Method and system for monitoring wind-induced vibration of wind driven generator tower

Abstract

The invention discloses a method and a system for monitoring wind-induced vibration of a tower drum of a wind driven generator. The method comprises the following steps: acquiring data by using a multi-mode sensor; performing numerical simulation on the acquired data to obtain a training data set; constructing an initial wind-induced vibration monitoring model of the wind driven generator tower; training the initial wind-induced vibration monitoring model of the wind driven generator tower tube by adopting the training data set to obtain a wind-induced vibration monitoring model of the wind driven generator tower tube; and carrying out actual wind-induced vibration monitoring on the wind driven generator tower tube by using the wind-induced vibration monitoring model of the wind driven generator tower tube. According to the method, through multi-sensor cooperation, cross-modal data fusion, physical mechanism constraint and edge lightweight design, the operation state of the tower drum is reflected from multiple dimensions, the accuracy of fault diagnosis is greatly improved, and high-precision monitoring of wind power tower drum bolt looseness is achieved.

Description

Technical Field

[0001] The present invention belongs to the field of power generation facility monitoring, and in particular relates to a method and system for monitoring wind-induced vibration of a wind turbine tower. Background Art

[0002] Amidst the global energy transition, wind power, as a key component of clean, renewable energy, has experienced rapid growth. With the expansion of the wind power industry, wind turbines are becoming larger and more scalable, and their operating environments are becoming increasingly complex and diverse, encompassing strong winds along the coast and complex inland terrain at high altitudes. This development has led to increasingly prominent vibration and fatigue issues facing wind turbine towers.

[0003] During operation, wind turbine towers are constantly exposed to natural factors such as strong winds and turbulent airflow, which can easily cause vibration. Long-term exposure to this vibration can gradually accumulate fatigue damage in the tower structure, compromising its structural integrity and stability. In severe cases, this can lead to safety accidents such as tower cracking and collapse, resulting in significant economic losses and potentially endangering personnel safety.

[0004] During daily operation, the tower's swaying vibration can cause the bolts at the flange connection to loosen, a major cause of tower instability. Therefore, accurate and effective monitoring of wind-induced vibrations on wind turbine towers and real-time detection and early warning of flange bolt loosening are critical components in ensuring the safe and stable operation of wind power generation systems.

[0005] In the prior art, tower vibration monitoring devices mainly use vibration sensors, stress strain gauges and other technical means, but these technologies have the following deficiencies and defects in practical applications.

[0006] First, regarding vibration sensors, most traditional vibration sensors can only measure a single physical quantity, such as acceleration, velocity, or displacement. Taking the common piezoelectric vibration sensor as an example, although it is relatively sensitive to vibration, in the complex wind power generation environment, when affected by factors such as temperature changes and electromagnetic interference, the measurement accuracy will show significant deviations. Moreover, a single vibration measurement cannot fully reflect the stress and strain state within the tower structure, making it difficult to accurately assess the actual damage caused by vibration to the tower structure. In addition, the frequency response range of some vibration sensors is limited, and the broadband vibration signals generated by the wind turbine tower under different operating conditions cannot be fully collected and analyzed. This leads to the omission of key information and affects the accurate judgment of the tower vibration state.

[0007] Secondly, regarding stress and strain gauges, the commonly used resistance strain gauges present a complex installation process when monitoring tower stress and strain. They need to be precisely affixed to the bolt surface, placing extremely high demands on the placement and process. Any inaccuracy can affect measurement accuracy, making the current solution extremely inconvenient to install and maintain, seriously impacting the feasibility of using strain gauges to monitor bolt loosening. Furthermore, resistance strain gauges are susceptible to environmental humidity. In humid coastal wind farms, moisture can easily penetrate the strain gauge, changing its resistance and increasing measurement errors. Furthermore, traditional stress and strain gauges can experience zero-point drift over long-term use, significantly compromising the reliability of the measurement data and preventing them from providing stable and accurate data support for tower structural health assessments.

[0008] Finally, regarding data fusion and analysis, existing monitoring solutions, even when using both vibration sensors and stress and strain gauges, often simply record the data from both, lacking effective data fusion algorithms. This makes it difficult to fully explore the inherent connections between vibration and stress and strain data, making it difficult to accurately assess the overall structural health of the tower. Furthermore, existing data analysis methods mostly rely on simple time-domain analysis, insufficiently exploring complex frequency-domain characteristics and modal information, and unable to promptly identify potential vibration risks and structural hazards in the tower.

[0009] In summary, existing monitoring technologies are difficult to meet the needs of comprehensive and accurate monitoring of wind-induced vibrations of wind turbine towers. An innovative monitoring method and system are urgently needed to solve these problems. Summary of the Invention

[0010] In order to address the shortcomings of the existing technology, one of the purposes of the present invention is to provide a method for monitoring wind-induced vibration of a wind turbine tower. Through an innovative multi-sensor collaborative working mode and advanced data fusion and analysis methods, comprehensive and accurate monitoring of tower vibration and deformation, and whether the tower flange bolts are loose can be achieved, thereby effectively improving the reliability of tower structure health assessment and ensuring the safe and stable operation of wind turbines.

[0011] A second object of the present invention is to provide a system for implementing the wind-induced vibration monitoring method of a wind turbine tower.

[0012] The present invention provides a method for monitoring wind-induced vibration of a wind turbine tower, comprising the following steps:

[0013] S1. Collect data using multimodal sensors;

[0014] S2. Perform numerical simulation on the data collected in step S1 to obtain a training data set;

[0015] S3. Construct an initial wind turbine tower wind-induced vibration monitoring model;

[0016] S4. The initial wind turbine tower wind-induced vibration monitoring model obtained in step S4 is trained using the training data set obtained in step S2 to obtain a wind turbine tower wind-induced vibration monitoring model;

[0017] S5. Using the wind turbine tower wind-induced vibration monitoring model obtained in step S4, actual wind turbine tower wind-induced vibration monitoring is performed.

[0018] In step S1, the multimodal sensor includes: vibration sensors installed at key nodes of the tower, and strain gauges installed at stress concentration areas and connection parts of the tower structure;

[0019] The data acquired by the vibration sensor is a vibration sensor signal, which is expressed using the following formula: Where t is the time series, R is the real space domain of the feature vector, V is the original signal of the sensor, C V is the number of vibration sensor channels; T is the sample data cutoff timing; v t is the instantaneous vibration velocity at time t;

[0020] The data obtained by the strain gauge is the strain gauge signal, which is expressed using the following formula: Among them, S is the original strain signal, C S is the number of strain gauge channels; s t is the instantaneous strain at time t.

[0021] In step S3, the initial wind turbine tower wind-induced vibration monitoring model includes a preprocessing module, a feature extraction module, a feature fusion module, a fully connected layer module, and a monitoring result output module;

[0022] The initial wind turbine tower wind-induced vibration monitoring model preprocesses the input vibration signal data and strain signal data preprocessing module. After preprocessing, features are extracted through the feature extraction module, and then the extracted features are cross-modally fused through the feature fusion module. Finally, the fusion result is processed through the fully connected layer as the input of the monitoring result output module. The monitoring result output module calculates according to the input data to obtain the final output of the model.

[0023] The preprocessing module includes a dynamic wavelet noise reduction module and a working condition adaptive normalization module; the wind-borne high-frequency noise in the data is suppressed by dynamic wavelet noise reduction, and then working condition adaptive normalization is performed according to wind speed intervals to complete the data preprocessing.

[0024] The dynamic wavelet denoising module is expressed using the following formula:

[0025]

[0026] in, is the acceleration signal of the vibration sensor at time t; SWT is dynamic wavelet denoising; is the wavelet basis function; λ is the scale-related threshold; σ is the standard deviation of the noise component; is the Haar wavelet transform basis function; is the strain acceleration signal of the strain gauge at time t;

[0027] The working condition adaptive normalization module is expressed by the following formula:

[0028]

[0029] in, for Normalized value; for Normalized value; w is the current wind speed range; μ v (w) is the mean value of the vibration signal under the current working condition; σ v (w) is the mean square error of the vibration signal under the current working condition; μ s (w) is the mean value of the strain signal under the current working condition; σ s (w) is the mean square error of the strain signal under the current working condition.

[0030] The feature extraction module includes a vibration signal feature extraction module and a strain signal feature extraction module;

[0031] The vibration signal feature extraction module first uses a multi-scale convolution kernel group to extract features in the frequency dimension, and then enhances the resonance frequency band response through frequency attention pooling to obtain vibration signal features. The feature extraction of the multi-scale convolution kernel group in the frequency dimension is expressed by the following formula:

[0032] in, is the feature representation of the lth layer in the frequency dimension; k is the index of the frequency channel, and its value range is 1 to K; K is the total number of frequency channels; l is the convolution kernel level; is the one-dimensional convolution kernel corresponding to the k-th frequency channel in the l-th layer;

[0033] The frequency attention pooling is expressed using the following formula:

[0034] Among them, P v is the output result after frequency attention pooling; To perform the maximum pooling operation on the value of the kth frequency channel of the lth layer feature map; a k is a parameter, and q is the query vector to be learned; is the value of the l-th layer feature map at the k-th frequency channel;

[0035] The strain signal feature extraction module first converts the input preprocessed strain signal data into preload data through the physical encoding of the bolt preload based on a temperature-compensated linear model. It then extracts time series features from the preload data using LSTM to obtain strain signal features.

[0036] The physical coding of the bolt preload force is expressed using the following formula:

[0037] Among them, F t is the preload at time t; k(τ) is the elastic modulus at temperature τ; b(τ) is the constant at temperature τ;

[0038] The time series feature extraction of preload force data based on LSTM is expressed by the following formula:

[0039] in, is the output hidden state of the long short-term memory network (LSTM) at time step t; LSTM is the long short-term memory network; H s It is the feature set of the entire time series data after being processed by the LSTM network.

[0040] The feature fusion module specifically uses spatiotemporal alignment and attention fusion to fuse the input features across modalities, which can be expressed by the following formula:

[0041] in, is element-by-element addition; H fused is the fused feature; A is the attention weight matrix, and A is expressed using the following formula: d is the dimension of the feature vector.

[0042] The fully connected layer module is expressed using the following formula:

[0043]

[0044] Among them, H fused is the fused feature; W fc is the weight matrix to be learned; b fc is the bias vector to be learned; is the output vector, z0 is the healthy category score; z1 is the loose category score.

[0045] The output structure of the monitoring result output module includes the bolt health status judgment result, loosening probability, vibration severity index and preload estimation value;

[0046] The calculation formula of the loosening probability is: Among them, p is the loose probability; z0 is the healthy category score output by the fully connected layer module; z1 is the loose category score output by the fully connected layer module;

[0047] The calculation formula of the vibration severity index is: Among them, I v is the vibration severity index; RMS v RMS is the root mean square amplitude of the tower vibration; max is the maximum RMS value of the amplitude corresponding to the wind condition of force 12; FER is the amplitude coefficient corresponding to the wind condition of force 12;

[0048] The calculation formula of the estimated preload force is: in, is the estimated value of preload force; MLP() is multi-layer perceptron processing; H fused is the fused feature;

[0049] The judgment logic of the bolt health status judgment result is: If the loose probability p is greater than or equal to 0.9, it is determined to be loose, otherwise it is determined to be normal.

[0050] In step S4, the total loss function of the training is expressed by the following formula:

[0051] Among them, ξ is the total loss function; is the classification cross entropy loss function; ξ vib is the loss term related to vibration; ξ bolt is the loss item related to the bolts;

[0052] The vibration-related loss term ξ vib Based on the Newton-Euler equation, it is expressed as follows:

[0053] in, is the tower vibration amplitude; m is the first tower parameter; c is the second tower parameter; k is the third tower parameter; is the wind pressure on the windward side of the tower; λ1 is the Lagrange multiplier of the corresponding model;

[0054] The loss term ξ associated with the bolt bolt In order to generate the strain-preload nonlinear relationship library through finite element simulation, the following formula is used:

[0055] Among them, F i is the strain and preload data pair; λ2 is the Lagrange multiplier of the corresponding model; is the bolt preload value at time t predicted by the model; N is the total number of data in the strain-preload relationship library; interp() is the interpolation function, which is used to calculate the bolt preload value according to the given current strain value ε. t , in the strain-preload nonlinear relationship library {ε i ,F i} to perform interpolation calculation to obtain the corresponding preload force estimation value;

[0056] Categorical cross entropy loss function ξ CE , expressed using the following formula:

[0057]

[0058] Where y is the sample set: y=(y1,y2,...y C ); only the true category k satisfies y k =1, the rest y j =0(j≠k); In model prediction, the probability distribution of sample output is satisfy

[0059] Furthermore, the method of the present invention can also perform lightweight deployment on the wind turbine tower wind-induced vibration monitoring model, and then use the lightweight wind turbine tower wind-induced vibration monitoring model for monitoring; the lightweight deployment includes model compression and edge inference optimization;

[0060] The model compression includes knowledge distillation and dynamic quantization; the knowledge distillation uses the wind turbine tower wind-induced vibration monitoring model as a teacher network to guide the lightweight student network;

[0061] The total loss function of the knowledge distillation is expressed as follows:

[0062] Among them, ξ KD is the total loss function of knowledge distillation; α is a hyperparameter, and its value range is between 0 and 1; ξ MSE is the mean square error loss function; is the feature vector output by the teacher network at layer s; is the feature vector output by the student network at layer s;

[0063] The dynamic quantization maps the weights to 8-bit fixed-point numbers, which are expressed by the following formula:

[0064] Among them, w quant is the quantized weight value; w is the original weight value; w max is the maximum value in the weight set that needs to be quantized; w minis the minimum value in the weight set that needs to be quantized; round() is the rounding function;

[0065] The edge inference optimization is specifically to transform the fusion layer H fused Decompose into subgraphs that can be computed in parallel, adapt to embedded GPUs, and then process them from aspects such as task scheduling, data transmission, computing resource allocation, result integration, and performance optimization;

[0066] The task scheduling is to arrange the computing tasks of each subgraph by utilizing the multi-core architecture of the embedded GPU;

[0067] The data transmission is to optimize the data transmission process between subgraphs;

[0068] The computing resource allocation is to dynamically allocate computing resources of the embedded GPU, such as video memory and computing units, according to the needs of the subgraph;

[0069] The result integration is to integrate the results of each subgraph after the calculation is completed;

[0070] The performance monitoring and optimization is to monitor the performance indicators of the embedded GPU in real time during the calculation process, such as utilization rate, temperature, etc.

[0071] The memory is pre-allocated as a fixed input and output buffer to reduce dynamic memory overhead.

[0072] The present invention also provides a wind turbine tower wind-induced vibration monitoring system, the system comprising a multimodal sensor data acquisition module, a data simulation module, a model construction module, a model training module, and a wind turbine tower wind-induced vibration monitoring module;

[0073] The multimodal sensor data acquisition module uses the multimodal sensor to collect data and uploads the data to the data simulation module;

[0074] The data simulation module performs numerical simulation on the collected data based on the received data to obtain the training data set and uploads the data to the model training module;

[0075] The model building module builds the initial wind turbine tower wind-induced vibration monitoring model and uploads the data to the model training module;

[0076] The model training module trains the initial wind turbine tower wind-induced vibration monitoring model using the training data set based on the received data to obtain the wind turbine tower wind-induced vibration monitoring model, and uploads the data to the wind turbine tower wind-induced vibration monitoring module;

[0077] The wind turbine tower wind-induced vibration monitoring module performs actual wind turbine tower wind-induced vibration monitoring based on the received data and using the wind turbine tower wind-induced vibration monitoring model.

[0078] The present invention discloses a method and system for monitoring wind-induced vibration of a wind turbine tower. Through multi-sensor collaboration and cross-modal data fusion, physical mechanism constraints and edge lightweight design, the system reflects the tower operating status from multiple dimensions, greatly improves the accuracy of fault diagnosis, and realizes high-precision monitoring of wind turbine tower bolt loosening. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 Schematic diagram of the process of the present invention;

[0080] Figure 2 Schematic diagram of the system structure of the present invention;

[0081] Figure 3 This is a schematic diagram of the installation of a monitoring device in an embodiment of the present invention; the specific symbols in the figure are: 1. Monitoring system; 2. Vibration sensor; 3. Strain gauge; 4. Flange; 5. Tower. DETAILED DESCRIPTION

[0082] The present invention provides a method for monitoring wind-induced vibration of a wind turbine tower, the flow chart of which is as follows: Figure 1 As shown, the following steps are included:

[0083] S1. Collect data using multimodal sensors;

[0084] In step S1, the multimodal sensor includes: vibration sensors installed at key nodes of the tower, and strain gauges installed at stress concentration areas and connection parts of the tower structure;

[0085] The data acquired by the vibration sensor is a vibration sensor signal, which is expressed using the following formula: Where t is the time series, R is the real space domain of the feature vector, V is the original signal of the sensor, C V is the number of vibration sensor channels; T is the sample data cutoff timing; v t is the instantaneous vibration velocity at time t;

[0086] The data obtained by the flange strain gauge is the flange strain gauge signal, which is expressed using the following formula: Among them, S is the original strain signal, C S is the number of strain gauge channels; s t is the instantaneous strain at time t.

[0087] S2. Perform numerical simulation on the data collected in step S1 to obtain a training data set;

[0088] S3. Construct an initial wind turbine tower wind-induced vibration monitoring model;

[0089] In step S3, the initial wind turbine tower wind-induced vibration monitoring model includes a preprocessing module, a feature extraction module, a feature fusion module, a fully connected layer module, and a monitoring result output module;

[0090] The initial wind turbine tower wind-induced vibration monitoring model preprocesses the input vibration signal data and strain signal data preprocessing module. After preprocessing, features are extracted through the feature extraction module, and then the extracted features are cross-modally fused through the feature fusion module. Finally, the fusion result is processed through the fully connected layer as the input of the monitoring result output module. The monitoring result output module calculates according to the input data to obtain the final output of the model.

[0091] The preprocessing module includes a dynamic wavelet noise reduction module and a working condition adaptive normalization module; the wind-borne high-frequency noise in the data is suppressed by dynamic wavelet noise reduction, and then working condition adaptive normalization is performed according to wind speed intervals to complete the data preprocessing.

[0092] The dynamic wavelet denoising module is expressed using the following formula:

[0093]

[0094] in, is the acceleration signal of the vibration sensor at time t; SWT is dynamic wavelet denoising; is the wavelet basis function; λ is the scale-related threshold; σ is the standard deviation of the noise component; is the Haar wavelet transform basis function; is the instantaneous strain signal of the strain gauge at time t;

[0095] The working condition adaptive normalization module is expressed by the following formula:

[0096]

[0097] in, for Normalized value; for Normalized value; w is the current wind speed range; μ v (w) is the mean value of the vibration signal under the current working condition; σ v (w) is the mean square error of the vibration signal under the current working condition; μ s (w) is the mean value of the strain signal under the current working condition; σ s (w) is the mean square error of the strain signal under the current working condition.

[0098] The feature extraction module includes a vibration signal feature extraction module and a strain signal feature extraction module;

[0099] The vibration signal feature extraction module first uses a multi-scale convolution kernel group to extract features in the frequency dimension, and then enhances the resonance frequency band response through frequency attention pooling to obtain vibration signal features. The feature extraction of the multi-scale convolution kernel group in the frequency dimension is expressed by the following formula:

[0100] in, is the feature representation of the lth layer in the frequency dimension; k is the index of the frequency channel, and its value range is 1 to K; K is the total number of frequency channels; l is the convolution kernel level; is the one-dimensional convolution kernel corresponding to the k-th frequency channel in the l-th layer;

[0101] The frequency attention pooling is expressed using the following formula:

[0102] Among them, P v is the output result after frequency attention pooling; To perform the maximum pooling operation on the value of the kth frequency channel of the lth layer feature map; a k is a parameter, and q is the query vector to be learned; is the value of the l-th layer feature map at the k-th frequency channel;

[0103] The strain signal feature extraction module first converts the input preprocessed strain signal data into preload data through the physical encoding of the bolt preload based on a temperature-compensated linear model. It then extracts time series features from the preload data using LSTM to obtain strain signal features.

[0104] The physical coding of the bolt preload force is expressed using the following formula:

[0105] Among them, F t is the preload at time t; k(τ) is the elastic modulus at temperature τ; b(τ) is the constant at temperature τ;

[0106] The time series feature extraction of preload force data based on LSTM is expressed by the following formula:

[0107] in, is the output hidden state of the long short-term memory network (LSTM) at time step t; LSTM is the long short-term memory network; H s It is the feature set of the entire time series data after being processed by the LSTM network.

[0108] The feature fusion module specifically uses spatiotemporal alignment and attention fusion to fuse the input features across modalities, which can be expressed by the following formula:

[0109] in, is element-by-element addition; H fused is the fused feature; A is the attention weight matrix, and A is expressed using the following formula: d is the dimension of the feature vector.

[0110] The fully connected layer module is expressed using the following formula:

[0111]

[0112] Among them, H fused is the fused feature; W fc is the weight matrix to be learned; b fc is the bias vector to be learned; is the output vector, z0 is the healthy category score; z1 is the loose category score.

[0113] The output structure of the monitoring result output module includes the bolt health status judgment result, loosening probability, vibration severity index and preload estimation value;

[0114] The calculation formula of the loosening probability is: Among them, p is the loose probability; z0 is the healthy category score output by the fully connected layer module; z1 is the loose category score output by the fully connected layer module;

[0115] The calculation formula of the vibration severity index is: Among them, I v is the vibration severity index; RMS v RMS is the root mean square amplitude of the tower vibration; max is the maximum RMS value of the amplitude corresponding to the wind condition of force 12; FER is the amplitude coefficient corresponding to the wind condition of force 12;

[0116] The calculation formula of the estimated preload force is: in, is the estimated value of preload force; MLP() is multi-layer perceptron processing; H fused is the fused feature;

[0117] The judgment logic of the bolt health status judgment result is: If the loose probability p is greater than or equal to 0.9, it is determined to be loose, otherwise it is determined to be normal.

[0118] S4. The initial wind turbine tower wind-induced vibration monitoring model obtained in step S4 is trained using the training data set obtained in step S2 to obtain a wind turbine tower wind-induced vibration monitoring model;

[0119] In step S4, the total loss function of the training is expressed by the following formula:

[0120] Among them, ξ is the total loss function; is the classification cross entropy loss function; ξ vib is the loss term related to vibration; ξ bolt is the loss item related to the bolts;

[0121] The vibration-related loss term ξ vib Based on the Newton-Euler equation, it is expressed as follows:

[0122] in, is the tower vibration amplitude; m is the first tower parameter; c is the second tower parameter; k is the third tower parameter; is the wind pressure on the windward side of the tower; λ1 is the Lagrange multiplier of the corresponding model;

[0123] The loss term ξ associated with the bolt bolt In order to generate the strain-preload nonlinear relationship library through finite element simulation, the following formula is used:

[0124] Among them, F i is the strain and preload data pair; λ2 is the Lagrange multiplier of the corresponding model; is the bolt preload value at time t predicted by the model; N is the total number of data in the strain-preload relationship library; interp() is the interpolation function, which is used to calculate the bolt preload value according to the given current strain value ε. t , in the strain-preload nonlinear relationship library {ε i ,F i} to perform interpolation calculation to obtain the corresponding preload force estimation value;

[0125] Categorical cross entropy loss function ξ CE , expressed using the following formula:

[0126]

[0127] Where y is the sample set: y=(y1,y2,...y C ); only the true category k satisfies y k =1, the rest y j =0(j≠k); In model prediction, the probability distribution of sample output is satisfy

[0128]

[0129] S5. Using the wind turbine tower wind-induced vibration monitoring model obtained in step S4, actual wind turbine tower wind-induced vibration monitoring is performed.

[0130] Furthermore, the method of the present invention can also perform lightweight deployment on the wind turbine tower wind-induced vibration monitoring model, and then use the lightweight wind turbine tower wind-induced vibration monitoring model for monitoring; the lightweight deployment includes model compression and edge inference optimization;

[0131] The model compression includes knowledge distillation and dynamic quantization; the knowledge distillation uses the wind turbine tower wind-induced vibration monitoring model as a teacher network to guide the lightweight student network;

[0132] The total loss function of the knowledge distillation is expressed as follows:

[0133] Among them, ξ KD is the total loss function of knowledge distillation; α is a hyperparameter, and its value range is between 0 and 1; ξ MSE is the mean square error loss function; is the feature vector output by the teacher network at layer s; is the feature vector output by the student network at layer s;

[0134] The dynamic quantization maps the weights to 8-bit fixed-point numbers, which are expressed by the following formula:

[0135] Among them, w quant is the quantized weight value; w is the original weight value; w max is the maximum value in the weight set that needs to be quantized; w min is the minimum value in the weight set that needs to be quantized; round() is the rounding function;

[0136] The edge inference optimization is specifically to transform the fusion layer H fused Decompose into subgraphs that can be computed in parallel, adapt to embedded GPUs, and then process them from aspects such as task scheduling, data transmission, computing resource allocation, result integration, and performance optimization;

[0137] The task scheduling is to arrange the computing tasks of each subgraph by utilizing the multi-core architecture of the embedded GPU;

[0138] The data transmission is to optimize the data transmission process between subgraphs;

[0139] The computing resource allocation is to dynamically allocate computing resources of the embedded GPU, such as video memory and computing units, according to the needs of the subgraph;

[0140] The result integration is to integrate the results of each subgraph after the calculation is completed;

[0141] The performance monitoring and optimization is to monitor the performance indicators of the embedded GPU in real time during the calculation process, such as utilization rate, temperature, etc.

[0142] The memory is pre-allocated as a fixed input and output buffer to reduce dynamic memory overhead.

[0143] The present invention also provides a wind turbine tower wind-induced vibration monitoring system, the structural diagram of which is shown in FIG. Figure 2 As shown, the system includes a multimodal sensor data acquisition module, a data simulation module, a model building module, a model training module, and a wind turbine tower wind-induced vibration monitoring module;

[0144] The multimodal sensor data acquisition module uses the multimodal sensor to collect data and uploads the data to the data simulation module;

[0145] The data simulation module performs numerical simulation on the collected data based on the received data to obtain the training data set and uploads the data to the model training module;

[0146] The model building module builds the initial wind turbine tower wind-induced vibration monitoring model and uploads the data to the model training module;

[0147] The model training module trains the initial wind turbine tower wind-induced vibration monitoring model using the training data set based on the received data to obtain the wind turbine tower wind-induced vibration monitoring model, and uploads the data to the wind turbine tower wind-induced vibration monitoring module;

[0148] The wind turbine tower wind-induced vibration monitoring module performs actual wind turbine tower wind-induced vibration monitoring based on the received data and using the wind turbine tower wind-induced vibration monitoring model.

[0149] The method of the present invention is further described below with reference to an embodiment:

[0150] In a certain wind farm with 3MW units and a sampling rate of 1kHz, data is collected for 12 consecutive months. The monitoring device is set up as follows: Figure 3 shown.

[0151] By artificially inducing bolt loosening to set a fault sample, monitoring was performed using the wind turbine tower wind-induced vibration monitoring model in the method of the present invention, the lightweight wind turbine tower wind-induced vibration monitoring model in the method of the present invention, the traditional SVM model, and the single-mode CNN model. The results are shown in the following table:

[0152] Table 1 Comparison of monitoring performance between the method of the present invention and the prior art

[0153]

[0154] From the results in Table 1, it can be seen that the wind-induced vibration monitoring model for wind turbine towers disclosed in the present invention and the lightweight wind turbine tower wind-induced vibration monitoring model have higher accuracy and better inference delay effect.

[0155] Testing of monitoring performance under extreme conditions also revealed that the proposed method maintained an accuracy rate of over 91% even in typhoon conditions (wind speed 30 m / s), while the accuracy of a traditional single-modal CNN model was only 68%. Furthermore, in low-temperature conditions (minus 30 degrees Celsius), the proposed method, incorporating a temperature compensation model, reduced the false alarm rate to approximately 7%.

Claims

1. A method for monitoring wind-induced vibration of a wind turbine tower, characterized in that: The following steps are involved: S1. Collect data using multimodal sensors; S2. Perform numerical simulation on the data collected in step S1 to obtain a training data set; S3. Construct an initial wind turbine tower wind-induced vibration monitoring model; S4. The initial wind turbine tower wind-induced vibration monitoring model obtained in step S4 is trained using the training data set obtained in step S2 to obtain a wind turbine tower wind-induced vibration monitoring model; S5. Using the wind turbine tower wind-induced vibration monitoring model obtained in step S4, actual wind turbine tower wind-induced vibration monitoring is performed.

2. The method for monitoring wind-induced vibration of a wind turbine tower according to claim 1, wherein: In step S1, the multimodal sensor includes: vibration sensors installed at key nodes of the tower, and strain gauges installed at stress concentration areas and connection parts of the tower structure; the data obtained by the vibration sensor is a vibration sensor signal, which is expressed using the following formula: C V ={acceleration a, displacement x, ...}, where t is the time series, R is the real space domain of the eigenvector, V is the original signal of the sensor, and C V is the number of vibration sensor channels; T is the sample data cutoff timing; v t is the instantaneous vibration velocity at time t; The data obtained by the strain gauge is the strain gauge signal, which is expressed using the following formula: C S ={strain ε, temperature τ, ...}, where S is the original strain signal, C S is the number of strain gauge channels; s t is the instantaneous strain at time t.

3. The method for monitoring wind-induced vibration of a wind turbine tower according to claim 1, wherein: In step S3, the initial wind turbine tower wind-induced vibration monitoring model includes a preprocessing module, a feature extraction module, a feature fusion module, a fully connected layer module, and a monitoring result output module; The initial wind turbine tower wind-induced vibration monitoring model preprocesses the input vibration signal data and strain signal data preprocessing module. After preprocessing, the feature extraction module extracts features, and then the extracted features are cross-modally fused through the feature fusion module. Finally, the fusion result is processed through a fully connected layer as the input of the monitoring result output module. The monitoring result output module calculates based on the input data to obtain the final output of the model. The fully connected layer module is expressed using the following formula: Among them, H fused is the fused feature; W fc is the weight matrix to be learned; b fc is the bias vector to be learned; is the output vector, z0 is the healthy category score; z1 is the loose category score.

4. The method for monitoring wind-induced vibration of a wind turbine tower according to claim 3, wherein: The preprocessing module includes a dynamic wavelet noise reduction module and a working condition adaptive normalization module; dynamic wavelet noise reduction is used to suppress the high-frequency noise of wind in the data, and then working condition adaptive normalization is performed according to the wind speed interval to complete the data preprocessing; The dynamic wavelet denoising module is expressed using the following formula: in, is the acceleration signal of the vibration sensor at time t; SWT is dynamic wavelet denoising; is the wavelet basis function; λ is the scale-related threshold; σ is the standard deviation of the noise component; is the Haar wavelet transform basis function; is the instantaneous strain signal of the strain gauge at time t; The working condition adaptive normalization module is expressed by the following formula: in, for Normalized value; for Normalized value; w is the current wind speed range; μ v (w) is the mean value of the vibration signal under the current working condition; σ v (w) is the mean square error of the vibration signal under the current working condition; μ s (w) is the mean value of the strain signal under the current working condition; σ s (w) is the mean square error of the strain signal under the current working condition.

5. The method for monitoring wind-induced vibration of a wind turbine tower according to claim 3, wherein: The feature extraction module includes a vibration signal feature extraction module and a strain signal feature extraction module; The vibration signal feature extraction module first uses a multi-scale convolution kernel group to extract features in the frequency dimension, and then enhances the resonance frequency band response through frequency attention pooling to obtain vibration signal features. The feature extraction of the multi-scale convolution kernel group in the frequency dimension is expressed by the following formula: in, is the feature representation of the lth layer in the frequency dimension; k is the index of the frequency channel, and its value range is 1 to K; K is the total number of frequency channels; l is the convolution kernel level; is the one-dimensional convolution kernel corresponding to the k-th frequency channel in the l-th layer; The frequency attention pooling is expressed using the following formula: Among them, P v is the output result after frequency attention pooling; To perform the maximum pooling operation on the value of the kth frequency channel of the lth layer feature map; a k is a parameter, and q is the query vector to be learned; is the value of the l-th layer feature map at the k-th frequency channel; The strain signal feature extraction module first converts the input preprocessed strain signal data into preload data through the physical encoding of the bolt preload based on a temperature-compensated linear model. It then extracts time series features from the preload data using LSTM to obtain strain signal features. The physical coding of the bolt preload force is expressed using the following formula: Among them, F t is the preload at time t; k(τ) is the elastic modulus at temperature τ; b(τ) is the constant at temperature τ; The time series feature extraction of preload force data based on LSTM is expressed by the following formula: in, is the output hidden state of the long short-term memory network (LSTM) at time step t; LSTM is the long short-term memory network; H s It is the feature set of the entire time series data after being processed by the LSTM network.

6. The method for monitoring wind-induced vibration of a wind turbine tower according to claim 3, characterized in that: The feature fusion module specifically uses spatiotemporal alignment and attention fusion to fuse the input features across modalities, which can be expressed by the following formula: in, is element-by-element addition; H fused is the fused feature; A is the attention weight matrix, and A is expressed using the following formula: d is the dimension of the feature vector.

7. The method for monitoring wind-induced vibration of a wind turbine tower according to claim 3, wherein: The output structure of the monitoring result output module includes the bolt health status judgment result, loosening probability, vibration severity index and preload estimation value; The calculation formula of the loosening probability is: Among them, p is the loose probability; z0 is the healthy category score output by the fully connected layer module; z1 is the loose category score output by the fully connected layer module; The calculation formula of the vibration severity index is: Among them, I v is the vibration severity index; RMS v RMS is the root mean square amplitude of the tower vibration; max is the maximum RMS value of the amplitude corresponding to the wind condition of force 12; FER is the amplitude coefficient corresponding to the wind condition of force 12; The calculation formula of the estimated preload force is: in, is the estimated value of preload force; MLP() is multi-layer perceptron processing; H fused is the fused feature; The judgment logic of the bolt health status judgment result is: If the loose probability p is greater than or equal to 0.9, it is determined to be loose, otherwise it is determined to be normal.

8. The method for monitoring wind-induced vibration of a wind turbine tower according to claim 1, wherein: In step S4, the total loss function of the training is expressed by the following formula: Among them, ξ is the total loss function; is the classification cross entropy loss function; ξ vib is the loss term related to vibration; ξ bolt is the loss item related to the bolts; The vibration-related loss term ξ vib Based on the Newton-Euler equation, it is expressed as follows: in, is the tower vibration amplitude; m is the first tower parameter; c is the second tower parameter; k is the third tower parameter; is the wind pressure on the windward side of the tower; λ1 is the Lagrange multiplier of the corresponding model; The loss term ξ associated with the bolt bolt In order to generate the strain-preload nonlinear relationship library through finite element simulation, the following formula is used: Among them, F i is the strain and preload data pair; λ2 is the Lagrange multiplier of the corresponding model; is the bolt preload value at time t predicted by the model; N is the total number of data in the strain-preload relationship library; interp() is the interpolation function, which is used to calculate the bolt preload value according to the given current strain value ε. t , in the strain-preload nonlinear relationship library {ε i ,F i } to perform interpolation calculation to obtain the corresponding preload force estimation value; Categorical cross entropy loss function ξ CE , expressed using the following formula: Where y is the sample set: y=(y1,y2,...y C ); only the true category k satisfies y k =1, the rest y j =0(j≠k); In model prediction, the probability distribution of sample output is satisfy 9. The method for monitoring wind-induced vibration of a wind turbine tower according to claim 1, wherein: In step S5, the wind turbine tower wind-induced vibration monitoring model is lightweight deployed, and then the lightweight wind turbine tower wind-induced vibration monitoring model is used for monitoring; the lightweight deployment includes model compression and edge inference optimization; The model compression includes knowledge distillation and dynamic quantization; the knowledge distillation uses the wind turbine tower wind-induced vibration monitoring model as a teacher network to guide the lightweight student network; The total loss function of the knowledge distillation is expressed as follows: Among them, ξ KD is the total loss function of knowledge distillation; α is a hyperparameter, and its value range is between 0 and 1; ξ MSE is the mean square error loss function; is the feature vector output by the teacher network at layer s; is the feature vector output by the student network at layer s; The dynamic quantization maps the weights to 8-bit fixed-point numbers, which are expressed by the following formula: Among them, w quant is the quantized weight value; w is the original weight value; w max is the maximum value in the weight set that needs to be quantized; w min is the minimum value in the weight set that needs to be quantized; round() is the rounding function; The edge inference optimization is specifically to transform the fusion layer H fused Decompose into subgraphs that can be computed in parallel, adapt to embedded GPUs, and then process them from aspects such as task scheduling, data transmission, computing resource allocation, result integration, and performance optimization; The task scheduling is to arrange the computing tasks of each subgraph by utilizing the multi-core architecture of the embedded GPU; The data transmission is to optimize the data transmission process between subgraphs; The computing resource allocation is to dynamically allocate computing resources of the embedded GPU, such as video memory and computing units, according to the needs of the subgraph; The result integration is to integrate the results of each subgraph after the calculation is completed; The performance monitoring and optimization is to monitor the performance indicators of the embedded GPU in real time during the calculation process, such as utilization rate, temperature, etc. The memory is pre-allocated as a fixed input and output buffer to reduce dynamic memory overhead.

10. A system for implementing the method for monitoring wind-induced vibration of a wind turbine tower according to any one of claims 1 to 9, characterized in that: The system includes a multimodal sensor data acquisition module, a data simulation module, a model construction module, a model training module, and a wind turbine tower wind-induced vibration monitoring module; The multimodal sensor data acquisition module uses the multimodal sensor to collect data and uploads the data to the data simulation module; The data simulation module performs numerical simulation on the collected data based on the received data to obtain the training data set and uploads the data to the model training module; The model building module builds the initial wind turbine tower wind-induced vibration monitoring model and uploads the data to the model training module; The model training module trains the initial wind turbine tower wind-induced vibration monitoring model using the training data set based on the received data to obtain the wind turbine tower wind-induced vibration monitoring model, and uploads the data to the wind turbine tower wind-induced vibration monitoring module; The wind turbine tower wind-induced vibration monitoring module performs actual wind turbine tower wind-induced vibration monitoring based on the received data and using the wind turbine tower wind-induced vibration monitoring model.

Images