Failure evaluation system based on offshore wind turbine generator

Abstract

The invention relates to the technical field of offshore wind turbine generator operation and maintenance, and discloses a failure evaluation system based on an offshore wind turbine generator. Comprising a sensor data acquisition module, a data preprocessing module, a failure feature extraction unit, a failure probability calculation module, a maintenance decision module and the like. The system obtains parameters such as vibration and temperature of the wind turbine generator in real time, after preprocessing, failure features are extracted through a deep convolutional neural network, and a dynamic Bayesian network model is constructed to calculate the component failure probability. An optimal maintenance path is generated by using an improved ant colony optimization algorithm, and meanwhile, a dynamic risk assessment module, an abnormal signal detection module and a maintenance strategy optimization module are arranged. The method can accurately evaluate the unit failure risk, timely discover abnormity, optimize the maintenance decision, improve the operation reliability of the offshore wind turbine generator, reduce the operation and maintenance cost, and guarantee the stability and high efficiency of offshore wind power generation.

Description

Technical Field

[0001] The present invention relates to the technical field of offshore wind turbine operation and maintenance, and in particular to a failure assessment system based on offshore wind turbines. Background Art

[0002] With the growing global demand for clean energy, offshore wind power generation, as a key renewable energy source, has garnered widespread attention. Compared to onshore wind turbines, offshore wind turbines face more complex and harsh operating environments, such as strong winds, wave impacts, and salt spray erosion. This results in higher failure rates and greater operational and maintenance challenges.

[0003] Offshore wind turbines are exposed to complex environments for extended periods, subjecting key components to immense mechanical stress and environmental corrosion. Blades, for example, must withstand not only the aerodynamic forces generated by strong winds but also the impact of splashing waves and the erosion of salt spray, making them susceptible to cracks, wear, and other faults. Under high loads, frequent starts and stops, and the influence of high seawater humidity, the gearbox's lubrication system is prone to failure, leading to gear wear and tooth bonding. Generators also face risks such as heat dissipation difficulties and electrical insulation damage from salt spray. Once these failures occur, they can not only cause the wind turbine to shut down, impacting power supply, but also incur high repair costs.

[0004] Currently, the operation and maintenance of offshore wind turbines mostly rely on traditional scheduled maintenance and post-fault repair models. Scheduled maintenance is often performed at fixed intervals or operating hours, and it is impossible to accurately adjust maintenance strategies based on the actual operating status of the turbine. This leads to over-maintenance of some components in good condition, wasting significant manpower, material resources, and time. Meanwhile, some components with potential faults are not discovered and addressed in a timely manner, increasing the risk of sudden failures. Post-fault repair, however, is performed only after a fault has occurred, during which the wind turbine is shut down, resulting in significant power generation losses. For example, a serious gearbox failure can cause a wind turbine to be shut down for weeks or even months, resulting in significant economic losses for the wind farm.

[0005] Existing failure assessment methods lack accuracy and timeliness. Most rely solely on a single type of sensor data, failing to fully reflect the unit's operating status. Furthermore, data processing and analysis methods are relatively simplistic, making it difficult to accurately extract failure characteristics from complex monitoring data. This results in significant errors in the assessment of component failure probabilities and fails to provide a reliable basis for operational and maintenance decisions. Furthermore, the lack of effective optimization algorithms for maintenance route planning and decision-making makes it difficult to comprehensively consider multiple factors, such as unit location distribution, failure frequency, and environmental factors, making it difficult to develop efficient and economical maintenance plans. Summary of the Invention

[0006] The purpose of the present invention is to provide a failure assessment system based on offshore wind turbines to solve the problems raised in the above background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a failure assessment system for offshore wind turbines, the system comprising:

[0008] The sensor data acquisition module is used to obtain the vibration, temperature, speed and environmental parameters of the wind turbine in real time and mark the parameters as raw monitoring data;

[0009] A data preprocessing module is used to perform noise reduction processing on the original monitoring data to generate a standardized data sequence;

[0010] A failure feature extraction unit, used to extract a multi-dimensional failure feature vector from a standardized data sequence based on a deep convolutional neural network;

[0011] A failure probability calculation module, configured to construct a dynamic Bayesian network model based on the multidimensional failure feature vector and output component failure probability;

[0012] Maintenance decision module, used to generate the optimal maintenance path based on the improved ant colony optimization algorithm and send the decision instructions to the execution terminal;

[0013] The improved ant colony optimization algorithm is specifically improved as follows:

[0014] Extract unit location distribution and fault frequency from historical maintenance records and initialize the pheromone concentration matrix;

[0015] Introducing a dynamic volatilization factor to adjust the pheromone volatilization rate according to the ambient wind speed level;

[0016] If the path selection probability exceeds the preset threshold, the path is determined to be a valid maintenance plan.

[0017] Preferably, the specific method of the data preprocessing module is:

[0018] The original monitoring data is decomposed in the frequency domain by wavelet transform algorithm to separate high-frequency noise and low-frequency effective signals;

[0019] Filter the high-frequency noise components based on the adaptive threshold and reconstruct the denoised time domain signal;

[0020] The Z-score normalization method is used to normalize the reconstructed signal to generate a standardized data sequence.

[0021] Preferably, the failure feature extraction unit includes:

[0022] The time series segmentation submodule is used to divide the standardized data sequence into multiple data segments according to a fixed time window;

[0023] The convolution kernel generation submodule is used to optimize the convolution kernel parameters based on the gradient descent algorithm to match the vibration modal characteristics of the unit;

[0024] The feature fusion submodule is used to weightedly fuse multi-scale convolution outputs through the attention mechanism to generate a multi-dimensional failure feature vector.

[0025] Preferably, the dynamic Bayesian network model is constructed as follows:

[0026] Extract prior probability distribution and conditional dependencies from the component fault database to construct the network topology;

[0027] Iteratively update the node conditional probability table based on the expectation maximization algorithm to optimize network parameters;

[0028] The multi-dimensional failure feature vector is input into the network, and the posterior probability of component failure is calculated through the probability propagation algorithm.

[0029] Preferably, the dynamic volatility factor is adjusted by:

[0030] Obtain real-time wind speed data from the environmental monitoring module and calculate the wind speed fluctuation coefficient;

[0031] Based on the Sigmoid function, the wind speed fluctuation coefficient is mapped to the volatility factor range. The formula is:

[0032]

[0033] Among them, β(t) is the volatility factor at time t, v(t) is the real-time wind speed, and k and v0 are preset adjustment parameters.

[0034] Preferably, the system further comprises:

[0035] Dynamic risk assessment module, used to generate real-time risk levels based on the unit's operating status;

[0036] The method for generating the real-time risk level is:

[0037] Obtain component failure probability from the failure probability calculation module and construct a fuzzy membership function;

[0038] The membership is weighted and aggregated based on the fuzzy inference rule base to output the quantitative value of the risk level.

[0039] Preferably, the dynamic risk assessment module further includes:

[0040] The risk mitigation submodule is used to trigger graded warning signals according to risk levels;

[0041] The triggering method of the graded warning signal is:

[0042] If the risk level quantified value exceeds a first threshold, a local shutdown command is activated and a standby unit is started;

[0043] If the risk level quantification value exceeds the second threshold, an emergency repair request is sent to the monitoring center.

[0044] Preferably, the system further comprises:

[0045] Abnormal signal detection module, used to identify sudden abnormal signals based on the improved local outlier factor algorithm;

[0046] The specific implementation steps of the improved local outlier factor algorithm are:

[0047] Extract time domain statistics and frequency domain energy ratio from the standardized data sequence as detection features;

[0048] Dynamically adjust the neighborhood radius parameters according to the unit operation stage and calculate the local outlier factor score;

[0049] If the score exceeds the adaptive threshold, the current data segment is marked as abnormal.

[0050] Preferably, the method for determining the adaptive threshold is:

[0051] Calculate the mean and variance of the outlier factor scores from historical normal data;

[0052] Set initial thresholds based on principles and update the threshold range based on real-time data distribution.

[0053] Preferably, the system further comprises:

[0054] Maintenance strategy optimization module, used to iteratively update maintenance decision logic based on genetic algorithm;

[0055] The iterative method of the genetic algorithm is:

[0056] Encode the historical maintenance path as a chromosome population, and define the path length and risk value as the fitness function;

[0057] Retain high fitness individuals through tournament selection strategy, and perform crossover and mutation operations;

[0058] If the new generation population converges to the Pareto frontier, the optimal maintenance strategy set is output.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] The offshore wind turbine failure assessment system of the present invention has significant benefits in many aspects. In terms of data acquisition and processing, the sensor data acquisition module can obtain a wide range of wind turbine operating and environmental parameters in real time, covering vibration, temperature, speed, and other parameters, providing a solid data foundation for a comprehensive understanding of the turbine status. The data preprocessing module uses a wavelet transform algorithm, adaptive threshold filtering, and Z-score normalization to effectively remove noise interference and unify data scales, making subsequent analysis more accurate and improving data availability.

[0061] The failure feature extraction unit leverages a deep convolutional neural network to accurately extract multidimensional failure feature vectors from standardized data sequences. The time series segmentation submodule rationally divides data segments, the convolution kernel generation submodule optimizes kernel parameters to match the unit's vibration modal characteristics, and the feature fusion submodule utilizes an attention mechanism to weightedly fuse multi-scale convolution outputs. This multi-step collaborative approach significantly improves the accuracy and comprehensiveness of failure feature extraction, enabling earlier and more accurate detection of potential fault hazards.

[0062] The dynamic Bayesian network model constructed by the failure probability calculation module extracts the prior probability distribution and conditional dependencies from the component fault library to construct a topological structure. It optimizes network parameters through the expectation maximization algorithm and inputs multi-dimensional failure feature vectors to calculate the posterior probability of component failure. Compared with traditional evaluation methods, it can more accurately assess the component failure probability and provide a reliable basis for operation and maintenance decisions.

[0063] The maintenance decision module is based on an improved ant colony optimization algorithm. It initializes the pheromone concentration matrix from historical maintenance records, introduces a dynamic volatility factor to adjust the pheromone volatility rate according to the ambient wind speed level, and determines the effective maintenance plan by judging the path selection probability. It can quickly plan the optimal maintenance path, significantly improve maintenance efficiency, and reduce maintenance time and costs.

[0064] The dynamic risk assessment module generates a real-time risk rating based on the unit's operating status. It quantitatively assesses unit risk by constructing a fuzzy membership function and weighted aggregation based on a fuzzy inference rule base. The risk mitigation submodule triggers graded warning signals based on the risk level. When the risk is low, it initiates a local shutdown command and activates a backup unit. When the risk is high, it sends an emergency repair request to the monitoring center, effectively ensuring the safe and stable operation of the unit and reducing losses caused by failures.

[0065] The abnormal signal detection module, based on an improved local outlier factor algorithm, extracts multi-dimensional detection features from standardized data sequences, dynamically adjusts neighborhood radius parameters to calculate local outlier factor scores, and uses adaptive thresholds to determine abnormal conditions. This allows for timely detection of sudden abnormal signals and proactive measures to prevent failures. The maintenance strategy optimization module utilizes a genetic algorithm to iteratively update maintenance decision logic, encoding historical maintenance paths as chromosome populations. Using path length and risk value as fitness functions, the module searches for the optimal maintenance strategy set on the Pareto front through selection, crossover, and mutation operations, further enhancing the scientific and economic efficiency of maintenance decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a working principle diagram of the offshore wind turbine failure assessment system according to the present invention;

[0067] Figure 2 This is the working principle diagram of the failure feature extraction unit;

[0068] Figure 3 A diagram showing the working principle of dynamic risk assessment and risk mitigation;

[0069] Figure 4 Schematic diagram of working principle optimized for maintenance strategy. DETAILED DESCRIPTION

[0070] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0071] See also Figures 1-4 The present invention provides a failure assessment system for offshore wind turbines. The system primarily comprises a sensor data acquisition module, a data preprocessing module, a failure feature extraction unit, a failure probability calculation module, and a maintenance decision module. These modules work together to effectively assess offshore wind turbine failures and make maintenance decisions. The overall implementation scheme is as follows:

[0072] The sensor data acquisition module is responsible for acquiring real-time wind turbine vibration, temperature, speed, and environmental parameters. These parameters constitute the raw monitoring data. This module utilizes high-precision sensors. For example, vibration sensors are installed in key wind turbine components, such as blades, gearboxes, and generators, accurately capturing vibrations during operation. Temperature sensors are located in heat-generating areas to monitor component temperature fluctuations. A speed sensor is installed on the rotating shaft to measure the wind turbine's speed in real time. Environmental monitoring equipment also collects environmental parameters, including wind speed, direction, temperature, and air pressure, to ensure comprehensive and accurate raw data.

[0073] The data preprocessing module receives the raw monitoring data from the sensor data acquisition module and performs noise reduction on it to generate a standardized data sequence. First, the raw monitoring data is decomposed in the frequency domain using a wavelet transform algorithm to separate high-frequency noise from the low-frequency, valid signal. Then, the high-frequency noise component is filtered using an adaptive threshold to remove noise interference and reconstruct the denoised time domain signal. Finally, the reconstructed signal is normalized using the Z-score method to make different types of data comparable and facilitate subsequent analysis.

[0074] The failure feature extraction unit extracts multidimensional failure feature vectors from a standardized data sequence based on a deep convolutional neural network. This unit first divides the standardized data sequence into multiple data segments based on fixed time windows, preparing for the subsequent convolution operation. Next, it optimizes the convolution kernel parameters using a gradient descent algorithm to better match the unit's vibration modal characteristics. Finally, it uses an attention mechanism to weightedly fuse the multi-scale convolution outputs to produce a multidimensional failure feature vector containing rich information.

[0075] The failure probability calculation module constructs a dynamic Bayesian network model based on the multidimensional failure feature vector and outputs the component failure probability. This module extracts prior probability distributions and conditional dependencies from the component fault database to construct the network topology. It then iteratively updates the node conditional probability table using the expectation-maximization algorithm to optimize network parameters. Finally, the multidimensional failure feature vector is input into the network, and the posterior probability of component failure is calculated using a probability propagation algorithm.

[0076] The maintenance decision module generates the optimal maintenance path based on an improved ant colony optimization algorithm and sends the decision instructions to the execution terminal. The improved ant colony optimization algorithm extracts unit location distribution and fault frequency from historical maintenance records to initialize the pheromone concentration matrix. A dynamic volatility factor is introduced to adjust the pheromone volatility rate based on the ambient wind speed level. If the path selection probability exceeds a preset threshold, the path is considered a valid maintenance plan. This method enables the rapid and accurate planning of the optimal maintenance path, improving maintenance efficiency.

[0077] The implementation of the present invention will be further described below with reference to Examples 1 to 6.

[0078] Example 1:

[0079] In this embodiment, the specific implementation of the data preprocessing module and the failure feature extraction unit is further described.

[0080] The wavelet transform algorithm plays a key role in processing raw monitoring data in the data preprocessing module. Wavelet transform is a time-frequency analysis method that can analyze signals at different time and frequency resolutions. Raw monitoring data from offshore wind turbines contains signals with various frequency components, including low-frequency, effective signals reflecting the normal operation of the turbines, as well as high-frequency noise generated by environmental interference and equipment noise.

[0081] By performing frequency-domain decomposition on the raw monitoring data using a wavelet transform algorithm, these signals of varying frequencies can be separated. For example, during the decomposition process, an appropriate wavelet basis function can be selected, such as the db4 wavelet basis, which exhibits excellent time-frequency localization and is well suited to the characteristics of wind turbine data. This decomposition yields coefficients for different frequency subbands. The coefficients of the high-frequency subbands primarily contain noise information, while the coefficients of the low-frequency subbands retain valid information about turbine operation.

[0082] When filtering high-frequency noise components using an adaptive threshold, the threshold needs to be dynamically determined based on the characteristics of the data. A common approach is to determine the threshold based on the statistical characteristics of the noise. For example, the standard deviation of the high-frequency subband coefficients is calculated and the threshold is determined according to a specific empirical formula. Assuming the standard deviation of the high-frequency subband coefficients is σ, the threshold T = kσ, where k is an empirical constant typically between 2 and 3. This approach effectively removes high-frequency noise while preserving the characteristics of the low-frequency, significant signal.

[0083] Reconstructing the de-noised time-domain signal involves performing an inverse wavelet transform on the filtered high- and low-frequency subband coefficients. This transform converts the frequency-domain signal back to the time domain, yielding the de-noised signal. During the reconstruction process, the accuracy and stability of the transform must be ensured to ensure that the de-noised signal truly reflects the unit's operating status.

[0084] The Z-score normalization method is used to normalize the reconstructed signal. The Z-score normalization formula is: Where X is the raw data, μ is the mean, and σ is the standard deviation. This normalization process converts data of varying magnitudes and distributions into data with a mean of 0 and a standard deviation of 1, making subsequent analysis and processing more convenient.

[0085] The time series segmentation submodule within the failure feature extraction unit divides the standardized data sequence into multiple data segments based on fixed time windows. The selection of the time window requires a comprehensive consideration of the turbine's operating characteristics and the frequency of data changes. For example, for wind turbine vibration data, which fluctuates frequently, a shorter time window, such as 1 second, can be chosen. For temperature data, which fluctuates more slowly, a longer time window, such as 10 minutes, can be chosen. By properly selecting the time window, the efficiency of subsequent processing can be improved while preserving data characteristics.

[0086] The convolution kernel generation submodule optimizes the convolution kernel parameters based on the gradient descent algorithm. Gradient descent is a commonly used optimization algorithm that continuously adjusts parameters to gradually reduce the value of an objective function. In this module, the objective function can be the error between the predicted value and the true value. The convolution kernel parameters determine the convolution operation's ability to extract data features. The gradient descent algorithm continuously optimizes the convolution kernel parameters to better match the unit's vibration modal characteristics. For example, during training, data segments are fed into the convolutional neural network, and the error between the predicted value and the true value is calculated. The convolution kernel parameters are then adjusted based on the gradient of the error, and iterations are repeated until the error reaches a minimum.

[0087] The feature fusion submodule uses an attention mechanism to weightedly fuse multi-scale convolution outputs. The attention mechanism automatically learns the importance of convolution outputs at different scales, assigning higher weights to important features. In wind turbine failure feature extraction, convolutions of different scales can capture features at different levels. For example, small-scale convolutions can capture local details, while large-scale convolutions can capture overall trends. By weightedly fusing these convolution outputs at different scales through the attention mechanism, a multi-dimensional failure feature vector containing richer information can be obtained. For example, assuming there are three convolution outputs at three scales, F1, F2, and F3, the attention mechanism calculates weights w1, w2, and w3 for each, respectively, where w1+w2+w3=1. The fused feature vector F=w1F1+w2F2+w3F3. This emphasizes important features and improves the accuracy of failure feature extraction.

[0088] Example 2:

[0089] This embodiment describes in detail a method for constructing a dynamic Bayesian network model.

[0090] When building a dynamic Bayesian network model, the network topology is constructed by first extracting prior probability distributions and conditional dependencies from a component fault database. This database is obtained by collecting, organizing, and analyzing historical failure data from a large number of wind turbines. For example, for a wind turbine's gearbox, the database records the probability of gearbox failure under different operating conditions, as well as the correlation between gearbox failures and other component failures.

[0091] When extracting the prior probability distribution, we assume that the probability of a gearbox failure under normal operating conditions is P(G=1), where G represents the gearbox's fault state, G=1 indicating a failure, and G=0 indicating normal operation. The specific value of P(G=1) can be obtained through statistical analysis of historical data. Furthermore, it is necessary to extract the conditional dependencies between gearbox failure and other component failures, such as generator failure E. These dependencies are P(G=1|E=1) and P(G=1|E=0), representing the probability of a gearbox failure under generator failure and normal operation, respectively.

[0092] Based on these prior probability distributions and conditional dependencies, a network topology is constructed. The nodes in the network represent the components of the wind turbine, and the edges represent the dependencies between the components. For example, if a gearbox failure is dependent on a generator failure, the nodes representing the gearbox and generator are connected in the network and labeled with the corresponding conditional probabilities.

[0093] After the network topology is constructed, the node conditional probability table is iteratively updated to optimize network parameters using the expectation-maximization algorithm. The expectation-maximization algorithm consists of two steps: the E-step and the M-step. In the E-step, the posterior probability of each node is calculated based on the current network parameters. For example, for node G, given the observation data O, P(G|O) is calculated. In the M-step, the conditional probability table for each node is updated based on the posterior probability calculated in the E-step. Through continuous iteration of the E-step and the M-step, the network parameters gradually converge to the optimal values.

[0094] The multidimensional failure feature vector is input into the network, and the posterior probability of component failure is calculated through the probability propagation algorithm. There are many probability propagation algorithms, such as the joint tree algorithm and the variable elimination method. Taking the joint tree algorithm as an example, the dynamic Bayesian network is first converted into a joint tree structure, and then message passing is performed in the joint tree. The multidimensional failure feature vector is input into the joint tree as observation data, and the posterior probability of each component node is calculated through message passing. For example, the posterior probability P(G=1|O) of a gearbox failure is calculated, where O is the observation data represented by the multidimensional failure feature vector. In this way, the failure probability of a component can be accurately assessed, providing a basis for subsequent maintenance decisions.

[0095] Example 3:

[0096] This embodiment mainly introduces the method for adjusting the dynamic volatility factor.

[0097] The dynamic volatilization factor plays an important role in the improved ant colony optimization algorithm. It can adjust the pheromone volatilization rate according to the environmental wind speed level, thereby improving the search efficiency and accuracy of the algorithm.

[0098] Obtain real-time wind speed data from the environmental monitoring module and calculate the wind speed fluctuation coefficient. There are many ways to calculate the wind speed fluctuation coefficient, for example, calculating the ratio of the standard deviation of the wind speed to the mean over a period of time. Assuming that the wind speed data is v(t) in the time interval [t1, t2], the wind speed mean is Wind speed standard deviation Wind speed fluctuation coefficient

[0099] Based on the Sigmoid function, the wind speed fluctuation coefficient is mapped to the volatility factor range. The formula is: Where β(t) is the volatility factor at time t, which determines the pheromone's volatility rate. v(t) is the real-time wind speed, obtained from the environmental monitoring module. k and v0 are preset adjustment parameters. k controls the slope of the Sigmoid function, affecting the sensitivity of the volatility factor to wind speed changes; v0 is a reference wind speed value used to adjust the center position of the Sigmoid function.

[0100] For example, when wind speed fluctuations are large (i.e., the wind speed fluctuation coefficient ω is large), indicating unstable ambient wind speeds, the volatility factor β(t) calculated using the Sigmoid function will increase accordingly. A larger volatility factor means faster pheromone evaporation, which allows the ant colony to adapt more quickly to environmental changes and prevents the algorithm from falling into a local optimal solution. Conversely, when wind speed fluctuations are small, the volatility factor β(t) decreases, slowing the pheromone's evaporation. This facilitates the algorithm's refined search in a local area and helps it find a more optimal maintenance path.

[0101] In practical applications, the preset adjustment parameters k and v0 need to be appropriately adjusted based on the wind turbine's operating environment and historical data. Through multiple experiments and data analysis, an optimal set of parameter values ​​can be found to better adapt the dynamic volatility factor to varying wind speeds, improving the performance of the improved ant colony optimization algorithm and more accurately generating the optimal maintenance path.

[0102] Example 4:

[0103] This embodiment describes in detail the specific implementation of the dynamic risk assessment module and the risk mitigation submodule it contains.

[0104] The dynamic risk assessment module generates a real-time risk level based on the turbine's operating status. First, the failure probability of components is obtained from the failure probability calculation module, and a fuzzy membership function is constructed. The failure probability of different wind turbine components, such as blades, gearboxes, and generators, has varying degrees of impact on the overall turbine risk. For example, blade failure can cause severe turbine vibration or even shutdown, significantly impacting the turbine's risk. Meanwhile, failure of some auxiliary components has a relatively minor impact on the turbine's risk.

[0105] For each component, a corresponding fuzzy membership function is constructed according to the size of its failure probability. Taking the gearbox as an example, assuming that the failure probability of the gearbox is P, a triangular fuzzy membership function can be constructed. When P is less than a lower threshold P low When P is in P, the membership degree is 0, indicating that the risk is very low; low and P mid When P is equal to P, the membership increases linearly; when P is equal to P mid When P is in P mid and P high When P is greater than P high When , the membership is 0. Here P low 、P mid 、P high It is a threshold determined based on historical data and experience.

[0106] Based on a fuzzy inference rule base, membership degrees are weighted and aggregated to produce a quantitative risk level. This fuzzy inference rule base is built based on expert experience and extensive historical data. For example, if multiple key components have high membership degrees, the risk level should be increased accordingly. However, if only a few auxiliary components have high membership degrees, the increase in the risk level is relatively small. This weighted aggregation method comprehensively considers the fuzzy membership degrees of each component to produce a quantitative value that reflects the overall risk status of the unit.

[0107] The risk mitigation submodule triggers graded warning signals based on the risk level. If the quantified risk level exceeds a first threshold, a local shutdown command is activated and the backup unit is activated. This first threshold is determined based on the unit's safe operating standards and historical failure data. When the quantified risk level exceeds this threshold, it indicates a potential safety hazard, potentially causing serious damage to the unit. At this point, a local shutdown command is activated, halting operation of the unit where the faulty component is located to prevent further escalation of the fault. Simultaneously, the backup unit is activated to ensure that the entire wind farm's power output is not significantly impacted.

[0108] If the risk level exceeds the second threshold, an emergency repair request is sent to the monitoring center. The second threshold is typically higher than the first. When the risk level exceeds this threshold, it indicates the unit is in serious danger and requires immediate repair. Sending an emergency repair request to the monitoring center allows them to promptly dispatch maintenance personnel to perform on-site repairs, minimizing downtime and reducing financial losses.

[0109] Through the collaborative work of the dynamic risk assessment module and the risk mitigation sub-module, potential risks of wind turbines can be discovered in a timely manner, and corresponding measures can be taken to deal with them, thereby ensuring the safe and stable operation of wind turbines.

[0110] Example 5:

[0111] This embodiment focuses on the abnormal signal detection module and the method for determining the adaptive threshold.

[0112] The abnormal signal detection module identifies sudden abnormal signals based on the improved local outlier factor algorithm. First, the time domain statistics and frequency domain energy ratio are extracted from the standardized data sequence as detection features. The time domain statistics can include mean, standard deviation, peak value, etc. For example, the mean μ and standard deviation σ of the standardized data sequence within a certain time window are calculated. These statistics can reflect the central tendency and dispersion of the data. The frequency domain energy ratio is calculated by Fourier transforming the data to calculate the proportion of energy in different frequency bands to the total energy. For example, the frequency range is divided into multiple sub-bands, and the energy E of each sub-band is calculated. i , then the frequency domain energy ratio is Where n is the number of sub-bands.

[0113] The neighborhood radius parameter is dynamically adjusted based on the turbine's operating stage to calculate the local outlier factor score. Data distribution characteristics vary across different wind turbine operating stages. For example, the data's patterns of change and fluctuation range differ between the startup and stable operation phases. During the startup phase, as turbine parameters gradually stabilize, data fluctuations are greater. In this case, the neighborhood radius parameter can be appropriately increased. During the stable operation phase, when data is relatively stable, the neighborhood radius parameter can be appropriately decreased.

[0114] The calculation of the local outlier factor score is based on the distance and density relationship between the data point and other data points in its neighborhood. Assume that there are k data points in the neighborhood of data point x, and calculate the distance d(x,x) from x to these k data points. i ), i = 1, 2, ..., k. Then the local reachability density lrd(x) of the data points in the neighborhood is calculated, and the local outlier factor score LOF(x) is the average ratio of the local reachability density of x to the local reachability density of other data points in the neighborhood.

[0115] If the score exceeds the adaptive threshold, the current data segment is marked as abnormal. The adaptive threshold is determined by calculating the mean μ of the outlier factor score from the historical normal data. LOF and variance σ LOF The initial threshold is set based on the 3σ principle, that is, the initial threshold T0 = μ LOF +3σ LOF The 3σ principle is a commonly used statistical method. Under normal circumstances, the probability that a data point falls within the range of the mean plus or minus 3 times the standard deviation is about 99.7%, so data points outside this range are considered abnormal.

[0116] Sliding update threshold range based on real-time data distribution. As wind turbines operate, the distribution of data may change. For example, due to factors such as equipment aging and environmental changes, the fluctuation of data may gradually increase. Therefore, it is necessary to continuously update the threshold range based on real-time data. The mean and variance of the outlier factor scores of historical normal data can be recalculated at regular intervals, and then the threshold is updated. Assuming that the current time is t, and the time of the last threshold update is t-Δt, during the period [t-Δt, t], new normal data is collected and the mean is recalculated. and variance The updated threshold By dynamically updating the threshold in this way, we can better adapt to data changes and improve the accuracy of abnormal signal detection.

[0117] Example 6:

[0118] This embodiment mainly provides a detailed description of the maintenance strategy optimization module and the genetic algorithm on which it is based.

[0119] The maintenance strategy optimization module uses genetic algorithms to iteratively update the maintenance decision logic, aiming to find a better set of maintenance strategies, thereby improving the maintenance efficiency of offshore wind turbines, reducing operation and maintenance costs, and ensuring stable operation of the units.

[0120] The iterative process of the genetic algorithm begins by encoding historical maintenance paths into a chromosome population. Historical maintenance paths encompass the series of steps and sequences of operations performed during previous wind turbine maintenance. Converting these paths into chromosomes facilitates the genetic algorithm's operation. For example, using integer encoding, each maintenance operation in the maintenance path is assigned an integer, and the chromosomes are arranged in the order in which the operations are performed. Assuming a wind turbine has maintenance operations such as blade inspection, gearbox maintenance, and generator overhaul, represented by 1, 2, and 3, respectively, a maintenance path of "blade inspection first, generator overhaul second, and gearbox maintenance last" can be encoded as [1, 3, 2]. By encoding a large number of historical maintenance paths, an initial chromosome population is formed.

[0121] Define the path length and risk value as the fitness function. The path length reflects the complexity of the maintenance process and the time cost required. A shorter path means a more efficient maintenance process. The risk value is related to the failure probability of the wind turbine components. Considering that during the maintenance process, if the components with a high failure probability are maintained first, the overall risk of the unit can be reduced. For example, according to the failure probability of the component obtained by the failure probability calculation module, a risk weight is assigned to each component. Assuming that the failure probabilities of the blades, gearboxes, and generators are P1, P2, and P3 respectively, and the corresponding risk weights are w1, w2, and w3, the maintenance order of each component in the maintenance path is [i1,i2,…,i n], then the risk value calculation can be The path length is Length = n. The fitness function Fitness can comprehensively consider the path length and risk value, such as Here, α is a weight coefficient with a value between 0 and 1, which is used to balance the impact of path length and risk value on fitness. It can be adjusted according to actual needs. For example, through multiple experiments, the α value that achieves the best maintenance effect can be found.

[0122] A tournament selection strategy is used to retain individuals with high fitness and perform crossover and mutation operations. This strategy randomly selects a certain number of individuals from the population, such as k (k is typically 3-5), and selects the individuals with the highest fitness from these k individuals to enter the next generation. This process is repeated until the number of individuals in the next generation reaches a set value. This selection method ensures that individuals with high fitness have a greater chance of being retained, promoting the evolution of the population towards a more optimal state.

[0123] The crossover operation simulates the gene exchange process in biological inheritance. Randomly select two chromosomes from the current population as parents, for example, select chromosomes A = [1, 2, 3, 4] and B = [5, 6, 7, 8], and randomly determine a crossover point. Assume that the crossover point is 2. Then the offspring chromosome A generated after the crossover will be ′ =[1,2,7,8], B ′ =[5,6,3,4]. Through the crossover operation, excellent genes between different chromosomes can be combined, which may produce a more optimal maintenance path.

[0124] Mutation prevents the algorithm from converging prematurely and maintains population diversity. Genes within a chromosome are randomly altered with a certain mutation probability (e.g., 0.01-0.1). For example, for chromosomes [1, 2, 3, 4], if the mutation probability is 0.05 and a gene locus is selected for mutation, assuming the third locus is selected, the original gene 3 may randomly change to another legal maintenance operation code. Mutation introduces new gene combinations to the population, preventing the algorithm from falling into a local optimum.

[0125] If the new generation population converges to the Pareto front, the optimal maintenance strategy set is output. The Pareto front refers to a set of solutions in a multi-objective optimization problem that cannot further optimize a certain objective without sacrificing other objectives. In maintenance strategy optimization, path length and risk value are two conflicting objectives. When the population evolves to a certain level and a solution that can no longer optimize both objectives can be found, the population is considered to have converged to the Pareto front. At this point, the solutions on the Pareto front are the optimal maintenance strategies under different trade-offs. These strategies are grouped into a set and output. Operations and maintenance personnel can select the most appropriate maintenance strategy from the set based on actual conditions, such as maintenance resources and time constraints, to achieve more scientific and efficient maintenance of offshore wind turbines.

[0126] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0127] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A failure assessment system based on offshore wind turbines, characterized in that: include: The sensor data acquisition module is used to obtain the vibration, temperature, speed and environmental parameters of the wind turbine in real time and mark the parameters as raw monitoring data; A data preprocessing module is used to perform noise reduction processing on the original monitoring data to generate a standardized data sequence; A failure feature extraction unit, used to extract a multi-dimensional failure feature vector from a standardized data sequence based on a deep convolutional neural network; A failure probability calculation module, configured to construct a dynamic Bayesian network model based on the multidimensional failure feature vector and output component failure probability; Maintenance decision module, used to generate the optimal maintenance path based on the improved ant colony optimization algorithm and send the decision instructions to the execution terminal; The improved ant colony optimization algorithm is specifically improved as follows: Extract unit location distribution and fault frequency from historical maintenance records and initialize the pheromone concentration matrix; Introducing a dynamic volatilization factor to adjust the pheromone volatilization rate according to the ambient wind speed level; If the path selection probability exceeds the preset threshold, the path is determined to be a valid maintenance plan.

2. The offshore wind turbine failure assessment system according to claim 1, characterized in that: The specific method of the data preprocessing module is: The original monitoring data is decomposed in the frequency domain by wavelet transform algorithm to separate high-frequency noise and low-frequency effective signals; Filter the high-frequency noise components based on the adaptive threshold and reconstruct the denoised time domain signal; The Z-score normalization method is used to normalize the reconstructed signal to generate a standardized data sequence.

3. The offshore wind turbine failure assessment system according to claim 2, characterized in that: The failure feature extraction unit includes: The time series segmentation submodule is used to divide the standardized data sequence into multiple data segments according to a fixed time window; The convolution kernel generation submodule is used to optimize the convolution kernel parameters based on the gradient descent algorithm to match the vibration modal characteristics of the unit; The feature fusion submodule is used to weightedly fuse multi-scale convolution outputs through the attention mechanism to generate a multi-dimensional failure feature vector.

4. The offshore wind turbine failure assessment system according to claim 3, characterized in that: The construction method of the dynamic Bayesian network model is: Extract prior probability distribution and conditional dependencies from the component fault database to construct the network topology; Iteratively update the node conditional probability table based on the expectation maximization algorithm to optimize network parameters; The multi-dimensional failure feature vector is input into the network, and the posterior probability of component failure is calculated through the probability propagation algorithm.

5. The offshore wind turbine failure assessment system according to claim 4, characterized in that: The dynamic volatility factor is adjusted as follows: Obtain real-time wind speed data from the environmental monitoring module and calculate the wind speed fluctuation coefficient; Based on the Sigmoid function, the wind speed fluctuation coefficient is mapped to the volatility factor range. The formula is: Among them, β(t) is the volatility factor at time t, v(t) is the real-time wind speed, and k and v0 are preset adjustment parameters.

6. The offshore wind turbine failure assessment system according to claim 1, characterized in that: Also includes: Dynamic risk assessment module, used to generate real-time risk levels based on the unit's operating status; The method for generating the real-time risk level is: Obtain component failure probability from the failure probability calculation module and construct a fuzzy membership function; The membership is weighted and aggregated based on the fuzzy inference rule base to output the quantitative value of the risk level.

7. The offshore wind turbine failure assessment system according to claim 6, characterized in that: The dynamic risk assessment module also includes: The risk mitigation submodule is used to trigger graded warning signals according to risk levels; The triggering method of the graded warning signal is: If the risk level quantified value exceeds a first threshold, a local shutdown command is activated and a standby unit is started; If the risk level quantification value exceeds the second threshold, an emergency repair request is sent to the monitoring center.

8. The offshore wind turbine failure assessment system according to claim 1, characterized in that: Also includes: Abnormal signal detection module, used to identify sudden abnormal signals based on the improved local outlier factor algorithm; The specific implementation steps of the improved local outlier factor algorithm are: Extract time domain statistics and frequency domain energy ratio from the standardized data sequence as detection features; Dynamically adjust the neighborhood radius parameters according to the unit operation stage and calculate the local outlier factor score; If the score exceeds the adaptive threshold, the current data segment is marked as abnormal.

9. The offshore wind turbine failure assessment system according to claim 8, characterized in that: The method for determining the adaptive threshold is: Calculate the mean and variance of the outlier factor scores from historical normal data; Set initial thresholds based on principles and update the threshold range based on real-time data distribution.

10. The offshore wind turbine failure assessment system according to claim 1, characterized in that: Also includes: Maintenance strategy optimization module, used to iteratively update maintenance decision logic based on genetic algorithm; The iterative method of the genetic algorithm is: Encode the historical maintenance path as a chromosome population, and define the path length and risk value as the fitness function; Individuals with high fitness are retained through the tournament selection strategy, and crossover and mutation operations are performed; if the new generation population converges to the Pareto front, the optimal maintenance strategy set is output.

Images