A progressive health state assessment method for a wind turbine drive train

Abstract

The present application relates to a kind of wind turbine drive chain progressive health state evaluation method, first, wind turbine drive chain vibration signal, acoustic emission detection signal, each low-speed statistical value, each operating parameter statistical value are collected, by gray map and time sequence feature vector, design is mapped into uniform latent vector by cross-modal feature fusion encoder, and first generative adversarial network is learned normal latent vector distribution, and second generative adversarial network is learned from health to failure Continuous evolution path;Further, using the cross-modal feature fusion encoder and second generative adversarial network trained, by wind turbine drive chain latent vector, the nearest neighbor search reverse mapping is executed to obtain continuous health index, the problems of single data source, sample imbalance, health evaluation granularity in prior art are solved, continuous health index and residual life prediction are outputed only by small amount of fault sample, the precision and explainability of wind power transmission chain health state evaluation are significantly improved.

Description

Technical Field

[0001] This invention relates to a progressive health status assessment method for wind turbine drive trains, belonging to the technical field of wind power equipment health management and fault prediction. Background Technology

[0002] The wind turbine drivetrain is the core energy transfer system connecting the wind turbine and the generator, mainly composed of components such as the main shaft, main bearings, gearbox (planetary-level + cylindrical-level), couplings, and generator bearings. Due to the fluctuating and random nature of wind energy, the drivetrain is subjected to irregular directional loads, instantaneous strong impact loads, and alternating stresses over a long period, resulting in its failure rate accounting for the highest proportion of all wind turbine failures. In practical applications, downtime caused by the drivetrain accounts for more than 25% of the rated power generation time, with rolling bearing and gear failures being the most common failure modes. These failures include rolling bearing issues such as inner ring cracks, outer ring cracks, rolling element wear, cage breakage, plastic deformation, and fatigue spalling; gear issues such as tooth surface wear, tooth root cracks, broken teeth, and pitting; and shaft misalignment, imbalance, and rubbing.

[0003] Existing technologies propose methods such as using gradient boosting trees for fault diagnosis and using generative adversarial networks for anomaly detection, but they still have technical shortcomings: including a single data source and limited model architecture. The single data source means that only vibration signals or only acoustic emission signals are used, without deep integration of the two complementary signals as well as torque, temperature, and operating condition parameters, resulting in limited ability to represent complex composite fault modes. The limited model architecture means that only one generative adversarial network is used to learn the normal sample distribution, lacking continuous modeling of the fault evolution process and unable to output a continuous health index. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a progressive health status assessment method for wind turbine drivetrain, which integrates multi-source heterogeneous data fusion and dual generative adversarial network joint design to efficiently achieve health status assessment of wind turbine drivetrain.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention designs a progressive health status assessment method for wind turbine drive train. For a target wind turbine, the following steps A to C are performed to train the cross-modal feature fusion encoder and the second generative adversarial network corresponding to the target wind turbine; then, for the target wind turbine, step D is performed to determine the current state of the target wind turbine drive train.

[0006] Step A. Based on each historical moment, collect vibration signals, acoustic emission detection signals, low-speed statistical values, and statistical values ​​of various operating parameters of the target wind turbine drive chain for a preset duration. Construct vibration grayscale maps, acoustic emission time-spectrum grayscale maps, and time-series feature vectors of the target wind turbine drive chain for the corresponding historical moment. Combine these to form a single sample and determine the true health index corresponding to the sample. Then obtain the samples and their true health indices of the target wind turbine drive chain for each historical moment, and proceed to Step B.

[0007] Step B. Construct a cross-modal feature fusion encoder that processes the vibration grayscale image, acoustic emission time-spectrum grayscale image, and temporal feature vector in the sample respectively to obtain the corresponding features and perform feature fusion to output the corresponding latent vector;

[0008] Construct a first generative adversarial network for learning the normal latent vector distribution, and a second generative adversarial network for learning the continuous evolutionary path from health to degradation;

[0009] Then proceed to step C;

[0010] Step C. First, based on the normal sample set, jointly train the cross-modal feature fusion encoder and the first generative adversarial network. Then, fix the trained cross-modal feature fusion encoder and the trained first generative adversarial network, and train the second generative adversarial network based on the degraded sample set containing normal and abnormal samples to obtain the trained second generative adversarial network. Then proceed to step D.

[0011] Step D. Collect the vibration grayscale image, acoustic emission time spectrum grayscale image, and time-series feature vector of the target wind turbine drive chain at the current moment. Then, through the trained cross-modal feature fusion encoder and the trained second generative adversarial network, determine the state of the target wind turbine drive chain at the current moment by using the health index of the target wind turbine drive chain at the current moment.

[0012] As a preferred technical solution of the present invention: in step A, the following steps A1 to A4 are performed to obtain the samples and their true health index of the target wind turbine transmission chain corresponding to each historical moment.

[0013] Step A1. Perform the following steps for each historical moment, then proceed to Step A2;

[0014] Vibration signals of the target wind turbine drive chain at a preset duration are sampled at equal angles along the X and Y directions. The vibration signals in the X and Y directions are plotted as scatter plots using the axis trajectory method. Then, through bilinear interpolation and linear normalization, a vibration grayscale image of the target wind turbine drive chain at a historical moment is constructed.

[0015] Acoustic emission signal detection is performed on the target wind turbine drive chain for a preset duration to obtain the corresponding acoustic emission detection signal. After cubic spline interpolation and short-time Fourier transform, the dynamic range is compressed by taking the logarithm and then bilinear interpolation is performed. Finally, the signal is normalized to construct the acoustic emission time-spectrum grayscale map of the target wind turbine drive chain at the corresponding historical moment.

[0016] Collect low-speed statistical values ​​of the target wind turbine drive train for a preset duration, as well as statistical values ​​of various operating parameters of the target wind turbine. Then, construct a time-series feature vector of the target wind turbine drive train at a historical moment through standardization.

[0017] Step A2. Based on each historical moment, combine the vibration grayscale image, acoustic emission time spectrum grayscale image, and time series feature vector of the target wind turbine drive chain corresponding to the historical moment to form a single sample, thereby obtaining the samples of the target wind turbine drive chain corresponding to each historical moment, and then proceed to step A3.

[0018] Step A3. Based on the samples corresponding to each historical moment of the target wind turbine drive chain, determine whether there are any samples where the target wind turbine drive chain has failed. If so, define the true health index of the sample where the target wind turbine drive chain has failed as 1, and then proceed to step A4; otherwise, define the true health index of the samples corresponding to each historical moment of the target wind turbine drive chain as 0.

[0019] Step A4. For each of the remaining samples corresponding to the target wind turbine drivetrain, determine whether there is no failure in the target wind turbine drivetrain in the future time direction from the historical time. If yes, define the true health index of the sample corresponding to the target wind turbine drivetrain in the historical time as equal to 0; otherwise, obtain the interval between the historical time and the first time in the future time direction where the target wind turbine drivetrain fails. Execute as follows:

[0020] like If the true health index of the target wind turbine drive train corresponding to the historical moment is defined as 0;

[0021] like Then, the true health index of the target wind turbine drivetrain corresponding to a historical moment is defined as equal to... ;

[0022] This allows us to obtain the true health index of the target wind turbine drivetrain for each of the remaining historical time periods, among which... The base of the natural logarithm. This represents the degradation rate coefficient, which is preset to a positive real number. This represents the threshold for the duration of complete health.

[0023] As a preferred technical solution of the present invention: the cross-modal feature fusion encoder constructed in step B includes a cross-attention module, a gated fusion layer, and three branches. The first branch is a convolutional neural network with a preset number of layers, used to receive and process the vibration grayscale image in the sample and output the corresponding feature. The second branch is a convolutional neural network with a preset number of layers, used to receive and process the acoustic emission time-spectrum grayscale image in the sample and output the corresponding feature. The third branch is a fully connected network with a preset number of layers, used to receive and process the temporal feature vector in the sample and output the corresponding feature. The output of the first branch and the output of the second branch are connected to the input of the cross-attention module. The cross-attention module calculates the attention weights of the outputs of the first and second branches and performs weighted fusion on the outputs of the first and second branches to output the corresponding weighted fused feature. The output of the cross-attention module and the output of the third branch are connected to the input of the gated fusion layer. The gated fusion layer performs a gating mechanism to fuse the latent vector corresponding to the sample for the outputs of the cross-attention module and the third branch.

[0024] As a preferred embodiment of the present invention: the first generative adversarial network constructed in step B includes a first generator. With the first discriminator The first generator The first discriminator takes random noise as input and the corresponding pseudo-normal latent vector as output. The input is a normal latent vector, and the output is the probability that the input normal latent vector is a true normal latent vector.

[0025] The constructed second generative adversarial network includes a second generator. With the second discriminator The second generator The second discriminator takes the normal latent vector and various health indices within the health index range as input, and outputs the pseudo-degenerate latent vector of the normal latent vector with respect to the health index; The input is a degenerate latent vector, and the output is the probability that the input degenerate latent vector is the true degenerate latent vector and the predicted health index corresponding to the input degenerate latent vector.

[0026] As a preferred technical solution of the present invention: in step C, the parameters of the cross-modal feature fusion encoder and the parameters of the first generative adversarial network are initialized, and then the following steps C1 to C6 are executed to realize the joint training of the cross-modal feature fusion encoder and the first generative adversarial network, and to obtain the trained cross-modal feature fusion encoder and the trained first generative adversarial network.

[0027] Step C1. Based on the normal sample set composed of samples whose true health index is equal to 0, randomly sample a predetermined number of samples. Each sample is processed by a cross-modal feature fusion encoder to output corresponding true normal latent vectors;

[0028] Meanwhile, based on the standard normal distribution, random generation A noise sequence, generated by the first generator in the first generative adversarial network. Process the data and output the corresponding pseudo-normal latent vectors; then proceed to step C2.

[0029] Step C2. The first discriminator in the first generative adversarial network. Receive each real normal latent vector and each false normal latent vector respectively, process and output the probability that each input normal latent vector is a real normal latent vector; then proceed to step C3;

[0030] Step C3. Calculate the mean vector of each true normal latent vector output by the cross-modal feature fusion encoder to form the true normal mean latent vector for the current iteration. The results constitute the average discrimination probability of the true normal latent vector in the current iteration, and are calculated. The results constitute the average discrimination probability of the pseudo-normal latent vector in the current iteration, where, Indicates selection based on normal sample set In the nth sample Each sample is processed by a cross-modal feature fusion encoder to output the corresponding true normal latent vector. express After the first discriminator Process the output to determine the probability that it is a true normal latent vector. Indicates selection based on normal sample set In the nth sample Each sample is processed by a cross-modal feature fusion encoder to output the corresponding pseudo-normal latent vector. express After the first discriminator Process the output to determine the probability that it is the true normal latent vector; then proceed to step C4;

[0031] Step C4. Determine whether the following two conditions are met simultaneously. If yes, complete the joint training of the cross-modal feature fusion encoder and the first generative adversarial network to obtain the trained cross-modal feature fusion encoder and the trained first generative adversarial network; otherwise, proceed to step C5.

[0032] Condition 1. Under the first preset number of iterations from the current iteration to the historical time direction, the average discrimination probability of the true normal latent vector and the average discrimination probability of the false normal latent vector are all within the range of fluctuation around a preset proportion centered at 50% in each iteration;

[0033] Condition 2. In the second consecutive preset iteration from the current iteration towards the historical time direction, the Euclidean distance between the true normal mean latent vectors in two consecutive iterations of each group is less than the preset difference threshold.

[0034] Step C5. Fix the parameters in the cross-modal feature fusion encoder and the first generator. The parameters are defined by the following formula:

[0035] ;

[0036] To minimize the loss result Update the first discriminator to the target. Each parameter, Indicates the first discriminator The loss function is then determined, and then proceed to step C6;

[0037] Step C6. Fix the first discriminator Each parameter is defined by the following formula:

[0038] ;

[0039] To minimize the loss result To achieve this, update the parameters in the cross-modal feature fusion encoder and the first generator. Each parameter in the middle, Represents the first generator The loss function is calculated, and then the process returns to step C1 to proceed to the next iteration.

[0040] As a preferred technical solution of the present invention: in step C, based on the parameters of the cross-modal feature fusion encoder after fixed training and the parameters of the first generative adversarial network after training, the parameters of the second generative adversarial network are initialized, and the following steps C7 to C13 are executed to realize the training of the second generative adversarial network and obtain the trained second generative adversarial network;

[0041] Step C7. Based on the degenerate sample set consisting of samples with true health indices equal to 0, 1, and those between 0 and 1, randomly sample a predetermined number of samples. Each sample is processed by a cross-modal feature fusion encoder to output the corresponding true degradation latent vectors, and then proceeds to step C8;

[0042] Step C8. Based on the normal sample set composed of samples whose true health index is equal to 0, randomly sample a preset number of samples. Each sample is processed by a cross-modal feature fusion encoder to output corresponding true normal latent vectors;

[0043] Meanwhile, based on the standard normal distribution, random generation A noise sequence, generated by the first generator in the first generative adversarial network. The process is performed to output the corresponding pseudo-normal latent vectors;

[0044] The second generator in the second generative adversarial network is composed of each real normal latent vector and each false normal latent vector. Received Each normal latent vector is combined with a target health index randomly collected from the range of 0 to 1 to form a second generator. Received Then, proceed to step C9;

[0045] Step C9. By the second generator Receive its corresponding The system processes and outputs the pseudo-degenerate latent vector of the normal latent vector with respect to the target health index in each input combination, and then proceeds to step C10.

[0046] Step C10. By the second discriminator Receive each true degenerate latent vector and each false degenerate latent vector respectively, process and output the probability that the input degenerate latent vector is a true degenerate latent vector and the predicted health index corresponding to the input degenerate latent vector, and then proceed to step C11;

[0047] Step C11. Fix the second generator The parameters are defined by the following formula:

[0048] ;

[0049] ;

[0050] ;

[0051] To minimize Update the second discriminator to the target. The parameters in the middle, among which, Indicates selection based on the degraded sample set In the nth sample Each sample is processed by a cross-modal feature fusion encoder to output the corresponding true degradation latent vector. express After the second discriminator The output is the probability that the value is the true degenerate latent vector. express After the second discriminator Process the output and determine the corresponding predicted health index. Indicates selection based on the degraded sample set In the nth sample The true health index corresponding to each sample. This indicates the second generator in step C8. Received The first normal latent vector The pseudo-degenerate latent vector of a normal latent vector with respect to the target health index in the corresponding input combination. express After the second discriminator The output is the probability that the value is the true degenerate latent vector. This indicates the preset hyperparameters. express After the second discriminator Process the output and determine the corresponding predicted health index. This indicates the second generator in step C8. Received The first normal latent vector Each normal latent vector corresponds to a target health index in the input combination. Indicates the second discriminator The ability to discriminate against losses Indicates the second discriminator The true regression loss, This refers to the second generator. The regression loss is generated, and then proceed to step C12;

[0052] Step C12. Fix the second discriminator The parameters are further combined with the following formula:

[0053] ;

[0054] ;

[0055] To minimize Update the second generator to the target. The parameters in the middle, among which, This indicates the second generator in step C8. Received The first normal latent vector A normal latent vector, express The input combination with the target health index of 0, Indicates input combination via the second generator Process the output pseudo-degenerate latent vectors, Represents the L2 norm. This indicates the preset hyperparameters. This refers to the second generator. Generate adversarial loss, This refers to the second generator. The cyclic consistent loss is then calculated; then proceed to step C13;

[0056] Step C13. Use the following formula:

[0057] ;

[0058] Obtain the total loss of the current iteration. Then determine whether it satisfies the following condition: the total loss after a third consecutive preset number of iterations from the current iteration towards the historical time direction. If the average value is lower than the preset total loss threshold, then the training of the second generative adversarial network is completed and the trained second generative adversarial network is obtained; otherwise, return to step C7 to enter the next iteration.

[0059] As a preferred technical solution of the present invention: step D includes the following steps D1 to D4;

[0060] Step D1. Following the method in Step A, acquire the vibration grayscale image, acoustic emission time-spectrum grayscale image, and time-series feature vector of the target wind turbine drivetrain at the current moment. After training, process the cross-modal feature fusion encoder to obtain the latent vector of the target wind turbine drivetrain at the current moment. Then proceed to step D2;

[0061] Step D2. Based on the normal sample set, a cross-modal feature fusion encoder, after training, obtains the mean vectors between the true normal latent vectors corresponding to a predetermined number of samples, thus forming the normal distribution center latent vector. Then proceed to step D3;

[0062] Step D3. Divide the health index range of 0 to 1 into preset intervals to obtain a preset number. Discrete health index , Indicates the first A discrete health index, and a second generator in a second generative adversarial network. According to the following formula:

[0063] ;

[0064] Obtain the latent vector of the center of the normal distribution Regarding each discrete health index pseudo-degenerate latent vector Then proceed to step D4;

[0065] Step D4. Calculation A pseudo-degenerate latent vector Latent vectors corresponding to the current time of the target wind turbine drive train respectively The Euclidean distance between them is determined, and the smallest Euclidean distance is selected. The current state of the target wind turbine drivetrain is determined by using the actual health index of the target wind turbine drivetrain at the current moment, combined with the numerical range of the health index from 0 to 1 for each preset state.

[0066] As a preferred embodiment of the present invention, it further includes obtaining the real-time health index of the target wind turbine drivetrain based on the real-time execution step D, and triggering the execution of the following steps based on the occurrence of a real-time health index greater than 0:

[0067] Step a. Continue to step D to obtain the real-time health index of the target wind turbine drive train, and then proceed to step b;

[0068] Step b. For each real-time real health index of the target wind turbine drivetrain that is greater than 0, use linear regression to determine the time required for the real-time real health index to change to 1, that is, the remaining lifespan prediction of the target wind turbine drivetrain, output and proceed to step c.

[0069] Step c. Determine if the predicted remaining lifespan is greater than 0. If yes, return to step a; otherwise, trigger an alarm on the target wind turbine drive train and stop operation.

[0070] The progressive health status assessment method for wind turbine drivetrain described in this invention, compared with existing technologies, has the following technical advantages:

[0071] This invention designs a progressive health status assessment method for wind turbine drivetrains. First, it collects vibration signals, acoustic emission detection signals, low-speed statistical values, and statistical values ​​of various operating parameters of the wind turbine drivetrain, converting them into grayscale images and time-series feature vectors. A cross-modal feature fusion encoder is designed to map these into unified latent vectors. A first generative adversarial network (GAN) is constructed to learn the normal latent vector distribution, and a second GAN is constructed to learn the continuous evolution path from health to failure. Then, the trained cross-modal feature fusion encoder extracts the latent vectors of the wind turbine drivetrain, and the trained second GAN performs nearest neighbor search and inverse mapping to obtain a continuous health index, thus achieving status assessment. This method solves the problems of single data source, imbalanced samples, and coarse granularity in existing technologies. It achieves continuous health index and remaining life prediction with only a small number of fault samples, significantly improving the accuracy and interpretability of wind turbine drivetrain health status assessment. Attached Figure Description

[0072] Figure 1This is a schematic diagram of the training and application process of the progressive health status assessment method for wind turbine drive train designed in this invention. Detailed Implementation

[0073] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0074] The present invention discloses a progressive health status assessment method for wind turbine drivetrains. In practical applications, it is based on... Figure 1 As shown, for the target wind turbine, steps A to C are performed to train the cross-modal feature fusion encoder and the second generative adversarial network corresponding to the target wind turbine.

[0075] Step A. Based on each historical moment, collect vibration signals, acoustic emission detection signals, low-speed statistical values, and statistical values ​​of various operating parameters of the target wind turbine drive chain for a preset duration. Construct vibration grayscale maps, acoustic emission time-spectrum grayscale maps, and time-series feature vectors of the target wind turbine drive chain for the corresponding historical moment. Combine these to form a single sample and determine the true health index corresponding to the sample. Then obtain the samples and their true health indices for each historical moment of the target wind turbine drive chain, and proceed to Step B.

[0076] In practical applications, step A above is specifically designed to be executed as follows: steps A1 to A4, to obtain the samples and their true health index of the target wind turbine drive chain at each historical moment.

[0077] Step A1. Perform the following steps based on each historical moment, and then proceed to Step A2.

[0078] Vibration signals of the target wind turbine drive chain at a preset duration are sampled at equal angles along the X and Y directions. The vibration signals in the X and Y directions are plotted as scatter plots using the axis trajectory method. Then, bilinear interpolation and linear normalization are used to construct a vibration grayscale image of the target wind turbine drive chain at a historical moment. In practical applications, for example, each segment has 1024 continuous sampling points and a sampling frequency of 10kHz, which is converted into a 64×64 vibration grayscale image.

[0079] Acoustic emission signal detection is performed on the target wind turbine drivetrain for a preset duration to obtain the corresponding acoustic emission detection signal. After cubic spline interpolation and short-time Fourier transform, the dynamic range is compressed logarithmically and then bilinearly interpolated. Finally, the signal is normalized to construct the acoustic emission time-spectrum grayscale map of the target wind turbine drivetrain at the corresponding historical moment. In practical applications, the acoustic emission signal is broadband with a center frequency of 65 kHz. Each segment is sampled with 512 points, and a 64×64 acoustic emission time-spectrum grayscale map is generated after short-time Fourier transform.

[0080] The system collects low-speed statistics for the target wind turbine drivetrain at a preset duration, as well as statistical values ​​of various operating parameters of the target wind turbine. After standardization, it constructs a time-series feature vector corresponding to the historical moment of the target wind turbine drivetrain. In practical applications, low-speed statistics, such as torque, speed, bearing temperature, and gearbox oil temperature, are sampled with a period of 1 second. Operating parameters, such as wind speed, wind direction, and pitch angle, are obtained from the SCADA system with a sampling period of 7 seconds and interpolated to 1 second.

[0081] In practical applications, after the vibration signal, acoustic emission detection signal, various low-speed statistical values, and statistical values ​​of various operating parameters of the target wind turbine are collected, data cleaning is performed. This includes first removing data from abnormal periods such as shutdown and sensor failure, and then using nearest neighbor interpolation for missing values. After that, the corresponding vibration grayscale image, acoustic emission time spectrum grayscale image, and time series feature vector are obtained.

[0082] Step A2. Based on each historical moment, combine the vibration grayscale image, acoustic emission time spectrum grayscale image, and time series feature vector of the target wind turbine drive chain corresponding to the historical moment to form a single sample, thereby obtaining the samples of the target wind turbine drive chain corresponding to each historical moment, and then proceed to step A3.

[0083] Step A3. Based on the samples corresponding to each historical moment of the target wind turbine drive chain, determine whether there are any samples where the target wind turbine drive chain has failed. If so, define the true health index of the sample where the target wind turbine drive chain has failed as equal to 1, and then proceed to step A4; otherwise, define the true health index of the samples corresponding to each historical moment of the target wind turbine drive chain as equal to 0.

[0084] Step A4. For each of the remaining samples corresponding to the target wind turbine drivetrain, determine whether there is no failure in the target wind turbine drivetrain in the future time direction from the historical time. If yes, define the true health index of the sample corresponding to the target wind turbine drivetrain in the historical time as equal to 0; otherwise, obtain the interval between the historical time and the first time in the future time direction where the target wind turbine drivetrain fails. Execute as follows:

[0085] like If the true health index of the target wind turbine drive train corresponding to the historical moment is defined as 0;

[0086] like Then, the true health index of the target wind turbine drivetrain corresponding to a historical moment is defined as equal to... ;

[0087] This allows us to obtain the true health index of the target wind turbine drivetrain for each of the remaining historical time periods, among which... The base of the natural logarithm. This represents the degradation rate coefficient, which is preset to a positive real number. This represents the threshold for the duration of complete health.

[0088] The health index designed above is defined as follows: a health index of 0 indicates that the target wind turbine drivetrain is healthy; a health index of 1 indicates that the target wind turbine drivetrain is completely failed. For example, the health index is 0.9 one day before the target wind turbine drivetrain failure; 0.8 three days before the failure; 0.5 one month before the failure; and 0.2 three months before the failure.

[0089] Step B. Construct a cross-modal feature fusion encoder that processes the vibration grayscale image, acoustic emission time-spectrum grayscale image, and temporal feature vector in the sample respectively to obtain the corresponding features and perform feature fusion to output the corresponding latent vector; construct a first generative adversarial network for learning the normal latent vector distribution and a second generative adversarial network for learning the continuous evolution path from healthy to degenerate; then proceed to step C.

[0090] In the specific design, the cross-modal feature fusion encoder includes a cross-attention module, a gated fusion layer, and three branches. The first branch is a pre-defined 5-layer convolutional neural network used to process the vibration grayscale image in the sample and output the corresponding feature. The second branch is a pre-defined 5-layer convolutional neural network used to process the acoustic emission temporal spectrum grayscale image in the sample and output the corresponding feature. The third branch is a pre-defined 3-layer fully connected network used to process the temporal feature vector in the sample and output the corresponding feature. The outputs of the first and second branches are connected to the input of the cross-attention module, which calculates the attention weights of the first and second branch outputs and performs weighted fusion on the first and second branch outputs to output the corresponding weighted fused feature. The outputs of the cross-attention module and the third branch are connected to the input of the gated fusion layer, which performs a gating mechanism on the cross-attention module output and the third branch output to fuse the latent vector corresponding to the sample.

[0091] For the first generative adversarial network, the design specifically includes the first generator. With the first discriminator The first generator The first discriminator takes random noise as input and the corresponding pseudo-normal latent vector as output. The input is a normal latent vector, and the output is the probability that the input normal latent vector is a true normal latent vector.

[0092] For the second generative adversarial network, the design specifically includes a second generator. With the second discriminator The second generator The second discriminator takes the normal latent vector and various health indices within the health index range as input, and outputs the pseudo-degenerate latent vector of the normal latent vector with respect to the health index; The input is a degenerate latent vector, and the output is the probability that the input degenerate latent vector is the true degenerate latent vector and the predicted health index corresponding to the input degenerate latent vector.

[0093] Step C. Press Figure 1 As shown, firstly, based on the normal sample set, the cross-modal feature fusion encoder and the first generative adversarial network are jointly trained. Then, the trained cross-modal feature fusion encoder and the trained first generative adversarial network are fixed. Based on the degraded sample set containing normal and abnormal samples, the second generative adversarial network is trained to obtain the trained second generative adversarial network. Then, the process proceeds to step D.

[0094] In practical applications, initialize the parameters of the cross-modal feature fusion encoder and the parameters of the first generative adversarial network, and then execute the following steps C1 to C6 to achieve joint training of the cross-modal feature fusion encoder and the first generative adversarial network, and obtain the trained cross-modal feature fusion encoder and the trained first generative adversarial network.

[0095] Step C1. Based on the normal sample set composed of samples whose true health index is equal to 0, randomly sample a predetermined number of samples. Each sample is processed by a cross-modal feature fusion encoder, which outputs corresponding true normal latent vectors; simultaneously, based on a standard normal distribution, random latent vectors are generated. A noise sequence, generated by the first generator in the first generative adversarial network. Process the data and output the corresponding pseudo-normal latent vectors; then proceed to step C2.

[0096] Step C2. The first discriminator in the first generative adversarial network. Receive each real normal latent vector and each false normal latent vector respectively, process and output the probability that each input normal latent vector is a real normal latent vector; then proceed to step C3.

[0097] Step C3. Calculate the mean vector of each true normal latent vector output by the cross-modal feature fusion encoder to form the true normal mean latent vector for the current iteration. The results constitute the average discrimination probability of the true normal latent vector in the current iteration, and are calculated. The results constitute the average discrimination probability of the pseudo-normal latent vector in the current iteration, where, Indicates selection based on normal sample set In the nth sample Each sample is processed by a cross-modal feature fusion encoder to output the corresponding true normal latent vector. express After the first discriminator Process the output to determine the probability that it is a true normal latent vector. Indicates selection based on normal sample set In the nth sample Each sample is processed by a cross-modal feature fusion encoder to output the corresponding pseudo-normal latent vector. express After the first discriminator Process the output to determine the probability that it is a true normal latent vector; then proceed to step C4.

[0098] Step C4. Determine whether the following two conditions are met simultaneously. If yes, complete the joint training of the cross-modal feature fusion encoder and the first generative adversarial network to obtain the trained cross-modal feature fusion encoder and the trained first generative adversarial network; otherwise, proceed to step C5.

[0099] Condition 1. Under the first preset number of iterations from the current iteration to the historical time direction, the average discrimination probability of the true normal latent vector and the average discrimination probability of the false normal latent vector are all within the range of fluctuation around a preset proportion centered at 50% in each iteration;

[0100] Condition 2. In the second consecutive preset iteration from the current iteration towards the historical time direction, the Euclidean distance between the true normal mean latent vectors in two consecutive iterations of each group is less than the preset difference threshold.

[0101] Step C5. Fix the parameters in the cross-modal feature fusion encoder and the first generator. The parameters are defined by the following formula:

[0102] ;

[0103] To minimize the loss result Update the first discriminator to the target. Each parameter, Indicates the first discriminator The loss function is then used to proceed to step C6.

[0104] Step C6. Fix the first discriminator Each parameter is defined by the following formula:

[0105] ;

[0106] To minimize the loss result To achieve this, update the parameters in the cross-modal feature fusion encoder and the first generator. Each parameter in the middle, Represents the first generator The loss function is calculated, and then the process returns to step C1 to proceed to the next iteration.

[0107] Further fix the parameters in the cross-modal feature fusion encoder after training, as well as the parameters in the first generative adversarial network after training. Then initialize the parameters in the second generative adversarial network and execute steps C7 to C13 to train the second generative adversarial network and obtain the trained second generative adversarial network.

[0108] Step C7. Based on the degenerate sample set consisting of samples with true health indices equal to 0, 1, and those between 0 and 1, randomly sample a predetermined number of samples. Each sample is processed by a cross-modal feature fusion encoder to output the corresponding true degradation latent vectors, and then proceeds to step C8.

[0109] Step C8. Based on the normal sample set composed of samples whose true health index is equal to 0, randomly sample a preset number of samples. Each sample is processed by a cross-modal feature fusion encoder, which outputs corresponding true normal latent vectors; simultaneously, based on a standard normal distribution, random latent vectors are generated. A noise sequence, generated by the first generator in the first generative adversarial network. The process is performed to output the corresponding pseudo-normal latent vectors.

[0110] The second generator in the second generative adversarial network is composed of each real normal latent vector and each false normal latent vector. Received Each normal latent vector is combined with a target health index randomly collected from the range of 0 to 1 to form a second generator. Received The input combination is then processed, and then step C9 is performed.

[0111] Step C9. By the second generator Receive its corresponding The system processes and outputs the pseudo-degenerate latent vector of the normal latent vector with respect to the target health index for each input combination, and then proceeds to step C10.

[0112] Step C10. By the second discriminator Each true degenerate latent vector and each false degenerate latent vector are received, processed and output to represent the probability that the input degenerate latent vector is a true degenerate latent vector and the predicted health index corresponding to the input degenerate latent vector, and then proceed to step C11.

[0113] Step C11. Fix the second generator The parameters are defined by the following formula:

[0114] ;

[0115] ;

[0116] ;

[0117] To minimize Update the second discriminator to the target. The parameters in the middle, among which, Indicates selection based on the degraded sample set In the nth sample Each sample is processed by a cross-modal feature fusion encoder to output the corresponding true degradation latent vector. express After the second discriminator The output is the probability that the value is the true degenerate latent vector. express After the second discriminator Process the output and determine the corresponding predicted health index. Indicates selection based on the degraded sample set In the nth sample The true health index corresponding to each sample. This indicates the second generator in step C8. Received The first normal latent vector The pseudo-degenerate latent vector of a normal latent vector with respect to the target health index in the corresponding input combination. express After the second discriminator The output is the probability that the value is the true degenerate latent vector. This indicates the preset hyperparameters. express After the second discriminator Process the output and determine the corresponding predicted health index. This indicates the second generator in step C8. Received The first normal latent vector Each normal latent vector corresponds to a target health index in the input combination. Indicates the second discriminator The ability to discriminate against losses Indicates the second discriminator The true regression loss, This refers to the second generator. The regression loss is generated, and then the process proceeds to step C12.

[0118] Step C12. Fix the second discriminator The parameters are further combined with the following formula:

[0119] ;

[0120] ;

[0121] To minimize Update the second generator to the target. The parameters in the middle, among which, This indicates the second generator in step C8. Received The first normal latent vector A normal latent vector, express The input combination with the target health index of 0, Indicates input combination via the second generator Process the output pseudo-degenerate latent vectors, Represents the L2 norm. This indicates the preset hyperparameters. This refers to the second generator. Generate adversarial loss, This refers to the second generator. The cyclic consistent loss is then calculated; then proceed to step C13.

[0122] Step C13. Use the following formula:

[0123] ;

[0124] Obtain the total loss of the current iteration. Then determine whether it satisfies the following condition: the total loss after a third consecutive preset number of iterations from the current iteration towards the historical time direction. If the average value is lower than the preset total loss threshold, then the training of the second generative adversarial network is completed and the trained second generative adversarial network is obtained; otherwise, return to step C7 to enter the next iteration.

[0125] Following steps A to C above, the trained cross-modal feature fusion encoder and the trained second generative adversarial network corresponding to the target wind turbine are obtained. Further, in practical applications, for the target wind turbine, according to... Figure 1 As shown, step D is executed to determine the current state of the target wind turbine drive train.

[0126] Step D. Collect the vibration grayscale image, acoustic emission time spectrum grayscale image, and time-series feature vector of the target wind turbine drive chain at the current moment. Then, through the trained cross-modal feature fusion encoder and the trained second generative adversarial network, determine the state of the target wind turbine drive chain at the current moment by using the health index of the target wind turbine drive chain at the current moment.

[0127] In practical applications, the above step D is specifically designed to be executed as follows: steps D1 to D4.

[0128] Step D1. Following the method in Step A, acquire the vibration grayscale image, acoustic emission time-spectrum grayscale image, and time-series feature vector of the target wind turbine drivetrain at the current moment. After training, process the cross-modal feature fusion encoder to obtain the latent vector of the target wind turbine drivetrain at the current moment. Then proceed to step D2.

[0129] Step D2. Based on the normal sample set, a cross-modal feature fusion encoder, after training, obtains the mean vectors between the true normal latent vectors corresponding to a predetermined number of samples, thus forming the normal distribution center latent vector. Then proceed to step D3.

[0130] Step D3. Divide the health index range of 0 to 1 into preset intervals to obtain a preset number. Discrete health index , Indicates the first A discrete health index, and a second generator in a second generative adversarial network. According to the following formula:

[0131] ;

[0132] Obtain the latent vector of the center of the normal distribution Regarding each discrete health index pseudo-degenerate latent vector Then proceed to step D4.

[0133] Step D4. Calculation A pseudo-degenerate latent vector Latent vectors corresponding to the current time of the target wind turbine drive train respectively The Euclidean distance between them is determined, and the smallest Euclidean distance is selected. The current state of the target wind turbine drivetrain is determined by using the actual health index of the target wind turbine drivetrain at the current moment, combined with the numerical range of the health index from 0 to 1 for each preset state.

[0134] By executing design step D above, the target wind turbine drivetrain can be monitored in real time to obtain its real-time health index. Therefore, in practical applications, based on this real-time health index, a life prediction scheme for the target wind turbine drivetrain can be further designed. Specifically, during the real-time execution of step D, if the real-time health index is greater than 0, then... Figure 1 As shown, steps a to c are then triggered to achieve prediction.

[0135] Step a. Continue with step D to obtain the real-time health index of the target wind turbine drive train, and then proceed to step b.

[0136] Step b. For each real-time real health index of the target wind turbine drivetrain that is greater than 0, use linear regression to determine the time required for the real-time real health index to change to 1, that is, the remaining lifespan prediction of the target wind turbine drivetrain, output and proceed to step c.

[0137] Step c. Determine if the predicted remaining lifespan is greater than 0. If yes, return to step a; otherwise, trigger an alarm on the target wind turbine drive train and stop operation.

[0138] The progressive health status assessment method for wind turbine drivetrain designed in the above technical solution first collects vibration signals, acoustic emission detection signals, low-speed statistical values, and statistical values ​​of various operating parameters of the wind turbine drivetrain. These are converted into grayscale images and time-series feature vectors, respectively. A cross-modal feature fusion encoder is designed to map these into a unified latent vector. A first generative adversarial network is constructed to learn the normal latent vector distribution, and a second generative adversarial network is constructed to learn the continuous evolution path from health to failure. Then, the latent vectors of the wind turbine drivetrain are extracted using the trained cross-modal feature fusion encoder, and the continuous health index is obtained by performing nearest neighbor search and inverse mapping in combination with the trained second generative adversarial network. This achieves status assessment and solves the problems of single data source, imbalanced samples, and coarse granularity of health assessment in existing technologies. It realizes the output of continuous health index and remaining life prediction with only a small number of failure samples, significantly improving the accuracy and interpretability of wind turbine drivetrain health status assessment.

[0139] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A progressive health status assessment method for wind turbine drivetrain, characterized in that: For the target wind turbine, perform steps A to C to train the cross-modal feature fusion encoder and the second generative adversarial network corresponding to the target wind turbine; then, for the target wind turbine, perform step D to determine the current state of the transmission chain of the target wind turbine. Step A. Based on each historical moment, collect vibration signals, acoustic emission detection signals, low-speed statistical values, and statistical values ​​of various operating parameters of the target wind turbine drive chain for a preset duration. Construct vibration grayscale maps, acoustic emission time-spectrum grayscale maps, and time-series feature vectors of the target wind turbine drive chain for the corresponding historical moment. Combine these to form a single sample and determine the true health index corresponding to the sample. Then obtain the samples and their true health indices of the target wind turbine drive chain for each historical moment, and proceed to Step B. Step B. Construct a cross-modal feature fusion encoder that processes the vibration grayscale image, acoustic emission time-spectrum grayscale image, and temporal feature vector in the sample respectively to obtain the corresponding features and perform feature fusion to output the corresponding latent vector; Construct a first generative adversarial network for learning the normal latent vector distribution, and a second generative adversarial network for learning the continuous evolutionary path from health to degradation; Then proceed to step C; Step C. First, based on the normal sample set, jointly train the cross-modal feature fusion encoder and the first generative adversarial network. Then, fix the trained cross-modal feature fusion encoder and the trained first generative adversarial network, and train the second generative adversarial network based on the degraded sample set containing normal and abnormal samples to obtain the trained second generative adversarial network. Then proceed to step D. Step D. Collect the vibration grayscale image, acoustic emission time spectrum grayscale image, and time-series feature vector of the target wind turbine drive chain at the current moment. Then, through the trained cross-modal feature fusion encoder and the trained second generative adversarial network, determine the state of the target wind turbine drive chain at the current moment by using the health index of the target wind turbine drive chain at the current moment.

2. The progressive health status assessment method for wind turbine drive train according to claim 1, characterized in that: In step A, steps A1 to A4 are performed to obtain samples and their true health indices for each historical moment corresponding to the target wind turbine drive chain. Step A1. Perform the following steps for each historical moment, then proceed to step A2; Vibration signals of the target wind turbine drive chain at a preset duration are sampled at equal angles along the X and Y directions. The vibration signals in the X and Y directions are plotted as scatter plots using the axis trajectory method. Then, through bilinear interpolation and linear normalization, a vibration grayscale image of the target wind turbine drive chain at a historical moment is constructed. Acoustic emission signal detection is performed on the target wind turbine drive chain for a preset duration to obtain the corresponding acoustic emission detection signal. After cubic spline interpolation and short-time Fourier transform, the dynamic range is compressed by taking the logarithm and then bilinear interpolation is performed. Finally, the signal is normalized to construct the acoustic emission time-spectrum grayscale map of the target wind turbine drive chain at the corresponding historical moment. Collect low-speed statistical values ​​of the target wind turbine drive train for a preset duration, as well as statistical values ​​of various operating parameters of the target wind turbine. Then, construct a time-series feature vector of the target wind turbine drive train at a historical moment through standardization. Step A2. Based on each historical moment, combine the vibration grayscale image, acoustic emission time spectrum grayscale image, and time series feature vector of the target wind turbine drive chain corresponding to the historical moment to form a single sample, thereby obtaining the samples of the target wind turbine drive chain corresponding to each historical moment, and then proceed to step A3. Step A3. Based on the samples corresponding to each historical moment of the target wind turbine drive chain, determine whether there are any samples in which the target wind turbine drive chain has failed. If so, define the true health index of the sample in which the target wind turbine drive chain has failed as equal to 1, and then proceed to step A4. Otherwise, the true health index of the target wind turbine drive train corresponding to each historical moment is defined as 0. Step A4. For each of the remaining samples corresponding to the target wind turbine drivetrain, determine whether there is no failure in the target wind turbine drivetrain in the future time direction from the historical time. If yes, define the true health index of the sample corresponding to the target wind turbine drivetrain in the historical time as equal to 0; otherwise, obtain the interval between the historical time and the first time in the future time direction where the target wind turbine drivetrain fails. Execute as follows: like If the true health index of the target wind turbine drive train corresponding to the historical moment is defined as 0; like Then, the true health index of the target wind turbine drivetrain corresponding to a historical moment is defined as equal to... ; This allows us to obtain the true health index of the target wind turbine drivetrain for each of the remaining historical time periods, among which... The base of the natural logarithm. This represents the degradation rate coefficient, which is preset to a positive real number. This represents the threshold for the duration of complete health.

3. The progressive health status assessment method for wind turbine drive train according to claim 1, characterized in that: The cross-modal feature fusion encoder constructed in step B includes a cross-attention module, a gated fusion layer, and three branches. The first branch is a convolutional neural network with a preset number of layers, used to process the vibration grayscale image in the sample and output the corresponding feature. The second branch is a convolutional neural network with a preset number of layers, used to process the acoustic emission temporal spectrum grayscale image in the sample and output the corresponding feature. The third branch is a fully connected network with a preset number of layers, used to process the temporal feature vector in the sample and output the corresponding feature. The outputs of the first and second branches are connected to the input of the cross-attention module, which calculates the attention weights between the outputs of the first and second branches and performs weighted fusion on them, outputting the corresponding weighted fused feature. The outputs of the cross-attention module and the third branch are connected to the input of the gated fusion layer, which performs a gating mechanism to fuse the latent vector corresponding to the sample based on the outputs of the cross-attention module and the third branch.

4. The progressive health status assessment method for wind turbine drive train according to claim 1, characterized in that: The first generative adversarial network constructed in step B includes a first generator. With the first discriminator The first generator The first discriminator takes random noise as input and the corresponding pseudo-normal latent vector as output. The input is a normal latent vector, and the output is the probability that the input normal latent vector is a true normal latent vector. The constructed second generative adversarial network includes a second generator. With the second discriminator The second generator The second discriminator takes the normal latent vector and various health indices within the health index range as input, and outputs the pseudo-degenerate latent vector of the normal latent vector with respect to the health index; The input is a degenerate latent vector, and the output is the probability that the input degenerate latent vector is the true degenerate latent vector and the predicted health index corresponding to the input degenerate latent vector.

5. The progressive health status assessment method for wind turbine drive train according to claim 1, characterized in that: In step C, the parameters of the cross-modal feature fusion encoder and the parameters of the first generative adversarial network are initialized. Then, steps C1 to C6 are executed to achieve joint training of the cross-modal feature fusion encoder and the first generative adversarial network, and the trained cross-modal feature fusion encoder and the trained first generative adversarial network are obtained. Step C1. Based on the normal sample set composed of samples whose true health index is equal to 0, randomly sample a predetermined number of samples. Each sample is processed by a cross-modal feature fusion encoder to output corresponding true normal latent vectors; Meanwhile, based on the standard normal distribution, random generation A noise sequence, generated by the first generator in the first generative adversarial network. Process the data and output the corresponding pseudo-normal latent vectors; then proceed to step C2. Step C2. The first discriminator in the first generative adversarial network. Receive each real normal latent vector and each false normal latent vector respectively, process and output the probability that each input normal latent vector is a real normal latent vector; Then proceed to step C3; Step C3. Calculate the mean vector of each true normal latent vector output by the cross-modal feature fusion encoder to form the true normal mean latent vector for the current iteration. The results constitute the average discrimination probability of the true normal latent vector in the current iteration, and are calculated. The results constitute the average discrimination probability of the pseudo-normal latent vector in the current iteration, where, Indicates selection based on normal sample set In the nth sample Each sample is processed by a cross-modal feature fusion encoder to output the corresponding true normal latent vector. express After the first discriminator Process the output to determine the probability that it is a true normal latent vector. Indicates selection based on normal sample set In the nth sample Each sample is processed by a cross-modal feature fusion encoder to output the corresponding pseudo-normal latent vector. express After the first discriminator Process the output to determine the probability that it is the true normal latent vector; then proceed to step C4; Step C4. Determine whether the following two conditions are met simultaneously. If yes, complete the joint training of the cross-modal feature fusion encoder and the first generative adversarial network to obtain the trained cross-modal feature fusion encoder and the trained first generative adversarial network; otherwise, proceed to step C5. Condition 1. Under the first preset number of iterations from the current iteration to the historical time direction, the average discrimination probability of the true normal latent vector and the average discrimination probability of the false normal latent vector are all within the range of fluctuation around a preset proportion centered at 50% in each iteration; Condition 2. In the second consecutive preset iteration from the current iteration towards the historical time direction, the Euclidean distance between the true normal mean latent vectors in two consecutive iterations of each group is less than the preset difference threshold. Step C5. Fix the parameters in the cross-modal feature fusion encoder and the first generator. The parameters are defined by the following formula: ； To minimize the loss result Update the first discriminator to the target. Each parameter, Indicates the first discriminator The loss function is then determined, and then proceed to step C6; Step C6. Fix the first discriminator Each parameter is defined by the following formula: ； To minimize the loss result To achieve the goal, update the parameters of the cross-modal feature fusion encoder and the first generator. Each parameter in the middle, Represents the first generator The loss function is calculated, and then the process returns to step C1 to proceed to the next iteration.

6. The progressive health status assessment method for wind turbine drivetrain according to claim 5, characterized in that: In step C, based on the parameters of the cross-modal feature fusion encoder after fixed training and the parameters of the first generative adversarial network after training, the parameters of the second generative adversarial network are initialized, and steps C7 to C13 are executed to train the second generative adversarial network and obtain the trained second generative adversarial network. Step C7. Based on the degenerate sample set consisting of samples with true health indices equal to 0, 1, and those between 0 and 1, randomly sample a predetermined number of samples. Each sample is processed by a cross-modal feature fusion encoder to output the corresponding true degradation latent vectors, and then proceeds to step C8; Step C8. Based on the normal sample set composed of samples whose true health index is equal to 0, randomly sample a preset number of samples. Each sample is processed by a cross-modal feature fusion encoder to output corresponding true normal latent vectors; Meanwhile, based on the standard normal distribution, random generation A noise sequence, generated by the first generator in the first generative adversarial network. The process is performed to output the corresponding pseudo-normal latent vectors; The second generator in the second generative adversarial network is composed of each real normal latent vector and each false normal latent vector. Received Each normal latent vector is combined with a target health index randomly collected from the range of 0 to 1 to form a second generator. Received Then, proceed to step C9; Step C9. By the second generator Receive its corresponding The system processes and outputs the pseudo-degenerate latent vector of the normal latent vector with respect to the target health index in each input combination, and then proceeds to step C10. Step C10. By the second discriminator Receive each true degenerate latent vector and each false degenerate latent vector respectively, process and output the probability that the input degenerate latent vector is a true degenerate latent vector and the predicted health index corresponding to the input degenerate latent vector, and then proceed to step C11; Step C11. Fix the second generator The parameters are defined by the following formula: ； ； ； To minimize Update the second discriminator to the target. The parameters in the middle, among which, Indicates selection based on the degraded sample set In the nth sample Each sample is processed by a cross-modal feature fusion encoder to output the corresponding true degradation latent vector. express After the second discriminator The output is the probability that the value is the true degenerate latent vector. express After the second discriminator Process the output and determine the corresponding predicted health index. Indicates selection based on the degraded sample set In the nth sample The true health index corresponding to each sample This indicates the second generator in step C8. Received The first normal latent vector The pseudo-degenerate latent vector of a normal latent vector with respect to the target health index in the corresponding input combination. express After the second discriminator The output is the probability that the value is the true degenerate latent vector. This indicates the preset hyperparameters. express After the second discriminator Process the output and determine the corresponding predicted health index. This indicates the second generator in step C8. Received The first normal latent vector Each normal latent vector corresponds to a target health index in the input combination. Indicates the second discriminator The ability to discriminate against losses Indicates the second discriminator The true regression loss, This refers to the second generator. The regression loss is generated, and then proceed to step C12; Step C12. Fix the second discriminator The parameters are further combined with the following formula: ； ； To minimize Update the second generator to the target. The parameters in the middle, among which, This indicates the second generator in step C8. Received The first normal latent vector A normal latent vector, express The input combination with the target health index of 0, Indicates input combination via the second generator Process the output pseudo-degenerate latent vectors, Represents the L2 norm. This indicates the preset hyperparameters. This refers to the second generator. Generate adversarial loss, This refers to the second generator. The cyclic consistent loss is then calculated; then proceed to step C13; Step C13. Use the following formula: ； Obtain the total loss of the current iteration. Then determine whether it satisfies the following condition: the total loss after a third consecutive preset number of iterations from the current iteration towards the historical time direction. If the average value is lower than the preset total loss threshold, then the training of the second generative adversarial network is completed and the trained second generative adversarial network is obtained; otherwise, return to step C7 to enter the next iteration.

7. The progressive health status assessment method for wind turbine drive train according to claim 1, characterized in that: Step D includes the following steps D1 to D4; Step D1. Following the method in Step A, acquire the vibration grayscale image, acoustic emission time-spectrum grayscale image, and time-series feature vector of the target wind turbine drivetrain at the current moment. After training, process the cross-modal feature fusion encoder to obtain the latent vector of the target wind turbine drivetrain at the current moment. Then proceed to step D2; Step D2. Based on the normal sample set, a cross-modal feature fusion encoder, after training, obtains the mean vectors between the true normal latent vectors corresponding to a predetermined number of samples, thus forming the normal distribution center latent vector. Then proceed to step D3; Step D3. Divide the health index range of 0 to 1 into preset intervals to obtain a preset number. Discrete health index , Indicates the first A discrete health index, and a second generator in a second generative adversarial network. According to the following formula: ； Obtain the latent vector of the center of the normal distribution Regarding each discrete health index pseudo-degenerate latent vector Then proceed to step D4; Step D4. Calculation A pseudo-degenerate latent vector Latent vectors corresponding to the current time of the target wind turbine drive train respectively The Euclidean distance between them is determined, and the smallest Euclidean distance is selected. The current state of the target wind turbine drivetrain is determined by using the actual health index of the target wind turbine drivetrain at the current moment, combined with the numerical range of the health index from 0 to 1 for each preset state.

8. The progressive health status assessment method for wind turbine drivetrain according to claim 7, characterized in that: It also includes obtaining the real-time health index of the target wind turbine drivetrain based on the real-time execution step D, and triggering the following steps when the real-time health index is greater than 0: Step a. Continue to step D to obtain the real-time health index of the target wind turbine drive train, and then proceed to step b; Step b. For each real-time real health index of the target wind turbine drivetrain that is greater than 0, use linear regression to determine the time required for the real-time real health index to change to 1, that is, the remaining lifespan prediction of the target wind turbine drivetrain, output and proceed to step c. Step c. Determine if the predicted remaining lifespan is greater than 0. If yes, return to step a; otherwise, trigger an alarm on the target wind turbine drive chain and stop operation.

Images