A wind turbine generator health management method and system based on deep learning

Abstract

The present application relates to the technical field of wind turbine health management, and particularly relates to a wind turbine health management method and system based on deep learning. The method comprises the following steps: deploying an oil multi-parameter sensor in a hydraulic component of a wind turbine variable pitch system, collecting oil state monitoring data, and calculating the proportion of abnormal water content growth value per unit time; estimating the hydraulic oil lubrication performance degradation ratio according to the proportion, and then evaluating the damage of hydraulic power output, determining the pitch control offset, and identifying the risk factors of hydraulic power abnormality; finally, using the multilayer perceptron in the deep learning algorithm to design a health risk identification architecture, identifying the health risk of the hydraulic system of the wind turbine, and sending the result to the control center to realize the health management of the wind turbine. The present application improves the wind turbine health management technology and makes it more perfect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind turbine health management, and particularly relates to a wind turbine health management method and system based on deep learning. BACKGROUND

[0002] As a clean and efficient energy form, wind power has been paid more and more attention by more and more countries. As the core equipment for converting wind energy into electric energy, the stability and reliability of wind turbine directly affect the efficiency of energy production and environmental sustainable development. In order to ensure the efficient operation of wind turbine, health management has become an important research topic in the wind power industry. In the health management of wind turbine, the introduction of deep learning has brought revolutionary changes. Through deep learning model to deeply analyze the operation data of wind turbine, the real-time monitoring of the operation state of the unit can be realized, and the potential fault hidden danger can be found in time, so as to reduce the operation and maintenance cost and improve the power generation efficiency. The health management method based on deep learning can automatically process and analyze the operation data of the unit, and extract the key features affecting the health state of the unit. These features include but are not limited to mechanical failure, vibration anomaly, temperature anomaly, oil quality change, etc. By constructing a deep learning model with strong adaptability and good robustness, large-scale data generated by wind turbine can be effectively processed, and intelligent identification and prediction of the health state of the unit can be realized. However, the traditional wind turbine health management method based on deep learning has the problems of inaccurate identification of hydraulic power output loss, resulting in large control error of the variable pitch system, and inaccurate identification of hydraulic power abnormality preposition risk. SUMMARY

[0003] Therefore, it is necessary to provide a wind turbine health management method and system based on deep learning to solve at least one of the above technical problems.

[0004] To achieve the above purpose, a wind turbine health management method based on deep learning, the method comprising the following steps:

[0005] Step S1: deploying an oil multi-parameter sensor in the hydraulic component of the variable pitch system of the wind turbine, and then collecting oil state monitoring data; calculating the water content abnormal growth value proportion of the oil state monitoring data in unit time;

[0006] Step S2: estimating the hydraulic oil lubrication performance degradation ratio according to the water content abnormal growth value proportion to obtain the lubrication performance degradation ratio; evaluating the hydraulic power output loss according to the lubrication performance degradation ratio to generate hydraulic power output loss data;

[0007] Step S3: determining the pitch control offset based on the hydraulic power output reduction data, and then identifying the hydraulic power abnormality pre-warning risk factor to obtain the hydraulic power abnormality pre-warning risk factor;

[0008] Step S4: using a multilayer perceptron in a deep learning algorithm to design a health risk identification architecture for the hydraulic power abnormality pre-warning risk factor, to obtain the health risk identification architecture, and send it to the wind turbine generator control center to perform health management of the wind turbine generator.

[0009] The application also provides a wind turbine generator health management system based on deep learning, which is used to perform the wind turbine generator health management method based on deep learning as described above, and comprises:

[0010] The monitoring data acquisition module is used to deploy an oil multi-parameter sensor in the hydraulic component of the pitch system of the wind turbine generator, and then acquire oil state monitoring data; and calculate the water content abnormal growth value proportion in unit time.

[0011] The hydraulic power output reduction evaluation module is used to estimate the hydraulic oil lubrication performance degradation ratio according to the water content abnormal growth value proportion, to obtain the lubrication performance degradation ratio; and evaluate the hydraulic power output reduction according to the lubrication performance degradation ratio, to generate the hydraulic power output reduction data.

[0012] The pre-warning risk factor identification module is used to determine the pitch control offset based on the hydraulic power output reduction data, and then identify the hydraulic power abnormality pre-warning risk factor to obtain the hydraulic power abnormality pre-warning risk factor.

[0013] The health risk identification architecture design module is used to use a multilayer perceptron in a deep learning algorithm to design a health risk identification architecture for the hydraulic power abnormality pre-warning risk factor, to obtain the health risk identification architecture, and send it to the wind turbine generator control center to perform health management of the wind turbine generator.

[0014] The beneficial effects of the present application are that by deploying oil multi-parameter sensors in the hydraulic components of the wind turbine group variable pitch system and collecting real-time oil condition monitoring data, detailed real-time data support can be provided for the operating state of the hydraulic system. This step effectively captures the water content change in the hydraulic oil by calculating the abnormal growth value proportion of water content. Since water has a significant impact on the performance of hydraulic oil, timely monitoring and analyzing the water content change can help identify potential problems that may cause hydraulic system failure, ensuring the normal operation of the system and providing a data basis for subsequent lubrication performance degradation prediction. According to the abnormal growth value proportion of water content, the lubrication performance degradation ratio can be accurately estimated. The performance of lubricating oil will gradually deteriorate over time and with changes in working environment, which can lead to problems such as decreased hydraulic system efficiency and increased component wear. By calculating the lubrication performance degradation ratio, the actual lubricating capacity of the hydraulic oil can be scientifically evaluated, and combined with the hydraulic power output loss data, the performance of the hydraulic system can be determined in a timely manner to provide reliable decision-making basis for maintenance personnel and avoid premature failure of the system due to lubrication problems. Based on the hydraulic power output loss data, the variable pitch control offset is determined and the hydraulic power abnormality pre-position risk factor is identified, which helps to provide early warning for abnormal behavior of the variable pitch system. The reduction of hydraulic power output usually means that the system's working efficiency is declining, and even affects the power generation capacity of the wind turbine group. By identifying potential hydraulic power abnormality pre-position risk factors, preventive measures can be taken before the system fails, reducing equipment downtime and economic losses caused by sudden failures, and providing a scientific basis for subsequent fault diagnosis and repair. It should be noted that the hydraulic system is mainly used to provide power support for the wind turbine group. Specifically, the hydraulic system controls the flow and pressure of hydraulic oil to drive the action of the variable pitch system, and the quality, flow and pressure of the hydraulic oil directly affect the operation accuracy and response speed of the variable pitch system. After the oil in the hydraulic system is compressed, filtered and adjusted, it controls the angle of the wind turbine blades through the hydraulic cylinder to adapt to different wind speed conditions and optimize power generation efficiency. Therefore, the present application improves the traditional deep learning-based wind turbine group health management method, solves the problem of inaccurate identification of hydraulic power output loss in the traditional deep learning-based wind turbine group health management method, which causes large control errors of the variable pitch system, and improves the accuracy of the identification of the hydraulic power abnormality pre-position risk. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 A flowchart of a deep learning-based wind turbine group health management method;

[0016] Figure 2 For Figure 1 Detailed implementation step flow diagram of step S2 in the embodiment;

[0017] Figure 3 For Figure 2 Detailed implementation step flow diagram of step S23 in the embodiment;

[0018] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0019] Please refer to Figures 1 to 3 A wind turbine health management method based on deep learning, the method comprising the following steps:

[0020] Step S1: deploying an oil multi-parameter sensor in the hydraulic components in the variable pitch system of the wind turbine, and then collecting oil state monitoring data; calculating the oil state monitoring data, and the abnormal growth value proportion of water content in unit time;

[0021] Step S2: estimating the hydraulic oil lubrication performance degradation ratio according to the abnormal growth value proportion of water content, to obtain the lubrication performance degradation ratio; evaluating the hydraulic power output reduction according to the lubrication performance degradation ratio, to generate hydraulic power output reduction data;

[0022] Step S3: determining the variable pitch control offset based on the hydraulic power output reduction data, and then identifying the hydraulic power abnormality pre-risk factor, to obtain the hydraulic power abnormality pre-risk factor;

[0023] Step S4: using a multilayer perception machine in a deep learning algorithm to design a health risk identification architecture for the hydraulic power abnormality pre-risk factor, to obtain the health risk identification architecture, and sending it to the wind turbine control center to perform health management of the wind turbine.

[0024] In the embodiment of the present application, reference Figure 1 The is a step flow diagram of a wind turbine health management method based on deep learning, in this example, the wind turbine health management method based on deep learning comprises the following steps:

[0025] Step S1: deploying an oil multi-parameter sensor in the hydraulic components in the variable pitch system of the wind turbine, and then collecting oil state monitoring data; calculating the oil state monitoring data, and the abnormal growth value proportion of water content in unit time;

[0026] In the embodiment of the present application, three groups of oil multi-parameter sensors are arranged equidistantly along the main flow direction in the pipe section between the outlet of the main oil tank of the hydraulic circuit of the wind turbine variable pitch system and the inlet of the actuator, each group of sensors including a moisture content detection unit, a viscosity detection unit, a particle contamination detection unit and a temperature compensation unit, the sampling frequency of the sensors is set to one per second, the raw data is transmitted to the edge computing node through the RS485 bus, the edge computing node sorts the collected oil state monitoring data according to the time stamp and encapsulates it into a structured time series data table, each row in the table corresponds to a sampling time, and the fields include sampling time, moisture volume percentage, kinematic viscosity value, pollution level code and environmental temperature value; for the moisture volume percentage field, a sliding window algorithm is used, the window length is fixed at 3600 sampling points, i.e. one hour, and the step is 1, a first-order difference operation is performed on the moisture data in each window to obtain the moisture change per minute, and then the threshold value is selected for the difference sequence, the threshold value is set to 0.02% per minute, the data points exceeding the threshold value are marked as abnormal growth points, the proportion of the number of abnormal growth points in the total sampling points per hour is calculated, which is the water content abnormal growth value proportion per unit time, and finally the output is a time series with an hour as a unit, and each element represents the relative frequency of abnormal moisture growth in that hour.

[0027] Step S2: estimating the hydraulic oil lubrication performance degradation ratio according to the water content abnormal growth value proportion to obtain the lubrication performance degradation ratio; evaluating the hydraulic power output reduction according to the lubrication performance degradation ratio to generate hydraulic power output reduction data;

[0028] In the embodiment of the present application, the base viscosity value 46 mm2 / s and the anti-emulsification index not greater than 15 minutes separation completion of the hydraulic oil of this model at the time of factory delivery are extracted from the hydraulic oil technical manual as the basic performance parameters, the water content abnormal growth value proportion sequence is coupled with the base viscosity value, the molecular dynamics simulation method is used to construct the oil film molecular chain polarity balance loss process, the specific operation is to establish a three-dimensional lattice model containing base oil molecules, water molecules and additive molecules, the initial state is set as a stable oil film structure in a water-free environment, water molecules are gradually injected into the model, after injecting 0.1% volume fraction water molecules, 10 steps of Monte Carlo simulation are run, the number of hydrogen bond breakage between oil film molecules and the dipole moment deviation angle are recorded, and the polarity balance loss data matrix under different water contents is accumulated; based on the matrix, the oil film rupture event occurrence probability is fitted by using Poisson distribution, and the ratio of the number of oil film rupture per unit time to the total molecular logarithm is defined as the oil film rupture probability index; then the oil film rupture probability index and the polarity balance loss data are jointly input into the Navier-Stokes equation, the divergence characteristics of viscosity in the shear stress field with time are solved, and the dynamic viscosity loss divergence data are obtained, which is expressed as a viscosity decay rate curve per hour; finally, the dynamic viscosity loss divergence data and the oil film rupture probability index are weighted and fused, and the weight coefficients are 0.6 and 0.4 respectively, and the weighted result is the lubrication performance degradation ratio; further, the lubrication performance degradation ratio is input into the shear thinning constitutive equation, and the shear resistance performance reduction degree is derived, the nonlinear time series fitting of the reduction degree is carried out by using the cubic spline interpolation method, and the performance reduction time series increment fitting data are obtained; according to the data, combined with the Archard wear formula in the boundary lubrication theory, the material removal volume per unit area of the friction pair contact surface is calculated, and the friction disorder increment data are accumulated; finally, the lubrication performance degradation ratio and the friction disorder increment data are input into the multivariate grey correlation analysis model, the resolution coefficient is set to 0.5, the correlation coefficient matrix between the internal leakage dynamic increment value and the oil film bearing capacity decline gradient is calculated, the multi-parameter correlation structure is constructed according to the eigenvector corresponding to the maximum eigenvalue of the matrix, and then the hydraulic power output reduction data are output, which includes three dimensions of pressure loss percentage per hour, flow decay rate and power conversion efficiency reduction value.

[0029] Step S3: determining the pitch control offset based on the hydraulic power output reduction data, and then identifying the hydraulic power abnormality pre-warning risk factor to obtain the hydraulic power abnormality pre-warning risk factor;

[0030] In the embodiment of the present application, the hydraulic power output reduction data is introduced into the feature engineering processing module. Firstly, the pressure loss percentage, flow decay rate and power conversion efficiency reduction value fields are respectively subjected to Z-score standardization, and then the first two cumulative variance contribution rates exceeding 95% are extracted as the power output reduction feature structure by using principal component analysis. The feature structure is input into a one-dimensional convolutional neural network layer, the convolution kernel size is set to 5, the step is 1, the activation function is ReLU, and the power reduction feature convolution data is output after convolution operation. The recursive mean filtering algorithm is applied to the data, the window length is set to 12, and the output pressure attenuation recursive mean difference sequence is calculated. The sequence is substituted into the hydraulic cylinder kinematics equation set in reverse, the equation set includes the piston displacement differential equation and the velocity differential equation, and the cylinder extension and retraction amount sequence and its first derivative, i.e. the extension and retraction rate sequence, are obtained by simultaneous solution. The variances of the two sequences with respect to the nominal values are calculated to obtain the retraction amount variance and the retraction rate variance. The two variance values are input into the pitch angle mapping function, which is a pre-calibrated cubic polynomial regression model, the input is the weighted sum of the retraction amount variance and the rate variance, and the output is the upper and lower limits of the pitch angle deviation angle interval. The LSTM recurrent neural network is used for sequence memory modeling on the interval sequence, the number of hidden layer units is set to 64, the training round is 100, the loss function is mean square error, and the network parameters are frozen after training. The real-time deviation angle interval is input, and the network output is the deviation angle memory learning data. Finally, the data is difference-operated with the target pitch angle in the current control command, and the result is the variable pitch control deviation; the variable pitch control deviation and the power reduction feature convolution data are spliced into a joint feature vector, which is input into the isolated forest anomaly detection algorithm. The number of trees is set to 200, the subsampling ratio is 0.8, and the anomaly score threshold is set to 0.7. The feature dimensions corresponding to the samples with scores higher than the threshold are identified as the hydraulic power abnormality pre-risk factors, and the output is a binary mask vector indicating which feature dimensions have pre-risk.

[0031] Step S4: using a multilayer perceptron in a deep learning algorithm to design a health risk identification architecture for the hydraulic power abnormality pre-risk factors, obtaining a health risk identification architecture, and sending it to a wind turbine generator control center to perform health management of the wind turbine generator.

[0032] In the embodiment of the present application, the hydraulic power abnormality pre-position risk factor vector is input into a feature embedding layer, which is composed of a fully connected neural network, the input dimension is equal to the pre-position risk factor dimension, the output dimension is fixed at 128, and the activation function is tanh. After embedding transformation, the pre-position risk learning factor is obtained. The factor is input into a linear discriminant analysis module, and the between-class scatter matrix and the within-class scatter matrix are calculated according to the historical normal samples and abnormal samples. The first 10 eigenvectors that maximize the Fisher criterion function are selected as the projection direction, and the risk learning linear discriminant factor is obtained after projection. The factor is sent as an input feature into a multi-layer perceptron architecture, which includes an input layer, three hidden layers and an output layer. The number of input layer neurons is 10, the number of first hidden layer neurons is 64, the activation function is ReLU, the number of second hidden layer neurons is 32, the activation function is ReLU, the number of third hidden layer neurons is 16, the activation function is sigmoid, the number of output layer neurons is 1, and the activation function is sigmoid, which is used to output the health risk probability value. The network training adopts the cross-entropy loss function, the optimizer selects Adam, the initial value of the learning rate is 0.001, the batch size is 32, the training data is derived from the historical labeled sample library, and the sample label is 0 for healthy state and 1 for impending failure. After training, the complete network structure and weight parameters are exported as an ONNX format file, which is uploaded to the health management server of the wind turbine control center through the industrial Ethernet protocol. After the server loads the architecture, it receives the pre-position risk factor data from the edge computing node in real time, performs forward inference calculation, outputs the health risk probability, and triggers the early warning instruction or automatic maintenance scheduling instruction according to the preset threshold 0.5, thereby realizing the closed-loop health management of the wind turbine.

[0033] Step S1 includes the following steps:

[0034] Step S11: deploying an oil multi-parameter sensor in the hydraulic components in the variable pitch system of the wind turbine generator set, and then collecting oil state monitoring data;

[0035] Step S12: timestamp embedding is performed on the oil state monitoring data to generate oil state monitoring time series data;

[0036] Step S13: extracting the abnormal growth trend of water content in the oil state monitoring time series data;

[0037] Step S14: calculating the water content abnormal growth trend, and the water content abnormal growth value proportion in unit time.

[0038] In the embodiment of the present application, three groups of industrial-grade online oil multi-parameter sensors are evenly installed in the straight pipe section downstream of the filter of the main oil circuit of the hydraulic system and before the inlet of the servo valve in the circumferential direction. Each group of sensors integrates a capacitive moisture detection probe, a rotary viscometer module, a laser particle counter, and a platinum resistance temperature sensor. The sampling period is fixed at 1000 milliseconds. The original data is uploaded to the programmable logic controller through the Modbus RTU protocol. The collected parameters include the moisture volume percentage, the 40°C kinematic viscosity value, the NAS 1638 contamination level, and the real-time oil temperature. All data are buffered in the controller internal register in hexadecimal encoding form.

[0039] The programmable logic controller has a built-in real-time clock module. When data collection is triggered, the current date and time are read synchronously, and the oil condition data are packaged to form a structured record. Each record contains nine fields arranged in a fixed order, including the timestamp, moisture value, viscosity value, contamination level, temperature value, and three sensor number identifiers. The data packet is pushed to the edge server through the TCP / IP protocol and stored as a CSV format file. The file naming rule is "unit number_date_hour.csv". The oil condition monitoring time series data set is constructed in strict physical time order.

[0040] The sliding window first-order difference algorithm is applied to the moisture volume percentage field in the CSV file. The window width is set to 60 consecutive sampling points, i.e., one minute. The moisture change between adjacent points is calculated. After generating the difference sequence, the threshold judgment method is used to screen the abnormal growth segment. The threshold is set to a moisture increase of more than 0.02% per minute. The difference value that meets the condition is marked as 1, and the rest is marked as 0. Morphological dilation operation is performed on the binary sequence. The structure element length is set to 5. The adjacent abnormal points are merged to form a continuous abnormal interval. The output is a list of water content abnormal growth trends consisting of the starting timestamp and the duration.

[0041] Take one hour as the statistical unit. Traverse all the water content abnormal growth trend records in this hour. Add the duration of each trend to get the total abnormal minutes. Divide by 60 to get the water content abnormal growth value proportion in the unit time of this hour. The value range is limited between 0 and 1. The result is written into a new time series data table with the field structure of "statistical starting time, abnormal proportion value". This table is used as the direct input source for subsequent lubrication performance degradation analysis. The data update frequency is one batch processing calculation per hour at the whole point.

[0042] Step S2 includes the following steps:

[0043] Step S21: Obtain the basic performance parameters of the hydraulic oil;

[0044] Step S22: estimating the hydraulic oil lubrication performance degradation ratio according to the abnormal growth value proportion of the water content, to obtain the lubrication performance degradation ratio;

[0045] Step S23: quantifying the wear aggravation degree of the hydraulic component based on the lubrication performance degradation ratio;

[0046] Step S24: performing hydraulic power output reduction evaluation according to the lubrication performance degradation ratio and the wear aggravation degree, to generate hydraulic power output reduction data.

[0047] As an example of the present application, referring to Figure 2 In the present example, the step S2 includes:

[0048] Step S21: obtaining the basic performance parameters of the hydraulic oil;

[0049] In the embodiment of the present application, the factory technical document of the oil used by the current hydraulic system is called from the wind turbine generator set equipment history database, the basic performance indicators clearly recorded in the type test report issued by the recognized detection agency are extracted, including the initial kinematic viscosity value measured under standard temperature conditions, the emulsion resistance separation time measured according to the industry standard method, the copper strip corrosion test grade, the oxidation stability test results, and the pour point and flash point temperature values, all data fields are stored in a structured data table in a fixed format, each record in the table contains the oil batch number, the detection date, the detection unit qualification number, the measured values of each performance and the corresponding test method national standard number, the data table is completely loaded into the memory cache area when the system starts, and is bound with the current running unit through the unique oil identification code, to ensure that the basic parameters referred to in the subsequent analysis have legal traceability and batch consistency, and to prevent the use of estimated values or default values instead of measured data.

[0050] Step S22: estimating the hydraulic oil lubrication performance degradation ratio according to the abnormal growth value proportion of the water content, to obtain the lubrication performance degradation ratio;

[0051] In the embodiment of the present application, the water content anomaly growth value proportion sequence output in step S14 is input into the physical simulation deduction platform together with the initial viscosity and emulsion resistance data in the hydraulic oil basic performance parameter table. The platform is internally provided with an oil film structure evolution engine constructed based on the principle of molecular dynamics. The engine dynamically simulates the process of water molecules invading the intermolecular chain gap of base oil molecules according to the input water content growth rate, records the oil film interfacial tension change trajectory, the degree of disorder of polar group arrangement and the hydrogen bond network fracture frequency, and simulates the process in steps of 0.1% water content. Each step performs 10 times of energy minimization iteration operation, and outputs the oil film stability degradation characteristic matrix under different water content states. Then the matrix is sent to a rupture event statistic device. The statistic device determines the oil film rupture probability level per unit time under the current water content disturbance intensity according to the oil film failure event frequency table established based on historical bench tests through the table lookup interpolation method. The rupture probability level and the oil film structure disorder characteristics are jointly input into a viscosity response deducer. The deducer calls a pre-calibrated viscosity attenuation trend surface under shear environment. The surface is generated by fitting the measured data of a rotary viscometer under controllable water content of the same type of oil. The output is a nonlinear downward trend of dynamic viscosity over time. Finally, the viscosity downward trend and the rupture probability level are sent to a weighted fusion device for numerical synthesis according to the preset weight distribution rule, and a lubrication performance degradation ratio sequence is output. The sequence is updated every hour, and the numerical range is limited between zero and one. Data exceeding the boundary is automatically truncated and triggers a data anomaly alarm log.

[0052] Step S23: quantifying the wear aggravation degree of the hydraulic component based on the lubrication performance degradation ratio;

[0053] In the embodiment of the present application, the lubrication performance degradation ratio sequence is sent to a shear resistance mapping module. The module internally stores a hydraulic oil shear resistance degradation reference table collected by a laboratory bench under standard working conditions. Each row in the table corresponds to a degradation ratio interval, and the viscosity retention rate and energy dissipation growth rate of the oil under the action of a fixed shear rate in the interval are recorded in the column. The shear resistance attenuation level corresponding to the current degradation ratio is located in the table through linear interpolation. Then the level and the current working pressure, action frequency and stroke length parameters of the hydraulic component are jointly input into a wear deduction engine. The engine is based on a material removal amount statistical distribution library accumulated through a large number of friction pair bench tests. The library is stored by component material, surface roughness, load level, etc. Each category contains a monotonic mapping relationship between shear resistance attenuation level and material wear volume per unit time. The wear increment value under the current comprehensive working condition is determined through a double-variable table lookup method. The increment value is accumulated by hour to form a friction disorder increment data sequence. Each record in the sequence contains a time stamp, a unique component code, a cumulative wear volume, a current increment value and a working condition combination identification code. The data is written into a wear state time series database for real-time calling by a subsequent power output evaluation module.

[0054] Step S24: Perform hydraulic power output reduction evaluation according to the lubrication performance degradation ratio and the wear aggravation degree, and generate hydraulic power output reduction data.

[0055] In the embodiment of the present application, the friction disorder increment data is sent to a mechanical clearance evolution calculation unit, which is embedded with an experience mapping relationship library of clearance expansion based on measured data training. The relationship library is stored by component type, and each type contains a monotonically increasing corresponding table between wear volume and clearance expansion amount. The clearance increase value corresponding to the current wear amount is determined by the table lookup interpolation method. The clearance increase value is input into an internal leakage deducer. The deducer calls a pre-established leakage flow response surface according to the hydraulic system topology and sealing form. The surface is fitted by measured data of leakage amount corresponding to different clearance values under fixed pressure difference conditions, and outputs the additional internal leakage volume flow caused by clearance expansion per unit time. Simultaneously, the lubrication performance degradation ratio is sent to a bearing capacity evaluator. The evaluator calls an oil film strength degradation feature library, which records the deformation depth and recovery delay time of the oil film under standard load at different degradation stages. According to this, the bearing capacity decline gradient is output. The internal leakage volume flow and the bearing capacity decline grade are jointly input into a multivariate grey correlation processor. The processor calculates the similarity weight of the two and the typical failure mode according to the historical fault sample library, and constructs a comprehensive correlation structure. Finally, the structure drives a three-channel linear fusion device to output three indicators: system pressure maintenance capability decline percentage, main circuit effective flow attenuation percentage, and energy conversion efficiency reduction percentage. The three data are generated once an hour, packaged as structured records with time stamps and unit numbers, and written into a hydraulic power output reduction data table for subsequent risk identification module to call.

[0056] Step S22 includes the following steps:

[0057] Step S221: Extract the base viscosity and anti-emulsification property in the basic performance parameters of the hydraulic oil;

[0058] Step S222: Perform oil film intermolecular chain polarity balance loss process simulation on the base viscosity and anti-emulsification property according to the abnormal growth value proportion of water content, to obtain oil film intermolecular chain polarity balance loss data;

[0059] Step S223: Quantify the oil film rupture probability index based on the polarity balance loss data;

[0060] Step S224: Perform power viscosity loss divergence evaluation according to the oil film rupture probability index and the polarity balance loss data, to obtain power viscosity loss divergence data;

[0061] Step S225: Perform hydraulic oil lubrication performance degradation ratio estimation according to the power viscosity loss divergence data and the oil film rupture probability index, to obtain the lubrication performance degradation ratio.

[0062] In the embodiment of the present application, the two core indicators of basic viscosity and emulsion resistance are extracted from the basic performance parameters of hydraulic oil, which respectively represent the internal flow resistance and water phase separation capacity of the hydraulic oil. In order to ensure the stability of the data, the viscosity value is continuously monitored by using an online rheometer at constant temperature, and the value is taken at a constant shear rate in a constant temperature environment of 40 degrees Celsius, and the viscosity fluctuation range within 3 hours is recorded to confirm the balanced characteristics of the viscosity. The emulsion resistance is determined by combining constant speed stirring and standing method. First, the oil sample and deionized water are mixed in proportion, stirred for 10 minutes in a constant temperature water bath, and then left to stand. The water-oil separation rate is calculated by the ratio of the corresponding layer thickness to the standing time. After collecting the detection results of the two types, time sequence data structure is formed, and a unified parameter table is established with time node as index. Each row in the parameter table records the viscosity value and emulsion resistance rate corresponding to the same time, and the time resolution of the data is kept at minute level. Then, by comparing the sampling curves of different working periods through data structured arrangement, the basic stable section of the performance fluctuation of the hydraulic oil is confirmed, so as to determine the reference state value of the basic viscosity and emulsion resistance of the hydraulic oil, and provide input data for subsequent simulation operation.

[0063] On the basis of the determined basic viscosity and emulsion resistance data, the abnormal growth rate of water content obtained in the previous step is used as an influencing factor to quantitatively simulate the process of loss of polarity balance between the molecular chains of the oil film in the hydraulic oil. First, the water content growth curve and the viscosity change curve are superimposed on the time axis, and the viscosity change amplitude in the water content sudden increase period is identified by time section matching. The system calculates the energy distribution deviation of the oil molecular chain structure after being disturbed by the polar molecules (water content) in these periods, and determines the degree of loss of polarity stability by comparing the decline amplitude of the emulsion resistance rate in each period. Then, according to the coupling change relationship between the emulsion resistance rate and the viscosity decline, the delamination imbalance duration and recovery delay of the oil film after being invaded by micro water are extracted as the key feature indexes of the loss of polarity balance. For continuous period data, the statistical analysis method of moving time window is used to calculate the cumulative value of the polarity deviation of the oil film in each time window, forming the sequence of intermolecular polarity balance loss. Through the information superposition of the sequence, the molecular chain rupture trend caused by the increase of water content and the molecular rearrangement process in the oil film are depicted, and finally the intermolecular polarity balance loss data of the oil film are obtained, which include time index, cumulative imbalance value and recovery delay index, and are used to evaluate the depth of water resistance degradation of the hydraulic oil.

[0064] After obtaining the oil film molecular chain polarity balance loss data, the oil film rupture trend is quantitatively calculated to reflect the stability of the lubricating film in the running stage. First, according to the time series of the polarity balance loss, the stable state of the oil film is classified and calibrated, and the imbalance intensity change of adjacent monitoring intervals is divided into three categories: weak fluctuation, sustained deviation and structural rupture. Then, by analyzing the time overlap length between the local viscosity attenuation and the anti-emulsification recovery lag in the oil film, the formation probability of the rupture point is judged. When the viscosity drop continues to exceed the normal fluctuation interval and the corresponding anti-emulsification fails to recover in a short time, it is marked as a potential rupture time. According to the distribution density and duration of all potential rupture times, the oil film instability interval distribution is formed on the time axis. Subsequently, the cumulative statistical method is used to comprehensively calculate the number, duration and fluctuation amplitude of these instability intervals, and the oil film rupture probability index is derived. The whole process relies on the time continuity of the oil film polarity balance loss data, and determines the time law of oil film rupture by statistically analyzing the imbalance accumulation and viscosity attenuation correlation strength. The final generated oil film rupture probability index is output in the form of three indicators: time period label, rupture trend grade and oil film recovery ratio, which is used for the subsequent comprehensive judgment stage of lubrication degradation ratio.

[0065] After the oil film rupture probability index and the polarity balance loss data are extracted, the dynamic viscosity loss divergence evaluation is performed to quantitatively depict the flow resistance attenuation trend of the hydraulic oil. The evaluation process first divides the time axis into multiple fixed time periods based on the oil film rupture probability distribution in the continuous monitoring period. The corresponding polarity balance loss value and temperature change record are superimposed in each time period to analyze the change of the flow structure of the hydraulic oil after the molecular chain is disturbed. By calculating the deviation amplitude of the viscosity change in adjacent time periods, the stable fluctuation interval and the diffusion attenuation interval are distinguished. For the diffusion attenuation interval, the difference accumulation method is used to calculate the divergence rate of the viscosity behavior of the oil to describe the expansion process of the dynamic viscosity of the hydraulic oil to the low value direction. Then, the viscosity change of the circulating pipeline in the hydraulic system under different pressure conditions is statistically segmented, and the viscosity drop amplitude and the growth rate of the rupture index are compared synchronously to identify the coupling correlation strength between viscosity attenuation and oil film rupture. According to the numerical value of the correlation strength, the dynamic viscosity loss divergence data table is generated. The data table includes time index, divergence section number, viscosity drop amplitude interval and coupling strength value, which reflects the continuous weakening of the flow performance of the hydraulic oil in different working stages and provides complete input data structure for the next step of lubrication performance degradation ratio estimation.

[0066] After the divergence data of the dynamic viscosity loss is formed, the comprehensive estimation of the lubrication performance degradation ratio of the hydraulic oil is performed by taking the data and the previously obtained oil film rupture probability index as the core input. This step takes time series superposition analysis as the main line, and each period of dynamic viscosity loss value and rupture probability value are corresponded one by one, and a synchronous mapping relationship table is established in the order of time increment. Then, according to the division of the running period of the hydraulic system, the energy transmission efficiency of each period is calculated, and the weakening trend of the lubrication transmission capacity is quantitatively represented by the reduction of the pressure output and the flow output in the viscosity drop period. By averaging the viscous resistance diffusion rate in each cycle, the time proportion of the weakening of the lubrication layer integrity is determined, and the lubrication performance degradation speed of the hydraulic oil under long period condition is determined. Then, the local decline rate of each cycle is accumulated as a cumulative attenuation sequence, and the change curve of the lubrication performance with time is established, and the proportion of the co-variation interval of the oil film rupture and the viscosity loss is calculated as the final lubrication performance degradation index. The generated hydraulic oil lubrication performance degradation ratio data set includes monitoring cycle number, performance decline percentage interval, energy transmission weakening grade and time correlation identifier, reflecting the overall lubrication attenuation amplitude and system health deviation degree of the hydraulic oil under long time operation, and providing quantitative basis for the hydraulic power output reduction calculation.

[0067] Step S23 comprises the following steps:

[0068] Step S231: simulate and deduce the shear resistance performance degradation degree based on the lubrication performance degradation ratio;

[0069] Step S232: nonlinear time sequence increment fitting is performed on the shear resistance performance degradation degree to obtain performance degradation time sequence increment fitting data;

[0070] Step S233: boundary friction disorder increment derivation of the hydraulic component is performed based on the performance degradation time sequence increment fitting data to obtain friction disorder increment data;

[0071] Step S234: the wear aggravation degree of the hydraulic component is quantified according to the friction disorder increment data.

[0072] As an example of the present application, referring to Figure 3 in this example, the step S23 comprises:

[0073] Step S231: simulate and deduce the shear resistance performance degradation degree based on the lubrication performance degradation ratio;

[0074] In the embodiment of the application, the lubrication performance degradation ratio is taken as an input core index to deduce and analyze the shear resistance performance degradation degree of the hydraulic system of the wind turbine generator set. First, the pressure fluctuation curve, temperature rise gradient curve and flow stability record of each operation period are collected from the hydraulic execution system, and these data are synchronized and aligned in time sequence. Taking the lubrication performance degradation ratio as a benchmark, the fluid pressure drop and inner wall shear rate change of the corresponding period are compared to judge the weakening range of the rheological property of the hydraulic oil under the action of shear, the overlap degree between the viscosity attenuation interval and the internal shear rate change interval is used to determine the shear resistance decline trend caused by the instability of the molecular structure of the hydraulic oil, then the cycle number and amplitude change of the pressure pulsation in the operation cycle are calculated to calculate the shear action energy density weakening proportion in the hydraulic pipeline, and the shear resistance performance degradation level interval is divided accordingly. The performance degradation rate of each stage is calculated by superimposed analysis of multiple time windows, and the shear resistance performance degradation degree sequence is output as the input data basis for subsequent time series incremental fitting.

[0075] Step S232: nonlinear time series incremental fitting is performed on the shear resistance performance degradation degree to obtain performance degradation time series incremental fitting data;

[0076] In the embodiment of the application, the shear resistance performance degradation degree sequence obtained in step S231 is taken as input to perform nonlinear time series incremental fitting on the shear attenuation process of the hydraulic system. First, the degradation degree is divided into continuous equal-width sampling sections according to time index, the change amplitude of the shear resistance performance decline curve is extracted in each sampling section, and the time gradient is calculated. Then, the local incremental value of the performance decline in each time period is calculated by using the sliding window method, the dynamic decreasing trend of the shear resistance of the hydraulic oil is determined by comparing the numerical difference of adjacent time windows. The time series data is input into the nonlinear regression algorithm, and the increments of each section are smoothed in a time window progressive manner to extract the main fluctuation law showing periodic attenuation. On this basis, the slope change of the curve in different stages is twice fitted to deduce the incremental sequence of the shear performance change rate, and the performance degradation time series incremental fitting data is obtained by cumulative superposition. The data set contains data fields such as time node number, incremental amplitude, duration index and incremental fluctuation direction, which are used to describe the time series change characteristics of the shear resistance performance degradation of the hydraulic oil.

[0077] Step S233: boundary friction disorder increment derivation of the hydraulic component is performed based on the performance degradation time series incremental fitting data to obtain friction disorder increment data;

[0078] In the embodiment of the present application, the performance degradation time increment fitting data is taken as input, and the derivation and calculation of the boundary friction disorder increment of the hydraulic component are carried out to reveal the random distribution characteristics of the friction force change between components. First, according to the time alignment principle, the pressure feedback data of the hydraulic cylinder, valve group and oil pump are matched with the performance degradation time increment curve to extract the output fluctuation data of each component at the corresponding time. The amplitude of the flow and pressure curve change in each period is discretized and grouped, the unstable range of the pressure response in the continuous time period is calculated, and the fluctuation intensity of the friction coefficient caused by the shear performance decline is judged. Then, the adjacent difference sequence of the friction coefficient fluctuation is measured by the time difference method, and a discrete signal set of the boundary friction change is formed. The signal sequence is subjected to variance statistics and frequency distribution analysis, and the concentration interval of the local friction fluctuation between components is identified. The fluctuation peak value and its frequency are extracted from all period data, and the change trajectory of the friction disorder degree is judged in combination with the shear performance decline amplitude to finally generate the friction disorder increment data. The data set records the component name, time index, fluctuation intensity level and disorder distribution density, which is used to support the quantitative analysis of wear aggravation.

[0079] Step S234: quantifying the wear aggravation degree of the hydraulic component according to the friction disorder increment data.

[0080] In the embodiment of the present application, the wear aggravation degree of the hydraulic component is quantitatively calculated based on the friction disorder increment data. First, the friction disorder increment data is classified and counted according to the component category, and the working load, pressure fluctuation frequency and corresponding time cumulative running amount of each component are extracted. According to the frequency and duration of the disorder fluctuation in the increment data, the friction stress change trend of each component is divided into stages. Then, the friction intensity distribution in adjacent stages is compared to determine the time proportion of the stable interval and the mutation interval, and the growth stage of the wear development rate is identified. The friction energy consumption increment of each component is calculated by the cumulative summation method of time series, and the component wear aggravation rate is derived according to the slope change of the energy consumption accumulation curve. Taking the hydraulic pump, oil cylinder and control valve as the main calculation objects, the wear increase rate of each component is combined according to the system working weight to generate a comprehensive data set representing the wear aggravation degree of the entire hydraulic system. The data table includes component number, wear intensity level, increase rate value and corresponding working interval, which provides a basic quantitative index for the hydraulic power output performance degradation evaluation stage.

[0081] Step S24 includes the following steps:

[0082] Step S241: according to the wear aggravation degree, the component fitting gap expansion value is evolved, and then the dynamic increment value of the internal leakage is calculated by probability;

[0083] Step S242: determining the oil film bearing capacity decline gradient according to the lubrication performance degradation ratio;

[0084] Step S243: performing multi-parameter gray correlation processing according to the internal leakage dynamic increment value and the oil film bearing capacity gradient to generate a multi-parameter correlation structure;

[0085] Step S244: performing hydraulic power output reduction evaluation based on the multi-parameter correlation structure to generate hydraulic power output reduction data.

[0086] In the embodiment of the present application, after obtaining the hydraulic component wear aggravation degree data, first, the working load, running time and material properties of the core friction pair parts of each component are classified and analyzed. According to the wear aggravation degree result, the actual surface wear depth of the hydraulic pump, valve group and execution cylinder is mapped to the geometric fitting relationship, and the gap expansion trend is calculated through geometric expansion. This process takes the component running cycle as the horizontal axis time sequence, and superimposes the distance change between the contact surfaces caused by friction and wear in each sampling period to form a gap expansion numerical sequence. Then, by comparing the pressure fluctuation in the hydraulic oil flow path with the oil leakage rate data at the sliding seal, the gap expansion sequence and the fluid leakage rate change are analyzed synchronously, and the internal leakage change proportion corresponding to the gap increase is extracted. In order to confirm the dynamic characteristics of the leakage rate under random fluctuation conditions, the proportion sequence is grouped, the leakage duration probability in the adjacent interval is calculated, and the internal leakage dynamic increment data is generated based on the time variation law. The internal leakage dynamic increment data table records the hydraulic pump outlet flow loss, oil cylinder return port pressure fluctuation amplitude, valve core sealing leakage frequency and corresponding time section, which is used to accurately depict the fluid leakage trend caused by wear in the hydraulic system.

[0087] According to the lubrication performance degradation ratio, the oil film bearing capacity gradient of the hydraulic oil is determined. In this process, first, the pressure output history record, fluid temperature curve and oil viscosity change sequence of the main circuit of the hydraulic system are collected, and these data are coupled and aligned according to the lubrication performance degradation ratio. By analyzing the synchronicity of pressure output and oil temperature rise in the performance degradation stage, the rate of oil film thickness reduction with temperature rise is evaluated. When the oil film flow resistance reduction caused by viscosity reduction exceeds the system stability boundary, the bearing capacity reduction amplitude in this period is extracted in real time and recorded as a gradient data node. By comparing the cumulative change of the bearing capacity reduction trend in multiple cycles, the sustained attenuation level of the oil film support state is determined. The gradient calculation process combines the dividing line between the pressure stable interval and the instability interval to extract the stable transition point from each running cycle, compare the relative difference of the oil film shear capacity before and after the period, and obtain the oil film bearing capacity gradient sequence on the continuous time axis. This sequence reflects the segmented rate of the hydraulic oil support force and shear load in the degradation process, providing input conditions for subsequent multi-parameter gray analysis.

[0088] The dynamic leakage data obtained in step S241 is combined with the oil film carrying capacity decline gradient determined in step S242, and multi-parameter gray correlation processing analysis is carried out. First, the two types of data are matched according to the time index and the working component to form a two-dimensional data array in time parallel. Taking the leakage increment value as the main variable and the oil film support decline gradient as the contrast variable, the synchronous difference of the change in the unit time period is calculated. Then, the cumulative difference method is used to compare the increment amplitude change, the difference value is converted into a gray scale value, and a quantitative distribution interval is constructed through the gray mapping process. For the independent working condition data of the hydraulic pump and the oil cylinder and other core components, the concentrated distribution range of the gray scale value is extracted in segments, and the frequency distribution curve corresponding to each gray scale is counted to identify the correlation strength between the leakage rate and the carrying decline. Then, the gray mean value offset of the same component in different running stages is calculated through time sequence superposition, which represents the coupling degree of fluid loss and support capacity decline. After integrating all the calculation results, a multi-parameter correlation structure is generated, which contains parameter index number, time period identifier, gray level and double-variable correlation density fields, etc., for describing the dynamic coupling behavior of the hydraulic circuit under lubrication degradation.

[0089] Based on the multi-parameter correlation structure generated in step S243, the hydraulic power output reduction evaluation is carried out and the hydraulic power output reduction data is generated. First, the time period with a higher gray level in the correlation structure is defined as the low-efficiency area of the hydraulic transmission, and the corresponding pressure output value and flow feedback data are extracted from it. By comparing the baseline output curve in the same time period, the power output reduction and the fluid energy loss amplitude are determined. Then, according to the gray density distribution in the correlation structure, the energy transmission deviation of each gray interval is weighted and counted to generate a power decay mean sequence. This sequence is developed in time sequence, showing the dynamic evolution of output reduction in the running period. Then, by smoothing interpolation, the power reduction trend in each continuous time zone is analyzed to determine the energy decay rate of the hydraulic system at different stages, and the output proportion of the pump body, valve group and oil cylinder is labeled according to the system components. Finally, the decay values of each component are integrated into the total output reduction index to generate the hydraulic power output reduction data. The data is output in a structured form, including time index, energy loss proportion, main power circuit number and decay type label, providing quantitative basis for the input of the health monitoring and control logic layer of the variable pitch system.

[0090] Step S3 includes the following steps:

[0091] Step S31: Perform feature structure analysis on the hydraulic power output reduction data to generate power output reduction feature structure;

[0092] Step S32: Perform convolution processing on the power output reduction feature structure to obtain power reduction feature convolution data;

[0093] Step S33: determining the pitch control offset based on the power reduction feature convolution data;

[0094] Step S34: identifying the hydraulic power abnormality pre-warning risk factor of the power reduction feature convolution data according to the pitch control offset, to obtain the hydraulic power abnormality pre-warning risk factor.

[0095] In the embodiment of the present application, the hydraulic power output reduction data is taken as the input basis to analyze the feature structure of the hydraulic output change of the pitch system of the wind turbine generator set. First, a dynamic data matrix is established with time as the vertical axis and output power as the horizontal axis, and by analyzing the curve slope of the energy output and time distribution in the matrix, the main change section and the stable section of the power reduction are identified. Then, based on the output data of the hydraulic pump, the execution oil cylinder and the control valve, the output trend of each component is separately counted, and the relative attenuation ratio of energy transmission is calculated. Then, the number of peaks and troughs in the output reduction process is observed by the sliding window statistical method to determine the periodic characteristics of the output reduction. The fluctuation characteristics appearing in different time periods are divided into short-period disturbance and long-term recession, and are labeled in the matrix. According to the above analysis, the multi-dimensional feature data such as the pressure response delay of the hydraulic system, the flow feedback deviation and the output energy attenuation rate are aligned, to form a three-dimensional structured feature set, in which each dimension represents a time node, a component type and an energy distribution intensity. After structure induction and correlation analysis, the output power output reduction feature structure data set is output, which provides a clear input feature distribution for subsequent convolution feature extraction.

[0096] Based on the power output reduction feature structure generated in step S31, the output feature sequence of the hydraulic system is processed by convolution. First, a multi-layer feature array with time as the sequence main shaft is established, and the output reduction structure value of each time node is mapped to a continuous sequence. A two-dimensional local scan is performed on the sequence, and the local average value of the energy change gradient is extracted in each time period to obtain the spatial correlation feature of the power reduction. Then, according to the convolution calculation logic, the convolution kernel is slid on the feature array with a fixed step size, and the average difference of the energy change value in the local section extracted each time and its front and back time is compared to identify the significant mutation feature in the reduction process. By multiple superimposed convolution operations, energy change information at different levels is refined, and feature enhancement is formed in the time and amplitude directions. In order to ensure the stability of the convolution processing, the abnormal fluctuation interval of the data is interpolated and smoothed, so that the gradient of the convolution response is more continuous. The finally output power reduction feature convolution data includes the power decay intensity, the energy mutation section identifier and the time sequence local enhancement feature, which provides the basic feature input for the accurate determination of the pitch control offset in the next step.

[0097] According to the power reduction feature convolution data obtained in step S32, the determination of the control offset of the variable pitch system of the wind turbine generator set is carried out. First, the convolution data is aligned with the pitch angle position measurement signal, the pitch control pressure feedback and the actuator displacement signal according to the time index. Then, the hydraulic response delay time of different blade groups in the same period is compared to detect the output difference caused by the inconsistent action of the control valve. The high-intensity section of the energy fluctuation in the convolution feature is compared with the instantaneous offset section of the pitch angle change rate to extract the pitch angle change position corresponding to the power reduction peak, and the offset trend of the control system is identified therefrom. The relative time lag between the target position and the actual response position of the variable pitch system is calculated, and the offset difference of the three groups of blades in the same period is counted. The synchronization error of the pitch control of each blade is analyzed through the concentration degree of the offset distribution, the synchronization deviation time is associated with the output decay intensity, and the offset ratio of the control response under the energy constraint is judged. The finally obtained variable pitch control offset data consists of the offset amplitude, the time lag and the offset duration, which provides key input basis for identifying the pre-risk factors of hydraulic power abnormalities.

[0098] After determining the variable pitch control offset, the variable pitch control offset is combined with the power reduction feature convolution data to extract the pre-risk factors of the hydraulic power abnormalities. First, the offset time sequence and the output reduction time sequence are superimposed, and the relative positions of the offset peak and the reduction peak are aligned on the time axis. Then, the stability of the energy output before and after the offset peak is detected, and the abnormal energy transmission period is determined by calculating the pressure fluctuation frequency, the flow variation amplitude and the actuator response delay in the precursor section. The main hydraulic parameters of all abnormal periods are extracted, including the pressure decay rate, the flow unevenness coefficient and the mechanical response amplitude difference, and these data are mapped to the corresponding rows and columns of the risk index matrix. Through statistical analysis of the synergy between the columns in the matrix, the hydraulic links that have shown a potential energy loss trend before the offset occurs are identified. The part of data after normalization sorting constitutes the pre-risk factors of the hydraulic power abnormalities, including the main risk parameters, the occurrence time interval, the component number of the energy offset source and the risk level label, which are used as the input basis in the subsequent health state identification architecture design stage.

[0099] Step S33 includes the following steps:

[0100] Step S331: calculating the output pressure decay recursive mean difference based on the power reduction feature convolution data;

[0101] Step S332: calculating the actuator extension and retraction amount and the extension and retraction rate offset variance by pressure decay recursive mean difference back calculation;

[0102] Step S333: evaluating the pitch angle offset angle interval according to the actuator extension and retraction amount and the extension and retraction rate offset variance;

[0103] Step S334: offset angle memory learning is performed on the pitch angle offset angle interval to obtain offset angle memory learning data;

[0104] Step S335: the pitch control offset is determined through the offset angle memory learning data.

[0105] In the embodiment of the present application, after the power loss feature convolution data is generated, the output pressure time sequence of the hydraulic system is taken as the main analysis object, the average difference of the pressure values in different time windows is calculated, and the output pressure decay recursive average difference is obtained. First, the pressure value sequence of each time section is extracted from the convolution data, and the pressure average value change rate in the adjacent window is calculated in a sliding time window manner. According to the time advancing sequence, the change rate data is accumulated and superimposed to form a continuous recursive average difference sequence, which is used to reflect the smooth trend of pressure decay. In order to identify the stable area and abnormal area of system performance decay, a difference calculation is performed on the average difference sequence, the change law of the decay rate in each time period is detected, and the time distribution interval of the fluctuation peak point is extracted. Then, the recursive average difference and the flow feedback signal are paired synchronously to verify the response delay degree of the hydraulic system in the pressure loss stage, and the smoothing coefficient of the recursive average difference curve is adjusted to make the smooth change trend of the output pressure decay more continuous. The finally output pressure decay recursive average difference data contains time layer identification, average difference change amplitude and curve fluctuation period, which is used to provide basic input for oil cylinder running state back calculation.

[0106] After obtaining the pressure decay recursive average difference data, according to the physical dependence relationship between pressure and hydraulic actuator stroke, the back calculation of the stroke amount and stroke rate offset variance of the hydraulic cylinder is performed. First, the stroke time sequence of the cylinder body is established, the pressure recursive average difference sequence is completely aligned with the oil cylinder position sensing signal in the time dimension. The change trend of the pressure decay signal and the expansion and contraction characteristics of the oil cylinder displacement curve in the time period are compared, and the pressure drop rate is mapped to the expansion and contraction response time difference. The total expansion and contraction amount of the oil cylinder in multiple time periods is determined by the difference accumulation method. Then, the expansion and contraction rate data of the continuous time period is calculated, the rate change sequence is divided by time, the fluctuation variance of the rate value in each time window is obtained, and the strength of the hydraulic response stability offset is reflected. The relative amplitude of the rate difference between adjacent sampling sections is calculated to identify the abnormal interval in the speed change. The output of this step is the expansion and contraction amount sequence and the expansion and contraction rate offset variance data, which contains time index, displacement increment, rate deviation level and response delay index, and provides quantitative input basis for the evaluation of the pitch control offset angle interval.

[0107] The cylinder extension and retraction amount and the extension and retraction rate offset variance obtained in step S332 are combined to evaluate the pitch angle offset angle interval. First, the output sequence of the blade angle control command is obtained in the pitch feedback sensing system, and is paired with the cylinder extension and retraction amount curve for consistency. By comparing the numerical difference between the set instruction angle and the actual feedback angle, the offset response interval in the pitch control system is identified. To determine the concentration range of the offset angle, the cylinder extension and retraction amount is mapped by the main shaft rotation ratio to obtain the mechanical interval corresponding to the blade offset angle. Then, the offset angle and the retraction rate deviation in each time period are analyzed synchronously to detect the correlation between the offset angle in the rate rising stage and the energy loss. According to the cylinder offset duration, response lag time and offset amplitude distribution, the angle interval is divided into three sections: high offset, medium offset and stable section. The divided interval is subjected to section statistics, the proportion of each type of offset time and the angle boundary position are calculated, and the pitch angle offset angle interval data is formed, which provides regionalized input for the pitch control offset memory learning process.

[0108] After the pitch angle offset angle interval data is generated, offset memory learning is performed to establish the dynamic correlation relationship of the pitch offset with the running schedule. First, all offset angle interval sequences are arranged in chronological order with time as the main axis to form a time sequence matrix, each row corresponding to an offset occurrence, including the offset start point, end point and amplitude range. Then, the similarity between consecutive offset intervals is calculated, and the difference proportion of the offset sequence in the adjacent period is accumulated and superimposed to form an offset repetition rate sequence. This sequence is used to judge the repetition deviation in the execution of the system control command. The frequency of repeated occurrence of the offset angle in a specific time period is then calculated by the sliding window method, and the high-frequency region is extracted and its duration is recorded. The formed repeated offset interval is weighted and accumulated to obtain an offset persistence index, and the result sequence of the index changing with time is stored to form the offset angle memory learning data. The data set contains offset amplitude, repetition frequency, stability level and historical correlation degree, which is used in the subsequent offset amount determination stage to realize dynamic tracing of the offset trend.

[0109] The offset angle memory learning data output in step S334 is calculated to determine the pitch control offset. First, the high-frequency angle interval range in the offset memory sequence is calculated, the offset amplitude data with high repetition degree in multiple time periods is extracted, and the time accumulation proportion is counted. The offset angle data of the three oil cylinders of the same blade group are compared synchronously, the synchronization error of the offset response of each oil cylinder is calculated, and a spatial multi-point offset distribution map is formed. Then, the maximum offset amplitude and its duration in each time period are counted, and the offset peak value in the energy output loss larger interval is identified. The offset peak value interval is overlapped and analyzed with the energy decline in the hydraulic output convolution data, and the time corresponding relationship between the pitch angle change and the hydraulic energy recession is extracted. The offset amplitude is time-smoothed by a multi-cycle cumulative average method to determine the center value and fluctuation range of the overall offset trend, and finally the pitch control offset data is output, including the offset angle center value, time duration, offset direction category and offset level index, which provides input conditions for the identification of hydraulic power abnormal risk factors.

[0110] Step S4 includes the following steps:

[0111] Step S41: learning features of the pre-position risk factors of the hydraulic power abnormality to obtain pre-position risk learning factors;

[0112] Step S42: performing linear discriminant analysis on the pre-position risk learning factors to obtain risk learning linear discriminant factors;

[0113] Step S43: using a multilayer perceptron in a deep learning algorithm to design a health risk identification architecture for the risk learning linear discriminant factors to obtain a health risk identification architecture, and sending the health risk identification architecture to a wind turbine generator set control center to perform health management of the wind turbine generator set.

[0114] In the embodiment of the present application, the abnormal hydraulic power pre-position risk factor is taken as the input data source, and its internal correlation law is extracted by feature learning. First, the pre-position risk factor data is arranged in time sequence, and a risk sequence matrix containing pressure fluctuation, flow deviation, oil temperature anomaly, energy loss rate, pitch deviation amplitude and other multi-dimensional indexes is established. The rows of the matrix correspond to the time nodes, and the columns correspond to the physical quantity categories. Then, the change trend of each index over time is calculated, and the mean square deviation and change amplitude of each index in different time periods are calculated by sliding time window to judge the sensitivity of the index to system fluctuation. According to the sensitivity value, the index values are mapped to a unified interval by normalization method, so that different risk sources have comparability. Then, the correlation between multi-dimensional indexes is analyzed, the dominant and passive relationship between risk factors is identified by calculating the dynamic cooperative change proportion between adjacent columns, for example, the triggering effect of pressure drop on energy attenuation or the delayed response of flow change to oil temperature rise. The above feature relationships are integrated into a feature cooperation matrix, and the feature combination with significant cooperation is extracted as the core influence group. After traversal statistics, the comprehensive weighted pre-position risk learning factor is output, which describes the key variable set of risk triggering of the hydraulic system under different working conditions and provides basic input for subsequent discriminant analysis.

[0115] The obtained pre-position risk learning factor is subjected to linear discriminant analysis, and the high-dimensional complex risk data is mapped into a discriminable discriminant vector to form a risk learning linear discriminant factor. First, the training set and the validation set are constructed, and the risk learning factors of each time period are classified according to the running state, including stable state, slight abnormality and serious abnormality. Then, the mean vector and the within-sample variance of each category factor set are calculated, and the best discriminant direction is determined by comparing the distribution characteristics of the class difference and the class difference. The feature direction with larger total variance between classes is used for mapping, and the original multi-dimensional factor is compressed to form a low-dimensional classifiable combination vector. The combination vector is the risk learning linear discriminant factor, which can form a separable feature space under different risk categories. In order to ensure the consistency of the sample, long-time continuous monitoring data is selected, and the classification boundary is dynamically corrected by sliding update. When a new risk factor sample is input, its projection distance in the linear discriminant space can be judged by measuring its projection distance in the linear discriminant space. The final output risk learning linear discriminant factor includes risk category number, linear weight sequence and projection interval distribution, which is used as a direct input variable of the health risk identification architecture.

[0116] After obtaining the linear discriminant factor of risk learning, a health risk identification architecture is designed using a multilayer perception algorithm and deployed to a wind turbine control center. First, the number of input layer nodes is determined according to the number of risk learning linear discriminant factors and the data dimension, and the weight values in each discriminant vector are input into the input unit of the perception machine in chronological order. The multi-layer structure of the hidden layer is set, and the input signal is weighted and summed by the non-linear nodes in each layer to extract the complex non-linear mapping relationship between high-dimensional features. Each layer of nodes receives the output of the previous layer, and through continuous activation and inhibition calculation, a multi-stage feature extraction process is formed. The output layer classifies and determines according to the risk category label of the training data, and the identification result is divided into four categories: normal operation, energy anomaly, hydraulic attenuation and control deviation. In each iteration, the deviation between the system output and the actual detection label is calculated, and then the connection weights between nodes are adjusted through the error back propagation mechanism, so that the error of the network gradually decreases, and the identification result tends to be stable. After training, the generated health risk identification architecture includes the full connection path relationship from input to output and the corresponding weight parameter matrix. The architecture data is packaged in standard binary form and transmitted to the wind turbine control center through the industrial communication bus. After receiving, the control center analyzes the hydraulic power data in real time according to the architecture and automatically executes the health management logic, so as to realize the classification and continuous monitoring of the hydraulic system health status within the operation cycle.

[0117] The application also provides a wind turbine health management system based on deep learning, which is used to execute the wind turbine health management method based on deep learning as described above, and the wind turbine health management system based on deep learning comprises:

[0118] A monitoring data acquisition module is configured to deploy an oil multi-parameter sensor in a hydraulic component in a variable pitch system of a wind turbine, and then acquire oil state monitoring data; and calculate the water content abnormal growth value proportion of the oil state monitoring data within a unit time.

[0119] A hydraulic power output reduction evaluation module is configured to estimate a hydraulic oil lubrication performance degradation ratio based on the water content abnormal growth value proportion, to obtain the lubrication performance degradation ratio; and evaluate the hydraulic power output reduction based on the lubrication performance degradation ratio, to generate hydraulic power output reduction data.

[0120] A front risk factor identification module is configured to determine a variable pitch control deviation based on the hydraulic power output reduction data, and then identify a hydraulic power abnormality front risk factor, to obtain the hydraulic power abnormality front risk factor.

[0121] The health risk identification architecture design module is configured to utilize a multilayer perceptron in a deep learning algorithm to design a health risk identification architecture for the hydraulic power abnormality preposed risk factor, and send the health risk identification architecture to a wind turbine generator system control center to perform health management of the wind turbine generator system.

[0122] The above description is merely that of the embodiments of the application, to enable any person skilled in the art to understand or implement the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A deep learning-based health management method for wind turbine generator sets, characterized in that, Includes the following steps: Step S1: Deploy multi-parameter oil sensors in the hydraulic components of the wind turbine pitch system to collect oil condition monitoring data; Calculate the percentage of abnormal increase in water content per unit time from oil condition monitoring data; Step S2: Estimate the deterioration ratio of hydraulic oil lubrication performance based on the percentage of abnormal increase in water content to obtain the deterioration ratio of lubrication performance; evaluate the reduction in hydraulic power output based on the deterioration ratio of lubrication performance to generate hydraulic power output reduction data. Step S3: Determine the pitch control offset based on the hydraulic power output loss data, and then identify the hydraulic power anomaly precursor risk factors to obtain the hydraulic power anomaly precursor risk factors. Step S4: Use the multilayer perceptron in the deep learning algorithm to design a health risk identification architecture for the anterior risk factors of hydraulic power anomalies, obtain the health risk identification architecture, and send it to the wind turbine control center to perform health management of the wind turbine. Step S2 includes the following steps: Step S21: Obtain basic performance parameters of hydraulic oil; Step S22: Estimate the deterioration ratio of hydraulic oil lubrication performance based on the abnormal increase in water content and the basic performance parameters of hydraulic oil to obtain the deterioration ratio of lubrication performance. Step S23: Quantify the degree of wear aggravation of hydraulic components based on the lubrication performance degradation ratio; Step S24: Evaluate the reduction in hydraulic power output based on the lubrication performance degradation ratio and wear aggravation degree, and generate hydraulic power output reduction data; Step S22 includes the following steps: Step S221: Extract the basic viscosity and demulsibility from the basic performance parameters of the hydraulic oil; Step S222: Based on the abnormal increase in water content, simulate the polar balance loss process between oil film molecular chains for the basic viscosity and demulsibility to obtain polar balance loss data between oil film molecular chains. Step S223: Quantify the oil film rupture probability index based on the polarity balance loss data; Step S224: Based on the oil film rupture probability index and polarity balance loss data, perform dynamic viscosity loss divergence assessment to obtain dynamic viscosity loss divergence data; Step S225: Estimate the deterioration ratio of hydraulic oil lubrication performance based on dynamic viscosity loss divergence data and oil film rupture probability index to obtain the deterioration ratio of lubrication performance.

2. The deep learning-based wind turbine health management method according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Deploy multi-parameter oil sensors in the hydraulic components of the wind turbine pitch system to collect oil condition monitoring data. Step S12: Embed timestamps into the oil condition monitoring data to generate oil condition monitoring time series data; Step S13: Extract the abnormal growth trend of water content from the oil condition monitoring time series data; Step S14: Calculate the abnormal growth trend of water content and the percentage of abnormal growth in water content per unit time.

3. The deep learning-based wind turbine health management method according to claim 1, characterized in that, Step S23 includes the following steps: Step S231: Based on the lubrication performance degradation ratio, simulate and deduce the degree of shear resistance reduction; Step S232: Perform nonlinear time-series incremental fitting on the degree of shear resistance reduction to obtain time-series incremental fitting data of performance reduction; Step S233: Based on the time-series incremental fitting data of performance loss, derive the boundary friction disorder increment of hydraulic components to obtain friction disorder increment data; Step S234: Quantify the degree of wear aggravation of hydraulic components based on the aforementioned friction disorder increment data.

4. The deep learning-based health management method for wind turbine generators according to claim 1, characterized in that, Step S24 includes the following steps: Step S241: Based on the degree of wear aggravation, the value of the expansion of the mating clearance of the evolved components is calculated, and then the dynamic increment value of internal leakage is calculated probabilistically. Step S242: Determine the gradient of oil film load-bearing capacity reduction based on the aforementioned lubrication performance degradation ratio; Step S243: Perform multi-parameter grayscale correlation processing based on the internal leakage dynamic increment value and the oil film bearing capacity decrease gradient to generate a multi-parameter correlation structure; Step S244: Based on the multi-parameter correlation structure, perform hydraulic power output loss assessment and generate hydraulic power output loss data.

5. The deep learning-based health management method for wind turbine generators according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Perform feature structure analysis on the hydraulic power output loss data to generate a power output loss feature structure; Step S32: Perform convolution processing on the power output loss feature structure to obtain power loss feature convolution data; Step S33: Determine the pitch control offset based on the convolutional data of power loss reduction features; Step S34: Identify the hydraulic power anomaly precursor risk factors by performing a power loss feature convolution data on the pitch control offset, and obtain the hydraulic power anomaly precursor risk factors.

6. The deep learning-based health management method for wind turbine generators according to claim 5, characterized in that, Step S33 includes the following steps: Step S331: Calculate the recursive mean difference of output pressure attenuation based on the convolutional data of the dynamic loss reduction features; Step S332: Calculate the cylinder extension / retraction amount and the variance of the extension / retraction rate deviation by recursively back-calculating the pressure decay. Step S333: Evaluate the pitch angle offset range based on the cylinder extension / retraction amount and the variance of the extension / retraction rate offset; Step S334: Perform offset memory learning on the pitch angle offset angle range to obtain offset angle memory learning data; Step S335: Determine the pitch control offset by using the offset angle memory learning data.

7. The deep learning-based health management method for wind turbine generators according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Perform feature learning on the pre-risk factors of hydraulic power anomalies to obtain pre-risk learning factors; Step S42: Perform linear discriminant analysis on the pre-learning risk factors to obtain the linear discriminant factors for risk learning; Step S43: Use the multilayer perceptron in the deep learning algorithm to design a health risk identification architecture for learning linear discriminant factors of risk, obtain the health risk identification architecture, and send it to the wind turbine control center to perform health management of the wind turbine.

8. A deep learning-based health management system for wind turbine generators, characterized in that, For executing the deep learning-based wind turbine health management method as described in claim 1, the deep learning-based wind turbine health management system includes: The monitoring data acquisition module is used to deploy multi-parameter oil sensors in the hydraulic components of the wind turbine pitch system to collect oil condition monitoring data; and to calculate the percentage of abnormal increase in water content per unit time. The hydraulic power output reduction assessment module is used to estimate the deterioration ratio of hydraulic oil lubrication performance based on the percentage increase in water content, so as to obtain the lubrication performance deterioration ratio; and to assess the hydraulic power output reduction based on the lubrication performance deterioration ratio, thereby generating hydraulic power output reduction data. The pre-risk factor identification module is used to determine the pitch control offset based on the hydraulic power output loss data, and then to identify the pre-risk factors of hydraulic power anomaly to obtain the pre-risk factors of hydraulic power anomaly. The health risk identification architecture design module is used to design a health risk identification architecture for the anterior risk factors of hydraulic power anomalies using a multilayer perceptron in deep learning algorithms. The resulting health risk identification architecture is then sent to the wind turbine control center to perform health management of the wind turbine.

Images