A wind turbine gearbox diagnosis method based on semi-supervised momentum prototype network

Abstract

The present invention relates to a wind turbine gearbox diagnosis method based on a semi-supervised momentum prototype network, and belongs to the field of fault diagnosis. The method includes two processes: semi-supervised momentum prototype network training and fault diagnosis and identification; 1. Semi-supervised momentum prototype network training: 2. Wind turbine gearbox fault diagnosis: The present invention combines pseudo-label learning in semi-supervision with semi-supervised momentum prototype network in small sample learning, and proposes a wind turbine gearbox fault diagnosis method based on a semi-supervised momentum prototype network. The method makes full use of the fault information contained in the unlabeled samples, greatly reduces the overfitting phenomenon of the prototype network due to the scarcity of labeled samples, and enhances the stability and accuracy of the semi-supervised momentum prototype network during fine-tuning.

Description

Technical Field

[0001] The invention belongs to the field of fault diagnosis and relates to a wind turbine gearbox diagnosis method based on a semi-supervised momentum prototype network. Background Art

[0002] Wind turbines operate under harsh conditions such as alternating loads all year round, which can easily lead to failures in key components of wind turbines. Among them, wind turbine gearbox failures cause the longest downtime, the highest maintenance costs, and the greatest economic losses. Therefore, real-time fault diagnosis and monitoring of wind turbine gearboxes should be carried out, and timely maintenance strategies should be adopted when faults occur to effectively control the further development of faults.

[0003] Wind turbine gearbox fault diagnosis methods based on deep learning have developed rapidly in the past few years. They use the powerful feature dimension reduction and pattern recognition capabilities of neural networks to perform fault diagnosis and identification of vibration signals. Compared with traditional diagnostic algorithms, deep learning has more powerful high-dimensional, nonlinear abstract data feature extraction capabilities and more accurate pattern recognition capabilities, without the need for manual feature extraction. Deep learning methods such as autoencoders (AE), convolutional neural networks (CNN) and recurrent neural networks (RNN) have been widely used in the field of wind turbine gearbox fault diagnosis with sufficient labeled samples and have shown good performance. However, in practical applications, expensive labeled samples are not easy to obtain, and the scarcity of labeled samples leads to extremely low fault identification accuracy. Therefore, it is of great significance to conduct research on wind turbine gearbox fault diagnosis under the conditions of limited labeled samples and sufficient unlabeled samples. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a wind turbine gearbox diagnosis method based on a semi-supervised momentum prototype network. The present invention introduces pseudo-label learning in semi-supervised learning into small sample learning and makes targeted improvements, thereby proposing a wind turbine gearbox fault diagnosis method based on a semi-supervised momentum prototype network. This method utilizes the powerful embedding expression ability and small sample classification ability of the semi-supervised momentum prototype network to construct a feature mapping space and prototypes. Through the similarity between unlabeled samples and prototypes in the feature mapping space and a threshold selection strategy based on Monte Carlo uncertainty, unlabeled samples are given high-confidence pseudo-labels and are regarded as labeled samples for retraining the semi-supervised momentum prototype network.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] A wind turbine gearbox diagnosis method based on a semi-supervised momentum prototype network comprises the following steps:

[0007] S1: Semi-supervised momentum prototype network training;

[0008] S2: Fault diagnosis and identification.

[0009] Optionally, the S1 is specifically:

[0010] S11: Sample set production;

[0011] S12: Based on a limited set of labeled samples, a semi-supervised momentum prototype network is trained to learn feature maps and prototypes.

[0012] S13: Based on the trained semi-supervised momentum prototype network, pseudo-label learning is performed on the unlabeled sample set, pseudo-label information after threshold selection is given to the unlabeled samples, and the scale of the labeled sample set is continuously expanded to obtain the expanded sample set E;

[0013] S14: Based on the expanded sample set E, fine-tune the semi-supervised momentum prototype network, optimize the feature mapping of the semi-supervised momentum prototype network, and modify the prototype based on the momentum prototype.

[0014] Optionally, S11 is specifically as follows: normalizing the original vibration signal sample set containing K types of fault categories into standardized one-dimensional samples, and dividing a small number of labeled samples into a support set S and a query set Q, as a semi-supervised momentum prototype network training sample set L for S12; and making the unlabeled samples into a pseudo-label learning sample set U for S13.

[0015] Optionally, S12 specifically includes: training the semi-supervised momentum prototype network based on the semi-supervised momentum prototype network training sample set L:

[0016] Semi-supervised Momentum Prototype Network Utilizing Feature Extractor The sample x in the semi-supervised momentum prototype network training sample set L i Embed into a specific feature space:

[0017]

[0018] For the cth type of fault, the prototype P is generated using the support set S c :

[0019]

[0020] Calculate the distance between the sample in the query set Q and each prototype in the feature space, and use the calculated distance as the predicted output p of the sample after softmax:

[0021]

[0022] The semi-supervised momentum prototype network is trained by back propagation through the loss function loss:

[0023] loss = -lg p(y = k | x)

[0024] Through the above process, the semi-supervised momentum prototype network continuously optimizes the feature mapping and prototype of the semi-supervised momentum prototype network.

[0025] Optionally, S13 specifically comprises: based on the trained semi-supervised momentum prototype network, performing pseudo-label learning on the pseudo-label learning sample set U, using MC-Dropout in the document ArXiv preprint arXiv:1506.02142,Jun.2015. “Dropoutas a Bayesian Approximation:Representing Model Uncertainty in Deep Learning.” to measure the uncertainty of the prediction of the semi-supervised momentum prototype network, and using this as an evaluation index for threshold selection, giving unlabeled samples high-confidence pseudo-labels, and continuously expanding the scale of the labeled sample set;

[0026] Use the semi-supervised momentum prototype network after adding MC-Dropout to perform I forward propagations on the unlabeled samples in the pseudo-label learning sample set U to obtain I predictions, and calculate the mean μ and variance σ of the I prediction probability vectors 2 :

[0027]

[0028] Using the variance σ describing the degree of dispersion 2 Calculate the standard deviation φ of I forward passes:

[0029]

[0030] The information entropy ψ of the I-th forward propagation is calculated using the mean μ describing the central tendency:

[0031]

[0032] The Monte Carlo uncertainty θ of the semi-supervised momentum prototype network prediction is calculated using the standard deviation φ and information entropy ψ:

[0033]

[0034] Among them, λ is a hyperparameter that adjusts the weight between the information entropy ψ and the standard deviation φ; based on the calculated Monte Carlo uncertainty θ predicted by the semi-supervised momentum prototype network, the model output of the unlabeled samples with high confidence is selected to generate pseudo labels after the set threshold θ′, and the scale of the labeled samples is expanded to obtain the expanded sample set E.

[0035] Optionally, the step S14 specifically includes: based on the expanded sample set E, performing fine-tuning training on the trained semi-supervised momentum prototype network to optimize the mapping;

[0036] Suppose that in the nth round of training batches during fine-tuning training, the theoretical prototype calculated using pseudo-label samples and labeled samples is for:

[0037]

[0038] Suppose that in the nth round of training batches during fine-tuning training, the actual adopted prototype P obtained by momentum prototype calculation is n

[0039]

[0040] Among them, η is the prototype momentum factor; based on the above-mentioned prototype update algorithm based on momentum prototype, the prototype is corrected during fine-tuning training to extract a more accurate prototype; finally, the semi-supervised momentum prototype network and the corrected prototype after fine-tuning training are obtained.

[0041] Optionally, S2 is specifically:

[0042] S21: processing of data to be diagnosed and identified;

[0043] S22: Input the processed data to be diagnosed and identified into the semi-supervised momentum prototype network after fine-tuning training to carry out diagnosis and identification, and output the diagnosis results.

[0044] Optionally, S21 specifically includes: cutting the original vibration signal data to be diagnosed and identified into one-dimensional samples X of standard length.

[0045] Optionally, the S22 is specifically as follows: inputting the processed sample X to be diagnosed and identified into the semi-supervised momentum prototype network after fine-tuning training, and Embed the sample X into the feature space to obtain the sample feature F:

[0046]

[0047] Calculate the distance between the prototypes corrected by the sample feature F in the feature space, and use it as the diagnostic recognition result p of the sample after softmax:

[0048]

[0049] The beneficial effects of the present invention are as follows: the present invention combines pseudo-label learning in semi-supervised learning with semi-supervised momentum prototype network in small sample learning, and proposes a wind turbine gearbox fault diagnosis method based on semi-supervised momentum prototype network, which makes full use of the fault information contained in unlabeled samples and greatly reduces the overfitting phenomenon of the prototype network due to the scarcity of labeled samples; the present invention adopts a pseudo-label threshold selection method based on Monte Carlo uncertainty to measure the confidence of the pseudo-label, increases the accuracy of the pseudo-label, and avoids the use of erroneous pseudo-labels for noisy training; the present invention proposes a momentum prototype method, which reduces the prototype jump problem occurring in the prototype update through the momentum algorithm, and enhances the stability and accuracy of the semi-supervised momentum prototype network during fine-tuning.

[0050] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0052] Figure 1 A flowchart of a specific embodiment of the present invention;

[0053] Figure 2 This is a schematic diagram of the prototype network structure;

[0054] Figure 3 It is a structural diagram of the feature extractor;

[0055] Figure 4 This is a schematic diagram of the semi-supervised momentum prototype network structure proposed in the present invention;

[0056] Figure 5 A schematic diagram of a pseudo-label threshold selection strategy based on Monte Carlo uncertainty of the present invention;

[0057] Figure 6 This is a schematic diagram of the prototype jump phenomenon that occurs during the prototype update process;

[0058] Figure 7 This is a schematic diagram of vibration signals collected in the gearbox of a wind turbine generator set; Figure 7 (a) is the vibration signal H; Figure 7 (b) is the vibration signal BW; Figure 7 (c) is the vibration signal TP; Figure 7 (d) is the vibration signal BF; Figure 7(e) is the vibration signal CC1; Figure 7 (f) is the vibration signal CC2;

[0059] Figure 8 A schematic diagram showing the comparison of diagnostic accuracy of different fault diagnosis and identification methods;

[0060] Fig. 9 This is the vibration signal waveform diagram collected when the wind turbine gearbox is in a healthy state; Fig. 9 (a) is the vibration signal CNN; Fig. 9 (b) is the vibration signal SiaNet; Fig. 9 (c) is the vibration signal MatNet; Fig. 9 (d) is the vibration signal RelaNet; Fig. 9 (e) is the vibration signal ProNet; Fig. 9 (f) is the vibration signal PLL; Fig. 9 (g) is the vibration signal MSSSL; Fig. 9 (h) is the vibration signal RNSSL; Fig. 9 (i) is the vibration signal MPNSSL. DETAILED DESCRIPTION

[0061] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0062] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0063] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0064] like Figure 1 As shown, the present invention discloses a wind turbine gearbox fault diagnosis method based on a semi-supervised momentum prototype network, comprising the following steps:

[0065] S1, semi-supervised momentum prototype network training;

[0066] S11, training sample set preparation;

[0067] S12. Train the semi-supervised momentum prototype network based on a limited set of labeled samples;

[0068] S13. Based on the trained semi-supervised momentum prototype network, pseudo-label learning is performed on the unlabeled sample set;

[0069] S14. Fine-tune the semi-supervised momentum prototype network based on the expanded sample set;

[0070] S2, fault diagnosis and identification;

[0071] S21, processing of data to be diagnosed and identified;

[0072] S22, the data to be diagnosed and identified is input into the semi-supervised momentum prototype network, and the network outputs the result;

[0073] The semi-supervised momentum prototype network is a small sample semi-supervised learning model. The semi-supervised momentum prototype network mainly includes the basic semi-supervised momentum prototype network structure, threshold selection module and momentum prototype module;

[0074] The basic semi-supervised momentum prototype network conducts small sample learning based on a small number of labeled samples, aiming at pre-training for subsequent pseudo-label learning; its main structure includes feature extractor, prototype calculation layer, distance measurement module and output layer, such as Figure 2As shown in Figure 1, the feature extractor contains multiple convolutional layers, pooling layers, BN layers, activation layers, and fully connected layers. The convolutional layer extracts global features from the original input, the pooling layer compresses the extracted features, the BN layer normalizes the features, the activation layer enhances the model's ability to extract nonlinear features, and the fully connected layer maps the learned feature representation to the designed prototype feature space, as shown in Figure 1. Figure 3 As shown. Feature Extractor The sample x in the semi-supervised momentum prototype network training sample set L i Embed into a specific feature space:

[0075]

[0076] The prototype calculation layer constructs the prototype using the support set mapped into the feature space;

[0077]

[0078] The distance measurement layer is used to calculate the distance between the test sample x and each prototype in the feature space;

[0079]

[0080] The output layer converts the distance obtained by the distance measurement layer into a prediction vector through the softmax layer as the output of the network;

[0081]

[0082] The threshold selection module is used to select high-confidence predictions as pseudo-labels of unlabeled samples when the semi-supervised momentum prototype network performs pseudo-label learning, and then treats the unlabeled samples with pseudo-labels as labeled samples to expand the sample set; it mainly includes the confidence and threshold selection of uncertainty, such as Figure 4 As shown; Figure 5 A schematic diagram of a pseudo-label threshold selection strategy based on Monte Carlo uncertainty of the present invention; Figure 6 Schematic diagram of the prototype jump phenomenon that occurs during the prototype update process.

[0083] Confidence index extraction Carry out I forward propagation on the unlabeled samples in the pseudo-label learning sample set U to obtain I predictions, and calculate the mean μ and variance σ of the I prediction probability vectors 2 :

[0084]

[0085] Using the variance σ describing the degree of dispersion 2 Calculate the standard deviation φ of I forward passes:

[0086]

[0087] The information entropy ψ of the I-th forward propagation is calculated using the mean μ describing the central tendency:

[0088]

[0089] The Monte Carlo uncertainty θ of the semi-supervised momentum prototype network prediction is calculated using the standard deviation φ and information entropy ψ:

[0090]

[0091] Among them, λ is a hyperparameter that adjusts the weight between information entropy ψ and standard deviation φ.

[0092] Threshold selection is based on the set threshold θ′ for model prediction Make a selection:

[0093]

[0094] Through the threshold selection module, the semi-supervised momentum prototype network selects the predictions of unlabeled samples with high confidence to generate pseudo labels, and repeats the above steps to expand the scale of labeled samples to obtain the expanded sample set E.

[0095] Since samples with pseudo labels differ from labeled samples in feature distribution, when the semi-supervised momentum prototype network is fine-tuned based on the expanded sample set E, the prototypes calculated in different training batches have large fluctuations in feature space; even, because the network gives the wrong pseudo labels to the difficult samples during the sample set expansion stage, the prototypes calculated from the unlabeled samples with the wrong pseudo labels have a larger jump. To prevent the prototype jump phenomenon in the above prototype update, the momentum prototype uses the prototype of the current batch for fine-tuning, and at the same time accumulates the historical prototype changes to guide the generation of the current batch of prototypes, which greatly reduces the amplitude of the prototype jump.

[0096] Theoretical prototype calculated using pseudo-label samples and labeled samples in the nth training batch during fine-tuning training for:

[0097]

[0098] Suppose that in the nth round of training batches during fine-tuning training, the actual adopted prototype P obtained by momentum prototype calculation is n

[0099]

[0100] Where η is the prototype momentum factor. Based on the above prototype update algorithm based on momentum prototype, the prototype is corrected during fine-tuning training to extract a more accurate prototype. Finally, the semi-supervised momentum prototype network and the corrected prototype after fine-tuning training are obtained.

[0101] In order to verify the effectiveness of the fault diagnosis and identification method disclosed in the present invention, a comparative test was conducted using the vibration signal of a certain type of wind turbine gearbox in a certain wind farm; the main structure of a certain type of wind turbine gearbox includes a first-stage planetary reducer and a two-stage parallel shaft reducer, and a condition monitoring system (CMS) was used to collect wind turbine gearbox vibration data at a sampling frequency of 25600Hz. A total of 6 types of wind turbine gearbox vibration signals in different health states were collected. In order to obtain a sufficient number of samples, for each type, 600K (1K=1024) data before the equipment failure downtime point was taken according to the on-site fault inspection and maintenance report, and 600 samples were obtained by dividing each sample into 1024 samples, such as Figure 7 As shown in the figure, 30 samples are randomly selected as labeled training samples, 100 samples are selected as test samples, and the remaining samples are discarded with label information as unlabeled samples. Figure 7 (a) is the vibration signal H; Figure 7 (b) is the vibration signal BW; Figure 7 (c) is the vibration signal TP; Figure 7 (d) is the vibration signal BF; Figure 7 (e) is the vibration signal CC1; Figure 7 (f) is the vibration signal CC2; the specific description of the sample set is shown in Table 1.

[0102] Table 1 Description of wind turbine gearbox sample set

[0103]

[0104] Based on the above sample sets, the semi-supervised momentum prototype network proposed in the present invention is compared with nine other methods, including CNN, Siamese Network (SiaNet) described in the document 32nd International Conference Machine Learning, 2015. “Siamese Neural Networks for One-Shot Image Recognition.”, Matching Network (MatNet) described in the document NIPS., 2016. “Matching Networks for One Shot Learning.”, Relation Network (RelaNet) described in the document CVPR., 2018. “Learning to Compare: Relation Networkfor Few-Shot Learning.”, Prototypical Network (semi-supervised momentum prototype network) described in the document NIPS., 2017. “Prototypical Networks for Few-shot Learning.”, and ICML., 2013. “Pseudo-Label: The Simple and Efficient Semi-Supervised Learning Method for Deep Neural "A multi-stage semi-supervised learning approach for intelligent fault diagnosis of rolling bearing using dataaugmentation and metric learning." Multi-Stage Semi-Supervised Learning (MSSSL) described in IEEE Trans.Instru.Meas., Jan. 2021. "A relation-based semi-supervised method for box gear fault diagnosis with limited labeled samples." Relation Network based Semi-Supervised Learning (RNSSL).To ensure the fairness of the experiment, the above methods use the same feature extractor. The specific parameter settings of the feature extractor are shown in Table 2.

[0105] Table 2 Structural parameters of semi-supervised momentum prototype network feature extractor

[0106]

[0107] In order to reduce the impact of randomness of experimental data on experimental results, five random experiments were conducted. The experimental results are as follows: Figure 8 And as shown in Table 3.

[0108] Table 3 Experimental results

[0109]

[0110] As can be seen from Table 3, the classic convolutional neural network (CNN) has the worst effect. The main reason is that when the model has a small amount of training data, the real distribution of the data in the high-dimensional space cannot be fully reflected, resulting in serious overfitting. The small sample learning method simulates the ability of humans to acquire knowledge from a small number of examples through similarity calculation and category promotion. Compared with CNN, its diagnostic performance has been significantly improved. The optimal semi-supervised momentum prototype network in small sample learning has improved the average accuracy of CNN by about 14.72%. PLL and MSSSL in semi-supervised learning use the classic convolutional neural network to pre-train in a small number of samples and then give unlabeled data pseudo-labels for network retraining. They make full use of a large amount of unlabeled data for sample expansion to enhance the performance of the network. Compared with the small sample learning method that only uses a small number of samples, there is a significant improvement. Among them, PLL directly uses the prediction of the pre-trained model as a pseudo-label, which is worse than MSSSL based on cluster center membership prediction. Compared with PLL and MSSSL, RNSSL and MPNSSL in small sample semi-supervised learning use the relational network and semi-supervised momentum prototype network in the small sample learning method in the pre-training process. Compared with traditional supervised learning, it is more suitable for training in small sample situations, and the prediction model with better performance is used for pseudo-label estimation, which is improved by 13.01%. Compared with RNSSL, MPNSSL adopts a certain threshold selection strategy and momentum prototype in the pseudo-label learning stage to make the use of pseudo-labels more reasonable. Experimental results show that the semi-supervised momentum prototype network proposed in the present invention is significantly better than other methods.

[0111] In order to verify the advancedness of the semi-supervised momentum prototype network of the present invention, t-SNE is used to visualize the output layer of the feature extractor of each method. Fig. 9 As shown in the figure, after the original vibration signal is extracted by CNN, the feature points are randomly distributed in the two-dimensional visualization diagram. Fig. 9 (a) is the vibration signal CNN; Fig. 9(b) is the vibration signal SiaNet; Fig. 9 (c) is the vibration signal MatNet; Fig. 9 (d) is the vibration signal RelaNet; Fig. 9 (e) is the vibration signal ProNet; Fig. 9 (f) is the vibration signal PLL; Fig. 9 (g) is the vibration signal MSSSL; Fig. 9 (h) is the vibration signal RNSSL; Fig. 9 (i) is the vibration signal MPNSSL. Fig. 9 As shown in (a), the health status of the sample cannot be distinguished from the visualization. In contrast, the fault diagnosis method based on small sample learning has a better visualization effect, but Fig. 9 (b), Fig. 9 (c), Fig. 9 (d), Fig. 9 As shown in (e), there are still a considerable number of feature points that overlap, which ultimately leads to low fault recognition accuracy. Compared with small sample learning, pseudo-label learning in semi-supervised learning uses model prediction to give pseudo-labels to unlabeled data for training, making full use of available data. Fig. 9 (f), Fig. 9 As shown in (g), a better classification effect was achieved, but because the model was blindly confident in the wrong pseudo-labels, the model was trained with noise, and some feature points overlapped, resulting in low model fault recognition accuracy. In comparison, the small sample semi-supervised diagnosis method based on the semi-supervised momentum prototype network proposed in this paper has better feature extraction and fault identification capabilities, and has achieved good clustering performance. It can cluster the feature points of different fault types well and has obvious boundaries. The visualization results show that the semi-supervised momentum prototype network proposed in this invention has a good clustering effect on sample features;

[0112] Aiming at application scenarios where fault labeled data is scarce, the present invention proposes a wind turbine gearbox fault diagnosis method based on a semi-supervised momentum prototype network. Based on small sample learning, this method makes full use of unlabeled data to improve the robustness of the model, and obtains a classifier with better classification performance than that obtained by training only with labeled data.

[0113] Aiming at the clustering hypothesis in semi-supervised learning, a small sample semi-supervised learning method based on semi-supervised momentum prototype network is proposed. The unlabeled data is given more reasonable pseudo-labels according to the similarity between the sample and the prototype in the feature space.

[0114] Aiming at the pseudo-label confidence problem, an improved pseudo-label screening strategy is proposed. Compared with the traditional method of directly giving unlabeled data for prediction as pseudo-labels, this method only labels samples with higher confidence, thereby improving the accuracy of pseudo-labels and thus greatly improving the accuracy of the final fault classification.

[0115] Aiming at the prototype jumping problem existing in the semi-supervised momentum prototype network when introducing semi-supervision into small samples, a momentum prototype is proposed to reduce the oscillation amplitude in the prototype generation process.

[0116] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A wind turbine gearbox diagnosis method based on a semi-supervised momentum prototype network, characterized in that: The method comprises the following steps: S1: Semi-supervised momentum prototype network training; S2: Fault diagnosis and identification; The S1 is specifically: S11: Sample set production; S12: Based on a limited set of labeled samples, a semi-supervised momentum prototype network is trained to learn feature maps and prototypes. S13: Based on the trained semi-supervised momentum prototype network, pseudo-label learning is performed on the unlabeled sample set, pseudo-label information after threshold selection is given to the unlabeled samples, and the scale of the labeled sample set is continuously expanded to obtain the expanded sample set E; S14: Based on the expanded sample set E, fine-tune the semi-supervised momentum prototype network, optimize the feature mapping of the semi-supervised momentum prototype network, and modify the prototype based on the momentum prototype; The S11 is specifically as follows: normalizing the original vibration signal sample set containing K types of fault categories into standardized one-dimensional samples, and dividing a small number of labeled samples into a support set S and a query set Q, as a semi-supervised momentum prototype network training sample set L for S12; making the unlabeled samples into a pseudo-label learning sample set U for S13; The S12 specifically includes: training the semi-supervised momentum prototype network based on the semi-supervised momentum prototype network training sample set L: Semi-supervised Momentum Prototype Network Utilizing Feature Extractor The sample x in the semi-supervised momentum prototype network training sample set L i Embed into a specific feature space: For the cth type of fault, the prototype P is generated using the support set S c : Calculate the distance between the sample in the query set Q and each prototype in the feature space, and use the calculated distance as the predicted output p of the sample after softmax: The semi-supervised momentum prototype network is trained by back propagation through the loss function loss: loss = -lg p(y = k | x) Through the above process, the semi-supervised momentum prototype network continuously optimizes the feature mapping and prototype of the semi-supervised momentum prototype network; The S13 is specifically as follows: based on the trained semi-supervised momentum prototype network, pseudo-label learning is performed on the pseudo-label learning sample set U, MC-Dropout is used to measure the uncertainty of the semi-supervised momentum prototype network prediction, and this is used as a judgment indicator for threshold selection, giving the unlabeled samples high-confidence pseudo-labels, and continuously expanding the scale of the labeled sample set; Use the semi-supervised momentum prototype network after adding MC-Dropout to perform I forward propagations on the unlabeled samples in the pseudo-label learning sample set U to obtain I predictions, and calculate the mean μ and variance σ of the I prediction probability vectors 2 : Using the variance σ describing the degree of dispersion 2 Calculate the standard deviation φ of I forward passes: The information entropy ψ of the I-th forward propagation is calculated using the mean μ describing the central tendency: The Monte Carlo uncertainty θ of the semi-supervised momentum prototype network prediction is calculated using the standard deviation φ and information entropy ψ: Among them, λ is a hyperparameter that adjusts the weight between information entropy ψ and standard deviation φ; based on the calculated Monte Carlo uncertainty θ predicted by the semi-supervised momentum prototype network, the model output of the unlabeled sample with high confidence is selected to generate pseudo labels after the set threshold θ′, and the scale of the labeled samples is expanded to obtain the expanded sample set E; The S14 specifically includes: based on the expanded sample set E, fine-tuning the trained semi-supervised momentum prototype network to optimize the mapping; Suppose that in the nth round of training batch in fine-tuning training, the theoretical prototype calculated using pseudo-label samples and labeled samples is for: Suppose that in the nth round of training batches during fine-tuning training, the actual adopted prototype P obtained by momentum prototype calculation is n Wherein, η is the prototype momentum factor; based on the above-mentioned prototype update algorithm based on momentum prototype, the prototype is corrected during fine-tuning training to extract a more accurate prototype; finally, the semi-supervised momentum prototype network and the corrected prototype after fine-tuning training are obtained; The S2 is specifically: S21: processing of data to be diagnosed and identified; S22: Input the processed data to be diagnosed and identified into the semi-supervised momentum prototype network after fine-tuning training to carry out diagnosis and identification, and output the diagnosis result; The S21 specifically includes: cutting the original vibration signal data to be diagnosed and identified into a one-dimensional sample X of standard length; S22 specifically includes: inputting the processed sample X to be diagnosed and identified into the semi-supervised momentum prototype network after fine-tuning training, and Embed the sample X into the feature space to obtain the sample feature F: Calculate the distance between the prototypes corrected by the sample feature F in the feature space, and use it as the diagnostic recognition result p of the sample after softmax:

Images