Wind turbine gearbox fault diagnosis method and system based on multi-modal data fusion

Abstract

The present application relates to fan equipment management technical field, especially in kind of wind turbine gear box fault diagnosis method and system based on multi-modal data fusion, including to the key structure identification of wind turbine gear box, and mark it as gear box key structure piece;Collect the vibration signal of each gear box key structure piece, and the impeller strain signal of wind turbine;The aerodynamic load feature is extracted to the impeller strain signal;The aerodynamic load feature is input to the load transmission evaluation model constructed in advance, obtains load influence transmission coefficient set;Based on load influence transmission coefficient set, the impeller influence correction is carried out to each vibration signal, and vibration signal correction data set is obtained;Multi-dimensional feature extraction is carried out to vibration signal correction data set, and the vibration feature after extraction is converted into three-dimensional vibration feature matrix;The three-dimensional vibration feature matrix is identified by the preset gear box fault identification model. It can eliminate the interference of aerodynamic load, improve the accuracy of fault diagnosis.

Description

Technical Field

[0001] This invention relates to the technical field of wind turbine equipment management, and in particular to a method and system for diagnosing wind turbine gearbox faults based on multimodal data fusion. Background Technology

[0002] As an important component of renewable energy, wind power has seen its installed capacity continue to expand, and the reliability and safety of wind power equipment have become key challenges for the industry's development. As the core component for transmitting wind energy, wind turbine gearboxes are subjected to harsh operating environments such as high speed, heavy load, and complex wind conditions for a long time. Faults such as gear wear, bearing pitting, and tooth surface scuffing occur frequently, accounting for more than 60% of all wind turbine failures. This seriously affects the unit's operating efficiency and may lead to safety accidents.

[0003] Existing wind turbine gearbox fault diagnosis methods primarily rely on the analysis of vibration signals from a single source, identifying gearbox fault characteristics through time-domain, frequency-domain, or time-frequency-domain feature extraction methods. While these methods can identify obvious gearbox faults to some extent, they neglect the dynamic impact of wind turbine rotor aerodynamic loads on gearbox vibration characteristics. These rotor aerodynamic loads are transmitted to the gearbox through the transmission system, exacerbating the vibration response of critical structural components. However, existing methods do not quantify the impact of this transmission process on the vibration signal, resulting in vibration characteristics containing non-fault load interference, which can easily lead to misdiagnosis of gearbox faults. Summary of the Invention

[0004] This invention provides a wind turbine gearbox fault diagnosis method and system based on multimodal data fusion that can eliminate aerodynamic load interference and improve fault diagnosis accuracy, effectively solving the problems in the background art.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a method for fault diagnosis of wind turbine gearboxes based on multimodal data fusion, comprising:

[0006] Key structural components of the wind turbine gearbox are identified and marked as key structural parts of the gearbox.

[0007] Vibration signals of each key structural component of the gearbox and the impeller strain signal of the wind turbine are collected; the impeller strain signal and the vibration signal are collected at the same time.

[0008] Extract aerodynamic load characteristics from the impeller strain signal;

[0009] The aerodynamic load characteristics are input into a pre-built load transmission evaluation model to obtain a set of load influence transmission coefficients;

[0010] Based on the load influence transmission coefficient set, impeller influence correction is performed on each vibration signal to obtain a vibration signal correction dataset.

[0011] Multidimensional feature extraction is performed on the vibration signal correction dataset, and the extracted vibration features are converted into a three-dimensional vibration feature matrix;

[0012] The gearbox fault type is obtained by using a preset gearbox fault identification model to identify faults in the three-dimensional vibration feature matrix.

[0013] In conjunction with the first aspect, in one possible design, the identification of key structures in the wind turbine gearbox and the marking of these key structural components as gearbox key structural parts includes:

[0014] Stress concentration factors at various structural locations were calculated using gearbox dynamics simulation. The structural locations with stress concentration factors greater than a preset threshold are selected; The calculation formula is:

[0015] ;

[0016] in, This represents the maximum stress value of the structure. This is the standard stress value;

[0017] Based on the statistical data of actual gearbox operation and maintenance failures, the structural parts whose failure frequency ranks within a preset position are marked as the intersection of the two as key structural components of the gearbox.

[0018] In conjunction with the first aspect, in one possible design, vibration signals of each key structural component of the gearbox, as well as strain signals of the wind turbine impeller, are collected, including:

[0019] Multiple orthogonal acceleration sensors are deployed on the surface of each key structural component of the gearbox to collect vibration signals;

[0020] Multiple strain gauges are uniformly deployed circumferentially at the connection between the impeller hub and the blades to collect impeller strain signals; the strain signal quantization formula is:

[0021] ;

[0022] in, This represents the change in strain gauge length. The initial length of the strain gauge; the sampling frequency is consistent with the sampling frequency of the vibration signal and the timestamp synchronization error. .

[0023] In conjunction with the first aspect, in one possible design, aerodynamic load characteristics are extracted from the impeller strain signal, including:

[0024] Time-domain features of the impeller strain signal are extracted to obtain strain peak value, strain root mean square and strain kurtosis;

[0025] Frequency domain features were extracted from the impeller strain signal to obtain the strain dominant frequency amplitude and strain spectrum entropy;

[0026] The time-domain features and frequency-domain features are fused to form an aerodynamic load feature vector, which serves as the aerodynamic load feature.

[0027] In conjunction with the first aspect, in one possible design, the pre-built load transmission evaluation model construction method includes:

[0028] Collect sample signals of impeller strain and corresponding vibration of key structural components of the gearbox under different wind speeds and pitch angles of the wind turbine.

[0029] A mapping function between impeller strain and vibration response is established based on the dynamic equations of the transmission system. ,in For vibration response, For impeller strain, The transfer coefficient is to be optimized;

[0030] Model training was performed using impeller strain sample signals, vibration sample signals, and mapping relationship functions. The model parameters were iteratively optimized using a particle swarm optimization algorithm to complete the construction of the load transmission evaluation model and obtain a set of load influence transmission coefficients. .

[0031] In conjunction with the first aspect, in one possible design, impeller influence correction is performed on each vibration signal based on the load influence transmission coefficient set, including:

[0032] For the vibration signals of each key structural component of the gearbox, calculate the load disturbance component. ;

[0033] Subtracting the load disturbance component from the original vibration signal yields the corrected vibration signal. The correction formula is as follows:

[0034] ;

[0035] in, For the corrected vibration signal, Original vibration signal, For the first The load transfer coefficient of a key structural component This represents the time-domain quantization value of the aerodynamic load characteristics.

[0036] In conjunction with the first aspect, in one possible design, multidimensional feature extraction is performed on the vibration signal correction dataset, including:

[0037] Time-domain features are extracted from the corrected vibration signal to obtain the vibration peak factor, impulse factor, and margin factor.

[0038] Frequency domain features are extracted from the corrected vibration signal to obtain the peak value and frequency amplitude of the vibration power spectral density;

[0039] Time-frequency domain feature extraction is performed on the corrected vibration signal to obtain the decomposed energy entropy vector;

[0040] The above features are recombined according to spatial coordinates to form a three-dimensional vibration feature matrix.

[0041] In conjunction with the first aspect, in one possible design, the pre-defined gearbox fault identification model training method includes:

[0042] Construct a gearbox fault sample set containing different fault types. The fault sample set includes a three-dimensional vibration feature matrix and fault labels corresponding to the fault types.

[0043] The fault sample set is divided into a training set and a validation set according to a predetermined ratio;

[0044] An optimizer is used to train the model, and the number of training rounds and the conditions for stopping training are set. Training is completed when the model reaches the preset fault identification accuracy standard.

[0045] In conjunction with the first aspect, in one possible design, a pre-defined gearbox fault identification model is used to identify faults in the three-dimensional vibration feature matrix, thereby obtaining gearbox fault types, including:

[0046] The three-dimensional vibration feature matrix is ​​standardized to eliminate the dimensional differences between features of different dimensions;

[0047] The standardized three-dimensional vibration feature matrix is ​​input into the preset gearbox fault identification model;

[0048] The model outputs the predicted probability of various types of faults, and selects the category with the highest probability as the preliminary fault identification result;

[0049] The initial identification results are verified for confidence. When the confidence exceeds the preset threshold, it is determined to be the final gearbox fault type. When the confidence does not exceed the preset threshold, a prompt for review is output and the feature data is retained for model optimization.

[0050] Secondly, the present invention also provides a wind turbine gearbox fault diagnosis system based on multimodal data fusion, comprising:

[0051] The critical structure identification module is used to identify the critical structures of the wind turbine gearbox and mark them as critical structural components of the gearbox.

[0052] The signal acquisition module is used to acquire vibration signals of each key structural component of the gearbox, as well as the strain signals of the wind turbine impeller.

[0053] The aerodynamic load feature extraction module is used to extract aerodynamic load features from the impeller strain signal;

[0054] The load influence transmission coefficient acquisition module is used to input aerodynamic load characteristics into a pre-built load transmission evaluation model to obtain a set of load influence transmission coefficients;

[0055] The vibration signal correction module is used to correct the impeller effect on each vibration signal based on the load influence transmission coefficient set, and obtain the vibration signal correction dataset.

[0056] The multidimensional feature extraction and conversion module is used to extract multidimensional features from the vibration signal correction dataset and convert the extracted vibration features into a three-dimensional vibration feature matrix.

[0057] The fault identification module is used to identify faults in the three-dimensional vibration feature matrix using a preset gearbox fault identification model, thereby obtaining the gearbox fault type.

[0058] The technical solution of this invention can achieve the following technical effects:

[0059] By collecting impeller strain signals to extract aerodynamic load characteristics and quantifying the impact of loads on vibration signals based on a load transmission evaluation model, it is possible to correct impeller load interference in vibration signals and eliminate interference from non-faulty aerodynamic loads on vibration characteristics. By combining key structural component identification and multi-dimensional vibration feature extraction, a three-dimensional vibration feature matrix is ​​constructed and diagnosed through a fault identification model, which can improve the accuracy and reliability of wind turbine gearbox fault identification and reduce the fault misjudgment rate caused by load interference under complex wind conditions. Attached Figure Description

[0060] Figure 1 This is a flowchart of the wind turbine gearbox fault diagnosis method based on multimodal data fusion in this invention;

[0061] Figure 2 This is a flowchart illustrating the construction process of the load transmission evaluation model in this invention;

[0062] Figure 3 This is a flowchart illustrating the training process of the gearbox fault identification model in this invention.

[0063] Figure 4 This is a structural diagram of the wind turbine gearbox fault diagnosis system based on multimodal data fusion in this invention. Detailed Implementation

[0064] This application will now be described with reference to the accompanying drawings.

[0065] like Figure 1 As shown, the wind turbine gearbox fault diagnosis method based on multimodal data fusion of the present invention specifically includes the following steps:

[0066] Step S1: Identify the key structures of the wind turbine gearbox and mark them as key structural components of the gearbox;

[0067] Step S2: Collect vibration signals of each key structural component of the gearbox, as well as the impeller strain signal of the wind turbine; the impeller strain signal and vibration signal are collected at the same time.

[0068] Step S3: Extract aerodynamic load characteristics from the impeller strain signal;

[0069] Step S4: Input the aerodynamic load characteristics into the pre-built load transmission evaluation model to obtain the load influence transmission coefficient set;

[0070] Step S5: Based on the load influence transmission coefficient set, perform impeller influence correction on each vibration signal to obtain the vibration signal correction dataset;

[0071] Step S6: Extract multidimensional features from the vibration signal correction dataset and convert the extracted vibration features into a three-dimensional vibration feature matrix;

[0072] Step S7: Use the preset gearbox fault identification model to identify faults in the three-dimensional vibration feature matrix and obtain the gearbox fault type.

[0073] In embodiments of the present invention, by fusing two multimodal data sources—vibration signals from key gearbox structural components and impeller strain signals—accurate quantification and correction of the impact of impeller aerodynamic loads are achieved. Specifically, aerodynamic load features are extracted from the impeller strain signals, and the influence coefficient of load transmission on gearbox vibration is determined using a load transmission evaluation model. Based on this, non-faulty load interference in the vibration signals is corrected, eliminating spurious interference components caused by aerodynamic load transmission in the vibration features. The corrected vibration signal dataset more accurately reflects the fault characteristics of the gearbox. Combining multidimensional feature extraction and three-dimensional feature matrix construction improves the accuracy of the preset fault identification model in identifying gearbox fault features, reduces the risk of misjudgment caused by the dynamic influence of aerodynamic loads on single vibration signal analysis, and thus improves the accuracy of wind turbine gearbox fault diagnosis.

[0074] In some embodiments of the present invention, for step S1, key structural identification of the wind turbine gearbox is performed and it is marked as a key structural component of the gearbox, specifically including:

[0075] Step S11: Calculate the stress concentration factor at each structural location through gearbox dynamics simulation. Filter out structural locations with stress concentration factors greater than a preset threshold; The calculation formula is:

[0076] ;

[0077] in, This represents the maximum stress value generated by the structure under dynamic load, expressed in megapascals (MPA), and is obtained through finite element simulation or experimental measurement. K represents the standard stress value of a structure under ideal uniform stress conditions, expressed in megapascals (MPa), calculated based on theoretical loads and geometric dimensions. t This represents the stress concentration factor, which is dimensionless. The larger the value, the more significant the stress concentration phenomenon at that location, and the more likely fatigue damage or crack propagation will occur. This formula is used to quantify the local stress amplification effect of a structure and identify weak areas that are prone to high stress under dynamic loads.

[0078] Step S12: Based on the statistical data of actual maintenance failures of the gearbox, the top 5 structural parts with the highest failure frequency are marked as the key structural components of the gearbox.

[0079] In embodiments of the present invention, the wind turbine gearbox has a complex structure, and different parts experience varying forces and failure probabilities during operation. By identifying key structural components, targeted data collection and fault diagnosis can be performed, improving diagnostic efficiency and accuracy. Gearbox dynamics simulation refers to the process of establishing a digital model of the gearbox using computer software to simulate its dynamic mechanical behavior under actual operating conditions, including the interaction between components such as gear meshing, bearing support, and shaft rotation, and analyzing key dynamic characteristics such as vibration, impact, stress distribution, and dynamic load transmission. Gearbox actual operation and maintenance fault statistics refer to data obtained by recording, collecting, and analyzing various faults that occur during the actual operation and maintenance of the gearbox. By comprehensively considering theoretical simulation and actual operation and maintenance data, the key structural components most prone to failure and with the greatest impact on overall operation of the gearbox can be accurately identified, avoiding indiscriminate monitoring and analysis of the entire gearbox, thus saving resources and time.

[0080] In some embodiments of the present invention, for step S2, collecting vibration signals of each key structural component of the gearbox and the impeller strain signal of the wind turbine specifically includes:

[0081] Multiple orthogonal acceleration sensors are deployed on the surface of each key structural component of the gearbox to collect vibration signals;

[0082] Multiple strain gauges are uniformly deployed circumferentially at the connection between the impeller hub and the blades to collect impeller strain signals; the strain signal quantization formula is:

[0083] ;

[0084] in, This represents the change in strain gauge length. The initial length of the strain gauge is given; the sampling frequency is consistent with the sampling frequency of the vibration signal; this formula converts the resistance change of the strain gauge into a quantifiable strain value, directly characterizing the magnitude and distribution of the aerodynamic load on the impeller.

[0085] In embodiments of this invention, the selection of the accelerometer sensor must consider the measurement range, sensitivity, frequency response, and the temperature and vibration resistance of the installation environment. Piezoelectric or MEMS types can be selected to accurately capture vibration signals. When the accelerometer sensor moves with the measured structure, the internal mass block generates a force proportional to the acceleration due to inertia. This force is transmitted through an elastic element and induces deformation. The deformation is then converted into a measurable electrical signal by a sensitive element such as a piezoelectric element or capacitor, thereby enabling the detection of the magnitude and direction of the acceleration. The selection of the strain gauge must consider the strain range, sensitivity coefficient, linearity, and compatibility with the impeller material. Foil strain gauges can be selected to ensure stable output of strain signals under dynamic loads on the hub. When the strain gauge is attached to the surface of the measured component, the deformation of the component under stress will cause the strain gauge to expand and contract synchronously, resulting in changes in the length and cross-sectional area of ​​the strain gauge. The material resistance will change with the change in geometry. By measuring the relative change in resistance value, the strain magnitude of the component can be indirectly calculated according to the strain signal quantification formula. Vibration signals of key structural components directly reflect the vibration of the gearbox, while impeller strain signals can reflect the impact of impeller aerodynamic loads on the system. Both signals must be acquired at the same time to accurately analyze their relationship.

[0086] In some embodiments of the present invention, for step S3, extracting aerodynamic load features from the impeller strain signal specifically includes:

[0087] Step S31: Extract time-domain features from the impeller strain signal to obtain the strain peak value. root mean square strain and strain kurtosis ;

[0088] Step S32: Extract frequency domain features from the impeller strain signal to obtain the amplitude of the dominant strain frequency. and strain spectrum entropy ;

[0089] Step S33: Fuse the time-domain features and frequency-domain features to form an aerodynamic load feature vector. As a characteristic of aerodynamic loads.

[0090] In an embodiment of the present invention, the strain peak value is determined by... Calculations reflect the instantaneous maximum aerodynamic load impact; root mean square strain is obtained through... Calculations characterize the energy level of load fluctuations; strain kurtosis is achieved through... The strain kurtosis value reflects the intensity of the impact component in the signal. A higher strain kurtosis value indicates a more significant transient impact component in the signal, corresponding to abrupt load impact events in aerodynamic loads, and helps identify load characteristics under extreme wind conditions; strain dominant frequency amplitude. Identify the amplitude corresponding to the highest energy frequency in the spectrum, reflecting the main vibration frequency components of the aerodynamic load; strain spectrum entropy through... calculate( For the first The power ratio of each frequency component characterizes the complexity of the spectral distribution. By extracting aerodynamic load information contained in the impeller strain signal, the impact of aerodynamic load on the gearbox can be quantified, thus considering this factor in fault diagnosis and improving diagnostic accuracy. Furthermore, by extracting and fusing features from multiple perspectives in the time and frequency domains, the aerodynamic load characteristics reflected by the impeller strain signal can be comprehensively described, providing a clear and quantitative feature representation for subsequent analysis of the impact of aerodynamic load on gearbox vibration signals, and helping to more accurately separate fault-related features.

[0091] like Figure 2 As shown, in some embodiments of the present invention, step S4 involves inputting aerodynamic load characteristics into a pre-built load transmission evaluation model to obtain a set of load influence transmission coefficients. The pre-built load transmission evaluation model is constructed through the following steps:

[0092] Collect sample signals of impeller strain and corresponding vibration of key structural components of the gearbox under different wind speeds and pitch angles of the wind turbine.

[0093] A mapping function between impeller strain and vibration response is established based on the dynamic equations of the transmission system. ,in For vibration response, For impeller strain, The transfer coefficient is to be optimized;

[0094] Model training was performed using impeller strain sample signals, vibration sample signals, and mapping relationship functions. The model parameters were iteratively optimized using a particle swarm optimization algorithm until the model prediction error was less than 5%, thus completing the construction of the load transmission evaluation model and obtaining a set of load influence transmission coefficients. .

[0095] In embodiments of the present invention, the load influence transmission coefficient set includes multiple load influence transmission coefficients, which respectively represent the vibration influence of the impeller aerodynamic load on each key structural component of the gearbox through the transmission system. The load transmission evaluation model is used to quantify the relationship between impeller strain and vibration response of key structural components of the gearbox, thereby obtaining the load influence transmission coefficients to correct the vibration signal and eliminate aerodynamic load interference. The model is trained with a large amount of sample data under different working conditions, and the model parameters are optimized using a particle swarm optimization algorithm, so that the established load transmission evaluation model can accurately reflect the true relationship between impeller strain and gearbox vibration response. The resulting load influence transmission coefficient set has high accuracy and reliability.

[0096] In some embodiments of the present invention, regarding step S5, impeller influence correction is performed on each vibration signal based on the load influence transmission coefficient set to obtain a vibration signal correction dataset, specifically including:

[0097] Step S51: Calculate the load disturbance component for the vibration signal of each key structural component of the gearbox. ;

[0098] Step S52: Subtract the load interference component from the original vibration signal to obtain the corrected vibration signal. The correction formula is:

[0099] ;

[0100] in, The corrected vibration signal is expressed in m / s. 2 or g; The original vibration signal is expressed in m / s. 2 or g; For the first The load transfer coefficient of a key structural component; The time-domain quantization value of the aerodynamic load characteristics is given by the eigenvector F. a The corrected formula is obtained by interpolation or reconstruction and its dimensions are consistent with the vibration signal. It removes non-fault interference components caused by impeller aerodynamic load from the original vibration signal, so that the remaining signal reflects the gearbox's own operating status and fault characteristics more purely.

[0101] In an embodiment of the present invention, the vibration signal correction dataset includes vibration signals of each key structural component of the gearbox after correction by its corresponding load influence transmission coefficient. By correcting the vibration signal, non-fault interference caused by impeller aerodynamic load can be effectively removed from the vibration signal, so that the corrected vibration signal can more realistically reflect the actual vibration of the key structural component of the gearbox and reduce the possibility of subsequent fault diagnosis misjudgment due to load interference.

[0102] In some embodiments of the present invention, step S6 involves multidimensional feature extraction of the vibration signal correction dataset, specifically including:

[0103] Step S61: Extract time-domain features from the corrected vibration signal to obtain the vibration peak factor. Pulse factor and margin factor ;

[0104] Step S62: Extract frequency domain features from the corrected vibration signal to obtain the peak value of the vibration power spectral density. and 3rd harmonic amplitude ;

[0105] Step S63: Extract time-frequency domain features from the corrected vibration signal to obtain the decomposed energy entropy vector. ;

[0106] Step S64: Reorganize the above features according to spatial coordinates to form a three-dimensional vibration feature matrix. ,in The length of the time dimension corresponds to the number of time steps in the feature sequence. This represents the length of the frequency dimension, corresponding to the number of dimensions of the frequency domain or time-frequency domain features. The number of spatial dimensions corresponds to the number of different sensors or feature types. This three-dimensional matrix structure can completely preserve the information of vibration signals in the three dimensions of time, frequency, and space, providing structured input for deep learning models and enhancing the ability to identify complex fault modes.

[0107] In an embodiment of the present invention, the vibration peak factor The ratio of peak value to root mean square value, impulse factor. The ratio of peak value to average value reflects localized impact failure in gears or bearings; margin factor. The ratio of peak value to kurtosis, peak value of vibrational power spectral density. The peak value with the highest energy in the spectrum, 3rd harmonic amplitude. The amplitude corresponding to the third harmonic of the gear meshing frequency can be used to locate faults related to the gear meshing frequency; the energy entropy vector in the time-frequency domain. It can quantify the degree of energy distribution disorder of signals in different frequency bands, effectively identify signal non-stationarity caused by early weak faults; through multi-dimensional feature extraction, it can cover the transient impact characteristics of vibration signals in the time domain, the frequency distribution of harmonic / fault characteristics in the frequency domain, and the dynamic changes of energy distribution in the time and frequency domain, avoiding the omission of key fault information by single-dimensional features, and improving the accuracy and reliability of fault diagnosis.

[0108] like Figure 3As shown, in some embodiments of the present invention, step S7 involves using a preset gearbox fault identification model to identify faults in the three-dimensional vibration feature matrix and obtain the gearbox fault type. The preset gearbox fault identification model is obtained through the following steps:

[0109] Construct a gearbox fault sample set containing different fault types. The fault sample set includes a three-dimensional vibration feature matrix and fault labels corresponding to the fault types.

[0110] The fault sample set is divided into a training set and a validation set according to a predetermined ratio;

[0111] An optimizer is used to train the model, and the number of training rounds and the conditions for stopping training are set. Training is completed when the model reaches the preset fault identification accuracy standard.

[0112] In embodiments of the present invention, the training set and validation set are typically divided empirically, such as 7:3 or 8:2, to ensure that the training set is large enough to support learning and the validation set is large enough for objective evaluation. The optimizer is used to adjust the internal parameters of the model so that the error between the model's prediction results and the actual labels becomes smaller and smaller. Optimizers such as Adam and SGD can be selected. By constructing a rich set of fault samples and reasonably dividing the training set and validation set, the model can fully learn the characteristics of various fault types. The use of the optimizer and reasonable training settings help the model converge to the optimal solution, thereby improving the model's fault identification accuracy and generalization ability, enabling it to accurately identify the fault type of the gearbox in practical applications.

[0113] In some embodiments of the present invention, regarding step S7, a preset gearbox fault identification model is used to identify faults in the three-dimensional vibration feature matrix to obtain the gearbox fault type, specifically including:

[0114] Step S71: Standardize the three-dimensional vibration feature matrix to eliminate the dimensional differences of features in different dimensions;

[0115] Step S72: Input the standardized three-dimensional vibration feature matrix into the preset gearbox fault identification model;

[0116] Step S73: The model outputs the predicted probability of various types of faults, and selects the category with the highest probability as the preliminary fault identification result;

[0117] Step S74: Perform confidence verification on the preliminary identification results. When the confidence level is greater than or equal to the preset threshold, it is determined as the final gearbox fault type. When the confidence level is less than the preset threshold, output a prompt for review and retain the feature data for model optimization.

[0118] In embodiments of the present invention, dimensional differences refer to the different units or numerical magnitudes of different features. If left untreated, the model may be influenced by features with large numerical values, while ignoring important features with smaller magnitudes. By using mathematical methods to transform all features to a uniform numerical range, such as mapping feature values ​​to between 0 and 1, or converting them to a normal distribution with a mean of 0 and a standard deviation of 1, the influence weight of each feature on the model can be ensured fairly. By calculating a probability value for each possible fault type, the probability that the gearbox belongs to a certain type of fault can be represented. At the same time, confidence verification can further improve the reliability of fault identification results and avoid misjudgment.

[0119] like Figure 4 As shown, the present invention also provides a wind turbine gearbox fault diagnosis system based on multimodal data fusion, which specifically includes the following modules;

[0120] The critical structure identification module is used to identify the critical structures of the wind turbine gearbox and mark them as critical structural components of the gearbox.

[0121] The signal acquisition module is used to acquire vibration signals of each key structural component of the gearbox, as well as the strain signals of the wind turbine impeller.

[0122] The aerodynamic load feature extraction module is used to extract aerodynamic load features from the impeller strain signal;

[0123] The load influence transmission coefficient acquisition module is used to input aerodynamic load characteristics into a pre-built load transmission evaluation model to obtain a set of load influence transmission coefficients;

[0124] The vibration signal correction module is used to correct the impeller effect on each vibration signal based on the load influence transmission coefficient set, and obtain the vibration signal correction dataset.

[0125] The multidimensional feature extraction and conversion module is used to extract multidimensional features from the vibration signal correction dataset and convert the extracted vibration features into a three-dimensional vibration feature matrix.

[0126] The fault identification module is used to identify faults in the three-dimensional vibration feature matrix using a preset gearbox fault identification model, thereby obtaining the gearbox fault type.

[0127] In embodiments of the present invention, vibration signals of key structural components of the gearbox and strain signals of the impeller are collected. Aerodynamic load features are extracted and transmission coefficients are obtained through a load transmission evaluation model. Vibration signals are corrected to eliminate interference from non-faulty loads. Multidimensional features are then extracted and converted into a three-dimensional matrix. Faults are diagnosed through a fault identification model. Its advantages include considering the dynamic impact of impeller aerodynamic loads on gearbox vibration, quantifying and correcting the interference of load transmission on vibration signals, reducing misjudgments; integrating multimodal data to improve feature completeness; and enhancing fault feature identification through multidimensional feature extraction and three-dimensional matrix conversion, thereby improving the accuracy and reliability of gearbox fault diagnosis.

[0128] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for fault diagnosis of wind turbine gearboxes based on multimodal data fusion, characterized in that, include: Key structural components of the wind turbine gearbox are identified and marked as key structural parts of the gearbox. Vibration signals of each key structural component of the gearbox and the impeller strain signal of the wind turbine are collected; the impeller strain signal and the vibration signal are collected at the same time. Extract aerodynamic load characteristics from the impeller strain signal; The aerodynamic load characteristics are input into a pre-built load transmission evaluation model to obtain a set of load influence transmission coefficients; Based on the load influence transmission coefficient set, impeller influence correction is performed on each vibration signal to obtain a vibration signal correction dataset. The vibration signal correction dataset includes vibration signals of each key structural component of the gearbox after correction by its corresponding load influence transmission coefficient. By correcting the vibration signals, non-fault interference caused by impeller aerodynamic load is removed from the vibration signals. Based on the load influence transmission coefficient set, impeller influence correction is performed on each of the vibration signals, including: calculating the load interference component for the vibration signal of each key structural component of the gearbox. Subtract the load interference component from the original vibration signal to obtain the corrected vibration signal. The correction formula is as follows: ； in, For the corrected vibration signal, Original vibration signal, For the first The load transfer coefficient of a key structural component This represents the time-domain quantization value of the aerodynamic load characteristics; Multidimensional feature extraction is performed on the vibration signal correction dataset, and the extracted vibration features are converted into a three-dimensional vibration feature matrix; The gearbox fault type is obtained by using a preset gearbox fault identification model to identify faults in the three-dimensional vibration feature matrix.

2. The wind turbine gearbox fault diagnosis method based on multimodal data fusion according to claim 1, characterized in that, The process of identifying key structures in the wind turbine gearbox and marking them as key structural components includes: Stress concentration factors at various structural locations were calculated using gearbox dynamics simulation. The structural locations with stress concentration factors greater than a preset threshold are selected; The calculation formula is: ； in, This represents the maximum stress value of the structure. This is the standard stress value; Based on the statistical data of actual gearbox operation and maintenance failures, the structural parts whose failure frequency ranks within a preset position are marked as the intersection of the two as key structural components of the gearbox.

3. The wind turbine gearbox fault diagnosis method based on multimodal data fusion according to claim 1, characterized in that, The acquisition of vibration signals from each of the key structural components of the gearbox, and the strain signals from the wind turbine rotor, includes: Multiple orthogonal acceleration sensors are deployed on the surface of each key structural component of the gearbox to collect vibration signals; Multiple strain gauges are uniformly deployed circumferentially at the connection between the impeller hub and the blades to collect impeller strain signals; the quantization formula for the strain signals is: ； in, This represents the change in strain gauge length. The initial length of the strain gauge; the sampling frequency is consistent with the sampling frequency of the vibration signal and the timestamp synchronization error. .

4. The wind turbine gearbox fault diagnosis method based on multimodal data fusion according to claim 1, characterized in that, The extraction of aerodynamic load features from the impeller strain signal includes: Time-domain features of the impeller strain signal are extracted to obtain strain peak value, strain root mean square and strain kurtosis; Frequency domain feature extraction was performed on the impeller strain signal to obtain the strain dominant frequency amplitude and strain spectrum entropy; The time-domain features and frequency-domain features are fused to form an aerodynamic load feature vector, which is used as the aerodynamic load feature.

5. The wind turbine gearbox fault diagnosis method based on multimodal data fusion according to claim 1, characterized in that, The pre-built load transmission evaluation model construction method includes: Collect sample signals of impeller strain and corresponding vibration of key structural components of the gearbox under different wind speeds and pitch angles of the wind turbine. A mapping function between impeller strain and vibration response is established based on the dynamic equations of the transmission system. ,in For vibration response, For impeller strain, The transfer coefficient is to be optimized; The impeller strain sample signal, vibration response sample signal, and mapping function are used for model training. The model parameters are iteratively optimized using a particle swarm optimization algorithm to complete the construction of the load transmission evaluation model and obtain the load influence transmission coefficient set. .

6. The wind turbine gearbox fault diagnosis method based on multimodal data fusion according to claim 1, characterized in that, The multidimensional feature extraction of the vibration signal correction dataset includes: Time-domain features are extracted from the corrected vibration signal to obtain the vibration peak factor, impulse factor, and margin factor. Frequency domain features are extracted from the corrected vibration signal to obtain the peak value of the vibration power spectral density and the set frequency amplitude value; Time-frequency domain feature extraction is performed on the corrected vibration signal to obtain the decomposed energy entropy vector; The above features are recombined according to spatial coordinates to form a three-dimensional vibration feature matrix.

7. The wind turbine gearbox fault diagnosis method based on multimodal data fusion according to claim 1, characterized in that, The preset gearbox fault identification model training method includes: Construct a gearbox fault sample set containing different fault types. The fault sample set includes a three-dimensional vibration feature matrix and fault labels corresponding to the fault types. The fault sample set is divided into a training set and a validation set according to a predetermined ratio; An optimizer is used to train the model, and the number of training rounds and the conditions for stopping training are set. Training is completed when the model reaches the preset fault identification accuracy standard.

8. The wind turbine gearbox fault diagnosis method based on multimodal data fusion according to claim 1, characterized in that, The method of using a preset gearbox fault identification model to identify faults in the three-dimensional vibration feature matrix to obtain gearbox fault types includes: The three-dimensional vibration feature matrix is ​​standardized to eliminate the dimensional differences between features of different dimensions; The standardized three-dimensional vibration feature matrix is ​​input into the preset gearbox fault identification model; The model outputs the predicted probability of various types of faults, and selects the category with the highest probability as the preliminary fault identification result; The initial identification results are verified for confidence. When the confidence exceeds the preset threshold, it is determined to be the final gearbox fault type. When the confidence does not exceed the preset threshold, a prompt for review is output and the feature data is retained for model optimization.

9. A wind turbine gearbox fault diagnosis system based on multimodal data fusion, wherein the system is applied to the wind turbine gearbox fault diagnosis method based on multimodal data fusion as described in claim 1, characterized in that, The system includes: The critical structure identification module is used to identify the critical structures of the wind turbine gearbox and mark them as critical structural components of the gearbox. The signal acquisition module is used to acquire vibration signals of each key structural component of the gearbox, as well as the rotor strain signal of the wind turbine. The aerodynamic load feature extraction module is used to extract aerodynamic load features from the impeller strain signal; The load influence transmission coefficient acquisition module is used to input the aerodynamic load characteristics into a pre-built load transmission evaluation model to obtain a set of load influence transmission coefficients; The vibration signal correction module is used to perform impeller influence correction on each vibration signal based on the load influence transmission coefficient set to obtain a vibration signal correction dataset. The multidimensional feature extraction and conversion module is used to extract multidimensional features from the vibration signal correction dataset and convert the extracted vibration features into a three-dimensional vibration feature matrix. The fault identification module is used to identify faults in the three-dimensional vibration feature matrix using a preset gearbox fault identification model to obtain the gearbox fault type.

Images