Wind power plant layered multi-source data fusion processing method and system based on edge calculation

Abstract

The invention discloses a wind power plant layered multi-source data fusion processing method and system based on edge calculation. The method comprises the steps of collecting multi-source data; performing timestamp alignment on the multi-source data by adopting a sliding window alignment algorithm in combination with an interpolation method, calculating an arithmetic mean value of the data in a window through a sliding window with a preset unified period, and outputting a unified frequency data set; taking a uniform frequency data set output by the data layer as input, extracting unit operation event information, monitoring equipment alarm information, time domain features, frequency domain features and time frequency domain features, and performing fusion processing on the extracted features by adopting attention mechanism fusion, association rule mining and clustering fusion algorithms to generate feature information; and feature information output by the feature layer and real-time operation data of the SCADA system are obtained, and a control instruction is generated based on multi-level threshold setting, performance degradation index evaluation, decision confidence judgment and economical efficiency evaluation. According to the invention, the problems of data island, shallow fusion level and resource waste in the prior art are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wind turbine units, and particularly relates to a wind farm hierarchical multi-source data fusion processing method and system based on edge computing. BACKGROUND

[0002] With the rapid development of wind power technology and the continuous increase of wind power installed capacity, a large number of monitoring equipment such as SCADA systems, vibration sensors and laser radars are deployed on wind turbine units, and a large amount of multi-source data is formed by real-time collection of multiple types of data, which puts forward higher requirements on processing technology.

[0003] The limitations of the existing data fusion method are as follows: 1. Data island problem: state, vibration, weather and other types of data belong to independent systems, lack of correlation analysis, and cannot fully understand the health status of the equipment.

[0004] 2. Shallow fusion level: data fusion is mostly performed in the cloud after the event, which cannot be quickly completed locally on the wind turbine unit, and it is difficult to discover early faults and link intervention in time.

[0005] 3. Resource waste: a large amount of normal data without exception is uploaded with core data, occupying cloud computing and storage resources, and increasing management costs.

[0006] Therefore, it is necessary to develop a new wind farm hierarchical multi-source data fusion processing method and system based on edge computing. SUMMARY

[0007] The purpose of the application is to provide a wind farm hierarchical multi-source data fusion processing method and system based on edge computing, which can solve the problems of data island, shallow fusion level and resource waste.

[0008] In the first aspect, the wind farm hierarchical multi-source data fusion processing method based on edge computing comprises the following steps: Data layer fusion step: deploy edge processing devices, realize partition access of wind farm field layer multi-source data acquisition equipment through independent network cards, the multi-source data includes unit operation data, environmental weather data, equipment state monitoring data, maintenance management data and power grid operation data; use a sliding window alignment algorithm combined with an interpolation method to align the time stamps of multi-source data with different collection frequencies, calculate the arithmetic mean value of data in the window through a preset sliding window with a unified period, and output a unified frequency data set; Feature layer fusion step: Taking the unified frequency dataset output from the data layer as input, extract unit operation event information, monitoring equipment alarm information, time domain features, frequency domain features, and time-frequency domain features. Use attention mechanism fusion, association rule mining, and clustering fusion algorithms to fuse the extracted features and generate feature information. The feature information includes feature vectors, status identifier sets, feature rule sets, and feature importance ranking results. The decision-making layer fusion steps are as follows: acquire the feature information output by the feature layer and the real-time operation data of the SCADA system; generate control commands, maintenance strategies, performance optimization schemes and early warning information based on multi-level threshold settings, performance degradation index evaluation, decision confidence judgment and economic evaluation; and send the decision execution results back to the data layer through a feedback mechanism to achieve closed-loop optimization.

[0009] One possible implementation, in the data layer fusion step, includes the following specific implementation of the sliding window alignment algorithm: When a new data point flows into the window, it is determined whether the window is full. If the window is full, the oldest data point is removed and the new data point is added. The arithmetic mean of all data in the window is calculated. If the window is not full, the arithmetic mean of all data in the current window is directly calculated and the arithmetic mean is output as uniform frequency data.

[0010] The above technical solution clearly defines the execution logic of the sliding window algorithm, ensuring accurate alignment of data from different acquisition frequencies and resolving the issue of asynchronous time for multi-source data. By dynamically updating the window data and calculating the mean, a unified frequency dataset is output, providing a standardized data foundation for subsequent feature extraction, improving the consistency and reliability of data fusion, and ensuring the accuracy of upper-level analysis results.

[0011] In one possible implementation, during the data layer fusion step, for monitoring needs below 100Hz, blade root load, tower top displacement, tower load, and tilt angle parameters are collected using blade root load monitoring equipment, BeiDou positioning and orientation equipment for air clearance, tower load sensing equipment, and tilt angle sensing equipment. For high-frequency data needs from 1Hz to 100kHz, high-frequency data acquisition is completed using a CMS system, a transmission chain vibration monitoring system, a blade vibration monitoring system, and oil monitoring sensors. Simultaneously, video and audio monitoring equipment are used to achieve visualization and acoustic signal monitoring. Image processing technology is employed for blade clearance measurement and identification of abnormal features such as abnormal high temperatures or fires. Acoustic signal processing involves converting the audio signal into a Mel spectrum using Fast Fourier Transform to analyze energy change characteristics.

[0012] In the above technical solution, dedicated data acquisition equipment is matched according to frequency requirements to achieve full coverage monitoring of key unit parameters. High-frequency data accurately captures subtle equipment anomalies, low-frequency data ensures comprehensive perception of operational status, and audio and video monitoring supplements visual information. Abundant data sources provide sufficient basis for subsequent data integration, improve the ability to identify potential equipment faults, and reduce the risk of missed detections.

[0013] In one possible implementation, the feature extraction step in the feature layer fusion step specifically includes: The time-domain features are statistical characteristics of mean, variance, peak value, and kurtosis; The frequency domain features are the spectral features and envelope spectrum features obtained through FFT transformation; The time-frequency domain features are obtained through wavelet transform or short-time Fourier transform; Unit operation event information includes fault information, unit operating status, and warning information.

[0014] In the above technical solution, the system extracts multi-dimensional features, covering event information and multi-domain statistical features, to comprehensively depict the equipment's operating status. Time-domain, frequency-domain, and time-frequency-domain features are complementary, enabling the capture of different types of operational anomalies. By combining event information to establish a correlation between features and equipment status, rich input is provided for subsequent fusion algorithms, improving the effectiveness and relevance of the feature layer's output results.

[0015] In one possible implementation, in the feature layer fusion step, the association rule mining uses an experience base combined with the FP-Growth algorithm to mine the association relationships between features and form a feature rule set; Cluster fusion employs feature compression clustering or k-nearest neighbor method to perform cluster analysis on features and identify unit operating status patterns; Attention mechanism fusion highlights key features and suppresses secondary features by calculating attention weights for different features.

[0016] In the above technical solution, the three algorithms work synergistically: the attention mechanism focuses on core features, enhancing the influence of key information; association rule mining discovers hidden relationships between features, forming knowledge rules; and clustering fusion achieves state pattern recognition. This algorithmic combination improves the depth and accuracy of feature fusion, generating feature information of appropriate dimensionality, providing high-quality input to the decision-making layer, and reducing the difficulty of decision-making.

[0017] In one possible implementation, the performance degradation index in the decision-level fusion step is calculated based on principal component analysis, specifically as follows: The system comprehensively evaluates CMS online monitoring and early warning information, voiceprint sensor system data, video early warning information, unit power curve, blade root load assessment results, and number of faults, and performs weighted calculations according to preset weights. When the calculation results exceed preset thresholds, the preventive maintenance process is triggered.

[0018] In the above technical solution, the PCA-based performance degradation index is combined with multi-source assessment data to achieve a quantitative assessment of equipment status through weighted calculation. The participation of multi-dimensional data in the assessment improves the comprehensiveness of the results, and the preset threshold triggering of maintenance processes makes operation and maintenance more targeted. It can detect equipment performance degradation trends in advance, transforming passive maintenance into proactive prevention and reducing downtime due to failures.

[0019] In one possible implementation, in the decision-level fusion step, a decision confidence threshold is set. When the decision confidence is lower than the decision confidence threshold, the system does not perform automatic decision-making and triggers a manual intervention request or a supplementary data collection task.

[0020] In the above technical solution, setting the decision confidence threshold ensures decision reliability and avoids erroneous decisions due to data issues. When the confidence level is insufficient, an intervention or supplementary data collection mechanism is triggered. This prevents the system from blindly executing instructions and causing risks, while also improving the accuracy of subsequent decisions by supplementing data, thus balancing the efficiency and security of automated decision-making.

[0021] In one possible implementation, the economic assessment in the decision-making fusion step is achieved through power allocation adjustment. For faulty units with low health factors, the total power demand of the wind farm at the next moment is allocated according to the proportion of the current available power of each healthy unit to the total available power of the entire farm.

[0022] In the above technical solution, power is dynamically allocated based on the health status of the units, avoiding overload operation of faulty units while fully utilizing the power generation capacity of healthy units. This approach optimizes the power resource allocation of wind farms, improves overall power generation efficiency, reduces losses from faulty units, lowers maintenance costs, and achieves a balance between operational economy and equipment safety.

[0023] One possible implementation method is that the feedback mechanism of the decision-making layer fusion step is as follows: compare the decision execution results with the actual operating effect of the unit, optimize the decision model and various threshold settings, re-input the feedback data into the data layer, and perform a new round of data fusion cycle.

[0024] In the above technical solution, the feedback mechanism forms a closed-loop optimization system. It verifies the rationality of decisions through actual results, continuously optimizes models and thresholds, and improves the system's decision-making capabilities. Feedback data return makes data fusion cycles more targeted, enabling the system to self-iterate and upgrade. As operating time increases, data processing and decision-making accuracy continuously improve, adapting to the complex operating environment of wind farms.

[0025] Secondly, the wind farm hierarchical multi-source data fusion processing device of the present invention includes a memory and a processor. The memory stores a computer-readable program, which, when called by the processor, can execute the wind farm hierarchical multi-source data fusion processing method as described in the present invention.

[0026] The present invention has the following technical effects: Addressing the issues of data silos, shallow integration, and resource waste in existing systems, this invention breaks down data barriers through a layered integration architecture, enabling multi-source data correlation analysis. Edge computing supports rapid local processing, avoiding cloud latency and allowing for timely detection of early faults. Data layer preprocessing and feature layer extraction reduce invalid data uploads and lower cloud resource consumption. Bottom-up aggregation and closed-loop feedback comprehensively improve device health awareness and decision-making timeliness, reducing operation and maintenance costs. Attached Figure Description

[0027] Figure 1 This is a hardware architecture diagram of the wind farm hierarchical multi-source data fusion processing method based on edge computing described in the embodiments of this application; Figure 2 This is a flowchart of the wind farm hierarchical multi-source data fusion processing method based on edge computing described in the embodiments of this application; Figure 3 This is a flowchart of the sliding window alignment algorithm in the edge computing-based wind farm hierarchical multi-source data fusion processing method described in the embodiments of this application; Figure 4 This is a block diagram illustrating the principle of data layer fusion in the edge computing-based multi-source data fusion processing method for wind farms described in the embodiments of this application. Figure 5 This is a schematic diagram of the principle of the edge computing-based wind farm hierarchical multi-source data fusion processing system described in the embodiments of this application. Detailed Implementation

[0028] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention and not for limiting the scope of protection of the present invention.

[0029] like Figure 1 and Figure 2As shown in the figure, this application discloses a method for hierarchical multi-source data fusion processing of wind farms based on edge computing. The method is based on a bottom-up data aggregation architecture and realizes the processing and application of multi-source data of wind farms through hierarchical fusion of data layer, feature layer, and decision layer. Specifically, it includes the following steps: Data layer fusion steps: Deploy edge processing devices to enable partitioned access of multi-source data acquisition devices at the wind farm site layer through independent network cards. The multi-source data includes unit operation data, environmental meteorological data, equipment status monitoring data, maintenance management data, and power grid operation data. Use a sliding window alignment algorithm combined with interpolation to align the timestamps of multi-source data with different acquisition frequencies. Calculate the arithmetic mean of the data within the window through a sliding window with a preset uniform period, and output a uniform frequency dataset.

[0030] Feature layer fusion steps: Taking the unified frequency dataset output from the data layer as input, extract unit operation event information, monitoring equipment alarm information, time domain features, frequency domain features, and time-frequency domain features. Use attention mechanism fusion, association rule mining, and clustering fusion algorithms to fuse the extracted features and generate feature information. Feature information includes feature vectors, status identifier sets, feature rule sets, and feature importance ranking results.

[0031] The decision-making layer fusion steps are as follows: acquire the feature information output by the feature layer and the real-time operation data of the SCADA system; generate control commands, maintenance strategies, performance optimization schemes and early warning information based on multi-level threshold settings, performance degradation index evaluation, decision confidence judgment and economic evaluation; and send the decision execution results back to the data layer through a feedback mechanism to achieve closed-loop optimization.

[0032] The following is a detailed description of this application: like Figure 1 As shown, the overall system architecture includes multi-source data acquisition devices, edge processing devices, and a cloud platform. The multi-source data acquisition devices are used for data collection, while the edge processing devices are used for short-term, real-time, high-frequency, and intelligent control. The cloud platform is used for long-term, trend prediction, lifetime prediction, diagnostic functions, and safety margin evaluation.

[0033] The overall data flow of the system exhibits a bottom-up convergence process: Raw data is collected by multi-source data acquisition devices deployed on-site and transmitted to edge processing devices via the wind farm's internal network. After preliminary processing, the data enters the feature layer for feature extraction and fusion. Finally, a comprehensive evaluation result and operational instructions are generated at the decision layer. Throughout the process, the data undergoes a transformation from concrete to abstract and from local to global.

[0034] This application constructs a cloud-edge-device collaborative data processing architecture. The wind farm's hierarchical multi-source data fusion architecture includes data layer fusion, feature layer fusion, and decision layer fusion. This hierarchical structure follows the principle of progressive refinement, gradually extracting and condensing information from raw data acquisition to advanced decision generation. It effectively integrates multi-source heterogeneous data such as SCADA (Supervisory Control and Data Acquisition) data, CMS (Condition Monitoring System) information, meteorological information, and equipment condition monitoring data, significantly improving the wind farm's operating efficiency, fault prediction capabilities, and decision-making intelligence level.

[0035] I. Data Layer Fusion Steps 1. The data sources for the wind farm multi-source data fusion system are extremely extensive, including: 1) Operating data of the wind turbine generator set: power output, speed, pitch angle, yaw position, wind speed, wind direction, etc., are usually collected by the SCADA system, with frequencies ranging from 1Hz to 100kHz.

[0036] 2) Environmental meteorological data: temperature, air pressure, humidity, etc., collected at a frequency of 1 to 10 minutes per time.

[0037] 3) Equipment condition monitoring data: vibration (transmission chain vibration, blade vibration, etc.), temperature, oil analysis, etc., with a high sampling rate (up to 100kHz).

[0038] 4) Maintenance and management data: fault records, repair history, component replacement records, etc.

[0039] 5) Power grid operation data: voltage, current, frequency, power factor, etc., with the acquisition frequency usually at the second or minute level.

[0040] 2. Data Layer Fusion: The edge processing device enables different devices to access different partitions through independent network interface cards (NICs) without affecting the original data network transmission. Each device resolves the issue of inconsistent data acquisition frequencies from multiple sources through data timestamp alignment and buffering mechanisms.

[0041] For data of different frequencies, the time alignment method mainly adopts the sliding window alignment algorithm, which dynamically adjusts the size of the alignment window according to the data timestamp and the acquisition frequency to achieve high-precision time alignment of multi-source data. For high-frequency data (such as vibration data), interpolation is used to achieve time alignment with low-frequency data (such as SCADA data).

[0042] Using a sliding window with a preset uniform period, calculate the arithmetic mean (SMA) of the aligned data and output a uniform frequency dataset: SMA = (x1 + x2 + ... + x n ) / n Where, x1 to x n'n' represents the data value within the window. Different sampling frequencies have different input periods, so the corresponding window data will have different periods. The shorter the sampling period, the smaller the moving average can be calculated. For sampling periods of 1s, 10s, etc., only the average value on the order of minutes or 10 minutes can be calculated.

[0043] 3. In the data layer fusion step, the specific implementation of the sliding window alignment algorithm includes: starting the calculation; when a new data point flows into the window, determining whether the window is full; if so, the oldest data point is moved out of the window, the new data point is added to the window, and the arithmetic mean of all data in the new window is calculated; otherwise, the arithmetic mean of all data in the current window is calculated; and the output arithmetic mean is used as the unified frequency data, see [link to relevant documentation]. Figure 3 .

[0044] 4. In the data fusion step, for monitoring needs below 100Hz, blade root load monitoring equipment, airspace Beidou positioning and orientation equipment, tower load sensing equipment, and tilt angle sensing equipment are added to collect parameters such as turbine blade root load, tower top displacement, tower load, and tilt angle. For high-frequency data acquisition needs from 1Hz to 100kHz, independent high-frequency monitoring systems are implemented, including the CMS system, drivetrain vibration monitoring, blade vibration monitoring, pitch bolt preload monitoring, and tower bolt preload monitoring. These are combined with oil monitoring and displacement monitoring sensors to achieve accurate high-frequency data acquisition. Simultaneously, video and audio monitoring equipment are integrated to achieve real-time video monitoring and audio data acquisition, comprehensively covering visualization and sound signal monitoring scenarios. (See [link to relevant documentation]). Figure 4 .

[0045] II. Feature Layer Fusion Steps Feature layer fusion is the intermediate processing layer of the wind farm data fusion system. It is mainly responsible for extracting valuable feature information from the preprocessed data and performing feature-level fusion. The input data is the standardized dataset output by the data layer fusion.

[0046] 1. In the feature layer fusion step, feature extraction specifically includes: The time-domain features are statistical characteristics of mean, variance, peak value, and kurtosis.

[0047] The frequency domain features are the spectral features and envelope spectrum features obtained through FFT transformation.

[0048] The time-frequency domain features are obtained through wavelet transform or short-time Fourier transform.

[0049] Unit operation event information includes fault information, unit operating status, and warning information.

[0050] 2. Feature fusion algorithm and logical judgment: Fluctuation threshold judgment: For core operating parameters such as speed and power, fluctuation threshold judgment is performed to identify whether there are abnormal situations where fluctuations exceed the standard.

[0051] Fault information association: When the unit experiences faults such as excessive vibration, key operating data such as pitch, yaw, speed, transmission chain CMS vibration, blade CMS vibration and load are collected simultaneously to establish the correspondence between faults and associated parameters.

[0052] Feature importance assessment: Based on the fused key data such as rotational speed frequency domain features and vibration frequency domain features, the importance of each feature is quantitatively assessed.

[0053] In the feature layer fusion step, the association rule mining adopts an experience base combined with the FP-Growth algorithm to mine the association relationships between features and form a feature rule set.

[0054] Cluster fusion employs feature compression clustering or k-nearest neighbor method to perform cluster analysis on features and identify unit operating status patterns.

[0055] Attention mechanism fusion highlights key features and suppresses secondary features by calculating attention weights for different features.

[0056] 3. The output data of feature layer fusion is a high-dimensional feature vector or feature identifier set, including: Feature vector: A multidimensional feature array that represents the operating status of a device.

[0057] Status label: Status label based on feature clustering or classification results (such as normal, abnormal, degenerate, etc.).

[0058] Feature rule set: Association rules and confidence scores between features (mined using algorithms such as FP-Growth).

[0059] Feature importance ranking: Ranking of the contribution of each feature to the state assessment.

[0060] The output data format is typically a feature matrix, a feature description file in JSON format, or a compressed file, containing information such as feature name, feature value, time range, and confidence level. This layer's output provides refined and condensed feature information for the decision-level fusion, significantly reducing data dimensionality and improving processing efficiency.

[0061] III. Steps for Integrating Decision-Making Levels 1. Data source: Directly obtained from the SCADA system, including operating data such as power output, speed, pitch angle, yaw position, wind speed, and wind direction of the wind turbine generator, with a collection frequency of 1Hz~100kHz (covering data at the second and minute levels).

[0062] Key feature information provided by the edge processor includes frequency domain features, time domain features, and abnormal information such as alarms and faults.

[0063] As the highest level of the wind farm data fusion system, the decision-making layer fusion, based on the feature vectors, status identifiers, feature rule sets, and feature importance ranking results output by the feature layer, realizes comprehensive decision-making and optimized control of equipment. The core is to trigger corresponding level maintenance strategies by setting multi-level thresholds (early warning, alarm, danger) for parameters such as vibration and temperature.

[0064] 2. Data Transmission and Storage Specifications Feature value data: collected in real time with a 30-second cycle for rapid on-site alarm; transmitted in point table format via Modbus protocol, supporting push from CMS server to SCADA / field control server, or active acquisition by SCADA / field control server (see appendix for data point table sample).

[0065] Waveform file data: A complete set of files is generated every 2 hours, which is dedicated to equipment fault diagnosis and precise analysis; it is transmitted to the preset storage path of the back-end server via SFTP protocol and saved in CSV / TXT format, and can also be retrieved from the field server.

[0066] Transmission logic description: Because the waveform file data is large, shortening the transmission interval will lead to bandwidth congestion. Therefore, feature value data transmission is used to achieve real-time monitoring.

[0067] Threshold settings: The feature value alarm threshold is temporarily configured as shown in the table below. It will be optimized and adjusted after accumulating three months of blade vibration data across the entire field.

[0068] 3. Performance Evaluation Indicators Performance degradation metrics are calculated based on principal component analysis (PCA), and maintenance decisions are triggered when these metrics exceed a preset threshold (e.g., 0.8).

[0069] Core logic: Comprehensively evaluate factors such as CMS online monitoring and early warning information (weight 1), voiceprint sensor system data (weight 2), video early warning information (weight 3), wind turbine power curve (weight 4), blade root load assessment results (weight 5), and number of failures (weight 6). If the degradation index calculated by weighting exceeds the preset threshold (e.g., 0.8), the preventive maintenance process will be initiated.

[0070] For example, the comprehensive evaluation result = CMS online monitoring and early warning information * weight 1 + voiceprint sensor system data * weight 2 + video early warning information * weight 3 + wind turbine power curve * weight 4 + blade root load evaluation result * weight 5 + number of failures * weight 6.

[0071] 4. Confidence threshold A decision confidence threshold (typically 0.9-0.95) is set to assess the reliability of the decision. If key data is missing or there are contradictions in the data from different sensors, the decision confidence will decrease. When the confidence is below the threshold (e.g., 0.9), the system will not execute automatic decision-making, but will instead trigger two types of actions: requesting manual intervention or initiating a supplementary data collection task.

[0072] 5. Economic evaluation indicators Calculate economic indicators such as power generation efficiency, maintenance costs, and revenue forecasts to provide a basis for operational decisions.

[0073] Core logic: Based on the health status of the turbines, the power demand of the wind farm is reallocated to the faulty turbines with low health factors. The total power demand of the wind farm at the next moment is allocated to other healthy turbines according to the proportion of the available power of each healthy turbine to the total available power of the entire farm at the current moment. The power allocation formula is as follows: P D,Qi (k+1)=P D (k+1)*P A,Qi (k) / P A (k) P D (k+1) represents the total power demand of the wind farm at the next time step k+1.

[0074] P A，Qi (k) represents the available power of the wind turbine Qi at the current time k.

[0075] P D,Qi (k+1) represents the available power of the wind turbine Qi at the next time step k+1.

[0076] P A (k) represents the total available power of the wind farm at the current time k.

[0077] 6. Output Decision and Feedback Control The output of decision-level fusion includes four core outcomes: Control commands: Control commands that adjust operating parameters such as wind turbine pitch angle and yaw angle.

[0078] Maintenance strategy: Maintenance recommendations (such as routine maintenance, emergency repairs, and component replacement) generated based on equipment condition assessment.

[0079] Performance optimization solutions: Overall performance optimization solutions for wind farms (such as power distribution adjustment and operation mode switching).

[0080] Warning information: Equipment failure warnings, performance degradation warnings, and other alerts.

[0081] The output data format is a structured control instruction set, maintenance work order, or decision report, which is transmitted to the monitoring system or control system for execution through standard interfaces such as Modbus-TCP.

[0082] 7. Feedback Mechanism The feedback mechanism is a core component of the decision-making level: by comparing the results of decision execution with the actual operating effects, the decision-making model and various threshold settings are continuously optimized to form a closed-loop optimization system; the feedback data is re-inputted into the data layer to start a new round of data fusion cycle, realizing the continuous learning and iterative optimization of the system.

[0083] like Figure 5 As shown in the embodiments of this application, a wind farm hierarchical multi-source data fusion processing device includes a memory and a processor. The memory stores a computer-readable program, and when the computer-readable program is called by the processor, it can execute the wind farm hierarchical multi-source data fusion processing method of this application.

[0084] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for hierarchical multi-source data fusion processing of wind farms based on edge computing, characterized in that, The method is based on a bottom-up data aggregation architecture, which realizes the processing and application of multi-source data from wind farms through hierarchical fusion of data layer, feature layer, and decision layer. Specifically, it includes the following steps: Data layer fusion steps: Deploy edge processing devices to achieve partitioned access of multi-source data acquisition devices at the wind farm site layer through independent network cards. The multi-source data includes unit operation data, environmental meteorological data, equipment status monitoring data, maintenance management data, and power grid operation data. Use a sliding window alignment algorithm combined with interpolation to align the timestamps of multi-source data with different acquisition frequencies. Calculate the arithmetic mean of the data within the window using a sliding window with a preset uniform period to output a uniform frequency dataset. Feature layer fusion step: Taking the unified frequency dataset output from the data layer as input, extract unit operation event information, monitoring equipment alarm information, time domain features, frequency domain features, and time-frequency domain features. Use attention mechanism fusion, association rule mining, and clustering fusion algorithms to fuse the extracted features and generate feature information. The feature information includes feature vectors, status identifier sets, feature rule sets, and feature importance ranking results. The decision-making layer fusion steps are as follows: acquire the feature information output by the feature layer and the real-time operation data of the SCADA system; generate control commands, maintenance strategies, performance optimization schemes and early warning information based on multi-level threshold settings, performance degradation index evaluation, decision confidence judgment and economic evaluation; and send the decision execution results back to the data layer through a feedback mechanism to achieve closed-loop optimization.

2. The wind farm hierarchical multi-source data fusion processing method based on edge computing according to claim 1, characterized in that, The specific implementation of the sliding window alignment algorithm in the data layer fusion step includes: When a new data point flows into the window, it is determined whether the window is full. If the window is full, the oldest data point is removed and the new data point is added. The arithmetic mean of all data in the window is calculated. If the window is not full, the arithmetic mean of all data in the current window is directly calculated and the arithmetic mean is output as uniform frequency data.

3. The wind farm hierarchical multi-source data fusion processing method based on edge computing according to claim 1, characterized in that, In the data fusion step, for monitoring needs below 100Hz, the blade root load, tower top displacement, tower load, and tilt angle parameters of the unit are collected through blade root load monitoring equipment, airspace Beidou positioning and orientation equipment, tower load sensing equipment, and tilt angle sensing equipment; for high-frequency data needs from 1Hz to 100kHz, high-frequency data acquisition is completed through CMS system, transmission chain vibration monitoring system, blade vibration monitoring system, and oil monitoring sensor, while combining video and audio monitoring equipment to achieve visualization and acoustic signal monitoring.

4. The wind farm hierarchical multi-source data fusion processing method based on edge computing according to claim 1, characterized in that, In the feature layer fusion step, feature extraction specifically includes: The time-domain features are statistical characteristics of mean, variance, peak value, and kurtosis; The frequency domain features are the spectral features and envelope spectrum features obtained through FFT transformation; The time-frequency domain features are obtained through wavelet transform or short-time Fourier transform; Unit operation event information includes fault information, unit operating status, and warning information.

5. The wind farm hierarchical multi-source data fusion processing method based on edge computing according to claim 1, characterized in that, In the feature layer fusion step, the association rule mining adopts an experience base combined with the FP-Growth algorithm to mine the association relationship between features and form a feature rule set. Cluster fusion employs feature compression clustering or k-nearest neighbor method to perform cluster analysis on features and identify unit operating status patterns; Attention mechanism fusion highlights key features and suppresses secondary features by calculating attention weights for different features.

6. The wind farm hierarchical multi-source data fusion processing method based on edge computing according to claim 1, characterized in that, In the decision-level fusion step, the performance degradation index is calculated based on principal component analysis, specifically as follows: The system comprehensively evaluates CMS online monitoring and early warning information, voiceprint sensor system data, video early warning information, unit power curve, blade root load assessment results, and number of faults, and performs weighted calculations according to preset weights. When the calculation results exceed preset thresholds, the preventive maintenance process is triggered.

7. The wind farm hierarchical multi-source data fusion processing method based on edge computing according to claim 1, characterized in that, In the decision-making layer fusion step, a decision confidence threshold is set. When the decision confidence is lower than the decision confidence threshold, the system does not perform automatic decision-making and triggers a manual intervention request or a supplementary data collection task.

8. The wind farm hierarchical multi-source data fusion processing method based on edge computing according to claim 1, characterized in that, In the decision-making integration step, the economic assessment is achieved through power allocation adjustment. For faulty units with low health factors, the total power demand of the wind farm at the next moment is allocated according to the proportion of the current available power of each healthy unit to the total available power of the entire farm.

9. The wind farm hierarchical multi-source data fusion processing method based on edge computing according to claim 1, characterized in that, The feedback mechanism of the decision-making layer fusion step is as follows: compare the decision execution results with the actual operating effect of the unit, optimize the decision model and various threshold settings, re-input the feedback data into the data layer, and carry out a new round of data fusion cycle.

10. A wind farm hierarchical multi-source data fusion processing device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer-readable program, which, when invoked by the processor, can execute the wind farm hierarchical multi-source data fusion processing method as described in any one of claims 1 to 9.

Images