A fan blade intelligent monitoring and optimization method based on deep learning

Abstract

The application discloses a kind of fan blade intelligent monitoring and optimization method based on deep learning, it is related to wind power generation technology field, including: the fatigue accumulation value corresponding to each historical period strain data is calculated;According to the fatigue state prediction model, the current fatigue state of each local area is predicted, and according to the current fatigue state, a plurality of candidate load adjustment strategies are generated by pre-training reinforcement learning model;Determine the short-term trend change characteristics of fan blade fatigue, and according to the difference between the short-term trend change characteristics and the current fatigue state, the target load adjustment strategy is determined and screened, and the active load control signal is generated to implement active adjustment to the local area load of fan blade;The application realizes high-precision prediction and active load adjustment of fan blade fatigue state by deep learning and multi-scale convolution feature extraction, so as to significantly improve the service life and operation safety of fan blade.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of wind power generation technology, in particular to a wind turbine blade intelligent monitoring and optimization method based on deep learning. BACKGROUND

[0002] As an important part of renewable energy, the operation reliability and safety of wind turbine generators are crucial for ensuring power supply. In the process of wind power generation, wind turbine blades, as the core components, are subjected to complex aerodynamic loads and environmental influences for a long time, and are prone to fatigue damage. Especially in the process of high wind speed and frequent start-stop, the fatigue damage of the blades will gradually accumulate, eventually leading to blade damage or performance degradation. Therefore, monitoring the fatigue state of the wind turbine blades and performing timely maintenance and optimization adjustment are important measures to improve the operation efficiency and safety of the wind turbine generators.

[0003] Currently, the monitoring methods for wind turbine blades mainly include monitoring systems based on vibration, strain sensors and image processing technology. However, the existing monitoring systems still have problems such as insufficient precision, unstable prediction effect and inability to accurately adjust the load in real time. These problems make the existing technology unable to comprehensively solve the problems of real-time monitoring, prediction and active adjustment of wind turbine blade fatigue damage.

[0004] In view of the above technical problems, the existing technology still lacks an effective solution that can accurately and low-power monitor the fatigue state of the wind turbine blades in real time and intelligently adjust the load according to the monitoring results. SUMMARY

[0005] The purpose of the present application is to provide a wind turbine blade intelligent monitoring and optimization system and method based on deep learning to solve the problems in the background art.

[0006] In order to achieve the above purpose, the present application provides the following technical solutions:

[0007] The present application provides a wind turbine blade intelligent monitoring and optimization method based on deep learning, comprising:

[0008] S101: According to the historical strain data of the local area of the wind turbine blade, the fatigue accumulation value corresponding to each historical period strain data is calculated by combining the fatigue accumulation rate calculation method, and the historical strain data is taken as the input and the fatigue accumulation value is taken as the output to train the fatigue state prediction model;

[0009] S102: According to the fatigue state prediction model, the current fatigue state of each local area is predicted, and according to the current fatigue state, a plurality of candidate load adjustment strategies are generated by pre-training the reinforcement learning model;

[0010] S103: Collect real-time vibration data of the fan blade, determine the short-term trend change characteristics of the fan blade fatigue, and select a target load adjustment strategy from the plurality of candidate load adjustment strategies according to the difference between the short-term trend change characteristics and the current fatigue state;

[0011] S104: According to the determined target load adjustment strategy, an active load control signal is generated by the fan active load control system to implement active adjustment of the local area load of the fan blade.

[0012] In the above technical solution, the technical effects and advantages provided by the present application are as follows:

[0013] The present application realizes high-precision prediction of the fan blade fatigue state through the fatigue state prediction model based on deep learning. By using the multi-scale convolution feature extraction structure and combining historical strain data and fatigue cumulative values, the strain characteristics of each local area of the fan blade can be fully captured, thereby obtaining more accurate fatigue prediction results. This improvement significantly improves the accuracy of fatigue monitoring in the prior art, avoids the potential risks caused by insufficient strain data processing and low prediction accuracy in traditional monitoring methods, and ensures the safety and stability of the fan blade during long-term operation.

[0014] The present application can realize active load adjustment of the fan blade by constructing a load adjustment strategy based on deep reinforcement learning. By designing the reinforcement learning state space, action space and reward function, and training the pre-trained reinforcement learning model based on historical fatigue data, the system can generate a plurality of candidate load adjustment strategies, and quickly respond to the current fatigue state of the fan blade to select the most suitable adjustment strategy. This innovation can significantly improve the service life of the fan blade, reduce damage or performance degradation caused by excessive fatigue, and improve the economy and stability of the wind turbine generator set. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0016] Figure 1 The flowchart of the present application is a kind of intelligent monitoring and optimization method for fan blade based on deep learning;

[0017] Figure 2 The framework diagram of the present application is a kind of intelligent monitoring and optimization system for fan blade based on deep learning. DETAILED DESCRIPTION

[0018] Example implementations are now described with reference to the drawings. Example implementations can, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these example implementations are provided so that this disclosure will be thorough and complete, and will fully convey the inventive aspects to those skilled in the art. The figures are not necessarily to scale, and the same reference numerals in different figures designate the same or similar components. For the clarity's sake, the description is divided into sections, and the sections are arranged in a certain order, but the arrangement of the sections is not meant to limit the scope of the disclosure.

[0019] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more example implementations. In the following description, numerous specific details are provided to give a thorough understanding of example implementations. One skilled in the relevant art will recognize, however, that the techniques disclosed can be practiced without one or more of the specific details, or with other methods, components, materials, and so forth. In other instances, well-known structures, materials, or actions are not shown or described in detail in order to avoid obscuring aspects of the disclosure.

[0020] Example 1

[0021] As Figure 1 shown, the embodiment disclosed provides a deep learning-based intelligent monitoring and optimization method for fan blades, including:

[0022] S101: According to the historical strain data of the local area of the fan blade, the fatigue accumulation value corresponding to the strain data of each historical period is calculated by combining the fatigue accumulation rate calculation method, and the historical strain data is taken as the input and the fatigue accumulation value is taken as the output to train and obtain a fatigue state prediction model;

[0023] In implementation, the training method of the fatigue state prediction model includes:

[0024] The historical strain data of the local area of the fan blade is collected, the historical strain data is segmented according to a fixed time length, and strain data samples corresponding to multiple historical periods are obtained;

[0025] It should be noted that the local area of the fan blade in this embodiment is a typical fatigue-prone area determined by finite element simulation analysis of the blade structure and statistical analysis of actual operation monitoring data, and specifically includes but is not limited to: the connection area of the blade root and the hub; the joint area of the blade main beam and the skin; the maximum bending moment section area of the middle section of the blade; the area near the leading edge of the blade tip.

[0026] Further need to explain is that first in each of the above local area surface mounting a plurality of high-precision fiber Bragg grating strain sensor (FBG sensor) for long-term operation in real time, continuous collection of historical strain data of fan blade, the historical strain data includes the following three specific content:

[0027] Strain amplitude: the micro tensile or compressive deformation of the material surface of the local area of the fan blade;

[0028] Strain frequency: the number of strain cycles extracted by rainflow counting method per unit time, indicating the number of strain cycles experienced by the local area of the fan blade;

[0029] Strain waveform signal: two-dimensional time series data sequence composed of sampling time point (unit: second) and corresponding strain amplitude.

[0030] Subsequently, in order to facilitate data processing and model training, the historical strain data of long period is divided into a plurality of independent historical period strain data samples in units of fixed time length (such as 60 minutes) to constitute the model training data set, wherein each strain data sample is a two-dimensional data form of "time series data points x number of measuring points".

[0031] According to the Miner fatigue damage criterion and the fatigue S-N curve of the fan blade material, the fatigue cumulative value corresponding to each historical period strain data sample is calculated respectively;

[0032] In specific implementation, in order to obtain the target output of model training, i.e. the fatigue cumulative value of each historical period, the method of combining the classical Miner fatigue damage criterion with the fatigue S-N curve of the fan blade material is adopted for calculation, and the specific process is fully and clearly disclosed as follows:

[0033] Firstly, the rainflow counting method is used to identify the cycles and extract the strain amplitude and actual cycle number of each cycle in the strain waveform signal of each historical period.

[0034] Secondly, the fatigue cumulative value is calculated based on the Miner fatigue damage criterion calculation formula:

[0035] ;

[0036] In the formula: The fatigue cumulative value (dimensionless) is used to represent the fatigue damage degree of the local area of the blade, and the value range of D is from 0 to 1, and when D=1, it indicates that the fatigue life of the material has been completely exhausted; The actual number of i-th strain cycles is obtained by rainflow counting method; The fatigue life cycle number corresponding to the strain amplitude based on the material fatigue S-N curve; The total number of all fatigue cycle grades is determined by the rainflow counting method.

[0037] Further need to explain, the fan blade material fatigue S-N curve refers to the blade material (such as glass fiber reinforced composite material, GFRP) under different strain amplitude conditions, the strain-life relationship curve obtained by fatigue test, its specific mathematical expression is as follows:

[0038] ;

[0039] In the formula: N is the fatigue life cycle number under the strain amplitude C and m are material parameters obtained by a large number of material fatigue tests; C and m are material parameters obtained by a large number of material fatigue tests; ;

[0040] A multi-scale convolution feature extraction structure is constructed, and convolution operation is performed on each historical period strain data sample to extract initial fatigue features corresponding to each scale;

[0041] Specifically, the specific construction method of the multi-scale convolution feature extraction structure comprises:

[0042] A plurality of parallel convolution branches with different convolution kernel sizes are constructed respectively, and a dilated convolution layer, a depth separable convolution layer and an attention mechanism module are sequentially arranged in each parallel convolution branch;

[0043] In specific implementation, the embodiment constructs a multi-scale convolution feature extraction structure for extracting multi-scale fatigue features from each historical period strain data sample, which comprises a plurality of parallel convolution branches, each convolution branch corresponds to a convolution kernel size (for example, 3×3, 5×5, 7×7), and the input data dimension is 6000 time points×8 measuring points (10 minutes sampling, 10Hz sampling frequency). The specific branch structure is:

[0044] Branch 1 (3×3 kernel): input 8 channels→dilated convolution (d=1)→output 64 channels→depth separable convolution→output 64 channels→CA module→output 64 channels;

[0045] Branch 2 (5×5 kernel): input 8 channels→dilated convolution (d=2)→output 128 channels→depth separable convolution→output 128 channels→CA module→output 128 channels;

[0046] Branch 3 (7×7 kernel): input 8 channels→dilated convolution (d=4)→output 256 channels→depth separable convolution→output 256 channels→CA module→output 256 channels.

[0047] In each parallel convolution branch, the initial convolution features of the corresponding scale are obtained through the dilated convolution layer, and then the initial convolution features are further convolved by using the depth separable convolution layer to obtain the refined fatigue features of the corresponding scale.

[0048] In this embodiment, the dilated convolution is used for initial feature extraction, and the dilated convolution of each convolution branch is specifically defined as follows:

[0049] ;

[0050] In the formula: is the convolution output feature; is the historical period strain data sample input; is the convolution kernel weight; is the dilation factor, which is 1, 2 and 4 respectively, to obtain different receptive field features; is the convolution kernel size, which is 3x3, 5x5 and 7x7 respectively.

[0051] In specific implementation, in order to further extract refined fatigue features, a depth separable convolution layer is arranged in each convolution branch, which includes two-step operations of “depth convolution” and “point-wise convolution”:

[0052] In the formula, the depth convolution operation formula is:

[0053] ;

[0054] In the formula: is the convolution output feature; is the cth channel of the corresponding input feature; is the depth convolution kernel; M and N are the depth convolution kernel size, specifically 3x3; the point-wise convolution uses a 1x1 convolution kernel to fuse the feature channels:

[0055] ;

[0056] In the formula: is the point-wise convolution output feature; C is the number of input feature channels; is the point-wise convolution kernel weight.

[0057] For the refined fatigue features obtained in each parallel convolution branch, the attention mechanism module is used to calculate the attention weight of the feature channel, and each channel of the refined fatigue features is weighted and adjusted according to the attention weight, so as to obtain the attention feature of each convolution branch;

[0058] In order to strengthen the fatigue feature representation of the key channel, a channel attention (CA) module is arranged in each parallel convolution branch, and the specific disclosure is as follows:

[0059] First, the spatial compression of each channel feature is performed by global average pooling (GAP) to obtain the global average value of each channel;

[0060] Then, two fully connected layers (FC) are used for channel feature mapping, and finally the attention weights of each feature channel are obtained through the sigmoid activation function:

[0061] ;

[0062] In the formula: is the channel attention weight; is the global average pooling; is the fully connected mapping; is the rectified linear unit function; is the Sigmoid function; is the input refined feature channel;

[0063] After obtaining the channel attention weight, multiply the refined fatigue feature by channel by channel to obtain the attention feature of each convolution branch, which is the multi-scale fatigue feature.

[0064] According to the correlation between different scale fatigue features, the fusion weight is determined, and the fusion fatigue feature of the strain data of each historical period is obtained. The fusion fatigue feature is used as the model input, and the fatigue cumulative value is used as the model output. The fatigue state prediction model is trained and obtained;

[0065] Specifically, the fusion weight is determined according to the correlation between different scale fatigue features, and the fusion fatigue feature of the strain data of each historical period is obtained, including:

[0066] According to the correlation between each scale fatigue feature, the corresponding correlation coefficient is calculated;

[0067] In specific implementation, according to the correlation between each scale fatigue feature, the fusion weight is determined and the feature fusion is performed; specifically, according to the correlation between the fatigue features of multiple scales (i.e. the attention features of different convolution branches), the Pearson correlation coefficient calculation method is used to calculate the correlation coefficient between each scale fatigue feature:

[0068] ;

[0069] In the formula: is the Pearson correlation coefficient of the jth scale fatigue feature and the fatigue cumulative value; is the fatigue feature value of the ith sample under the jth scale (which has been compressed to a scalar by global average pooling); is the mean value of the fatigue features of all samples under the jth scale; is the fatigue accumulation value of the i th sample; is the mean value of fatigue accumulation values of all samples; is the total number of samples.

[0070] According to the numerical value of the calculated correlation coefficient, the fusion weight corresponding to each scale fatigue feature is determined;

[0071] Specifically, the fusion weight corresponding to each scale fatigue feature is obtained according to the above calculated correlation coefficient, and the Softmax normalization method is used to determine the fusion weight:

[0072] ;

[0073] In the formula: is the fusion weight corresponding to the j th scale fatigue feature; is the correlation coefficient of the j th scale fatigue feature and the target output (fatigue accumulation value); K is the number of scales (for example, 3 scales).

[0074] According to the determined fusion weight, the fatigue features of different scales are weighted and fused respectively to obtain the fusion fatigue features of the strain data of each historical period;

[0075] Finally, according to the above fusion weight, the fatigue features of different scales are weighted and fused to obtain the fusion fatigue features corresponding to the strain data of each historical period:

[0076] ;

[0077] In the formula: is the fused fatigue feature; is the j th scale fatigue feature.

[0078] S102: According to the fatigue state prediction model, the current fatigue state of each local region is predicted, and according to the current fatigue state, a plurality of candidate load adjustment strategies are generated by pre-training a reinforcement learning model;

[0079] In implementation, the training method of the reinforcement learning model comprises:

[0080] According to the historical fatigue state data of the local region of the fan blade, the fatigue accumulation value changes of each local region before and after the implementation of the historical load adjustment strategy (the state after the implementation of the historical load adjustment strategy is obtained by simulating the OpenFAST fan simulation software) are determined, and the state space and action space of the reinforcement learning model are constructed;

[0081] In a specific implementation, first, according to the historical fatigue state data collected by each local region of the fan blade in different historical periods, the fatigue cumulative value changes of each local region before and after the implementation of the historical load adjustment strategy are determined, and the state space and action space of the reinforcement learning model are constructed based on the fatigue cumulative value changes, specifically as follows:

[0082] The historical fatigue state data specifically includes the fatigue cumulative values of each local region obtained in the historical operation process;

[0083] The state space is defined as a state vector composed of the fatigue cumulative values of each local region, for example:

[0084] ;

[0085] Among them, denote the fatigue cumulative values corresponding to the time t of the blade root, the blade tip, the blade edge and the blade center, respectively;

[0086] The action space is defined as different parameter combinations of the implemented load adjustment strategy, specifically, for example: the attack angle adjustment amplitude (for example , etc.) of the fan blade, the yaw angle adjustment range (for example , etc.), the change interval of the pitch angle (for example ), and the speed adjustment range (for example, the rated speed ), and the action space is represented as:

[0087] ;

[0088] Among them: is the attack angle adjustment amplitude corresponding to the i-th strategy; is the yaw angle adjustment range; is the change interval of the pitch angle; is the speed adjustment amplitude;

[0089] It should be noted that the attack angle adjustment amplitude is converted into the local region load change amount through the aerodynamic load distribution function, and the conversion formula is as follows:

[0090] ;

[0091] In the formula: is the load change amount of region j; is the region aerodynamic coefficient (for example, blade root: 0.8, blade tip: 1.2); is the lift curve slope (example value 0.12 / deg); is the equivalent area of the region (blade root: 2.5 , blade tip: 0.8 ), is the air density; for the wind speed;

[0092] Based on the amplitude of the change of the fatigue accumulation value, the fatigue damage propagation rate of each local region is determined, and a reward function of the reinforcement learning model is constructed;

[0093] Specifically, the construction method of the reward function of the reinforcement learning model comprises:

[0094] According to the fatigue accumulation values of each local region before and after the implementation of the historical load adjustment strategy, the change amplitude of the fatigue accumulation rate after the implementation of the load adjustment strategy relative to before the implementation is calculated, and the change amplitude is taken as the basic reward value of each local region;

[0095] In specific implementation, according to the fatigue accumulation values of each local region before and after the implementation of the historical load adjustment strategy, the change amplitude of the fatigue accumulation rate after the implementation of the load adjustment strategy relative to before the implementation is calculated, and the change amplitude is defined as the basic reward value, specifically:

[0096] First, the fatigue accumulation rate is defined as:

[0097] ;

[0098] Then, the basic reward value is calculated as:

[0099] ;

[0100] In the formula: is the fatigue accumulation value of the i th local region at time t ; is the time interval (such as 10 minutes) before and after the implementation of the strategy, is the basic reward value corresponding to the i th local region; is the fatigue accumulation rate in the period [t- , t] before the implementation of the strategy: ; is the fatigue accumulation rate in the period [t, t+ ] before the implementation of the strategy: .

[0101] According to the difference between the current fatigue accumulation value of each local region and the set fatigue safety threshold, the fatigue sensitivity coefficient corresponding to each local region is calculated;

[0102] Specifically, the determination method of the fatigue sensitivity coefficient comprises:

[0103] According to the difference between the fatigue limit value of the local region material of the fan blade and the current fatigue accumulation value, the fatigue margin corresponding to each local region is determined;

[0104] In a specific implementation, the fatigue margin is determined according to the difference between the current fatigue accumulation value (dimensionless value, range 0~1) and the material fatigue damage limit (i.e. damage limit value 1 when the fatigue life is exhausted), and the dimensionless definition is as follows:

[0105] ;

[0106] In the formula: is the fatigue margin of the i-th local region; is the fatigue limit value; is the current fatigue accumulation value of the i-th local region.

[0107] According to the size of the fatigue margin, the initial sensitivity coefficient corresponding to each local region is calculated;

[0108] Further, the initial sensitivity coefficient corresponding to each local region is calculated according to the size of the determined fatigue margin, for example, calculated by using an exponential function as follows:

[0109] ;

[0110] In the formula: is the initial sensitivity coefficient; is a constant to prevent zero, and the value range is, for example, 0.01, is the upper limit of the sensitivity coefficient, for example, the value is 10.

[0111] According to the wind speed level and load level of the blade during operation, a dynamic amplification factor of the fatigue state is determined;

[0112] Further, according to the wind speed level (for example, I~V level) and load level (for example, light load, medium load and heavy load) of the blade during operation, a dynamic amplification factor is determined through experiments, and the specific calculation formula is, for example:

[0113] ;

[0114] In the formula: is the dynamic amplification factor; , is a coefficient obtained according to actual wind turbine test data, and the numerical representation of the wind speed level and the load level is obtained through real-time sensor data (for example 0.1, 0.3); , are the numerical representation of the wind speed level and the load level, respectively;

[0115] The dynamic amplification factor and the initial sensitivity coefficient are multiplied one by one to obtain the fatigue sensitivity coefficient of each local region;

[0116] Finally, the dynamic amplification factor is multiplied by the initial sensitivity coefficient one by one to obtain the fatigue sensitivity coefficient:

[0117] ;

[0118] According to the weighting calculation of the basic reward value and the fatigue sensitivity coefficient one by one, a comprehensive reward value is obtained, and the comprehensive reward value is used to construct the reward function of the reinforcement learning model;

[0119] According to the weighting calculation of the above-mentioned basic reward value and the fatigue sensitivity coefficient one by one, a comprehensive reward value is obtained, and the reward function of the reinforcement learning model is constructed as follows:

[0120] .

[0121] The state space, action space and reward function constructed are used as inputs of the reinforcement learning model, and a deep reinforcement learning algorithm is used to train the reinforcement learning model to obtain a pre-trained reinforcement learning model for real-time generation of candidate load adjustment strategies;

[0122] In specific implementation, the state space, action space and reward function constructed are used as inputs, and a deep reinforcement learning algorithm (such as DDPG algorithm) is used to train the reinforcement learning model, specifically including: taking the state space vector as input; the action space vector is output; and the comprehensive reward value is used as feedback evaluation value.

[0123] Iterative training is performed until convergence to obtain a pre-trained reinforcement learning model for real-time generation of candidate load adjustment strategies.

[0124] S103: Collecting real-time vibration data of the fan blade, determining the short-term trend change characteristics of the fan blade fatigue, and according to the difference between the short-term trend change characteristics and the current fatigue state, screening and determining a target load adjustment strategy from the plurality of candidate load adjustment strategies;

[0125] In implementation, the screening and determining method of the target load adjustment strategy includes:

[0126] Collecting real-time vibration data of the fan blade, and determining the short-term trend change characteristics of the fan blade fatigue according to the real-time vibration data;

[0127] In specific implementation, first, a plurality of three-axis vibration sensors installed at different positions on the surface of the fan blade are used to collect real-time vibration data of local areas (blade root, blade tip, blade edge and blade center) of the fan blade, and the real-time vibration data specifically includes time domain waveform signal data of vibration and corresponding real-time sampling frequency, vibration amplitude and vibration acceleration data, and is sent to the control center through a wireless transmission module (ZigBee protocol).

[0128] The method for determining the short-term trend change characteristic of the fan blade fatigue specifically comprises:

[0129] The vibration data of the local area of the fan blade is collected in real time, and the corresponding frequency domain characteristic parameters are obtained through a fast Fourier transform method.

[0130] In a specific implementation, the vibration time domain waveform data of the local area of the fan blade collected in real time is converted into a frequency domain signal through a fast Fourier transform (FFT) method to obtain frequency domain characteristic parameters, including a frequency spectrum line, a frequency spectrum amplitude and a vibration energy spectrum.

[0131] Exemplarily, the calculation formula of the fast Fourier transform is specifically as follows:

[0132] ;

[0133] In the formula, f is a frequency domain spectrum; is a frequency domain spectrum; is the nth time domain data point of real-time sampling; N is the number of sampling points in each sampling period; and k is a discrete frequency index of the spectrum.

[0134] According to the frequency domain characteristic parameters, the vibration main frequency, the frequency spectrum amplitude and the corresponding vibration energy of the local area of the blade at the current time are extracted, and real-time vibration characteristic values are calculated and obtained.

[0135] Further, according to the frequency domain characteristic parameters, the vibration main frequency (i.e., the frequency corresponding to the highest amplitude in the spectrum) at the current time, the corresponding frequency spectrum amplitude and the corresponding vibration energy value of the local area of the blade at the current time are extracted, and the real-time vibration characteristic values are further calculated and obtained.

[0136] In a specific implementation, the real-time vibration characteristic values are as follows:

[0137] ;

[0138] In the formula, f is a frequency domain spectrum; is a real-time vibration characteristic value at the current time; is a maximum value of the frequency spectrum amplitude at the current time; is a vibration main frequency at the current time; is a vibration energy value corresponding to the current time; is a corresponding reference amplitude, frequency and energy (determined according to experimental data, for example, a historical mean value, a maximum value or a standard value).

[0139] According to the real-time vibration characteristic values and the historical vibration characteristic values of the local area of the fan blade, a vibration characteristic value change rate in the current period is calculated, and the vibration characteristic value change rate is determined as a short-term trend change characteristic of the fan blade fatigue.

[0140] It should be noted that the vibration characteristic value change rate is calculated according to the difference between the real-time vibration characteristic value and the blade local region historical vibration characteristic value (vibration characteristic value of the previous period), and the calculation formula is as follows:

[0141]

[0142] In the formula: is the vibration characteristic value change rate in the current period; is the real-time vibration characteristic value at the current moment; is the historical vibration characteristic value at the previous moment; is the vibration characteristic value update period (for example, 5 minutes or 10 minutes);

[0143] According to the numerical difference between the short-term trend change characteristic and the current fatigue state predicted by the fatigue state prediction model, the difference level is determined;

[0144] In specific implementation, according to the numerical difference between the determined short-term trend change characteristic and the current fatigue state predicted by the aforementioned real-time fatigue state prediction model, the specific difference level is determined:

[0145] Specifically, the calculation formula of the numerical difference is:

[0146]

[0147]

[0148] In the formula: is the difference value at the current moment; is the fatigue accumulation value corresponding to the current fatigue state predicted by the real-time fatigue state prediction model; represents the current fatigue accumulation value predicted by the fatigue state prediction model;

[0149] It should be noted that: in order to ensure the comparability between the vibration characteristic value change rate and the fatigue accumulation value change rate, the embodiment is based on a large amount of historical experimental data fitting to establish a linear mapping relationship between the two, which is used to convert the vibration characteristic change rate into the fatigue damage change rate, as follows:

[0150] First, the fatigue damage change rate corresponding to the vibration characteristic change is calculated:

[0151]

[0152] Among them, the parameters , are determined according to historical data fitting, for example , .

[0153] Then the numerical difference is calculated:​​​​

[0154] ;

[0155] Exemplarily, according to the size of the numerical difference, the difference level is set as follows:

[0156] Low difference level (I level): if , the difference level is defined as "low", indicating that the short-term change is basically consistent with the predicted state, and the blade fatigue development is smooth;

[0157] Medium difference level (II level): if , the difference level is defined as "medium", indicating that there is a significant difference in blade fatigue, and the strategy adjustment intensity should be improved;

[0158] High difference level (III level): if , the difference level is defined as "high", indicating that the fatigue state is seriously beyond the prediction, and emergency load adjustment intervention is needed.

[0159] For each candidate load adjustment strategy, the implementation effect of the candidate load adjustment strategy is evaluated according to the pre-constructed multi-factor fuzzy weight comprehensive evaluation system, and the strategy evaluation value corresponding to each candidate strategy is obtained;

[0160] The specific construction method of the multi-factor fuzzy weight comprehensive evaluation system includes:

[0161] According to the implementation record of the historical load adjustment strategy, the fatigue state data of the local area of the blade before and after the implementation of the strategy is collected, and a plurality of evaluation indexes corresponding to the implementation effect of the load adjustment strategy are determined;

[0162] First, according to the implementation record of the historical load adjustment strategy, the fatigue state data (including fatigue cumulative value change rate, vibration amplitude change rate, and fatigue damage propagation rate) of the local area of the fan blade before and after the implementation of the strategy is collected, and these data are used as a plurality of evaluation indexes for evaluating the implementation effect of the load adjustment strategy.

[0163] A fuzzy evaluation factor set is constructed, which includes the fatigue damage propagation rate, the fatigue cumulative value change rate, and the vibration amplitude change rate of the local area of the blade;

[0164] The analytic hierarchy process is used to calculate the pairwise comparison matrix between each evaluation index, and the initial weight is modified after consistency check to obtain the final weight corresponding to each evaluation index;

[0165] It should be noted that the construction method of the pairwise comparison matrix in the analytic hierarchy process is as follows:

[0166] ;

[0167] For example, the pairwise comparison matrix given by the expert is (exemplary):

[0168] ;

[0169] The matrix is subjected to consistency test by eigenvalue method, and the consistency ratio (CR) is required to be ≤ 0.1, otherwise the matrix is modified and the weight is recalculated.

[0170] The membership degree corresponding to each evaluation index is determined based on the fuzzy membership function, and the maximum weight and the membership degree are subjected to fuzzy matrix operation to construct the multi-factor fuzzy weight comprehensive evaluation system.

[0171] It can be understood that the fuzzy membership function is selected as a trapezoidal or triangular fuzzy membership function.

[0172] Exemplarily, taking the trapezoidal fuzzy membership function of the fatigue accumulation value change rate as an example, it is defined as:

[0173] ;

[0174] wherein, is the fatigue accumulation value change rate (%), a, b, c, d are characteristic parameter values of the membership function; in this embodiment, after calculating the cumulative distribution function CDF of the fatigue accumulation value change rate index by statistically analyzing a large amount of historical load adjustment experimental data, the 5%, 25%, 75% and 95% quantiles of the CDF are taken as example values, which are 1%, 3%, 6% and 8% respectively, and the trapezoidal fuzzy membership function parameters are determined as: .

[0175] The priority coefficient corresponding to each candidate load adjustment strategy is determined according to the difference level, and the comprehensive priority score is obtained by weighted calculation of the priority coefficient and the corresponding strategy evaluation value, and the candidate load adjustment strategy with the highest comprehensive priority score is determined as the target load adjustment strategy.

[0176] Further, the priority coefficient corresponding to the difference level can be adjusted according to the blade safety requirement and the actual control effect, for example:

[0177] If the fan blade is in a working condition with low fatigue safety margin, the priority coefficient of the difference level can be appropriately increased to respond timely and quickly (such as the exemplary Table 1 below);

[0178] Table 1: Coefficient assignment table

[0179]

[0180] If the fatigue safety margin is high, the priority coefficient can be appropriately reduced to reduce the frequency of load adjustment action and prolong the service life of the system.

[0181] Exemplarily, the comprehensive priority score calculation formula is:

[0182] ;

[0183] In the formula: is the comprehensive priority score of the candidate strategy i; is the priority coefficient corresponding to the strategy i; is the strategy evaluation value corresponding to the candidate strategy i; denotes a normalization function, for example, Min-Max or Softmax normalization.

[0184] The candidate load adjustment strategy with the highest comprehensive priority score is determined as the target load adjustment strategy, thereby ensuring that the load adjustment strategy selection process is reasonable, efficient and accurate.

[0185] S104: According to the determined target load adjustment strategy, an active load control signal is generated through the fan active load control system to implement active adjustment of the local area load of the fan blade;

[0186] It should be noted that the purpose of this step is to generate accurate control instructions according to the optimal load adjustment strategy selected in the foregoing steps, so as to effectively adjust the load state of the local area of the fan blade in real time, thereby inhibiting the expansion of blade fatigue damage and prolonging the service life of the blade.

[0187] In implementation, the active load control signal generation method comprises:

[0188] According to the determined target load adjustment strategy, the load adjustment amplitude and the load adjustment direction of each local area of the blade;

[0189] It should be noted that the target load adjustment strategy is a precise load adjustment instruction set for different local areas of the blade, which specifically includes the following contents:

[0190] Load adjustment amplitude: the absolute change amount or the relative change percentage of the load of the local area of the blade to be adjusted, which is specifically expressed in units of N·m or %. In specific implementation, by establishing a dynamic mapping relationship between the overall load change of the global adjustment parameter (such as the angle of attack, the yaw angle, etc.) and the load change of each local area based on the local aerodynamic coefficient matrix obtained through CFD simulation calculation, the specific calculation formula is: ; wherein, is the overall load change amount caused by the adjustment amplitude of the angle of attack, the yaw angle, etc.; is the aerodynamic characteristic coefficient of the i-th local area of the blade, which is obtained by looking up the data table through CFD simulation calculation or wind tunnel test;

[0191] Load adjustment direction: both load increase and load decrease directions.

[0192] For example, the current load of a certain area of the fan blade close to the root area is , and the determined target load adjustment strategy indicates that it needs to be reduced by 5%. The load adjustment range is ; the load adjustment direction is the load reduction direction.

[0193] The specific calculation formula is as follows:

[0194] ;

[0195] In the formula: is the load adjustment range of the i-th local area of the blade; is the current load value of the i-th local area of the blade; is the load adjustment percentage of the i-th local area of the blade, and a positive value represents an increase in load, and a negative value represents a decrease in load.

[0196] It should be noted that the data table formed by the load adjustment instruction data of each local area is shown in Table 2 (exemplary) as follows:

[0197]

[0198] According to the current fan operating condition data, determine the execution timing of the target load adjustment strategy;

[0199] It should be noted that the operating condition data includes but is not limited to wind speed level, wind direction angle, fan operating speed, current blade angle position, load level, and operating environment temperature, etc.

[0200] In specific implementation, first, the above-mentioned operating condition data is collected, and then according to the real-time state of the current operating condition data and the requirements of the target load adjustment strategy, the specific execution timing of each load adjustment instruction is calculated.

[0201] The execution timing determination method is as follows:

[0202] Set the preset operating condition threshold, including wind speed threshold (such as rated wind speed), speed threshold (such as rated speed), blade position angle threshold;

[0203] According to the real-time collected operating condition data, determine and confirm that when the fan operating condition meets the preset threshold, it is the starting time of the load adjustment strategy execution;

[0204] It should be noted that, in order to avoid the difficulty of multi-condition coupling execution, the execution conditions meet the following priority order:

[0205] After the wind speed meets the threshold condition, it is executed preferentially;

[0206] If the wind speed does not meet the condition, but the rotation speed and the blade position both meet the condition, the method can also be executed.

[0207] If only the blade position meets the condition, the method is delayed to the next period for re-judgment.

[0208] If the delay time exceeds 30 minutes, no matter whether the working condition meets the threshold condition, the method is forced to be executed in the next sampling period, and the adjustment amplitude at the time of execution is reduced to 50% of the original value (to prevent sudden working condition impact).

[0209] Exemplarily, the fan working condition data table is as shown in Table 3 (example):

[0210]

[0211] After the working condition meets the condition, the time is marked as the load adjustment strategy execution time .

[0212] In addition, it should be noted that the duration of the load adjustment strategy is exemplarily set as: the starting time is determined, and the strategy implementation period is 10 minutes.

[0213] That is:

[0214] .

[0215] In the formula: is the current time, is the delay time determined according to the working condition data.

[0216] The determined load adjustment amplitude, load adjustment direction and execution time sequence are taken as inputs of the active load control system, and the active load control signal is obtained through real-time calculation of the active load control system.

[0217] It should be noted that the active load control system in the step is a real-time control system installed in the fan control cabinet, which specifically includes: a controller mainboard, a data interface module, a control algorithm module and a signal output module.

[0218] In specific implementation, the load adjustment amplitude, load adjustment direction and execution time sequence of each local area determined in the previous step are input into the controller mainboard through the data interface module of the active load control system in the data input format of

[0219] Specifically, the active load control system implements PID control algorithm calculation according to the control algorithm module of

[0220] .

[0221] In the formula: is the output value of the active load control signal, in units of voltage (V) or duty cycle (%); is the absolute difference between the load adjustment target value and the current actual load, in units of N·m, represents The error between the load adjustment target and the actual load at the moment is calculated as follows:

[0222] First, determine the target load value:

[0223] ;

[0224] Then calculate the error value:

[0225] ;

[0226] In the formula: is the load adjustment percentage, a positive value indicating an increase in load, a negative value indicating a decrease in load, is the current actual load; are the proportional coefficient, integral coefficient, and differential coefficient of the PID controller, respectively.

[0227] It should be noted that the initial setting value of the PID parameters (see Table 4 below for an example) is:

[0228]

[0229] In addition, the active load control signal output is in the form of a PWM pulse width modulation signal, with a PWM frequency of 1 kHz. The duty cycle (D) is calculated as follows:

[0230] ;

[0231] In the formula: is the maximum output control voltage, with a typical value of 10V; for example: if the calculation result is 5V, then the corresponding PWM duty cycle is 50%.

[0232] The active load control signal is sent to the active control actuator of the fan to drive the local area load of the fan blades to implement active adjustment;

[0233] It should be noted that the active control actuator in this step is an active control device installed at the root or inside of the fan blades, specifically: a hydraulic brake control device; or an electric active control device (such as a servo motor or an electric push rod).

[0234] In specific implementation, the active load control system outputs the active load control signal to the drive unit in the active control actuator through the PWM signal output interface.

[0235] It should be noted that the driving unit includes a signal receiving module (PWM signal input port) and a motor driving module (such as a PWM driver or a hydraulic pump controller).

[0236] Further, the driving unit converts the received PWM duty cycle signal into specific load adjustment execution actions, including execution amplitude (such as motor rotation angle or hydraulic cylinder piston movement displacement) and action execution speed.

[0237] The quantitative relationship of action execution (exemplary) is as follows:

[0238] When the PWM duty cycle is 50%, the displacement amplitude executed by the active control actuator is 50% of the rated displacement;

[0239] When the PWM duty cycle is 100%, the displacement amplitude executed by the active control actuator reaches the rated displacement.

[0240] Exemplary execution displacement and PWM duty cycle relationship is shown in the following Table 5:

[0241] Table 5: Execution displacement and PWM duty cycle relationship data table

[0242]

[0243] Finally, it should be noted that the active control actuator collects the displacement feedback signal of the actuator in real time when executing the load adjustment action, forming a closed loop control to ensure the accurate implementation of the local area load adjustment action of the fan blade, and real-time correction of the error in the load control process.

[0244] Embodiment 2

[0245] As Figure 2 shown, the part not detailed in this embodiment is shown in Embodiment 1, and this embodiment discloses a fan blade intelligent monitoring and optimization system based on deep learning, comprising:

[0246] The data prediction module 201 is configured to calculate the fatigue accumulation value corresponding to the strain data of each historical period according to the historical strain data of the local area of the fan blade, combine the fatigue accumulation rate calculation method, and train the fatigue state prediction model by taking the historical strain data as input and the fatigue accumulation value as output.

[0247] The strategy generation module 202 is configured to predict the current fatigue state of each local area according to the fatigue state prediction model, and generate a plurality of candidate load adjustment strategies by a pre-trained reinforcement learning model according to the current fatigue state.

[0248] The strategy screening module 203 is configured to collect real-time vibration data of the fan blade, determine a short-term trend change feature of fatigue of the fan blade, and screen a target load adjustment strategy from the plurality of candidate load adjustment strategies according to a difference between the short-term trend change feature and a current fatigue state.

[0249] The blade adjustment module 204 is configured to generate an active load control signal through the active load control system of the fan according to the determined target load adjustment strategy, so as to implement active adjustment on the local area load of the fan blade.

[0250] The above formulas are all dimensionless numerical calculations, the formulas are obtained by software simulation of a large amount of data to obtain a formula of the most recent real situation, and the preset parameters, weights and threshold values in the formulas are set by a person skilled in the art according to actual conditions.

[0251] The above embodiments can be realized wholly or partially by software, hardware, firmware or any other combination. When realized by software, the above embodiments can be realized wholly or partially in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the flow or function described in the embodiments of the present application is wholly or partially generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network or other programmable devices. The computer instructions can be stored in a computer readable storage medium or transferred from one computer readable storage medium to another, for example, the computer instructions can be transferred from one website, computer, server or data center to another website, computer, server or data center through a wired network or a wireless network. The computer readable storage medium can be any available medium accessible by a computer or a data storage device such as a server, data center and the like containing one or more available medium collections. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD) or a semiconductor medium. The semiconductor medium can be a solid state disk.

[0252] The above only describes some exemplary embodiments of the present application by way of illustration, and it is needless to say that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present application. Therefore, the above figures and descriptions are illustrative in nature and should not be understood as limiting the scope of protection of the claims of the present application.

Claims

1. A method for intelligent monitoring and optimization of a fan blade based on deep learning, characterized in that, The method comprises the following steps: S101: According to the historical strain data of the local area of the fan blade, the fatigue accumulation value corresponding to the strain data of each historical period is calculated by combining the fatigue accumulation rate calculation method, and a fatigue state prediction model is trained by taking the historical strain data as input and the fatigue accumulation value as output; The training method of the fatigue state prediction model comprises: Collecting historical strain data of the local area of the fan blade, segmenting the historical strain data according to a fixed time length, and obtaining strain data samples corresponding to multiple historical periods; According to the Miner fatigue damage criterion and the fatigue S-N curve of the fan blade material, the fatigue accumulation value corresponding to each historical period strain data sample is calculated respectively; A multi-scale convolution feature extraction structure is constructed, and convolution operation is performed on each historical period strain data sample to extract initial fatigue features corresponding to each scale; According to the correlation between fatigue features of different scales, a fusion weight is determined, and a fusion fatigue feature of each historical period strain data is obtained by fusion, and a fatigue state prediction model is trained by taking the fusion fatigue feature as input and the fatigue accumulation value as output; S102: According to the fatigue state prediction model, the current fatigue state of each local area is predicted, and a plurality of candidate load adjustment strategies are generated by a pre-trained reinforcement learning model according to the current fatigue state; S103: Collecting real-time vibration data of the fan blade, determining the short-term trend change characteristics of the fan blade fatigue, and selecting a target load adjustment strategy from the plurality of candidate load adjustment strategies according to the difference between the short-term trend change characteristics and the current fatigue state; S104: According to the determined target load adjustment strategy, an active load control signal is generated through the fan active load control system to implement active adjustment of the local area load of the fan blade.

2. The deep learning based intelligent monitoring and optimization method for wind turbine blades according to claim 1, characterized in that, The specific construction method of the multi-scale convolution feature extraction structure comprises: A plurality of parallel convolution branches with different convolution kernel sizes are constructed respectively, and a dilated convolution layer, a depth separable convolution layer and an attention mechanism module are arranged in each parallel convolution branch; In each parallel convolution branch, the initial convolution feature of the corresponding scale is obtained through the dilated convolution layer, and the initial convolution feature is further convolved by using the depth separable convolution layer to obtain the refined fatigue feature of the corresponding scale; The attention weights of the feature channels are calculated by the attention mechanism module for the refined fatigue features obtained in each parallel convolution branch, and each channel of the refined fatigue feature is weighted and adjusted according to the attention weight to obtain the attention feature of each convolution branch.

3. The deep learning based intelligent monitoring and optimization method for wind turbine blades according to claim 2, characterized in that, The method for determining the fusion weight according to the correlation between fatigue features of different scales and obtaining the fusion fatigue feature of each historical period strain data comprises: According to the correlation between the fatigue features of different scales, the corresponding correlation coefficients are calculated respectively; According to the numerical value of the calculated correlation coefficient, the fusion weight corresponding to each scale fatigue feature is determined; According to the determined fusion weight, the fatigue features of different scales are weighted and fused respectively to obtain the fusion fatigue feature of each historical period strain data.

4. The deep learning based intelligent monitoring and optimization method for wind turbine blades according to claim 3, characterized in that, The training method of the reinforcement learning model comprises: According to the historical fatigue state data of the local area of the fan blade, the fatigue cumulative value changes of each local area before and after the implementation of the historical load adjustment strategy are determined, and a state space and an action space of a reinforcement learning model are constructed; Based on the amplitude of the fatigue cumulative value changes, the fatigue damage expansion rate of each local area is determined, and a reward function of the reinforcement learning model is constructed; The state space, the action space and the reward function constructed are taken as inputs of the reinforcement learning model, a deep reinforcement learning algorithm is used to train the reinforcement learning model, and a pre-trained reinforcement learning model for generating a candidate load adjustment strategy in real time is obtained.

5. The deep learning based intelligent monitoring and optimization method for wind turbine blades according to claim 4, characterized in that, The method for constructing the reward function of the reinforcement learning model comprises: According to the fatigue cumulative values of each local area before and after the implementation of the historical load adjustment strategy, the change amplitude of the fatigue cumulative rate after the implementation of the load adjustment strategy relative to that before the implementation is calculated, and the change amplitude is taken as a basic reward value of each local area; According to the difference between the current fatigue cumulative value of each local area and a set fatigue safety threshold, a fatigue sensitivity coefficient corresponding to each local area is calculated; The basic reward value and the fatigue sensitivity coefficient are weighted one by one to obtain a comprehensive reward value, and the comprehensive reward value is used to construct the reward function of the reinforcement learning model.

6. The deep learning based intelligent monitoring and optimization method for wind turbine blades according to claim 5, characterized in that, The method for determining the fatigue sensitivity coefficient comprises: According to the difference between the fatigue limit value of the material of the local area of the fan blade and the current fatigue cumulative value, a fatigue margin corresponding to each local area is determined; According to the fatigue margin, an initial sensitivity coefficient corresponding to each local area is calculated; A dynamic amplification factor of the fatigue state is determined according to the wind speed grade and the load grade during the operation of the blade; The dynamic amplification factor and the initial sensitivity coefficient are multiplied one by one to obtain the fatigue sensitivity coefficient of each local area.

7. The deep learning based intelligent monitoring and optimization method of wind turbine blades according to claim 6, characterized in that, The method for screening and determining the target load adjustment strategy comprises: Real-time vibration data of the fan blade are collected, and short-term trend change characteristics of the fatigue of the fan blade are determined according to the real-time vibration data; According to the numerical difference between the short-term trend change characteristics and the current fatigue state predicted by the fatigue state prediction model, a difference grade is determined; For each candidate load adjustment strategy, the implementation effect of the candidate load adjustment strategy is evaluated according to a pre-constructed multi-factor fuzzy weight comprehensive evaluation system, and a strategy evaluation value corresponding to each candidate strategy is obtained; According to the difference grade, a priority coefficient corresponding to each candidate load adjustment strategy is determined, and the priority coefficient and the corresponding strategy evaluation value are weighted to obtain a comprehensive priority score, and the candidate load adjustment strategy with the highest comprehensive priority score is determined as the target load adjustment strategy.

8. The deep learning based intelligent monitoring and optimization method for wind turbine blades according to claim 7, characterized in that, The method for determining the short-term trend change characteristics of the fatigue of the fan blade specifically comprises: Real-time vibration data of the local area of the fan blade are collected, and corresponding frequency domain characteristic parameters are obtained through a fast Fourier transform method; According to the real-time vibration characteristic value and the historical vibration characteristic value of the local area of the fan blade, a vibration characteristic value change rate in the current period is calculated, and the vibration characteristic value change rate is determined as the short-term trend change characteristics of the fatigue of the fan blade.

9. The deep learning based intelligent monitoring and optimization method for wind turbine blades according to claim 8, characterized in that, The specific construction method of the multi-factor fuzzy weight comprehensive evaluation system comprises: According to the implementation record of the historical load adjustment strategy, fatigue state data of the local area of the blade before and after the implementation of the strategy is collected, and a plurality of evaluation indexes corresponding to the implementation effect of the load adjustment strategy are determined; A fuzzy evaluation factor set is constructed, and the fuzzy evaluation factor set includes a fatigue damage propagation rate, a fatigue cumulative value change rate, and a vibration amplitude change rate of the local area of the blade; Initial weights of the evaluation indexes are determined according to an expert scoring method, a two-by-two comparison matrix between the evaluation indexes is calculated by using an analytic hierarchy process, the initial weights are corrected after consistency checking, and final weights corresponding to the evaluation indexes are obtained; Membership degrees corresponding to the evaluation indexes are determined based on a fuzzy membership function, the final weights and the membership degrees are subjected to fuzzy matrix operation, and a multi-factor fuzzy weight comprehensive evaluation system is constructed.

Images