A method for predicting the residual life of key components of a wind turbine under multiple working conditions

Abstract

The application relates to the technical field of wind power generation equipment predictive maintenance, in particular to a wind power generation key component multi-working condition residual life prediction method S and a target domain sample set D T ; a feature extractor module is constructed; a multi-feature adaptive module is constructed; a migratable attention mechanism module is constructed; a residual service life predictor module is constructed; a domain adaptation module is constructed; a residual service life prediction model is trained; residual service life of a wind power generation key component under a target domain working condition is predicted, target domain test sample sets are input into the established network model, and residual service life prediction results of the key component at each time are output. The wind power generation key component multi-working condition residual life prediction method proposes a migratable attention mechanism module, the module can dynamically activate degradation features with high migration in model training, so that the generalization capability of the model is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of wind power generation equipment predictive maintenance, in particular to a wind turbine key component multi-working condition residual life prediction method. BACKGROUND

[0002] Wind energy is a kind of renewable energy with great potential, which has attracted widespread attention from countries around the world. Wind turbine is the core component of wind power generation system, which realizes the conversion of mechanical energy to electrical energy. Wind turbines are mostly installed in remote and harsh environments such as offshore, plateau and mountainous areas, and the operating conditions are extremely harsh, which causes the performance of wind turbine components to easily degrade. At present, with the continuous increase of installed capacity of wind power generation, the loss caused by wind turbine bearing failure is also increasing. Ensuring the stable, safe and reliable operation of wind turbines has become the primary problem faced by the wind power industry. Residual service life prediction, also known as residual service life prediction, refers to the time that the equipment can ensure the safe and reliable operation of the machine under the specified operating conditions. Timely understanding of the residual service life of key components of equipment is an important method to reduce production losses and can save maintenance costs throughout the life cycle. Therefore, predicting the residual service life of key components of wind turbines will ensure the safe and reliable operation of the unit, reduce downtime caused by serious accidents, and thus improve the economy of the unit operation. With the rapid development of global wind power industry, the residual service life prediction technology of key components of wind turbines will become more and more important.

[0003] From the current published data, there are related researches that realize the residual service life prediction of key components of wind turbines through machine learning or deep learning methods, such as the construction method of wind turbine bearing residual service life prediction model (patent number: CN202010416839.9), wind turbine residual life prediction method based on multi-channel separable residual neural network (patent number: CN202210031823.5), etc. The existing technical solutions still have the following shortcomings:

[0004] 1. The existing method mostly uses time domain analysis and frequency domain analysis to extract the characteristic parameters of the signal, and uses fitting means for life regression analysis. However, the correlation between the extracted features and the bearing life is not strong, and the prediction performance is limited;

[0005] 2. The existing method uses deep learning to adaptively mine hidden degradation features from vibration signals. However, the working conditions of wind turbines are variable, and the difference in distribution caused by the variable working conditions will affect the generalization performance of the model. SUMMARY

[0006] The purpose of the present application is to overcome the defects of the prior art, provide a wind turbine key component multi-working condition residual life prediction method, aiming to improve the prediction accuracy and generalization ability of the wind turbine key component residual life prediction model under multi-working condition.

[0007] To achieve the above purpose, the present application provides the following technical scheme: a wind turbine key component multi-working condition residual life prediction method, comprising the following steps:

[0008] S1: constructing a source domain sample set D S and a target domain sample set D T :

[0009] Select the data of the wind turbine key component under two different working conditions from the historical database as the training set and the test set. The training set data is composed of labeled source domain data D S and unlabeled target domain data D T . The test set data is composed of unlabeled target domain data;

[0010] S2: constructing a feature extractor module M F ;

[0011] S3: constructing a multi-representation adaptive module M MR ;

[0012] S4: constructing a transferable attention mechanism module;

[0013] S5: constructing a residual service life predictor module M RP ;

[0014] S6: constructing a domain adaptation module M DA ;

[0015] Calculate the conditional distribution difference loss of the weighted deep degradation feature AF extracted in S4, and obtain the conditional distribution difference loss L DA ;

[0016] S7: training the residual service life prediction model;

[0017] S8: residual service life prediction of wind turbine key components under target domain working condition, input the target domain test sample set into the established network model, and output the residual service life prediction result of the key components at each time.

[0018] Preferably, the feature extractor module in S2 is constructed, specifically comprising the following steps:

[0019] (2a) build a convolutional neural network stacked and connected by four structure same sub-modules as a feature extractor module M FEach sub-module is composed of a separate convolutional layer, a batch normalization layer, a nonlinear activation function layer, and a pooling layer.

[0020] (2b) input the source domain sample set D S and the target domain sample set D T into the feature extractor module M F to extract the deep degradation feature F.

[0021] Preferably, the S3 constructs a multi-representation adaptive module, specifically including the following steps:

[0022] (3a) build an InceptionA network with four different scales connected in parallel as the multi-representation adaptive module M MR ;

[0023] (3b) input the deep degradation feature F into the multi-representation adaptive module M MR to obtain four deep degradation features F s1 , F s2 , F s3 and F s4 of different scales.

[0024] Preferably, the S4 constructs a transferable attention mechanism module, specifically including the following steps:

[0025] (4a) build four domain classifiers composed of four fully connected networks and Softmax activation function layers;

[0026] (4b) input the scale i deep degradation feature F si extracted in S3 into the domain classifier corresponding to this scale to obtain the domain classification probability P d and the domain classification label l d ;

[0027] (4c) perform gradient flipping on the scale i deep degradation feature F si extracted in S3 and calculate the entropy value H using the following formula:

[0028]

[0029] (4d) convert the entropy value H into the attention weight W si using the following formula:

[0030]

[0031] (4e) multiply the deep degradation feature F si and the attention weight W si corresponding to this deep degradation feature F si +1 to obtain the weighted deep degradation feature AF si;

[0032] (4f) The domain classification label l d and the domain actual class label are used to calculate the transferable attention loss L A .

[0033] Preferably, the remaining useful life predictor module is constructed in S5, specifically comprising the following steps:

[0034] (5a) A remaining useful life predictor composed of a one-layer fully connected network is built;

[0035] (5b) The weighted deep degradation features AF of the source domain extracted in S4 are input into the remaining useful life predictor to obtain the predicted remaining useful life prediction value Y S of the source domain;

[0036] (5c) The predicted remaining useful life prediction value Y S of the source domain and the actual remaining useful life prediction value Y S R are compared using mean square error to obtain the prediction loss L R .

[0037] Preferably, the remaining useful life prediction model is trained in S7, specifically comprising the following steps:

[0038] (7a) Set the learning rate and the number of model iterations;

[0039] (7b) The condition monitoring signals of the source domain sample set D S and the target domain sample set D T are sequentially input into the feature extractor module M F to obtain the deep source domain features and target domain features F, which are then sequentially input into the feature extractor module M F , the multi-representation adaptive module M MR , the transferable attention mechanism module, the remaining useful life predictor module M RP and the domain adaptation module M DA , respectively, to calculate the prediction loss L R , the conditional distribution difference loss L DA and the transferable attention loss L A ;

[0040] (7c) According to the prediction loss L R , the conditional distribution difference loss L DA and the transferable attention loss L A obtained in step 6, the total loss function L is calculated as follows:

[0041]

[0042] wherein λ A is the transferable attention loss L A is the penalty coefficient λ DA is the conditional distribution difference loss L DA is the penalty coefficient;

[0043] (7d) iterates the parameters of the feature extractor module M F , the multi-representation adaptive module M MR , the transferable attention mechanism module, and the remaining useful life predictor module M RP in turn using the Adam optimization algorithm until the maximum number of iterations is reached, obtaining the trained remaining useful life prediction model.

[0044] Compared with the prior art, the beneficial effects of the present application are:

[0045] 1. The wind turbine key component multi-condition remaining life prediction method proposes a multi-representation transferable attention network for predicting the remaining useful life of the key components of the wind turbine under different conditions. The model can extract and align the distribution of degradation features extracted from different representation structures, thereby enriching high-order degradation information.

[0046] 2. The wind turbine key component multi-condition remaining life prediction method proposes a transferable attention mechanism module. This module can dynamically activate degradation features with high transferability during model training, thereby improving the generalization ability of the model. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0048] Fig. 1 is the flowchart for the implementation of the present application;

[0049] Fig. 2 is the multi-representation adaptive module MMR diagram of the present application. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0051] In the present application, unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connecting", "fixing" and the like should be understood in a broad sense, for example, can be fixed connection, can also be detachable connection, or integral; can be mechanical connection, can also be electrical connection; can be directly connected, can also be indirectly connected through an intermediate medium, can be the internal communication of two elements or the interaction relationship between two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0052] Please refer to Figs. 1-2 The present application provides a technical solution: a wind turbine key component multi-working condition residual life prediction method, comprising the following steps:

[0053] S1: Constructing source domain sample set D S and target domain sample set D T :

[0054] Selecting the data of the wind turbine key components under two different working conditions from the historical database as the training set and the test set. The training set data is composed of labeled source domain data D S and unlabeled target domain data D T . The test set data is composed of unlabeled target domain data;

[0055] S2: Constructing feature extractor module M F ;

[0056] S3: Constructing multi-representation adaptive module M MR ;

[0057] S4: Constructing a transferable attention mechanism module;

[0058] S5: Constructing residual service life predictor module M RP ;

[0059] S6: Constructing domain adaptation module M DA ;

[0060] Calculating the conditional distribution difference loss of the weighted deep degradation features AF extracted in S4 to obtain the conditional distribution difference loss L DA ;

[0061] S7: Training the residual service life prediction model;

[0062] S8: Residual service life prediction of wind turbine key components under target domain working condition, input the target domain test sample set into the established network model, and output the residual service life prediction results of the key components at each time.

[0063] The technical idea for achieving the object of the present application is that, first, state monitoring data of key components of a wind turbine under two different working conditions are collected to obtain a source domain sample set D S and a target domain sample set D T ; then, the source domain sample set D S and the target domain sample set D T are input into a feature extractor module M F to extract a deep degradation feature F; secondly, the deep degradation feature F is input into a multi-feature adaptive module M MR to obtain four deep degradation features F s1 , F s2 , F s3 and F s4 of different scales. Each deep degradation feature F si is input into a migratability attention mechanism module corresponding to the degradation feature to obtain an attention weight W si and a migratable attention loss L A of the deep degradation feature. The deep degradation feature F si is multiplied by the attention weight W si +1 corresponding to the deep degradation feature F si to obtain a weighted deep degradation feature AF si . The weighted deep degradation feature AF si is input into a remaining useful life predictor module M RP and a domain adaptation module M DA to obtain a predicted remaining useful life prediction value Y and a conditional distribution difference loss L DA , respectively. The predicted remaining useful life prediction value Y S in the source domain is compared with a real remaining useful life prediction value Y S R to obtain a prediction loss L R . The prediction loss L R , the migratable attention loss L A and the conditional distribution difference loss L DA are added to obtain a total loss. Finally, the Adam optimization algorithm is used to iteratively update the module parameters in sequence to minimize the total loss, complete the training of the remaining useful life prediction model, and input the target domain test sample set into the established network model to obtain the final remaining useful life prediction result.

[0064] Step 1, obtaining a source domain sample set and a target domain sample set.

[0065] The embodiment adopts a laboratory bearing full-cycle degradation dataset. Bearing state monitoring data is collected by two acceleration sensors fixed on the bearing outer ring, and the sampling frequency is 25.6 kHz. Two different working condition datasets are used in the case study, as shown in Table 1. Different working conditions are set as source domains in turn, and the other working condition is set as the target domain. Two transfer tasks are set respectively, as shown in Table 2.

[0066] Table 1

[0067]

[0068] Table 2

[0069]

[0070] Step 2, constructing a feature extractor module M F :

[0071] Firstly, a four-layer convolutional neural network stacked by four sub-modules with the same structure is built as the feature extractor module M F The structure of each sub-module is composed of a separation convolution layer, a batch normalization layer, a nonlinear activation function layer and a pooling layer. The number of convolution kernels of each layer is [4, 8, 16, 32], the kernel size is set to 3x1, and the pooling layer size is set to 3x1.

[0072] Secondly, the source domain sample set D S and the target domain sample set D T are input into the feature extractor module M F to extract deep degradation features F;

[0073] Step 3, constructing a multi-representation adaptive module M MR :

[0074] Firstly, an InceptionA network connected in parallel by four different scales is built as the multi-representation adaptive module M MR , as shown in Fig. 2 .

[0075] Then, the deep degradation features F are input into the multi-representation adaptive module M MR to obtain four deep degradation features F s1 , F s2 , F s3 and F s4 of different scales.

[0076] Step 4, constructing a transferable attention mechanism module:

[0077] First, four domain classifiers are built, each consisting of a 4-layer fully connected network and a Softmax activation function layer. The nodes of the fully connected layers are set to 288-80-20-2.

[0078] Then, the scale i depth degradation feature F extracted in step 3 si Input the domain classifier corresponding to this scale to obtain the domain classification probability P. d and domain category tags l d ;

[0079] Secondly, the scale i depth degradation features F extracted in step 3 are... si Gradient flipping is performed, and the entropy value H is used for calculation, as follows:

[0080]

[0081] Next, the entropy value H is converted into attention weights W. si The calculation formula is as follows:

[0082]

[0083] Finally, the deep degradation feature F si and the deep degradation feature F si The corresponding attention weight W si Multiplying by 1 yields the weighted depth degradation feature (AF). si ;

[0084] Simultaneously, through negative log-likelihood loss, the domain classification label l d The transferable attention loss L is calculated based on the actual category label of the domain. A ;

[0085] Step 5, construct the remaining useful life predictor module M RP :

[0086] First, a remaining lifetime predictor consisting of a single-layer fully connected network is built, with the nodes of the fully connected layer set to 288-1.

[0087] Then, the weighted depth degradation feature AF of the source domain extracted in step (4) is input into the remaining lifetime predictor to obtain the predicted remaining lifetime value Y of the source domain. S ;

[0088] Secondly, the mean square error is used to predict the remaining useful life Y in the source domain. S And the actual remaining useful life prediction value Y S R By comparison, the predicted loss L is obtained. R ;

[0089] Step 6, constructing the domain adaptation module M DA , the conditional distribution difference loss L of the weighted depth degradation feature AF extracted in step (4) is calculated, and the conditional distribution difference loss L is obtained DA .

[0090] Step 7, training the remaining useful life prediction model:

[0091] First, set the learning rate to 0.0005 and the model iteration number to 500;

[0092] Then, the state monitoring signals of the source domain sample set D S and the target domain sample set D T are input into the feature extractor module M F to obtain the deep source domain feature and the target domain feature F, and then input into the feature extractor module M F , the multi-representation adaptive module M MR , the transferable attention mechanism module, the remaining useful life predictor module M RP and the domain adaptation module M DA respectively, to calculate the prediction loss L R , the conditional distribution difference loss L DA and the transferable attention loss L A ;

[0093] Secondly, according to the prediction loss L R , the conditional distribution difference loss L DA and the transferable attention loss L A obtained in step 6, the total loss function L is calculated, and the calculation method is as follows:

[0094]

[0095] Wherein, λ A is the transferable attention loss L A penalty coefficient, which is set to 0.01 in the embodiment, and λ DA is the conditional distribution difference loss L DA penalty coefficient, which is set to 10 in the embodiment;

[0096] Finally, the parameters of the feature extractor module M F , the multi-representation adaptive module M MR , the transferable attention mechanism module and the remaining useful life predictor module M RP are iterated in turn by using the Adam optimization algorithm until the maximum iteration number is reached, and the trained remaining useful life prediction model is obtained.

[0097] Step 8, the remaining useful life of the key components of the wind turbine under the target domain working condition is predicted, and the target domain test sample set is input into the established network model, and the remaining useful life prediction results of the key components at each time are output.

[0098] The effect of the present application is further described below in combination with simulation experiments:

[0099] 1. Simulation experiment conditions:

[0100] The hardware platform of the simulation experiment of the present application is: the central processing unit is Intel(R) Core(TM) i5-7500 CPU, the main frequency is 3.40 GHZ, and the memory is 16G.

[0101] The software platform of the simulation experiment of the present application is: WINDOWS 7 operating system and Python 3.7.

[0102] 2. Simulation content and result analysis:

[0103] The simulation experiment of the present application is respectively using the method of the present application and four existing technologies, including the remaining useful life prediction method based on non-adversarial transfer learning and based on adversarial transfer learning, including using conditional adaptive network (CAN) based on conditional maximum mean difference, transferable convolutional neural network (TCNN) based on multi-core maximum mean difference, conditional domain adaptive network (CDAN)

[40] and joint domain adaptive network (JAN) based on joint maximum mean difference. In addition, the mean absolute error MAE, the root mean square error RMSE, and the R2 index are used as evaluation indexes, and the comparison results are shown in Table 3:

[0104] Table 3

[0105]

[0106] It can be found from Table 4 that, according to the MAE, RMSE, and R2 index results, the remaining useful life prediction model of the present application has the optimal index performance for predicting the remaining useful life of the bearing of the key components of the wind turbine. And compared with other methods, the model of the present application has greatly improved in the index, indicating that the present application has good performance in predicting the remaining useful life of the key components of the wind turbine under the multi-working condition scene, and has high accuracy and generalization ability.

[0107] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and implementations, it is to be understood that the terminology used is for the purpose of descriptive clarity and that it should be taken in a descriptive sense and not a limiting sense.

[0108] While the embodiments of the application have been shown and described herein, it is to be understood that the application is not limited to these embodiments. Rather, many modifications, changes, substitutions, and alterations can be made to the embodiments of the application without departing from the spirit and scope of the application as defined by the appended claims and their equivalents.

Claims

1. A method for predicting the remaining life of a key component of a wind turbine under multiple working conditions, characterized in that: The method comprises the following steps: S1 : Constructing source domain sample set D S and target domain sample set D T : The data of key components of a wind turbine under two different working conditions are selected from a historical database as a training set and a test set; the training set data are composed of labeled source domain data D S and unlabeled target domain data D T ; the test set data are composed of unlabeled target domain data; S2: constructing a feature extractor module M F ; S3: Constructing multi-characteristic adaptive module M MR ; S4: constructing a migratable attention mechanism module; S5: Constructing the remaining useful lifetime predictor module M RP ; S6: Constructing a domain adaptation module M DA ; A conditional distribution difference loss of the weighted depth degradation feature AF extracted in S4 is calculated to obtain a conditional distribution difference loss L DA ; S7: training the remaining useful life prediction model; S8: performing remaining useful life prediction on the key components of the wind turbine under the target domain working condition, inputting the target domain test sample set into the established network model, and outputting the remaining useful life prediction results of the key components at each time; The S3 comprises the following steps of constructing a multi-representation adaptive module: (3a) building a InceptionA network with four different scales connected in parallel as the multi-representation adaptive module M MR ; (3b) inputting the depth-degraded feature F into the multi-representation adaptive module M MR , to obtain four depth-degraded features F s1 , F s2 , F s3 and F s4 of different scales; The S4 comprises the following steps of constructing a migratable attention mechanism module: (4a) building four domain classifiers composed of four layers of full connection networks and Softmax activation function layers; (4b) Scale i depth degradation feature F extracted in S3 si Input the domain classifier corresponding to the scale to obtain the domain classification probability P d And domain classification label l d ; (4c) scale i depth degradation features F extracted in S3 si Gradient reversal is performed and entropy H is calculated using the following formula: (4d) converting the entropy value H into an attention weight W si The calculation formula is as follows: (4e) multiplying the depth degradation feature F si and the depth degradation feature F si corresponding attention weight W si +1 to obtain a weighted depth degradation feature AF si ; (4f) The domain classification label l d and the domain actual class label to compute the transferable attention loss L A .

2. The method of claim 1, wherein the method further comprises: The S2 comprises the following steps of constructing a feature extractor module: (2a) building a convolutional neural network stacked by four sub-modules with the same structure as a feature extractor module M F Each sub-module is composed of a separate convolutional layer, a batch normalization layer, a nonlinear activation function layer, and a pooling layer. (2b) a source domain sample set D S and a target domain sample set D T an input feature extractor module M F extracting deep degradation features F.

3. The method of claim 2, wherein the method further comprises: The S5 comprises the following steps of constructing a remaining useful life predictor module: (5a) building one remaining useful life predictor composed of one layer of full connection network; (5b) input the weighted depth degradation features AF of the source domain extracted in S4 into the remaining useful life predictor to obtain a predicted source domain remaining useful life prediction value Y S ; (5c) Using the mean squared error to compare the predicted remaining useful life prediction Y S and the true remaining useful life prediction Y S R to obtain a prediction loss L R .

4. The method of claim 3, wherein the method further comprises: The S7 comprises the following steps of training the remaining useful life prediction model: (7a) setting the learning rate and the number of model iterations; (7b) sequentially input the state monitoring signals of the source domain sample set D S and the target domain sample set D T to the feature extractor module M F to obtain the deep source domain features and target domain features F, and then sequentially input the deep source domain features and target domain features F to the feature extractor module M F , the multi-representation adaptive module M MR , the migratable attention mechanism module, the remaining useful life predictor module M RP and the domain adaptation module M DA , respectively, to calculate the prediction loss L R , the conditional distribution difference loss L DA and the migratable attention loss L A ; (7c) the prediction loss L obtained according to step 6 R , the conditional distribution difference loss L DA , and the transferable attention loss L A , the total loss function L is calculated as follows: Where, λ A It is a transferable attention loss L A Penalty coefficient, λ DA It is the conditional distribution difference loss L DA Penalty coefficient; (7d) iteratively using Adam optimization algorithm to the parameters of the feature extractor module M F , multi-representation adaptive module M MR , transferable attention mechanism module, remaining useful life predictor module M RP until the maximum number of iterations is reached, obtaining a trained remaining useful life prediction model model.

Images