Real-time monitoring method for running states of wind power generation host and tower drum

Abstract

The invention relates to the technical field of wind power generation system monitoring, and discloses a real-time monitoring method for the running state of a wind power generation host and a tower drum, which comprises the following steps of: installing sensors at key parts such as bearings and blade connecting parts of the host and the tower drum to collect running data samples; a risk assessment model and a state distribution model are respectively constructed by adopting deep learning and a clustering analysis algorithm, and a comprehensive monitoring model is formed through fusion. And solving a key performance index threshold boundary by using an optimization algorithm based on gradient descent, screening an early warning scheme set according to a dynamic priority decision-making method, implementing a maintenance strategy, and feeding back and adjusting according to new data. According to the method, data can be accurately collected, the operation state can be comprehensively evaluated, faults can be timely warned, stable operation of equipment is guaranteed through closed-loop maintenance, system reliability is enhanced through redundancy check of a fusion processing layer, and reliability, stability and economical efficiency of a wind power generation system are effectively improved.

Description

Technical Field

[0001] The invention relates to the technical field of wind power generation system monitoring, and in particular to a real-time monitoring method for the operating status of a wind power generation main engine and a tower. Background Art

[0002] Against the backdrop of global energy transformation, wind power generation, as a clean and renewable way to obtain energy, is becoming increasingly important in the power supply system. In recent years, the installed capacity of wind power generation has continued to rise, and the scale of wind farms has continued to expand and their distribution has become more and more extensive. However, as the core and key components of the wind power generation system, the operating stability and reliability of the wind turbine main engine and tower are directly related to the power generation efficiency, economic benefits and stability of the power supply of the entire wind farm. However, in actual operation, the main engine and tower face extremely complex and harsh working environments, which makes their operating status monitoring a very challenging task. There are many problems with existing monitoring technologies and they need to be improved urgently.

[0003] Wind turbines and towers are operated under severe weather conditions such as high altitude, strong wind, low temperature, and dust for a long time. For example, in wind farms in the north, the temperature in winter can be as low as minus 20 to 30 degrees Celsius. The viscosity of the lubricating oil in the main engine increases, affecting the normal lubrication between components and causing increased mechanical wear. In coastal areas, high humidity and salt spray environments can easily cause corrosion of the tower and weaken its structural strength. At the same time, the main engine and tower are also subject to alternating loads during operation. During the rotation of the blades, they not only have to overcome air resistance, but also have to cope with frequent changes in wind direction and wind speed. This makes the blade connection parts, main engine bearings and other key components always in a complex stress state, which is very easy to cause fatigue damage, and then lead to failures. According to statistics, the downtime of wind turbines caused by main engine and tower failures can reach tens or even hundreds of hours per year on average, seriously affecting the power generation income of wind farms.

[0004] At present, although there are various monitoring methods for the operating status of wind turbine mainframes and towers, they all have certain limitations. Traditional sensor-based monitoring technologies, such as relying only on simple temperature sensors and vibration sensors to monitor the operating status of equipment, can often only obtain data in a single dimension and cannot fully reflect the overall operating status of the equipment. For example, by only monitoring the temperature, it is impossible to know the abnormal vibration of equipment components, which may be an important signal that the equipment is about to fail. Moreover, when these sensors work in harsh environments for a long time, they are prone to problems such as reduced accuracy and signal drift, which greatly reduces the reliability of the monitoring data.

[0005] Monitoring methods based on data analysis also face difficulties. Some monitoring methods based on threshold judgment rely on pre-set fixed thresholds to determine whether the equipment is faulty. However, the operating conditions of wind power generation equipment are complex and changeable. Under different wind speeds, wind directions, loads and other conditions, the normal operating parameter range of the equipment will also change. Fixed thresholds are difficult to adapt to these complex operating conditions and are prone to false alarms or missed alarms. For example, in strong winds, the vibration amplitude of the tower will naturally increase, but it may still be within a safe range. If it is judged according to conventional thresholds, it may be misjudged as a fault, causing unnecessary downtime for maintenance and increasing operation and maintenance costs.

[0006] In addition, most existing monitoring systems lack the ability to effectively assess potential failure risks. They can only issue an alarm when a failure has already occurred or obvious failure characteristics appear. They cannot predict the occurrence of a failure in advance, resulting in the inability to take effective preventive measures in a timely manner. Waiting until a failure occurs before repairing it will not only be costly, but will also cause long power outages and affect the stability of power supply. For example, when the gearbox of the main engine is severely worn, repairs require the replacement of a large number of parts, which consumes a lot of manpower, material resources and time, causing huge economic losses to the wind farm.

[0007] In terms of the comprehensive utilization of monitoring data, existing technologies are also insufficient. There is a lack of effective fusion mechanism between different types of monitoring data, and it is impossible to form a comprehensive and accurate understanding of the operating status of the equipment. For example, vibration data and temperature data are collected and analyzed by different monitoring systems respectively, and there is no association between the two. It is impossible to comprehensively judge the operating status of the equipment from multiple angles, and it is difficult to find fault clues hidden behind complex data. With the continuous development of wind power generation technology, the single-unit capacity of wind turbines is getting larger and larger, and the tower height is constantly increasing, which puts higher requirements on the operating status monitoring of the main engine and tower. The existing monitoring technology can no longer meet the needs of efficient and stable operation of wind farms. There is an urgent need for a more comprehensive, accurate and intelligent monitoring method that can grasp the operating status of wind power main engines and towers in real time and accurately, discover potential fault risks in a timely manner, and formulate reasonable maintenance strategies to ensure the reliable operation of wind power generation systems, reduce operation and maintenance costs, and improve the economic and social benefits of wind farms. Summary of the invention

[0008] The object of the present invention is to provide a method for real-time monitoring of the operating status of a wind power generator main engine and a tower, so as to solve the problems raised in the above-mentioned background technology.

[0009] To achieve the above object, the present invention provides the following technical solution: a real-time monitoring method for the operating status of a wind power generator main engine and a tower, the method comprising: F1. Collect operation data samples using sensors placed at key locations of wind turbine mainframes and towers; F2. Use deep learning algorithms to build a risk assessment model that can infer the potential failure risks of the main engine and tower based on the operating data samples; use cluster analysis algorithms to build a state distribution model that can reflect the distribution characteristics of different operating states of the main engine and tower based on the operating data samples; connect the outputs of the state distribution model and the risk assessment model to a fusion processing layer to form a comprehensive monitoring model; F3. Solve the key performance indicator threshold boundary of the comprehensive monitoring model, and select the early warning solution set that matches the safe operation requirements in the key performance indicator threshold boundary according to the dynamic priority decision method; F4. Implement the corresponding maintenance strategy according to the early warning plan set. If the subsequent monitoring shows that the operating status of the main engine and the tower has returned to normal, stop the operation. Otherwise, feed back the newly collected operating data samples as input information to the comprehensive monitoring model, and repeat F3 to F4.

[0010] The real-time monitoring method of the present invention collects operating data by arranging sensors at key locations, and uses deep learning algorithms and cluster analysis algorithms to respectively construct risk assessment models and state distribution models, which work together to comprehensively evaluate the operating conditions of the main engine and tower. The outputs of the two models are fused to form a comprehensive monitoring model, so as to more accurately judge the operating status. After that, the threshold boundaries of key performance indicators are solved, and appropriate early warning schemes are screened out. Finally, maintenance strategies are implemented based on the early warnings, and whether to continue monitoring and adjustment is determined based on the subsequent operating status, forming a complete, closed-loop real-time monitoring system to ensure the safe and stable operation of the wind power main engine and tower.

[0011] Preferably, in step F1, sensors are installed at the bearings, blade connection parts, and tower bottom of the wind power main engine and tower according to a preset scientific layout plan to collect operation data samples. Sensors are installed at key parts such as the bearings, blade connection parts, and tower bottom of the wind power main engine and tower because the operating state of the bearings is directly related to the rotation stability of the main engine, the blade connection parts affect the normal operation of the blades, and the bottom of the tower is under great pressure and vibration. The operating data of these parts can effectively reflect the operating status of the overall equipment. Through the preset scientific layout plan, it can be ensured that the sensors can comprehensively and accurately collect various types of operation data, providing a reliable basis for subsequent analysis.

[0012] Preferably, the input of the comprehensive monitoring model is various real-time operating data of the wind power host and tower, and the output is quantitative indicators that characterize the quality of the operating status and the corresponding fault risk level. The comprehensive monitoring model uses the real-time operating data of the wind power host and tower as input, and these data cover a variety of parameters during the operation of the equipment. After calculation and analysis within the model, the output quantitative indicators that characterize the quality of the operating status can intuitively reflect the current operating quality of the equipment, and the corresponding fault risk level can allow operation and maintenance personnel to quickly understand the possibility of equipment failure, so that they can take corresponding measures in a timely manner.

[0013] Preferably, the key performance indicator threshold boundary is composed of a set of quantitative indicator values ​​that can accurately distinguish different operating states, and is obtained by iteratively solving the comprehensive monitoring model through an optimization algorithm based on gradient descent. The specific formula is:

[0014] in is the cost function, is the predicted value, is the true value, is the regularization parameter, are model parameters, is the sample size, is the number of parameters, by continuously adjusting Make Minimize to determine the threshold boundary. Determining the threshold boundary of key performance indicators is crucial to accurately determine the operating status of the equipment. This set of quantitative indicator values ​​is obtained by iteratively solving the comprehensive monitoring model using the above formula using the gradient descent optimization algorithm. In the formula, the cost function The difference between the predicted value and the true value and the regularization term of the model parameters are comprehensively considered. , minimize the cost function, and then determine the threshold boundaries that can accurately distinguish different operating states, providing an accurate reference standard for subsequent early warning and decision-making.

[0015] Preferably, the optimization algorithm based on gradient descent is a stochastic gradient descent algorithm. During the iteration process, the formula for updating the parameters is:

[0016] in is the learning rate. The stochastic gradient descent algorithm is a commonly used optimization algorithm, which is used in the present invention to update the parameters of the comprehensive monitoring model. During the iteration process, the parameters are updated according to the above formula. Learning rate It controls the step size of each parameter update, and its size affects the convergence speed and effect of the algorithm. The algorithm continuously iterates and updates the parameters to optimize the model in the direction of minimizing the cost function, thereby more accurately determining the threshold boundary of the key performance indicator.

[0017] Preferably, the construction of the comprehensive monitoring model includes: Using a multi-layer perceptron to fit the operation data samples collected in step F1, and dividing the operation data samples into a training sample set and a verification sample set according to a specific ratio; Based on the training sample set, the cross entropy loss function and L2 regularization term are used as the two objectives in the dual-objective optimization problem; and the adaptive moment estimation optimization algorithm is used to obtain the convergence boundary of the corresponding risk assessment model training error, and then the dynamic priority decision method is used to select a predetermined number of candidate solution sets in the convergence boundary of the risk assessment model training error, and the network parameters of the risk assessment model are determined according to the final performance of the candidate sets on the verification data; according to the performance of the models corresponding to different solution sets on the verification sample set, a set of network parameters with the best performance is finally selected to obtain the risk assessment model; A variational autoencoder that simulates the distribution characteristics of the original operation data samples is built, and network parameters are trained based on the training sample set. When the preset training termination condition is reached, the decoder in the variational autoencoder is used as the state distribution model; then the state distribution model and the output of the trained risk assessment model are connected to a fusion processing layer to form a comprehensive monitoring model. The construction of the comprehensive monitoring model first uses a multilayer perceptron to fit the operation data samples, and divides the training sample set and the verification sample set to prepare for subsequent training and evaluation of the model. When constructing the risk assessment model, the cross entropy loss function and the L2 regularization term are used as optimization targets, and the adaptive moment estimation optimization algorithm is used to find the convergence boundary of the training error. Then, the candidate solution set is screened by the dynamic priority decision method, and the network parameters are determined according to the verification data performance, so as to obtain a risk assessment model with excellent performance. For the state distribution model, a variational autoencoder is built and trained, and the decoder that meets the conditions is used as the state distribution model. Finally, the outputs of the two models are connected to the fusion processing layer to form a comprehensive monitoring model, which has more powerful monitoring capabilities.

[0018] Preferably, the fusion processing layer introduces an attention mechanism to achieve fusion, and calculates the respective attention scores for the operating state distribution feature vector output by the state distribution model and the fault risk quantification value output by the risk assessment model. Introducing the attention mechanism in the fusion processing layer can calculate the respective attention scores according to the importance of the operating state distribution feature vector output by the state distribution model and the fault risk quantification value output by the risk assessment model. This enables the model to pay more attention to important information and ignore relatively minor information when fusing information, thereby improving the fusion effect and more accurately reflecting the actual operating status of the equipment.

[0019] Preferably, the fusion processing layer establishes a redundant check module during operation to cross-validate the fused data. The fused comprehensive vector is divided into a check subset and a comparison subset according to a certain ratio, and the pre-trained check model is used to extract features and compare the two subsets respectively. If the difference between the two exceeds the preset threshold, it is determined that there may be an abnormality in the fusion process, triggering the error correction process, and retraining the fusion processing layer parameters using historical normal fusion data. The redundant check module of the fusion processing layer provides a guarantee for the accuracy of data fusion. During operation, the fused comprehensive vector is divided into a check subset and a comparison subset, and features are extracted and compared by using a pre-trained check model. Once the difference between the two exceeds the preset threshold, it means that there may be a problem in the fusion process. At this time, the error correction process is triggered, and the fusion processing layer parameters are retrained using historical normal fusion data to ensure the reliability of the fusion data and the accuracy of the monitoring results.

[0020] Preferably, in step F1, the collected operation data samples are subjected to outlier elimination processing, and smoothing and noise reduction are performed using a sliding window algorithm. After collecting the operation data samples, outlier elimination processing of the data can avoid interference of abnormal data on subsequent analysis and model training, and ensure the authenticity and reliability of the data. And using a sliding window algorithm for smoothing and noise reduction can remove noise in the data, making the data smoother and more stable, and improving the data quality.

[0021] Compared with the prior art, the present invention has the following beneficial effects: The present invention uses a deep learning algorithm to construct a risk assessment model, which can deeply explore the potential laws in the operating data samples and accurately infer the potential failure risks of the main engine and tower. Taking a certain wind farm as an example, by learning a large amount of historical fault data and normal operating data, the risk assessment model can accurately predict the possibility of equipment failure under different operating conditions. The state distribution model constructed by the cluster analysis algorithm can clearly reflect the distribution characteristics of different operating states of the main engine and tower, such as the classification of the operating state of the equipment under different wind speeds and load conditions. The outputs of these two models are connected to the fusion processing layer, and the attention mechanism is introduced to perform weighted fusion according to the importance of the operating state distribution feature vector and the fault risk quantification value to form a comprehensive monitoring model. The operating state quantitative indicators and fault risk levels output by this model can comprehensively and accurately reflect the actual operating status of the equipment and provide a reliable basis for subsequent decision-making.

[0022] The present invention solves the threshold boundaries of key performance indicators of the comprehensive monitoring model through an optimization algorithm based on gradient descent, and can accurately distinguish different operating states of the equipment. Taking the vibration amplitude indicator as an example, by continuously iterating and optimizing the threshold boundaries determined, it is possible to accurately judge whether the vibration of the equipment is within the normal range. The early warning scheme set is screened according to the dynamic priority decision-making method, taking into full account factors such as the severity of the fault, the impact on equipment operation, and maintenance costs. For faults that may cause serious consequences and have low maintenance costs, high-priority warnings are issued first to remind operation and maintenance personnel to take timely measures. This greatly improves the pertinence and effectiveness of fault warnings, allowing operation and maintenance personnel to respond in time before the fault occurs, take corresponding maintenance measures, avoid the expansion of faults, and reduce equipment downtime.

[0023] The present invention implements maintenance strategies according to the early warning scheme set, and feeds back the operation data after maintenance to the comprehensive monitoring model to form a closed-loop monitoring and maintenance system. If the operating status of the equipment returns to normal, the operation is stopped; if not, the monitoring, evaluation and maintenance process is repeated. For example, when the temperature of the main engine bearing is detected to be abnormally high, the early warning scheme instructs the bearing to be inspected and lubricated. After processing, the newly collected data is input into the comprehensive monitoring model again for evaluation. If the temperature returns to normal, it means that the maintenance measures are effective; if it is still abnormal, the cause is further analyzed and the maintenance strategy is adjusted. This closed-loop maintenance strategy ensures that the equipment is always in a good operating state, improves the reliability and stability of the equipment, extends the service life of the equipment, reduces the operation and maintenance costs, and provides a strong guarantee for the continuous and stable operation of the wind power generation system.

[0024] The fusion processing layer of the present invention establishes a redundant verification module to cross-validate the fused data. The fused comprehensive vector is divided into a verification subset and a comparison subset, and feature extraction and comparison analysis are performed respectively using a pre-trained verification model. Once the difference between the two exceeds a preset threshold, it is determined that there may be an abnormality in the fusion process and the error correction process is triggered, and the fusion processing layer parameters are retrained using historical normal fusion data. This mechanism effectively avoids monitoring misjudgments caused by data fusion errors, greatly enhances the reliability and stability of the entire monitoring system, ensures the accuracy of the monitoring results, and makes operation and maintenance decisions more reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a working principle diagram of the real-time monitoring method of the present invention; Figure 2 A flow chart for the construction of a comprehensive monitoring model; Figure 3 Flowchart for cross-validation of fused data by the redundancy check module. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] See also Figure 1-3 The present invention provides a real-time monitoring method for the operating status of a wind power generator and a tower, aiming to achieve accurate monitoring of the operating status of the wind power generator and the tower, timely discover potential fault risks and take effective maintenance measures. The overall implementation scheme is as follows: F1: Install sensors at key locations of wind turbine mainframes and towers according to a preset scientific layout plan to collect operation data samples. These key locations include bearings, blade connection locations, tower bottom, etc. Sensors collect various types of data, such as vibration data, temperature data, stress data, etc., to provide a basis for subsequent analysis.

[0028] F2: Use deep learning algorithms to build a risk assessment model, infer the potential failure risks of the main engine and tower based on the operating data samples; use cluster analysis algorithms to build a state distribution model to reflect the distribution characteristics of different operating states of the main engine and tower. Then connect the output of the state distribution model and the risk assessment model to the fusion processing layer to form a comprehensive monitoring model. This comprehensive monitoring model takes various real-time operating data of wind power main engines and towers as input, and outputs quantitative indicators that characterize the quality of the operating state and the corresponding fault risk level.

[0029] F3: Solve the threshold boundary of the key performance indicators of the comprehensive monitoring model. The threshold boundary consists of a set of quantitative indicator values ​​that can accurately distinguish different operating states. It is obtained by iteratively solving the comprehensive monitoring model based on the gradient descent optimization algorithm. In the threshold boundary of the key performance indicators, the early warning solution set that matches the safe operation requirements is selected according to the dynamic priority decision method.

[0030] F4: Implement the corresponding maintenance strategy according to the early warning plan set. If the subsequent monitoring shows that the operating status of the main engine and tower has returned to normal, stop the operation; if it has not returned to normal, feed the newly collected operating data samples as input information to the comprehensive monitoring model, and repeat F3 to F4 for continuous monitoring and maintenance to ensure the safe and stable operation of wind power generation equipment.

[0031] The implementation of the present invention is further described below in conjunction with Examples 1 to 5.

[0032] Embodiment 1:

[0033] This embodiment describes in detail the installation locations of sensors at key locations of the wind turbine main engine and tower and the data collection preprocessing process to ensure that accurate and effective operation data samples are collected, providing a reliable data basis for subsequent monitoring and analysis.

[0034] In a wind turbine generator in a wind farm, sensors are installed according to a preset scientific layout plan. High-precision vibration sensors and temperature sensors are selected for the bearings of the main engine. For example, a vibration sensor is installed at the main bearing of the generator, which can accurately measure the vibration amplitude, frequency and other parameters of the bearing in the radial and axial directions. At the same time, a temperature sensor is installed to monitor the temperature changes of the bearing in real time. This is because during the operation of the bearing, abnormal changes in vibration and temperature are often early signals of failure. By accurately measuring these parameters, potential problems with the bearing can be discovered in a timely manner.

[0035] Install strain gauge sensors and displacement sensors at the blade connection. Take the connection between the blade root and the hub as an example, paste [specific model of strain gauge] strain gauge to measure the stress changes at the connection under different working conditions. Install [specific model of displacement sensor] displacement sensor to monitor the tiny displacement of the blade connection. During the high-speed rotation of the blade, the connection is subjected to huge loads. Abnormal changes in stress and displacement may cause the blade to loosen or even fall off, so monitoring these parameters is crucial.

[0036] At the bottom of the tower, install a pressure sensor and an acceleration sensor. Install [Specific model of pressure sensor] pressure sensor on the foundation ring at the bottom of the tower to monitor the pressure at the bottom of the tower in real time. Install [Specific model of acceleration sensor] acceleration sensor on the outer wall at the bottom of the tower to measure the vibration acceleration of the tower under wind load. The bottom of the tower is a key part that withstands wind force and the weight of the equipment itself. Changes in pressure and acceleration reflect the structural health of the tower.

[0037] During the data collection process, the collected operation data samples are processed for outliers. For example, if the vibration amplitude data collected by the vibration sensor at a certain moment is far beyond the normal operating range and has poor correlation with other sensor data, it is judged to be an outlier and will be removed. At the same time, the sliding window algorithm is used for smoothing and noise reduction. Assuming that the sliding window size is 5, the data collected by the temperature sensor is processed. Taking the temperature data sequence [25, 26,24, 30, 27, 28, 26] collected in a certain period of time as an example, in the first window [25, 26, 24, 30, 27], the average value of the data in the window is calculated as (25 + 26 + 24 + 30 + 27) ÷ 5 = 26.4, and the average value is used as the smoothed data at the center of the window (i.e., the third data position). The windows are slided in sequence to process the entire data sequence to obtain a smoothed temperature data sequence, effectively removing noise interference in the data and improving data quality.

[0038] Embodiment 2:

[0039] This embodiment describes in detail the construction process of the comprehensive monitoring model, including the construction of the risk assessment model and the state distribution model and the formation of the fusion processing layer. Taking the operation data samples collected from a large wind farm as an example, a multi-layer perceptron is used to fit the collected operation data samples. Assuming that the collected operation data samples contain data in multiple dimensions such as vibration, temperature, wind speed, and power, these data are first normalized and mapped to the [0, 1] interval to facilitate model training. Then, the operation data samples are divided into a training sample set and a verification sample set in a ratio of 8:2.

[0040] Based on the training sample set, a risk assessment model is constructed. The cross entropy loss function and L2 regularization term are used as the two objectives in the dual-objective optimization problem. The cross entropy loss function can measure the difference between the model prediction value and the true value, and the L2 regularization term is used to prevent the model from overfitting. The adaptive moment estimation optimization algorithm (Adam algorithm) is used to obtain the convergence boundary of the risk assessment model training error. During the training process, the learning rate is set to 0.001, β1 = 0.9, β2 = 0.999, ε = 1e-8 and other parameters. As the training progresses, the model continuously adjusts the network parameters so that the sum of the cross entropy loss function and the L2 regularization term gradually decreases. When the training error converges to a certain extent, the dynamic priority decision method is used to select a predetermined number (such as 10) of candidate solution sets in the convergence boundary of the risk assessment model training error. According to the final performance of the candidate set on the verification data, such as accuracy, recall rate and other indicators, the network parameters of the risk assessment model are determined. After multiple experiments and comparisons, a set of network parameters with the best performance is finally selected to obtain the risk assessment model.

[0041] Build a variational autoencoder that simulates the distribution characteristics of the original operation data samples. The variational autoencoder consists of an encoder and a decoder. The encoder maps the input operation data samples to a low-dimensional space, and the decoder reconstructs the low-dimensional vector into the original data sample. Network parameter training is performed based on the training sample set, and the maximum number of training rounds is set to 100. When the reconstruction error is less than the preset threshold (such as 0.01), the preset training termination condition is reached. At this time, the decoder in the variational autoencoder is used as a state distribution model. The state distribution model can learn the potential distribution characteristics of the operation data samples and reflect the distribution of different operating states of the host and tower.

[0042] Finally, the output of the state distribution model and the trained risk assessment model are connected to a fusion processing layer to form a comprehensive monitoring model. The fusion processing layer fuses the operating state distribution feature vector output by the state distribution model with the fault risk quantification value output by the risk assessment model, and outputs a quantitative indicator representing the quality of the operating state and the corresponding fault risk level.

[0043] Embodiment 3:

[0044] This embodiment describes in detail how the fusion processing layer introduces the attention mechanism to achieve the fusion of the state distribution model and the risk assessment model output, and ensures the accuracy and reliability of the fusion data by establishing a redundant verification module to improve the performance of the comprehensive monitoring model.

[0045] In the fusion processing layer of the comprehensive monitoring model, the attention mechanism is introduced to achieve fusion. Taking the monitoring data at a certain moment as an example, the operating state distribution feature vector output by the state distribution model is [0.2, 0.3, 0.1, 0.4], and the fault risk quantification value output by the risk assessment model is 0.6. First, the state distribution feature vector and the fault risk quantification value are linearly transformed to obtain two new vectors. Assume that the state distribution feature vector is transformed to [0.15, 0.25, 0.08, 0.32] and the fault risk quantification value is transformed to [0.55] through the linear transformation matrices W1 and W2. Then, the respective attention scores are calculated. The dot product method is used to calculate the dot product of the transformed vector of the state distribution feature vector and the transformed vector of the fault risk quantification value to obtain the attention score. For example, (0.15×0.55)+(0.25×0.55)+(0.08×0.55)+(0.32×0.55)= 0.429. According to the attention score, the state distribution feature vector and the fault risk quantization value are weighted and fused to obtain the fused comprehensive vector. Assuming that the weights corresponding to the attention score are 0.429 and 1 - 0.429 = 0.571, the fused comprehensive vector is [0.2×0.429 + 0.6×0.571, 0.3×0.429 + 0.6×0.571, 0.1×0.429 + 0.6×0.571, 0.4×0.429 + 0.6×0.571] = [0.473, 0.504, 0.413, 0.527].

[0046] During the operation of the fusion processing layer, a redundant verification module is established to cross-validate the fused data. The fused comprehensive vector is divided into a verification subset and a comparison subset in a ratio of 6:4. For example, if the fused comprehensive vector is [0.473, 0.504, 0.413, 0.527], then the verification subset is [0.473, 0.504], and the comparison subset is [0.413, 0.527]. Use the pre-trained verification model to perform feature extraction and comparison analysis on the two subsets respectively. The verification model can be a simple neural network, which can extract the characteristic pattern of the fused data by learning a large amount of normal fusion data. Assume that the features extracted by the verification model for the verification subset are [0.4, 0.5], and the features extracted for the comparison subset are [0.42, 0.53]. Calculate the difference between the two. If the Euclidean distance is used for calculation, the difference value is If the difference between the two exceeds the preset threshold (such as 0.05), it is determined that there may be an abnormality in the fusion process, triggering the error correction process. In the error correction process, the historical normal fusion data is used to retrain the fusion processing layer parameters, adjust the weights of the attention mechanism and other parameters, ensure that the fusion processing layer can accurately fuse data, and improve the reliability of the comprehensive monitoring model.

[0047] Embodiment 4:

[0048] This embodiment describes in detail the process of solving the threshold boundaries of the key performance indicators of the comprehensive monitoring model, and determines the quantitative indicator values ​​that can accurately distinguish different operating states through an optimization algorithm based on gradient descent.

[0049] Taking the comprehensive monitoring model of a certain type of wind turbine main engine and tower as an example, the threshold boundary of key performance indicators is solved. The threshold boundary of key performance indicators is composed of a set of quantitative indicator values ​​that can accurately distinguish different operating states. It is obtained by iteratively solving the comprehensive monitoring model through the optimization algorithm based on gradient descent. The formula used is:

[0050] in is the cost function, is the predicted value, is the true value, is the regularization parameter, are model parameters, is the sample size, is the number of parameters.

[0051] Assume that in a certain training stage, the number of samples m = 100, the number of model parameters n = 5, and the regularization parameter β = 0.01. Initially, the value of the model parameter θ is randomly set to [0.1, 0.2, 0.3, 0.4, 0.5]. For each sample i, calculate the predicted value With the true value The difference between For example, for sample 1, , the predicted value is calculated by the comprehensive monitoring model = 0.7, true value = 0.8, then the contribution of this sample to the cost function is . Calculate the contribution of all samples to the cost function in turn and add the regularization term , get the current cost function value .

[0052] The stochastic gradient descent algorithm is used to update the parameters in the iterative process. The formula for updating the parameters is:

[0053] in is the learning rate, assuming that the learning rate α = 0.001. Calculate the cost function For each parameter The partial derivative of For example, for the parameter , calculate the partial derivatives by the chain rule, and then update the parameters according to the update formula Update all parameters in sequence to complete one iteration.

[0054] In the iterative process, the value of the parameter $θ$ is continuously adjusted so that the cost function Gradually decreases. After multiple iterations (such as 1000 iterations), when the cost function When convergence reaches a certain degree (such as the change is less than 0.0001), the parameters determined at this time are The corresponding quantitative index values ​​are the threshold boundaries of key performance indicators. These threshold boundaries can accurately distinguish the different operating states of wind turbine mainframes and towers, and provide a basis for the subsequent screening of early warning plans based on the dynamic priority decision-making method.

[0055] Embodiment 5:

[0056] This embodiment describes in detail the process of screening the early warning solution set according to the comprehensive monitoring model output and the key performance indicator threshold boundary, and implementing the maintenance strategy according to the early warning solution set.

[0057] In the real-time monitoring system of a certain wind farm, the comprehensive monitoring model takes various real-time operating data of the wind turbine main engine and tower as input, and outputs quantitative indicators that characterize the quality of the operating status and the corresponding fault risk level. For example, at a certain moment, the quantitative indicators of the operating status output by the comprehensive monitoring model are [0.6, 0.7, 0.5, 0.8], and the corresponding fault risk level is "medium risk". According to the threshold boundary of the key performance indicator obtained in advance, it is judged whether the current operating status exceeds the normal range. Assume that for a certain key performance indicator, the threshold boundary is [0.4, 0.6], and the current quantitative value of the indicator is 0.7, which exceeds the threshold boundary, indicating that there is a potential fault risk.

[0058] According to the dynamic priority decision method, the early warning scheme set that matches the safety operation requirements is screened within the threshold boundary of the key performance indicators. The dynamic priority decision method comprehensively considers factors such as the fault risk level, the impact of the fault on the equipment operation, and the maintenance cost. For example, for faults with "medium risk" that have a greater impact on the equipment operation and a lower maintenance cost, the early warning scheme of immediate shutdown for inspection and repair is preferred; for faults with "low risk" that have a smaller impact on the equipment operation, the early warning scheme of enhanced monitoring and handling during regular equipment maintenance is selected. After screening, it is determined that the early warning scheme set for the current situation is "immediately shut down to inspect the generator bearings, and replace the bearings if wear is found."

[0059] Implement the corresponding maintenance strategy according to the early warning scheme set. When the early warning of "stop the machine immediately to check the generator bearings, and replace the bearings if wear is found" is issued, the staff will shut down the wind turbine in accordance with the operating procedures, and then check the generator bearings. If the inspection finds that the bearings are slightly worn, replace them with new ones according to the early warning scheme. After replacing the bearings, continue to monitor the operating status of the main engine and tower. If the subsequent monitoring shows that the operating status has returned to normal, such as parameters such as vibration and temperature are within the normal range, and the fault risk level is reduced to "low risk", stop the operation. If the operating status has not returned to normal, the newly collected operating data samples are fed back to the comprehensive monitoring model as input information, and the process of solving the threshold boundaries of key performance indicators, screening the early warning scheme set, and implementing maintenance strategies is repeated until the operating status returns to normal, ensuring the safe and stable operation of wind power generation equipment.

[0060] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0061] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A real-time monitoring method for the operating status of a wind power generator and a tower, characterized in that: The method comprises: F1. Collect operation data samples using sensors placed at key locations of wind turbine mainframes and towers; F2. Use deep learning algorithms to build a risk assessment model that can infer the potential failure risks of the main engine and tower based on the operating data samples; use cluster analysis algorithms to build a state distribution model that can reflect the distribution characteristics of different operating states of the main engine and tower based on the operating data samples; connect the outputs of the state distribution model and the risk assessment model to a fusion processing layer to form a comprehensive monitoring model; F3. Solve the key performance indicator threshold boundary of the comprehensive monitoring model, and select the early warning solution set that matches the safe operation requirements in the key performance indicator threshold boundary according to the dynamic priority decision method; F4. Implement the corresponding maintenance strategy according to the early warning plan set. If the subsequent monitoring shows that the operating status of the main engine and the tower has returned to normal, stop the operation. Otherwise, feed back the newly collected operating data samples as input information to the comprehensive monitoring model and repeat F3 to F4.

2. A real-time monitoring method for the operating status of a wind power generator and a tower according to claim 1, characterized in that: In step F1, sensors are installed at the bearings of the wind turbine main engine and tower, the blade connection parts, and the bottom of the tower according to a preset scientific layout plan to collect operation data samples.

3. According to the real-time monitoring method for the operating status of a wind power host and tower as described in claim 1, the input of the comprehensive monitoring model is various real-time operating data of the wind power host and tower, and the output is quantitative indicators characterizing the quality of the operating status and the corresponding fault risk level.

4. A real-time monitoring method for the operating status of a wind power generator and a tower according to claim 1, characterized in that: The key performance indicator threshold boundary is composed of a set of quantitative indicator values ​​that can accurately distinguish different operating states, and is obtained by iteratively solving the comprehensive monitoring model through an optimization algorithm based on gradient descent. The specific formula is: ； in is the cost function, is the predicted value, is the true value, is the regularization parameter, are model parameters, is the sample size, is the number of parameters, by continuously adjusting Make Minimize to determine the threshold boundary.

5. A real-time monitoring method for the operating status of a wind power generator and a tower according to claim 4, characterized in that: The optimization algorithm based on gradient descent is a stochastic gradient descent algorithm. During the iteration process, the formula for updating parameters is: ； in is the learning rate.

6. A real-time monitoring method for the operating status of a wind power generator and a tower according to claim 1, characterized in that: The construction of the comprehensive monitoring model includes: Using a multi-layer perceptron to fit the running data samples collected in step F1, and dividing the running data samples into a training sample set and a validation sample set according to a specific ratio; Based on the training sample set, the cross entropy loss function and L2 regularization term are used as the two objectives in the dual-objective optimization problem; and the adaptive moment estimation optimization algorithm is used to obtain the convergence boundary of the corresponding risk assessment model training error, and then the dynamic priority decision method is used to select a predetermined number of candidate solution sets in the convergence boundary of the risk assessment model training error, and the network parameters of the risk assessment model are determined according to the final performance of the candidate sets on the verification data; according to the performance of the models corresponding to different solution sets on the verification sample set, a set of network parameters with the best performance is finally selected to obtain the risk assessment model; A variational autoencoder that simulates the distribution characteristics of the original operating data samples is built, and network parameters are trained based on the training sample set. When the preset training termination condition is reached, the decoder in the variational autoencoder is used as the state distribution model. The state distribution model and the output of the trained risk assessment model are then connected to a fusion processing layer to form a comprehensive monitoring model.

7. A real-time monitoring method for the operating status of a wind power generator and a tower according to claim 6, characterized in that: The fusion processing layer introduces an attention mechanism to achieve fusion, and calculates respective attention scores for the operating state distribution feature vector output by the state distribution model and the fault risk quantification value output by the risk assessment model.

8. A real-time monitoring method for the operation status of a wind power generator and a tower according to claim 7, characterized in that: The fusion processing layer establishes a redundant verification module during operation, cross-validates the fused data, divides the fused comprehensive vector into a verification subset and a comparison subset according to a certain ratio, and uses a pre-trained verification model to perform feature extraction and comparison analysis on the two subsets respectively. If the difference between the two exceeds a preset threshold, it is determined that there may be an abnormality in the fusion process, triggering an error correction process, and retraining the fusion processing layer parameters using historical normal fusion data.

9. A real-time monitoring method for the operation status of a wind power generator and a tower according to claim 1, characterized in that: In step F1, outliers are eliminated from the collected operation data samples, and a sliding window algorithm is used to smooth and reduce noise.

Images