Offshore wind power fan structure vibration prediction method and system based on time sequence model

Abstract

The invention belongs to the technical field of offshore wind power equipment monitoring, and particularly provides an offshore wind power fan structure vibration prediction method and system based on a time sequence model, and the method comprises the steps: arranging an acceleration sensor and a strain sensor to obtain a vibration signal, and obtaining the environment parameters of wind speed, wave height and flow velocity at the same time; performing short-time Fourier transform on the original vibration signal by adopting a Hanning window function to obtain a two-dimensional time-frequency matrix; performing normalization processing on the time-frequency matrix, and generating a training sample with a shape of (C, T, F) through a sliding window mode; a CNN-LSTM model is constructed, a CNN layer is used for feature extraction, and an LSTM layer is used for time sequence modeling; and training the CNN-LSTM model, outputting a future vibration prediction value, and if the prediction value exceeds a preset safety threshold, outputting an alarm signal. According to the method, the accuracy and robustness of vibration prediction are improved by fusing spatial feature extraction and time sequence dependence modeling, so that safe and stable operation of the offshore wind turbine generator is guaranteed.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of offshore wind power equipment monitoring, and in particular relates to a method and system for offshore wind turbine structure vibration prediction based on a time series model. BACKGROUND

[0002] As a core component of clean energy, offshore wind power has been rapidly developing worldwide in recent years. Offshore wind turbine units are exposed to complex marine environments for a long time, and their foundation pile and tower structure need to continuously withstand wind load, wave load, ocean current action and dynamic load generated by the operation of the unit itself, which is prone to periodic or random vibration. Such vibration not only causes fatigue damage accumulation and structural performance degradation of the pile and tower, significantly shortening the service life of the unit, but also may cause resonance with the main engine of the wind turbine at a specific frequency, leading to abnormal operation of the equipment and even serious safety accidents. Therefore, accurate monitoring and early prediction of the vibration state of the pile and tower are key technical means to ensure the safe and stable operation of offshore wind turbine units, prolong the service life and reduce operation and maintenance costs.

[0003] Currently, research on offshore wind turbine pile and tower vibration prediction mainly focuses on three technical paths: first, finite element modeling and dynamics analysis method, which relies on complex boundary condition setting and a large number of prior assumptions, making it difficult to accurately reflect the variability of marine environments (such as random fluctuations in wind speed and wave height), resulting in a large deviation between the prediction results and actual working conditions; second, frequency domain or time domain analysis method based on signal processing, which can extract local features of vibration signals, but cannot effectively model the nonlinear relationships and strong coupling characteristics existing in long time series data, limiting the prediction ability; third, traditional machine learning prediction methods (such as support vector machines and random forests), which rely on manually designed features and have insufficient generalization performance, making it difficult to adapt to complex and multi-dimensional working condition changes of offshore wind power.

[0004] In summary, the existing technology has three major defects in offshore wind turbine pile and tower vibration prediction: first, it is difficult to simultaneously consider local feature extraction and long-term dependence modeling, resulting in insufficient prediction accuracy; second, the model has poor adaptability to marine environmental disturbances, and the prediction stability decreases under complex working conditions; third, it lacks multi-time span early warning capability based on accurate prediction, and cannot provide effective support for operation and maintenance decisions. Therefore, there is an urgent need for a vibration prediction technology that has high prediction accuracy, strong environmental adaptability and early warning function to provide protection for the safe and reliable operation of offshore wind turbine units. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a method and system for offshore wind turbine structure vibration prediction based on a time sequence model, which fuses spatial feature extraction and time sequence dependent modeling to improve the accuracy and robustness of vibration prediction, thereby ensuring the safe and stable operation of offshore wind turbine units.

[0006] To solve the above technical problems, the technical solution adopted by the present application is: a method for offshore wind turbine structure vibration prediction based on a time sequence model, comprising the following steps: Step S1, data acquisition: acquiring vibration signals through acceleration sensors and strain sensors arranged on pile foundations and tower drums, and simultaneously acquiring wind speed, wave height and flow rate environmental parameters; Step S2, time-frequency transformation: performing short-time Fourier transformation on original vibration signals using a Hanning window function to obtain a two-dimensional time-frequency matrix with a dimension of TxF, wherein T is a time window length, F is a frequency resolution, and the frequency resolution F=T / F is f s / T w , f s is a sampling frequency, T w is a window length; Step S3, data preprocessing: performing Z-score normalization processing on the time-frequency matrix, generating training samples with a shape of (C,T,F) through a sliding window method, C is a channel number and contains vibration signals of different measuring points, simultaneously identifying abnormal data using a 3σ criterion and completing missing values through a linear interpolation method; Step S4, constructing a CNN-LSTM model: The CNN layer is used for feature extraction, and the time-frequency matrix is input into a convolutional neural network, which includes a first convolutional layer, a second convolutional layer, a 2x2 maximum pooling layer, a batch normalization layer and a flattening layer; The LSTM layer is used for time sequence modeling, and the convolutional feature sequence is input into a double-layer LSTM network, and an attention mechanism module is arranged between the two layers of LSTM.

[0007] Step S5, model training and optimization: using mean square error as a loss function and using an Adam optimizer to train the model; Step S6, prediction output: the trained CNN-LSTM model, the LSTM output is mapped to future vibration prediction values through a fully connected layer, and if the prediction value exceeds a preset safety threshold, an alarm signal is output.

[0008] In a preferred solution, in the step S3, the step size of the sliding window is set as T / 2, the time window length T of the training sample is in the range of 100-1000 sampling points, and the number of channels C is in the range of 2-8, corresponding to vibration signals of measuring points with different heights and different orientations.

[0009] In a preferred solution, in the step S4, in the CNN layer, the convolution kernel size of the first convolution layer is 3x3, the number is 32, and the ReLU activation function is used; the convolution kernel size of the second convolution layer is 5x5, the number is 64, and the ReLU activation function is used, and the step size of the two convolution layers is 1, and the padding mode is SAME padding.

[0010] In a preferred solution, in the step S4, in the double-layer LSTM network, the number of hidden units of the first layer LSTM is 128, which is used to model short-term time sequence dependence; the number of hidden units of the second layer LSTM is 64, which is used to model long-term time sequence dependence, and both layers of LSTM use dropout layer to suppress overfitting, and the dropout probability is set to 0.2-0.5.

[0011] In a preferred solution, in the step S4, the attention mechanism module adopts additive attention mechanism, dynamically allocates the contribution of different time step features to the prediction result by calculating the weight coefficient of each time step feature, and the weight coefficient is normalized by the sigmoid function.

[0012] In a preferred solution, in the model training and optimization of the step S5, the initial learning rate of the Adam optimizer is set to 0.001-0.01, and the learning rate is adjusted by an exponential decay strategy, and the learning rate is decayed to 0.9 times of the original learning rate every 10 training rounds.

[0013] In a preferred solution, in the step S5, L2 regularization is introduced, the regularization coefficient is 0.0001-0.001, and the early stopping strategy is combined, and the training is stopped when the validation set loss does not decrease for 5 consecutive rounds.

[0014] In a preferred solution, in the step S6, the time span of the future vibration prediction value includes three levels of 1 second, 10 seconds, and 60 seconds.

[0015] In a preferred solution, the preset safety threshold is determined based on the structural fatigue limit, material strength parameter and safety factor of the offshore wind turbine pile foundation and tower cylinder, including a strain threshold and an acceleration threshold.

[0016] In a preferred solution, the calculation method of the strain threshold is: 1) Fatigue allowable strain e f Calculation: ; in: s f,comp This represents the actual fatigue limit of the component. E The elastic modulus of the material; n f The fatigue safety factor is set at 1.2-1.5. ; in, s -1 The symmetrical cyclic fatigue limit of the material; β The surface condition coefficient refers to the steel structure of offshore wind power after anti-corrosion treatment. β =0.85-0.95, and 0.7-0.85 for severely corroded areas; E The dimension factor represents the thickness of the component section. t When ≤20mm, e =0.9-1.0, t When >20mm e =0.75-0.89; K t The stress concentration factor refers to the welds and flange connections between the pile foundation and the tower. K t =1.8-3.5, butt welds are 2.0-2.5, fillet welds are 2.8-3.5; 2) Allowable strain for static strength e s calculate: ; Among them: s s Material yield strength; n s The static strength safety factor is set at 1.1-1.3. 3) Final strain threshold e th The smaller of the allowable fatigue strain and the allowable strength strain is expressed as: e th =min( e f , e s ).

[0017] In a preferred embodiment, the acceleration threshold is calculated as follows: 1) Allowable acceleration due to fatigue calculate: ; in, p The density of the material; LThe distance from the installation location of the accelerometer to the point of mass concentration of the structure; K d The dynamic load factor indicates that offshore wind power vibration is simple harmonic or random. K d =1.2-1.8; 2) Allowable static acceleration calculate: ; 3) Final acceleration threshold The formula is the smaller of the allowable acceleration due to labor and the allowable acceleration due to static strength: .

[0018] This invention also provides a time-series model-based offshore wind turbine structural vibration prediction system, used to perform the above-described time-series model-based offshore wind turbine structural vibration prediction method, comprising: Data acquisition module: includes accelerometer, strain sensor and environmental parameter acquisition unit. Accelerometer is deployed at the top of pile foundation, middle of tower and top of tower. Strain sensor is deployed at weld joints between pile foundation and tower and stress concentration points. Environmental parameter acquisition unit is used to collect wind speed, wave height and flow velocity data in real time. Time-frequency conversion module: used to convert the original signal into a time-frequency matrix; Data preprocessing module: used for normalization and sample generation; Feature extraction module: includes a CNN layer for feature extraction, which inputs the time-frequency matrix into the convolutional neural network, which includes a first convolutional layer, a second convolutional layer, a 2×2 max pooling layer, a batch normalization layer and a flattening layer; Temporal modeling module: includes LSTM layers for temporal modeling. Convolutional feature sequences are input into a two-layer LSTM network, and an attention mechanism module is set between the two LSTM layers. Prediction and Early Warning Module: Used to output vibration prediction values ​​over multiple time spans and trigger an early warning when the values ​​exceed a threshold.

[0019] In a preferred embodiment, the time-frequency conversion module incorporates a Hanning window function and a short-time Fourier transform algorithm, enabling it to convert the input sampling frequency according to the frequency. f s and window length T w The frequency resolution F is automatically calculated, the original vibration signal is converted into a two-dimensional time-frequency matrix of T×F, and the time dimension identifier and frequency dimension identifier of the matrix are stored.

[0020] In a preferred embodiment, the data preprocessing module incorporates a Z-score normalization algorithm, a 3σ outlier identification algorithm, and a linear interpolation completion algorithm. It can automatically normalize the time-frequency matrix output by the time-frequency transformation module, identify and remove outlier data, and complete missing values. Then, it generates training and validation samples of (C,T,F) shape through a sliding window, with the ratio of training samples to validation samples being 7:3 to 8:2.

[0021] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method for predicting the structural vibration of offshore wind turbines based on a time-series model.

[0022] The present invention provides a method and system for predicting the structural vibration of offshore wind turbines based on a time-series model, which has the following significant advantages: 1. By using a CNN-LSTM fusion model, both time-frequency local feature extraction and long-term temporal dependency modeling are achieved simultaneously. Combined with an attention mechanism to dynamically optimize feature weights, the prediction accuracy of complex vibration signals is significantly improved.

[0023] STFT, combined with the Hanning window function, converts a one-dimensional vibration signal into a T×F two-dimensional time-frequency matrix, fully preserving the joint time-frequency characteristics and avoiding information loss in single time-domain or frequency-domain analysis.

[0024] 2. To address data anomalies and missing data caused by random fluctuations in marine environment (wind speed, wave height), this invention constructs a full-process data preprocessing system. Combined with multi-point data acquisition, it enhances the model's anti-interference capability. The preprocessing stage employs Z-score normalization to eliminate dimensional differences, the 3σ criterion to accurately identify abnormal data, and linear interpolation to complete missing values, ensuring data quality. Accelerometers / strain sensors are deployed at different heights in pile foundations and towers, as well as stress concentration areas such as welds, with 2-8 channels covering key structures and providing comprehensive data dimensions.

[0025] 3. Design a multi-time span (1 second, 10 seconds, 60 seconds) prediction output mechanism, combined with a safety threshold calculation method based on fatigue limit and static strength dual control, to achieve graded early warning from emergency response to long-term assessment; the early warning signal is triggered by both sound and light and remote communication to ensure timely response by operation and maintenance personnel and effectively avoid safety risks such as resonance and fatigue damage.

[0026] 4. The model training process employs regularization and early stopping strategies to avoid overfitting, resulting in high training efficiency; the transfer learning function significantly shortens the adaptation time of the fresh air unit; and preventive maintenance is achieved through early warning, reducing unplanned downtime losses and lowering operation and maintenance costs by more than 30%. Attached Figure Description

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the method of the present invention; Detailed Implementation To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0028] Example 1: like Figure 1 As shown, a method for predicting the structural vibration of offshore wind turbines based on a time-series model includes the following steps: Step S1, Data Acquisition: Vibration signals are acquired by using acceleration sensors and strain sensors installed on the pile foundation and tower, and environmental parameters such as wind speed, wave height, and flow velocity are also acquired.

[0029] Among them, the sampling frequency of the accelerometer f s The sampling frequency range is 100-1000Hz, and the sampling frequency of the strain sensor is consistent with that of the accelerometer. In the environmental parameter acquisition unit, the anemometer has a range of 0-70m / s and an accuracy of ±0.3m / s, the wave height meter has a range of 0-30m and an accuracy of ±0.1m, and the current meter has a range of 0-5m / s and an accuracy of ±0.05m / s. The vibration signal sampling interval is 1ms, and the environmental parameter sampling interval is 1s.

[0030] Step S2, Time-Frequency Transform: Perform a Short-Time Fourier Transform (STFT) on the original vibration signal using a Hanning window function to obtain a two-dimensional time-frequency matrix of dimension T×F, where T is the time window length, F is the frequency resolution, and the frequency resolution F= f s / T w , f s Sampling frequency, T w The length of the window.

[0031] The original vibration signal is subjected to a short-time Fourier transform (STFT) using a Hanning window function, converting the one-dimensional time-domain vibration signal into a two-dimensional time-frequency matrix of dimension T×F. Here, T is the time window length, F is the frequency resolution, and the frequency resolution F = ... f s / T w , f s Sampling frequency, T w The length of the window.

[0032] Window length T w The value range is 0.1-1s, ensuring that each time window contains at least one complete vibration signal cycle.

[0033] Step S3, Data Preprocessing: The time-frequency matrix is ​​normalized using Z-score, and training samples of shape (C,T,F) are generated using a sliding window method, where C is the number of channels and contains vibration signals from different measurement points. At the same time, the 3σ criterion is used to identify abnormal data and missing values ​​are filled in using linear interpolation.

[0034] The specific steps are as follows: 1) Normalization process: The Z-score normalization algorithm is applied to the time-frequency matrix. The normalization formula is as follows: ; in, m The mean of all samples in the training set. s The standard deviation of all samples in the training set is used for both the validation and test sets. m and s To avoid data leaks.

[0035] 2) Handling outliers and missing values: The 3σ criterion is used to identify outlier data; data exceeding... m ±3 s If an anomaly is detected, missing values ​​are filled in using linear interpolation. The interpolation formula is: , x i For missing data points, x i−3 , x i+3 Three normal sample points are taken before and after the missing point.

[0036] 3) Sample generation: Training and validation samples of shape (C,T,F) are generated using a sliding window method. The sliding window step size is set to T / 2 (i.e., 50-500 sampling points), and the overlap rate of adjacent samples is 50%. The number of channels C ranges from 2 to 8, corresponding to vibration signals of measuring points at different heights and orientations. The training and validation samples are divided in chronological order, with a ratio of 7:3 to 8:2, and the number of training samples generated under each working condition is no less than 1000.

[0037] Step S4: Construct the CNN-LSTM model: 1. CNN layer: used for feature extraction, the time-frequency matrix is ​​input into the convolutional neural network, the convolutional neural network includes a first convolutional layer, a second convolutional layer, a 2×2 max pooling layer, a batch normalization layer and a flattening layer.

[0038] First convolutional layer: The weights are initialized using He normality, the kernel size is 3×3 with 32 kernels, the stride is 1, the padding method is SAME (when the input size is odd, the number of zeros on the left and right differs by 1 to ensure that the output size is consistent with the input), and the activation function is ReLU with leakage (negative slope 0.01-0.05).

[0039] The second convolutional layer uses He normal initialization for weights, with a kernel size of 5×5, a number of 64 kernels, a stride of 1, a padding method of SAME, and an activation function of leaky ReLU (negative slope 0.01-0.05).

[0040] 2×2 max pooling layer: Takes the maximum value of the data in the window. If there is a NaN value in the window, it is skipped and the maximum value of the remaining values ​​is taken. The size after pooling is (T / / 2)×(F / / 2) (rounded down).

[0041] Batch normalization layer: Calculates the mean of the convolutional output for each batch. m b and variance s b 2 According to the formula After standardization, multiply by the scaling factor γ (initially 1) and the bias β (initially 0), and update γ and β synchronously during training. .

[0042] Flattening layer: The 64-channel × T' × F' (T' = T / / 2, F' = F / / 2) feature map output from the second convolutional layer is flattened into a time series form of [number of samples, T', 64 × F'], which is then used as the input to the LSTM layer.

[0043] 2. LSTM layer: Used for temporal modeling, it inputs the convolutional feature sequence into a two-layer LSTM network, with an attention mechanism module between the two LSTM layers.

[0044] The LSTM unit uses orthogonally initialized weights, the hidden state activation function is tanh, the gate activation function is sigmoid, dropout is applied only to the input gate and forget gate (probability 0.2-0.5), and gradient explosion is prevented by gradient norm clipping (L2 norm is limited to within 1.0).

[0045] The first LSTM layer has 128 hidden units, an input dimension of 64×F', and an output dimension of 128. It is used to model short-term temporal dependencies.

[0046] The second LSTM layer has 64 hidden units, 128 input dimensions, and 64 output dimensions, and is used to model long-term temporal dependencies.

[0047] Attention mechanism module: An additive attention mechanism is adopted, which dynamically allocates the contribution of different time step features to the prediction result by calculating the weight coefficient of each time step feature. The weight coefficient is normalized by the sigmoid function, ranging from 0 to 1, and the sum of the weights of all time steps is 1. The weight matrix is ​​updated synchronously with the model training.

[0048] Step S5, Model Training and Optimization: Use mean squared error as the loss function and train the model using the Adam optimizer.

[0049] The Adam optimizer has an initial learning rate of 0.001-0.01 and uses an exponential decay strategy to adjust the learning rate, which decays to 0.9 times its original value every 10 training epochs.

[0050] L2 regularization is introduced, with a regularization coefficient of 0.0001-0.001. Combined with an early stopping strategy, training is stopped when the validation set loss does not decrease for 5 consecutive rounds.

[0051] The specific description is as follows: 1. Loss Function: The mean squared error (MSE) is used as the loss function, and the calculation formula is as follows: ; in: N For the sample size, y i This represents the actual vibration value. These are predicted values.

[0052] 2. Optimizer: The model is trained using the Adam optimizer, with an initial learning rate set to 0.001-0.01, β1=0.9, β2=0.999, and epsilon=1e-7. An exponential decay strategy is used to adjust the learning rate (learning rate in the k-th round = initial learning rate × 0.9). (k / 10) ).

[0053] 3. Regularization and Early Stopping Strategy: L2 regularization is introduced (applied only to CNN convolutional kernel weights and LSTM hidden layer weights, with a regularization coefficient of 0.0001-0.001). Regularization loss = λ × Σw² (λ is the regularization coefficient), and total loss = MSE + regularization loss. Combined with the early stopping strategy, training is stopped when the validation set loss does not decrease for 5 consecutive rounds and the decrease is less than 1e-5. The batch size during training is set to 32-128, adjusted according to the GPU memory size.

[0054] Step S6: Predicted Output: The trained CNN-LSTM model outputs LSTM outputs and maps them to predicted future vibration values ​​through a fully connected layer. If the predicted value exceeds a preset safety threshold, an alarm signal is output.

[0055] The trained CNN-LSTM model maps the LSTM output to vibration prediction values ​​for three time spans: 1 second, 10 seconds, and 60 seconds, through a fully connected layer (64 input dimensions and 3 output dimensions). The fully connected layer uses He normal weight initialization and contains a 32-neuron hidden layer (ReLU activation) and a 3-neuron output layer (no activation function, directly outputting vibration amplitude, unit: m / s² or με).

[0056] The preset safety threshold is determined based on the structural fatigue limit, material strength parameters and safety factor of the offshore wind turbine pile foundation and tower, including strain threshold and acceleration threshold.

[0057] 1. The method for calculating the strain threshold is as follows: 1) Allowable fatigue strain e f calculate: ; in: s f,comp This represents the actual fatigue limit of the component. E The elastic modulus of the material; n f The fatigue safety factor is set at 1.2-1.5. ; in, s -1 The symmetrical cyclic fatigue limit of the material; β The surface condition coefficient refers to the steel structure of offshore wind power after anti-corrosion treatment. β =0.85-0.95, and 0.7-0.85 for severely corroded areas; E The dimension factor represents the thickness of the component section. t When ≤20mm, e =0.9-1.0, t When >20mm e =0.75-0.89; K t The stress concentration factor refers to the welds and flange connections between the pile foundation and the tower. K t =1.8-3.5, butt welds are 2.0-2.5, fillet welds are 2.8-3.5; 2) Allowable strain for static strength e s ; Among them: s s Material yield strength; n s The static strength safety factor is set at 1.1-1.3. 3) Final strain threshold e th The smaller of the allowable fatigue strain and the allowable strength strain is expressed as: e th =min( e f , e s ).

[0058] 2. The calculation method for the acceleration threshold is as follows: 1) Allowable acceleration due to fatigue calculate ; in, p The density of the material; L The distance from the installation location of the accelerometer to the point of mass concentration of the structure; K d The dynamic load factor indicates that offshore wind power vibration is simple harmonic or random. K d =1.2-1.8; 2) Allowable static acceleration calculate: ; 3) Final acceleration threshold The formula is the smaller of the allowable acceleration due to labor and the allowable acceleration due to static strength: .

[0059] In practice, before the wind turbine is put into operation, the component material testing report ( s -1 , s s ), structural design drawings ( L , K t ) and marine environmental assessment report ( β Determine the basic parameters and substitute them into the above formula to calculate the initial threshold.

[0060] Adjust parameters every 3 months based on actual operation and maintenance data (such as maximum vibration value during non-damage operation, material corrosion detection results, etc.). β (Adjusted as corrosion deepens), threshold recalculated; When minor structural damage occurs (such as weld cracks), the safety factor needs to be increased. n f The threshold is recalculated and a more stringent warning logic is triggered.

[0061] In practice, the following implementation steps shall be adopted: Data acquisition: Vibration signals of the pile foundation and tower are continuously collected by 6 accelerometers and 6 strain sensors. The environmental parameter acquisition unit collects wind speed, wave height and flow velocity data simultaneously. The vibration signal sampling interval is 1ms, the environmental parameter sampling interval is 1s, and the data caching unit caches 24 hours of data in real time.

[0062] Time-frequency transformation: Apply the Hanning window function to the acquired raw vibration signal to perform a short-time Fourier transform, generate a two-dimensional time-frequency matrix with a dimension of 250×1000, and store the timestamp and frequency value identifier of the matrix.

[0063] Data preprocessing: Z-score normalization was performed on the time-frequency matrix, outliers were identified and removed using the 3σ criterion, and missing values ​​were filled in using linear interpolation; 4000 training samples, 1000 validation samples, and 750 test samples with a shape of (6,250,1000) were generated using a sliding window.

[0064] Model training: Input the training samples into the CNN-LSTM fusion model, use the mean squared error as the loss function, and train for 30 rounds using the Adam optimizer; during the training process, the validation set loss reaches its minimum and stabilizes in the 22nd round, triggering the early stopping strategy, and saving the model parameter file (approximately 120MB in size).

[0065] Threshold calculation: Strain threshold calculation: s f,comp =170×0.9×0.85 / 2.2≈59.8 MPa ; e f =59.8 / (206×103×1.4)≈205 m ; e s =345 / (206×103×1.2)≈1380 m ; final e th =min(205,1380)=205 m .

[0066] Acceleration threshold calculation: sf =59.8 / 1.4≈42.7 MPa ; a f =42.7×10⁶ / (7850×25×1.6)≈13.5 m / s 2 ; s s =345 / 1.2≈287.5 Mpa ; a s =287.5×10⁶ / (7850×25×1.6)≈91.3 m / s 2 ; final a th =min(13.5,91.3)=13.5 m / s 2 .

[0067] Prediction and Early Warning: The test set data is input into the trained model, which outputs vibration prediction values ​​for the next 1 second, 10 seconds, and 60 seconds. The test results show that the MSE for the 1-second span is 0.0042 and R² is 0.983, the MSE for the 10-second span is 0.0085 and R² is 0.967, and the MSE for the 60-second span is 0.017 and R² is 0.952, all of which meet the performance requirements. When the vibration acceleration of a certain test sample is predicted to be 14.2 m / s² after 1 second (exceeding the threshold of 13.5 m / s²), the system simultaneously triggers an audible and visual alarm and a remote communication alarm, sending an early warning message to the monitoring platform and the operation and maintenance terminal.

[0068] Model Deployment and Iteration: The trained model is converted into ONNX format and deployed to an edge computing device. It is connected to the data acquisition module and the prediction and early warning module via gigabit Ethernet to achieve real-time prediction. The threshold is calibrated every 3 months using operation and maintenance data, and the model performance is reviewed every 6 months. If the accuracy drops by more than 15%, it is retrained based on newly acquired data.

[0069] Example 2: A time-series model-based offshore wind turbine structural vibration prediction system is used to execute the time-series model-based offshore wind turbine structural vibration prediction method described in Example 1, comprising: Data acquisition module: includes accelerometer, strain sensor and environmental parameter acquisition unit. Accelerometer is deployed at the top of pile foundation, middle of tower and top of tower. Strain sensor is deployed at weld seam between pile foundation and tower and stress concentration point. Environmental parameter acquisition unit is used to collect wind speed, wave height and flow velocity data in real time.

[0070] The system includes accelerometers, strain gauges, and environmental parameter acquisition units. The accelerometers are piezoelectric accelerometers with a range of ±50g, accuracy of ±0.1%FS, and a sampling depth of ≥16 bits. They are deployed at the top of the pile foundation (2 sensors, measuring x and y directions), the bottom of the pile foundation (2 sensors, measuring x and y directions), the middle of the tower (2 sensors, measuring x and y directions), and the top of the tower (2 sensors, measuring x and y directions), for a total of 8 sensors. The strain gauges are resistance strain gauges with a range of ±2000με, accuracy of ±0.5%FS, and a sampling depth of ≥16 bits. They are deployed at stress concentration points such as the circumferential welds, longitudinal welds, and flange connections between the pile foundation and the tower, for a total of 4-8 sensors. The environmental parameter acquisition unit includes an anemometer, wave height meter, and current meter, installed at the top of the tower (anemometer) and in the sea area near the wind turbine (wave height meter and current meter), respectively. The module also has a 10GB-100GB data cache unit to support temporary data storage and recovery during network interruptions.

[0071] Time-frequency conversion module: used to convert the original signal into a time-frequency matrix.

[0072] The time-frequency conversion module incorporates a Hanning window function and a short-time Fourier transform algorithm, enabling it to convert input sampling frequency according to the frequency. f s and window length T w The system automatically calculates the frequency resolution F, converts the original vibration signal into a T×F two-dimensional time-frequency matrix, and stores the time and frequency dimension identifiers of the matrix. The data preprocessing module is used for normalization and sample generation.

[0073] Data preprocessing module: Used for normalization and sample generation, it has built-in Z-score normalization algorithm, 3σ outlier identification algorithm and linear interpolation completion algorithm. It can automatically normalize the time-frequency matrix output by the time-frequency transformation module, identify and remove outlier data, and fill in missing values. Then, it generates training samples and validation samples of (C,T,F) shape through a sliding window. The ratio of training samples to validation samples is 7:3~8:2.

[0074] Feature extraction module: includes a CNN layer for feature extraction, which inputs the time-frequency matrix into the convolutional neural network. The convolutional neural network includes a first convolutional layer, a second convolutional layer, a 2×2 max pooling layer, a batch normalization layer, and a flattening layer.

[0075] Temporal modeling module: includes LSTM layers for temporal modeling. Convolutional feature sequences are input into a two-layer LSTM network, and an attention mechanism module is set between the two LSTM layers.

[0076] Prediction and Early Warning Module: Used to output vibration prediction values ​​over multiple time spans and trigger an early warning when the values ​​exceed a threshold.

[0077] Example 3: This embodiment discloses a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the method for predicting the structural vibration of offshore wind turbines based on a time-series model as described in Embodiment 1.

[0078] The computer-readable storage medium is a solid-state drive (SSD), a hard disk drive (HDD), a USB flash drive, an optical disc, a read-only memory (ROM), or a random access memory (RAM) on which computer programs are stored.

[0079] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for offshore wind turbine structure vibration prediction based on a time series model, characterized in that, The method comprises the following steps: Step S1, data acquisition: obtaining vibration signals through acceleration sensors and strain sensors arranged on the pile foundation and the tower drum, and simultaneously obtaining environmental parameters such as wind speed, wave height and flow rate; Step S2, time-frequency transformation: a short-time Fourier transform is performed on the original vibration signal by using a Hanning window function to obtain a two-dimensional time-frequency matrix with a dimension of TxF, wherein T is a time window length, F is a frequency resolution, and the frequency resolution F = f / T, f s / T w , f s is a sampling frequency, T w is a window length; Step S3, data preprocessing: performing Z-score normalization processing on the time-frequency matrix, generating training samples in the shape of (C, T, F) through a sliding window method, C being the number of channels and containing vibration signals of different measuring points, and simultaneously identifying abnormal data through a 3σ criterion and completing missing values through a linear interpolation method; Step S4, constructing a CNN-LSTM model: The CNN layer is used for feature extraction, and the time-frequency matrix is input into a convolutional neural network, which comprises a first convolutional layer, a second convolutional layer, a 2x2 maximum pooling layer, a batch normalization layer and a flattening layer; The LSTM layer is used for time series modeling, and the convolutional feature sequence is input into a double-layer LSTM network, and an attention mechanism module is arranged between the two layers of LSTM; Step S5, model training and optimization: using mean square error as a loss function and using an Adam optimizer to train the model; Step S6, prediction output: the trained CNN-LSTM model, the LSTM output is mapped into a future vibration prediction value through a fully connected layer, and if the prediction value exceeds a preset safety threshold, an alarm signal is output.

2. The method for offshore wind turbine structure vibration prediction based on timing model according to claim 1, characterized in that, In the step S3, the step length of the sliding window is set as T / 2, the time window length T of the training sample is in the range of 100-1000 sampling points, and the number of channels C is in the range of 2-8, corresponding to the vibration signals of measuring points at different heights and different directions.

3. The method for offshore wind turbine structure vibration prediction based on timing model according to claim 1, characterized in that, In the step S4, in the CNN layer, the convolution kernel size of the first convolutional layer is 3x3, the number is 32, and a ReLU activation function is used; the convolution kernel size of the second convolutional layer is 5x5, the number is 64, and a ReLU activation function is used, and the step length of the two convolutional layers is 1, and the padding mode is SAME padding.

4. The method for offshore wind turbine structure vibration prediction based on timing model according to claim 1, characterized in that, In the step S4, in the double-layer LSTM network, the number of hidden units of the first layer of LSTM is 128, which is used for modeling short-term time series dependence; the number of hidden units of the second layer of LSTM is 64, which is used for modeling long-term time series dependence, and both layers of LSTM use a dropout layer to suppress overfitting, and the dropout probability is set to 0.2-0.

5.

5. The method for offshore wind turbine structure vibration prediction based on timing model according to claim 1, characterized in that, In the step S4, the attention mechanism module adopts an additive attention mechanism, calculates the weight coefficient of each time step feature, dynamically allocates the contribution degree of different time step features to the prediction result, and normalizes the weight coefficient through a sigmoid function.

6. The method for offshore wind turbine structure vibration prediction based on timing model according to claim 1, characterized in that, In the model training and optimization of the step S5, the initial learning rate of the Adam optimizer is set to 0.001-0.01, an exponential decay strategy is used to adjust the learning rate, and the learning rate is decayed to 0.9 times of the original value every 10 training rounds.

7. The method for offshore wind turbine structure vibration prediction based on timing model according to claim 1, characterized in that, In the step S5, L2 regularization is introduced, the regularization coefficient is 0.0001-0.001, and an early stopping strategy is combined, and the training is stopped when the loss of the validation set does not decrease for 5 consecutive rounds.

8. The method for offshore wind turbine structure vibration prediction based on timing model according to claim 1, characterized in that, In the step S6, the time span of the future vibration prediction value includes three levels of 1 second, 10 seconds and 60 seconds.

9. The method for offshore wind turbine structure vibration prediction based on timing model according to claim 1, characterized in that, The preset safety threshold is determined based on the structural fatigue limit, material strength parameter and safety factor of the offshore wind turbine pile foundation and tower drum, and includes a strain threshold and an acceleration threshold.

10. The method for offshore wind turbine structure vibration prediction based on timing model according to claim 9, characterized in that, The calculation method of the strain threshold is: 1) Fatigue Allowable Strain The calculation method of the acceleration threshold is: f Calculation: ； wherein: The method for performing offshore wind turbine structural vibration prediction based on the time sequence model according to any one of claims 1-11 comprises: f,comp is the actual fatigue limit of the component; E is the modulus of elasticity of the material; n f is the fatigue safety factor, which is taken as 1.2-1.5; ； Wherein, A data acquisition module: including an acceleration sensor, a strain sensor and an environmental parameter acquisition unit, the acceleration sensor is arranged at the top of the pile foundation, the middle of the tower drum and the top of the tower drum, the strain sensor is arranged at the weld and stress concentration position of the pile foundation and the tower drum; the environmental parameter acquisition unit is used for real-time acquisition of wind speed, wave height and flow rate data; -1 is the material symmetry cycle fatigue limit; β is the surface state coefficient, the steel structure of offshore wind power is treated by corrosion protection β = 0.85-0.95, 0.7-0.85 is taken in the serious corrosion area; A time-frequency transformation module: used for converting the original signal into a time-frequency matrix; is the size coefficient, the section thickness of the component t ≤20mm, A data preprocessing module: used for normalization and sample generation; =0.9-1.0, t >20mm A feature extraction module: including a CNN layer, used for feature extraction, inputting the time-frequency matrix into a convolutional neural network, the convolutional neural network includes a first convolutional layer, a second convolutional layer, a 2x2 maximum pooling layer, a batch normalization layer and a flattening layer; =0.75-0.89; K t is the stress concentration coefficient, the welding seam of the pile foundation and the tower drum, the flange connecting part K t =1.8-3.5, 2.0-2.5 is taken for the butt welding seam, 2.8-3.5 is taken for the fillet welding seam; 2) Static strength allowable strain A time sequence modeling module: including an LSTM layer, used for time sequence modeling, inputting the convolutional feature sequence into a double-layer LSTM network, and an attention mechanism module is arranged between the two layers of LSTM; s Calculation: ； Wherein: is A prediction and early warning module: used for outputting vibration prediction values in multiple time spans, and triggering early warning when the threshold is exceeded. s The material yield strength; n s The static strength safety factor is 1.1-1.

3. 3) final strain threshold The data preprocessing module is built-in Z-score normalization algorithm, 3σ outlier identification algorithm and linear interpolation completion algorithm, which normalizes the time-frequency matrix output by the time-frequency transformation module, identifies and eliminates abnormal data, and after completing the missing values, generates (C, T, F) shaped training samples and verification samples through a sliding window, the ratio of training samples to verification samples is 7:3-8:

2. th The smaller of the fatigue allowable strain and the strength allowable strain, expressed as: The computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize the method for offshore wind turbine structural vibration prediction based on the time sequence model according to any one of claims 1-11. th = min( ​ f , ​ s ).

11. The method of claim 9, wherein the method further comprises: ​ 1) Fatigue Allowable Acceleration Calculation: ； wherein ​ is the material density; L is the distance from the acceleration sensor mounting position to the structural mass center point; K d is the dynamic load factor, the offshore wind vibration is a simple harmonic or random vibration, K d = 1.2 - 1.8; 2) static strength allowable acceleration Calculation: ； 3) Final acceleration threshold The lesser of the acceleration and static strength allowable acceleration for the crew, which is given by the formula: 。 12. A system for offshore wind turbine structure vibration prediction based on a time series model, characterized by, ​ ​ ​ ​ ​ ​ ​ 13. A timing model based offshore wind turbine structure vibration prediction system according to claim 12, wherein, The time-frequency conversion module is built-in with Hanning window function and short-time Fourier transform algorithm, can convert the input sampling frequency f s and window length T w The frequency resolution F is automatically calculated, the original vibration signal is converted into a two-dimensional time-frequency matrix of TxF, and the time dimension identifier and the frequency dimension identifier of the matrix are stored.

14. A timing model based offshore wind turbine structure vibration prediction system according to claim 12, wherein, ​ 15. A computer-readable storage medium, characterized in that, ​

Images