Method for monitoring stability of evaporative cooling heat dissipation system of wind driven generator
By combining variational mode decomposition with particle swarm optimization and a prediction model fused with KAN-Transformer network, the instability problem of the boiling cooling system of wind turbine generators is solved, achieving efficient fault detection and real-time early warning, and ensuring stable system operation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-08
- Publication Date
- 2026-03-24
AI Technical Summary
The existing boiling cooling system of wind turbine generators is unstable during the vapor-liquid two-phase flow process, which leads to vibration, thermal fatigue and structural damage. Existing monitoring methods are difficult to effectively identify the stability boundary and provide real-time early warning. Furthermore, the utilization efficiency of multi-source sensor data is low, and modeling methods are difficult to take into account both time-series characteristics and long-term dependence.
A variational mode decomposition method based on particle swarm optimization is used to reduce noise in multi-source sensor data. Combined with the prediction models of KAN network and Transformer network, the similarity of temperature data is calculated by dynamic time warping algorithm, and an adaptive threshold is set for fault early warning.
It improves the sensitivity and response speed of fault detection, ensures system stability, adapts to the dynamic changes in the operating status of wind turbine units, and enables early warning and real-time monitoring.
Smart Images
Figure CN121723192A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind turbine technology, specifically to a method for monitoring the stability of a wind turbine evaporative cooling system, used to achieve stability monitoring and fault early warning of the wind turbine cooling system. Background Technology
[0002] Currently, wind turbine stators mostly use air cooling or liquid cooling for heat dissipation, relying on fluid convection heat transfer. However, as the power of the units increases, traditional cooling methods are gradually becoming limited in their heat exchange capacity. In contrast, boiling heat exchange technology, due to its ability to absorb a large amount of latent heat during phase change, has higher heat dissipation efficiency and temperature uniformity, making it suitable for the cooling needs of high power density motors.
[0003] The constant temperature during boiling cooling helps remove a significant amount of heat without increasing the heat source temperature, thereby reducing motor temperature rise, improving temperature field distribution, and significantly enhancing heat dissipation performance and operational stability. This technology offers the following advantages: (1) It has strong adaptability to heat load and maintains efficient heat dissipation under different working conditions; (2) Increasing the power density of the motor allows it to withstand a larger line load, which is beneficial for reducing its size and weight; (3) The medium has high safety, low boiling point, strong insulation, low leakage risk and simple operation and maintenance.
[0004] Although boiling heat exchange technology has been widely used in various industrial fields, instability issues still exist in its vapor-liquid two-phase flow process. Fluctuations in parameters such as mass flow rate, pressure, and wall temperature can lead to vibration, thermal fatigue, and structural damage, threatening the stable operation of the system. Therefore, this invention proposes a stability monitoring method based on field operating data to identify the stability boundaries of the boiling cooling system under different operating conditions and achieve effective early warning.
[0005] Despite the use of various signal processing and modeling methods to analyze the operating status in existing wind turbine cooling system stability monitoring technologies, the following shortcomings still exist: (1) The utilization efficiency of multi-source sensor data is low and the feature extraction capability is limited. Most existing methods rely on traditional signal processing techniques, which make it difficult to fully explore the deep information in high-dimensional and multivariate monitoring data. Especially when dealing with the complex operating environment and nonlinear dynamic behavior of wind power sites, the feature extraction effect is not ideal, and redundant data interferes with the model performance, which can easily lead to misjudgment or missed judgment.
[0006] (2) Modeling methods are difficult to take into account both time-series characteristics and long-term dependencies. Traditional fault detection methods usually adopt a single time-series modeling structure, which makes it difficult to capture both short-term local changes and long-term evolution trends in the operating data at the same time. The model does not respond to changes in system state in a timely manner, resulting in delayed early warning.
[0007] (3) Insufficient online monitoring and model updating capabilities make it difficult to achieve high-precision real-time early warning. Most existing systems lack dynamic adjustment capabilities after model deployment, and cannot optimize and update the model based on real-time field data. This makes it difficult to adapt to the dynamic changes in the operating status of wind turbine units, affecting the accuracy of stability identification and the practical application value of the system. Summary of the Invention
[0008] To address the aforementioned technical problems, this invention provides a method for monitoring the stability of a wind turbine's evaporative cooling system. This invention utilizes both on-site data acquisition and model prediction data for fault detection, enabling operational monitoring and fault early warning of the motor system. The method includes: acquiring multi-source sensor data from the wind turbine generator set and using a particle swarm optimization variational mode decomposition (VMD) method to reduce noise; standardizing data length through equal-interval decimation downsampling and cubic spline interpolation; and using a prediction model based on the fusion of KAN and Transformer networks to accurately predict temperature data for a future period (e.g., within 10 minutes). By introducing a dynamic time warping (DTW) algorithm, the similarity between the spliced temperature data and historical normal data is calculated, and combined with an adaptive threshold method, faults are monitored and warned in real time. When the DTW distance of the spliced temperature data exceeds the threshold range, the system triggers a fault warning. This method enables early warning of potential faults, effectively improving the sensitivity and response speed of fault detection, and ensuring the safe and stable operation of the system.
[0009] The present invention adopts the following technical solution: A method for monitoring the stability of a wind turbine evaporative cooling system includes: Collect multi-source sensor data from the cooling system of wind turbine generator sets; Preprocessing of multi-source sensors: The variational mode decomposition method with parameters optimized by particle swarm optimization is used to reconstruct the multi-source sensor data, thereby reducing the noise of the multi-source sensor data. The length of the multi-source sensor data is unified by equal interval extraction and cubic spline interpolation. A prediction model based on the fusion of KAN and Transformer networks is used to predict temperature data for future time periods by inputting preprocessed multi-source sensor data into the prediction model. Specifically, the KAN network replaces the feedforward neural network in the traditional Transformer network and is responsible for performing nonlinear transformations on the features processed by the self-attention mechanism. An adaptive normal state threshold range is set, and the DTW distance between the stitched temperature data based on the prediction results of the prediction model and the historical normal data is calculated by the dynamic time warping algorithm. A warning is triggered when the normal state threshold range is exceeded.
[0010] In one embodiment, the multi-source sensor data includes motor stator temperature, heat dissipation system mass, flow rate and pressure, and rotor speed; wherein the temperature data refers to motor stator temperature.
[0011] In one embodiment, the reconstruction of multi-source sensor data specifically includes: The variational mode decomposition method is used to analyze the data from each sensor. Decomposed into multiple eigenmode functions with finite bandwidth Furthermore, variational constraints are used to minimize the bandwidth of each intrinsic mode function.
[0012] In one embodiment, the variational mode decomposition method is used to process the data from each sensor. Decomposed into multiple eigenmode functions with finite bandwidth Furthermore, the bandwidth of each intrinsic mode function is minimized through variational constraints, specifically including: Will Decomposed into a set of K eigenmode functions The k-th eigenmode function The center frequency is The variational constraints of the decomposition process are: , ; in, This represents the convolution operation. For impulse functions, Represents the time differential symbol. Represents the imaginary unit. Represents pi (π). Represents a time variable. Denotes the F2 norm; Introducing Lagrange multipliers and secondary penalty factor Considering the constraints of minimizing the intrinsic mode function bandwidth and signal reconstruction error, a Lagrange augmented function is constructed; the alternating direction multiplier method is used for iterative optimization, including updating the frequency domain representation of the intrinsic mode function corresponding to the sensor data, updating the center frequency of the intrinsic mode function corresponding to the sensor data, and updating the Lagrange multipliers; After completing the variational mode decomposition, it will be performed according to the center frequency. The intrinsic mode functions are arranged in order from low frequency to high frequency, and then the time-domain components of the intrinsic mode functions are linearly superimposed one by one until the signal energy of the reconstructed sensor data accounts for more than 98% of the signal energy of the original sensor data. The Lagrange augmented function for: .
[0013] In one embodiment, the frequency domain representation of the intrinsic mode function corresponding to the updated sensor data specifically includes: ; For frequency, For the first Each eigenmode function in Frequency domain representation after +1 iterations Representing sensor data Frequency domain representation after Fourier transform, For the first Frequency domain representation of the intrinsic mode functions As a weighting factor, No. The center frequencies of the intrinsic mode functions; The center frequency of the intrinsic mode function corresponding to the updated sensor data specifically includes: ; For the first Each eigenmode function in The center frequency after +1 iterations; The updated Lagrange multipliers are: ; For the Lagrange multipliers in the nth iteration, The step size parameter is used to control the update speed.
[0014] In one embodiment, the variational mode decomposition method after parameter optimization using the particle swarm optimization algorithm specifically includes: S21, define the particle swarm: N is the particle swarm size, and the i-th particle is represented as... ; This represents the number of decomposition modes of the i-th particle. This represents the penalty factor for the i-th particle; S22, Initialize particle positions: The initial position of each particle is randomly generated within the corresponding range; S23, Fitness Function Construction: Perform variational mode decomposition on the position information of each particle, and calculate the energy entropy of each particle based on the obtained intrinsic mode functions. The smaller the energy entropy, the less noise component there is in the sensor data signal. The optimization objective is... The formula for calculating energy entropy is: , ; It is the first Normalized energy distribution of each intrinsic mode function It is the frequency component of the k-th eigenmode function in the frequency domain. It is the energy distribution of the k-th eigenmode function at each frequency; S24, update particle velocity and position: ; ; ω is the inertia weight, c1 and c2 are constants, and r1 and r2 are random numbers. and These are the individual optimal particle position and the global optimal particle position, respectively. Let be the velocity of the i-th particle after the nth iteration. This represents the position of the i-th particle after the nth iteration. S25 introduces Cauchy distribution variation to perturb the current global optimal particle position, thus avoiding getting trapped in a local optimum; S26. Repeat the above steps iteratively until the maximum number of iterations is reached or the particle fitness value reaches the convergence condition, and output the optimal variational mode decomposition parameters.
[0015] In one embodiment, the method of unifying the length of multi-source sensor data using equal-interval extraction and cubic spline interpolation specifically includes: For multi-source sensor data with a sampling rate higher than the target acquisition frequency, anti-aliasing filtering is used for filtering, and the cutoff frequency of the filter is set to half of the target sampling frequency. The multi-source sensor data after anti-aliasing filtering is downsampled to reduce the sampling rate of the multi-source sensor data to the target acquisition frequency. For multi-source sensor data with a sampling rate lower than the target acquisition frequency, cubic spline interpolation is used to interpolate the multi-source sensor data to the target sampling frequency.
[0016] In one embodiment, the use of a prediction model based on the fusion of KAN and Transformer networks, inputting preprocessed multi-source sensor data into the prediction model to predict temperature data for future time periods, specifically includes: The preprocessed multi-source sensor data is input into the encoder part of the Transformer network, and then processed through a self-attention mechanism. Capture input multi-source sensor data Global dependencies: .
[0017] The output features of the Transformer network; Passed to KAN network Perform nonlinear transformation: ; in, The output features of the KAN network are further passed to the prediction layer, mapping the high-dimensional feature space to a numerical output, which outputs temperature data for future time periods. : ; Prediction layer It is a mapping function consisting of two fully connected layers.
[0018] In one embodiment, the step of calculating the DTW distance between the stitched temperature data based on the prediction model and historical normal data using a dynamic time warping algorithm, and triggering a warning when the distance exceeds the normal state threshold range, specifically includes: The collected temperature data and the predicted temperature data are stitched together to obtain stitched temperature data, which is then compared with historical normal data stored in the database. The Dynamic Time Warping (DTW) distance between the stitched temperature data and the historical normal data is calculated using a dynamic time warping algorithm. When the DTW distance exceeds the normal state threshold range, a warning is triggered.
[0019] In one embodiment, setting the adaptive normal state threshold range specifically includes: The Dynamic Time Warping (DTW) distance between multiple sets of stitched temperature data and historical normal data was calculated using the Dynamic Time Warping (DTW) algorithm, and the mean value of all DTW distances was calculated. and standard deviation Calculate the initial threshold interval; the initial threshold interval for normal state is... ; The mean and standard deviation of the DTW distance between the spliced temperature data and historical normal data are updated in real time using a recursive algorithm during the operation of the wind turbine generator cooling system. The recursive formula for the mean is: ; The DTW distance between the current stitched temperature data and the historical normal data. This represents the average DTW distance between all spliced temperature data and historical normal data at the previous moment. The current moment; Variance recursive formula: ; This is the variance of the DTW distance between all spliced temperature data at the current moment and the historical normal data. The feature value of the DTW distance between all spliced temperature data at the current moment and the historical normal data; The normal state threshold range is .
[0020] Compared with the prior art, the beneficial technical effects of the present invention are: 1. This invention combines particle swarm optimization (PSO) and variational mode decomposition (VMD). By automatically selecting the optimal parameters of VMD, it can effectively extract key information from signals, remove redundant data, and avoid feature loss that may be caused by traditional methods.
[0021] 2. This invention integrates the KAN network and the Transformer network into the Kansformer prediction model, which improves the ability of the time series prediction model to handle long-term dependencies and short-term fluctuations.
[0022] 3. This invention concatenates online collected data with the output data of the prediction model, enhancing the ability to predict future trends of the data. It uses a dynamic time warping algorithm to calculate the similarity between the collected data and the fault data, and uses a recursive method to adaptively update the threshold, thereby improving the accuracy and sensitivity of fault detection.
[0023] 4. This invention has strong adaptability and engineering practice value. The method is applicable to real-time monitoring and fault prediction of wind turbines. By analyzing the data of the motor cooling system in real time, it can provide early warning of potential faults, ensure the safety and reliability of the equipment, and has the ability to dynamically adjust according to environmental changes to ensure the continuous and stable operation of the system. Attached Figure Description
[0024] Figure 1 This is a flowchart of a method in one embodiment of the present invention; Figure 2 This is a schematic diagram of the overall structure of the Transformer network in one embodiment of the present invention; Figure 3 This is a schematic diagram of the overall structure of the KAN network in one embodiment of the present invention; Figure 4 This is a schematic diagram of a single-head attention calculation process in one embodiment of the present invention; Figure 5 This is a schematic diagram of the multi-head attention calculation process in one embodiment of the present invention; Figure 6 This is a schematic diagram of the overall structure of the prediction model in one embodiment of the present invention. Detailed Implementation
[0025] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.
[0026] like Figure 1 As shown, a method for monitoring the stability of a wind turbine evaporative cooling system according to the present invention includes the following steps: S1 collects multi-source sensor data from the wind turbine generator cooling system; S2, preprocessing of multi-source sensors: using variational mode decomposition method with parameters optimized by particle swarm optimization algorithm to reconstruct multi-source sensor data, remove high-frequency noise from multi-source sensor data, and use equal interval extraction and cubic spline interpolation to unify the length of multi-source sensor data; S3 uses a prediction model based on the fusion of KAN and Transformer networks. Preprocessed multi-source sensor data is input into the prediction model to predict temperature data for future time periods. KAN network replaces the feedforward neural network in the traditional Transformer network and is responsible for performing nonlinear transformations on the features processed by the self-attention mechanism. S4 sets an adaptive normal state threshold range. The DTW distance between the stitched temperature data based on the prediction results of the prediction model and the historical normal data is calculated by the dynamic time warping algorithm. A warning is triggered when the normal state threshold range is exceeded.
[0027] In one embodiment, the multi-source sensor data includes motor stator temperature, heat dissipation system mass, flow rate and pressure, and rotor speed.
[0028] This invention uses multi-source sensors installed on wind turbine generators to collect various data during the operation of the cooling system, including temperature, pressure, flow rate, and rotational speed, with a collection time of no less than 30 minutes and continuous updates.
[0029] In one embodiment, the reconstruction of multi-source sensor data specifically includes: The variational mode decomposition method is used to analyze the data from each sensor. Decomposed into multiple eigenmode functions with finite bandwidth Furthermore, variational constraints are used to minimize the bandwidth of each intrinsic mode function.
[0030] In one embodiment, the variational mode decomposition method is used to process the data from each sensor. Decomposed into multiple eigenmode functions with finite bandwidth , representing the components of the signal at different frequency bandwidths, and minimizing the bandwidth of each intrinsic mode function through variational constraints, specifically including: Will Decomposed into a set of K eigenmode functions The k-th eigenmode function The center frequency is The variational constraints of the decomposition process are: , ; in, This represents the convolution operation. For impulse functions, Represents the time differential symbol. Represents the imaginary unit. Represents pi (π). Represents a time variable. Denotes the F2 norm; Introducing Lagrange multipliers and secondary penalty factor Considering the constraints of minimizing the intrinsic mode function bandwidth and signal reconstruction error, a Lagrange augmented function is constructed; the alternating direction multiplier method is used for iterative optimization, including updating the frequency domain representation of the intrinsic mode function corresponding to the sensor data, updating the center frequency of the intrinsic mode function corresponding to the sensor data, and updating the Lagrange multipliers; After completing the variational mode decomposition, it will be performed according to the center frequency. The intrinsic mode functions are arranged in order from low frequency to high frequency, and then the time-domain components of the intrinsic mode functions are linearly superimposed one by one until the signal energy of the reconstructed sensor data accounts for more than 98% of the signal energy of the original sensor data. The Lagrange augmented function for: .
[0031] In one embodiment, the frequency domain representation of the intrinsic mode function corresponding to the updated sensor data specifically includes: ; For frequency, For the first Each eigenmode function in Frequency domain representation after +1 iterations Representing sensor data Frequency domain representation after Fourier transform, For the first Frequency domain representation of the intrinsic mode functions As a weighting factor, No. The center frequencies of the intrinsic mode functions; The center frequency of the intrinsic mode function corresponding to the updated sensor data specifically includes: ; For the first Each eigenmode function in The center frequency after +1 iterations; The updated Lagrange multipliers are: ; For the Lagrange multipliers in the nth iteration, The step size parameter is used to control the update speed.
[0032] In one embodiment, the variational mode decomposition method after parameter optimization using the particle swarm optimization algorithm specifically includes: S21, define the particle swarm: N is the particle swarm size, and the i-th particle is represented as... ; This represents the number of decomposition modes of the i-th particle. This represents the penalty factor for the i-th particle; S22, Initialize particle positions: The initial position of each particle is randomly generated within the corresponding range; S23, Fitness Function Construction: Perform variational mode decomposition on the position information of each particle, and calculate the energy entropy of each particle based on the obtained intrinsic mode functions. The smaller the energy entropy, the less noise component there is in the sensor data signal. The optimization objective is... The formula for calculating energy entropy is: , ; It is the first Normalized energy distribution of each intrinsic mode function It is the frequency component of the k-th eigenmode function in the frequency domain. It is the energy distribution of the k-th eigenmode function at each frequency; S24, update particle velocity and position: ; ; ω is the inertia weight, c1 and c2 are constants, and r1 and r2 are random numbers. and These are the individual optimal particle position and the global optimal particle position, respectively. Let be the velocity of the i-th particle after the nth iteration. This represents the position of the i-th particle after the nth iteration. S25 introduces Cauchy distribution variation to perturb the current global optimal particle position, thus avoiding getting trapped in a local optimum; S26. Repeat the above steps iteratively until the maximum number of iterations is reached or the particle fitness value reaches the convergence condition, and output the optimal variational mode decomposition parameters.
[0033] In one embodiment, the method of unifying the length of multi-source sensor data using equal-interval extraction and cubic spline interpolation specifically includes: Anti-aliasing filtering: For multi-source sensor data with a sampling rate higher than the target acquisition frequency, anti-aliasing filtering is used for filtering. The cutoff frequency of the filter is set to half of the target sampling frequency to ensure that the signal does not alias.
[0034] The multi-source sensor data, after anti-aliasing filtering, is downsampled to reduce the sampling rate to the target acquisition frequency; the target sampling frequency is then set. The purpose of downsampling is to reduce the sampling rate of a signal to a target frequency. The downsampled signal... The calculation formula is as follows: .
[0035] For multi-source sensor data with a sampling rate lower than the target acquisition frequency, cubic spline interpolation is used to interpolate the multi-source sensor data to the target sampling frequency.
[0036] The interpolation function has the following form: ; in, It is an interpolation function. These are the coefficients of the cubic spline interpolation function.
[0037] A segment of multi-source wind turbine data collected on site , For the data dimension, sequences with sampling rates higher than the set acquisition frequency are downsampled using anti-aliasing filtering and equal-interval decimation; assuming the signal... Perform low-pass filtering; the filter's cutoff frequency is... (Generally take) the following): ; in, yes High-frequency sequences in It is the signal after anti-aliasing filtering. It is the impulse response function of the low-pass filter.
[0038] Will According to the set sampling frequency Equal-interval downsampling was performed to obtain Assume the sampling frequency is... Set the target sampling frequency as The downsampled signal for: ; in, It is the downsampled signal, i.e. , To reduce the sampling rate.
[0039] for Mid- and low-frequency sequences Perform cubic spline interpolation to the target sampling frequency. For every two adjacent sampling points and The cubic spline interpolation function is described above.
[0040] Calculate the target sampling frequency Next new time point Through interpolation function Calculate the new sequence ; Data from different sensors are merged after being processed to a uniform length and then normalized to normalize each column of data to the range [-1, 1].
[0041] The Transformer network is used to capture long-term dependencies in multi-source sensor data. Specifically, it is a deep learning architecture based on a self-attention mechanism, with a core structure consisting of an encoder and a decoder. The encoder processes the input sequence through multiple self-attention layers and a feedforward neural network, capturing global features and long-term dependencies in the input data. The decoder, based on the features output by the encoder and combined with the positional information of the target sequence, generates predictions through a similar self-attention mechanism and feedforward network. The Transformer network adds order information to the input sequence through positional encoding, enabling the model to identify the relative order of each position based on the parallel computation of the multi-head attention mechanism. The masked multi-head attention mechanism in the decoder is used to process the target sequence, ensuring that the model relies only on predictions from previous time steps.
[0042] KAN networks exhibit high interpretability when handling nonlinear relationships. Specifically, based on the Kolmogorov-Arnold theorem, KAN networks simplify the learning process by representing high-dimensional functions as a combination of multiple one-dimensional functions. Unlike traditional multilayer perceptrons (MLPs), KAN networks do not linearly combine all input features before inputting them into activation functions. Instead, they directly input activation functions into each input feature, then sum these activation functions to obtain a multidimensional space, which is then passed to the next layer.
[0043] The prediction model based on the fusion of KAN and Transformer networks specifically includes: using a KAN network to replace the feed-forward network (FFN) in the traditional Transformer, responsible for performing nonlinear transformations on the features processed by the self-attention mechanism. The KAN network is better able to handle more complex feature mappings in both the encoder and decoder. Specifically, it includes: 1) Input the preprocessed multidimensional data into the prediction model. Multidimensional data from each acquisition channel, after being standardized in length and normalized, is used as input to the KAN-Transformer fusion model. The prediction model uses 30 minutes of acquired data to predict the motor stator temperature and the stability of the cooling system within the next 10 minutes.
[0044] 2) Replace the FFN in the original Transformer with a KAN network: In the traditional Transformer network, the FFN (Feed-Forward Network) layer is used to perform non-linear transformations on the features processed by the self-attention mechanism.
[0045] 3) Input multi-source sensor data into the encoder part of the Transformer network. Transformer's self-attention mechanism. Used to capture input sequence The global dependencies in the Transformer network are used to handle temporal features in the input data. The encoder output of the Transformer network... It can be represented as: .
[0046] 4) Input the output of the Transformer network into the KAN network: The output of the Transformer encoder It is passed as input to the KAN network. The KAN network performs a nonlinear transformation on the input features, as shown in the following formula: ; in, It is the output after processing by the KAN network, which is then passed to the prediction layer.
[0047] 5) The prediction layer outputs the prediction results for future time steps. This output is a multi-dimensional time series data, representing the predicted motor stator temperature for several future time steps.
[0048] Preferably, the KAN network in the prediction model includes an input layer, a hidden layer, and an output layer. Each layer undergoes a nonlinear transformation through an activation function and passes the transformation result to the next layer, simulating the function representation process in the Kolmogorov-Arnold theorem.
[0049] During the training of the prediction model, the model is optimized by comparing the error between the predicted results and the true values. The loss function used is mean squared error (MSE).
[0050] In one embodiment, the step of determining the stability of the heat dissipation system using a dynamic time warping algorithm, setting an adaptive normal state threshold range, and triggering a warning when the normal state threshold range is exceeded specifically includes: After stitching together the collected multi-source sensor data and predicted data, the stitched temperature data is obtained. This data is then compared with historical normal data stored in the database, and a dynamic time warping algorithm is used to calculate the stitched temperature data. and historical normal data The DTW distance between them is used to trigger a warning when the DTW distance exceeds the normal state threshold range.
[0051] and DTW distance The calculation process is as follows: ; These represent the spliced temperature data at time t and the historical normal data, respectively.
[0052] The Dynamic Time Warping (DTW) algorithm finds the optimal alignment path between two time series to obtain the minimum DTW distance.
[0053] In one embodiment, setting the adaptive normal state threshold range specifically includes: The Dynamic Time Warping (DTW) distance between the stitched temperature data and historical normal data was calculated using the Dynamic Time Warping (DTW) algorithm, and the corresponding mean was also calculated. and standard deviation Calculate the initial threshold interval; the initial threshold interval for normal state is... ; The mean and standard deviation of the DTW distance between the spliced temperature data and historical normal data are updated in real time using a recursive algorithm during the operation of the wind turbine generator cooling system. The recursive formula for the mean is: ; The DTW distance between the current stitched temperature data and the historical normal data. This represents the average DTW distance between all spliced temperature data and historical normal data at the previous moment. The current moment; Standard deviation recursive formula: ; This is the variance of the DTW distance between all spliced temperature data at the current moment and the historical normal data. The feature value of the DTW distance between all spliced temperature data at the current moment and the historical normal data; The normal state threshold range is .
[0054] Furthermore, by incorporating an online stability discrimination module and employing a dynamic update strategy, this invention enables the monitoring system to adjust its prediction model in real time, rapidly responding to any abnormal changes occurring in the wind power generator system. This module not only analyzes multi-source signals from the wind power generator system in real time but also optimizes the model based on real-time data, improving the accuracy of fault prediction. Compared to traditional methods, the monitoring system of this invention can react more promptly and accurately before a fault occurs, effectively reducing false alarms and missed alarms, and enhancing the reliability and stability of fault diagnosis.
[0055] Matters not covered in this invention are common knowledge.
[0056] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.
[0057] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.
[0058] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0059] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.
[0060] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. A method for monitoring the stability of a wind turbine evaporative cooling system, characterized in that, include: Collect multi-source sensor data from the cooling system of wind turbine generator sets; Preprocessing of multi-source sensors: The variational mode decomposition method with parameters optimized by particle swarm optimization is used to reconstruct the multi-source sensor data, thereby reducing the noise of the multi-source sensor data. The length of the multi-source sensor data is unified by equal interval extraction and cubic spline interpolation. A prediction model based on the fusion of KAN and Transformer networks is used to predict temperature data for future time periods by inputting preprocessed multi-source sensor data into the prediction model. Specifically, the KAN network replaces the feedforward neural network in the traditional Transformer network and is responsible for performing nonlinear transformations on the features processed by the self-attention mechanism. An adaptive normal state threshold range is set, and the DTW distance between the stitched temperature data based on the prediction results of the prediction model and the historical normal data is calculated by the dynamic time warping algorithm. A warning is triggered when the normal state threshold range is exceeded.
2. The method for monitoring the stability of a wind turbine evaporative cooling system according to claim 1, characterized in that, The multi-source sensor data includes motor stator temperature, heat dissipation system mass, flow rate and pressure, and rotor speed; wherein, the temperature data refers to motor stator temperature.
3. The method for monitoring the stability of a wind turbine evaporative cooling system according to claim 1, characterized in that, The reconstruction of multi-source sensor data specifically includes: The variational mode decomposition method is used to analyze the data from each sensor. Decomposed into multiple eigenmode functions with finite bandwidth Furthermore, variational constraints are used to minimize the bandwidth of each intrinsic mode function.
4. The method for monitoring the stability of a wind turbine evaporative cooling system according to claim 3, characterized in that, The variational mode decomposition method is used to analyze the data from each sensor. Decomposed into multiple eigenmode functions with finite bandwidth Furthermore, the bandwidth of each intrinsic mode function is minimized through variational constraints, specifically including: Will Decomposed into a set of K eigenmode functions The k-th eigenmode function The center frequency is The variational constraints of the decomposition process are: ， ； in, This represents the convolution operation. For impulse functions, Represents the time differential symbol. Represents the imaginary unit. Represents pi (π). Represents a time variable. Denotes the F2 norm; Introducing Lagrange multipliers and secondary penalty factor Considering the constraints of minimizing the intrinsic mode function bandwidth and signal reconstruction error, a Lagrange augmented function is constructed; the alternating direction multiplier method is used for iterative optimization, including updating the frequency domain representation of the intrinsic mode function corresponding to the sensor data, updating the center frequency of the intrinsic mode function corresponding to the sensor data, and updating the Lagrange multipliers; After completing the variational mode decomposition, it will be performed according to the center frequency. The intrinsic mode functions are arranged in order from low frequency to high frequency, and then the time-domain components of the intrinsic mode functions are linearly superimposed one by one until the signal energy of the reconstructed sensor data accounts for more than 98% of the signal energy of the original sensor data. The Lagrange augmented function for: 。 5. The method for monitoring the stability of a wind turbine evaporative cooling system according to claim 4, characterized in that, The frequency domain representation of the intrinsic mode function corresponding to the updated sensor data specifically includes: ； For frequency, For the first Each eigenmode function in Frequency domain representation after +1 iterations Representing sensor data Frequency domain representation after Fourier transform, For the first Frequency domain representation of the intrinsic mode functions As a weighting factor, No. The center frequencies of the intrinsic mode functions; The center frequency of the intrinsic mode function corresponding to the updated sensor data specifically includes: ； For the first Each eigenmode function in The center frequency after +1 iteration; The updated Lagrange multipliers are: ； For the Lagrange multipliers in the nth iteration, The step size parameter is used to control the update speed.
6. The method for monitoring the stability of a wind turbine evaporative cooling system according to claim 1, characterized in that, The variational mode decomposition method optimized by the particle swarm optimization algorithm specifically includes: S21, define the particle swarm: N is the particle swarm size, and the i-th particle is represented as... ; This represents the number of decomposition modes of the i-th particle. This represents the penalty factor for the i-th particle; S22, Initialize particle positions: The initial position of each particle is randomly generated within the corresponding range; S23, Fitness Function Construction: Perform variational mode decomposition on the position information of each particle, and calculate the energy entropy of each particle based on the obtained intrinsic mode functions. The smaller the energy entropy, the less noise component there is in the sensor data signal. The optimization objective is... The formula for calculating energy entropy is: ， ； It is the first Normalized energy distribution of each intrinsic mode function It is the frequency component of the k-th eigenmode function in the frequency domain. It is the energy distribution of the k-th eigenmode function at each frequency; S24, update particle velocity and position: ； ； ω is the inertia weight, c1 and c2 are constants, and r1 and r2 are random numbers. and These are the individual optimal particle position and the global optimal particle position, respectively. Let be the velocity of the i-th particle after the nth iteration. This represents the position of the i-th particle after the nth iteration. S25 introduces Cauchy distribution variation to perturb the current global optimal particle position, thus avoiding getting trapped in a local optimum; S26. Repeat the above steps iteratively until the maximum number of iterations is reached or the particle fitness value reaches the convergence condition, and output the optimal variational mode decomposition parameters.
7. The method for monitoring the stability of a wind turbine evaporative cooling system according to claim 1, characterized in that, The method of using equidistant extraction and cubic spline interpolation to unify the length of multi-source sensor data specifically includes: For multi-source sensor data with a sampling rate higher than the target acquisition frequency, anti-aliasing filtering is used for filtering, and the cutoff frequency of the filter is set to half of the target sampling frequency. The multi-source sensor data after anti-aliasing filtering is downsampled to reduce the sampling rate of the multi-source sensor data to the target acquisition frequency. For multi-source sensor data with a sampling rate lower than the target acquisition frequency, cubic spline interpolation is used to interpolate the multi-source sensor data to the target sampling frequency.
8. The method for monitoring the stability of a wind turbine evaporative cooling system according to claim 1, characterized in that, The prediction model, which uses a fusion of KAN and Transformer networks, inputs preprocessed multi-source sensor data into the model to predict temperature data for future time periods. Specifically, this includes: The preprocessed multi-source sensor data is input into the encoder part of the Transformer network, and then processed through a self-attention mechanism. Capture input multi-source sensor data Global dependencies: ； The output features of the Transformer network; Passed to KAN network Perform nonlinear transformation: ； in, The output features of the KAN network are further passed to the prediction layer, mapping the high-dimensional feature space to a numerical output, which outputs temperature data for future time periods. : ； Prediction layer It is a mapping function consisting of two fully connected layers.
9. The method for monitoring the stability of a wind turbine evaporative cooling system according to claim 1, characterized in that, The method involves calculating the DTW distance between the stitched temperature data based on the prediction model and historical normal data using a dynamic time warping algorithm. A warning is triggered when the distance exceeds the normal state threshold range. Specifically, this includes: The collected temperature data and the predicted temperature data are stitched together to obtain stitched temperature data, which is then compared with historical normal data stored in the database. The Dynamic Time Warping (DTW) distance between the stitched temperature data and the historical normal data is calculated using a dynamic time warping algorithm. When the DTW distance exceeds the normal state threshold range, a warning is triggered.
10. The method for monitoring the stability of a wind turbine evaporative cooling system according to claim 9, characterized in that, The setting of the adaptive normal state threshold range specifically includes: The Dynamic Time Warping (DTW) distance between multiple sets of stitched temperature data and historical normal data was calculated using the Dynamic Time Warping (DTW) algorithm, and the mean value of all DTW distances was calculated. and standard deviation Calculate the initial threshold interval; the initial threshold interval for normal state is... ; The value is a constant; during the operation of the wind turbine generator cooling system, a recursive algorithm is used to update the mean and standard deviation of the DTW distance between the spliced temperature data and historical normal data in real time; the mean recursive formula is: ； The DTW distance between the current stitched temperature data and the historical normal data. This represents the average DTW distance between all spliced temperature data and historical normal data at the previous moment. The current moment; Variance recursive formula: ； This is the variance of the DTW distance between all spliced temperature data at the current moment and the historical normal data. The feature value of the DTW distance between all spliced temperature data at the current moment and the historical normal data; The normal state threshold range is .