Wind turbine gearbox wear state evaluation method and system fusing vibration and oil

Abstract

The present application belongs to the technical field of gear box wear state evaluation. A wind turbine gear box wear state evaluation method and system are provided, which fuse vibration and oil. A Markov transition field is used to convert wind turbine gear box vibration signals into time-frequency domain images. The time-frequency domain images are resampled to obtain vibration two-dimensional time-frequency images with the same pixels as the abrasive particle ring two-dimensional images. The abrasive particle ring two-dimensional images and the vibration two-dimensional time-frequency images are down-sampled respectively, and the down-sampled abrasive particle ring two-dimensional images and the down-sampled vibration two-dimensional time-frequency images are fused. The fusion results are up-sampled. The wear state evaluation results are obtained according to the up-sampled results. The present application avoids feature selection bias caused by subjective factors, and improves the accuracy and robustness of wear state recognition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of gear box wear state evaluation, and particularly relates to a wind turbine gear box wear state evaluation method and system fusing vibration and oil. BACKGROUND

[0002] The statements in this section merely provide background technology related to the present application and do not necessarily constitute prior art.

[0003] In the field of gear system health monitoring, the abrasive particle and vibration online monitoring technologies have become key means for evaluating the health condition of the gear box. The abrasive particle monitoring reflects the wear state by analyzing the relative concentration change of the abrasive particles in the oil, and further explores the wear mechanism through abrasive particle morphology analysis; and the vibration monitoring can track the change trend in the process of tooth surface wear in real time, and exhibits different sensitivity from the abrasive particle monitoring. However, the wear process of the gear system has obvious stages, and the interaction mechanism between the tooth surface wear and the dynamics behavior is complex, resulting in a complex nonlinear relationship between the tooth surface wear state and the vibration and abrasive particle characteristics and the sensitivity difference between the characteristics, and this mechanism complexity leads to the obvious difference in sensitivity of a single monitoring means in different wear stages, and increases the uncertainty of the wear state evaluation. Therefore, it is difficult to realize the comprehensive and accurate evaluation of the tooth surface wear state of the gear system by relying on a single monitoring means.

[0004] The multi-source sensing information fusion technology has significant potential in realizing the high reliability evaluation and life prediction of mechanical equipment, and provides a new idea for the wear state evaluation of the gear system. Compared with a single monitoring characteristic, the abrasive particle and vibration characteristics exhibit obvious complementarity in evaluating the tooth surface wear state, and the evaluation method fusing the two characteristics can effectively improve the accuracy and reliability of the evaluation. The current mainstream fusion method relies on the manual extraction of time-frequency characteristics (mean, kurtosis, etc.) and oil characteristics (concentration, morphology, etc.), and this subjective characteristic selection not only loses the detailed information of the original data, but also is difficult to capture the deep nonlinear correlation between the vibration waveform and the abrasive particle image. The existing feature layer or decision layer fusion architecture cannot directly utilize the complementary characteristics of the multi-source data due to the reliance on the prior knowledge to define the weight, resulting in a low early wear recognition rate and a high false alarm rate. SUMMARY

[0005] In order to solve the problems of insufficient cross-modal feature mining, too much artificial intervention and limited fusion level, the application provides a wind turbine gearbox wear state evaluation method and system fusing vibration and oil, adopts Markov transition field to convert the wind turbine gearbox vibration signal into a time-frequency domain image, resamples the time-frequency domain image to obtain a vibration two-dimensional time-frequency graph with the same pixels as the abrasive particle ring two-dimensional image, and directly carries out automatic processing and comprehensive analysis on the vibration two-dimensional time-frequency graph and the abrasive particle ring two-dimensional image, so that the feature selection deviation caused by subjective factors is avoided, the potential correlation between different data modes is mined, the details and complex characteristics in the original information can be better preserved, the correlation and complementarity between features can be fully mined, and the accuracy and robustness of wear state recognition are improved.

[0006] In order to achieve the above-mentioned purpose, the application adopts the following technical solutions:

[0007] In the first aspect, the application provides a wind turbine gearbox wear state evaluation method fusing vibration and oil.

[0008] A wind turbine gearbox wear state evaluation method fusing vibration and oil comprises the following processes:

[0009] Synchronously acquired wind turbine gearbox vibration signals and abrasive particle ring two-dimensional images used for representing abrasive particle concentration changes are acquired;

[0010] Markov transition field is adopted to convert the wind turbine gearbox vibration signals into time-frequency domain images, and the time-frequency domain images are resampled to obtain vibration two-dimensional time-frequency graphs with the same pixels as the abrasive particle ring two-dimensional images;

[0011] The abrasive particle ring two-dimensional images and the vibration two-dimensional time-frequency graphs are down-sampled respectively, the down-sampled abrasive particle ring two-dimensional images and the down-sampled vibration two-dimensional time-frequency graphs are fused, and the fusion results are up-sampled;

[0012] According to the up-sampled results, wear state evaluation results are obtained.

[0013] In the second aspect, the application provides a wind turbine gearbox wear state evaluation system fusing vibration and oil.

[0014] A wind turbine gearbox wear state evaluation system fusing vibration and oil comprises:

[0015] An image acquisition unit is configured to acquire synchronously acquired wind turbine gearbox vibration signals and abrasive particle ring two-dimensional images used for representing abrasive particle concentration changes;

[0016] A two-dimensional time-frequency diagram generation unit is configured to convert a wind turbine gearbox vibration signal into a time-frequency domain image by using a Markov transition field, and to resample the time-frequency domain image to obtain a vibration two-dimensional time-frequency diagram having the same pixels as the abrasive particle ring two-dimensional image.

[0017] A feature fusion unit is configured to down-sample the abrasive particle ring two-dimensional image and the vibration two-dimensional time-frequency diagram respectively, and to fuse the down-sampled abrasive particle ring two-dimensional image and the down-sampled vibration two-dimensional time-frequency diagram, and to up-sample the fusion result.

[0018] A wear state generation unit is configured to obtain a wear state evaluation result according to the up-sampled result.

[0019] In a third aspect, the present application provides a computer device, comprising a processor and a computer readable storage medium.

[0020] The processor is adapted to execute a computer program.

[0021] The computer readable storage medium has a computer program stored therein, and the computer program is executed by the processor to implement the wind turbine gearbox wear state evaluation method fusing vibration and oil as described in the first aspect of the present application.

[0022] In a fourth aspect, the present application provides a computer readable storage medium having a computer program stored therein, and the computer program is adapted to be loaded and executed by a processor to implement the wind turbine gearbox wear state evaluation method fusing vibration and oil as described in the first aspect of the present application.

[0023] Compared with the prior art, the present application has the following beneficial effects:

[0024] The present application innovatively provides a wind turbine gearbox wear state evaluation method fusing vibration and oil, which converts one-dimensional vibration time series signals into two-dimensional time-frequency images by Markov transition field technology, completely retains signal dynamic evolution characteristics (such as state transition probability distribution), and avoids information loss caused by traditional artificial feature extraction.

[0025] The present application innovatively provides a wind turbine gearbox wear state evaluation method fusing vibration and oil, which constructs a double-branch U-Net fusion model, directly fuses two-dimensional time-frequency images and oil original images in the encoder, and uses a convolutional neural network to autonomously mine cross-modal correlation features (such as coupling rules of wear impact energy distribution and abrasive particle morphology change), thereby significantly improving feature representation ability.

[0026] This invention innovatively provides a method for assessing the wear status of wind turbine gearboxes by integrating vibration and oil. From the input of two-dimensional time-frequency images and original oil images to the output of wear status, no manual intervention is required, and it supports the real-time deployment of wind power field online monitoring systems.

[0027] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0028] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0029] Figure 1 A flowchart illustrating a method for evaluating the wear condition of a wind turbine gearbox that integrates vibration and oil, provided as an exemplary embodiment of the present invention;

[0030] Figure 2 A schematic diagram of a dual-branch neural network provided as an exemplary embodiment of the present invention;

[0031] Figure 3 This is a schematic diagram comparing wear assessment results of different signal confusion matrices, provided as an exemplary embodiment of the present invention. Figure 3 (A) in the diagram is a schematic diagram corresponding to multi-source data fusion. Figure 3 (B) in the diagram is a schematic diagram corresponding to the vibration signal. Figure 3 (C) in the diagram represents the corresponding abrasive particles in the oil.

[0032] Figure 4 A schematic diagram of a wind turbine gearbox wear condition assessment system that integrates vibration and oil parameters, provided as an exemplary embodiment of the present invention;

[0033] Figure 5 A schematic diagram of a computer device provided for an exemplary embodiment of the present invention. Detailed Implementation

[0034] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0035] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0036] As described in the background, the existing fusion method needs to manually extract the features of different monitoring information for feature layer and decision layer fusion, which is greatly affected by human subjectivity, resulting in low diagnosis efficiency and poor accuracy. Compared with feature layer and decision layer fusion, data level fusion does not need to rely on human feature extraction, and directly processes and analyzes the vibration signals and oil images in the form of raw data, thereby avoiding the feature selection bias caused by subjective factors, mining the potential correlation between different data modalities, and better retaining the details and complex characteristics of the original information through data level fusion, so that the correlation and complementarity between features can be fully mined. In addition, data level fusion can also realize the deep integration of cross-modal information, provide more comprehensive data input for machine learning models, and further improve the accuracy and robustness of fault pattern recognition.

[0037] Some researchers have proposed a fault diagnosis method combining vibration, oil and noise feature fusion, aiming to improve the fault diagnosis ability of wind power gear speed increasing box under complex working conditions. This patent uses deep learning and DS evidence theory to fuse the non-stationary features of vibration signals, sound pressure level and frequency spectrum features of noise signals, and particle type and particle size distribution features of oil, and constructs a comprehensive evaluation model of fault state. However, the above research relies on multiple detection methods, including vibration, noise and oil sensing information, and does not conduct deep mining and optimization on a single information source, especially in the fineness of data fusion; some researchers have proposed a multi-source and multi-feature fusion diagnosis method for aero-engine main bearing faults, which uses multiple sensors (including vibration sensors, temperature sensors and acoustic sensors) to collect aero-engine main bearing operating state data, processes the monitoring data through signal denoising, feature extraction and machine learning algorithm, thereby realizing the detection and positioning of mechanical equipment faults. However, this research relies more on algorithm processing after feature extraction in the fusion of monitoring information, and does not involve the underlying fusion strategy of multi-sensor data; some researchers have proposed a method and test bench design for gear box fault diagnosis using multi-source sensor data fusion technology. This research collects the vibration signals, noise signals, temperature signals, displacement signals and oil data of the lubricating oil of the gear box, combines the multi-sensor data fusion technology based on BP neural network and D-S evidence theory, and realizes the diagnosis of gear box faults. However, this research mainly relies on feature level and decision level fusion, and does not focus on optimizing the synergy of vibration and lubricating oil monitoring information in data level fusion.

[0038] Therefore, the present application proposes a wind turbine gear box wear state evaluation method combining vibration and oil, as shown in Figure 1 The method comprises the following processes:

[0039] S101: Acquire the synchronously collected vibration signal of the wind turbine gearbox and a two-dimensional image of the abrasive ring used to characterize the change in abrasive particle concentration;

[0040] S102: The vibration signal of the wind turbine gearbox is converted into a time-frequency domain image using a Markov transfer field. The time-frequency domain image is resampled to obtain a vibration two-dimensional time-frequency image with the same pixels as the two-dimensional image of the abrasive ring.

[0041] S103: Downsample the two-dimensional image of the abrasive ring and the two-dimensional time-frequency diagram of the vibration respectively, fuse the downsampled two-dimensional image of the abrasive ring and the downsampled two-dimensional time-frequency diagram of the vibration, and upsample the fusion result;

[0042] S104: The wear condition assessment results are obtained based on the upsampling results.

[0043] In S101 of this implementation, specifically, it includes:

[0044] By conducting a full-life wear test on a wind turbine gearbox, and combining online vibration sensors and high-throughput online oil abrasive image monitoring sensors, real-time monitoring of the gearbox under different wear conditions is achieved. The vibration sensors and high-throughput online oil abrasive image monitoring sensors are used to synchronously collect the vibration characteristics and oil abrasive information of the gearbox under different wear conditions in real time. The collected data are then labeled according to the actual wear conditions to form a labeled vibration and oil abrasive dataset.

[0045] Specifically, S102 of this implementation includes:

[0046] In mechanical systems, the concentration of wear particles is directly proportional to the degree of wear of the friction pair. The higher the concentration of wear particles in the lubricating oil, the more severe the wear. A high-throughput online oil wear particle image monitoring sensor can directly acquire a two-dimensional image of the wear particle ring, characterizing changes in wear particle concentration. Simultaneously, the synchronously acquired vibration signal is one-dimensional time-series data, requiring data processing and spatial alignment with the wear particle image data to facilitate subsequent state evaluation of the intelligent network model. This invention uses a Markov transfer field (MTF) to convert the time-series vibration signal into a time-frequency domain image, and further resamples it to obtain a two-dimensional time-frequency vibration image with the same pixel count as the wear particle image. Specifically, this includes:

[0047] Signal discretization: Discretizing vibration signals Further discretization into a sequence of finite states Through quantization function Generate discrete state values:

[0048] (1);

[0049] in, For discrete interval boundaries, For state values, Representing the i-th discrete state value, Representing the i-th vibration signal value.

[0050] Transition probability matrix: use Markov chain model to calculate the transition probability between each state , describe the dynamic evolution characteristics of the time series signal:

[0051] (2);

[0052] Generate MTF time-frequency domain graph: by logarithmic transformation of the transition probability matrix, generate Markov transition field image (MTF) containing time domain and frequency domain information. The matrix after logarithmic transformation Has the following form:

[0053] (3);

[0054] Where, is a smoothing factor to avoid zero value problems in logarithmic calculation, and the row index and column index of the matrix Correspond to the horizontal coordinate (X axis, representing time) and vertical coordinate (Y axis, representing frequency) of the MTF image pixel respectively.

[0055] The generated Markov transition field image (MTF) is in the form of a two-dimensional image, which can effectively show the dynamic change characteristics of the vibration signal, and is convenient for subsequent intelligent diagnosis model to perform feature extraction and classification.

[0056] Time-frequency graph resampling: in order to ensure that the Markov transition field image (MTF) is the same as the oil abrasive particle image pixel, the Markov transition field image (MTF) is resampled using an interpolation algorithm to obtain a vibration two-dimensional time-frequency graph with the same pixel as the oil abrasive particle image.

[0057] In S103 and S104 in the present implementation, specifically, it includes:

[0058] The present application adopts U-Net network architecture to evaluate the wear state of wind power gear box. The U-Net architecture is composed of encoder (downsampling part), skip connection and decoder (upsampling part), which can effectively extract and fuse information from different modalities of vibration and oil. The model architecture is shown in Figure 2 , and the model design principles include:

[0059] Double-branch independent downsampling: vibration and oil branches are respectively subjected to 5 layers of downsampling, and the end features are fused;

[0060] Fusion position: channel concatenation after 5th layer downsampling (at minimum resolution);

[0061] Symmetric decoder: restore resolution by 5th layer upsampling, skip connection from the same layer encoder (i.e. first layer downsampling skip connection with first layer upsampling, second layer downsampling skip connection with second layer upsampling, third layer downsampling skip connection with third layer upsampling, fourth layer downsampling skip connection with fourth layer upsampling).

[0062] More specifically, the input includes data from two channels of vibration and oil:

[0063] Input 1: , representing a vibration two-dimensional time-frequency diagram generated by a vibration signal, H represents the height of the image, W represents the width of the image, and R represents the real number field;

[0064] Input 2: , representing an oil abrasive particle image;

[0065] The two inputs enter the respective independent encoder branches, and the encoder branches include five downsampling layers, each of which includes a 2D convolution layer (Conv2D), a maximum pooling (MaxPool) layer, etc.

[0066] The convolution block expression of the first layer is:

[0067] (4);

[0068] In the formula, is the output feature map of the first layer downsampling layer; represents a convolution operation for extracting oil / vibration signal features, etc. represents BatchNorm, i.e. batch normalization; is a ReLU activation function; and are the convolution kernel weights and bias of the first layer; the pooling operation uses maximum pooling to reduce the feature map size. The final extracted and output features of each branch are:

[0069]

[0070] (5);

[0071] (6);

[0072] wherein, 𝐶 represents the number of channels increasing layer by layer (for example: 16→32→64→128→256), represents the height of the output feature, ​​​width of the representative output feature, representing a two-dimensional time-frequency diagram of the vibration encoded output feature, representing an oil particle image encoded output feature.

[0073] After extracting the key features of vibration and oil through the double-branch network, deep feature fusion is performed at the minimum resolution to maximize the retention of global statistical correlation of vibration-oil and avoid local noise interference caused by early fusion (such as misjudgment of oil bubbles as wear). The fused feature map is used as the input of the decoder, and the fused feature is represented as:

[0074] (7);

[0075] The decoder performs upsampling and classification reconstruction on the fused feature map. The decoder uses transposed convolution (Transposed) and jump connection to reconstruct spatial features. Through the deconvolution layer, the spatial resolution of the oil / vibration image is gradually restored for learning the correlation between oil / vibration features and wear state. Among them, the deconvolution operation of the first layer is:

[0076] (8);

[0077] wherein, is the transposed convolution, denotes the transposed convolution, is the convolution kernel weight and bias of the first layer, and the pooling operation uses maximum pooling to reduce the feature map size, and the output size is restored layer by layer (e.g., 256→128→64→32→16).

[0078] The same layer features in the encoder are spliced with the deconvolution results of the current layer of the decoder to enhance the edge and spatial semantic information, which is represented as:

[0079] (9);

[0080] The output layer adopts a four-level progressive feature compression and classification architecture, and its method flow is as follows:

[0081] The spatial aggregation layer (feature compression) performs global spatial aggregation on the three-dimensional feature tensor output by the decoder to generate a channel feature vector:

[0082] (10);

[0083] wherein, denotes the feature value of the th channel at position . ​

[0084] The dense layer maps the feature vector processed by the spatial aggregation layer to a hidden representation space through linear transformation, which is used to further learn the association between the "wear state" and the "image feature", and can be expressed as:

[0085] (11);

[0086] wherein, is the output vector of the spatial aggregation layer; and is the weight and bias of the dense layer.

[0087] In the wind turbine gearbox wear data scenario, the dropout layer is used to effectively alleviate the overfitting problem of the model in the training set, and to improve the generalization ability of identifying different wear stages of the wind turbine gearbox, which can be expressed as:

[0088] (12);

[0089] wherein, is the dropout probability.

[0090] The softmax layer converts the real number vector output by the dense layer into a multi-class probability distribution, and each value represents the probability that the input sample belongs to a certain class (wear degree), which can be expressed as:

[0091] (13);

[0092] wherein, is the probability of belonging to the class (such as severe wear); represents the number of categories of classification, which is 4 (running-in, mild, abnormal, severe) in the present application, represents the result of the dropout layer processing of the wear state, , represents the result of the dropout layer processing of the wear state.

[0093] The model training process constructed by the present application adopts the cross-entropy loss function, which is used to learn the feature distribution of the running-in, mild, abnormal and severe wear states of the gearbox;

[0094] (14);

[0095] wherein, is the real label encoding.

[0096] In order to make the network find the optimal solution of wear state recognition faster, the application uses Adam to update optimization, which can be expressed as:

[0097] (15);

[0098] In the formula, is the weight of the first step; indicates the learning rate; , respectively, the first order momentum (mean) and the second order momentum (square mean) of the gradient; indicates a minimum constant to prevent division by zero.

[0099] After the model training is completed, the model prediction accuracy is evaluated through the fan gear box validation set data, and the main evaluation indexes include wear feature classification accuracy, confusion matrix and the like. Through these evaluation indexes, the accuracy of the model in the actual application process to the different wear stages of the gear box can be prepared and evaluated, the model performance can be optimized, and a basis can be provided for wear state classification.

[0100] The wind turbine gear box wear state evaluation method based on oil particle and vibration multi-source information fusion analysis of the application comprises the steps of collecting gear box vibration and oil particle signals, preprocessing signals, building an intelligent model network, and evaluating the health state of sample data. Figure 1

[0101] Firstly, the application carries out a full life cycle wear test of a wind turbine gear box test bench, and collects full life cycle monitoring signals through an acceleration vibration sensor and a high-throughput oil particle image online monitoring sensor. The experimental duration is 284 hours, and vibration and oil online monitoring data are automatically and synchronously collected at intervals of 20 minutes during the experimental duration, and 851 groups of data are accumulated. During the monitoring process, the vibration signal sampling frequency is 2.56 kHz, and the sampling time length is 3 s. The oil particle ring image can be directly generated through a high-throughput oil particle image online monitoring sensor CMOS1 imaging system. These data cover the complete process of the gear system from the running-in period, normal wear, abnormal wear to severe wear.

[0102] Secondly, according to the sampling frequency and sampling time of the vibration signal, 7680 data points of each vibration signal can be calculated within 3 seconds. 7680 data points (sampling frequency 2.56 kHz) collected within 3 seconds are discretized into 5 state intervals (contour coefficient 0.81) through K-means clustering, and a 5*5 Markov transition matrix is generated. In order to retain the state transition probability characteristics, a two-step interpolation method is used, first, the matrix is expanded to 16*16 through cubic interpolation, and then pixel replication is used to enlarge to 64*64 and pair with the oil image. ​

[0103] Again, the application divides the processed 851 vibration time-frequency maps and wear particle pattern samples according to the actual wear state of the gearbox, and divides them into 4 categories of wear states. According to the wear stage, it is divided into 16 groups of running-in period, 644 groups of light wear, 116 groups of abnormal wear and 72 groups of severe wear. In order to alleviate the imbalance of categories, stratified sampling is used (75% of each category): 12 groups of running-in, 483 groups of light wear, 87 groups of abnormal wear and 54 groups of severe wear, a total of 596 groups; the remaining 255 groups of data are used as the test set, 4 groups of running-in, 161 groups of light wear, 29 groups of abnormal wear and 18 groups of severe wear, to verify the accuracy and robustness of the method.

[0104] Finally, a dual-channel encoder design (vibration branch / oil branch) is adopted, and each branch contains 5 levels of feature extraction layers; each level of down-sampling layer adopts a composite structure of "3x3 convolution→batch normalization→ReLU→2x2 max pooling"; the number of channels increases exponentially (16→32→64→128→256), which controls the computational complexity while ensuring the depth of feature extraction.

[0105] After the 5th layer of down-sampling (feature map size 2x2x256), feature splicing is performed; 1x1 convolution is used for cross-channel information interaction, and the spliced 512-dimensional features are reduced to 256-dimensional. The decoder gradually recovers the resolution through 5 layers of transpose convolution, and is connected with the encoder features of the same layer through jump connection, and finally outputs 64x64x16 of the reconstructed features.

[0106] The model training uses Adam optimizer, the initial learning rate is set to , the momentum parameters , . The loss function selects the cross-entropy loss function to accurately distinguish the four states of running-in, light wear, abnormal wear and severe wear. In order to prevent overfitting, Dropout regularization (dropout rate 20%) is used in the training process and the early stopping mechanism is set: when the validation set loss decreases by less than for 10 consecutive epochs, the training is automatically terminated, and the maximum training period is limited to 500 times.

[0107] In the model evaluation stage, the reserved 255 independent test samples are used to verify the performance. Through the calculation of overall classification accuracy, the precision and recall of each category and other quantitative indicators, and the combination of confusion matrix, the recognition ability of the model in different wear stages is comprehensively evaluated.

[0108] In order to highlight the advantages of the multi-source data fusion method proposed in the application, the performance of the single-modal model using only vibration signals or oil wear particle images is tested simultaneously, as shown in Figure 3 In the confusion matrix, the horizontal axis is the real wear state of the gearbox, and the vertical axis is the wear state evaluated by the model, Figure 3(A) in (A) is a schematic diagram corresponding to multi-source data fusion (accuracy 97%), Figure 3 (B) in (B) is a schematic diagram corresponding to vibration signal (accuracy 71%), Figure 3 (C) in (C) is a schematic diagram corresponding to oil particle (accuracy 79%). Figure 3 The numbers in each cell of the table represent how many times the system evaluates as another state when the true state is a certain state. The numbers on the diagonal of the table represent the number of times the system evaluates correctly, and the numbers in other cells represent the number of times the system evaluates incorrectly.

[0109] The experimental results show that the dual-branch fusion model of the application achieves a recognition accuracy of 97% on the test set, which is significantly improved compared to the single-modal model method (vibration about 71%, oil about 79%), verifying the effectiveness of the data-level fusion strategy.

[0110] The experimental results show that the method proposed in the application has better recognition accuracy and robustness compared to traditional single-oil and single-vibration signal analysis methods, and its successful application will provide more reliable and efficient technical support for health monitoring and fault diagnosis of mechanical equipment.

[0111] Figure 4 A wind turbine gearbox wear state evaluation system fusing vibration and oil is shown, comprising:

[0112] The image acquisition unit 401 is configured to acquire the synchronously collected wind turbine gearbox vibration signal and the particle ring two-dimensional image for representing the change of particle concentration;

[0113] The two-dimensional time-frequency diagram generation unit 402 is configured to convert the wind turbine gearbox vibration signal into a time-frequency domain image by using a Markov transition field, and resample the time-frequency domain image to obtain a vibration two-dimensional time-frequency diagram having the same pixels as the particle ring two-dimensional image;

[0114] The feature fusion unit 403 is configured to down-sample the particle ring two-dimensional image and the vibration two-dimensional time-frequency diagram respectively, fuse the down-sampled particle ring two-dimensional image and the down-sampled vibration two-dimensional time-frequency diagram, and up-sample the fusion result;

[0115] The wear state generation unit 404 is configured to obtain a wear state evaluation result according to the up-sampled result

[0116] It can be understood that the above-mentioned units can be combined into one or several other units respectively or entirely, or some of the units can be further split into a plurality of units with smaller functions to constitute, which can achieve the same operation without affecting the implementation of the technical effects of the embodiments of the present application. The above-mentioned units are divided based on logical functions, and in actual application, the function of one unit can also be implemented by a plurality of units, or the functions of a plurality of units are implemented by one unit. In other embodiments of the present application, the system can also include other units, and in actual application, these functions can also be assisted by other units to achieve, and can be achieved by cooperation of a plurality of units.

[0117] According to another embodiment of the present application, the system described in the embodiment can be constructed by running a computer program (including program code) capable of performing each step involved in the corresponding method of the present application on a general computing device such as a computer including processing elements and storage elements such as a Central Processing Unit (CPU), a Random Access Memory (RAM), a Read Only Memory (ROM), etc., the computer program can be recorded on a computer readable recording medium, and loaded into the above-mentioned computing device through the computer readable recording medium and run therein.

[0118] Figure 5 An electronic device is shown, which includes a processor 501, a communication interface 502, and a computer readable storage medium 503. Among them, the processor 501, the communication interface 502 and the computer readable storage medium 503 can be connected through a bus or other means.

[0119] Among them, the communication interface 502 is used to receive and send data, the computer readable storage medium 503 can be stored in the memory of the electronic device, the computer readable storage medium 503 is used to store computer programs, the computer programs include program instructions, and the processor 501 is used to execute the program instructions stored in the computer readable storage medium 503.

[0120] The processor 501 is the computing core and control core of the electronic device, which is suitable for implementing one or more instructions, and is particularly suitable for loading and executing one or more instructions to realize a corresponding method flow or a corresponding function.

[0121] The processor 501 is configured to perform the following process:

[0122] Obtaining the vibration signal of the wind turbine gearbox synchronously collected and the two-dimensional image of the abrasive particle ring for representing the change of the abrasive particle concentration;

[0123] The wind turbine gearbox vibration signal is converted into a time-frequency domain image by using a Markov transition field, the time-frequency domain image is resampled, and a vibration two-dimensional time-frequency image with the same pixels as the abrasive particle ring two-dimensional image is obtained;

[0124] The abrasive particle ring two-dimensional image and the vibration two-dimensional time-frequency image are respectively down-sampled, the down-sampled abrasive particle ring two-dimensional image and the down-sampled vibration two-dimensional time-frequency image are fused, and the fusion result is up-sampled;

[0125] The wear state evaluation result is obtained according to the up-sampling result.

[0126] The application further provides a computer readable storage medium, which is a memory device in an electronic device and is used for storing programs and data.

[0127] In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and the instructions can be one or more computer programs (including program codes).

[0128] In one embodiment, one or more instructions are stored in the computer readable storage medium; the processor loads and executes the one or more instructions stored in the computer readable storage medium to realize the following process:

[0129] The wind turbine gearbox vibration signal and the abrasive particle ring two-dimensional image for representing abrasive particle concentration change are acquired by synchronous acquisition;

[0130] The wind turbine gearbox vibration signal is converted into a time-frequency domain image by using a Markov transition field, the time-frequency domain image is resampled, and a vibration two-dimensional time-frequency image with the same pixels as the abrasive particle ring two-dimensional image is obtained;

[0131] The abrasive particle ring two-dimensional image and the vibration two-dimensional time-frequency image are respectively down-sampled, the down-sampled abrasive particle ring two-dimensional image and the down-sampled vibration two-dimensional time-frequency image are fused, and the fusion result is up-sampled;

[0132] The wear state evaluation result is obtained according to the up-sampling result.

[0133] The application further provides a computer program product or computer program, which comprises computer instructions stored in a computer readable storage medium.

[0134] Obtain the wind turbine gearbox vibration signals synchronously collected and the abrasive particle ring two-dimensional image for representing the abrasive particle concentration change;

[0135] Convert the wind turbine gearbox vibration signals into time-frequency domain images by using a Markov transition field, resample the time-frequency domain images to obtain vibration two-dimensional time-frequency images with the same pixels as the abrasive particle ring two-dimensional image;

[0136] Downsample the abrasive particle ring two-dimensional image and the vibration two-dimensional time-frequency image respectively, fuse the downsampled abrasive particle ring two-dimensional image and the downsampled vibration two-dimensional time-frequency image, and upsample the fusion result;

[0137] Obtain the wear state evaluation result according to the upsampled result.

[0138] Those skilled in the art can understand that the units and algorithm steps of each example described in combination with the embodiments disclosed in the present application can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized by hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0139] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions according to the embodiments of the present application are generated. The computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in or transmitted by a computer readable storage medium. The computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through a wired (for example, coaxial cable, optical fiber, digital line) or wireless (for example, infrared, wireless, microwave, etc.) manner. The computer readable storage medium can be any available medium that can be accessed by a computer or a data processing device such as a server, data center, etc. integrated with one or more available media. The available media can be magnetic media (for example, floppy disk, hard disk, magnetic tape), optical media (for example, DVD), or semiconductor media (for example, solid state disk) and the like.

[0140] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A wind turbine gearbox wear condition assessment method that fuses vibration and oil data, characterized in that, The method comprises the following steps: obtaining wind turbine gearbox vibration signals synchronously collected and abrasive ring two-dimensional images for representing abrasive concentration changes; converting the wind turbine gearbox vibration signals into time-frequency domain images by using a Markov transition field, resampling the time-frequency domain images to obtain vibration two-dimensional time-frequency images having the same pixels as the abrasive ring two-dimensional images; downsampling the abrasive ring two-dimensional images and the vibration two-dimensional time-frequency images respectively, fusing the downsampled abrasive ring two-dimensional images and the downsampled vibration two-dimensional time-frequency images, and upsampling the fusion result; obtaining a wear state evaluation result according to the upsampling result; converting the wind turbine gearbox vibration signals into time-frequency domain images by using a Markov transition field, comprising: discretizing the wind turbine gearbox vibration signals into a sequence of finite states, and generating discrete state values by using a quantization function; calculating transition probabilities between each state value by using a Markov chain model according to the discrete state values; performing logarithmic transformation on the transition probabilities to generate a Markov transition field image containing time domain and frequency domain information; calculating transition probabilities between each state value by using a Markov chain model, comprising: where s t+1 represents the (t+1)th signal in a sequence of finite states, s t represents the tth signal in a sequence of finite states, q i represents the ith discrete state value, q j represents the jth discrete state value, count is a counting function; obtaining a wear state evaluation result according to the upsampling result, comprising: a spatial aggregation layer performs global spatial aggregation on the upsampling result to generate a channel feature vector; a fully connected layer maps the channel feature vector to a hidden representation space through linear transformation; a Dropout layer processes the hidden representation space; a Softmax layer performs probability calculation on the processing result of the Dropout layer to obtain the probability of each wear state, and takes the wear state with the maximum probability as the final wear state evaluation result; the spatial aggregation layer performs global spatial aggregation on the upsampling result to generate a channel feature vector, comprising: where H, W represent the height and width of the vibration two-dimensional time-frequency map, C is the number of channels, U i,j,k represents the feature value of the kth channel at position (i, j); The fully connected layer maps the channel feature vector through a linear transformation to a hidden representation space h k comprising: h k = σ(W fc · z k + b fc ); where W fc and b fc represent the fully connected layer weights and bias, and σ represents the activation function; the Dropout layer processes the hidden representation space, comprising: h′ k = Dropout(h k , p), p = 0.2; wherein p is a dropout probability; the Softmax layer performs probability calculation on the processing result of the Dropout layer, comprising: wherein, the probability that the wear state belongs to the k-th class, K denotes the number of classes of wear states, h j represents the result of the Dropout layer processing for the j-th wear state, j = 1,..., K, h k represents the result of the Dropout layer processing for the k-th wear state.

2. The wind turbine gearbox wear state evaluation method for fusing vibration and oil, according to claim 1, characterized in that, the logarithmic transformation on the transition probability matrix comprises: M(i, j) = log(P(i, j) + ∈); wherein ∈ is a smoothing factor, and M(i, j) represents the logarithmic transformation result of the i th discrete state value and the j th discrete state value.

3. A wind turbine gearbox wear condition assessment system that fuses vibration and oil data, characterized in that, The wind turbine gearbox wear state evaluation method for fusing vibration and oil, according to claim 1 or 2, comprising: an image acquisition unit configured to obtain wind turbine gearbox vibration signals synchronously collected and abrasive ring two-dimensional images for representing abrasive concentration changes; a two-dimensional time-frequency image generation unit configured to convert the wind turbine gearbox vibration signals into time-frequency domain images by using a Markov transition field, resample the time-frequency domain images to obtain vibration two-dimensional time-frequency images having the same pixels as the abrasive ring two-dimensional images; The feature fusion unit is configured to: down-sample the abrasive particle ring two-dimensional image and the vibration two-dimensional time-frequency diagram respectively, fuse the down-sampled abrasive particle ring two-dimensional image and the down-sampled vibration two-dimensional time-frequency diagram, and up-sample the fusion result; The wear state generation unit is configured to obtain a wear state evaluation result according to the up-sampled result.

4. A computer device, comprising: The method comprises: a processor and a computer readable storage medium; a processor adapted to execute a computer program; a computer readable storage medium, wherein a computer program is stored in the computer readable storage medium, and the computer program is executed by the processor to implement the fusion vibration and oil wear state evaluation method of the wind turbine gearbox according to claim 1 or 2.

5. A computer readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is adapted to be loaded and executed by the processor to implement the fusion vibration and oil wear state evaluation method of the wind turbine gearbox according to claim 1 or 2.

Images