Wind driven generator bolt loosening early warning system and method

Abstract

The invention relates to the technical field of wind driven generator bolt monitoring, and discloses a wind driven generator bolt looseness early warning system and method, and the method comprises the steps: collecting bolt stress, displacement, acoustics and other multi-dimensional data through a distributed sensor array, and inputting a deep belief network DBN model to predict a bolt looseness risk value after wavelet transform processing. And optimizing the risk assessment model through a simulated annealing algorithm, dividing risk levels, and constructing a layered response decision model to realize closed-loop management and control. The bolt loosening risk can be accurately monitored and predicted, the evaluation error is reduced, the decision and execution efficiency is improved, and continuous optimization of the system is realized. Compared with a traditional mode, the prediction accuracy is greatly improved, the maintenance time is shortened, the equipment failure rate is reduced, fault shutdown and maintenance cost are effectively reduced, safe and stable operation of the wind driven generator is guaranteed, and economic benefits and safety are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of bolt monitoring for wind turbines, and specifically to a bolt loosening warning system and method for wind turbines. Background Art

[0002] In the field of wind power generation, the stable operation of wind turbines is crucial for power supply. Bolts, as key components connecting various parts of wind turbines, their fastening states directly affect the safety and reliability of the entire equipment. However, due to the usually complex and harsh operating environment of wind turbines, bolt loosening has become a frequent and intractable problem.

[0003] From the perspective of the operating environment, wind turbines are mostly installed in open outdoor areas, and are subjected to the combined effects of strong winds, vibrations, temperature changes, humidity and other factors for a long time. Strong winds cause the wind turbine blades to rotate at high speeds, generating huge torques and alternating loads, which are transmitted to the bolts through the connecting components, causing the bolts to bear frequent tensile and shear stresses. For example, in coastal areas or mountainous areas with high wind speeds, the wind forces acting on wind turbines are more significant, and the stress changes borne by the bolts are more intense. At the same time, the temperature changes brought about by the day-night temperature difference and seasonal alternation will cause relative displacements between the bolts and the connected components due to different thermal expansion coefficients, gradually weakening the pre-tightening force of the bolts. In addition, a humid environment is likely to cause bolt corrosion, reducing its mechanical strength and further increasing the risk of bolt loosening. According to incomplete statistics, among wind turbines operating for a certain number of years, a considerable proportion of equipment has experienced bolt loosening problems, which not only affects power generation efficiency, but may also lead to serious safety accidents.

[0004] Currently, there are many limitations in the detection and warning means for bolt loosening of wind turbines. Traditional manual detection methods mainly rely on maintenance personnel to regularly inspect the wind turbines, and judge whether the bolts are loose by visually observing the appearance of the bolts and using simple tools such as wrenches to check the fastening degree. This method is not only inefficient and labor-intensive, but also the detection results are easily affected by the subjective factors of the maintenance personnel, and it is difficult to detect early slight loosening signs. Especially for bolts with relatively hidden installation positions that are difficult to directly observe, manual detection is extremely prone to missed detections. In addition, the cycle of manual inspection is relatively long, and between two inspections, bolt loosening problems may gradually develop and deteriorate, and effective maintenance measures cannot be taken in time.

[0005] With the development of technology, some sensor-based monitoring methods have gradually been applied. These methods collect relevant data of bolts during operation by installing stress sensors, displacement sensors, acoustic sensors, etc. near or on the bolts, and then analyze the state of the bolts. However, these existing monitoring technologies still face many challenges in practical applications. In terms of data processing, the collected raw data is often affected by various interference factors such as environmental noise and the vibration of the equipment itself, resulting in low data quality and making it difficult to accurately extract the effective features of the bolt state. In terms of model construction and prediction, the existing prediction models and evaluation models have insufficient adaptability to complex working conditions and cannot accurately predict the bolt loosening risk and evaluate the risk level accurately. For example, when some models face the bolt loosening situation under the coupling action of multiple factors, the prediction accuracy is low, and misjudgment or missed judgment is likely to occur. Moreover, the existing monitoring systems have defects in the decision-making and feedback mechanisms, lacking a perfect hierarchical response decision model, unable to make reasonable maintenance decisions in a timely manner according to different risk levels, and it is also difficult to effectively collect maintenance effect data and feedback it for model update, making it difficult to achieve closed-loop control of the bolt loosening risk. Summary of the Invention

[0006] The purpose of the present invention is to provide a wind turbine bolt loosening early warning system and method to solve the problems raised in the above background technology.

[0007] To achieve the above purpose, the present invention provides the following technical solution: A wind turbine bolt loosening early warning method, the method includes:

[0008] Collect multi-dimensional data of the wind turbine bolts through a distributed sensor array, the distributed sensor array includes a stress sensor, a displacement sensor, and an acoustic sensor; use wavelet transform to denoise and extract features from the collected raw data to obtain a bolt state feature vector; input the feature vector into a pre-trained deep belief network DBN model, the deep belief network is composed of multiple restricted Boltzmann machines RBMs stacked, including an input layer, multiple hidden layers, and an output layer, and optimize the parameters by combining the backpropagation algorithm with contrastive divergence, and output the bolt loosening risk prediction value;

[0009] Construct a risk assessment model based on the risk prediction value, take maximizing the risk assessment accuracy and minimizing the assessment error as the optimization objectives, and use the simulated annealing optimization algorithm to optimize the model hyperparameters; generate the bolt loosening risk level according to the optimized risk assessment model.

[0010] A hierarchical response decision-making model is established based on the risk level. This model includes a decision-making layer, an execution layer, and a feedback layer. The decision-making layer formulates maintenance decisions according to the risk level. The execution layer is responsible for executing maintenance tasks and monitoring the maintenance process in real time. The feedback layer collects maintenance effect data and feeds it back to the decision-making layer for model update, realizing the closed-loop control of the bolt loosening risk.

[0011] Preferably, the bolt state feature vector is input into a pre-trained deep belief network (DBN) model, and the parameters are optimized by combining the backpropagation algorithm with contrastive divergence to output the bolt loosening risk prediction value. Specifically, it includes:

[0012] Obtain the bolt state feature vector processed by wavelet transform, which includes stress change characteristics, displacement fluctuation characteristics, and acoustic signal characteristics. Based on this feature vector, construct an input space, and construct an output space with different prediction categories of bolt loosening risks.

[0013] Initialize the parameters of each layer of the deep belief network (DBN), including the weights and biases of each restricted Boltzmann machine (RBM). Use the contrastive divergence algorithm to perform unsupervised pre-training on each RBM layer by layer to optimize the weights and biases of each layer.

[0014] Fine-tune the pre-trained DBN model. Calculate the error between the predicted value and the true value through the backpropagation algorithm, and then adjust the parameters of the entire network. Use the stochastic gradient descent method to update the parameters, and randomly select a batch of samples from the training data for training in each iteration.

[0015] After multiple rounds of training, input the test data into the trained DBN model to output the bolt loosening risk prediction value, which represents the probability of the bolt loosening.

[0016] Preferably, when using the contrastive divergence algorithm to perform unsupervised pre-training on each RBM layer by layer:

[0017] Adopt the momentum method to accelerate convergence. By accumulating the gradient information of previous iteration steps, adjust the gradient direction of the current iteration to make the pre-training process tend to the optimal solution faster. At the same time, to avoid falling into local optima, introduce a random restart mechanism. After pre-training a certain number of rounds, randomly reset the weights of some RBMs and start training again.

[0018] Preferably, construct a risk assessment model based on the risk prediction value, use the simulated annealing optimization algorithm to optimize the hyperparameters of the model, and generate the bolt loosening risk level according to the optimized risk assessment model, including:

[0019] Construct a multi-objective function for risk assessment. This function consists of the evaluation accuracy rate and the evaluation error. The evaluation accuracy rate is calculated by the ratio of the number of correctly predicted bolt loosening samples to the total number of samples, and the evaluation error is measured by the mean square error between the predicted risk value and the actual risk value.

[0020] Build the model constraint conditions based on multiple objective functions, including model complexity constraints and data range constraints. The model complexity constraints are used to limit the number and value range of model parameters, and the data range constraints ensure that the input data is within a reasonable interval;

[0021] Use the simulated annealing algorithm to optimize the hyperparameters of the risk assessment model, and set the initial temperature and cooling rate parameters; randomly generate a set of hyperparameters as the initial solution, and calculate the objective function value of this solution; in each iteration, accept a worse solution with a certain probability, generate a new solution by perturbing the current solution and calculate the objective function value of the new solution; as the temperature decreases, gradually reduce the probability of accepting a worse solution, and when the temperature reaches the termination condition, obtain the optimal hyperparameters;

[0022] Adjust the risk assessment model according to the optimized hyperparameters, input the bolt loosening risk prediction value into the model, and divide the risk level according to the set risk threshold.

[0023] Preferably, in the process of optimizing the hyperparameters of the risk assessment model using the simulated annealing algorithm:

[0024] Use parallel computing technology to accelerate the search process, start multiple simulated annealing processes simultaneously, each process independently explores different regions of the hyperparameter space, exchanges information regularly, and aggregates the better solutions; and design an adaptive cooling strategy to dynamically adjust the cooling rate according to the quality of the currently searched solution, and speed up the cooling when the quality of the solution improves slowly, so as to prompt the algorithm to converge to the global optimum faster.

[0025] Preferably, the decision-making layer formulates maintenance decisions based on the risk level, including:

[0026] Use the decision tree algorithm to build a maintenance decision model, and use the risk level, operation duration, and maintenance history of the bolt as the input features of the decision tree;

[0027] Set decision nodes and leaf nodes for the decision tree. The decision nodes make conditional judgments based on the input features, and the leaf nodes correspond to different maintenance decisions;

[0028] Train the decision tree with training data, use the information gain index to select the optimal partitioning feature, and build a reasonable decision tree structure;

[0029] When new bolt risk level information is obtained, the decision tree model outputs the corresponding maintenance decision according to the trained structure.

[0030] Preferably, the execution layer is responsible for executing maintenance tasks and monitoring the maintenance process in real time, including:

[0031] Formulate a detailed maintenance execution plan based on the maintenance decision, including maintenance tool selection, maintenance personnel allocation, and maintenance step planning;

[0032] During maintenance, the status data of bolts and surrounding components are collected in real time by sensors, including stress changes and displacement changes during maintenance;

[0033] The Kalman filter algorithm is used to process the collected dynamic data, predict the change trend of bolt status, and timely detect abnormal situations during maintenance;

[0034] According to the maintenance execution plan and real-time monitoring results, the maintenance operations are dynamically adjusted.

[0035] Preferably, the feedback layer collects maintenance effect data and feeds it back to the decision-making layer for model update, including:

[0036] Collect the operation data of the bolts after maintenance, including stress, displacement and acoustic data, as well as the operation status data for a period of time afterwards;

[0037] Analyze and process the collected data, evaluate the effectiveness of maintenance measures, and calculate the change of bolt loosening risk before and after maintenance;

[0038] Feed the maintenance effect data back to the decision-making layer, and the decision-making layer updates the maintenance decision model and risk assessment model according to the feedback data;

[0039] The stochastic gradient descent online learning algorithm is used to adjust the model parameters.

[0040] Preferably, the arrangement of the distributed sensor array follows the principle of giving priority to key mechanical parts:

[0041] Identify the high stress concentration areas, easy vibration parts and frequently failed points in the structure of the wind turbine, and preferentially install stress sensors, displacement sensors and acoustic sensors at these key parts; use finite element analysis software to simulate and analyze the wind turbine, and optimize the specific installation angles and positions of the sensors according to the simulation results.

[0042] Preferably, a wind turbine bolt loosening warning system for implementing the above-mentioned wind turbine bolt loosening warning method includes:

[0043] A data acquisition and preprocessing unit for collecting multi-dimensional data of wind turbine bolts through a distributed sensor array, and using wavelet transform to denoise and extract features from the collected raw data to obtain bolt status feature vectors;

[0044] A risk prediction unit for inputting the bolt status feature vectors into a pre-trained deep belief network DBN model, optimizing the parameters by means of the backpropagation algorithm combined with contrastive divergence, and outputting bolt loosening risk prediction values;

[0045] A risk assessment unit, which is used to construct a risk assessment model based on risk prediction values, optimize the hyperparameters of the model using a simulated annealing optimization algorithm, and generate the bolt loosening risk level according to the optimized risk assessment model;

[0046] A decision execution feedback unit, which is used to establish a hierarchical response decision model based on the risk level, including a decision-making layer, an execution layer, and a feedback layer. The decision-making layer formulates maintenance decisions based on the risk level; the execution layer is responsible for executing maintenance tasks and monitoring the maintenance process in real time; the feedback layer collects maintenance effect data and feeds it back to the decision-making layer for model update.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] Collect multi-dimensional data of bolts through a distributed sensor array, use wavelet transform for denoising and feature extraction, and the deep belief network DBN model outputs accurate bolt loosening risk prediction values. This model is composed of multiple restricted Boltzmann machines RBMs stacked together, and the parameters are optimized by the backpropagation algorithm combined with contrastive divergence, which can effectively learn complex data features. Taking a wind farm as an example, before using the present invention, traditional detection methods often missed early bolt loosening hazards, resulting in failures. After adopting the present invention, the bolt loosening risk is accurately predicted in advance, and the prediction accuracy is greatly improved, from 60% before to more than 90%, successfully avoiding multiple potential failures and ensuring the stable operation of the fan.

[0049] Construct a risk assessment model based on risk prediction values, optimize the hyperparameters with a simulated annealing optimization algorithm, and maximize the risk assessment accuracy and minimize the assessment error. In this process, a multi-objective function is constructed, comprehensively considering the assessment accuracy and error, and setting constraints on the model complexity and data range. Compared with traditional assessment methods, the present invention can more accurately divide the bolt loosening risk level. In tests in different wind farms, the risk assessment error is reduced by about 40%, providing a scientific and accurate basis for maintenance decisions and making maintenance work more targeted.

[0050] Establish a hierarchical response decision model based on the risk level. The decision-making layer uses a decision tree algorithm to formulate maintenance decisions based on the bolt risk level, operation duration, and maintenance history, making the decisions more scientific and reasonable. The execution layer formulates a detailed maintenance plan accordingly, covering tool selection, personnel allocation, and step planning, and uses sensors to monitor the maintenance process in real time, processes data using the Kalman filter algorithm, and adjusts maintenance operations in a timely manner. In actual maintenance work, the maintenance efficiency is significantly improved, the average maintenance time is shortened by about 30%, ensuring the efficient and orderly progress of maintenance work, reducing equipment downtime, and improving power generation efficiency.

[0051] The feedback layer collects maintenance effect data, analyzes and evaluates the effectiveness of maintenance measures, calculates the change in bolt loosening risk before and after maintenance, and feeds back to the decision-making layer to update the maintenance decision-making model and risk assessment model. The random gradient descent online learning algorithm is used to adjust the model parameters to achieve closed-loop control of bolt loosening risk. With the continuous accumulation of data and the continuous update and optimization of the model, the system's prediction and evaluation of bolt loosening risk are more accurate, the maintenance decision-making is more scientific and reasonable. Under long-term operation, the equipment failure rate is reduced by about 50%, greatly improving the reliability and stability of wind turbine operation.

[0052] The early warning system and method of the present invention effectively reduce fan failures and downtime caused by bolt loosening, and reduce maintenance costs and power generation losses. At the same time, early warning and scientific maintenance decision-making avoid serious safety accidents caused by bolt loosening, ensuring the safety of operation and maintenance personnel and the safety of equipment and property. From the overall operation of the wind farm, within one year after applying the present invention, the maintenance cost of a certain wind farm is reduced by about 20%, and the power generation loss is reduced by about 15%, significantly improving economic benefits, and no safety accidents caused by bolt loosening occur, achieving a win-win situation of safety and economy. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is the working principle diagram of the bolt loosening early warning method for the wind turbine described in the present invention;

[0054] Figure 2 It is the step diagram of parameter optimization by combining the backpropagation algorithm with contrastive divergence to output the bolt loosening risk prediction value;

[0055] Figure 3 It is the step diagram of generating the bolt loosening risk level according to the optimized risk assessment model

[0056] Figure 4 It is the step diagram of the decision-making layer formulating maintenance decisions based on the risk level. DETAILED DESCRIPTION OF THE INVENTION

[0057] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0058] Please refer to Figures 1-4 , the present invention provides a technical solution: a bolt loosening early warning method for a wind turbine, the method includes:

[0059] With the help of a distributed sensor array, multi-dimensional data such as stress, displacement, and acoustics of the bolts of a wind turbine are collected. Subsequently, wavelet transform is used to denoise and extract features from the collected original data, removing interference information and refining feature vectors that can accurately reflect the bolt status, providing a reliable basis for subsequent analysis.

[0060] The processed bolt status feature vectors are input into a pre-trained deep belief network (DBN) model. This model is composed of multiple stacked restricted Boltzmann machines (RBMs) and has powerful feature learning and pattern recognition capabilities. Through parameter optimization using the backpropagation algorithm combined with contrastive divergence, the model can output the predicted value of bolt loosening risk, quantifying the possibility of bolt loosening.

[0061] A risk assessment model is constructed based on the risk prediction values. To achieve the goals of maximizing the accuracy of risk assessment and minimizing the assessment error, a simulated annealing optimization algorithm is used to optimize the hyperparameters of the model. The optimized model divides the bolt loosening risk into different levels according to the set risk threshold, intuitively presenting the risk status of the bolts.

[0062] A hierarchical response decision-making model is established based on the risk level, covering the decision-making layer, the execution layer, and the feedback layer. The decision-making layer formulates targeted maintenance decisions according to the risk level, such as arranging maintenance time, determining maintenance personnel and tools, etc.; the execution layer is responsible for efficiently executing maintenance tasks and using sensors to monitor the status data of bolts and surrounding components in the process; the feedback layer collects maintenance effect data and feeds it back to the decision-making layer for updating the maintenance decision model and the risk assessment model, forming a closed-loop control of bolt loosening risk and continuously improving the accuracy and reliability of the early warning system.

[0063] Example 1:

[0064] This example elaborates in detail the training process and prediction mechanism of the deep belief network (DBN) model in bolt loosening risk prediction to ensure that the model can accurately output the predicted value of bolt loosening risk and provide reliable data support for subsequent risk assessment and decision-making. The specific implementation methods include:

[0065] In a large wind farm, multiple wind turbines are selected for monitoring. After obtaining the bolt status feature vectors processed by wavelet transform, the input-output space is constructed. Taking one of the wind turbines as an example, its bolt status feature vectors include stress change features, displacement fluctuation features, and acoustic signal features. These features are reasonably encoded to construct the input space. According to historical data and expert experience, the bolt loosening risk is divided into three prediction categories: low, medium, and high, and the output space is constructed accordingly.

[0066] Initialize the parameters of each layer of the Deep Belief Network (DBN), including the weights and biases of each Restricted Boltzmann Machine (RBM). To accelerate the training speed, the momentum method is used to accelerate convergence. During the training process, record the gradient information of the previous iteration steps. For example, at the 10th iteration, add the gradient information of the 9th iteration to the current gradient according to a certain ratio (such as 0.9) to adjust the gradient direction of the current iteration, so that the pre-training process converges to the optimal solution faster. At the same time, to avoid falling into local optima, a random restart mechanism is introduced. Set that after every 50 rounds of pre-training, randomly reset the weights of some RBMs (such as 10%) and start training again.

[0067] Use the contrastive divergence algorithm to perform layer-by-layer unsupervised pre-training on each RBM. During the pre-training process, for a certain RBM layer, input a batch of bolt state feature vectors, and update the weights and biases through the contrastive divergence algorithm. Assume that after 100 rounds of pre-training, the reconstruction error of this RBM layer for the input data reaches a stable state, that is, it is considered that the pre-training of this layer is completed. Perform pre-training on each RBM layer in turn to complete the pre-training stage of the entire DBN model.

[0068] Fine-tune the pre-trained DBN model, and calculate the error between the predicted value and the true value through the backpropagation algorithm. Randomly select a batch of samples from the training data, for example, select 100 samples each time for training. Use the stochastic gradient descent method to update the parameters. In a certain iteration, calculate the prediction error of this batch of samples, and adjust the network parameters according to the error. After multiple rounds of training, input the test data into the trained DBN model. For example, input the bolt state feature vectors of another wind turbine as test data, and the model outputs the predicted value of the bolt loosening risk. This predicted value represents the probability of bolt loosening. Assume that the output result is 0.3, which means that the bolt has a 30% probability of loosening.

[0069] Example 2:

[0070] This example focuses on the optimization strategy adopted when using the contrastive divergence algorithm to perform layer-by-layer unsupervised pre-training on each RBM. Through the momentum method and the random restart mechanism, improve the model training efficiency, avoid falling into local optima, and enhance the generalization ability and prediction accuracy of the model. The specific implementation methods include:

[0071] On the basis of Example 1, further optimize the training process of the contrastive divergence algorithm. In a certain wind turbine bolt monitoring project, train the RBM layer of the Deep Belief Network (DBN). During the implementation of the momentum method, set the momentum factor to 0.9. In the initial stage of training, due to the relatively unstable gradient direction, the momentum method can effectively smooth the gradient change. For example, at the 1st round of training, the calculated gradient is g1, and the weight update amount is Δw1 = -η *g1 (where η is the learning rate, assumed to be 0.01). In the second round of training, the gradient is g2, and at this time the weight update amount is Δw2 = -η * g2 + 0.9 * Δw1. By accumulating the gradient information of previous iteration steps, the weight update direction becomes more stable, accelerating the convergence speed.

[0072] The introduction of the random restart mechanism effectively prevents the model from falling into a local optimum. It is set that after every 80 rounds of pre-training, the training effect of the model is checked. If in three consecutive checks, the error reduction rate of the model on the validation set is less than 0.01%, it is considered that the model may have fallen into a local optimum. At this time, 20% of the weights of the RBM are randomly reset. For example, for a DBN model with 5 RBM layers, randomly select 1 of the RBM layers, re-initialize its weights randomly, and then retrain from the current round. After retraining, the error of the model on the validation set decreases significantly, indicating that the random restart mechanism effectively breaks the local optimum and enables the model to continue exploring a better solution space.

[0073] By combining the momentum method and the random restart mechanism, experimental verification is carried out on the bolt condition monitoring data of multiple wind turbines. The results show that compared with the model without these optimization strategies, the training time of the model with optimization strategies is shortened by about 30%, and the prediction accuracy on the test set is increased by about 5%, effectively improving the performance and efficiency of the model.

[0074] Example 3:

[0075] This example details the process of constructing a risk assessment model based on risk prediction values and the specific steps of optimizing the model hyperparameters using the simulated annealing optimization algorithm to ensure that the risk assessment model can accurately generate the bolt loosening risk level. The specific implementation methods include:

[0076] In a wind farm in a certain area, a risk assessment model is constructed based on bolt loosening risk prediction values. A multi-objective function for risk assessment is constructed. Bolt data of 100 wind turbines in this area are collected, including samples with known loosening conditions and samples in normal operation. By calculating the ratio of the number of correctly predicted bolt loosening samples to the total number of samples, the evaluation accuracy is obtained. For example, in an evaluation, the total number of samples is 100, and the number of correctly predicted loosening samples is 80, then the evaluation accuracy is 80%. The evaluation error is measured by the mean square error between the predicted risk value and the actual risk value. Assume that the actual risk value of a certain bolt is 0.2 and the predicted risk value is 0.25, then the error of this sample is (0.25 - 0.2) ∧ 2 = 0.0025, and the average of the errors of all samples is calculated to obtain the evaluation error.

[0077] Construct the model constraint conditions based on multiple objective functions. In terms of model complexity constraints, limit the number of parameters of the risk assessment model to no more than 100, and the value range of the parameters is between [-10, 10]. The data range constraint ensures that the input data is within a reasonable range. For example, the value range of the stress change characteristic is [-50, 50] MPa, and the value range of the displacement fluctuation characteristic is [-10, 10] mm.

[0078] Use the simulated annealing algorithm to optimize the hyperparameters of the risk assessment model. Set the initial temperature to 1000 and the cooling rate to 0.95. Randomly generate a set of hyperparameters as the initial solution, such as the learning rate of 0.01, the regularization coefficient of 0.001, etc. Calculate the objective function value of this solution. Suppose the calculated objective function value is 0.5. In each iteration, accept a worse solution with a certain probability. For example, generate a new solution with an objective function value of 0.55. Calculate the acceptance probability (such as according to the formula of the simulated annealing algorithm, the acceptance probability is exp((0.5 - 0.55) / 1000)). If the acceptance probability is greater than a random number between 0 and 1 (such as 0.8), then accept this worse solution. Generate a new solution by perturbing the current solution, such as adjusting the learning rate to 0.011 and the regularization coefficient to 0.0011, and calculate the objective function value of the new solution. As the temperature decreases, gradually reduce the probability of accepting a worse solution. When the temperature reaches the termination condition (such as the temperature is lower than 1), obtain the optimal hyperparameters.

[0079] Adjust the risk assessment model according to the optimized hyperparameters, and input the bolt loosening risk prediction value into the model. Divide the risk levels according to the set risk thresholds. For example, set the risk thresholds to 0.3 and 0.7. When the risk prediction value is less than 0.3, it is classified as a low risk level; when the risk prediction value is between 0.3 and 0.7, it is classified as a medium risk level; when the risk prediction value is greater than 0.7, it is classified as a high risk level.

[0080] Example 4:

[0081] This example focuses on describing the specific implementation methods of using parallel computing technology and adaptive cooling strategies in the process of optimizing the hyperparameters of the risk assessment model by the simulated annealing algorithm, improving the optimization efficiency and accelerating the speed of the algorithm converging to the global optimal solution. The specific implementation methods include:

[0082] Based on Example 3, the simulated annealing algorithm is further optimized. In a bolt loosening risk assessment project of a large-scale wind farm, parallel computing technology is used to accelerate the search process. A parallel computing environment with 4 computing nodes is built, and 4 simulated annealing processes are started simultaneously. Each process independently explores different regions of the hyperparameter space. For example, Process 1 explores the region where the learning rate is in [0.001, 0.01] and the regularization coefficient is in [0.0001, 0.001]; Process 2 explores the region where the learning rate is in [0.01, 0.1] and the regularization coefficient is in [0.001, 0.01], etc.

[0083] During the operation of each process, information is exchanged regularly. It is set that every 100 iterations, each process will send the currently found better solution to a central node for summarization. The central node compares and filters these better solutions, and broadcasts the better solutions to each process. For example, at the 100th iteration, Process 1 finds a solution with an objective function value of 0.45, and Process 2 finds a solution with an objective function value of 0.48. After comparison by the central node, the solution of Process 1 is broadcast to other processes, and each process continues to explore based on this.

[0084] An adaptive cooling strategy is designed to dynamically adjust the cooling rate according to the quality of the currently found solution. At the initial stage of the operation of the simulated annealing algorithm, the quality of the solution improves relatively fast. At this time, the cooling rate is maintained at 0.95. When the decrease in the objective function value of the solution is less than 0.005 in 50 consecutive iterations, it is judged that the quality of the solution improves slowly, and the cooling rate is increased, such as adjusting the cooling rate to 0.9. Through this adaptive adjustment, the algorithm can converge to the global optimal solution faster.

[0085] Through experimental verification, under the same risk assessment model and data conditions, the simulated annealing algorithm using parallel computing technology and adaptive cooling strategy, compared with the algorithm without these optimization measures, the optimization time is shortened by about 40%, and the accuracy of the finally obtained risk assessment model on the test set is improved by about 3%, effectively improving the optimization effect of the simulated annealing algorithm and the performance of the risk assessment model.

[0086] Example 5:

[0087] This example elaborates in detail the specific operation mechanisms of the decision-making layer, the execution layer, and the feedback layer in the hierarchical response decision-making model, realizes the effective control of the bolt loosening risk, and ensures the rationality of maintenance decisions, the efficient execution of maintenance tasks, and the continuous optimization of the model. The specific implementation methods include:

[0088] In a certain wind farm, a hierarchical response decision-making model is operated. The decision-making layer formulates maintenance decisions based on the risk level, and uses the decision tree algorithm to construct a maintenance decision-making model. Data such as the risk level, operation duration, and maintenance history of bolts of multiple wind turbines in the power plant are collected as input features of the decision tree. For example, the risk level of a certain bolt is high, the operation duration is 5000 hours, and the last maintenance time was 1000 hours ago.

[0089] Decision nodes and leaf nodes are set for the decision tree. The decision nodes make conditional judgments based on the input features. For example, the decision node judges whether the risk level is high. If so, it further judges whether the operation duration exceeds 4000 hours; the leaf nodes correspond to different maintenance decisions. For example, when the risk level is high and the operation duration exceeds 4000 hours, the maintenance decision corresponding to the leaf node is to immediately arrange a comprehensive overhaul. The decision tree is trained with training data, and the information gain index is used to select the optimal partitioning feature to construct a reasonable decision tree structure. When new bolt risk level information is obtained, the decision tree model outputs the corresponding maintenance decision according to the trained structure.

[0090] The execution layer is responsible for executing maintenance tasks and monitoring the maintenance process in real time. A detailed maintenance execution plan is formulated based on the maintenance decision. For example, for the bolts that need to be comprehensively overhauled immediately as mentioned above, appropriate maintenance tools are selected, such as torque wrenches, bolt loosening agents, etc.; maintenance personnel with corresponding skills and experience are deployed; maintenance steps are planned, including first checking the tightening degree of the bolts, and then disassembling and replacing the loose bolts.

[0091] During the maintenance process, sensors are used to collect the status data of the bolts and surrounding components in real time. For example, the stress change during the maintenance process is collected through a stress sensor, and the displacement change is collected through a displacement sensor. The Kalman filter algorithm is used to process the collected dynamic data to predict the change trend of the bolt status. Suppose that during the maintenance process, the stress data of the bolt collected by the sensor fluctuates. The Kalman filter algorithm predicts the change of the bolt stress in the future period according to the historical data and the current measurement value. If the prediction result shows an abnormal increase in stress, the abnormal situation during the maintenance process is detected in time. According to the maintenance execution plan and the real-time monitoring results, the maintenance operation is dynamically adjusted. For example, if it is found that the bolt is difficult to disassemble, the maintenance method is adjusted in time, and measures such as heating or using a stronger loosening agent are taken.

[0092] The feedback layer collects maintenance effect data and feeds it back to the decision-making layer for model update. The operating data of the bolts after maintenance is collected, including stress, displacement, and acoustic data, as well as the operating status data for a period of time afterwards. For example, within 1 month after maintenance, the operating data of the bolts is continuously monitored. The collected data is analyzed and processed to evaluate the effectiveness of the maintenance measures, and the change in the bolt loosening risk before and after maintenance is calculated. Suppose the bolt loosening risk level was high before maintenance and dropped to low after maintenance, indicating that the maintenance measures are effective.

[0093] The maintenance effect data is fed back to the decision-making layer, and the decision-making layer updates the maintenance decision model and the risk assessment model according to the feedback data. The online learning algorithm of stochastic gradient descent is used to adjust the model parameters. For example, according to the actual operating data of the bolts after maintenance, the division conditions of some nodes in the decision tree model are adjusted, or the hyperparameters of the risk assessment model are adjusted to improve the accuracy and adaptability of the model and achieve closed-loop control of the bolt loosening risk.

[0094] The present invention also includes a bolt loosening warning system for a wind turbine, which is used to implement the above-mentioned bolt loosening warning method for a wind turbine, and includes:

[0095] A data acquisition and preprocessing unit, which is used to collect multi-dimensional data of the bolts of the wind turbine through a distributed sensor array, and use wavelet transform to denoise and extract features from the collected original data to obtain a bolt state feature vector;

[0096] A risk prediction unit, which is used to input the bolt state feature vector into a pre-trained deep belief network DBN model, optimize the parameters by means of the backpropagation algorithm combined with contrastive divergence, and output a bolt loosening risk prediction value;

[0097] A risk assessment unit, which is used to construct a risk assessment model based on the risk prediction value, use the simulated annealing optimization algorithm to optimize the model hyperparameters, and generate a bolt loosening risk level according to the optimized risk assessment model;

[0098] A decision execution feedback unit, which is used to establish a hierarchical response decision model based on the risk level, including a decision-making layer, an execution layer, and a feedback layer. The decision-making layer formulates maintenance decisions according to the risk level; the execution layer is responsible for executing the maintenance tasks and monitoring the maintenance process in real time; the feedback layer collects maintenance effect data and feeds it back to the decision-making layer for model update.

[0099] The implementation manner of this system refers to the above-mentioned embodiments and will not be elaborated in the description.

[0100] It should be noted that in this article, relational terms such as first and second are only used 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 "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or elements inherent to such process, method, article or device.

[0101] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for warning of bolt loosening in a wind turbine, characterized in that The method includes: Collecting multi-dimensional data of the bolts of a wind turbine through a distributed sensor array, where the distributed sensor array includes a stress sensor, a displacement sensor, and an acoustic sensor; using wavelet transform to denoise and extract features from the collected raw data to obtain a bolt state feature vector; inputting the feature vector into a pre-trained deep belief network (DBN) model, where the deep belief network is composed of multiple restricted Boltzmann machines (RBMs) stacked together, including an input layer, multiple hidden layers, and an output layer, and optimizing the parameters by combining backpropagation algorithm with contrastive divergence to output a bolt loosening risk prediction value; Constructing a risk assessment model based on the risk prediction value, with the maximization of risk assessment accuracy and the minimization of assessment error as the optimization objectives, and using the simulated annealing optimization algorithm to optimize the model hyperparameters; generating a bolt loosening risk level according to the optimized risk assessment model; Establishing a hierarchical response decision-making model based on the risk level, where the model includes a decision-making layer, an execution layer, and a feedback layer; the decision-making layer formulates maintenance decisions based on the risk level; the execution layer is responsible for executing the maintenance tasks and monitoring the maintenance process in real time; the feedback layer collects the maintenance effect data and feeds it back to the decision-making layer for model update to achieve closed-loop control of the bolt loosening risk.

2. The method for warning of bolt loosening of a wind turbine according to claim 1, wherein Inputting the bolt state feature vector into a pre-trained deep belief network (DBN) model and optimizing the parameters by combining backpropagation algorithm with contrastive divergence to output a bolt loosening risk prediction value, which specifically includes: Obtaining the bolt state feature vector processed by wavelet transform, which includes stress change features, displacement fluctuation features, and acoustic signal features; constructing an input space based on the feature vector and constructing an output space with different prediction categories of bolt loosening risks; Initializing the parameters of each layer of the deep belief network (DBN), including the weights and biases of each restricted Boltzmann machine (RBM); using the contrastive divergence algorithm to perform layer-by-layer unsupervised pre-training on each RBM to optimize the weights and biases of each layer; Fine-tuning the pre-trained DBN model, calculating the error between the predicted value and the true value through the backpropagation algorithm, and then adjusting the parameters of the entire network; using the stochastic gradient descent method to update the parameters, and randomly selecting a batch of samples from the training data for training in each iteration; After multiple rounds of training, inputting the test data into the trained DBN model to output a bolt loosening risk prediction value, which represents the probability of the bolt loosening.

3. A method for warning of bolt loosening in a wind turbine according to claim 2, characterized in that, When using the contrastive divergence algorithm to perform layer-by-layer unsupervised pre-training on each RBM: Using the momentum method to accelerate convergence, adjusting the gradient direction of the current iteration by accumulating the gradient information of the previous iteration steps to make the pre-training process tend to the optimal solution faster; at the same time, to avoid falling into local optima, a random restart mechanism is introduced, and after a certain number of rounds of pre-training, the weights of some RBMs are randomly reset and the training starts again.

4. A method for warning of bolt loosening in a wind turbine according to claim 1, characterized in that, Constructing a risk assessment model based on the risk prediction value, using the simulated annealing optimization algorithm to optimize the model hyperparameters, and generating a bolt loosening risk level according to the optimized risk assessment model, including: Construct a multi-objective function for risk assessment, which consists of assessment accuracy and assessment error. The assessment accuracy is calculated by the ratio of the number of correctly predicted bolt loosening samples to the total number of samples, and the assessment error is measured by the mean square error between the predicted risk value and the actual risk value. Based on the multi-objective function, construct model constraints, including model complexity constraints and data range constraints. The model complexity constraints are used to limit the number and value range of model parameters, and the data range constraints ensure that the input data is within a reasonable interval. Use the simulated annealing algorithm to optimize the hyperparameters of the risk assessment model, and set the initial temperature and cooling rate parameters; randomly generate a set of hyperparameters as the initial solution, and calculate the objective function value of this solution; in each iteration, accept a worse solution with a certain probability, generate a new solution by perturbing the current solution and calculate the objective function value of the new solution; as the temperature decreases, gradually reduce the probability of accepting a worse solution, and when the temperature reaches the termination condition, obtain the optimal hyperparameters. Adjust the risk assessment model according to the optimized hyperparameters, input the predicted value of bolt loosening risk into the model, and divide the risk levels according to the set risk threshold.

5. A method for warning of bolt loosening in a wind turbine according to claim 4, characterized in that, In the process of optimizing the hyperparameters of the risk assessment model using the simulated annealing algorithm, it specifically includes: Utilize parallel computing technology to accelerate the search process. Start multiple simulated annealing processes simultaneously. Each process independently explores different regions of the hyperparameter space, exchanges information regularly, and aggregates the better solutions; moreover, design an adaptive cooling strategy to dynamically adjust the cooling rate according to the quality of the currently found solution. When the quality of the solution improves slowly, increase the cooling rate to prompt the algorithm to converge to the global optimum faster.

6. A method for warning of bolt loosening in a wind turbine according to claim 1, characterized in that, The decision-making layer formulates maintenance decisions based on the risk levels, including: Use the decision tree algorithm to construct a maintenance decision-making model, taking the risk level, operating duration, and maintenance history of the bolt as the input features of the decision tree. Set decision nodes and leaf nodes for the decision tree. The decision nodes make conditional judgments based on the input features, and the leaf nodes correspond to different maintenance decisions. Train the decision tree with the training data, select the optimal splitting feature using the information gain metric, and construct a reasonable decision tree structure. When new bolt risk level information is obtained, the decision tree model outputs the corresponding maintenance decision according to the trained structure.

7. A method for warning of bolt loosening in a wind turbine according to claim 6, characterized in that, The execution layer is responsible for executing maintenance tasks and monitoring the maintenance process in real time, including: Based on the maintenance decision, formulate a detailed maintenance execution plan, including maintenance tool selection, maintenance personnel allocation, and maintenance step planning. During the maintenance process, use sensors to collect the status data of the bolt and surrounding components in real time, including stress changes and displacement changes during the maintenance process. Use the Kalman filter algorithm to process the collected dynamic data, predict the change trend of the bolt state, and timely detect abnormal situations during the maintenance process. According to the maintenance execution plan and real-time monitoring results, dynamically adjust the maintenance operations.

8. A method for warning of bolt loosening in a wind turbine according to claim 7, characterized in that, The feedback layer collects maintenance effect data and feeds it back to the decision-making layer for model update, including: Collect the operating data of the bolt after maintenance, including stress, displacement, and acoustic data, as well as the operating status data for a subsequent period of time. Analyze and process the collected data, evaluate the effectiveness of maintenance measures, and calculate the change in the bolt loosening risk before and after maintenance; Feed the maintenance effect data back to the decision-making level, and the decision-making level updates the maintenance decision model and risk assessment model according to the feedback data; Adopt the stochastic gradient descent online learning algorithm to adjust the model parameters.

9. A method for warning of bolt loosening in a wind turbine according to claim 1, characterized in that, The arrangement of the distributed sensor array follows the principle of giving priority to key mechanical parts: Identify the high stress concentration areas, easy-to-vibrate parts and frequently failed points in the wind turbine structure, and preferentially install stress sensors, displacement sensors and acoustic sensors at these key parts; use finite element analysis software to simulate the wind turbine, and optimize the specific installation angles and positions of the sensors according to the simulation results.

10. A wind turbine bolt loosening warning system for implementing the method described in any one of claims 1-9, characterized in that, including: A data acquisition and preprocessing unit, which is used to collect multi-dimensional data of the bolts of the wind turbine through a distributed sensor array, and use wavelet transform to denoise and extract features from the collected raw data to obtain a bolt state feature vector; A risk prediction unit, which is used to input the bolt state feature vector into a pre-trained deep belief network DBN model, optimize the parameters by means of the backpropagation algorithm combined with contrastive divergence, and output the bolt loosening risk prediction value; A risk assessment unit, which is used to construct a risk assessment model based on the risk prediction value, use the simulated annealing optimization algorithm to optimize the model hyperparameters, and generate a bolt loosening risk level according to the optimized risk assessment model; A decision execution feedback unit, which is used to establish a hierarchical response decision model based on the risk level, including a decision-making layer, an execution layer and a feedback layer. The decision-making layer formulates maintenance decisions according to the risk level; the execution layer is responsible for executing maintenance tasks and monitoring the maintenance process in real time; the feedback layer collects maintenance effect data and feeds it back to the decision-making layer for model update.

Images