Fault early warning method and system for rolling bearing of offshore wind turbine generator

Abstract

The invention discloses an offshore wind turbine generator rolling bearing fault early warning method and system, and belongs to the technical field of wind power monitoring. Since different data parameters often have different dimensions and orders of magnitude, the used measurement units may affect data analysis, resulting in completely different results. In order to avoid dependence on measurement unit selection, the data is transformed to fall into a smaller common interval, and meanwhile, the model can be converged more quickly. Then, an improved LightGBM algorithm is adopted to carry out feature extraction on the offshore wind turbine generator rolling bearing data obtained after data transformation, ISOMAP is adopted to replace Euclidean distance in multi-dimensional scale transformation with geodesic distance, distance distribution of data points is kept unchanged, low-dimensional embedding in a high-dimensional space is mined, manifold learning dimensionality reduction of original data is completed, and the original data is obtained. And finally, a support vector machine is adopted to realize fault identification more conveniently. Therefore, the method provided by the invention can solve the problems in the prior art.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wind power monitoring and early warning, and relates to a fault early warning method and system for a rolling bearing of an offshore wind turbine generator. BACKGROUND

[0002] At present, offshore units present the trend of larger single machine capacity and larger impeller diameter, and are installed in deeper water. In addition, offshore wind turbine generators are located in complex and changeable offshore environments, and the atmospheric area has high humidity and high salt spray, the splash area has frequent dry and wet alternation, the underwater area is soaked in seawater for a long time and has serious water organism attachment, which seriously challenges the long-term safe and stable operation of offshore wind power equipment.

[0003] The main components with high failure rate of wind turbine generators are converters, generators, rolling bearings and the like, which account for about 80%-90% of the total cost of wind turbine generators, and the average troubleshooting time of these failures is about 1-7 days, of which the blade can be up to about 10 days, which seriously restricts the output of wind turbine generators. With the installation of deeper water and the distance from the shore farther and farther, the complex operating environment of offshore and the high-difficulty maintenance conditions will also lead to longer maintenance time.

[0004] At present, the existing fault diagnosis method of offshore wind turbine generators is mainly offline, that is, fault diagnosis generally occurs after failure or during regular maintenance, which has significant lag in fault discovery and diagnosis, resulting in further increase of maintenance cost. Due to the harsh offshore operating environment, real-time fault diagnosis is very important for offshore wind turbine generators, and therefore fault early warning is increasingly valued. Effective fault early warning can warn before failure occurs and take measures to prevent failure.

[0005] The SCADA system of offshore wind turbine generators records dozens of operating parameters collected by different sensors, but although the existing fault diagnosis research considers the correlation between various operating parameters of the unit, it only retains input parameters higher than the set correlation threshold, and the threshold setting depends on expert experience, which has great subjectivity. At the same time, retaining only part of the input parameters will cause the loss of parameter correlation information, resulting in reduced accuracy of normal behavior monitoring and early warning. SUMMARY

[0006] The purpose of the present application is to solve the problems in the prior art and provide a fault early warning method and system for a rolling bearing of an offshore wind turbine generator.

[0007] To achieve the above purpose, the following technical solutions are adopted:

[0008] The fault early warning method for a rolling bearing of an offshore wind turbine generator provided by the present application comprises the following steps:

[0009] The pre-processed offshore wind turbine rolling bearing data is acquired, data transformation is performed on the pre-processed offshore wind turbine rolling bearing data, and the transformed offshore wind turbine rolling bearing data is acquired;

[0010] An improved LightGBM model is used to extract features from the transformed offshore wind turbine rolling bearing data, and wind turbine rolling bearing features are acquired;

[0011] The wind turbine rolling bearing features are dimensionally reduced, and a support vector machine is used to identify the reduced features to acquire a fault rolling bearing, thereby realizing wind turbine rolling bearing fault early warning.

[0012] Preferably, the method for acquiring pre-processed offshore wind turbine rolling bearing data is as follows:

[0013] The offshore wind turbine rolling bearing data is sequentially subjected to data filtering, data auditing, data desensitization, and data cleaning, thereby acquiring pre-processed offshore wind turbine rolling bearing data.

[0014] Preferably, the pre-processed offshore wind turbine rolling bearing data is subjected to linear transformation using min-max standardization.

[0015] Preferably, the improved LightGBM model optimizes the tree growth strategy as follows:

[0016] First, a leaf-growth algorithm with depth limitation is used;

[0017] Second, a depth limitation is added to the leaf-growth algorithm.

[0018] Preferably, the step of dimensionally reducing the wind turbine rolling bearing features is as follows:

[0019] For an original data set X = [x1, x2, …, xn], calculate the Euclidean distance d n , T ∈R n×m between all sample points d ij ,i,j = 1,2, …, n, determine the k nearest neighbors of x i and construct a neighborhood graph matrix G n×n = {d ij}, and n is the number of data points.

[0020] Based on the neighborhood graph matrix G n×n , use Dijkstra's algorithm to determine the shortest distance between any two points as the geodesic distance, and construct a geodesic matrix D G (i,j) = {d G (i,j)}.

[0021] computing a distance matrix corresponding to the geodesic matrix centralizing the Gram matrix wherein H is a centralizing matrix;

[0022] spectrally decomposing the matrix B into B=PΛP T wherein Λ=diag(λ1, λ2, …, λ n ), λ1≥λ2≥…≥λ n is a diagonal matrix of eigenvalues of B, and P is an orthogonal eigenvector matrix of B, and can reduce m-dimensional high-dimensional data to k-dimensional data, 1≤k≤m, P k is the first k columns of P

[0023] Preferably, the calculation method of Dijkstra algorithm is as follows:

[0024]

[0025] Preferably, the centralizing matrix I is an n-dimensional unit matrix, and the symmetric matrix E=[1,…n] T ∈R n .

[0026] The application provides a sea wind turbine rolling bearing fault early warning system, which comprises:

[0027] A data preprocessing module is configured to acquire preprocessed sea wind turbine rolling bearing data, perform data transformation on the preprocessed sea wind turbine rolling bearing data, and acquire transformed sea wind turbine rolling bearing data.

[0028] A rolling bearing feature extraction module is configured to extract features from the transformed sea wind turbine rolling bearing data by using an improved LightGBM model, and acquire wind turbine rolling bearing features.

[0029] A fault identification module is configured to perform dimension reduction processing on the wind turbine rolling bearing features, identify the dimension-reduced features by using a support vector machine, acquire a fault rolling bearing, and realize wind turbine rolling bearing fault early warning.

[0030] An electronic device comprises a memory and a processor, the memory stores a computer program, and the processor realizes the steps of the sea wind turbine rolling bearing fault early warning method when executing the computer program.

[0031] A computer readable storage medium stores a computer program, and the computer program realizes the steps of the sea wind turbine rolling bearing fault early warning method when executed by a processor.

[0032] Compared with the prior art, the present application has the following beneficial effects:

[0033] The offshore wind turbine rolling bearing fault early warning method provided by the present application avoids the dependence on the selection of the measurement unit by transforming the data so that it falls within a smaller common interval, and this also facilitates faster convergence of the model. The improved LightGBM algorithm is used to extract features from the offshore wind turbine rolling bearing data after data transformation, ISOMAP is used to replace the Euclidean distance in the multidimensional scaling transformation by using geodetic distance to maintain the distance distribution of the data points unchanged, to mine the low-dimensional embedding in the high-dimensional space, to complete the manifold learning dimension reduction of the original data, and to facilitate the use of support vector machines to realize fault recognition. Therefore, the method provided by the present application can solve the problems existing in the prior art.

[0034] Further, the min-max standardization method is used to perform linear transformation on the original data, and the purpose is to keep the data in a consistent standardized form.

[0035] The offshore wind turbine rolling bearing fault early warning system provided by the present application realizes the early warning of the rolling bearing fault of the wind turbine by dividing the system into a data preprocessing module, a rolling bearing feature extraction module and a fault recognition module. The modularization idea is used to make each module independent of each other, facilitating unified management of each module. BRIEF DESCRIPTION OF DRAWINGS

[0036] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope, and for those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0037] Figure 1 The offshore wind turbine rolling bearing fault early warning method flow chart of the present application.

[0038] Figure 2 The unit data chart containing the shutdown state of the present application.

[0039] Figure 3 The unit data chart containing the temperature anomaly of the present application.

[0040] Figure 4 The data comparison chart before and after cleaning of the present application.

[0041] Figure 5A LightGBM sub-node histogram splitting schematic diagram of the present application.

[0042] Figure 6 A level-wise strategy diagram of the present application.

[0043] Figure 7 A leaf-wise strategy diagram of the present application.

[0044] Figure 8 A GBDT training flowchart of the present application.

[0045] Figure 9 A top ten feature importance ranking diagram of the present application.

[0046] Figure 10 An ISOMAP dimensionality reduction degradation trend index diagram of the present application.

[0047] Figure 11 A plurality of dimensionality reduction method degradation trend index diagrams of the present application ((a) is a PCA degradation trend index, (b) is a LLE degradation trend index, (c) is a T-SNE degradation trend index, and (d) is a UMAP degradation trend index).

[0048] Figure 12 A offshore wind turbine rolling bearing fault early warning system diagram of the present application.

[0049] Figure 13 A structural schematic diagram of an electronic device. DETAILED DESCRIPTION

[0050] The present application will be further described in detail below with reference to the accompanying drawings:

[0051] A offshore wind turbine rolling bearing fault early warning method provided by the present application, as shown in the figure, comprises the following steps: Figure 1

[0052] S1, obtaining pre-processed offshore wind turbine rolling bearing data, performing data transformation on the pre-processed offshore wind turbine rolling bearing data, and obtaining transformed offshore wind turbine rolling bearing data;

[0053] The method for obtaining the pre-processed offshore wind turbine rolling bearing data is as follows:

[0054] The offshore wind turbine rolling bearing data is sequentially subjected to data filtering, data auditing, data desensitization and data cleaning, so as to obtain the pre-processed offshore wind turbine rolling bearing data.

[0055] The pre-processed offshore wind turbine rolling bearing data is subjected to linear transformation by using min-max standardization.

[0056] ​S2, using the improved LightGBM model to extract features from the transformed offshore wind turbine rolling bearing data, and obtaining wind turbine rolling bearing features;

[0057] The improved LightGBM model optimizes the tree growth strategy as follows:

[0058] First, the leaf growth algorithm with depth limitation is used;

[0059] Second, a depth limit is added to the leaf growth algorithm.

[0060] S3, dimensionality reduction processing is performed on the wind turbine rolling bearing features, and support vector machines are used to identify the dimensionality reduced features to obtain the fault rolling bearing, and realize the fault early warning of the wind turbine rolling bearing.

[0061] The steps of dimensionality reduction processing of the wind turbine rolling bearing features are as follows:

[0062] For the original data set X=[x1,x2,…,x n ] T ∈R n×m , the Euclidean distance d ij ,i,j=1,2,…,n between all sample points is calculated, k nearest neighbors of x i are determined and a neighborhood graph matrix G n×n ={d ij} is constructed, and n is the number of data points;

[0063] Based on the neighborhood graph matrix G n×n , the shortest distance between any two points is determined as the geodesic distance using Dijkstra algorithm, and a geodesic matrix D G (i,j)={d G (i,j)} is constructed;

[0064] The calculation method of Dijkstra algorithm is as follows:

[0065] The distance matrix corresponding to the geodesic matrix is calculated Centered Gram matrix Where H is the centering matrix; the centering matrix I is an n-dimensional identity matrix, and the symmetric matrix E=[1,…n] T ∈R n .

[0066] Spectral decomposition is performed on the matrix B B=P∧P T , where ∧=diag(λ1,λ2,…,λ n ), λ1≥λ2≥…≥λ nis a diagonal matrix of eigenvalues of B, P is an orthogonal eigenvector matrix of B, and P k is the first k columns of P

[0067] The method is described in detail as follows:

[0068] Step 1, data preprocessing

[0069] (1) Original data access

[0070] Original data access is a process of extracting data from a data source. The original data access stage mainly focuses on:

[0071] 1) Understand the data volume: By understanding the data volume (batch processing or streaming data), determine the storage method after data access according to the data volume size and platform planning.

[0072] 2) Understand the scope of data acquisition: such as full acquisition or incremental acquisition. Full extraction is similar to data mirroring or data replication, which extracts the data in the table or view from the database without changing the data. Incremental extraction refers to the extraction of newly added or modified data in the database after the last extraction.

[0073] 3) Understand the metadata: including two parts: one is to look at the metadata, including field explanation, data source, code table and all other information describing the data; if the data is multi-dimensional, understand the relationship between the data. Two is to extract a part of the data and use manual viewing to have a direct understanding of the data itself and preliminarily find some problems to prepare for the subsequent processing.

[0074] (2) Data filtering

[0075] Data filtering is to preliminarily implement filtering operation on the data in the business data that does not meet the application rules or invalid data, so as to unify the data standard.

[0076] (3) Data auditing rules

[0077] Data auditing rules include: data missing, data duplication, illegal code cleaning, illegal value cleaning, data format cleaning, data reference, data consistency, data relationship, data value domain, record number checking, business constraint checking, etc.

[0078] 1) Data missing

[0079] The field is required to be non-empty, and the data in the field needs to be checked. Such as null value, incomplete data, etc. For time series data with large data volume, the data volume value in a unit time can be used to judge whether there is a null value, and the auditing rule configuration can be performed accordingly.

[0080] 2) Data dead value

[0081] For time series data, the field data needs to be checked for changes. For example, dead value. For large amounts of time series data, the number of different values in a unit of time can be used to determine the number of different values. Based on this, the dead value audit rule can be configured.

[0082] 3) Data duplication

[0083] Data sources may cause data duplication under certain abnormal conditions, such as non-unique conditions, multiple triggers, retransmission, etc. Depending on the data type and business characteristics, there may also be single-field duplication, multi-field duplication, etc.

[0084] 4) Illegal code, illegal value cleaning

[0085] Illegal code problems include illegal code, code and data standards, etc. Illegal value problems include value errors, format errors, extra characters, and garbled characters, etc. Depending on the specific situation, it needs to be checked and corrected.

[0086] 5) Data format audit

[0087] Check if the format of the attribute value in the table is correct to measure its accuracy.

[0088] Including but not limited to: address format audit, case conversion, full-width and half-width conversion, special character audit, mobile phone number audit, ID card audit, ID conversion, date format, time format, region format, fixed phone format, organization format, currency format, extra characters, garbled characters, and other business custom formats.

[0089] 6) Data reference audit

[0090] Data reference is mainly to verify specific fields according to standard specifications. It includes id reference audit, code reference audit, and business attribute reference audit.

[0091] 7) Data consistency audit

[0092] Data consistency verification includes single-table data consistency audit and multi-table data consistency audit.

[0093] 8) Data relationship

[0094] Data relationship includes column data logical relationship audit, table data logical relationship audit, and multi-table data logical relationship audit.

[0095] 9) Data value range audit

[0096] Data value range audit includes value range audit, maximum value audit, and business logic value range audit.

[0097] 10) Record number check

[0098] Refers to the total number of data between the relevant data of each system or the fluctuation check of the daily data volume in the data table.

[0099] 11) Business constraint check

[0100] Business constraint check should be determined together with business personnel during implementation. Business personnel consider data check rules from the correctness, consistency, effectiveness, and other aspects of business.

[0101] (4) Data audit steps

[0102] 1) Null value audit

[0103] Null value audit mainly uses the amount of data in a unit of time to audit whether there is a null value, such as the audit rule is configured as more than 3000 data volume in an hour. The specific steps are as follows:

[0104] ① Create a null value audit table, including point number and quantity, and other key fields.

[0105] ② Read the hour data table, aggregate the data by point number, and store it in the null value audit table in step ①, such as the data volume of AAAA.BBBB in this hour is 3600.

[0106] ③ Configure the process of step ② into the scheduling process, execute it once an hour, and store the data volume of each point number per hour.

[0107] ④ Configure the audit rule for the null value audit table in step ①, such as the data volume field needs to be greater than 3000. If the data in this field appears to be less than 3000, it is considered that there is a null value in the data.

[0108] ⑤ Configure the audit rule into the scheduling process in step ③, and you can view the hour data that does not meet the audit rule in the audit result.

[0109] 2) Dead value audit

[0110] Dead value audit mainly uses the number of different data values in a unit of time to audit dead values, such as the audit rule is configured as more than 2 data volume in an hour, indicating that the number of different data values in a unit of time is greater than 2, which means that the measurement point does not exist. The specific steps are as follows:

[0111] ① Create a dead value audit table, including point number, value, and quantity, and other key fields.

[0112] ② Read the hour data table, aggregate the number according to the point number and data value, and store it in the dead value audit table in step ①. For example, if the point number is AAAA.BBBB and the value is "66" in this hour, the data volume is "10".

[0113] ③ Configure the process of step ② into the scheduling process, execute it once an hour, and store the data volume of each value of each point number per hour.

[0114] ④ Configure the audit rules for the empty value audit table in step ①, such as the data volume field must be greater than 2. If the data in this field appears a value greater than 2, it is considered as a dead value.

[0115] ⑤ Configure the audit rules into the scheduling process of step ③, and view the hour data that does not meet the audit rules in the audit results.

[0116] Step 2, data desensitization

[0117] (1) Desensitization plan

[0118] It is advisable to make overall planning for data desensitization work, develop a complete data desensitization work plan, and train relevant parties who may come into contact with desensitized data on data desensitization procedures, and regularly evaluate and maintain the content of data desensitization procedures. When developing a data desensitization work plan, the following factors should be considered:

[0119] 1) Clearly define the sensitive data management department and its security responsibilities and obligations;

[0120] 2) Establish a classification and grading system for sensitive data, and regularly maintain and update the desensitization tool operation and management system;

[0121] 3) Establish a data security management mechanism, such as code security, audit security, and security management;

[0122] 4) Regularly train relevant parties on data desensitization work;

[0123] 5) Develop a complete sensitive data use approval process to ensure the safe and compliant use of sensitive data;

[0124] 6) Clearly define the data desensitization process, including identifying sensitive data, identifying sensitive data, and determining desensitization methods.

[0125] (2) Desensitization process

[0126] 1) Identify sensitive data

[0127] It is advisable to thoroughly sort out the information contained in the data and clearly identify the sensitive information. Identifying sensitive data includes but is not limited to:

[0128] ① Clearly define the scope of data desensitization work;

[0129] ②Data in the working range are sorted and classified;

[0130] ③A sensitive data location and relationship database is established to save the location of sensitive data and the association between sensitive data and original data;

[0131] ④According to business needs, manual or automatic identification methods are selected, and the applicability of the identification methods to mainstream database systems, data warehouse systems, file systems, and new storage systems in cloud computing environments is considered;

[0132] ⑤Data discovery tools are selected, and their scalability is considered. The discovery logic of sensitive data can be customized according to business needs;

[0133] ⑥The field name, field type, field length, and assignment specification of sensitive information are clearly defined;

[0134] ⑦Using anti-association methods, find the mapping of another sensitive field that can be inferred from some non-sensitive fields, and identify these non-sensitive fields, for example: in the scenario where the identity card number can be inferred from the date of birth, the date of birth needs to be identified.

[0135] 2) Identify sensitive data

[0136] After identifying sensitive data, it is appropriate to identify the format, location, and other information of sensitive data as soon as possible. The selection of identification methods should consider the following factors:

[0137] ①Sensitive data identification information can flow with sensitive data;

[0138] ②Sensitive data identification information is not easily deleted and tampered with by malicious attackers;

[0139] ③Consider convenience and security to make it easy to identify data after identification;

[0140] ④Support sensitive identification for different data types (such as static data and dynamic data);

[0141] ⑤All non-sensitive fields that may generate sensitive data are also identified, for example: in patient treatment records, the correspondence between the name and the condition is hidden, and the "name" is transformed as a sensitive field. However, if the "name" can be derived from the uniqueness of a certain "address", the "address" needs to be identified and desensitized.

[0142] 3) Determine the desensitization scenario

[0143] Based on the identification of sensitive data, determine the desensitization scenario, which includes but is not limited to:

[0144] ① Static desensitization: After the original data is desensitized once, the desensitized result data can be used multiple times;

[0145] ② Dynamic desensitization: A data desensitization method that screens data according to different user needs, requiring the system to have security measures to ensure that users cannot bypass the data desensitization level to directly access sensitive data.

[0146] 4) Define desensitization rules

[0147] Define desensitization rules and solidify common data desensitization rules to avoid repeated definition. The current desensitization rule in Huaneng field is to replace the station information with an asterisk to shield it. In the future, this rule can be adjusted, such as processing the station information according to the defined rules.

[0148] 5) Perform desensitization operation

[0149] Desensitization operation needs to follow relevant laws and regulations, industry supervision specifications or standards such as personal privacy protection and data security protection. Personal sensitive information security should follow the relevant provisions in GB / T 35273. According to the defined data desensitization rules, data desensitization operation includes but is not limited to:

[0150] ① Define desensitization rules, such as desensitizing provinces or stations represented by letters to numbers;

[0151] ② Establish rule configuration information table to facilitate subsequent rule adjustment.

[0152] ③ Establish desensitization library table with the same structure but different names;

[0153] ④ Match the configuration information in step ② with the desensitization rules and configure them into the data scheduling process;

[0154] ⑤ Monitor and analyze the desensitization process;

[0155] ⑥ Automatic running of desensitization tasks.

[0156] Step 3, data cleaning

[0157] Through cleaning and reconstruction of global equipment historical operation data, on this basis, big data analysis and machine learning modeling are carried out, focusing on real-time state evaluation and short-term (24 hours) deterioration trend prediction method of equipment.

[0158] The equipment that has completed big data modeling mainly includes the transmission chain, blades, tower of the wind turbine generator and the like. The big data cleaning refers to the process of converting the original data into consistent data that can be used for modeling. The data in the real world is generally incomplete and inconsistent, and direct data mining often results in unsatisfactory results. Low-quality data will lead to low-quality mining results. In order to improve the quality of modeling evaluation, the data generated by the global equipment is cleaned and transformed accordingly, and the consistency of all data in the application modeling is maintained.

[0159] The existing unit data can be continuous in time, but the unit has downtime, data acquisition anomaly and the like, and this part of data has no gain for state evaluation, as shown in Figure 2 At the same time, the data section with abnormal fluctuation of temperature and the like, as shown in Figure 3 can be cleaned by statistical method, and the effect is as shown in Figure 4

[0160] Step 4, data transformation

[0161] Different data parameters often have different dimensions and orders of magnitude, and the used measurement units may affect data analysis, thereby leading to completely different results. In order to avoid the dependence on the selection of measurement units, the data should be standardized, or called normalized. This involves transforming the data to fall into a smaller common interval, such as [0.0, 1.0], and at the same time, this also facilitates faster convergence of the model.

[0162] The standardization method standardizes the data and attempts to give all attributes equal weight, and thus is suitable for the case without prior knowledge of data. For the classification algorithm involving neural network or the classification and clustering based on distance measurement, standardization is very useful. Standardization is also suitable for the case without prior knowledge of data. There are many data normalization methods, such as min-max standardization, z-score standardization and decimal standardization. After the above standardization processing, the unit limitation of the data is removed, and the original data is converted into dimensionless pure numerical value, i.e. each index value is at the same order of magnitude, so that the indexes with different units or orders of magnitude can be compared, weighted and the like for comprehensive evaluation and analysis.

[0163] The present application adopts min-max standardization, also called deviation standardization, for linear change of the original data. Assuming that min and max are the minimum value and the maximum value of attribute A respectively. The min-max standardization is calculated by:

[0164]

[0165] Here, the same min and max values are used for the training data and the validation data, and the same standardization form is maintained.

[0166] ​Step 5, feature extraction using improved LightGBM model

[0167] LightGBM algorithm is an engineering practice method of GBDT, which is a distributed and high-performance gradient boosting framework based on decision tree, and can be used for a series of machine learning tasks such as classification and regression. LightGBM adopts a single-side gradient sampling strategy, focusing on samples with large gradients in model calculation, reducing the calculation of a large number of small gradient samples, thereby speeding up the training process of the model.

[0168] LightGBM uses a histogram algorithm to bin the original continuous features into discrete features, uses reshaped data to store the original floating-point data and index data, reduces the data memory occupation and calculation complexity; when splitting the decision tree, as shown in Figure 5 , the parent node histogram can be directly subtracted from the child node histogram to obtain another child node histogram, thereby reducing the model calculation amount.

[0169] The generation of decision tree child nodes includes level-wise and leaf-wise strategies, as shown in Figure 6 and Figure 7 . Specifically, each time a layer of leaf nodes is generated, the next layer of nodes is generated, and the leaf node with the maximum split gain is generated each time. LightGBM uses the leaf-wise strategy to generate child nodes, which can reduce the calculation amount, and in combination with the maximum depth limit of the decision tree, it can prevent overfitting in the generation of the decision tree. The LightGBM model measures the importance of features by the role of the features in the decision tree splitting, and the feature importance is equal to the number of times the feature is used as a splitting node and the split gain is positive during model training.

[0170] Gradient boosting tree combines boosting strategy and decision tree model, and obtains the final prediction result by training multiple rounds of decision trees and fusing the trained model. The boosting strategy uses the base model to fit the training data, and continues to use the base model to fit the residual between the true data and the predicted data, and repeatedly until the accuracy requirement is met, and finally the results of the multiple base models are weighted and fused to obtain the final result. GBDT uses CART tree as the base model, continuously fits the residual, and finally obtains the prediction result by weighting, and its training process is shown in Figure 8 .

[0171] For a training data set T = {(x1, y1), (x2, y2), …, (x N , y N )}, x i ∈R n , y i ∈R, N is the number of training set samples, and the loss function is L(y, f(x)). The GBDT algorithm is as follows:

[0172] 1) Initialize the decision tree:

[0173]

[0174] 2) For m = 1, 2, ..., M:

[0175] (a) For i = 1, 2, ..., N, calculate using the following formula:

[0176]

[0177] (b) for r mi Fit a regression tree to obtain the leaf node region R of the m-th tree. mi j = 1, 2, ..., J

[0178] (c) For j = 1, 2, ..., J, calculate the following formula:

[0179]

[0180] (d) Calculation results for the first m trees:

[0181]

[0182] 3) Obtain the final regression tree:

[0183]

[0184] Step 6: Data Dimensionality Reduction

[0185] ISOMAP, a globally optimized manifold learning dimensionality reduction method, is based on the core idea of ​​replacing Euclidean distance in Multidimensional Scaling (MDS) with geodesic distance, preserving the distance distribution of data points, and mining low-dimensional embeddings in the high-dimensional space to complete manifold learning dimensionality reduction of the original data. In other words, it seeks the feature mapping f: X→Y from the high-dimensional space to the low-dimensional space. The specific computational steps of the ISOMAP algorithm are as follows:

[0186] (1) For the original dataset X = [x1, x2, ..., x n ] T ∈R n×m Calculate the Euclidean distance d between all sample points. ij For i,j = 1, 2, ..., n, determine x using the KNN or ε-radius method. i Find the k nearest neighbors and construct the neighborhood graph matrix G. n×n ={d ij}, where n is the number of data points.

[0187] (2) Based on neighborhood graph matrix G n×n, the shortest distance between any two points is determined as the geodesic distance using Dijkstra algorithm to construct the geodesic matrix D G (i,j) = {d G (i,j)}, the calculation method of Dijkstra algorithm is shown in the following formula.

[0188]

[0189] (3) The distance matrix corresponding to the geodesic matrix is centered Gram matrix B, and the calculation method is shown in the following formula, wherein H is the centering matrix I is an n-dimensional unit matrix, and E = [1, … n] T ∈R n .

[0190]

[0191] (4) The matrix B is spectrally decomposed B = P ∧ P T , ∧ = diag (λ1, λ2, …, λ n ), λ1≥ λ2≥ … ≥ λ n is the eigenvalue diagonal matrix of B, and P is the orthogonal eigenvector matrix of B. The m-dimensional high-dimensional data can be reduced to k-dimensional data, 1≤ k≤ m, P k is the first k columns of P:

[0192]

[0193] Step 7, using support vector machine to identify the characteristics, and obtaining the fault condition of the rolling bearing.

[0194] The following examples are verified:

[0195] Feature selection based on LightGBM importance ranking

[0196] Through the above screening strategy, the multi-dimensional feature set Q is obtained, and the features are screened through the LightGBM model, the features with high importance ranking are retained, the features with low contribution to model training are deleted, and the redundancy of features is reduced. The training data uses 982x196-dimensional feature Q as input variable and 0-1 equally divided sequence as output variable, and the specific parameters of LightGBM model are shown in Table 1:

[0197] Table 1 LightGBM model parameters

[0198] Variable Parameter value Variable Parameter value boosting_type gbdt seed 2021 objective regression n_estimators 100 num_leaves 32 lamda_1 0.4 feature_fraction 0.9 lamda_2 0.5 bagging_fraction 0.9 num_boost_round 2000 learining_rate 0.05 verbose_eval 100 min_data_in_leaf 100 early_stopping_rounds 500

[0199] Train the LightGBM model on the feature set Q, and the top ten features in the feature importance ranking are as follows Figure 9As shown in the figure. Through the LightGBM feature importance analysis, when the feature dimension is high, there are a large number of features with low or even zero importance, which contribute little to model training, but introduce a lot of redundant information in subsequent feature fusion. Therefore, the LightGBM model importance is used to screen the feature set Q. After screening, the feature set P has only 50 dimensions, and part of the reserved features are shown in Table 2.

[0200] Table 2 Partial features of feature set P

[0201]

[0202]

[0203] The feature set P obtained by LightGBM has 50 dimensions, which is still difficult to intuitively represent the degradation state of the rolling bearing, and needs to be further fused by a dimension reduction method to construct a degradation index to represent the bearing degradation state. The advantages and disadvantages of the indexes constructed by PCA, LLE, T-SNE, ISOMAP and UMAP dimension reduction methods are judged by using the comprehensive health evaluation index. The monotonicity, robustness, trend, distinguishability and comprehensive evaluation index of the degradation index constructed by different dimension reduction methods are shown in Table 3, and the degradation trend of ISOMAP and the comparative dimension reduction method is shown in Figure 10 and Figure 11 .

[0204] Table 3 Health index of different dimension reduction methods

[0205] Dimension reduction method Monotonicity Robustness Trend Identifiability Comprehensive health evaluation index PCA 0.0112 0.7437 0.8231 0.8239 0.6005 LLE 0.0071 0.5372 0.1630 0.1758 0.2208 T-SNE 0.0377 0.6944 0.8704 0.8373 0.6100 ISOMAP 0.0133 0.7804 0.8719 0.9004 0.6419 UMAP 0.0133 0.7581 0.8365 0.8594 0.6168

[0206] From Table 3, it can be observed that the robustness, trend and distinguishability of the degradation index extracted by the ISOMAP dimension reduction method are higher than those of other dimension reduction methods, and the monotonicity is lower than that of the degradation index extracted by the T-SNE dimension reduction method. Therefore, the ISOMAP method is selected to fuse the features to represent the degradation state of the rolling bearing.

[0207] Example 2

[0208] The offshore wind turbine rolling bearing fault early warning system provided by the application, as shown in Figure 12 , comprises a data preprocessing module, a rolling bearing feature extraction module and a fault recognition module.

[0209] The data preprocessing module is used to acquire preprocessed offshore wind turbine rolling bearing data, perform data transformation on the preprocessed offshore wind turbine rolling bearing data, and acquire transformed offshore wind turbine rolling bearing data.

[0210] The rolling bearing feature extraction module is configured to extract features of the transformed offshore wind turbine rolling bearing data by using an improved LightGBM model, and obtain rolling bearing features of the wind turbine.

[0211] The fault identification module is configured to perform dimension reduction processing on the rolling bearing features of the wind turbine, identify the dimension-reduced features by using a support vector machine, obtain the rolling bearing with faults, and realize early warning of rolling bearing faults of the wind turbine.

[0212] Embodiment 3

[0213] Please refer to Figure 13 The electronic device 100 comprises a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on the at least one processor 102, and at least one communication bus 104.

[0214] The memory 101 can be used to store the computer program 103, and the processor 102 can realize the steps of the offshore wind turbine rolling bearing fault early warning method of embodiment 1 by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101. The memory 101 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application program required by a function (such as a sound playing function, an image playing function, etc.), etc.; the data storage area can store data (such as audio data) created according to the use of the electronic device 100, etc. In addition, the memory 101 can include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device.

[0215] The at least one processor 102 can be a central processing unit (CPU), and can also be other general-purpose processors, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic component, a discrete hardware component, or the like. The processor 102 can be a microprocessor or the processor 102 can also be any conventional processor or the like, which is a control center of the electronic device 100 and connects various parts of the entire electronic device 100 through various interfaces and lines.

[0216] The memory 101 in the electronic device 100 stores a plurality of instructions to implement a method for early warning of offshore wind turbine rolling bearing failure, and the processor 102 can execute the plurality of instructions to implement:

[0217] Obtaining preprocessed offshore wind turbine rolling bearing data, performing data transformation on the preprocessed offshore wind turbine rolling bearing data, and obtaining transformed offshore wind turbine rolling bearing data;

[0218] Performing feature extraction on the transformed offshore wind turbine rolling bearing data using an improved LightGBM model to obtain wind turbine rolling bearing features;

[0219] Performing dimensionality reduction processing on the wind turbine rolling bearing features, identifying the dimensionality reduced features using a support vector machine, obtaining a rolling bearing in failure, and implementing early warning of wind turbine rolling bearing failure.

[0220] Embodiment 4

[0221] The modules / units integrated in the electronic device 100, if implemented in the form of software function units and sold or used as independent products, can be stored in a computer readable storage medium. Based on such understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. When the computer program is executed by a processor, the steps of each method embodiment described above can be implemented. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or some intermediate forms. The computer readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, and read-only memory (ROM).

[0222] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0223] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system), and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that carries out the functions specified in one or more flows and / or blocks.

[0224] These computer program instructions can also be stored in a computer readable storage medium that can direct the computer or other programmable data processing devices to work in a specific manner, so that the instructions stored in the computer readable storage medium produce a manufactured product including instruction apparatus, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that carries out the functions specified in one or more flows and / or blocks.

[0225] These computer program instructions can also be loaded into a computer or other programmable data processing devices, so that a series of operational steps are performed on the computer or other programmable data processing devices to generate a computer implemented process, so that the instructions executed on the computer or other programmable data processing devices provide a process for implementing the functions specified in the flowchart Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0226] The above merely provides the preferred embodiments of the present application, but is not intended to limit the present application. The present application can have various modifications and changes for those skilled in the art. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for early warning of rolling bearing failure in offshore wind turbines, characterized in that, Includes the following steps: Acquire preprocessed rolling bearing data of offshore wind turbines, perform data transformation on the preprocessed rolling bearing data of offshore wind turbines, and obtain transformed rolling bearing data of offshore wind turbines. An improved LightGBM model was used to extract features from the transformed offshore wind turbine rolling bearing data to obtain the features of the wind turbine rolling bearings. The characteristics of rolling bearings in wind turbine units are reduced in dimension, and a support vector machine is used to identify the reduced characteristics to obtain faulty rolling bearings, thereby enabling early warning of rolling bearing faults in wind turbine units.

2. The method for early warning of rolling bearing failure in offshore wind turbines according to claim 1, characterized in that, The method for obtaining preprocessed rolling bearing data for offshore wind turbines is as follows: The rolling bearing data of offshore wind turbines is subjected to data filtering, data auditing, data desensitization and data cleaning in sequence to obtain preprocessed rolling bearing data of offshore wind turbines.

3. The method for early warning of rolling bearing failure in offshore wind turbines according to claim 1, characterized in that, The preprocessed rolling bearing data of offshore wind turbines were linearly transformed using min-max normalization.

4. The method for early warning of rolling bearing failure in offshore wind turbines according to claim 1, characterized in that, The improved LightGBM model optimizes the tree growth strategy as follows: First, a leaf-based growth algorithm with depth restrictions was used; Second, a depth constraint was added on top of the leaf growth algorithm.

5. The method for early warning of rolling bearing failure in offshore wind turbines according to claim 1, characterized in that, The steps for dimensionality reduction processing of the rolling bearing features of wind turbine units are as follows: For the original dataset Calculate the Euclidean distance d between all sample points. ij Given i, j = 1, 2, ..., n, determine x. i Find the k nearest neighbors and construct the neighborhood graph matrix G. n×n ={d ij }, where n is the number of data points; Based on the neighborhood graph matrix G n×n Dijkstra's algorithm is used to determine the shortest distance between any two points as the geodesic distance, and a geodesic matrix D is constructed. G (i,j)={d G (i,j)}; Calculate the distance matrix-centered Gram matrix corresponding to the geodesic matrix. Where H is the centered matrix; Perform spectral decomposition on matrix B: B = P∧P T , where ∧=diag(λ1,λ2,…,λ n ), λ1≥λ2≥…≥λ n Let P be the eigenvalue diagonal matrix of B, and let P be the orthogonal eigenvector matrix of B. This matrix can reduce m-dimensional high-dimensional data to k-dimensional data, where 1 ≤ k ≤ m. k The first k columns of P 6. The method for early warning of rolling bearing failure in offshore wind turbines according to claim 5, characterized in that, The calculation method of Dijkstra's algorithm is as follows:

7. The method for early warning of rolling bearing failure in offshore wind turbines according to claim 5, characterized in that, Centralized matrix Let I be an n-dimensional identity matrix, and let E = [1,…n] be a symmetric matrix. T ∈R n .

8. A fault early warning system for rolling bearings of offshore wind turbines, characterized in that, include: The data preprocessing module is used to acquire preprocessed offshore wind turbine rolling bearing data, perform data transformation on the preprocessed offshore wind turbine rolling bearing data, and acquire transformed offshore wind turbine rolling bearing data. A rolling bearing feature extraction module is used to extract features from the transformed offshore wind turbine rolling bearing data using an improved LightGBM model, thereby obtaining the rolling bearing features of the wind turbine. The fault identification module is used to perform dimensionality reduction processing on the features of the rolling bearings of the wind turbine, and to identify the dimensionality reduction features using a support vector machine to obtain the faulty rolling bearings, thereby realizing early warning of rolling bearing faults in wind turbines.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the offshore wind turbine rolling bearing fault early warning method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the offshore wind turbine rolling bearing fault early warning method as described in any one of claims 1 to 7.

Images