Wind power plant equipment state on-line monitoring system

Abstract

The invention relates to the technical field of wind power plant equipment monitoring, and discloses a wind power plant equipment state online monitoring system. The system comprises a real-time acquisition module, a multi-modal fusion module, a dynamic segmentation module, a preliminary evaluation module, a deep analysis module and a maintenance trigger module. The real-time acquisition module acquires a multi-source sensor data stream, and the multi-modal fusion module performs adaptive weighted fusion on the multi-source sensor data stream to generate a comprehensive signal feature set. The dynamic segmentation module segments a feature set according to a load rate to generate a running state feature sequence, and the preliminary evaluation module performs health degree scoring by using a deep belief network. The deep analysis module performs modeling on abnormal time period data through a space-time diagram neural network to generate an equipment state label, and the maintenance triggering module triggers a real-time maintenance strategy instruction according to the equipment state label. The system can accurately monitor the equipment state of the wind power plant in real time, improve the fault diagnosis accuracy, realize intelligent maintenance decision, and improve the operation reliability and economy of the wind power plant.

Description

Technical Field

[0001] This invention relates to the field of wind farm equipment monitoring technology, specifically to an online monitoring system for the status of wind farm equipment. Background Technology

[0002] As global demand for clean energy continues to rise, wind power, as an important renewable energy source, is playing an increasingly important role in the power supply system. Wind farms are expanding in scale, with more equipment and increasingly complex structures. From wind turbine blades, gearboxes, and generators to various transmission lines and transformer equipment, any failure in any component can lead to power outages, power generation losses, or even safety accidents.

[0003] Traditional wind farm equipment monitoring methods rely heavily on manual inspections, which have significant limitations. Manual inspections are typically conducted on a fixed schedule, making it impossible to detect sudden anomalies in equipment operation in real time. Equipment failures are often insidious and sudden; potential faults may have already appeared and gradually worsened between inspections, but these are difficult to detect in time. For example, wear and tear on internal components of a wind turbine gearbox may initially manifest as only slight vibrations or temperature changes. Manual inspections often struggle to accurately detect these subtle anomalies based on experience. Once the fault progresses to a severe stage, repair costs will increase significantly, potentially requiring the replacement of the entire gearbox.

[0004] While some wind farms are equipped with simple monitoring devices, their functions are limited, monitoring only a few parameters and failing to comprehensively reflect the equipment's operating status. For example, monitoring only temperature while ignoring key indicators such as vibration and current makes it difficult to accurately diagnose and predict equipment failures. Moreover, the data collected by these single-parameter monitoring devices is mostly discrete and independent, lacking the ability to fuse and analyze multi-source data, thus failing to grasp the overall operating status of the equipment.

[0005] Furthermore, existing monitoring systems rely heavily on simple threshold judgments or empirical models for equipment health assessment and fault diagnosis. Threshold judgments depend too heavily on pre-set fixed values ​​and cannot adapt to changes in equipment operating characteristics under different conditions. Empirical models are limited by the experience of technicians; differences in experience among technicians can lead to inconsistent assessment results, making it difficult to achieve accurate equipment condition assessment and fault prediction, and thus failing to meet the requirements of efficient and reliable wind farm operation.

[0006] In terms of maintenance strategies, traditional monitoring systems cannot trigger maintenance commands in a timely and accurate manner based on the actual condition of the equipment. Repairs are typically only carried out after equipment failure, a reactive approach that not only results in significant power generation losses but may also exacerbate equipment damage due to delayed repairs. Furthermore, there is a lack of effective preventative maintenance measures for potential faults, preventing the optimization of equipment operating parameters or the scheduling of maintenance plans in advance, thus failing to fully realize the equipment's lifespan and economic benefits. Therefore, the development of an efficient and intelligent online monitoring system for wind farm equipment is urgently needed to address the numerous problems currently existing in wind farm equipment monitoring and maintenance, ensuring the stable and efficient operation of wind farms. Summary of the Invention

[0007] The purpose of this invention is to provide an online monitoring system for the status of wind farm equipment to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: an online monitoring system for the status of wind farm equipment, the system comprising: The real-time acquisition module is used to synchronously acquire multi-source sensor data streams from wind farm equipment. The sensor data streams include vibration signals, temperature distribution, current waveforms, and environmental parameters. The multimodal fusion module is used to perform adaptive weighted fusion of the multi-source sensor data streams based on the covariance matrix to generate a comprehensive signal feature set; The dynamic segmentation module performs variable time window segmentation on the comprehensive signal feature set based on the equipment operating load rate, generating multiple operating state feature sequences; The preliminary assessment module inputs the operating status feature sequence into a pre-trained deep belief network to score the device health and outputs the health index for each time period. The deep analysis module filters abnormal time periods based on the health index and inputs the corresponding comprehensive signal features into a pre-trained spatiotemporal graph neural network for multi-scale correlation modeling to generate device status labels. The maintenance triggering module triggers real-time maintenance strategy instructions based on the device status label.

[0009] Preferably, the specific steps of the multimodal fusion module include: Extract the frequency domain energy distribution characteristics and the spatial gradient characteristics of the temperature distribution of the vibration signal; Harmonic component decomposition is performed on the current waveform to generate the current distortion coefficient; Calculate the weight allocation values ​​for each sensor data based on the covariance matrix; Based on the weighted values, the frequency domain energy distribution characteristics, spatial gradient characteristics, and current distortion coefficient are linearly weighted and fused to generate a comprehensive signal feature set.

[0010] Preferably, the implementation logic of the dynamic segmentation module includes: The time window length is dynamically adjusted according to the equipment operating load rate. Under high load rate, the window length is shortened to the first set value, and under low load rate, the window length is extended to the second set value. A Gaussian weighted strategy is used to smooth the features in the overlapping areas of adjacent time windows; The smoothed feature sequence is segmented by running state clustering based on the density clustering algorithm.

[0011] Preferably, the construction steps of the deep belief network include: The operating state feature sequence is input into a restricted Boltzmann machine for unsupervised feature learning; The network parameters are fine-tuned using the backpropagation algorithm to generate hidden layer feature representations; A sigmoid activation function is introduced into the output layer to map the hidden layer features to the health index range.

[0012] Preferably, the modeling steps of the spatiotemporal graph neural network include: A static topology graph is constructed using equipment components as nodes and physical connections as edges; A time decay function is introduced to dynamically update the state dependency weights between nodes; A spatiotemporal embedding vector is generated by aggregating the multi-time-period state features of neighboring nodes through a graph attention mechanism. The spatiotemporal embedding vector is input into a fully connected layer for state label classification.

[0013] Preferably, the method for weighting the covariance matrix includes: Calculate the covariance matrix between each sensor's data and historical fault data; Eigenvalue decomposition is performed on the covariance matrix to extract the principal component directions; The fusion weights of each sensor's data are determined based on the variance contribution rate of the principal component direction.

[0014] Preferably, the implementation steps of the density clustering algorithm include: Calculate the local density and relative distance of the smoothed feature sequence; Cluster centers were determined by density peak detection. Non-centroid points are clustered and assigned based on a truncation distance threshold.

[0015] Preferably, the training process of the restricted Boltzmann machine includes updating the connection weights between the visible layer and the hidden layer using a contrastive divergence algorithm.

[0016] Preferably, the design of the time decay function includes: The exponential decay coefficient is defined based on the physical wear characteristics of equipment components; The attenuation coefficient is multiplied by the time interval to generate a dynamic weight adjustment factor; Dependencies that exceed a preset time threshold are weighted to zero.

[0017] Preferably, the triggering logic for the real-time maintenance strategy instruction includes: If the equipment status label is critically abnormal, the equipment shutdown protocol will be initiated and an emergency maintenance work order will be sent. If the equipment status label indicates a potential risk, a preventative maintenance plan is generated and the power generation quota is adjusted. Synchronize maintenance operation records to the equipment lifecycle management database.

[0018] Compared with the prior art, the beneficial effects of the present invention are: At the data acquisition and fusion level, the real-time acquisition module synchronously acquires multi-source sensor data streams from wind farm equipment, covering various aspects such as vibration signals, temperature distribution, current waveforms, and environmental parameters, comprehensively reflecting the equipment's operating status. The multi-modal fusion module organically combines data from different types of sensors through adaptive weighted fusion based on the covariance matrix. For example, it extracts the frequency domain energy distribution characteristics of vibration signals, the spatial gradient characteristics of temperature distribution, and decomposes the current waveform into harmonic components to generate current distortion coefficients. Then, based on the weights determined by the covariance matrix, it performs linear weighted fusion to generate a comprehensive signal feature set. This fusion method fully explores the inherent connections between multi-source data, overcomes the limitations of single-parameter monitoring, improves the accuracy and completeness of the data, and lays a solid foundation for subsequent precise analysis of equipment status.

[0019] The dynamic segmentation module uses variable time windows to segment the comprehensive signal feature set based on the equipment's operating load rate, offering strong flexibility and adaptability. Shortening the time window length under high load and lengthening it under low load allows for more effective capture of changes in the equipment's operating status under different conditions. A Gaussian weighting strategy is used to smooth the features in overlapping areas of adjacent time windows, avoiding analytical errors caused by sudden data changes. Combining density clustering algorithms with operational state clustering of the smoothed feature sequences enables precise segmentation of different equipment operating stages, providing clear data segmentation criteria for in-depth analysis of equipment operating patterns and health assessment.

[0020] The preliminary assessment module utilizes a pre-trained deep belief network to score the equipment health of operational status feature sequences, outputting a health index for each time period. The deep belief network performs unsupervised feature learning by inputting the operational status feature sequences into a restricted Boltzmann machine, then fine-tunes the network parameters using a backpropagation algorithm. Finally, a sigmoid activation function is introduced into the output layer to map to the health index range. Compared to traditional simple threshold judgments and empirical models, this approach more accurately and objectively reflects the equipment's health status, enabling a quantitative assessment of equipment condition and facilitating timely monitoring of equipment operation by maintenance personnel.

[0021] The deep analysis module filters abnormal time periods based on a health index and inputs the corresponding comprehensive signal features into a pre-trained spatiotemporal graph neural network for multi-scale correlation modeling to generate equipment status labels. The spatiotemporal graph neural network constructs a static topology graph with equipment components as nodes and physical connections as edges. A time decay function is introduced to dynamically update the state dependency weights between nodes. A graph attention mechanism aggregates the multi-time-period state features of neighboring nodes to generate spatiotemporal embedding vectors, which are then classified using fully connected layers. This modeling approach fully considers the spatial connections between equipment components and the temporal changes in state, enabling in-depth exploration of the potential causes and development trends of abnormal equipment states, thus improving the accuracy and reliability of fault diagnosis.

[0022] The maintenance trigger module triggers real-time maintenance strategy commands based on equipment status tags, achieving intelligent integration from equipment status monitoring to maintenance decision-making. If the equipment status tag indicates a severe anomaly, the equipment shutdown protocol is immediately activated and an emergency repair work order is pushed, minimizing the risk of equipment damage and accident losses. If it indicates a potential risk, a preventative maintenance plan is generated and the power generation quota is adjusted to eliminate potential hazards in advance, ensuring stable equipment operation while optimizing power generation efficiency. Maintenance operation records are synchronized to the equipment's full lifecycle management database, facilitating the traceability and analysis of equipment maintenance history and providing strong data support for subsequent maintenance strategy optimization and equipment performance improvement.

[0023] Through the collaborative work of the above modules, the system enables real-time, comprehensive, and accurate monitoring and intelligent maintenance decision-making for wind farm equipment. It can effectively reduce equipment downtime due to failure, lower maintenance costs, increase wind farm power generation and economic benefits, ensure the safe, stable, and efficient operation of wind farms, and promote the sustainable development of the wind power industry. Attached Figure Description

[0024] Figure 1 This is a schematic diagram illustrating the working principle of the online monitoring system for wind farm equipment status described in this invention. Figure 2 This is a schematic diagram of the multimodal fusion process; Figure 3 A flowchart illustrating the workflow for implementing dynamic segmentation logic; Figure 4 A flowchart for the design of the time decay function. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Please see Figures 1-4 This invention provides an online monitoring system for the status of wind farm equipment, the overall implementation of which is as follows: Real-time data acquisition module: This module deploys various types of sensors on wind farm equipment, such as vibration sensors, temperature sensors, current sensors, and environmental sensors, to synchronously acquire multi-source sensor data streams from the wind farm equipment. Vibration sensors collect vibration signals from the equipment, reflecting the operating status of its components. Temperature sensors acquire temperature distribution across different parts of the equipment, providing a clear picture of heat generation during operation. Current sensors collect current waveforms, allowing analysis of the equipment's power consumption characteristics. Environmental sensors collect environmental parameters such as wind speed, wind direction, and humidity, which affect equipment operation. Through the coordinated operation of these sensors, real-time, synchronous acquisition of multi-source data from the wind farm equipment is achieved.

[0027] Multimodal fusion module: This module performs adaptive weighted fusion based on the covariance matrix on the multi-source sensor data streams acquired by the real-time acquisition module to generate a comprehensive signal feature set. Specifically, it first extracts the frequency domain energy distribution features of the vibration signal. These features reflect the energy proportions of different frequency components of the vibration signal, helping to analyze the frequency characteristics of equipment vibration. Simultaneously, it extracts the spatial gradient features of the temperature distribution. These features show the temperature variation trend at different locations on the equipment, helping to identify areas of abnormal temperature changes. Harmonic component decomposition is performed on the current waveform to generate the current distortion coefficient. This coefficient measures the degree to which the current waveform deviates from an ideal sine wave, reflecting the power quality of the equipment. Then, based on the covariance matrix, the weight allocation values ​​of each sensor data are calculated. The frequency domain energy distribution features, spatial gradient features, and current distortion coefficient are then linearly weighted and fused according to these weight allocation values, ultimately generating a comprehensive signal feature set containing multi-source data features, achieving effective fusion of multi-source data.

[0028] The dynamic segmentation module segments the comprehensive signal feature set using variable time windows based on the equipment's operating load rate. In actual operation, the equipment's operating load rate is constantly changing. When the load rate is high, the equipment's operating state may change more frequently. In this case, the time window length is shortened to a first set value to more precisely capture changes in equipment state. When the load rate is low, the equipment's operating state is relatively stable, and the time window length is extended to a second set value to reduce data processing volume. For overlapping areas of adjacent time windows, a Gaussian weighting strategy is used for feature smoothing to avoid data abrupt changes caused by time window switching. Finally, a density clustering algorithm is used to cluster and segment the smoothed feature sequence according to its operating state, classifying the equipment's operating state into different categories to provide a basis for subsequent evaluation.

[0029] Preliminary evaluation module: The operating status feature sequence generated by the dynamic segmentation module is input into a pre-trained deep belief network to score the device health. The deep belief network performs unsupervised feature learning by inputting the operating status feature sequence into a restricted Boltzmann machine, mining latent features in the data. Then, the network parameters are fine-tuned through backpropagation to generate hidden layer feature representations, enabling the network to better fit the data. A sigmoid activation function is introduced into the output layer to map the hidden layer features to the health index range, thereby outputting the health index for each time period, intuitively reflecting the device's health status at different times.

[0030] The deep analysis module filters out abnormal time periods based on the health index output by the preliminary assessment module. The comprehensive signal features corresponding to these abnormal time periods are then input into a pre-trained spatiotemporal graph neural network for multi-scale correlation modeling. The spatiotemporal graph neural network constructs a static topology graph with device components as nodes and physical connections as edges, reflecting the structural relationships between device components. A time decay function is introduced to dynamically update the state dependency weights between nodes, considering the changes in device state over time. A graph attention mechanism aggregates the multi-time-segment state features of neighboring nodes, generating spatiotemporal embedding vectors to capture the device's feature information in the spatiotemporal dimension. Finally, the spatiotemporal embedding vectors are input into a fully connected layer for state label classification, generating device state labels to further clarify the device's operating status.

[0031] Maintenance Trigger Module: Based on the equipment status tags generated by the deep analysis module, real-time maintenance strategy instructions are triggered. If the equipment status tag is "Severe Anomaly," it indicates a serious risk of equipment failure. In this case, the equipment shutdown protocol is activated to ensure equipment safety, and an emergency repair work order is pushed out to arrange for professional personnel to carry out emergency repairs. If the equipment status tag is "Potential Risk," a preventative maintenance plan is generated to perform maintenance on the equipment in advance, while adjusting the power generation quota to minimize the impact on power generation while ensuring equipment safety. Finally, the maintenance operation records are synchronized to the equipment's full lifecycle management database to provide data support for subsequent equipment management and maintenance.

[0032] The implementation of the present invention will be further described below with reference to Examples 1 to 5.

[0033] Example 1: In online monitoring systems for wind farm equipment, multimodal fusion modules play a crucial role. When processing vibration signals, the Fast Fourier Transform (FFT) algorithm is used to convert the time-domain vibration signal to the frequency domain, calculate the energy values ​​of different frequency bands, and thus obtain the frequency domain energy distribution characteristics. For example, let the vibration signal be... The frequency domain signal is obtained after FFT transformation. Frequency domain energy distribution characteristics Through formula Calculation, where Indicates frequency, The energy spectrum of the frequency domain signal is calculated by this formula. This reflects the proportion of each frequency component in the total energy.

[0034] For the extraction of spatial gradient features of temperature distribution, the distribution location of temperature sensors on the device is used as coordinates to calculate the ratio of the temperature difference between adjacent sensors to the distance. This feature reflects the rate of temperature change in space.

[0035] When decomposing the harmonic components of a current waveform, Fourier series expansion is used to decompose the current waveform into the fundamental frequency and each harmonic component. Let the current waveform be... Its Fourier series expansion is , in The DC component, For the first The amplitude of the second harmonic. The fundamental angular frequency, For harmonic order, For the first The phase of the second harmonic. The current distortion coefficient is obtained by calculating the ratio of the amplitude of each harmonic to the amplitude of the fundamental wave. , Used to measure the degree of distortion in a current waveform.

[0036] When calculating the weight allocation values ​​for each sensor data based on the covariance matrix, the covariance matrix between each sensor data and the historical fault data is first calculated. Let the sensor data vector be... Historical fault data vector is covariance matrix elements The calculation formula is , in For the number of data samples, and They are respectively and The mean of the covariance matrix. Perform eigenvalue decomposition to extract principal component directions. Assume that eigenvalue decomposition yields eigenvalues. and the corresponding feature vector The fusion weights of each sensor's data are determined based on the variance contribution rate along the principal component direction. fusion weight according to Allocation can be performed, for example, through a certain normalization process, to ensure... Finally, based on these weights, the frequency domain energy distribution characteristics, spatial gradient characteristics, and current distortion coefficient are linearly weighted and fused to generate a comprehensive signal feature set, effectively integrating multi-source sensor data features and providing more comprehensive information for subsequent equipment status analysis.

[0037] Example 2: The dynamic segmentation module is a crucial component in the online monitoring system for wind farm equipment status, effectively dividing the operating status of the equipment. Regarding the dynamic adjustment of the time window length based on the equipment's operating load rate, the operating load rate can be obtained by measuring the ratio of the equipment's actual output power to its rated power. Assuming the equipment's rated power is... The actual output power is Then the load rate When the load rate Higher than the set high load rate threshold When the time window length is shortened to the first set value, When the load rate Below the set low load rate threshold When this happens, the time window length is extended to the second set value. For example, in a wind farm, the high load factor threshold... Low load threshold First setting value Seconds, second set value Seconds. In this way, when the equipment is running under high load, data can be collected and analyzed more frequently to detect potential problems in a timely manner; when running under low load, the data collection frequency can be appropriately reduced to decrease the amount of data processing.

[0038] For overlapping regions between adjacent time windows, a Gaussian weighted strategy is used for feature smoothing. Let the feature value of the overlapping region be... The Gaussian weighting function is ,in The mean of the Gaussian function can be set as the center position of the overlapping region; The standard deviation is used to control the width of the weighting function and is adjusted according to the characteristics of the actual data. For each feature value in the overlapping region... Multiply by the corresponding weighting value Then, a weighted average is performed to obtain smoothed feature values, which effectively avoids data mutations caused by time window switching and makes the data more continuous and stable.

[0039] When performing state-based clustering segmentation on smoothed feature sequences using density-based clustering algorithms, the local density and relative distance of the smoothed feature sequences are first calculated. For each data point... Local density Through formula Calculation, where The total number of data points. For data points and The distance between them can be calculated using Euclidean distance. The cutoff distance is a pre-defined parameter used to control the calculation range of local density. Relative distance. Defined as data point The distance between the nearest data point and the data point with a higher local density, i.e. Cluster centers are determined by density peak detection, and local density is selected. Larger and relative distance Larger data points are also used as cluster centers. This is based on a cutoff distance threshold. Non-centroids are clustered and assigned to the clusters of their nearest centroids. This density-based clustering algorithm can classify the sequence of equipment operating status features into different categories, clearly demonstrating the differences in equipment characteristics under different operating conditions and providing strong support for equipment status assessment.

[0040] Example 3: When constructing a deep belief network, the running state feature sequence generated by the dynamic segmentation module is first input into a Restricted Boltzmann Machine (RBM) for unsupervised feature learning. The RBM consists of a visible layer and hidden layers. It is assumed that the visible layer has... Each neuron corresponds to an input state feature; the hidden layer has The energy function of a restricted Boltzmann machine is defined as neurons. , in Visible layer neurons With hidden layer neurons Connection weights between them Visible layer neurons The bias, Hidden layer neurons The bias.

[0041] The training process of a Restricted Boltzmann Machine (RBM) employs a contrastive divergence algorithm to update the connection weights between the visible and hidden layers. During training, the running state feature sequence is first input to the visible layer. According to probability Sample the hidden layer neurons to obtain the hidden layer state. Then, based on probability... The visible layer is reconstructed to obtain the reconstructed visible layer state. Then based on probability Obtain the reconstructed hidden state Connection weights The update formula is ,in The learning rate is a pre-set parameter used to control the step size of weight updates. Through multiple iterations of training, the Restricted Boltzmann Machine (RBM) learns the latent features of the running state feature sequence.

[0042] After learning via a Restricted Boltzmann Machine (RBM), the network parameters are fine-tuned using backpropagation. The features learned by the RBM are used as input to subsequent network layers. The error between the network output and the true labels (which can be determined unsupervised during pre-training, such as clustering based on data similarity) is calculated. Then, based on this error, the weights and biases of each layer in the network are adjusted using backpropagation to better fit the data and generate more effective hidden feature representations. A Sigmoid activation function is introduced into the output layer to map the hidden features to a health index range. The Sigmoid function expression is: ,in As input values, the hidden layer features are fed into the Sigmoid function, and its output is mapped to... The range serves as a health index. For example, the closer the output value is to 1, the higher the health of the device, and the closer it is to 0, the lower the health of the device. This allows for a quantitative score of the device's health at different time periods, providing an intuitive basis for the preliminary assessment of the device's status.

[0043] Example 4: Spatiotemporal graph neural networks are used for multi-scale correlation modeling in the deep analysis module of a wind farm equipment condition online monitoring system. When constructing the spatiotemporal graph neural network, a static topology graph is built using equipment components as nodes and physical connections as edges. For example, in a wind turbine, components such as blades, hubs, gearboxes, and generators are considered as nodes, and edges are determined based on their mechanical and electrical connections. Assume the equipment has… Each component is represented by an adjacency matrix. Represents a static topology diagram, if components and components If there is a connection, then ,otherwise .

[0044] A time decay function is introduced to dynamically update the state dependency weights between nodes. An exponential decay coefficient is defined based on the physical wear and tear characteristics of the equipment components. The wear patterns of different components vary and can be determined through analysis of historical operating data and study of physical characteristics. Let the state dependency weight between two nodes be... The time interval is The dynamic weight adjustment factor is Then the update formula for the state-dependent weights is: Meanwhile, for those exceeding a preset time threshold... The dependency relationships are weighted to zero, that is, when hour, This effectively takes into account changes in equipment status over time, highlighting the impact of recent status on the current status.

[0045] A graph attention mechanism is used to aggregate multi-time-period state features of neighboring nodes, generating a spatiotemporal embedding vector. The graph attention mechanism weights and aggregates the features of neighboring nodes by calculating the attention coefficients between nodes. For each node... Its attention coefficient Through formula Calculation, where It is a learnable weight matrix. and These are nodes and nodes eigenvectors, It is a node The set of neighboring nodes is used, and LeakyReLU is an activation function used to increase the non-linear expressive power of the network. The node is obtained by weighted summation of the features of the neighboring nodes through attention coefficients. aggregation features The aggregated features from different time periods are combined to generate a spatiotemporal embedding vector.

[0046] Finally, the spatiotemporal embedding vector is input into a fully connected layer for state label classification. The fully connected layer consists of multiple neurons, each connected to all elements of the input spatiotemporal embedding vector. Let the weight matrix of the fully connected layer be... , bias is The input spatiotemporal embedding vector is The output of the fully connected layer By analyzing the output Softmax classification is performed to obtain device status labels. The output value is then converted into the probability of belonging to different status label categories using the Softmax function. The category with the highest probability is the device status label, thereby achieving accurate classification and in-depth analysis of device status.

[0047] Example 5: The maintenance triggering module is a crucial component in the online monitoring system for wind farm equipment status, ensuring the normal operation of the equipment. After the deep analysis module generates equipment status tags, the maintenance triggering module initiates corresponding real-time maintenance strategy commands based on the tag type.

[0048] If the equipment status label indicates a severe anomaly, it means the equipment is highly likely to have experienced a malfunction that affects normal operation or even poses a safety hazard. In this case, the system will immediately activate the equipment shutdown protocol. Upon activation, the system will send a shutdown command to the equipment's control system according to a pre-set procedure. This command will gradually stop the operation of each component of the equipment; for example, it will first control the blade pitch to reduce wind energy capture, then stop the generator's power generation, and simultaneously disconnect relevant electrical connections to ensure the equipment is in a safe shutdown state. At the same time, the system will push an emergency maintenance work order to the maintenance personnel's work terminal. The emergency maintenance work order records detailed basic equipment information, such as equipment model, serial number, and location, as well as a specific description of the abnormal status, including the time of the anomaly and potentially affected components. Upon receiving the work order, maintenance personnel can quickly understand the equipment problem, bring the appropriate tools and equipment, and rush to the site for emergency repairs, minimizing the impact of equipment failure on wind farm operation.

[0049] If the equipment status label is "potential risk," it indicates that although the equipment can still operate, there are certain hidden dangers that require proactive maintenance to prevent malfunctions. In this case, the system will generate a preventative maintenance plan. This plan will meticulously plan the maintenance time, content, and required resources based on the equipment's historical operating data, current status, and maintenance experience. For example, for a potential risk in the gearbox, the maintenance plan might be scheduled during a recent period of low wind speeds, involving maintenance operations such as oil checks and replacements, and gear wear inspections. Simultaneously, to minimize the impact on power generation while ensuring equipment safety, the system will adjust the power generation quota. By reducing the equipment's power generation, it operates under a relatively lower load, reducing the burden on the equipment and decreasing the likelihood of potential risks turning into actual failures. When adjusting the power generation, the system comprehensively considers the overall power generation of the wind farm, the operating status of other equipment, and the grid's needs to ensure the overall power generation efficiency and stability of the wind farm.

[0050] Whether performing emergency repairs or preventative maintenance, the system synchronizes maintenance operation records to the equipment's full lifecycle management database. These records contain detailed information such as the maintenance time, personnel, specific operations performed, and replaced parts. These records are crucial for the subsequent management and maintenance of the equipment. On one hand, managers can review these records to understand the equipment's maintenance history, evaluate maintenance effectiveness, and provide a reference for developing subsequent maintenance plans. On the other hand, when equipment problems arise, technicians can analyze the maintenance records to quickly locate potential issues, improving troubleshooting efficiency. Simultaneously, these records also help assess the overall operational status of the equipment, providing a basis for decisions regarding equipment upgrades and replacements, thereby achieving effective management of the equipment's entire lifecycle and ensuring the long-term stable operation of wind farm equipment.

[0051] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0052] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An online monitoring system for the status of wind farm equipment, characterized in that, The system includes: The real-time acquisition module is used to synchronously acquire multi-source sensor data streams from wind farm equipment. The sensor data streams include vibration signals, temperature distribution, current waveforms, and environmental parameters. The multimodal fusion module is used to perform adaptive weighted fusion of the multi-source sensor data streams based on the covariance matrix to generate a comprehensive signal feature set; The dynamic segmentation module performs variable time window segmentation on the comprehensive signal feature set based on the equipment operating load rate, generating multiple operating state feature sequences; The preliminary assessment module inputs the operating status feature sequence into a pre-trained deep belief network to score the device health and outputs the health index for each time period. The deep analysis module filters abnormal time periods based on the health index and inputs the corresponding comprehensive signal features into a pre-trained spatiotemporal graph neural network for multi-scale correlation modeling to generate device status labels. The maintenance triggering module triggers real-time maintenance strategy instructions based on the device status label.

2. The online monitoring system for wind farm equipment status as described in claim 1, characterized in that, The specific steps of the multimodal fusion module include: Extract the frequency domain energy distribution characteristics and the spatial gradient characteristics of the temperature distribution of the vibration signal; Harmonic component decomposition is performed on the current waveform to generate the current distortion coefficient; Calculate the weight allocation values ​​for each sensor data based on the covariance matrix; Based on the weighted values, the frequency domain energy distribution characteristics, spatial gradient characteristics, and current distortion coefficient are linearly weighted and fused to generate a comprehensive signal feature set.

3. The online monitoring system for wind farm equipment status as described in claim 1, characterized in that, The implementation logic of the dynamic segmentation module includes: The time window length is dynamically adjusted according to the equipment operating load rate. Under high load rate, the window length is shortened to the first set value, and under low load rate, the window length is extended to the second set value. A Gaussian weighted strategy is used to smooth the features in the overlapping areas of adjacent time windows; The smoothed feature sequence is segmented by running state clustering based on the density clustering algorithm.

4. The online monitoring system for wind farm equipment status as described in claim 1, characterized in that, The construction steps of the deep belief network include: The operating state feature sequence is input into a restricted Boltzmann machine for unsupervised feature learning; The network parameters are fine-tuned using the backpropagation algorithm to generate hidden layer feature representations; A sigmoid activation function is introduced into the output layer to map the hidden layer features to the health index range.

5. The online monitoring system for wind farm equipment status as described in claim 1, characterized in that, The modeling steps of the spatiotemporal graph neural network include: A static topology graph is constructed using equipment components as nodes and physical connections as edges; A time decay function is introduced to dynamically update the state dependency weights between nodes; A spatiotemporal embedding vector is generated by aggregating the multi-time-period state features of neighboring nodes through a graph attention mechanism. The spatiotemporal embedding vector is input into a fully connected layer for state label classification.

6. The online monitoring system for wind farm equipment status as described in claim 2, characterized in that, The weighting method for the covariance matrix includes: Calculate the covariance matrix between each sensor's data and historical fault data, perform eigenvalue decomposition on the covariance matrix, and extract the principal component directions; The fusion weights of each sensor's data are determined based on the variance contribution rate of the principal component direction.

7. The online monitoring system for wind farm equipment status as described in claim 3, characterized in that, The implementation steps of the density clustering algorithm include: Calculate the local density and relative distance of the smoothed feature sequence, determine the cluster center point by density peak detection, and classify the non-center points into clusters based on the truncation distance threshold.

8. The online monitoring system for wind farm equipment status as described in claim 4, characterized in that, The training process of the restricted Boltzmann machine includes updating the connection weights between the visible layer and the hidden layer using a contrastive divergence algorithm.

9. The online monitoring system for wind farm equipment status as described in claim 5, characterized in that, The design of the time decay function includes: An exponential decay coefficient is defined based on the physical wear and tear of equipment components. The decay coefficient is multiplied by the time interval to generate a dynamic weight adjustment factor. Dependencies that exceed a preset time threshold are weighted to zero.

10. The online monitoring system for wind farm equipment status as described in claim 1, characterized in that, The triggering logic for the real-time maintenance strategy command includes: If the equipment status label is critically abnormal, the equipment shutdown protocol will be initiated and an emergency maintenance work order will be sent. If the equipment status label indicates a potential risk, a preventative maintenance plan is generated and the power generation quota is adjusted. Synchronize maintenance operation records to the equipment lifecycle management database.

Images