Reliability analysis system for wind power generation bearing based on multivariate degradation analysis

Abstract

The application provides a wind power generation bearing reliability analysis system based on multivariate degradation analysis, relates to the field of wind power generation, and comprises the following modules: a position determination module, which is used for determining multiple vibration monitoring positions of a wind power generation bearing; a model construction module, which is used for constructing multiple first unconditional binary Copula functions and multiple conditional binary Copula functions with different orders based on the multiple vibration monitoring positions through a Vine-Copula algorithm, constructing a first reliability evaluation model, and then constructing multiple second unconditional binary Copula functions and a second reliability evaluation model; a vibration monitoring module, which is used for collecting vibration sequences of the multiple vibration monitoring positions of the wind power generation bearing; and a reliability analysis module, which is used for calculating a first reliability through the second reliability evaluation model, judging whether a bearing reliability risk occurs, calculating a wind power generation bearing reliability through the first reliability evaluation model if the bearing reliability risk occurs, and having the advantages of improving the accuracy and efficiency of wind power generation bearing reliability analysis.

Description

Technical Field

[0001] This invention relates to the field of wind power generation, and in particular to a wind power bearing reliability analysis system based on multivariate degradation analysis. Background Technology

[0002] The main shaft, yaw, and pitch bearings of wind turbine generators are core load-bearing components of the drivetrain. They are exposed to extreme operating environments such as strong gusts in mountainous areas, high salt spray near the coast, and large temperature differences in deserts, continuously enduring the combined effects of variable loads, strong impacts, and alternating torques. Their degradation and failure account for more than 40% of all failures in the wind turbine drivetrain system, making them a core cause of unplanned downtime and soaring maintenance costs. As wind power installed capacity continues to increase and existing units gradually enter the later stages of service, accurate reliability assessment of wind turbine bearings has become a key technical requirement for wind farms to achieve condition-based maintenance and reduce the cost per kilowatt-hour.

[0003] The current mainstream method for determining the reliability of wind turbine bearings in the industry generally adopts a technical logic of directly comparing the vibration increment at a single point with a fixed failure threshold: the operation and maintenance system only extracts the real-time vibration amplitude increment at a single measuring point and compares it with the difference of a preset unified failure threshold. When the vibration increment exceeds the threshold, the bearing is determined to have entered a failure state. This method has significant inherent defects in practical engineering applications: the operating load of wind turbine units is not a constant value, and the normal vibration benchmark of the bearing naturally fluctuates under different wind speeds and different power generation. Relying solely on the absolute value of the vibration increment at a single point can easily misjudge normal vibration increments under high load conditions as failure characteristics. At the same time, it can also mask early hidden damage caused by the simultaneous accumulation of multi-dimensional degradation parameters, ultimately leading to a large number of false alarms and missed alarms.

[0004] Therefore, there is a need to provide a wind turbine bearing reliability analysis system based on multivariate degradation analysis to improve the accuracy and efficiency of wind turbine bearing reliability analysis. Summary of the Invention

[0005] This invention provides a wind turbine bearing reliability analysis system based on multivariate degradation analysis, comprising: a location determination module for determining multiple vibration monitoring locations of the wind turbine bearing; a model building module for constructing multiple first unconditional binary Copula functions and multiple conditional binary Copula functions of different orders based on the multiple vibration monitoring locations using the Vine-Copula algorithm, and constructing a first reliability assessment model based on the multiple first unconditional binary Copula functions and multiple conditional binary Copula functions of different orders; the model building module is further used to construct multiple second unconditional binary Copula functions based on the first reliability assessment model and the multiple vibration monitoring locations, and construct a second reliability assessment model based on the multiple second unconditional binary Copula functions; a vibration monitoring module for collecting vibration sequences from the multiple vibration monitoring locations of the wind turbine bearing; a reliability analysis module for calculating a first reliability based on the vibration sequences from the multiple vibration monitoring locations using the second reliability assessment model; and the reliability analysis module is further used to determine whether a bearing reliability risk has occurred based on the first reliability, and if so, to calculate the reliability of the wind turbine bearing using the first reliability assessment model.

[0006] Furthermore, multiple vibration monitoring locations of the wind turbine bearing are determined, including: collecting test data corresponding to various test conditions, wherein the test data includes vibration sequences at multiple locations of the wind turbine bearing; and determining multiple vibration monitoring locations from the multiple locations of the wind turbine bearing based on the test data corresponding to various test conditions.

[0007] Furthermore, based on test data corresponding to various test conditions, multiple vibration monitoring locations are determined from multiple locations on the wind turbine bearing. This includes: for any two locations, calculating the average Kendall rank correlation coefficient between the two locations based on the test data corresponding to various test conditions; selecting multiple valid locations from multiple locations on the wind turbine bearing based on the average Kendall rank correlation coefficient between the two locations; for any two valid locations, calculating the average similarity between the two valid locations based on the test data corresponding to various test conditions; and selecting multiple vibration monitoring locations from multiple valid locations based on the average similarity between the two valid locations.

[0008] Furthermore, using the Vine-Copula algorithm, multiple first unconditional binary Copula functions and multiple conditional binary Copula functions of different orders are constructed based on multiple vibration monitoring locations. A first reliability assessment model is then constructed based on these first unconditional binary Copula functions and multiple conditional binary Copula functions of different orders. This model includes: for any two vibration monitoring locations, calculating the average Kendall rank correlation coefficient between the two locations based on processed test data corresponding to various test conditions; constructing a vine tree based on the average Kendall rank correlation coefficient between any two vibration monitoring locations; and constructing multiple first unconditional binary Copula functions and multiple conditional binary Copula functions of different orders based on the vine tree to build the first reliability assessment model.

[0009] Furthermore, based on the first reliability assessment model and multiple vibration monitoring locations, multiple second unconditional binary Copula functions are constructed, including: initializing the population based on the average Kendall rank correlation coefficient of any two vibration monitoring locations, where each individual includes multiple location pairs, and each location pair includes two vibration monitoring locations; using a parameter-adaptive genetic algorithm, iteratively optimizing the population according to the first reliability assessment model to determine the optimal multiple location pairs; and constructing multiple second unconditional binary Copula functions based on the optimal multiple location pairs.

[0010] Furthermore, the population is initialized based on the average Kendall rank correlation coefficient of any two vibration monitoring locations, including: determining the sampling probability of the location pair corresponding to any two vibration monitoring locations based on the average Kendall rank correlation coefficient of any two vibration monitoring locations; and initializing the population based on the sampling probability of the location pair corresponding to any two vibration monitoring locations.

[0011] Furthermore, the population is iteratively optimized using a parameter-adaptive genetic algorithm based on the first reliability assessment model to determine the optimal multiple position pairs, including determining the elite retention ratio in each iteration.

[0012] Furthermore, using a parameter-adaptive genetic algorithm, the population is iteratively optimized based on the first reliability assessment model to determine the optimal multiple position pairs. This includes: calculating the multivariate joint reliability corresponding to each experimental condition based on test data corresponding to various experimental conditions using the first reliability assessment model; in each iteration, for each individual, constructing a second unconditional binary Copula function for the multiple position pairs included in the individual, determining the independent joint reliability of the second unconditional binary Copula function for the multiple position pairs included in the individual for each experimental condition, and determining the fitness of the individual based on the multivariate joint reliability corresponding to each experimental condition and the independent joint reliability of the second unconditional binary Copula function for the multiple position pairs included in the individual.

[0013] Furthermore, based on multiple second unconditional binary Copula functions, a second reliability assessment model is constructed, including: determining the independent joint reliability of each output of the second unconditional binary Copula function for each test condition according to the test data corresponding to multiple test conditions; calculating the multivariate joint reliability corresponding to each test condition using the first reliability assessment model based on the test data corresponding to multiple test conditions; constructing a multivariate nonlinear regression model, wherein the independent variables of the multivariate nonlinear regression model include the independent joint reliability calculated by multiple second unconditional binary Copula functions, and the dependent variable of the multivariate nonlinear regression model is the first reliability; and solving for the parameters of the multivariate nonlinear regression model based on the independent joint reliability and multivariate joint reliability output of each second unconditional binary Copula function for each test condition to construct the second reliability assessment model.

[0014] Furthermore, based on the first reliability, it is determined whether a bearing reliability risk has occurred, including: determining whether the first reliability is less than the first reliability threshold; if so, it is determined that a bearing reliability risk has occurred; if not, it is determined whether a bearing reliability risk has occurred based on the reliability calculated by multiple second unconditional binary Copula functions.

[0015] Compared with existing technologies, the wind power bearing reliability analysis system based on multivariate degradation analysis provided by this invention has at least the following beneficial effects: Addressing the industry pain points of existing wind turbine bearing reliability assessment technologies, such as high false alarm and missed alarm rates due to single-point threshold judgment, high computational barriers to deploying traditional high-dimensional Vine-Copula models, and poor real-time performance, this system presents an innovative end-to-end solution from optimized measurement point layout to dual-layer verification output, offering multiple significant engineering benefits. Firstly, this system uses Kendall's rank correlation coefficient to quantitatively screen vibration monitoring locations, moving away from the traditional extensive point deployment model that relies on operational experience. The selected measurement point combinations retain the full-dimensional degradation characteristics of the bearing while minimizing information redundancy. Compared to conventional point deployment schemes, the number of measurement points can be reduced by 30%, thus reducing the computational overhead for subsequent modeling from the data source. Secondly, this system innovatively constructs a two-layer modeling architecture that combines a full-precision first reliability assessment model with a lightweight second reliability assessment model. It not only fully captures the nonlinear and asymmetric interdependence and coupling characteristics between multiple measurement points through Vine-Copula, avoiding the assessment bias caused by the assumption of variable independence, but also generates a lightweight model that maintains more than 98% assessment accuracy while taking only 1 / 10 of the computation time of the original full-dimensional model. It can be directly deployed on low-computing-power devices at the edge of the wind farm and can complete real-time assessment without transmitting massive vibration data back to the cloud. Attached Figure Description

[0016] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein: Figure 1 This is a schematic diagram of a wind power bearing reliability analysis system based on multivariate degradation analysis, as shown in some embodiments of this specification. Figure 2 This is a flowchart illustrating the determination of multiple vibration monitoring locations according to some embodiments of this specification; Figure 3 This is a schematic diagram of the process for constructing a second reliability assessment model according to some embodiments of this specification; Figure 4 This is a schematic diagram showing the results of the first reliability assessment model and the test results of the second reliability assessment model according to some embodiments of this specification. Detailed Implementation

[0017] To more clearly illustrate the technical solutions of the embodiments in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some examples or embodiments of this specification. For those skilled in the art, these drawings can be applied to other similar scenarios without creative effort. Unless obvious from the context or otherwise specified, the same reference numerals in the drawings represent the same structures or operations.

[0018] Figure 1 This is a schematic diagram of a wind turbine bearing reliability analysis system based on multivariate degradation analysis, as shown in some embodiments of this specification. Figure 1 As shown, the wind power bearing reliability analysis system based on multivariate degradation analysis can include a location determination module, a model building module, a vibration monitoring module, and a reliability analysis module.

[0019] The location determination module is used to determine multiple vibration monitoring locations of the wind turbine bearing.

[0020] Specifically, it includes: Test data were collected for various test conditions, including vibration sequences at multiple locations of the wind turbine bearing. Based on test data corresponding to various test conditions, multiple vibration monitoring locations were determined from multiple positions of the wind turbine bearing.

[0021] Specifically, the various test conditions are layered simulation scenarios that fully reproduce the real service environment of wind turbine units, covering the entire spectrum of operating states from low wind speed standby and rated power operation to strong gust impact and alternating load loading. These include: 1. Rated load steady-state condition, where the bearing is loaded to 100% of the design rated dynamic load, and the speed is kept stable within the rated operating range of the unit, simulating the normal state of stable daily power generation; 2. Variable speed and variable load transient condition, where the axial and radial loading forces and shaft speed are dynamically adjusted according to the IEC standard wind power load spectrum to reproduce the load fluctuations caused by gusts and sudden changes in wind speed; 3. Extreme temperature coupling condition, which switches between environmental temperature ranges of -40℃ to 80℃ to simulate the bearing operating state in extreme environments such as cold regions and high-temperature deserts; 4. Degradation simulation condition of wind power bearings at different wear stages.

[0022] Multiple locations on the wind turbine bearing may include, but are not limited to, the top and side of the bearing housing, the outer and inner rings of the bearing, the motor drive / non-drive end bearing housing, the impeller side, the base, and the corresponding bearing mounting position on the gearbox housing. Vibration acceleration sensors are installed at multiple locations on the wind turbine bearing to synchronously collect vibration sequences from these locations. Each vibration sequence consists of vibration accelerations at multiple moments. All vibration acceleration sensors are tightened to standard torque to prevent loose installation from introducing interference signals. The sampling frequency is set to no less than 25.6 kHz, and the duration of each vibration sequence acquisition is no less than 120 seconds.

[0023] Figure 2 This is a flowchart illustrating the determination of multiple vibration monitoring locations according to some embodiments of this specification, such as... Figure 2 As shown, in some embodiments, based on test data corresponding to various test conditions, multiple vibration monitoring locations are determined from multiple locations of the wind turbine bearing, including: For any two locations, the average Kendall rank correlation coefficient between the two locations is calculated based on test data corresponding to various test conditions. Based on the average Kendall rank correlation coefficient between any two positions, multiple valid positions are selected from multiple positions of the wind turbine bearing. For any two valid locations, the average similarity between the two valid locations is calculated based on the test data corresponding to various test conditions. Multiple vibration monitoring locations are selected from a pool of valid locations based on the average similarity between any two valid locations.

[0024] Specifically, for any two locations and each test condition, the Kendall rank correlation coefficient of the vibration sequences at the two locations is calculated. For any two locations, the average Kendall rank correlation coefficient of the vibration sequences at each location for each test condition is calculated to obtain the average Kendall rank correlation coefficient of the two locations. The value ranges from [-1, 1]. The closer the value is to 1, the stronger the positive dependence of the vibration signals at the two locations; the closer it is to -1, the stronger the negative dependence; and the closer it is to 0, the weaker the dependence.

[0025] For each location, determine whether the absolute value of the average Kendall rank correlation coefficient between that location and at least one other location is greater than an absolute value threshold (e.g., 0.3). If yes, the location is valid; otherwise, it is invalid and discarded. Wind turbine bearings operate in complex environments. Of the original multiple locations, some are affected by factors such as electromagnetic interference from the nacelle, installation locations deviating from the vibration transmission path, and poor adaptability to operating conditions. The collected vibration sequences at these locations show almost no monotonic correlation with other locations, failing to effectively capture the overall degradation trend of the bearing, and are therefore invalid. By setting an absolute value threshold of 0.3 to filter valid locations, these isolated locations without stable correlation with other locations can be eliminated, retaining valid locations that can form stable vibration response synchronization with at least one other location. This avoids the waste of acquisition resources, redundant data storage, and interference with subsequent analysis caused by invalid locations, while ensuring that the retained set of locations can be analyzed through cross-location vibration correlation.

[0026] For any two valid positions and each test condition, calculate the cosine similarity of the vibration sequences at the two valid positions. For any two valid positions, average the cosine similarity of the vibration sequences at each valid position for each test condition to obtain the average similarity between the two valid positions.

[0027] For any two valid locations, if the average similarity between the two valid locations is greater than a similarity threshold (e.g., 70%), then one of the valid locations is discarded. After discarding, the remaining valid locations are used as multiple vibration monitoring locations. Discarding redundant locations with highly homogeneous vibration responses avoids information redundancy caused by high overlap of data from multiple locations, significantly reducing the computational costs of subsequent data acquisition, storage, and analysis, and also minimizing resource waste caused by repeated measurement points.

[0028] The model building module is used to construct multiple first unconditional binary Copula functions and multiple conditional binary Copula functions of different orders based on multiple vibration monitoring locations using the Vine-Copula algorithm. Based on the multiple first unconditional binary Copula functions and multiple conditional binary Copula functions of different orders, a first reliability assessment model is constructed.

[0029] Specifically, it includes: For any two vibration monitoring locations, the average Kendall rank correlation coefficient between the two vibration monitoring locations is calculated based on the processed test data corresponding to various test conditions. A vine tree is constructed based on the average Kendall rank correlation coefficient between any two vibration monitoring locations; Based on the vine tree, multiple first unconditional binary Copula functions and multiple conditional binary Copula functions of different orders are constructed to build the first reliability evaluation model.

[0030] Specifically, the calculation method for the average Kendall rank correlation coefficient of two vibration monitoring locations based on the processed test data corresponding to various test conditions is similar to the method for calculating the average Kendall rank correlation coefficient of two locations, and will not be repeated here.

[0031] A vine is a hierarchical tree structure used to describe dependencies between multiple variables. It decomposes complex high-dimensional dependencies into combinations of multiple binary dependencies through a layer-by-layer approach, perfectly suited to the modeling needs of high-dimensional monitoring scenarios with multiple vibration monitoring points. The regular vine (R-vine) is selected as the basic structure, and the construction process uses the average Kendall rank correlation coefficient between two vibration monitoring locations as the core basis, specifically built step-by-step in three levels: The first layer is the main vine tree, which uses all vibration monitoring locations as nodes. The absolute value of the average Kendall rank correlation coefficient between any two nodes is used as the weight of the edge. The maximum spanning tree algorithm is used to filter the edges, prioritizing the nodes with the largest absolute value of the correlation coefficient. This ensures that the first layer of the tree can capture the strongest correlation among multiple vibration monitoring locations, thus avoiding the problem of strong dependencies being masked by weak dependencies.

[0032] The second and subsequent higher-order vine trees use the edges of the previous vine tree as nodes of the current layer. Edge weights are calculated based on the partial Kendall rank correlation coefficient under the corresponding conditional variables, and a new tree structure is generated again using the maximum spanning tree algorithm. With each new vine tree layer, the dependency relationship upgrades from unconditional binary dependency to first-order conditional binary dependency, then to second-order conditional binary dependency, and so on, until all n-1 layers of vine trees corresponding to the n vibration monitoring locations are constructed.

[0033] For example, the five vibration monitoring locations are: input shaft measuring point A, first-stage planetary carrier measuring point B, second-stage planetary carrier measuring point C, high-speed shaft measuring point D, and generator side measuring point E, totaling five measuring points. These correspond to four layers of vines, and their corresponding average Kendall rank correlation coefficients are shown in Table 1.

[0034]

[0035] The first-level main vine tree T1: Using A, B, C, D, and E as 5 original nodes, the absolute value of the average Kendall rank correlation coefficient is used as the edge weight. Kruskal's maximum spanning tree algorithm is used to filter edges: first, BC (0.85) with the largest weight is selected, then CD (0.78), followed by AB (0.72), and finally DE (0.61). The final edge set of T1 is {BC, CD, AB, DE}. The second layer of the vine tree T2: The four edges {BC, CD, AB, DE} of T1 are used as four new nodes in T2, denoted as N1(B,C), N2(C,D), N3(A,B), and N4(D,E). The partial Kendall rank correlation coefficients between these nodes are calculated as follows: Using the common measurement point of two nodes as a conditional variable, for example, if the common vibration monitoring location of N1 and N3 is B, the partial Kendall rank correlation coefficient between A and C with respect to B is calculated based on the average Kendall rank correlation coefficients between B and C, A and B, and A and C. The common measurement point of N1 and N2 is C, so the partial Kendall rank correlation coefficient between B and D with respect to C is calculated. The common measurement point of N2 and N4 is D, so the partial Kendall rank correlation coefficient between C and E with respect to D is calculated. The conditional correlation coefficients of N3 with other nodes are all below 0.3. Using the maximum spanning tree to select edges again, we obtain the edge set of T2 as {N1-N2, N2-N4, N1-N3}, corresponding to the first-order conditional dependencies: BD|C, CE|D, AC|B.

[0036] The third layer of the vine tree T3: Take the three edges {N1-N2, N2-N4, N1-N3} of T2 as the three new nodes of T3, denoted as M1(BD|C), M2(CE|D), and M3(AC|B). Calculate the second-order partial Kendall coefficients using the common condition variables of each pair of nodes. The maximum spanning tree selects two edges: {M1-M2, M1-M3}, corresponding to the second-order conditional dependencies: BE|{C,D} and AD|{B,C}.

[0037] Fourth layer of vine tree T4: Take the two edges {M1-M2, M1-M3} of T3 as two new nodes of T4, denoted as P1(BE|{C,D}) and P2(AD|{B,C}). Calculate the third-order partial Kendall rank correlation coefficients of A and E with respect to {B,C,D} under the common condition variable {C,D} of the two nodes, and obtain a unique edge P1-P2, corresponding to the third-order conditional dependency relation: AE|{B,C,D}.

[0038] After determining the topology of the vine tree, for each edge of each layer of the vine tree, a corresponding binary Copula function is matched to form a complete set of Copula functions covering unconditional dependence to higher-order conditional dependence: For all edges of the first-layer main vine tree, each edge corresponds to the unconditional dependence relationship between two vibration monitoring positions. Based on the average Kendall rank correlation coefficient corresponding to the edge, the function with the highest fitting accuracy is selected from commonly used binary Copula function families such as Gaussian Copula, Clayton Copula, Gumbel Copula, and Frank Copula through the AIC information criterion and empirical distribution fitting test. This type of Copula is defined as the first unconditional binary Copula function, which is used to directly describe the joint distribution law of vibration characteristics of two vibration monitoring positions.

[0039] For the edges of the vines in the second layer and above, each edge corresponds to the conditional dependency relationship between two vibration monitoring locations under one or more given vibration monitoring locations. Based on the partial Kendall rank correlation coefficient under the corresponding conditions, an optimal binary Copula function is individually matched for each edge. These functions are conditional binary Copula functions of different orders. For example, the edges of the second layer of the vines correspond to a first-order conditional binary Copula, the edges of the third layer correspond to a second-order conditional binary Copula, and so on, until all layers of edges have been matched.

[0040] The model building module is also used to construct multiple second unconditional binary Copula functions based on the first reliability assessment model and multiple vibration monitoring locations, and to construct a second reliability assessment model based on the multiple second unconditional binary Copula functions.

[0041] Specifically, it includes: The population is initialized based on the average Kendall rank correlation coefficient between any two vibration monitoring locations, where an individual consists of multiple location pairs, and each location pair consists of two vibration monitoring locations. Using a parameter-adaptive genetic algorithm, the population is iteratively optimized based on the first reliability evaluation model to determine the optimal multiple position pairs; Multiple second unconditional binary Copula functions are constructed based on the optimal multiple position pairs.

[0042] In some embodiments, the population is initialized based on the average Kendall rank correlation coefficient between any two vibration monitoring locations, including: Based on the average Kendall rank correlation coefficient of any two vibration monitoring locations, determine the sampling probability of the location pair corresponding to any two vibration monitoring locations. The population is initialized based on the sampling probability of the location pairs corresponding to any two vibration monitoring locations.

[0043] Specifically, for any two vibration monitoring locations, the ratio of the absolute value of the average Kendall rank correlation coefficient of the two vibration monitoring locations to the sum of the absolute values ​​of the average Kendall rank correlation coefficients of all two vibration monitoring locations can be calculated as the sampling probability of the location pair corresponding to the two vibration monitoring locations. This ensures that the sum of the sampling probabilities of all location pairs is strictly equal to 1, thus avoiding probability distribution discontinuities in subsequent sampling.

[0044] The population can be initialized according to the following procedure: First, determine the value range of the location pair based on the total number of vibration monitoring locations. For example, for a scenario with 5 vibration monitoring locations, the value range is [4, 10]. Before generating each new individual, an integer is randomly selected from the range of values ​​to represent the total number of position pairs contained in that individual. After determining the total number of position pairs contained in a single individual, roulette wheel sampling without replacement is performed based on the sampling probability of position pairs corresponding to any two vibration monitoring positions: the higher the absolute value of the average Kendall rank correlation coefficient between the two vibration monitoring positions, the greater the probability that the corresponding position pair will be selected. Each selected position pair is added to the position pair set of the current individual until the total number of position pairs contained in that individual is satisfied.

[0045] First, the range of possible position pairs for each individual is determined based on the total number of vibration monitoring locations. Then, the total number of position pairs for each individual is randomly selected. Finally, a roulette wheel sampling without replacement is performed based on the normalized sampling probability, giving position pairs with higher absolute values ​​of the average Kendall rank correlation coefficient a greater probability of being selected. The generated initial population naturally leans towards highly correlated position pairs, significantly reducing the proportion of low-correlation and meaningless individuals generated in traditional random initialization. This data-prior information-guided initialization method significantly compresses the traversal space of invalid solutions and significantly improves the iterative solution efficiency of the subsequent parameter adaptive genetic algorithm. Compared to traditional purely random initialization, the overall process of determining the optimal position pair and constructing the unconditional binary Copula function converges faster and is more stable.

[0046] In some embodiments, the population is iteratively optimized using a parameter-adaptive genetic algorithm based on a first reliability assessment model to determine optimal multiple position pairs, including: In each iteration, the proportion of elites retained is determined.

[0047] Specifically, in each iteration, for any two individuals, their similarity can be calculated. For example, the ratio of the intersection to the union of multiple position pairs included in the two individuals can be used as the similarity between them. The average similarity of any two individuals in the current iteration is then calculated to obtain the mean similarity. Based on the mean similarity, the proportion of elites to be retained is determined.

[0048] For example, the elite retention rate can be calculated using the following formula: , in, To retain the elite proportion for the current iteration, Reserve a minimum elite percentage (e.g., 5%). Reserve a percentage for the highest elite (e.g., 20%). The preset mean similarity value, for example, 0.5. The mean similarity value for the current iteration. The adjustment coefficient (e.g., 1, 2, or 3) controls the rate at which the elite retention rate increases as the mean similarity decreases. To prevent tiny positive numbers with a denominator of 0, The cutoff function forces the elite retention rate to be limited to the minimum elite retention rate. and the highest elite retention rate between.

[0049] In some embodiments, the population is iteratively optimized using a parameter-adaptive genetic algorithm based on a first reliability assessment model to determine optimal multiple position pairs, including: Using the first reliability assessment model, the multivariate joint reliability corresponding to each test condition is calculated based on the test data corresponding to various test conditions. In each iteration, for each individual, a second unconditional binary Copula function for multiple position pairs included in the individual is constructed. The independent joint reliability of the second unconditional binary Copula function for multiple position pairs included in the individual for each test condition is determined. Based on the multivariate joint reliability for each test condition and the independent joint reliability of the second unconditional binary Copula function for multiple position pairs included in the individual, the fitness of the individual is determined.

[0050] Specifically, for each test condition, for the vibration sequence at each vibration monitoring location, kernel density estimation is used to fit its marginal cumulative distribution function, mapping the acceleration in the vibration sequence to a uniform distribution space of [0,1], eliminating the interference of dimensional differences between different vibration monitoring locations on the dependency calculation. Based on the R-vine hierarchical topology already built in the first reliability assessment model, pre-trained unconditional binary Copula functions and conditional binary Copula functions of various orders are called to recursively calculate layer by layer starting from the first layer main vine tree: firstly, the binary joint distribution probability of adjacent vibration monitoring locations is obtained through multiple first unconditional binary Copula functions in the first layer, and then this result is used as a conditional variable to be input into the first-order conditional binary Copula function in the second layer, and the recursive calculation of the third-order, fourth-order and other higher-order conditional distributions is completed in sequence, and finally the high-dimensional joint distribution probability decomposition of all vibration measurement points is completed, and the multivariate joint cumulative distribution function of the vibration response of all vibration monitoring locations under the test condition is obtained. The entire recursive process strictly follows the R-vine dependency decomposition rule, and the type and parameters of the binary Copula function at each step use the optimal results selected in the previous stage through the AIC criterion. Combining the vibration failure thresholds of each vibration monitoring location obtained through experiments, the multivariate joint reliability under this working condition is calculated by numerical integration of the high-dimensional joint survival function, where the vibration response at all vibration monitoring locations does not exceed the safety threshold.

[0051] The method for constructing the second unconditional binary Copula function of multiple position pairs included in an individual is similar to the method for constructing multiple first unconditional binary Copula functions, and will not be repeated here.

[0052] For each test condition, the product of the reliability values ​​calculated by the second unconditional binary Copula function for all position pairs can be used as the independent joint reliability corresponding to the test condition. For example, for a test condition containing 3 independent position pairs, the independent reliability values ​​output by the corresponding second unconditional binary Copula function are 0.982, 0.976, and 0.991, respectively. The product of these three values ​​is 0.982 × 0.976 × 0.991 ≈ 0.949, therefore the independent joint reliability corresponding to the test condition is 0.949.

[0053] For each individual, the multivariate joint reliability and independent joint reliability corresponding to each test condition can be treated as two variables, substituted into the formula for calculating the correlation coefficient (e.g., Pearson correlation coefficient), and their absolute values ​​can be taken. The fitness of an individual is related to the absolute value of this correlation coefficient; for example, the larger the absolute value of the correlation coefficient, the higher the fitness of the individual.

[0054] Using the R-vine full-dimensional multivariate joint reliability as the benchmark truth, the location pair combination selected through iterative screening by genetic algorithm maximizes the absolute value of the Pearson correlation coefficient between its product-type independent joint reliability and the benchmark value. Under the premise of ensuring high consistency with high-dimensional full monitoring information, the modeling dimension is greatly compressed: there is no need to maintain a complex high-order condition Copula recursive link, and only a small number of independent binary Copulas can reproduce the dependency degradation law of all measurement points. The computational efficiency of traditional high-dimensional Vine-Copula is improved by more than 60%, which is fully adapted to the low computing power requirements of real-time reliability assessment at the edge of wind turbine bearings.

[0055] Figure 3 This is a flowchart illustrating the construction of a second reliability assessment model according to some embodiments of this specification, such as... Figure 3 As shown, in some embodiments, a second reliability assessment model is constructed based on multiple second unconditional binary Copula functions, including: Based on the test data corresponding to various test conditions, determine the independent joint reliability of the output of each second unconditional binary Copula function corresponding to each test condition; Using the first reliability assessment model, the multivariate joint reliability corresponding to each test condition is calculated based on the test data corresponding to various test conditions. Construct a multivariate nonlinear regression model, wherein the independent variables of the multivariate nonlinear regression model include the independent joint reliability calculated by multiple second unconditional binary Copula functions, and the dependent variable of the multivariate nonlinear regression model is the first reliability. Based on the independent joint reliability and multivariate joint reliability output of each second unconditional binary Copula function corresponding to each test condition, the parameters of the multivariate nonlinear regression model are solved, and a second reliability evaluation model is constructed.

[0056] Specifically, for example, a multivariate nonlinear regression model can be: , in, For the highest reliability, For bias terms, Let be the total number of the second unconditional binary Copula functions. For the first The first-order linear weights corresponding to the second unconditional binary Copula function. For the first The reliability of the output of the second unconditional binary Copula function. For the first The weights of the univariate quadratic terms corresponding to the second unconditional binary Copula function. For the first The second unconditional binary Copula function and the first The weights of the interaction terms between the second unconditional binary Copula functions For the first The reliability of the output of the second unconditional binary Copula function. The fitted residuals are those that conform to a normal distribution.

[0057] The Levenberg-Marquardt algorithm with physical constraints is used as the core iterator, balancing the local accuracy of gradient descent with the global convergence speed of the Gauss-Newton method. The process consists of three standardized steps: The first step is dataset preprocessing, dividing the test data for various experimental conditions into training and test sets in a 7:3 ratio; the second step is constrained parameter iteration, incorporating dual constraints: first, all weight parameters must be non-negative to avoid the physically unsound result where a decrease in the reliability of the second unconditional binary Copula function output leads to an increase in overall reliability; second, ... The first step is to add an L2 regularization term to prevent the multivariate nonlinear regression model from overfitting the operating noise in the training set. The iteration continues until the mean square error between the predicted value and the baseline true value is less than 0.005, at which point it automatically stops. The third step is model solidification and verification. The converged parameters and the multiple pre-trained second unconditional binary Copula functions are encapsulated into a whole to complete the construction of the second reliability assessment model. The model is then verified through a test set. The absolute difference between the output of the multivariate nonlinear regression model and the baseline output of the first reliability assessment model must not exceed a threshold (e.g., 0.1), and the absolute difference of more than 90% of the samples must be controlled within 0.05.

[0058] Figure 4 These are schematic diagrams illustrating the results of the first reliability assessment model and the experimental results of the second reliability assessment model, based on some embodiments shown in this specification. Figure 4 As shown, the solid blue line represents the multivariate joint reliability output by the first reliability assessment model, and the dashed red line represents the first reliability output by the second reliability assessment model. The two curves largely overlap, both exhibiting a stable nonlinear decay trend with increasing bearing operating time. Only extremely slight numerical deviations exist in localized small intervals, with no significant deviations throughout. This verifies that the second reliability assessment model constructed through multivariate nonlinear regression can accurately reproduce the calculation results of the high-dimensional first reliability assessment model, fully meeting the reliability assessment accuracy requirements for long-term wind turbine bearing operation.

[0059] This system breaks away from the traditional high-dimensional Copula modeling, which must strictly follow the nested rules of the vine tree hierarchy. It no longer requires calculating complex conditional partial correlation coefficients layer by layer. Instead, it directly uses the independent joint reliability of multiple second unconditional binary Copulas as input variables. This fully preserves the advantage of binary Copulas in characterizing the nonlinear and asymmetric vibration dependence between measurement points, while completely avoiding the problem of error accumulation layer by layer in traditional nested calculations. This significantly reduces the computational complexity of high-dimensional reliability.

[0060] The vibration monitoring module is used to collect vibration sequences from multiple vibration monitoring locations of the wind turbine bearing.

[0061] The reliability analysis module is used to calculate the first reliability based on the vibration sequence at multiple vibration monitoring locations using a second reliability assessment model.

[0062] Specifically, the reliability of multiple second unconditional binary Copula functions can be calculated based on the vibration sequences of multiple vibration monitoring locations of the wind turbine bearing. The reliability of these multiple second unconditional binary Copula functions can then be substituted into the second reliability assessment model, which outputs the first reliability.

[0063] The reliability analysis module is also used to determine whether a bearing reliability risk has occurred based on the first reliability level. If so, the reliability of the wind power bearing is calculated using the first reliability assessment model.

[0064] In some embodiments, determining whether a bearing reliability risk has occurred based on a first reliability level includes: Determine if the first reliability is less than the first reliability threshold (e.g., 0.92). If so, determine if a bearing reliability risk has occurred. If not, determine whether a bearing reliability risk has occurred based on the reliability calculated by multiple second unconditional binary Copula functions.

[0065] Specifically, the standard deviation of the reliability outputs of multiple second unconditional binary Copula functions is calculated. When the standard deviation is greater than the standard deviation threshold, it is determined that there is a risk of bearing reliability failure. This standard deviation threshold can be determined through full-life test data. For example, the standard deviation threshold can be 0.03. Under normal operating conditions, the degradation synchronization of each vibration monitoring position pair is strong, the reliability value deviation is extremely small, and the standard deviation will be stable within 0.02. When the wind turbine bearing has local hidden damage, the vibration abnormality of a single vibration monitoring position will break the degradation synchronization of the cross-measuring point pair, directly increasing the dispersion of the reliability of this group. When the standard deviation exceeds 0.03, it can be determined that there is an early hidden fault, triggering an early warning of bearing reliability risk. The reliability of the wind turbine bearing is calculated through the first reliability assessment model.

[0066] Finally, it should be understood that the embodiments described in this specification are merely illustrative of the principles of the embodiments described herein. Other variations may also fall within the scope of this specification. Therefore, alternative configurations of the embodiments described herein are intended to be illustrative rather than limiting, and should be considered consistent with the teachings of this specification. Accordingly, the embodiments described herein are not limited to those explicitly introduced and described herein.

Claims

1. A reliability analysis system for wind turbine bearings based on multivariate degradation analysis, characterized in that, include: The location determination module is used to determine multiple vibration monitoring locations of the wind turbine bearing; The model building module is used to construct multiple first unconditional binary Copula functions and multiple conditional binary Copula functions of different orders based on multiple vibration monitoring locations using the Vine-Copula algorithm. Based on the multiple first unconditional binary Copula functions and multiple conditional binary Copula functions of different orders, a first reliability assessment model is constructed. The model building module is also used to construct multiple second unconditional binary Copula functions based on the first reliability assessment model and multiple vibration monitoring locations, and to construct a second reliability assessment model based on the multiple second unconditional binary Copula functions; The vibration monitoring module is used to collect vibration sequences from multiple vibration monitoring locations of the wind turbine bearing; The reliability analysis module is used to calculate the first reliability based on the vibration sequence of multiple vibration monitoring locations using a second reliability assessment model. The reliability analysis module is also used to determine whether a bearing reliability risk has occurred based on the first reliability. If so, the reliability of the wind power bearing is calculated using the first reliability assessment model. Specifically, using the Vine-Copula algorithm, multiple first unconditional binary Copula functions and multiple conditional binary Copula functions of different orders are constructed based on multiple vibration monitoring locations. Based on these multiple first unconditional binary Copula functions and multiple conditional binary Copula functions of different orders, a first reliability assessment model is constructed, including: For any two vibration monitoring locations, the average Kendall rank correlation coefficient between the two vibration monitoring locations is calculated based on the processed test data corresponding to various test conditions. A vine tree is constructed based on the average Kendall rank correlation coefficient between any two vibration monitoring locations; Based on the vine tree, multiple first unconditional binary Copula functions and multiple conditional binary Copula functions of different orders are constructed to build the first reliability evaluation model; Based on the first reliability assessment model and multiple vibration monitoring locations, several second unconditional binary Copula functions are constructed, including: The population is initialized based on the average Kendall rank correlation coefficient between any two vibration monitoring locations, where an individual consists of multiple location pairs, and each location pair consists of two vibration monitoring locations. Using a parameter-adaptive genetic algorithm, the population is iteratively optimized based on the first reliability evaluation model to determine the optimal multiple position pairs; Based on the optimal multiple position pairs, construct multiple second unconditional binary Copula functions; Based on the first level of reliability, determine whether a bearing reliability risk has occurred, including: Determine if the first reliability is less than the first reliability threshold; if so, determine if a bearing reliability risk has occurred. If not, determine whether a bearing reliability risk has occurred based on the reliability calculated by multiple second unconditional binary Copula functions.

2. The wind turbine bearing reliability analysis system based on multivariate degradation analysis according to claim 1, characterized in that, Multiple vibration monitoring locations for wind turbine bearings were determined, including: Test data were collected for various test conditions, including vibration sequences at multiple locations of the wind turbine bearing. Based on test data corresponding to various test conditions, multiple vibration monitoring locations were determined from multiple positions of the wind turbine bearing.

3. The wind turbine bearing reliability analysis system based on multivariate degradation analysis according to claim 1, characterized in that, Based on test data corresponding to various test conditions, multiple vibration monitoring locations were determined from multiple positions of the wind turbine bearing, including: For any two locations, the average Kendall rank correlation coefficient between the two locations is calculated based on test data corresponding to various test conditions. Based on the average Kendall rank correlation coefficient between any two positions, multiple valid positions are selected from multiple positions of the wind turbine bearing. For any two valid locations, the average similarity between the two valid locations is calculated based on the test data corresponding to various test conditions. Multiple vibration monitoring locations are selected from a pool of valid locations based on the average similarity between any two valid locations.

4. The wind turbine bearing reliability analysis system based on multivariate degradation analysis according to claim 1, characterized in that, The population is initialized based on the average Kendall rank correlation coefficient between any two vibration monitoring locations, including: Based on the average Kendall rank correlation coefficient of any two vibration monitoring locations, determine the sampling probability of the location pair corresponding to any two vibration monitoring locations. The population is initialized based on the sampling probability of the location pairs corresponding to any two vibration monitoring locations.

5. The wind turbine bearing reliability analysis system based on multivariate degradation analysis according to claim 4, characterized in that, Using a parameter-adaptive genetic algorithm, the population is iteratively optimized based on a first reliability assessment model to determine multiple optimal position pairs, including: In each iteration, the proportion of elites retained is determined.

6. The wind turbine bearing reliability analysis system based on multivariate degradation analysis according to claim 5, characterized in that, Using a parameter-adaptive genetic algorithm, the population is iteratively optimized based on a first reliability assessment model to determine multiple optimal position pairs, including: Using the first reliability assessment model, the multivariate joint reliability corresponding to each test condition is calculated based on the test data corresponding to various test conditions. In each iteration, for each individual, a second unconditional binary Copula function for multiple position pairs included in the individual is constructed. The independent joint reliability of the second unconditional binary Copula function for multiple position pairs included in the individual for each test condition is determined. Based on the multivariate joint reliability for each test condition and the independent joint reliability of the second unconditional binary Copula function for multiple position pairs included in the individual, the fitness of the individual is determined.

7. The wind turbine bearing reliability analysis system based on multivariate degradation analysis according to any one of claims 2-6, characterized in that, A second reliability assessment model is constructed based on multiple second unconditional binary Copula functions, including: Based on the test data corresponding to various test conditions, determine the independent joint reliability of the output of each second unconditional binary Copula function corresponding to each test condition; Using the first reliability assessment model, the multivariate joint reliability corresponding to each test condition is calculated based on the test data corresponding to various test conditions. Construct a multivariate nonlinear regression model, wherein the independent variables of the multivariate nonlinear regression model include independent joint reliability calculated by multiple second unconditional binary Copula functions, and the dependent variable of the multivariate nonlinear regression model is the first reliability. Based on the independent joint reliability and multivariate joint reliability output of each second unconditional binary Copula function corresponding to each test condition, the parameters of the multivariate nonlinear regression model are solved, and a second reliability evaluation model is constructed.

Images