Wind turbine bearing fault diagnosis model online construction method and online updating device

Abstract

The application discloses a kind of wind generator bearing fault diagnosis model online construction method and online updating equipment, it is related to the bearing fault diagnosis field of wind generator.The method is according to the bearing state of transmission end and non-transmission end, runs wind generator under 9 different schemes, carries out bearing vibration data acquisition test, and processing is carried out to the collected data: first classification, then model is established respectively for transmission end and non-transmission end, construction method is similar, including determining model structure, model training and model performance test.Model online updating equipment mainly includes: vibration data acquisition module, core processing module, communication module and other auxiliary supporting equipment, using the device for line renewal, including model deployment, running data storage, model online updating and model performance regular test.The application is realized by using software and hardware combination, can accurately diagnose bearing transmission end, non-transmission end fault, and improve model field application performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of bearing fault diagnosis of wind turbines, in particular to a wind turbine bearing fault diagnosis model online construction method and online updating device. BACKGROUND

[0002] As one of the vulnerable parts in the transmission system of wind turbine generators, once the rolling bearing fails, a large amount of time and cost is needed for maintenance. Therefore, in view of the difficulty of fault diagnosis, experts and scholars use vibration, electrical signal analysis and pattern recognition algorithm to realize fault diagnosis, among which the vibration signal processing method based on deep learning is particularly widely used. Using multi-layer neural network architecture such as convolutional neural network and recurrent neural network to analyze the original vibration data and explore the useful information contained in the data, after training the appropriate deep neural network model, the function of fault analysis and diagnosis can be realized.

[0003] In practical application, by pre-preparing different types and degrees of bearing faults, collecting bearing vibration acceleration data under different operating conditions, using different network models to train and learn the fault data, and finally determining the model with the best performance to complete the algorithm modularization packaging and deployment, the deployed model can be used to diagnose the bearing fault. Due to the difference between the test data and the actual running data, which will affect the accuracy of the model diagnosis, therefore, the application of actual working condition data for model training will be more accurate, however, the actual data, especially the fault data, is difficult to obtain, even if it can be obtained, there is a lack of effective data, the application of test data to determine the parameters and structure of the model, on this basis, the introduction of field operation data to update and optimize the model, combined with test and field operation data to diagnose bearing fault, has an important role in improving the accuracy of diagnosis.

[0004] In the prior art, a patent with patent number CN112729825A collects acceleration data of a fault bearing in a laboratory, extracts vibration acceleration signals as model input, and applies a deep learning algorithm to establish a bearing fault recognition model. The test data accuracy is high, but the existing shortcomings are: the data collection is usually carried out in the laboratory, and the actual application effect generalization ability of the network model which performs particularly outstanding in the training cannot be completely simulated in the field running condition, so the identification accuracy of the actual running data often appears in the actual application. The phenomenon of false positives and false negatives. Another patent with patent number CN111458144A transforms and processes the bearing vibration signals of the generator during operation obtained from the data acquisition and monitoring control system of the wind farm at a certain time interval, calculates the characteristic parameters of the bearing to determine the bearing fault type, and adopts a random forest convolutional neural network to construct a fault diagnosis model. The defects of this patent are that the actual running data is used for training, but the fault type label is determined by feature calculation, which may be inaccurate or even wrong, affecting the performance of the model. In the process of operation and maintenance of the wind turbine, the model needs to be continuously optimized to improve the identification and diagnosis accuracy during the running period, and the real-time training and updating of the algorithm model cannot be carried out, and the optimized model needs to be deployed and upgraded regularly, which requires long-term maintenance cost and maintenance time. In the process of operation and maintenance of the wind turbine, the owner is often more concerned about whether the bearing is faulty and whether the bearing needs to be replaced, and the judgment of the bearing fault type can be further analyzed and determined combined with the operation data. Therefore, combined with the test data and the actual data, the fault diagnosis of the transmission end and the non-transmission end bearing of the wind turbine is carried out respectively, and the online construction of the algorithm model is imperative. SUMMARY

[0005] The present application provides a wind turbine bearing fault diagnosis model online construction method and online updating equipment to solve the problem of weak generalization ability and unsatisfactory application effect of the verified algorithm model in actual application.

[0006] The present application is realized by the following technical solutions: first, the present application designs a wind turbine bearing fault diagnosis model online construction method, and on this basis, designs a set of wind turbine bearing fault diagnosis model online updating equipment, which updates the model online in the equipment, as follows.

[0007] A wind turbine bearing fault diagnosis model online construction method includes the following steps:

[0008] I. Test data acquisition:

[0009] (1) Bearing fault pre-preparation:

[0010] For the common faults of bearing pre-preparation, the fault types include: bearing inner and outer ring electro-erosion, bearing inner and outer ring peeling, bearing outer ring peeling, bearing inner ring peeling, and the fault bearings are installed at the transmission end and non-transmission end of the wind turbine, a total of 9 operation schemes:

[0011] Scheme one: the transmission end bearing state is inner and outer ring electro-erosion, and the non-transmission end bearing state is normal;

[0012] Scheme two: the transmission end bearing state is normal, and the non-transmission end bearing state is inner and outer ring electro-erosion;

[0013] Scheme three: the transmission end bearing state is normal, and the non-transmission end bearing state is normal;

[0014] Scheme four: the transmission end bearing state is inner and outer ring peeling, and the non-transmission end bearing state is normal;

[0015] Scheme five: the transmission end bearing state is normal, and the non-transmission end bearing state is inner and outer ring peeling;

[0016] Scheme six: the transmission end bearing state is outer ring peeling, and the non-transmission end bearing state is normal;

[0017] Scheme seven: the transmission end bearing state is normal, and the non-transmission end bearing state is outer ring peeling;

[0018] Scheme eight: the transmission end bearing state is inner ring peeling, and the non-transmission end bearing state is normal;

[0019] Scheme nine: the transmission end bearing state is normal, and the non-transmission end bearing state is inner ring peeling;

[0020] (2) Test data collection and processing:

[0021] The wind turbine is installed on the test bench with a tilt angle of θ °, the rotor is short-circuited, and it is operated under no-load condition. Each scheme is operated at a constant speed n 1、 n 2、 n 3、 ……. n m The single-shaft vibration acceleration data of the transmission end and non-transmission end are collected for 10 minutes, a total of 9 m groups of original data;

[0022] The 9 m groups of test data are divided into two categories:

[0023] ① The non-transmission end is normal, and the transmission end is faulty original data, i.e. scheme one, scheme three, scheme four, scheme six, and scheme eight;

[0024] ②transmission end normal, non-transmission end fault original data, namely scheme two, scheme three, scheme five, scheme seven, scheme nine;

[0025] Considering the imbalance between positive and negative samples, the fault original data is under-sampled to ensure that the number of normal and fault samples is similar. The under-sampled fault original data is divided into time series data with a length of L, a total of Y parts, and the normal original data is also divided into time series data with a length of L, a total of X parts. Each data sequence is labeled with a fault label to form an original signal data set consisting of X+Y groups of time series data with fault labels. Discrete Fourier transform is performed on the original data sequence with a length of L to obtain frequency domain data with a length of L. The time domain and frequency domain data are merged, so each time series data changes from 1xL one-dimensional data to 2xL two-dimensional data. To implement the fast algorithm FFT of discrete Fourier transform, the signal sampling frequency Fs of the single-axis acceleration sensor should be greater than 2.56 times the maximum frequency of the signal;

[0026] II. Model establishment: models are established for the transmission end and the non-transmission end respectively, and the construction method is similar. The fault diagnosis model of the non-transmission end bearing is similar to that of the transmission end. The second type of data is used to establish the model. The model establishment method of the transmission end includes the following steps:

[0027] (1) Determine the model structure:

[0028] The first layer is a convolution layer, which inputs a 2xL grayscale image, filters for 32, strides=(1, 4), edge padding, and batch normalization, Relu activation, and maximum pooling. The second layer is a convolution layer, which inputs a 2xL grayscale image, filters for 32, strides=(1, 4), edge padding, and batch normalization, Relu activation, and maximum pooling. The third layer is a sequence conversion layer, which converts the convolution data into LSTM layer input data format. The fourth layer is a long short-term memory network layer LSTM (Long Short-Term Memory), which outputs a spatial dimension of 256 and returns all sequences. The fifth layer is a dropout layer with a dropout rate of 0.3. The sixth layer is a long short-term memory network layer LSTM, which outputs a spatial dimension of 32 and returns the last output in the sequence. The seventh layer is a dropout layer with a dropout rate of 0.3. The eighth layer is a fully connected layer using ReLu activation function and L2 regularization. The ninth layer is an output layer using a sigmod classifier for processing to realize fault data diagnosis.

[0029] (2) Model training:

[0030] Model training is to input data into the model after the framework structure of the model is determined, and the final algorithm model is obtained after training to judge the state of the bearing. The determination of the model structure is a mutual process with training. The structure listed in step (1) is the optimal structure compared after training; the first type of original data obtained by the test is divided into a training set, a validation set and a test set according to a certain proportion, a convolutional network is trained, an optimizer and an objective function are set, the loss rate converges in the training process, and a fault diagnosis model is generated after training. In order to improve the training efficiency and reduce the training time, the GPU is called to accelerate the training.

[0031] (3) Model performance test:

[0032] The test set data is imported into the generated fault diagnosis model to obtain the diagnosis result of the test set data, and the overall accuracy of the result is calculated.

[0033] On the basis of this method, the model trained by the test data is directly used for fault diagnosis of bearings of wind generators in field operation. Due to the poor generalization performance of the model, the output result accuracy is low, there are problems of false positives and false negatives, and even the model cannot be normally used. Therefore, the model needs to be processed and updated regularly to improve the accuracy of the diagnosis result. The update of the model is a long-term process, and the performance of the model gradually increases and improves with the continuous accumulation of field operation data. An online updating device for a wind generator bearing fault diagnosis model mainly includes a vibration signal acquisition module, a core processing module, a communication module and auxiliary supporting equipment, and the auxiliary supporting equipment includes power supply equipment and connection lines. The vibration signal acquisition module includes a vibration sensor, the vibration sensor is installed on a sensor seat at a transmission end and a non-transmission end of the wind generator, a data acquisition module of the vibration sensor realizes vibration data acquisition, and the acquired data is sent to the core processing module through Ethernet or other wireless communication methods. The core processing module includes a CPU and a GPU, and is an edge computing module or a ground server platform. Through the deployed software program, the vibration data is preprocessed and analyzed, and the fault diagnosis model is updated. The communication module realizes data interaction between the vibration signal acquisition module and the core processing module, and data interaction between the core processing module and an external system.

[0034] Preferably, the minimum sampling frequency of the data acquisition module of the vibration sensor is not less than 2.56 times the maximum frequency of the signal.

[0035] The model online updating method of the above-mentioned online updating device for a wind generator bearing fault diagnosis model, and the online construction method of the fault diagnosis model are adopted, and the specific steps include the following steps: the fault diagnosis model updating methods of the transmission end and the non-transmission end are the same, and the diagnosis model updating method of the transmission end includes the following steps:

[0036] ①Model deployment: deploy the transmission end bearing fault diagnosis model in the core processing module for bearing fault diagnosis, real-time receive the field transmission end vibration data of the vibration signal acquisition module, use the same method as step (2) of the fault diagnosis model online construction method to process the data into 2xL data sequence input to the model, which can realize the real-time fault diagnosis function of the transmission end bearing, but further processing and updating are needed due to low accuracy, false positives and other reasons;

[0037] ②Running data storage: add vibration data files of bearings in normal and fault states in the core processing module; normal data files store motor field normal operation data; fault data is motor field fault operation data, which can be further divided according to different fault types as the motor field operation time extends; to prevent the file system from occupying too much storage, fixed data volume and fixed cycle storage are used.

[0038] ③Model online update: due to the hysteresis of fault operation data, there is a problem of less or even no field operation fault data, so the model online update is carried out according to different methods according to the presence or absence of field operation fault data:

[0039] a. When there is only normal operation data or a small amount of fault operation data in addition to normal operation data, only the motor field normal operation data is extracted for features, the motor field normal operation data within a period of time is processed into 2xL data sequence input to the model for feature extraction, the data processing method uses the same method as step (2) of the fault diagnosis model online construction method, and the output of the full connection layer of the model is statistically analyzed when extracting to determine the full connection layer output range R of the motor field operation data under the normal state of the bearing, R is used as an index to evaluate the normal and fault of the field operation bearing, and R is considered as abnormal bearing state; Therefore, the model does not need to be trained again, only the last classification layer of the model needs to be removed, and the bearing state is diagnosed according to the output of the full connection layer, at this time the first model online update is completed, at this time the model can realize the diagnosis of bearing state normal and fault;

[0040] b. With the extension of running time, the motor running sample data under the bearing fault state gradually increases, and the running data amount of different bearing fault types can meet the requirements of model training. The field running data stored in the file system is also processed into 2xL time series and labeled, and is divided into training set, validation set and test set according to a certain proportion. The network model in the online construction method of the fault diagnosis model is retrained, and the training is stopped when the accuracy meets the set requirements or the training time reaches the set conditions. The original model is replaced by this model. Since normal and different fault types and different fault positions are used for training at this time, the model can not only determine the bearing state but also diagnose the fault type and determine the fault position. With the continuous accumulation of running data, the function of the model is continuously improved;

[0041] ④ Model performance test: periodically call the file storage data and process them into 2xL time series as test data to test the accuracy and false alarm rate of the bearing fault diagnosis model, and determine whether the model needs to be updated according to the accuracy and false alarm rate; when the motor fault diagnosis accuracy and false alarm rate exceed the set range, start the model updating function.

[0042] The above entire process of online updating can be realized by software without human intervention.

[0043] Compared with the prior art, the present application has the following advantages: (1) The test data and field running data are considered together, the model structure parameters are determined by the test data, and the online optimization and updating of the model are completed by using the field running data, solving the problem that the accuracy of the motor bearing fault diagnosis model is high in test data but low in actual deployment or even cannot be used normally; (2) Accurate diagnosis of bearing driving end and non-driving end fault can reduce bearing failure through maintenance methods such as generator maintenance, waste oil treatment, carbon deposition treatment and filter cotton replacement; (3) Under the condition of bearing fault, maintenance personnel can carry out state repair and preventive repair according to the bearing fault state, reduce planned repair and excessive repair, save maintenance cost and time, and improve maintenance efficiency. (4) The function of the model is gradually improved from diagnosing bearing normal and fault states at the beginning, and with the continuous accumulation of field running data, the fault type can be determined and the fault can be located. The entire online updating process is realized by software, which is more economical and intelligent. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 The original signal waveform diagram of part of the test data collected for the driving end and the non-driving end. At this time, the rotating speed is n 1.

[0045] Figure 2A model structure diagram constructed by the online construction method of the wind turbine bearing fault diagnosis model of the application.

[0046] Figure 3 An accuracy and loss rate evolution curve diagram in the training process of the application.

[0047] Figure 4 A test set diagnosis result diagram after model performance testing. DETAILED DESCRIPTION

[0048] The application will be further described below in combination with specific embodiments.

[0049] An online construction method of a wind turbine bearing fault diagnosis model, comprising the following steps:

[0050] I. Test data acquisition:

[0051] (1) Bearing fault preparation:

[0052] For common bearing pre-faults, the fault types include: bearing inner and outer ring electro-erosion, bearing inner and outer ring peeling, bearing outer ring peeling, and bearing inner ring peeling. The fault bearings are respectively installed at the driving end and non-driving end of the wind turbine, and there are a total of 9 operation schemes:

[0053] Scheme one: the driving end bearing state is inner and outer ring electro-erosion, and the non-driving end bearing state is normal;

[0054] Scheme two: the driving end bearing state is normal, and the non-driving end bearing state is inner and outer ring electro-erosion;

[0055] Scheme three: the driving end bearing state is normal, and the non-driving end bearing state is normal;

[0056] Scheme four: the driving end bearing state is inner and outer ring peeling, and the non-driving end bearing state is normal;

[0057] Scheme five: the driving end bearing state is normal, and the non-driving end bearing state is inner and outer ring peeling;

[0058] Scheme six: the driving end bearing state is outer ring peeling, and the non-driving end bearing state is normal;

[0059] Scheme seven: the driving end bearing state is normal, and the non-driving end bearing state is outer ring peeling;

[0060] Scheme eight: the driving end bearing state is inner ring peeling, and the non-driving end bearing state is normal;

[0061] Scheme nine: the driving end bearing state is normal, and the non-driving end bearing state is inner ring peeling;

[0062] (2) Test data collection and processing:

[0063] Tilting of wind turbines θ ° mounted on test bench, rotor short-circuited, no-load operation, constant rotational speed for each variant n 1、 n 2、 n 3、 ……. n m Collecting single-shaft vibration acceleration data of the driving end and the non-driving end for 10 minutes, a total of 9 m groups of original data; part of the data is shown in Figure 1 .

[0064] Divide the 9 m groups of test data into two categories:

[0065] ① non-driving end normal, driving end fault original data, namely scheme one, scheme three, scheme four, scheme six, scheme eight;

[0066] ② driving end normal, non-driving end fault original data, namely scheme two, scheme three, scheme five, scheme seven, scheme nine;

[0067] Considering the unbalanced situation of positive and negative samples, the fault original data is under-sampled to ensure that the number of normal and fault samples is similar; the under-sampled fault original data is divided into time series data with a length of L, a total of Y parts, and the normal original data is also divided into time series data with a length of L, a total of X parts; each data sequence is added with a fault label to form an original signal data set composed of X+Y groups of time series data with fault labels; the original data sequence with a length of L is subjected to discrete Fourier transform to obtain frequency domain data with a length of L, and the time domain and frequency domain data are combined, so that each time series data changes from 1

[0068] II. Model establishment: models are established for the driving end and the non-driving end respectively, and the construction method is similar. The non-driving end bearing fault diagnosis model is similar to the driving end fault diagnosis model, which is established using the second type of data. The model establishment method of the driving end includes the following steps:

[0069] (1) Determine the model structure:

[0070] The first layer is a convolutional layer, inputting a 2xL grayscale image, filters being 32, strides being (1, 4), edges being padded, and batch normalization, Relu activation, and maximum pooling being performed; the second layer is a convolutional layer, inputting a 2xL grayscale image, filters being 32, strides being (1, 4), edges being padded, and batch normalization, Relu activation, and maximum pooling being performed; the third layer is a sequence conversion layer, converting the data after convolution into an LSTM layer input data format; the fourth layer is a long short-term memory network layer LSTM (Long Short-Term Memory), with an output space dimension of 256 and returning all sequences; the fifth layer is a dropout layer with a dropout rate of 0.3; the sixth layer is a long short-term memory network layer LSTM, with an output space dimension of 32 and returning the last output in the sequence; the seventh layer is a dropout layer with a dropout rate of 0.3; the eighth layer is a fully connected layer using a ReLu activation function and L2 regularization; the ninth layer is an output layer using a sigmod classifier for processing to achieve fault data diagnosis; the overall network model is as shown in Figure 2 .

[0071] (2) Model training:

[0072] Model training is to input data to the model for training after the framework structure of the model is determined. The final algorithm model is obtained after training, which is used to judge the state of the bearing. The determination of the model structure is a mutual process with training. The structure listed in step (1) is the optimal structure obtained after comparison and training; the first type of original data obtained by experiment is divided into a training set, a validation set, and a test set according to a certain proportion, a convolutional network is trained, an optimizer and an objective function are set, and the loss rate converges during training. In order to improve the training efficiency and reduce the training time, GPU is called to accelerate the training, and the fault diagnosis model is generated after the training is completed; the accuracy rate and loss rate evolution curve during the training process is as shown in Figure 3 .

[0073] (3) Model performance test:

[0074] The test set data is imported into the generated fault diagnosis model to obtain the diagnosis result of the test set data, and the overall accuracy of the result is calculated, as shown in Figure 4 The overall accuracy is more than 97%.

[0075] On the basis of the above method, the model trained with test data is directly used for on-site operation of wind turbine bearing fault diagnosis. Due to the poor generalization performance of the model, the output result is low in accuracy, there are false positives and false negatives, and even the model cannot be used normally. Therefore, the model needs to be processed and updated regularly to improve the accuracy of the diagnosis result. The updating of the model is a long-term process, and with the continuous accumulation of on-site operation data, the performance of the model is gradually improved. An online updating device for a wind turbine bearing fault diagnosis model, the model online updating device mainly comprises a vibration signal acquisition module, a core processing module, a communication module and auxiliary supporting equipment, the auxiliary supporting equipment comprises power supply equipment and connecting lines etc.; the vibration signal acquisition module comprises a vibration sensor, the vibration sensor is installed on the sensor seat of the driving end and the non-driving end of the wind turbine, the data acquisition module of the vibration sensor realizes vibration data acquisition, and sends the acquired data to the core processing module through Ethernet or other wireless communication mode, the minimum sampling frequency of the data acquisition module of the vibration sensor is not less than 2.56 times of the maximum frequency of the signal; the core processing module comprises CPU and GPU, which is an edge computing module or a ground server platform, and realizes the pre-processing and analysis of vibration data through the deployed software program, and updates the fault diagnosis model; the communication module realizes the data interaction between the vibration signal acquisition module and the core processing module, and the data interaction between the core processing module and the external system.

[0076] The model online updating method of the above-mentioned online updating device for a wind turbine bearing fault diagnosis model, and the online construction method of the fault diagnosis model is adopted, and specifically comprises the following steps: the fault diagnosis model updating method of the driving end and the non-driving end is the same, the diagnosis model updating method of the driving end comprises the following steps:

[0077] ① Model deployment: deploy the driving end bearing fault diagnosis model in the core processing module for bearing fault diagnosis, real-time receive the driving end vibration data of the vibration signal acquisition module, adopt the same method as step (2) of the online construction method of the fault diagnosis model to process the data into 2xL data sequence and input into the model, so as to realize the real-time fault diagnosis function of the driving end bearing, but further processing and updating are needed due to low accuracy, false positives and other reasons;

[0078] ② Running data storage: increase the vibration data files of the bearing in normal and fault states in the core processing module; the normal data file stores the normal operation data of the motor on site; the fault data is the fault operation data of the motor on site, and with the extension of the operation time of the motor on site, the fault operation data can be further divided according to different fault types; in order to prevent the file system from occupying too much storage, fixed data volume and fixed period storage are adopted.

[0079] ③Model online updating: Because of the hysteresis of fault operation data, there is a problem of less or even no field operation fault data, therefore, according to the presence or absence of field operation fault data, model online updating is carried out according to different methods:

[0080] a. When there is only normal operation data or a small amount of fault operation data in addition to normal operation data, only the feature extraction of motor field normal operation data is carried out, the motor field normal operation data within a period of time is processed into a 2xL data sequence, which is input to the model for feature extraction, the data processing method adopts the same method as step (2) of the fault diagnosis model online construction method, the output of the full connection layer of the model is statistically analyzed during extraction, the full connection layer output range R of the motor field operation data of the bearing in the normal state is determined, R is taken as the index for evaluating the normal and fault state of the bearing, and R is considered as the abnormal state of the bearing; Therefore, the model does not need to be trained again, only the classification layer at the end of the model is removed, and the bearing state is diagnosed according to the output of the full connection layer, at this time, the first model online updating is completed, and the model can realize the diagnosis of the normal and fault state of the bearing;

[0081] b. With the extension of the operation time, the motor operation sample data of the bearing fault state gradually increases, the operation data of different fault types of the bearing can meet the requirements of model training, the field operation data stored in the file system is also processed into a 2xL time sequence and labeled, and is divided into a training set, a validation set and a test set according to a certain proportion, the network model in the fault diagnosis model online construction method is retrained, and the training is stopped when the accuracy meets the set requirements or the training time reaches the set conditions, and the original model is replaced by this model, because the normal and different fault types and different fault positions are used for training, the model can not only judge the bearing state but also diagnose the fault type and determine the fault position, and the function of the model is continuously improved with the continuous accumulation of operation data;

[0082] ④Model performance test: The data stored in the file is called regularly and processed into a 2xL time sequence as test data to test the accuracy and false alarm rate of the bearing fault diagnosis model, and whether the model needs to be updated is determined according to the accuracy and false alarm rate; When the motor fault diagnosis accuracy and false alarm rate exceed the set range, the model updating function is started.

[0083] The whole process of the above online updating can be realized by software without human intervention.

[0084] The scope of protection of the present application is not limited to the above specific embodiments, and the present application can have various modifications and changes for those skilled in the art, any modification, improvement and equivalent replacement within the concept and principle of the present application should be included in the protection scope of the present application.

Claims

1. A model online updating method of a wind turbine bearing fault diagnosis model online updating device, characterized in that: The online updating device is implemented in a wind turbine bearing fault diagnosis model, and the online updating device comprises a vibration signal acquisition module, a core processing module, a communication module and auxiliary supporting equipment, the auxiliary supporting equipment comprises a power supply device and a connecting line; the vibration signal acquisition module comprises a vibration sensor, the vibration sensor is installed on a sensor seat at a driving end and a non-driving end of a wind turbine, a data acquisition module of the vibration sensor realizes vibration data acquisition, and the acquired data is sent to the core processing module through Ethernet or wireless communication; the core processing module comprises a CPU and a GPU, and is an edge computing module or a ground server platform, and realizes pre-processing and analysis of vibration data through a deployed software program, and updates a fault diagnosis model; the communication module realizes data interaction between the vibration signal acquisition module and the core processing module, and data interaction between the core processing module and an external system; The online construction method of the fault diagnosis model comprises the following steps: I. Test data acquisition: (1) Bearing fault preparation: The common faults of the bearing are prepared, and the fault types include: bearing inner and outer ring electro-erosion, bearing inner and outer ring peeling, bearing outer ring peeling, and bearing inner ring peeling. The fault bearings are installed at the driving end and the non-driving end of the wind turbine, and there are 9 operation schemes in total: Scheme one: the driving end bearing state is inner and outer ring electro-erosion, and the non-driving end bearing state is normal; Scheme two: the driving end bearing state is normal, and the non-driving end bearing state is inner and outer ring electro-erosion; Scheme three: the driving end bearing state is normal, and the non-driving end bearing state is normal; Scheme four: the driving end bearing state is inner and outer ring peeling, and the non-driving end bearing state is normal; Scheme five: the driving end bearing state is normal, and the non-driving end bearing state is inner and outer ring peeling; Scheme six: the driving end bearing state is outer ring peeling, and the non-driving end bearing state is normal; Scheme seven: the driving end bearing state is normal, and the non-driving end bearing state is outer ring peeling; Scheme eight: the driving end bearing state is inner ring peeling, and the non-driving end bearing state is normal; Scheme nine: the driving end bearing state is normal, and the non-driving end bearing state is inner ring peeling; (2) Test data acquisition and processing: Tilting wind turbines θ ° mounted on test bench, rotor short-circuited, no-load operation, constant rotational speed for each variant n 1、 n 2、 n 3、 ……. n m Collecting single-shaft vibration acceleration data of the drive end and non-drive end for 10 minutes, a total of 9 × m Group raw data; The 9x m The test data is divided into two categories: ① Non-driving end normal, driving end fault original data, namely scheme one, scheme three, scheme four, scheme six and scheme eight; ② Driving end normal, non-driving end fault original data, namely scheme two, scheme three, scheme five, scheme seven and scheme nine; In consideration of the imbalance between positive and negative samples, the original fault data is under-sampled to ensure that the number of normal and fault samples is similar. The under-sampled original fault data is divided into Y parts of time series data with a length of L, and the normal original data is also divided into X parts of time series data with a length of L. A fault label is added to each data sequence to form an original signal data set consisting of X+Y groups of time series data with fault labels. Discrete Fourier transform is performed on the original data sequence with a length of L to obtain frequency domain data with a length of L. The time domain and frequency domain data are merged, so that each time series data changes from 1xL one-dimensional data to 2xL two-dimensional data. To realize the fast algorithm FFT of discrete Fourier transform, the signal sampling frequency Fs of the single-axis acceleration sensor should be greater than 2.56 times the maximum frequency of the signal; II. Model establishment: models are established for the transmission end and the non-transmission end respectively, and the construction method is the same. The model establishment method of the transmission end comprises the following steps: (1) Determine the model structure: The first layer is a convolution layer, which inputs a 2xL grayscale image, filters are 32, strides=(1, 4), the edges are filled, and batch normalization, Relu activation, and maximum pooling are performed; the second layer is a convolution layer, which inputs a 2xL grayscale image, filters are 32, strides=(1, 4), the edges are filled, and batch normalization, Relu activation, and maximum pooling are performed; the third layer is a sequence conversion layer, which converts the data after convolution into an LSTM layer input data format; the fourth layer is a long short-term memory network layer LSTM, which outputs a spatial dimension of 256 and returns all sequences; the fifth layer is a dropout layer with a dropout rate of 0.3; the sixth layer is a long short-term memory network layer LSTM, which outputs a spatial dimension of 32 and returns the last output in the sequence; the seventh layer is a dropout layer with a dropout rate of 0.3; the eighth layer is a fully connected layer using ReLu activation function and L2 regularization; the ninth layer is an output layer using a sigmod classifier for processing to realize fault data diagnosis; (2) Model training: Model training is to input data to the model after the framework structure of the model is determined. After training, the final algorithm model is obtained to judge the state of the bearing. The first type of original data obtained by experiment is divided into training set, validation set and test set in proportion. The convolution network is trained, the optimizer and objective function are set, and the loss rate converges during training. In order to improve the training efficiency and reduce the training time, GPU is called to accelerate the training. After training, the fault diagnosis model is generated; (3) Model performance test: The test set data is imported into the generated fault diagnosis model to obtain the diagnosis result of the test set data, and the overall accuracy of the result is calculated; The non-transmission end bearing fault diagnosis model is the same as the transmission end fault diagnosis model, which is established using the second type of data. The online updating method specifically comprises the following steps: the same method is used for updating the fault diagnosis model of the driving end and the non-driving end, and the diagnosis model updating method of the driving end comprises the following steps: ① Model deployment: the driving end bearing fault diagnosis model is deployed in the core processing module for bearing fault diagnosis, real-time receiving of the field driving end vibration data of the vibration signal collection module, the same method as step (2) of the fault diagnosis model online construction method is used to process the data into a 2xL data sequence and input into the model, so as to realize the real-time fault diagnosis function of the driving end bearing, and wait for further processing and updating; ② Running data storage: the vibration data files of the bearing in normal and fault states are added in the core processing module; the normal data file stores the normal running data of the motor in the field; the fault data is the fault running data of the motor in the field, and the fault running data is further divided according to different fault types as the running time of the motor in the field is prolonged; ③ Model online updating: the model online updating is performed according to different methods according to whether there is field running fault data: a. When there is only normal running data or a small amount of fault running data in addition to the normal running data, only the feature extraction of the normal running data of the motor in the field is performed, the normal running data of the motor in the field within a period of time is processed into a 2xL data sequence and input into the model for feature extraction, the data processing method adopts the same method as step (2) of the fault diagnosis model online construction method, the output of the full connection layer of the model is statistically analyzed during the extraction, the full connection layer output range R of the normal state of the bearing is determined, R is taken as an index for evaluating the normal and fault states of the running bearing, and R is considered as the abnormal state of the bearing; therefore, the model does not need to be trained again, only the classification layer at the end of the model is removed, the bearing state is diagnosed according to the output of the full connection layer, at this time, the first online updating of the model is completed, and at this time, the model can realize the diagnosis of the normal and fault states of the bearing; b. With the prolongation of the running time, the motor running sample data in the fault state of the bearing gradually increases, the running data of the bearing in different fault types can meet the requirements of model training, the field running data stored in the file system is also processed into a 2xL time sequence and labeled, and is divided into a training set, a validation set and a test set according to the proportion, the network model in the fault diagnosis model online construction method is retrained, the training is stopped when the accuracy meets the set requirements or the training time reaches the set condition, and the original model is replaced by the model, since the normal and different fault types and different fault positions are used for training, the model can not only judge the bearing state but also diagnose the fault type and determine the fault position, and the function of the model is continuously increased and improved with the continuous accumulation of the running data; (4) Model performance regular test: regular call file storage data and process into 2xL time series as test data to test bearing fault diagnosis model accuracy, false positive rate, and determine whether the model needs to be updated according to the accuracy and false positive rate; when the motor fault diagnosis accuracy and false positive rate exceed the set range, start the model updating function.

2. The model online updating method of the wind turbine bearing fault diagnosis model online updating device according to claim 1, characterized in that: In step 2, in order to prevent the file system from occupying too much storage, the file storage adopts fixed data volume and fixed period storage.

3. The model online updating method of the wind turbine bearing fault diagnosis model online updating device according to claim 2, characterized in that: The lowest sampling frequency of the data acquisition module of the vibration sensor is not less than 2.56 times of the maximum frequency of the signal.

Images