Wind power fault dynamic early warning method and system based on multi-source heterogeneous data fusion

Abstract

The invention relates to the field of fault early warning, in particular to a wind power fault dynamic early warning method and system based on multi-source heterogeneous data fusion. According to the method, multi-source data such as SCADA operation data, CMS vibration monitoring data and meteorological environment data of a wind turbine generator are collected in real time, standardization processing is carried out, and a multi-dimensional feature vector is constructed. And generating a fusion data set by using an adaptive weighted fusion algorithm, constructing a fault prediction model based on a deep convolutional neural network, and outputting a health state assessment value and a fault risk level in real time after historical fault sample supervised training. And when the risk level exceeds a threshold value, generating an early warning signal containing a fault type and a positioning and repairing suggestion, dynamically adjusting a monitoring parameter weight, iteratively updating a model, and realizing adaptive optimization of an early warning strategy. The problem that an existing method depends on single data source and multi-source data fusion is solved, and accurate dynamic early warning is achieved.

Description

Technical Field

[0001] The present invention relates to the field of fault warning, and in particular to a wind power fault dynamic warning method and system based on multi-source heterogeneous data fusion. Background Art

[0002] With the booming wind power industry, the efficient and stable operation of wind turbines is crucial for ensuring energy supply and reducing operation and maintenance costs. However, wind turbines are complex structures and operate in harsh environments. Frequent failures not only affect power generation efficiency but also lead to high maintenance costs. Traditional fault monitoring and early warning methods rely primarily on single data sources, such as operational data collected by SCADA systems or vibration monitoring data from CMSs. These methods have numerous limitations. First, a single data source cannot fully reflect the operational status of a wind turbine, easily overlooking potential fault hazards. Second, traditional methods, often based on empirical rules or simple statistical analysis, lack the ability to identify complex fault patterns, making it difficult to provide accurate early warnings. With the development of technologies such as big data and artificial intelligence, multi-source heterogeneous data fusion offers new approaches to wind power fault early warning. By integrating multiple data sources, such as SCADA operational data, CMS vibration monitoring data, and meteorological and environmental data, a more comprehensive picture of wind turbine operational status can be achieved, providing a richer information base for fault early warning. However, the application of multi-source heterogeneous data fusion to wind power fault early warning still faces numerous technical challenges. For example, different data sources have different data formats, sampling frequencies, and timestamps. Standardizing data processing and correlating cross-source features is a key issue. Furthermore, building an effective fault prediction model to fully utilize the fused data for accurate fault prediction is a technical challenge that needs to be addressed. This application addresses the aforementioned technical issues in existing wind power fault warning methods and proposes a dynamic wind power fault warning method and system based on multi-source heterogeneous data fusion. Summary of the Invention

[0003] The object of the present invention is to provide a wind power fault dynamic early warning method and system based on multi-source heterogeneous data fusion to solve the problems in the above-mentioned background technology.

[0004] To achieve the above objectives, the following technical solutions are adopted.

[0005] A dynamic early warning method for wind power faults based on multi-source heterogeneous data fusion includes the following steps: real-time collection of SCADA operation data, CMS vibration monitoring data and meteorological environment data of wind turbines to form a multi-source heterogeneous data stream, standardization of the multi-source heterogeneous data, and construction of a multi-dimensional feature vector with a timestamp; an adaptive weighted fusion algorithm is used to perform cross-source feature correlation analysis on the multi-dimensional feature vector to generate a fused data set; a fault prediction model is constructed based on a deep convolutional neural network, and the model is supervised and trained using historical fault samples to obtain a multi-level fault probability prediction function; the fused data set is input into the trained fault prediction model to output the health status assessment value and fault risk level of the unit's transmission chain components in real time; when the fault risk level exceeds a preset threshold, a warning signal containing the fault type, location information and repair suggestions is generated; the monitoring parameter weights are dynamically adjusted based on the warning results, and the newly added fault samples are iteratively updated to the fault prediction model to achieve adaptive optimization of the warning strategy.

[0006] Optionally, the adaptive weighted fusion algorithm includes establishing a cross-source data association matrix based on the Pearson correlation coefficient, calculating the dynamic weight coefficient of each data source and the target fault type; using a sliding time window to perform time domain alignment on the multidimensional feature vector, and generating a fused feature vector with time and space consistency through weighted superposition; and performing outlier removal and noise filtering on the fused data stream to form a standardized input sequence.

[0007] Optionally, the deep convolutional neural network adopts a dual-channel architecture, including: the first channel is used to process SCADA time series data, and an LSTM network with an attention mechanism is used to extract equipment operation status features; the second channel is used to process CMS high-frequency vibration signals, and a multi-layer wavelet convolution kernel is used to extract frequency domain fault features; the dual-channel outputs are feature spliced through a fully connected layer to output a multi-dimensional fault probability distribution.

[0008] Optionally, the determination of the fault risk level includes: constructing a dynamic warning threshold curve based on the health status assessment value, and triggering a three-level warning mechanism when the assessment value continuously exceeds the threshold curve and the duration reaches a preset window; the warning signal uses a fuzzy logic algorithm to integrate the fault probability, equipment criticality and maintenance resource availability to generate a differentiated emergency response strategy.

[0009] Optionally, it also includes: building a preventive maintenance knowledge base based on historical response data of early warning signals; dynamically optimizing unit maintenance cycles and spare parts replacement strategies based on maintenance records and equipment life prediction models in the knowledge base, to achieve a model transition from passive maintenance to predictive maintenance.

[0010] Optionally, the adaptive optimization of the model includes: when the confidence of a new fault sample exceeds a preset threshold, updating the model parameters using an online incremental learning algorithm; for low-confidence samples, triggering a manual labeling module and conducting adversarial training with historical data to improve the model's generalization ability.

[0011] Optionally, the multi-source heterogeneous data also includes: blade stress distribution data, gearbox oil detection data and power grid dispatch instruction data; by constructing a multimodal data mapping relationship, collaborative analysis of mechanical, electrical and environmental parameters can be achieved.

[0012] Optionally, it also includes: pushing the warning results to the mobile terminal in the form of visual charts, supporting users to customize warning rules; based on user feedback data, optimizing the warning threshold and model parameters through reinforcement learning algorithms to form a closed loop of human-machine collaborative decision-making.

[0013] The wind power fault dynamic early warning system based on multi-source heterogeneous data fusion includes:

[0014] Data acquisition module, used to obtain SCADA operation data, CMS vibration monitoring data and meteorological environment data in real time;

[0015] The data fusion module uses an adaptive weighting algorithm to perform cross-source feature correlation and standardization on multi-source heterogeneous data to generate a fused data set;

[0016] Model building module, including deep convolutional neural network and online learning unit, used to train fault prediction model and dynamically update parameters;

[0017] The early warning execution module triggers multi-level early warning signals based on the health status assessment values output by the model and generates equipment maintenance recommendations. The operation and maintenance optimization module builds a preventive maintenance knowledge base based on historical early warning data and dynamically adjusts the unit maintenance strategy.

[0018] The visual interaction module is used to display warning results and receive user feedback to achieve human-computer collaborative decision optimization.

[0019] Optionally, the data fusion module also includes: a cross-source data alignment unit, which is used to synchronize multi-source data in the time domain through a sliding time window; the model construction module adopts a dual-channel network architecture to process time series data and high-frequency vibration signals respectively; the operation and maintenance optimization module integrates a life prediction algorithm to dynamically generate spare parts procurement plans and maintenance priority lists based on the remaining life of the equipment.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] This invention achieves accurate dynamic early warning of wind turbine faults by collecting and standardizing multiple data sources in real time, using an adaptive weighted fusion algorithm to generate a fused dataset, and then building a fault prediction model using a deep convolutional neural network. This solves the problems mentioned in the background art, such as the existing methods' reliance on a single data source, resulting in insufficient fault identification capabilities, and the difficulty of fusing multi-source heterogeneous data. This method can comprehensively reflect the operating status of wind turbines, improve the accuracy of fault prediction, and thus effectively reduce the failure rate and maintenance costs, ensuring the stable operation of wind farms. The present invention improves and supplements the specific implementation of the adaptive weighted fusion algorithm, the dual-channel architecture design of the deep convolutional neural network, the fault risk level determination mechanism, the adaptive optimization of the early warning strategy, the construction of a preventive maintenance knowledge base, the adaptive optimization of the model, the expansion of multi-source heterogeneous data, and the closed-loop human-machine collaborative decision-making. This further enhances the performance and reliability of the fault early warning system, strengthens the system's intelligence and practicality, enables it to better adapt to the complex and changing wind power operating environment, and provides more powerful technical support for wind power operation and maintenance. DETAILED DESCRIPTION

[0022] The present invention will be described in detail below with reference to the embodiments. It should be noted that, unless there is a conflict, the embodiments and features in the embodiments of the present application can be combined with each other.

[0023] The following detailed description is an exemplary description, which is intended to provide further detailed description of the present invention. Unless otherwise indicated, all technical terms used in the present invention have the same meaning as those generally understood by those skilled in the art. The terms used in the present invention are only for describing specific embodiments, and are not intended to limit the exemplary embodiments according to the present invention.

[0024] This embodiment provides a dynamic wind turbine fault warning method and system based on multi-source heterogeneous data fusion. This method integrates multiple data sources and utilizes advanced data processing and analysis techniques to achieve accurate early warning and dynamic optimization of wind turbine faults. The following is a detailed implementation of this method.

[0025] First, various wind turbine operating data is collected in real time, including SCADA operating data, CMS vibration monitoring data, meteorological and environmental data, blade stress distribution data, gearbox oil test data, and grid dispatch instruction data. This data comes from a wide range of sources, covering mechanical, electrical, and environmental parameters of the wind turbine, and can comprehensively reflect the wind turbine's operating status. SCADA operating data primarily includes key operating parameters such as wind turbine power, speed, and temperature. It is collected in real time by the wind turbine's SCADA system, with a sampling frequency that can be set according to actual needs, such as once per second. CMS vibration monitoring data is obtained from vibration sensors installed on key wind turbine components. These sensors monitor component vibration in real time, typically at a high sampling frequency, such as hundreds of times per second, to capture subtle changes in vibration signals. Meteorological and environmental data, including wind speed, direction, temperature, and air pressure, is collected from meteorological monitoring stations installed at wind farms. This data is crucial for analyzing the wind turbine's operating environment and predicting potential failures. Blade stress distribution data is obtained from strain sensors installed on the blades. This data reflects the stress state of the blades under different operating conditions, helping to promptly detect problems such as blade fatigue damage. Gearbox oil test data is acquired through oil analysis sensors, including parameters such as oil temperature, viscosity, and impurity content. This data can reflect the lubrication condition and wear level of the gearbox. Grid dispatch command data is obtained from the grid dispatch system, including information such as the grid's power regulation instructions for wind turbines. This data helps analyze the operating status of wind turbines under different grid operating conditions.

[0026] The collected heterogeneous data from multiple sources differs in time and format, necessitating standardization. Specifically, data from each data source undergoes preprocessing, including data cleaning, format conversion, and timestamp alignment. Data cleaning primarily removes outliers and noise. For example, vibration data is filtered using a sliding average filter to remove abnormal vibration signals caused by sensor failure or external interference. Format conversion converts data from different sources into a unified format for subsequent processing. For example, wind speed units in meteorological data can be converted from meters per second to kilometers per hour to align with the units of other data sources. Timestamp alignment aligns data from different sources in chronological order to accurately analyze relationships between them. For example, a sliding time window algorithm is used to align SCADA data, CMS data, and meteorological data based on timestamps. This allows data at the same time point to be mapped, forming multidimensional feature vectors with timestamps. These multidimensional feature vectors contain various operating status information about the wind turbine at different time points, providing a foundation for further data fusion and fault prediction.

[0027] Next, an adaptive weighted fusion algorithm is used to perform cross-source feature correlation analysis on the multidimensional feature vectors to generate a fused dataset. The core of this algorithm lies in dynamically calculating the weight coefficients of each data source based on the degree of correlation between the different data sources and the target fault type, and then fusing the multi-source data through weighted superposition. Specifically, a cross-source data correlation matrix based on the Pearson correlation coefficient is first established to calculate the correlation between each data source and the target fault type. For example, by calculating the Pearson correlation coefficient between power parameters in SCADA data and gearbox faults, the importance of power parameters in predicting gearbox faults is determined. Then, based on the calculated correlation coefficient, a sliding time window is used to perform time-domain alignment on the multidimensional feature vectors, and a weighted superposition is performed to generate a fused feature vector with spatiotemporal consistency. During the weighted superposition process, data sources with a high correlation with the target fault type are assigned a higher weight, while data sources with a lower correlation are assigned a lower weight, thereby achieving effective fusion of multi-source data. In addition, the fused data stream is processed through outlier removal and noise filtering to further improve the data quality and reliability, forming a standardized input sequence that provides a high-quality data foundation for subsequent fault prediction model training.

[0028] A fault prediction model is constructed based on a deep convolutional neural network. The model is supervised and trained using historical fault samples to obtain a multi-level fault probability prediction function. The deep convolutional neural network employs a dual-channel architecture. The first channel processes SCADA time series data, employing an LSTM network with an attention mechanism to extract equipment operating status features. The second channel processes CMS high-frequency vibration signals, employing multi-layer wavelet convolution kernels to extract frequency-domain fault features. Specifically, in the first channel, the LSTM network effectively processes the long-term and short-term dependencies in time series data. By introducing the attention mechanism, the model can focus more on fault-related time series features, thereby improving fault prediction accuracy. For example, when processing wind turbine power fluctuation data, the attention mechanism can enable the model to focus on key features such as sudden power changes, which are often closely related to equipment failures. In the second channel, multi-layer wavelet convolution kernels perform multi-scale analysis on the CMS high-frequency vibration signals, extracting fault features in different frequency bands. For example, the wavelet convolution kernel can extract low-frequency vibration signal features, which may be associated with low-frequency equipment failure modes, as well as high-frequency vibration signal features, which may be associated with high-frequency equipment failure modes. The dual-channel output is concatenated using a fully connected layer, integrating the features extracted from both channels to produce a multi-dimensional fault probability distribution. For example, the model output might include multiple dimensions, such as the probability of a gearbox failure or the probability of a generator failure. The failure probability value for each dimension reflects the likelihood of a wind turbine component failing.

[0029] The fused dataset is fed into the trained fault prediction model, which outputs real-time health status assessments and fault risk levels for the turbine's drive train components. When the fault risk level exceeds a preset threshold, a warning signal is generated, including the fault type, location information, and repair recommendations. Specifically, based on the fault probability distribution output by the model, a health status assessment of the turbine's drive train components is calculated. This assessment reflects the component's current health status. For example, a weighted average is used to sum the probabilities of different fault types to produce a comprehensive health status assessment. Simultaneously, the fault risk level is determined based on the health status assessment and a preset threshold. When the fault risk level exceeds the preset threshold, indicating a high fault risk for the wind turbine, a warning signal is generated. The warning signal includes fault type information, such as gearbox fault or generator fault. By analyzing the fault probability distribution output by the model, the component with the highest probability of failure is identified as the fault type. The warning signal also includes location information, such as the specific location of the fault by analyzing the vibration signal's propagation path and time delay. Furthermore, the warning signal includes repair recommendations, providing appropriate repair measures and suggestions based on the fault type and severity, such as replacing damaged components or performing lubrication maintenance.

[0030] Based on the early warning results, the weights of monitoring parameters are dynamically adjusted, and newly added fault samples are iteratively updated to the fault prediction model, achieving adaptive optimization of the early warning strategy. Specifically, based on historical response data for early warning signals, the effectiveness of different monitoring parameters in fault early warning is analyzed, and the weights of these parameters are dynamically adjusted. For example, if a monitoring parameter has been shown to be highly correlated with fault occurrence in multiple early warnings, its weight is increased; conversely, if a monitoring parameter has not shown a clear correlation with faults in multiple early warnings, its weight is decreased. Simultaneously, an online incremental learning algorithm is used to update model parameters for newly added fault samples. When the confidence level of a newly added fault sample exceeds a preset threshold, it is directly added to the model training data, and the model parameters are updated using the online incremental learning algorithm, enabling the model to promptly learn new fault characteristics and patterns. For low-confidence samples, the manual annotation module is triggered, and professionals annotate and review the samples. These samples are then trained against historical data. This approach improves the model's generalization and accuracy, further optimizing the early warning strategy and enhancing the accuracy and reliability of fault early warnings.

[0031] In this embodiment, a dynamic early warning system for wind power faults based on multi-source heterogeneous data fusion is also provided. The system includes a data acquisition module, a data fusion module, a model building module, an early warning execution module, an operation and maintenance optimization module, and a visual interaction module. The data acquisition module is used to obtain multiple data sources such as SCADA operation data, CMS vibration monitoring data, and meteorological environment data in real time. The data fusion module uses an adaptive weighting algorithm to perform cross-source feature correlation and standardization processing on multi-source heterogeneous data to generate a fused data set. The model building module includes a deep convolutional neural network and an online learning unit for training fault prediction models and dynamically updating parameters. The early warning execution module triggers multi-level early warning signals based on the health status assessment value output by the model and generates equipment maintenance recommendations. The operation and maintenance optimization module builds a preventive maintenance knowledge base based on historical early warning data and dynamically adjusts the unit maintenance strategy. The visual interaction module is used to display early warning results and receive user feedback to achieve human-computer collaborative decision optimization.

[0032] The data fusion module also includes a cross-source data alignment unit, which synchronizes multi-source data in the time domain using a sliding time window. The model building module utilizes a dual-channel network architecture to process time series data and high-frequency vibration signals separately. The operation and maintenance optimization module integrates a lifespan prediction algorithm to dynamically generate spare parts procurement plans and maintenance priority lists based on the remaining lifespan of the equipment. The collaborative work of these modules enables accurate early warning and dynamic optimization of wind turbine failures, effectively improving the intelligence and efficiency of wind power operation and maintenance.

[0033] In practical applications, the system can be deployed in the wind farm's monitoring center, connecting to data with the wind turbine's SCADA system, CMS system, meteorological monitoring stations, and other equipment to collect and process various operational data in real time. Based on the set warning thresholds and strategies, the system automatically issues fault warning signals and pushes the warning results to the operator's mobile terminal or monitoring screen, while also providing detailed fault analysis reports and maintenance recommendations. Based on the warning signals and recommendations, operators can promptly take appropriate maintenance measures to prevent the occurrence or expansion of faults and ensure the stable operation of the wind farm. In addition, the system supports user-defined warning rules and parameter settings, allowing for flexible adjustment of warning strategies based on different wind farm operating conditions and management requirements, thereby improving the system's applicability and practicality.

[0034] In summary, the wind power fault dynamic early warning method and system based on multi-source heterogeneous data fusion provided in this embodiment achieves accurate early warning and dynamic optimization of wind turbine faults by integrating multiple data sources and utilizing advanced data processing and analysis technologies. This method and system can effectively improve the intelligence and efficiency of wind power operation and maintenance, reduce the occurrence rate of failures and maintenance costs, and is of great significance for ensuring the stable development of the wind power industry.

[0035] It is understood from common technical knowledge that the present invention may be implemented by other embodiments that do not depart from its spirit or essential features. Therefore, the embodiments disclosed above are, in all respects, merely illustrative and not exclusive. All modifications within the scope of the present invention or equivalent to the scope of the present invention are intended to be encompassed by the present invention.

Claims

1. A wind power fault dynamic early warning method based on multi-source heterogeneous data fusion is characterized by: The following steps are involved: Real-time collection of SCADA operation data, CMS vibration monitoring data, and meteorological environment data from wind turbines to form multi-source heterogeneous data streams; The multi-source heterogeneous data is standardized to construct a multi-dimensional feature vector with a timestamp; an adaptive weighted fusion algorithm is used to perform cross-source feature correlation analysis on the multi-dimensional feature vector to generate a fused data set; a fault prediction model is constructed based on a deep convolutional neural network, and the model is supervised and trained using historical fault samples to obtain a multi-level fault probability prediction function; the fused data set is input into the trained fault prediction model to output the health status assessment value and fault risk level of the unit's transmission chain components in real time; When the fault risk level exceeds a preset threshold, an early warning signal is generated containing the fault type, location information, and repair suggestions; The monitoring parameter weights are dynamically adjusted based on the early warning results, and the newly added fault samples are iteratively updated to the fault prediction model to achieve adaptive optimization of the early warning strategy.

2. The wind power fault dynamic early warning method based on multi-source heterogeneous data fusion according to claim 1 is characterized in that: The adaptive weighted fusion algorithm includes: establishing a cross-source data association matrix based on the Pearson correlation coefficient, calculating the dynamic weight coefficient of each data source and the target fault type; using a sliding time window to align the multidimensional feature vectors in the time domain, and generating a fused feature vector with temporal and spatial consistency through weighted superposition; and performing outlier removal and noise filtering on the fused data stream to form a standardized input sequence.

3. The wind power fault dynamic early warning method based on multi-source heterogeneous data fusion according to claim 1 is characterized in that: The deep convolutional neural network adopts a dual-channel architecture, including: the first channel is used to process SCADA time series data, using an LSTM network with an attention mechanism to extract equipment operating status features; the second channel is used to process CMS high-frequency vibration signals, using a multi-layer wavelet convolution kernel to extract frequency domain fault features; the dual-channel output is feature-concatenated through a fully connected layer to output a multi-dimensional fault probability distribution.

4. The wind power fault dynamic early warning method based on multi-source heterogeneous data fusion according to claim 1 is characterized in that: The determination of the fault risk level includes: constructing a dynamic warning threshold curve based on the health status assessment value, triggering a three-level warning mechanism when the assessment value continuously exceeds the threshold curve and the duration reaches a preset window; the warning signal uses a fuzzy logic algorithm to integrate the fault probability, equipment criticality and maintenance resource availability to generate a differentiated emergency response strategy.

5. The wind power fault dynamic early warning method based on multi-source heterogeneous data fusion according to claim 1 is characterized in that: Also includes: Build a preventive maintenance knowledge base based on historical response data of early warning signals; Based on the maintenance records and equipment life prediction model in the knowledge base, the unit maintenance cycle and spare parts replacement strategy are dynamically optimized to achieve a model transition from passive maintenance to predictive maintenance.

6. The wind power fault dynamic early warning method based on multi-source heterogeneous data fusion according to claim 1 is characterized in that: The adaptive optimization of the model includes: when the confidence of a new fault sample exceeds a preset threshold, an online incremental learning algorithm is used to update the model parameters; for low-confidence samples, a manual labeling module is triggered and adversarial training is performed with historical data to improve the model's generalization ability.

7. The wind power fault dynamic early warning method based on multi-source heterogeneous data fusion according to claim 1 is characterized in that: The multi-source heterogeneous data also includes: blade stress distribution data, gearbox oil detection data and power grid dispatch instruction data; by constructing a multi-modal data mapping relationship, collaborative analysis of mechanical, electrical and environmental parameters is achieved.

8. The wind power fault dynamic early warning method based on multi-source heterogeneous data fusion according to claim 1 is characterized in that: Also includes: Push warning results to mobile terminals in the form of visual charts, and support user-defined warning rules; Based on user feedback data, the warning threshold and model parameters are optimized through reinforcement learning algorithms to form a closed loop of human-machine collaborative decision-making.

9. A wind power fault dynamic early warning system based on multi-source heterogeneous data fusion, based on the wind power fault dynamic early warning method based on multi-source heterogeneous data fusion according to any one of claims 1 to 8, characterized in that: include: Data acquisition module, used to obtain SCADA operation data, CMS vibration monitoring data and meteorological environment data in real time; The data fusion module uses an adaptive weighting algorithm to perform cross-source feature correlation and standardization on multi-source heterogeneous data to generate a fused data set; Model building module, including deep convolutional neural network and online learning unit, used to train fault prediction model and dynamically update parameters; The early warning execution module triggers multi-level early warning signals based on the health status assessment values output by the model and generates equipment maintenance recommendations; The operation and maintenance optimization module builds a preventive maintenance knowledge base based on historical warning data and dynamically adjusts the unit maintenance strategy; The visual interaction module is used to display warning results and receive user feedback to achieve human-computer collaborative decision optimization.

10. The wind power fault dynamic early warning system based on multi-source heterogeneous data fusion according to claim 9 is characterized in that: The data fusion module also includes: a cross-source data alignment unit for performing time domain synchronization on multi-source data through a sliding time window; the model construction module adopts a dual-channel network architecture to process time series data and high-frequency vibration signals respectively; the operation and maintenance optimization module integrates a life prediction algorithm to dynamically generate spare parts procurement plans and maintenance priority lists based on the remaining life of the equipment.