, achieving real-time monitoring of turbine status through data acquisition and monitoring systems. Existing technologies typically utilize multi-source data acquisition methods, including vibration sensors, current and voltage transformers, SCADA systems, and environmental monitoring equipment, to obtain various parameter information during turbine operation. These systems perform preliminary data processing and then assess equipment status through statistical analysis or machine learning models. When anomalies are detected, an early warning mechanism is triggered, and maintenance personnel perform corresponding maintenance operations based on system prompts. Some advanced systems also establish historical fault case databases to assist in fault diagnosis and decision support.

[0003] In the process of realizing this invention, the inventors discovered that in the prior art, the wind farm operation and maintenance system has limited processing capabilities for multi-source heterogeneous data, making it difficult to effectively integrate data sources of different frequencies and properties. This leads to uncertainty in the system's state assessment under complex operating conditions, affecting the accuracy and timeliness of operation and maintenance decisions. Summary of the Invention

[0004] One of the objectives of this invention is to provide a management decision-making method and apparatus for unmanned wind farms based on multi-source data fusion, which can solve technical problems such as low utilization rate of multi-source heterogeneous data, insufficient accuracy of fault prediction, and lack of interpretable decision support.

[0005] To solve the above-mentioned technical problems, the embodiments of the present invention are implemented as follows: Firstly, a management decision-making method for unmanned wind farms based on multi-source data fusion is provided, including the following steps: Multimodal data is acquired and preprocessed to obtain standardized structured data, including high-frequency data, mid-frequency data, low-frequency data, and environmental data; A dynamic twin combining a physical mechanism model and a deep neural network is constructed. The standardized structural data is processed using the dynamic twin to obtain theoretical predictions and output fault residuals. A mapping relationship is established between the fault residuals and the standardized structure data, and a feature vector is output. The standardized structure data is dynamically fused using context-aware attention to obtain the fused feature vector. A causal knowledge graph is constructed based on the fused feature vectors, key features are screened to eliminate false associations, and the gating weights are dynamically adjusted. Fault inference is performed based on the dynamically adjusted gating weights, and fault prediction data is output. Based on the fault prediction data, historical fault cases are obtained by feature vectorization retrieval from the historical case database. The historical fault cases are then compared with the fault prediction data in a spatiotemporal manner to obtain relevant fault cases, and the confidence level of the relevant fault cases is calibrated. A causal attribution report is generated based on standardized structured data, causal knowledge graphs, and relevant failure cases. Multidimensional decision-making recommendations are formed based on failure prediction data and relevant failure cases.

[0006] Secondly, a management and decision-making device for unmanned wind farms based on multi-source data fusion is provided, comprising the following modules: Data acquisition module: used to acquire multimodal data and preprocess it to obtain standardized structured data, wherein the multimodal data includes high-frequency data, mid-frequency data, low-frequency data and environmental data; Fault Residual Acquisition Module: Used to construct a dynamic twin combining a physical mechanism model and a deep neural network, and to process the standardized structural data using the dynamic twin to obtain theoretical expectations and output fault residuals; Feature vector fusion module: used to establish a mapping relationship between fault residuals and the standardized structure data and output feature vectors, and dynamically fuse the standardized structure data using context-aware attention to obtain fused feature vectors; Fault prediction module: used to construct a causal knowledge graph based on the fused feature vector, filter key features to exclude false associations, dynamically adjust the gating weights, perform fault inference based on the dynamically adjusted gating weights, and output fault prediction data; Fault Case Acquisition Module: Used to retrieve historical fault cases from the historical case database based on the fault prediction data using feature vectorization, compare the historical fault cases with the fault prediction data in a spatiotemporal manner to obtain relevant fault cases, and calibrate the confidence level of the relevant fault cases. Report generation module: Used to generate causal tracing reports based on standardized structured data, causal knowledge graphs and relevant failure cases, and to form multi-dimensional decision-making suggestions based on failure prediction data and relevant failure cases.

[0007] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following: This application provides a management decision-making method and apparatus for unmanned wind farms based on multi-source data fusion. This method aims to address technical problems in existing unmanned wind farm operation and maintenance, such as low utilization of multi-source heterogeneous data, insufficient accuracy in fault prediction, and lack of interpretable decision support. The method acquires and preprocesses multimodal data covering high-frequency, mid-frequency, low-frequency, and environmental data. It then utilizes a dynamic twin combining physical mechanisms and deep neural networks for residual analysis. Furthermore, it employs context-aware attention mechanisms and causal knowledge graphs for feature fusion and causal inference, ultimately generating multi-dimensional decision recommendations by combining historical case libraries. This method enables accurate prediction, root cause analysis, and scientific decision-making for wind farm equipment failures, thereby improving the intelligent operation and maintenance level and economic efficiency of wind farms.

[0008] This application effectively solves the technical problems of low utilization rate of multi-source heterogeneous data, insufficient accuracy of fault prediction, and lack of interpretable decision support through the above-mentioned technical solution. It realizes accurate prediction, root cause tracing and scientific decision-making of wind farm equipment faults, thereby improving the intelligent operation and maintenance level and economic benefits of wind farms.

[0009] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0010] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0011] Figure 1 A flowchart illustrating the management decision-making method for unmanned wind farms based on multi-source data fusion, as provided in this embodiment of the invention.

[0012] Figure 2 This is a flowchart illustrating the process of acquiring multimodal data and preprocessing it to obtain standardized structured data in an embodiment of the present invention.

[0013] Figure 3 This is a flowchart of obtaining the fused feature vector in an embodiment of the present invention.

[0014] Figure 4 This is a flowchart of obtaining fault prediction data in an embodiment of the present invention.

[0015] Figure 5 This is a schematic diagram of a management and decision-making device for an unmanned wind farm that integrates multi-source data in an embodiment of the present invention.

[0016] Figure 6 This is an internal structural diagram of a computer device in an embodiment of the present invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to embodiments. It is to be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit the invention.

[0018] Please refer to the attached document. Figure 1 As shown in the figure, this embodiment provides a management decision-making method for unmanned wind farms based on multi-source data fusion, including the following steps: Step S001: Acquire multimodal data and preprocess it to obtain standardized structured data. The multimodal data includes high-frequency data, mid-frequency data, low-frequency data, and environmental data.

[0019] Understandably, data collected from different sampling mechanisms and sensing levels, with significantly varying temporal resolutions and physical semantic dimensions, is transformed from chaotic, multi-source data into standardized, structured data usable by subsequent models through preprocessing. This effectively improves data quality and computational efficiency, providing a fundamental input source for building a joint analysis framework across scales and physical domains.

[0020] The high-frequency data can be vibration acceleration signals, used to characterize the transient mechanical state of rotating components such as gearbox bearings and generator bearings; the mid-frequency data can be current and voltage signals, used to reflect the dynamic load characteristics during the power conversion process; the low-frequency data can be operating parameters such as power, speed, temperature, and pitch angle output by the SCADA system, used to characterize the macroscopic operating conditions of the equipment; and the environmental data can be wind speed, wind direction, air temperature, air pressure, and humidity, used to describe the external physical field state of the wind turbine. These four types of data come from different sampling mechanisms and sensing levels, and have significantly different temporal resolutions and physical semantic dimensions.

[0021] Please refer to the attached document. Figure 2 As shown in the embodiment, the step of acquiring multimodal data and preprocessing it to obtain standardized structured data includes: Step S010: Collect raw data using various high-precision sensors and monitoring systems deployed at key parts of the wind turbine. Acquire multimodal data from the high-frequency physical layer, mid-frequency process layer, low-frequency control layer, and environmental layer.

[0022] The high-frequency physical layer refers to the physical state signals of the equipment collected at a sampling rate of ≥20kHz. These can be vibration acceleration signals, such as data output from piezoelectric accelerometers installed on gearbox bearings and generator bearings, used to characterize the transient impact and high-frequency resonance characteristics of mechanical components. The mid-frequency process layer refers to electrical process quantity signals collected at a sampling rate of 1-10kHz, such as current and voltage signals, which can be data output from current transformers and voltage transformers, used to reflect the dynamic energy conversion process of the converter system and generator. The low-frequency control layer refers to the operating parameters of the control system with a sampling rate of less than 1Hz, such as power, speed, pitch angle, nacelle yaw angle, and key temperature values ​​output by the SCADA system, used to characterize the overall operating conditions of the wind turbine and the execution status of control commands. The environmental layer refers to the wind speed, wind direction, air temperature, air pressure, and humidity data collected synchronously by an ultrasonic anemometer, wind vane, thermo-hygrometer, and barometer on the top of the nacelle, used to characterize the external operating boundary conditions that affect the wind turbine load and performance.

[0023] The four layers of data—high-frequency physical layer, mid-frequency process layer, low-frequency control layer, and environmental layer—together constitute a complete multi-source observation system covering physical mechanisms, process responses, control logic, and environmental disturbances. Its hierarchical structure design makes each type of data independent and complementary in terms of time scale, physical meaning, and engineering origin, providing semantically clear and dimensionally complete raw inputs for subsequent fusion modeling.

[0024] Step S011: Perform spatiotemporal alignment and denoising preprocessing on the acquired multimodal data. Specifically, preprocess the acquired multi-source heterogeneous data to eliminate the impact of data quality on the model.

[0025] The spatiotemporal alignment refers to the process of uniformly mapping multimodal data from different sampling frequencies and different acquisition start times to a common time axis. This process includes: downsampling high-frequency vibration data to a time resolution consistent with that of mid-frequency data using resampling or sliding window aggregation, and then aligning mid-frequency, low-frequency, and environmental data to the same time grid using linear interpolation or nearest neighbor filling based on timestamp matching, thereby ensuring that the data of each layer are comparable and correlated at the same time point.

[0026] First, noise reduction is performed. For high-frequency vibration signals, wavelet threshold denoising is used, such as using the Daubechies wavelet basis to decompose the signal into multiple scales. High-frequency white noise is filtered out by setting a threshold, while retaining the impact component containing fault information. For low-frequency data such as SCADA, the Kalman filter algorithm is used to remove singularities caused by sensor jitter or transmission errors, resulting in smooth trend data.

[0027] Second, missing value imputation is performed. For data loss caused by network fluctuations, linear interpolation is used to calculate the estimated value of the missing point using data values ​​from adjacent time points before and after the missing point, in order to ensure the temporal continuity of the data.

[0028] Third, standardization is performed. This is to eliminate differences in dimensions (such as vibration amplitude). The influence of power (MW) on model calculations is investigated using the Min-Max normalization method, as shown in the formula: Map all data to the [0,1] interval, or use Z-Score standardization, formula: The data is transformed into a standard normal distribution, thus obtaining standardized structured data.

[0029] This application ensures the integrity of data sources by acquiring data in a layered manner, eliminates the misalignment and deviation of multi-source heterogeneous data in the time dimension by using spatiotemporal alignment, and improves the signal-to-noise ratio and stability of data in each layer by differential denoising. On this basis, the data in each layer can form a structured, low-noise, and computable standardized input under a unified time frame, providing a reliable data foundation for the theoretical expected output of dynamic twins, fault residual calculation and subsequent causal inference.

[0030] For example: Suppose we are performing data acquisition and processing on a certain type of 3MW wind turbine: High-frequency acquisition: The system acquires the vibration acceleration signal of the high-speed shaft bearing of the gearbox at a sampling rate of 25kHz. The original signal is mixed with high-frequency noise generated by the electromagnetic interference of the generator.

[0031] Low and medium frequency and environmental data acquisition: The system acquires the generator's three-phase current at 5kHz, and records the output power of 2.5MW, the speed of 1200rpm, the ambient wind speed of 8m / s, and the temperature of 15℃ every 10 seconds.

[0032] Noise reduction: The vibration signal was decomposed into four layers of wavelets, and a large amount of random noise was found in the detail coefficients. After processing with a soft threshold function, the reconstructed signal clearly showed the periodic impact pulses, indicating that there may be wear on the outer ring of the bearing.

[0033] Filling: It was found that the power data at 10:05:10 was missing. Linear interpolation was performed using 2.52MW at 10:05:00 and 2.48MW at 10:05:20 to fill in the missing data at that time, which is 2.5MW.

[0034] Standardization: Vibration data processed from the original range of 0-10g, current data and environmental data from the original range of 0-500A are all scaled to between 0 and 1 using the Min-Max algorithm.

[0035] Finally, a standardized structural data matrix containing vibration characteristics, current characteristics, and operating environment characteristics was generated, ready to be input into the dynamic twin for analysis.

[0036] Step S002: Construct a dynamic twin combining a physical mechanism model and a deep neural network. Use the dynamic twin to process the standardized structural data to obtain theoretical expectations and output fault residuals.

[0037] In a further embodiment, the construction of a dynamic twin combining a physical mechanism model and a deep neural network is as follows: a mathematical model of the device based on physical laws is used as the model foundation, and a dynamic twin is obtained by combining it with a deep neural network. The dynamic twin is used to output theoretical expected values ​​based on standardized structural data, and the fault residuals are obtained by comparing the theoretical expected values ​​with the standardized structural data.

[0038] In a further embodiment, the physical mechanism model is constructed based on the aerodynamic and mechanical dynamic equations of the wind turbine; the deep neural network includes an LSTM long short-term memory network or a GRU gated recurrent unit.

[0039] Understandably, to address the lack of physical interpretability in purely data-driven models and the difficulty of adapting purely physical models to complex and ever-changing operating conditions, resulting in low fault detection accuracy, a hybrid modeling dynamic twin is constructed. In this embodiment, a dynamic twin combining a physical mechanism model and a deep neural network is built. This dynamic twin processes the standardized structural data to obtain theoretical predictions and outputs fault residuals. This design fully utilizes the interpretability of the physical model and the flexibility of the data-driven model. The physical model ensures the basic prediction accuracy under known operating conditions and limits the divergence of the neural network in unreasonable regions, while the deep neural network effectively compensates for the physical model's shortcomings in modeling complex operating conditions. This dual mechanism enables the dynamic twin to accurately track the actual operating state of the wind turbine, and the output fault residuals can sensitively reflect even minor degradation in equipment performance.

[0040] Specific steps: First, the physical mechanism model is constructed based on the aerodynamic and mechanical dynamic equations of the wind turbine. For example, the wind energy conversion formula is used. As a foundation, among which Theoretical output power, air density, It can be calculated in real time based on data from air pressure and temperature sensors. For the area swept by the wind turbine, For wind speed, The wind energy utilization coefficient, Typically, the tip speed ratio and pitch angle The function can be obtained through table lookup or fitting formula. The model receives environmental input. Such as wind speed, air density, and control inputs Examples of parameters, such as pitch angle and torque setpoint, are used to describe the theoretical operating trajectory of the wind turbine in a healthy state, serving as the basis for the dynamic twin.

[0041] Secondly, the deep neural network includes an LSTM (Long Short-Term Memory) network or a GRU (Gated Recurrent Unit) to capture nonlinear residuals that cannot be precisely described by physical models, such as frictional losses in mechanical transmission chains, thermal characteristics of generators, and systematic biases in sensor systems. This network receives all or part of the standardized structured data. This includes high-frequency vibrations, temperature, etc. A complex mapping relationship between the input and the physical model output error is learned through multi-layer nonlinear transformations, serving as a supplement to the dynamic twin to improve approximation accuracy.

[0042] The theoretical expected total output of the dynamic twin is... It is generated by weighted superposition of two parts, and the formula is expressed as:

[0043] in, This represents the output of the physical mechanism model. This represents the corrected output of the deep neural network.

[0044] Finally, the fault residual By calculating the actual observed values Compared with theoretical expectations The deviation is obtained, and the calculation formula is:

[0045] To eliminate the influence of dimensions, the relative error formula can also be used. .when An anomaly is determined when the preset dynamic threshold is exceeded. The actual observed values ​​include real-time standardized data obtained from the SCADA system, such as actual power generation and actual rotational speed.

[0046] In this embodiment, by constructing a dynamic twin that combines physical mechanisms with deep learning, the interpretability of the physical model and the flexibility of the data-driven model are fully utilized. The physical model ensures the accuracy of basic predictions under known operating conditions, limiting the divergence of the neural network in unreasonable regions; while the deep neural network effectively compensates for the shortcomings of the physical model in modeling complex operating conditions. Through this dual mechanism, the dynamic twin can accurately track the actual operating status of the wind turbine, and the output fault residuals can keenly reflect the slight degradation of equipment performance, thereby significantly improving the accuracy and timeliness of fault detection and reducing the false alarm rate.

[0047] For example: Assuming a 2MW wind turbine of a certain model is in operation, the system acquires the current environmental data. With a wind speed of 10 m / s and an air density of Control data The pitch angle is 0 degrees (full engine operation).

[0048] Physical mechanism calculation: based on the formula Substitute the known parameters (assuming) At the optimal tip speed ratio of 0.45, the swept area is... (where the power is a constant), calculate the theoretical power. It is approximately 1800kW.

[0049] Neural Network Correction: Simultaneously, current gearbox vibration characteristics, lubricating oil temperature, and generator winding temperature are input into the trained LSTM model. Based on learned historical patterns, the LSTM model identifies that although the wind speed meets the standard, the low ambient temperature increases lubricating oil viscosity, thus increasing mechanical losses. Therefore, it outputs a correction value. It is -50kW.

[0050] Theoretical expected output: The dynamic twin fuses the above two parts to obtain the theoretical expected power. .

[0051] Residual calculation and fault diagnosis: The system reads the actual power at that moment. The power is 1600kW. Calculate the fault residual. .

[0052] Results Analysis: Under normal operating conditions, the residual error is typically within ±20kW, while a residual error of 150kW far exceeds the preset threshold. Since the physical model and neural network correction have already considered the effects of wind speed and temperature, the remaining huge power loss is most likely caused by a short circuit in the generator stator winding or mechanical jamming of the drive shaft. Therefore, the dynamic twin outputs this fault residual, triggering the subsequent fault warning process.

[0053] Step S003: Establish a mapping relationship between the fault residual and the standardized structure data and output a feature vector. Use context-aware attention to dynamically fuse the standardized structure data to obtain a fused feature vector.

[0054] Understandably, in the complex operating environment of wind farms, the importance of feature data from different sources for fault characterization varies significantly under different operating conditions. For example, vibration signals are extremely sensitive to mechanical faults under low-wind-speed, stable conditions, but under high-wind-speed, turbulent conditions, even normal wind disturbances can cause severe vibrations. If static feature splicing or average weighting is used, key fault features may be submerged by environmental noise, or normal fluctuations may be misjudged as faults. Traditional static fusion methods cannot dynamically distinguish the effectiveness of features based on the real-time operating environment (context), thus reducing the accuracy and robustness of fault diagnosis. In this embodiment, a mapping relationship is established between the fault residual and the standardized structured data, and a feature vector is output. Context-aware attention is used to dynamically fuse the standardized structured data to obtain a fused feature vector. This design achieves adaptive and intelligent feature extraction. The system no longer passively processes all input data but actively focuses on the most sensitive and reliable data features for fault diagnosis based on the current operating environment, while suppressing features that are heavily influenced by the environment or have a lot of noise, greatly improving the signal-to-noise ratio of fault features under complex and variable operating conditions.

[0055] That is, based on the current operating environment (situation), it automatically focuses on the data features most sensitive to fault diagnosis. For example, under high wind speed and turbulent conditions, it automatically reduces the fluctuation weight of vibration signals to avoid false alarms; while under low wind speed and stable conditions, it increases the weight of minor anomalies, thereby significantly improving the ability to express fault features.

[0056] Please refer to the attached document. Figure 3 As shown, the specific steps include: Step S030: Strictly align the fault residuals output in step S002 with the standardized structured data generated in step S001 according to the timestamps. This is to take into account that different data may have different sampling rates, such as vibration signals being high-frequency and SCADA being low-frequency; here, time window aggregation or upsampling methods are used to ensure that the data at each time step can correspond one-to-one, thereby constructing a high-dimensional initial feature vector containing multi-dimensional information.

[0057] Step S031: Introduce a context-aware attention mechanism, selecting state variables that reflect the current operating environment to construct a context vector. This mechanism uses environmental data as context variables, enabling the model to learn to assess the situation. Specifically, it selects state variables that reflect the current operating environment, such as wind speed. turbulence intensity Ambient temperature Etc., construct context vectors The embodiment calculates the attention weights for each feature channel by designing a sub-network containing a context encoder. This sub-network can output a set of weight coefficients based on the current context. The calculation formula is:

[0058] in, The weight matrix is ​​a learnable matrix. For bias terms, The function is used to ensure that the sum of the weights is 1.

[0059] Step S032: Use the calculated attention weights to perform a weighted summation on each channel in the high-dimensional initial feature vector to obtain the fused feature vector.

[0060] The example is to utilize the calculated attention weights. For each channel in the high-dimensional initial feature vector Perform a weighted summation to obtain the fused feature vector. The calculation formula is:

[0061] in, For the first Data from each feature channel. In this way, the model can dynamically adjust its focus on different features such as vibration, current, and temperature according to the environment.

[0062] In this embodiment, by introducing a context-aware attention mechanism, the system no longer passively processes all input data. Instead, it proactively focuses on the data features most sensitive and reliable for fault diagnosis based on the current operating environment, while suppressing features that are heavily influenced by the environment or have a lot of noise. This significantly improves the signal-to-noise ratio of fault features under complex and variable operating conditions, providing high-quality data support for subsequent causal knowledge graph construction and fault inference. This embodiment also achieves adaptive and intelligent feature extraction.

[0063] For example: Scenario 1: High wind speed entry phase.

[0064] The environmental data collected by the system shows a wind speed of 12 m / s and high turbulence intensity (situational variable). (Input attention subnetwork). At this point, normal gusts of wind cause large-amplitude, low-frequency vibrations in the blades and tower. Directly using this vibration data for diagnosis would cause serious interference. After computation, the context-aware attention mechanism assigns lower weights to the vibration signal features, for example... Instead, higher weights are assigned to relatively stable electrical parameters (such as generator current and voltage), for example... The fused feature vectors primarily reflect the stability of the electrical system, avoiding false alarms of "spurious faults" caused by wind disturbances.

[0065] Scenario 2: Low wind speed and stable operation phase.

[0066] Environmental data shows a wind speed of 5 m / s and stable airflow. At this time, if there were a pitting fault in the gearbox bearings, it would produce noticeable abnormal vibrations. The attention subnetwork recognizes the current quiet environment and assigns extremely high weights to the vibration signal features, for example... At this point, the faint fault vibration characteristics in the fused feature vector are amplified, enabling the system to keenly detect this early mechanical fault signal and issue a timely warning.

[0067] Through the above mechanism, regardless of the harsh or complex natural environment in which the wind turbine is located, this step can output the fused feature vector that is most conducive to accurately judging the fault.

[0068] Step S004: Construct a causal knowledge graph based on the fused feature vector, filter key features to eliminate false associations, dynamically adjust the gating weights, perform fault inference based on the dynamically adjusted gating weights, and output fault prediction data.

[0069] Understandably, most existing wind turbine fault diagnosis methods rely on statistical correlation for feature selection and fault determination. However, wind power systems are highly coupled and complex, often containing numerous spurious correlations between variables. For example, increased ambient temperature can lead to increased lubricating oil temperature, but this does not necessarily indicate a cooling system malfunction. Relying solely on correlations can easily misreport non-faulty environmental changes as equipment failures, resulting in a persistently high false alarm rate. Furthermore, wind turbine components experience performance degradation over time, and static diagnostic model parameters struggle to adapt to these dynamic changes, causing the model to fail in the later stages of turbine operation. Therefore, eliminating spurious correlations and constructing adaptive diagnostic models is crucial for improving decision-making accuracy.

[0070] In this embodiment, a causal knowledge graph is constructed based on the fused feature vectors. Key features are screened to eliminate false associations, and the gating weights are dynamically adjusted. Fault inference is performed based on the dynamically adjusted gating weights, and fault prediction data is output. By introducing the causal knowledge graph, the correlation of data mining can be elevated to the causality of logical reasoning, which significantly reduces false alarms caused by environmental coupling factors and improves the interpretability of diagnostic results. At the same time, the dynamic gating weighting mechanism gives the model the ability to adapt to the performance changes of the wind turbine throughout its entire life cycle.

[0071] Please refer to the attached document. Figure 4 The specific steps are as shown: Step S040: Based on historical fault data, maintenance logs, and physical rules constructed by domain experts, construct a causal knowledge graph containing three-layer entities: components, fault modes, and symptoms, to display information about each node and causal relationships.

[0072] Specifically, based on historical fault data, maintenance logs, and physical rules constructed by domain experts, a causal knowledge graph containing three layers of entities—"component-fault mode-symptom"—is pre-built. In this graph, nodes represent specific entities, such as gearboxes, bearing pitting, and abnormal vibration, while edges represent defined causal relationships, such as bearing pitting causing abnormal vibration.

[0073] Step S041: Map the fused feature vector to node attributes or activation states in the knowledge graph, and remove features without causal relationship based on the correlation between nodes in the knowledge graph and potential faulty nodes.

[0074] In a further embodiment, by aggregating information from neighboring nodes, the causal strength between input features and potential faulty nodes is obtained; based on the causal strength threshold, key features with strong causal relationships are selected, while features that only have statistical correlation but no causal relationship are eliminated.

[0075] Specifically, the fused feature vector output in step S003 is mapped to node attributes or activation states in a knowledge graph, and a graph neural network is used for multi-hop message passing on the graph. By aggregating information from neighboring nodes, the causal strength between the input features and potential faulty nodes is calculated. Based on the calculated causal strength threshold, key features with strong causal relationships are selected, such as a strong correlation between vibration at a specific frequency and gear tooth breakage. At the same time, features that only have statistical correlation but no causal relationship are eliminated. For example, if high ambient temperature explains the increase in lubricating oil temperature, the latter's independence as a fault feature is eliminated to avoid misjudging it as a cooling system fault.

[0076] Step S042: Introduce a dynamic gating weighting mechanism to automatically adjust the sensitivity to different fault modes based on the dynamic changes of the current input features.

[0077] Specifically, a dynamic gating weighting mechanism is introduced. This mechanism is similar to the gating logic in a Gated Recurrent Unit (GRU). A gating network is designed, whose input is the filtered key features. The gating network outputs a weight vector in the interval [0,1]. Used to dynamically adjust the parameter weights of the fault inference model. The formula can be expressed as: The model automatically adjusts its sensitivity to different fault modes based on the dynamic changes in the current input features.

[0078] Step S043: Perform fault inference based on the filtered features and the dynamically weighted model.

[0079] Specifically, fault inference is performed based on the selected features and the dynamically weighted model. This is achieved by using a Softmax classifier to output the probability distribution of various faults, such as gear tooth breakage, bearing pitting, and blade icing. ,in This represents the type of fault.

[0080] In this example, by introducing a causal knowledge graph, the correlation from data mining can be elevated to the causality of logical reasoning, significantly reducing false alarms caused by environmental coupling factors and improving the interpretability of diagnostic results. Simultaneously, the dynamic gating weighting mechanism endows the model with the ability to adapt to performance changes throughout the wind turbine's entire lifecycle, ensuring that the model maintains high-precision fault prediction capabilities even after equipment aging or operational condition drift.

[0081] For example: Suppose the system detects two signs in a certain fan: "gearbox temperature rises" and "vibration signal amplitude increases".

[0082] Causal graph analysis: After fusing feature vectors into the GNN, the knowledge graph begins reasoning.

[0083] Path 1: "Increased vibration amplitude" -> Spectrum query -> Strong correlation points to "Gear bearing pitting" or "Broken tooth", with a causal strength of 0.9 (key feature).

[0084] Path 2: "Gearbox temperature rise" -> Graph query -> Related to "Lubrication system failure" or "High ambient temperature".

[0085] Excluding spurious associations: GNN query of the environment nodes revealed that the current "ambient temperature" is within the normal range, and the "cooling system" operating parameters are normal. Therefore, it was determined that there is no causal relationship between "temperature rise" and the external cause of "high ambient temperature," and the causal strength between "temperature rise" and "gear wear" is low (only 0.2). In contrast, the graph shows that "gear wear" leads to frictional heat generation, and the causal strength between "temperature rise" and "gear wear" is 0.7.

[0086] Feature filtering and gating weighting: The system removes noise weights related to the environment and locks "vibration features" and "temperature features" as key features. Since vibration features are more sensitive to gear faults, the dynamic gating mechanism automatically adjusts the inference weight of the vibration channel to 0.8 and the weight of the temperature channel to 0.2.

[0087] Fault inference: The probability distribution output by the final model is as follows: , , .

[0088] Based on this, the system outputs fault prediction data, clearly indicating that the most likely fault is a broken gear tooth, rather than a false alarm of a lubrication failure.

[0089] Step S005: Based on the fault prediction data, perform feature vectorization retrieval from the historical case database to obtain historical fault cases, perform spatiotemporal comparison between the historical fault cases and the fault prediction data to obtain relevant fault cases, and perform confidence calibration on the relevant fault cases.

[0090] Understandably, relying solely on data-driven models, such as the fault probability distribution output in step S400, for decision-making often results in a black box effect. Furthermore, model predictions are typically based on statistical regularities and lack consideration for actual operational experience within specific time and spatial contexts. For instance, certain fault manifestations, such as abnormal vibration, may have drastically different causes in summer and winter, or certain wind turbine models may exhibit specific family-related defects in specific wind farm areas. Without verification using specific historical case backgrounds, this could lead to inflated prediction confidence or maintenance recommendations that are infeasible under specific time and spatial conditions.

[0091] In this embodiment, historical fault cases are obtained by feature vectorization retrieval from the historical case database based on the fault prediction data. The historical fault cases are compared with the fault prediction data in a spatiotemporal manner to obtain relevant fault cases, and the confidence level of the relevant fault cases is calibrated. This design makes full use of the valuable operation and maintenance experience accumulated by the wind farm over a long period of time, corrects the deviation of the pure data model under specific boundary conditions, and makes the fault prediction not only conform to the current signal characteristics, but also to historical patterns and spatiotemporal background.

[0092] Specific steps: Step S050: Establish a historical failure case library. Each case includes multimodal data characteristics before the failure occurred, failure type, handling measures and results.

[0093] Specifically, this involves establishing and maintaining a dynamically updated historical failure case library. Each case in this library is structured as a four-tuple. ,in The feature vector of multimodal data before the fault occurred. For fault type labels, Maintenance measures to be taken, The result is the processing output. Step S400 outputs fault prediction data, including the probability distribution of fault types. After obtaining the corresponding fused feature vector, the system performs vectorized encoding on the fused feature vector and uses it as the query vector. .

[0094] Step S051: Vectorize the fault prediction data features and use cosine similarity or Euclidean distance to search the case library, initially selecting the ten most similar historical cases.

[0095] Specifically, a preliminary search is performed in the historical case database. Cosine similarity or Euclidean distance is then used to calculate the relationship between the query vector and historical feature vectors in the case database. similarity between Based on the similarity scores, the top-K historical failure cases with the highest similarity were initially selected as the candidate set.

[0096] Step S052: Compare the differences between the current case and historical cases in terms of the time and location of occurrence to obtain relevant fault cases.

[0097] In this example, a spatiotemporal comparison mechanism is introduced to refine the selection of the Top-K candidate cases. A spatiotemporal feature vector is defined. ,in This includes the time of occurrence (month / season) and the cumulative operating time of the wind turbine; This includes the geographical region where the wind farm is located (e.g., offshore, mountainous, plains) and the specific wind turbine model. It also calculates the spatiotemporal distance between the current case and historical cases. ,in These are the weighting coefficients.

[0098] Step S053: Adjust the confidence level of the predicted probability based on the comparison results. If the spatiotemporal features are highly consistent, increase the confidence level; otherwise, decrease it.

[0099] In this example, confidence level calibration is required to improve the matching accuracy. Specifically, this is done based on spatiotemporal distance. For the initial prediction probability Make corrections. If the spatiotemporal comparison results show that the current operating conditions highly match a historical case, that is... Less than the preset threshold If the spatiotemporal characteristics are significantly different (e.g., a prediction of "leaf icing" but similar cases found all occur in summer, or in different climate regions), then the confidence level is lowered. The calibrated probability formula can be expressed as: ,in It is a regulating factor.

[0100] For example: Assume that step 104 predicts a 75% probability of "generator bearing wear" occurring on a certain wind turbine.

[0101] Preliminary search: The system retrieved three historical cases of bearing wear with highly similar vibration characteristics from the case database.

[0102] Case A: Occurred 3 years ago in the summer, the wind field was located in a plain area, and the aircraft type was X.

[0103] Case B: Occurred last winter, the wind farm is located in a high-altitude mountainous area, and the turbine model is X.

[0104] Case C: Occurred 6 months ago in winter, the wind farm is located in the same plain area, and the turbine type is X (current wind turbine).

[0105] Spatiotemporal comparison: Current time: Winter, the wind turbine has been operating for 3 years, located in a plain area.

[0106] Compared with Case A: The seasons do not match (summer vs. winter), and the distance is too far.

[0107] Compared with Case B: The season is the same, but the geographical environment is very different (mountainous area vs. plain), resulting in different heat dissipation and load characteristics.

[0108] Comparison with Case C: Same season, same geographical location, same aircraft type, similar runtime. Spatial-temporal distance. Extremely small, highly consistent.

[0109] Confidence calibration: Due to the existence of Case C, the system confirms that under this specific spatiotemporal context (winter plain environment), Type X wind turbines are indeed prone to this type of bearing wear, and the historical characteristics match well. Therefore, the system calibrates and increases the prediction confidence of "generator bearing wear" from 75% to 88%, and prioritizes the maintenance measures in Case C (such as replacing with a specific type of low-temperature grease) to generate decision recommendations.

[0110] Conversely, if similar cases found all occur in summer, the system will determine that the current vibration abnormality may be due to abnormal oil temperature and viscosity caused by low winter temperatures rather than actual wear, thereby reducing the probability of failure and suggesting that the heating system be observed first.

[0111] Step S006: Generate a causal tracing report based on standardized structured data, causal knowledge graphs, and relevant failure cases; and formulate multi-dimensional decision-making recommendations based on failure prediction data and relevant failure cases.

[0112] Understandably, while the system has already achieved high-precision fault prediction and confidence calibration in the preceding steps, the output data formats, such as probability distributions and feature vectors, remain too abstract for on-site maintenance personnel or intelligent decision-making terminals. Maintenance personnel not only need to know "what fault occurred," but also "why it happened" (root cause analysis) and "what to do" (decision execution). A lack of intuitive causal explanations and clear operational guidelines can lead to slow maintenance responses, inappropriate measures, and even safety accidents caused by misjudgments. Therefore, transforming complex model calculation results into interpretable and executable maintenance solutions is crucial for achieving intelligent closed-loop management of unmanned wind farms.

[0113] In this embodiment, a causal attribution report is generated based on standardized structured data, causal knowledge graphs, and relevant fault cases. Multidimensional decision-making suggestions are formed based on fault prediction data and relevant fault cases. This design realizes a closed loop from "data perception" to "intelligent decision-making". By generating a visual report containing causal logic, the application threshold of artificial intelligence models is greatly reduced, enabling operation and maintenance personnel to quickly understand the nature of the fault and trust the system's judgment. At the same time, the multidimensional decision-making suggestions are based on historical experience and real-time risk assessment, which can help operation and maintenance personnel find the optimal balance between safety, efficiency, and cost.

[0114] Specifically: First, by integrating the changing trends of key indicators in the standardized structural data, the causal reasoning paths in the causal knowledge graph, and relevant historical cases after calibration, a causal tracing report is generated. The causal tracing report includes the root cause of the failure, the evolution process, and the basis for the failure. In this example, by extracting the changing trends of key indicators from standardized structural data, such as temperature rise curves over time and the evolution of vibration spectra, and combining them with causal path chains inferred from a causal knowledge graph (e.g., blade icing, mass imbalance, increased vibration leading to load fluctuations), as well as factual descriptions from relevant historical cases after calibration, a richly illustrated causal attribution report is automatically constructed using Natural Language Generation (NLG) technology. The report includes visual charts, such as trend graphs and causal topology diagrams, detailing the root cause of the failure, its evolution, and the basis for the judgment.

[0115] Secondly, based on the severity of the fault prediction data and the successful handling experience of relevant historical cases, multi-dimensional decision-making suggestions are generated, including: immediate shutdown for maintenance, reduced operation for observation, adjustment of control parameters, or continued enhanced monitoring.

[0116] In this example, based on the probability distribution and confidence level of the fault prediction, combined with the real-time load of the wind turbine, the risk level of the fault is calculated. Subsequently, successful handling measures and their effects from relevant historical cases are retrieved, and multi-dimensional decision suggestions are generated using decision trees or reinforcement learning strategies. These decision suggestions include, but are not limited to: 1. Immediate shutdown for inspection: When the predicted probability of failure is extremely high (e.g., >90%) and the nature of the failure is catastrophic (e.g., broken gear teeth in the gearbox, fire risk), it is recommended to shut down the unit immediately to avoid damage to the unit.

[0117] 2. Reduced-rated operation observation: When the fault is confirmed to be of the gradual change type (such as early bearing wear) or the environment-induced type (such as temporary over-limit caused by high turbulence), and the risk is controllable, it is recommended to reduce the active power and speed, reduce the mechanical load, and keep the fan running under low load for close observation.

[0118] 3. Adjust control parameters: For specific faults (such as blade icing or wind deviation), it is recommended to automatically adjust the control logic, such as adjusting the pitch angle to reduce load, or activating the active yaw de-icing strategy.

[0119] 4. Continue to strengthen monitoring: When the probability of failure is low or the confidence level is insufficient, it is recommended not to intervene, but the sampling frequency should be increased and the inspection cycle should be shortened.

[0120] For example: Assume that step S005 confirms that the wind turbine has experienced a "blade icing" fault, with a confidence level of 92%.

[0121] Generate a causal attribution report: The system extracted the abnormal upward curves of "power fluctuation coefficient" and "vibration acceleration" over the past hour.

[0122] Based on the causal graph, the path was extracted: "low environmental temperature" High humidity -> ice formation on blade surface -> changes in aerodynamic shape -> mass imbalance -> abnormal vibration.

[0123] Matching historical cases: Referencing a similar case from Unit #15 of the same wind farm in 2022.

[0124] The output report states: "Ice formation on the blades due to low temperature and high humidity caused aerodynamic imbalance, resulting in excessive forward and backward vibration of the nacelle. Based on trend analysis, the icing rate is 5% per hour, and the vibration frequency coincides with the first natural frequency, posing a risk of resonance." (Vibration time-domain waveform and causal reasoning topology diagram are attached).

[0125] Formulate multi-dimensional decision-making recommendations: The system assesses the current wind speed as 8 m / s, and the vibration value is close to the shutdown threshold.

[0126] Recommendation 1 (preferred): Activate the de-icing system (hot air or pneumatic pulse) and adjust the pitch angle to feather position to reduce aerodynamic load. Observe whether the vibration decreases.

[0127] Recommendation 2 (Alternative): If the vibration does not improve significantly 10 minutes after the de-icing system is started, it is recommended to immediately execute the shutdown command to prevent blade breakage or tower collapse.

[0128] Recommendation 3 (follow-up): After the ambient temperature rises or de-icing is completed, it is recommended to conduct an external inspection of the blades and confirm that there is no damage before reconnecting to the grid.

[0129] The system pushed the above report and recommendations to the central control room, guiding maintenance personnel to perform de-icing operations and successfully preventing the fault from escalating.

[0130] The aforementioned management decision-making method first utilizes multimodal data fusion and dynamic twin technology to accurately capture fault characteristics. Then, it uses causal knowledge graphs to eliminate false associations, ensuring the accuracy of fault inference. Combining spatiotemporal comparison and confidence calibration with a historical case database further enhances the robustness of the prediction results. The resulting causal tracing report and multidimensional decision recommendations not only solve the problems of "difficult detection, slow diagnosis, and chaotic decision-making" in unmanned wind farm faults, but also significantly improve the intelligent operation and maintenance level and economic benefits of wind farms, demonstrating significant engineering application value.

[0131] For inventions based on the same inventive concept, please refer to the appendix. Figure 5 As shown, the second aspect discloses a management and decision-making device for unmanned wind farms based on multi-source data fusion. Data acquisition module 100: used to acquire multimodal data and preprocess it to obtain standardized structured data. The multimodal data includes high-frequency data, medium-frequency data, low-frequency data and environmental data. Fault Residual Acquisition Module 200: Used to construct a dynamic twin combining a physical mechanism model and a deep neural network, and to process standardized structure data using the dynamic twin to obtain theoretical expectations and output fault residuals; Feature vector fusion module 300: It is used to establish a mapping relationship between fault residuals and standardized structured data and output feature vectors. It uses context-aware attention to dynamically fuse standardized structured data to obtain fused feature vectors. Fault prediction module 400: It is used to construct a causal knowledge graph based on the fused feature vector, filter key features to eliminate false associations, dynamically adjust the gating weights, perform fault inference based on the dynamically adjusted gating weights, and output fault prediction data. Fault Case Acquisition Module 500: Used to retrieve historical fault cases from the historical case database based on fault prediction data using feature vectorization, compare historical fault cases with fault prediction data in a spatiotemporal manner to obtain relevant fault cases, and calibrate the confidence level of relevant fault cases. Report generation module 600: Used to generate causal tracing reports based on standardized structured data, causal knowledge graphs and relevant failure cases, and to form multi-dimensional decision-making suggestions based on failure prediction data and relevant failure cases.

[0132] The various modules in the aforementioned multi-source data fusion management and decision-making device for unmanned wind farms can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0133] In one embodiment, a computer-readable storage medium is provided, the computer device being a server, the internal structure of which can be as shown in the figure. Figure 6 As shown, the computer device includes a processor, memory, and network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores news data and data such as time decay factors. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements a multi-source data fusion management decision-making method for unmanned wind farms.

[0134] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0135] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. This disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims. Thus, if these modifications and variations of the invention fall within the scope of the claims of the invention and their equivalents, the invention is also intended to include these modifications and variations.

Claims

1. A management decision-making method for unmanned wind farms based on multi-source data fusion, characterized in that, Includes the following steps: Multimodal data is acquired and preprocessed to obtain standardized structured data, including high-frequency data, mid-frequency data, low-frequency data, and environmental data; A dynamic twin combining a physical mechanism model and a deep neural network is constructed. The standardized structural data is processed using the dynamic twin to obtain theoretical predictions and output fault residuals. A mapping relationship is established between the fault residuals and the standardized structure data, and a feature vector is output. The standardized structure data is dynamically fused using context-aware attention to obtain the fused feature vector. A causal knowledge graph is constructed based on the fused feature vectors, key features are screened to eliminate false associations, and the gating weights are dynamically adjusted. Fault inference is performed based on the dynamically adjusted gating weights, and fault prediction data is output. Based on the fault prediction data, historical fault cases are obtained by feature vectorization retrieval from the historical case database. The historical fault cases are then compared with the fault prediction data in a spatiotemporal manner to obtain relevant fault cases, and the confidence level of the relevant fault cases is calibrated. A causal attribution report is generated based on standardized structured data, causal knowledge graphs, and relevant failure cases. Multidimensional decision-making recommendations are formed based on failure prediction data and relevant failure cases.

2. The management decision-making method according to claim 1, characterized in that, The steps of acquiring multimodal data and preprocessing it to obtain standardized structured data are: acquiring multimodal data of the high-frequency physical layer, the mid-frequency process layer, the low-frequency control layer, and the environmental layer; Next, the acquired multimodal data is preprocessed with spatiotemporal alignment and denoising.

3. The management decision-making method according to claim 1, characterized in that, The construction of a dynamic twin combining a physical mechanism model and a deep neural network involves: using a mathematical model of the device based on physical laws as the model foundation, and combining it with a deep neural network to obtain a dynamic twin. The dynamic twin is used to output theoretical expected values ​​based on standardized structural data, and the theoretical expected values ​​are compared with the standardized structural data to obtain fault residuals.

4. The management decision-making method according to claim 3, characterized in that, The physical mechanism model is constructed based on the aerodynamic and mechanical dynamic equations of the wind turbine; the deep neural network includes an LSTM long short-term memory network or a GRU gated recurrent unit.

5. The management decision-making method according to claim 1, characterized in that, The process involves establishing a mapping relationship between the fault residuals and the standardized structured data, outputting a feature vector, and then dynamically fusing the standardized structured data using context-aware attention to obtain the fused feature vector: The fault residuals are strictly aligned with the standardized structure data according to timestamps; Introducing a context-aware attention mechanism, a context vector is constructed by selecting state variables that reflect the current operating environment. A context encoder is used to calculate the attention weights for each feature channel; The calculated attention weights are used to sum the values ​​of each channel in the high-dimensional initial feature vector to obtain the fused feature vector.

6. The management decision-making method according to claim 1, characterized in that, The process involves constructing a causal knowledge graph based on the fused feature vectors, filtering key features to eliminate false associations, dynamically adjusting the gating weights, performing fault inference based on the dynamically adjusted gating weights, and outputting fault prediction data. Based on historical fault data, maintenance logs, and physical rules constructed by domain experts, a causal knowledge graph containing three layers of entities—components, fault modes, and symptoms—is built to display information about each node and causal relationships. The fused feature vectors are mapped to node attributes or activation states in the knowledge graph, and features without causal relationship are removed based on the correlation between nodes in the knowledge graph and potential faulty nodes. A dynamic gating weighting mechanism is introduced to automatically adjust the sensitivity to different fault modes based on the dynamic changes in the current input characteristics. Fault inference is performed based on the selected features and the dynamically weighted model.

7. The management decision-making method according to claim 6, characterized in that, The method of removing features without causal relationships based on the correlation between nodes in the knowledge graph and potential faulty nodes is as follows: By aggregating information from neighboring nodes, the causal strength between input features and potential faulty nodes is obtained; based on the causal strength threshold, key features with strong causal relationships are selected, while features that only have statistical correlation but no causal relationship are removed.

8. The management decision-making method according to claim 1, characterized in that, The steps are as follows: First, retrieve historical fault cases from the historical case database using feature vectorization based on the fault prediction data. Then, perform a spatiotemporal comparison between the historical fault cases and the fault prediction data to obtain relevant fault cases. Finally, calibrate the confidence level of the relevant fault cases. Establish a historical failure case database, with each case containing multimodal data characteristics before the failure occurred, failure type, handling measures, and results; The fault prediction data features are vectorized, and cosine similarity or Euclidean distance is used to search the case library to initially select the ten most similar historical cases. By comparing the differences in the time and location of occurrence between current cases and historical cases, relevant failure cases can be obtained; The confidence level of the predicted probability is calibrated based on the comparison results. If the spatiotemporal features are highly consistent, the confidence level is increased; otherwise, it is decreased.

9. The management decision-making method according to claim 1, characterized in that, A causal attribution report is generated based on standardized structured data, causal knowledge graphs, and relevant failure cases. Multidimensional decision-making recommendations are then formed based on failure prediction data and relevant failure cases. By integrating the changing trends of key indicators in the standardized structural data, the causal reasoning paths in the causal knowledge graph, and relevant historical cases after calibration, a causal tracing report is generated, which includes the root cause of the failure, the evolution process, and the basis for the failure. Based on the severity of the fault prediction data and the successful handling experience of relevant historical cases, multi-dimensional decision-making suggestions are generated. These suggestions include: immediate shutdown for maintenance, reduced operation for observation, adjustment of control parameters, or continued enhanced monitoring.

10. A management and decision-making device for unmanned wind farms based on multi-source data fusion, characterized in that, Includes the following modules: Data acquisition module: used to acquire multimodal data and preprocess it to obtain standardized structured data, wherein the multimodal data includes high-frequency data, mid-frequency data, low-frequency data and environmental data; Fault Residual Acquisition Module: Used to construct a dynamic twin combining a physical mechanism model and a deep neural network, and to process the standardized structural data using the dynamic twin to obtain theoretical expectations and output fault residuals; Feature vector fusion module: used to establish a mapping relationship between fault residuals and the standardized structure data and output feature vectors, and dynamically fuse the standardized structure data using context-aware attention to obtain fused feature vectors; Fault prediction module: used to construct a causal knowledge graph based on the fused feature vector, filter key features to exclude false associations, dynamically adjust the gating weights, perform fault inference based on the dynamically adjusted gating weights, and output fault prediction data; Fault Case Acquisition Module: Used to retrieve historical fault cases from the historical case database based on the fault prediction data using feature vectorization, compare the historical fault cases with the fault prediction data in a spatiotemporal manner to obtain relevant fault cases, and calibrate the confidence level of the relevant fault cases. Report generation module: Used to generate causal tracing reports based on standardized structured data, causal knowledge graphs and relevant failure cases, and to form multi-dimensional decision-making suggestions based on failure prediction data and relevant failure cases.

Images