Flow field prediction method based on active transfer learning and incomplete supervision

Abstract

The invention provides a flow field prediction method based on active transfer learning and incomplete supervision, and belongs to the technical field of wing flow field prediction, and the method comprises the steps: training an initial transfer model through a first label data set, obtaining the control parameters of the transfer model, and carrying out the initialization setting of the control parameters of an initial student model and a teacher model; performing preheating training on the initial student model through the second label data set to obtain a pre-trained student model; performing active annotation on part of unlabeled data in the unlabeled data set to obtain sparse labeled physical field data; obtaining label-free sub-data in the label-free data set, and obtaining corresponding pseudo-physical field data through a teacher model; and training the pre-trained student model according to the real physical field data set and the predicted physical field data set, and obtaining the student model for predicting the physical data of the flow field around the wing. According to the method, the problem that complete label sample data is inconvenient to collect and process is solved, and the model training cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of wing flow field prediction, and in particular to a flow field prediction method based on active transfer learning and incomplete supervision. Background Art

[0002] The rich flow structure information contained in the flow field around an airfoil enables comprehensive evaluation of an airfoil's aerodynamic performance. Obtaining detailed flow fields under various freestream conditions helps improve airfoil design, optimization, and manufacturing. Deep learning methods have been widely used in fluid mechanics due to their powerful nonlinear fitting capabilities, feature representation capabilities, and rapid response capabilities.

[0003] Existing methods typically use supervised learning training to build deep learning models. Supervised learning typically requires a large number of fully labeled data samples. Fully labeled data samples are sample pairs consisting of different free flow conditions and corresponding complete flow fields. Collecting fully labeled data samples requires deploying a large number of sensors for monitoring, and must be obtained through physical experiments or numerical calculation methods. This process not only increases the difficulty of obtaining labeled data, but also incurs expensive labeling costs. In addition, due to the limitations of observation methods, some scenarios only allow sensors to be placed in a limited area. Therefore, only local flow field information can be collected, making it difficult to obtain a large number of fully labeled data samples for supervised training of deep learning models.

[0004] Therefore, a flow field prediction method based on active transfer learning and incomplete supervision is proposed. Summary of the Invention

[0005] In order to solve some or all of the technical problems existing in the above-mentioned prior art, the present invention provides a flow field prediction method based on active transfer learning and incomplete supervision, which is used to solve the problem that traditional technology is inconvenient to collect complete label data samples required for model training and has high acquisition cost.

[0006] The technical solutions of the present invention are as follows:

[0007] A flow field prediction method based on active transfer learning and incomplete supervision is provided, including:

[0008] Constructing the initial transfer model, initial student model and teacher model;

[0009] Obtain a first labeled dataset, a second labeled dataset, and an unlabeled dataset;

[0010] The initial migration model is trained using the first labeled dataset, and the control parameters of the initial migration model are optimized based on the deviation of the migration loss function analysis until the model converges, thereby obtaining the control parameters of the migration model.

[0011] Initializing the control parameters of the initial student model and the teacher model according to the migration model control parameters;

[0012] Preheat training the initial student model using a second label data set to obtain a pretrained student model;

[0013] Actively annotating part of the unlabeled data in the unlabeled dataset through a pre-trained student model based on an uncertainty query strategy to obtain corresponding sparsely annotated physical field data;

[0014] Obtain unlabeled sub-data in the unlabeled dataset that lacks corresponding sparsely annotated physical field data, and add corresponding pseudo physical field data to the unlabeled sub-data using the teacher model;

[0015] Constructing a real physical field dataset according to the second real physical field data, the sparsely annotated physical field data, and the pseudo physical field data corresponding to the second labeled data in the second labeled dataset;

[0016] The pre-trained student model performs prediction based on the second labeled data in the second labeled data set, the unlabeled data and the unlabeled sub-data in the unlabeled data set to obtain a predicted physical field data set, and trains the pre-trained student model by analyzing the deviation between the real physical field data set and the predicted physical field data set through a comprehensive loss function until the pre-trained student model converges to obtain a student model;

[0017] The obtained student model is used to predict the physical data of the flow field around the wing.

[0018] Furthermore, in the flow field prediction method based on active transfer learning and incomplete supervision provided by the present invention, the step of training the initial transfer model using the first labeled data set and optimizing the control parameters of the initial transfer model based on the deviation of the transfer loss function analysis until the model converges to obtain the transfer model control parameters includes:

[0019] Obtaining labeled data from the first labeled data set and splicing the quantile level based on the channel dimension, inputting the data into the initial migration model for processing, and obtaining predicted labeled physical field data;

[0020] Analyzing migration deviation data between the predicted label physical field data and the real label physical field data corresponding to the label data by using a migration loss function;

[0021]

[0022] Among them, L sqr is the migration loss function, and the first label dataset is is the i-th labeled data in the first labeled data set, For the initial migration model based on labeled data The quantile level during the ep-th training, For the initial migration model, the control parameter θ is based on the migration model according to the labeled data and quantile levels Output predicted label physical field data, There is labeled data in the first labeled dataset The corresponding real label physical field data, l τ Quantile loss function, τ r The quantile level set when training the initial migration model obeys the uniform distribution U[0,1], and E is the quantile loss function l τ Expected values obtained from conducting deviation assessments;

[0023] The control parameters of the initial migration model are optimized by minimizing the migration deviation data until a preset maximum number of iterations is reached or the model converges, thereby obtaining the migration model control parameters.

[0024] Furthermore, in the flow field prediction method based on active transfer learning and incomplete supervision provided by the present invention, the step of actively annotating part of the unlabeled data in the unlabeled dataset through a pre-trained student model based on an uncertainty query strategy to obtain corresponding sparsely annotated physical field data includes:

[0025] Analyzing the prediction uncertainty of the pre-trained student model based on the sample data;

[0026]

[0027] Where u(x) is the pre-trained student model f s The uncertainty of the prediction of sample data x, α is the significance level;

[0028] Annotating the unlabeled data at high uncertainty positions in the unlabeled dataset based on the prediction uncertainty to obtain sparsely annotated physical field data;

[0029]

[0030] Where M is a binary mask, To identify the unlabeled data in the unlabeled dataset Uncertainty prediction, and sort all unlabeled data in the unlabeled dataset in descending order of uncertainty, select the first m% of unlabeled data in the sequence for sparse physical annotation, and Q(m%) is the m% quantile of the uncertainty function u.

[0031] Furthermore, in the flow field prediction method based on active transfer learning and incomplete supervision provided by the present invention, the step of obtaining unlabeled sub-data in the unlabeled dataset that lacks corresponding sparsely annotated physical field data, and adding corresponding pseudo-physical field data to the unlabeled sub-data through the teacher model includes:

[0032] Obtaining unlabeled sub-data in the unlabeled data set, inputting the data into the teacher model for processing, and outputting corresponding pseudo physical field data;

[0033]

[0034] in, is pseudo-physical field data, For the teacher model f s Based on the teacher model control parameter θ t For unlabeled sub-data With a fixed quantile level τ c Concatenate the obtained data values.

[0035] Furthermore, in the flow field prediction method based on active transfer learning and incomplete supervision provided by the present invention, the steps of: the pre-trained student model performs prediction based on the second labeled data in the second labeled data set, the unlabeled data and the unlabeled sub-data in the unlabeled data set, obtaining a predicted physical field data set, and analyzing the deviation between the real physical field data set and the predicted physical field data set through a comprehensive loss function to train the pre-trained student model until the pre-trained student model converges to obtain a student model; including:

[0036] Obtaining second labeled data in the second labeled data set and unlabeled data in the unlabeled data set and inputting them into the pre-trained student model to obtain corresponding second predicted physical field data and sparse predicted physical field data;

[0037] Analyze the deviation between the second predicted physical field data and the second real physical field data, and between the sparse predicted physical field data and the sparse labeled physical field data through a supervised loss function to obtain supervised deviation data;

[0038]

[0039] Among them, L sup is the supervision loss function, D t is a supervised dataset constructed based on the second labeled dataset, the unlabeled data in the unlabeled dataset, and the corresponding sparsely annotated physical field data, x t is the second labeled data and unlabeled data in the supervised dataset, Ω is the area where the physical place is located, For the second real physical field data and sparsely labeled physical field data in the supervised dataset, The second predicted physical field data and sparse predicted physical field data output by the pre-trained student model;

[0040] Obtaining unlabeled sub-data in the unlabeled data set, inputting the data into the pre-trained student model, and obtaining corresponding predicted physical field data;

[0041] Analyze the deviation between pseudo physical field data and predicted physical field data through a pseudo loss function to obtain pseudo deviation data;

[0042]

[0043] Among them, L cr is a pseudo loss function, To use the unlabeled sub-data in the unlabeled dataset and the corresponding pseudo-physical field data The pseudo dataset constructed, M is the binary mask, ⊙ is the Hadamard product, Predicted physics data output by the pre-trained student model;

[0044] Construct a comprehensive loss function based on the supervised loss function and the pseudo loss function;

[0045] L=L sup +λL cr

[0046] Among them, L is the comprehensive loss function, λ is the weight value;

[0047] The pre-trained student model is trained by analyzing the deviation between the real physical field data set and the predicted physical field data set through a comprehensive loss function to obtain a student model.

[0048] Furthermore, in the flow field prediction method based on active transfer learning and incomplete supervision provided by the present invention, the step of obtaining the second labeled data in the second labeled data set and the unlabeled data in the unlabeled data set and inputting them into the pre-trained student model to obtain the corresponding second predicted physical field data and sparse predicted physical field data includes:

[0049] Obtaining second labeled data and unlabeled data in the supervised data set, inputting them into the pre-trained student model for processing, and outputting second predicted physical field data and sparse predicted physical field data;

[0050]

[0051] in, For the second predicted physical field data and sparse predicted physical field data, f s (xt ,τ s θ s ) is the student model f s Based on the student model control parameter θ s For the second labeled data and unlabeled data x t with random quantile level τ s Concatenate the obtained data values.

[0052] Furthermore, in the flow field prediction method based on active transfer learning and incomplete supervision provided by the present invention, the step of obtaining unlabeled sub-data in the unlabeled dataset and inputting the data into the pre-trained student model to obtain corresponding predicted physical field data includes:

[0053] Obtaining unlabeled sub-data in the unlabeled data set, inputting the data into the pre-trained student model for processing, and outputting predicted physical field data;

[0054]

[0055] in, To predict physical field data, f s (x t ,τ s θ s ) is the student model f s Based on the student model control parameter θ s For unlabeled sub-data with random quantile level τ s Concatenate the obtained data values.

[0056] Furthermore, in the flow field prediction method based on active transfer learning and incomplete supervision provided by the present invention, the first label data set includes label data and corresponding real label physical field data;

[0057] The second labeled data set includes second labeled data and second real physical field data;

[0058] The unlabeled dataset includes unlabeled data and unlabeled sub-data.

[0059] The main advantages of the technical solution of the present invention are as follows:

[0060] The flow field prediction method based on active transfer learning and incomplete supervision of the present invention only needs to obtain a small amount of labeled first label data set and second label data set and a large amount of unlabeled unlabeled data set to realize model training, and through active annotation and pseudo prediction of the unlabeled data set, it is used to train the pre-trained student model, so that the obtained student model has higher prediction accuracy, reduces the demand for the amount of required complete label data sample data, solves the problem of inconvenience in collecting and processing complete label sample data in traditional technology, thereby improving the training optimization efficiency of the pre-trained student model and reducing the training cost of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The drawings described herein are used to provide a further understanding of the embodiments of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0062] Figure 1 A flow chart of a flow field prediction method based on active transfer learning and incomplete supervision provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0063] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present invention and corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0064] The technical solutions provided by the embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0065] Example 1:

[0066] The embodiment of the present invention provides a flow field prediction method based on active transfer learning and incomplete supervision, referring to Figure 1 ,include:

[0067] Constructing the initial transfer model, initial student model and teacher model;

[0068] Obtain a first labeled dataset, a second labeled dataset, and an unlabeled dataset;

[0069] The initial migration model is trained using the first labeled dataset, and the control parameters of the initial migration model are optimized based on the deviation of the migration loss function analysis until the model converges, thereby obtaining the control parameters of the migration model.

[0070] Initialize the control parameters of the initial student model and the teacher model according to the transfer model control parameters;

[0071] Preheat training the initial student model using the second label data set to obtain a pretrained student model;

[0072] Based on the uncertainty query strategy, a pre-trained student model is used to actively annotate some unlabeled data in the unlabeled dataset to obtain the corresponding sparsely annotated physical field data;

[0073] Obtain unlabeled sub-data in the unlabeled dataset that lacks corresponding sparsely annotated physical field data, and add corresponding pseudo-physical field data to the unlabeled sub-data through the teacher model;

[0074] Constructing a real physical field dataset according to the second real physical field data, the sparsely annotated physical field data, and the pseudo physical field data corresponding to the second labeled data in the second labeled dataset;

[0075] The pre-trained student model makes predictions based on the second labeled data in the second labeled data set, the unlabeled data and the unlabeled sub-data in the unlabeled data set to obtain a predicted physical field data set, and the pre-trained student model is trained by analyzing the deviation between the real physical field data set and the predicted physical field data set through a comprehensive loss function until the pre-trained student model converges to obtain a student model;

[0076] The obtained student model is used to predict the physical data of the flow field around the wing.

[0077] In the above embodiment, the initial transfer model is trained using the first labeled dataset, and the control parameters of the initial transfer model are optimized based on the deviation of the transfer loss function analysis until the model converges, the transfer model control parameters are obtained, and the control parameters of the initial student model and the teacher model are initialized according to the transfer model control parameters;

[0078] Preheat training the initial student model using the second label data set to obtain a pretrained student model;

[0079] Based on the uncertainty query strategy, a pre-trained student model is used to actively annotate some unlabeled data in the unlabeled dataset to obtain the corresponding sparsely labeled physical field data; unlabeled sub-data that lacks corresponding sparsely labeled physical field data in the unlabeled dataset is obtained, and the corresponding pseudo-physical field data is added to the unlabeled sub-data through the teacher model;

[0080] The pre-trained student model is trained by analyzing the deviation between the real physical field dataset and the predicted physical field dataset through a comprehensive loss function until the pre-trained student model converges, and the student model is obtained and used to predict the physical data of the flow field around the wing.

[0081] In the above embodiments, the real physical field data set includes second real physical field data, sparsely labeled physical field data and pseudo physical field data; the predicted physical field data set is the second predicted physical field data, sparse predicted physical field data and predicted physical field data obtained by the pre-trained student model based on the second labeled data, unlabeled data and unlabeled sub-data.

[0082] In the above embodiment, the pre-training of the initial student model through the second label data set can call the migration loss function to perform model optimization to obtain the pre-trained student model.

[0083] The beneficial effects of the above technology are: supervised training of the initial migration model is realized through the first labeled data set, and the acquired migration model control parameters are actively migrated to the initial student model and the teacher model for initial configuration, and the initial student model is preheated and trained through the second labeled data set to obtain the pre-trained student model; based on the uncertainty query strategy, part of the unlabeled data in the unlabeled data set is actively annotated to obtain the corresponding sparsely labeled physical field data; unlabeled sub-data that lacks corresponding sparsely labeled physical field data in the unlabeled data set is obtained, and the corresponding pseudo-physical field data is added to the unlabeled sub-data through the teacher model, thereby realizing the annotation processing of the unlabeled data set; and based on the real physical field data set and the predicted physical field data set, it is realized The pre-trained student model is trained to obtain a student model for predicting the physical data of the flow field around the wing; compared with traditional technologies, the flow field prediction method based on active transfer learning and incomplete supervision proposed in the present invention only needs to obtain a small amount of labeled first label data sets and second label data sets, and a large amount of unlabeled unlabeled data sets to realize the training of the model, and through the active annotation and pseudo-prediction of the unlabeled data sets, it is used to train the pre-trained student model, so that the obtained student model has higher prediction accuracy, reduces the demand for the amount of complete label data sample data required, solves the problem of inconvenience in collecting and processing complete label sample data in traditional technologies, thereby improving the training optimization efficiency of the pre-trained student model and reducing the training cost of the model.

[0084] Example 2:

[0085] An embodiment of the present invention provides a flow field prediction method based on active transfer learning and incomplete supervision, comprising the steps of: training an initial transfer model using a first labeled dataset, and optimizing control parameters of the initial transfer model based on deviation analysis of a transfer loss function until the model converges, thereby obtaining the transfer model control parameters; and comprising:

[0086] Obtain labeled data from the first labeled data set and splice it with the quantile level based on the channel dimension, input it into the initial migration model for processing, and obtain predicted labeled physical field data;

[0087] The migration deviation data between the predicted labeled physical field data and the real labeled physical field data corresponding to the labeled data is analyzed through the migration loss function;

[0088]

[0089] Among them, L sqr is the migration loss function, and the first label dataset is is the i-th labeled data in the first labeled data set, For the initial migration model based on labeled data The quantile level during the ep-th training, For the initial migration model, the control parameter θ is based on the migration model according to the labeled data and quantile levels Output predicted label physical field data, There is labeled data in the first labeled dataset The corresponding real label physical field data, l τ Quantile loss function, τ r The quantile level set when training the initial migration model obeys the uniform distribution U[0,1], and E is the quantile loss function l τ Expected values obtained from conducting deviation assessments;

[0090] The control parameters of the initial migration model are optimized by minimizing the migration deviation data until a preset maximum number of iterations is reached or the model converges, and the migration model control parameters are obtained.

[0091] In the above embodiment, the labeled data in the first labeled data set is obtained and spliced with the quantile level based on the channel dimension, and input into the initial migration model for processing to obtain predicted label physical field data; the migration deviation data of the predicted label physical field data and the real label physical field data corresponding to the labeled data is analyzed by the migration loss function, and the control parameters of the initial migration model are optimized by minimizing the migration deviation data until the preset maximum number of iterations is reached or the model converges, and the migration model control parameters are obtained.

[0092] In the above embodiment, the quantile loss function is calculated as follows:

[0093]

[0094] in, It is an indicator function based on the fact that the predicted label physical field data is not greater than the actual label physical field data.

[0095] The beneficial effect of the above technology is that: through the labeled data in the first labeled data set and the corresponding real labeled physical field data, self-supervised training of the initial migration model is achieved based on the migration loss function, and then the migration model control parameters are obtained.

[0096] Example 3:

[0097] An embodiment of the present invention provides a flow field prediction method based on active transfer learning and incomplete supervision, comprising the following steps: actively annotating part of the unlabeled data in the unlabeled dataset through a pre-trained student model based on an uncertainty query strategy to obtain corresponding sparsely annotated physical field data; including:

[0098] Analyze the prediction uncertainty of the pre-trained student model based on sample data;

[0099]

[0100] Among them, u(x) is the pre-trained student model f s The uncertainty of the prediction of sample data x, α is the significance level;

[0101] Annotate the unlabeled data at high uncertainty locations in the unlabeled dataset based on prediction uncertainty to obtain sparsely annotated physical field data;

[0102]

[0103] Where M is a binary mask, For unlabeled data in unlabeled datasets Uncertainty prediction is performed, and all unlabeled data in the unlabeled dataset are sorted in descending order of uncertainty. The first m% of unlabeled data in the sequence are selected for sparse physical annotation, and Q(m%) is the m% quantile of the uncertainty function u.

[0104] In the above embodiment, the prediction uncertainty of the pre-trained student model is analyzed based on the sample data, and the uncertainty prediction is performed on all the unlabeled data in the unlabeled data set through the pre-trained student model, and the data are sorted in descending order. The pre-trained student model is used to perform sparse physical annotation on the first m% of the unlabeled data in the sequence to obtain sparsely annotated physical field data.

[0105] The beneficial effects of the above technology are: based on the uncertainty query strategy, sparse labeling of high-uncertainty unlabeled data in the unlabeled data set is realized, and sparsely labeled physical field data corresponding to the unlabeled data is obtained to solve the problem of insufficient labeled data and inability to effectively train the model. In addition, obtaining sparsely labeled physical field data labeled with high-uncertainty unlabeled data can limit the maximum prediction error of the obtained student model and improve the prediction accuracy of the student model.

[0106] Example 4:

[0107] An embodiment of the present invention provides a flow field prediction method based on active transfer learning and incomplete supervision, comprising the steps of obtaining unlabeled sub-data that lacks corresponding sparsely annotated physical field data in an unlabeled dataset, and adding corresponding pseudo-physical field data to the unlabeled sub-data through a teacher model; comprising:

[0108] Obtain unlabeled sub-data in the unlabeled dataset, input it into the teacher model for processing, and output the corresponding pseudo-physical field data;

[0109]

[0110] in, is pseudo-physical field data, For the teacher model f s Based on the teacher model control parameter θ t For unlabeled sub-data With a fixed quantile level τ c Concatenate the obtained data values.

[0111] In the above embodiment, the unlabeled sub-data in the unlabeled data set are spliced with the fixed quantile level and then input into the teacher model for processing to output pseudo physical field data.

[0112] In the above embodiment, the fixed quantile level τ c Set to 0.5.

[0113] The beneficial effect of the above technology is that by setting a fixed quantile level, it is equivalent to not interfering with the unlabeled sub-data. The teacher model processes the data after splicing the unlabeled sub-data with the fixed quantile level, and outputs more accurate pseudo-physical field data. The above technical solution adopts a consistency regularization method based on the quantile level to introduce disturbances without destroying the original unlabeled sub-data, thereby ensuring the consistency of the teacher model output.

[0114] Example 5:

[0115] An embodiment of the present invention provides a flow field prediction method based on active transfer learning and incomplete supervision, comprising the steps of: performing a pre-trained student model prediction based on second labeled data in a second labeled data set, unlabeled data in an unlabeled data set, and unlabeled sub-data to obtain a predicted physical field data set; and training the pre-trained student model by analyzing the deviation between the real physical field data set and the predicted physical field data set through a comprehensive loss function until the pre-trained student model converges to obtain a student model; and comprising:

[0116] Obtaining second labeled data in the second labeled data set and unlabeled data in the unlabeled data set and inputting them into the pre-trained student model to obtain corresponding second predicted physical field data and sparse predicted physical field data;

[0117] Analyze the deviation between the second predicted physical field data and the second real physical field data, and between the sparse predicted physical field data and the sparse labeled physical field data through a supervised loss function to obtain supervised deviation data;

[0118]

[0119] Among them, L sup is the supervision loss function, D t is a supervised dataset constructed based on the second labeled dataset, the unlabeled data in the unlabeled dataset, and the corresponding sparsely annotated physical field data, x t is the second labeled data and unlabeled data in the supervised dataset, Ω is the area where the physical place is located, For the second real physical field data and sparsely labeled physical field data in the supervised dataset, The second predicted physical field data and sparse predicted physical field data output by the pre-trained student model;

[0120] Obtain unlabeled sub-data in the unlabeled dataset, input it into the pre-trained student model, and obtain the corresponding predicted physical field data;

[0121] Analyze the deviation between pseudo physical field data and predicted physical field data through a pseudo loss function to obtain pseudo deviation data;

[0122]

[0123] Among them, L cr is a pseudo loss function, To use the unlabeled sub-data in the unlabeled dataset and the corresponding pseudo-physical field data The pseudo dataset constructed, M is the binary mask, ⊙ is the Hadamard product, Predicted physics data output by the pre-trained student model;

[0124] Construct a comprehensive loss function based on the supervised loss function and the pseudo loss function;

[0125] L=L sup +λL cr

[0126] Among them, L is the comprehensive loss function, λ is the weight value;

[0127] The pre-trained student model is trained by analyzing the deviation between the real physical field dataset and the predicted physical field dataset through a comprehensive loss function to obtain the student model.

[0128] In the above embodiment, the comprehensive function is constructed by the supervised loss function and the pseudo loss function. The deviation between the real physical field data set and the predicted physical field data set is analyzed by the comprehensive loss function to train the pre-trained student model and obtain the student model.

[0129] In the above embodiments, the second labeled data in the second labeled data set and the unlabeled data in the unlabeled data set are input into the pre-trained student model to obtain the second predicted physical field data and the sparse predicted physical field data; the deviation between the second predicted physical field data and the second real physical field data, and the deviation between the sparse predicted physical field data and the sparse labeled physical field data are analyzed through the supervised loss function to obtain supervised deviation data.

[0130] In the above embodiment, the unlabeled sub-data in the unlabeled data set are input into the pre-trained student model to obtain the corresponding predicted physical field data; the deviation between the pseudo physical field data and the predicted physical field data is analyzed by the pseudo loss function to obtain pseudo deviation data.

[0131] In the above embodiments, the deviation between the real physical field data set and the predicted physical field data set is analyzed based on the supervised deviation data and the pseudo deviation data to achieve the training of the pre-trained student model.

[0132] The beneficial effect of the above technology is that the training processing of the pre-trained student model is realized based on the constructed comprehensive loss function through a small amount of second labeled data in the second labeled data set and a large amount of unlabeled data and unlabeled sub-data in the unlabeled data set.

[0133] Example 6:

[0134] An embodiment of the present invention provides a flow field prediction method based on active transfer learning and incomplete supervision, comprising the steps of obtaining second labeled data in a second labeled data set and unlabeled data in an unlabeled data set and inputting them into a pre-trained student model to obtain corresponding second predicted physical field data and sparse predicted physical field data; and comprising:

[0135] Obtaining second labeled data and unlabeled data in the supervised data set, inputting them into the pre-trained student model for processing, and outputting second predicted physical field data and sparse predicted physical field data;

[0136]

[0137] in, For the second predicted physical field data and sparse predicted physical field data, f s (x t ,τs θ s ) is the student model f s Based on the student model control parameter θ s For the second labeled data and unlabeled data x t with random quantile level τ s Concatenate the obtained data values.

[0138] In the above embodiment, the second labeled data and the unlabeled data are spliced with random quantile levels and input into the pre-trained student model for processing to obtain the corresponding second predicted physical field data and sparse predicted physical field data.

[0139] Example 7:

[0140] The embodiment of the present invention provides a flow field prediction method based on active transfer learning and incomplete supervision, comprising the steps of obtaining unlabeled sub-data in an unlabeled dataset, inputting the data into a pre-trained student model, and obtaining corresponding predicted physical field data; and comprising:

[0141] Obtain unlabeled sub-data in the unlabeled dataset, input it into the pre-trained student model for processing, and output predicted physical field data;

[0142]

[0143] in, To predict physical field data, f s (x t ,τ s θ s ) is the student model f s Based on the student model control parameter θ s For unlabeled sub-data with random quantile level τ s Concatenate the obtained data values.

[0144] In the above embodiment, the unlabeled sub-data in the unlabeled data set are spliced with the random quantile level and then input into the pre-trained student model for processing to output the predicted physical field data.

[0145] In the above embodiment, the random quantile level τ s Obey the uniform distribution U[0,1].

[0146] Example 8:

[0147] An embodiment of the present invention provides a flow field prediction method based on active transfer learning and incomplete supervision, including:

[0148] A first labeled data set includes labeled data and corresponding real labeled physical field data;

[0149] A second labeled data set, comprising second labeled data and second real physical field data;

[0150] Unlabeled datasets include unlabeled data and unlabeled sub-data.

[0151] In the above embodiments, the first labeled data set includes a large amount of labeled data and real labeled physical field data, which are used for supervised training of the initial migration model.

[0152] In the above embodiments, the second labeled data set includes a small amount of second labeled data and second real physical field data, which are used for preheating training of the initial student model.

[0153] In the above embodiments, the unlabeled data in the unlabeled data is used for active annotation through the pre-trained student model to obtain sparsely labeled physical field data; the unlabeled sub-data is used for prediction through the teacher model to obtain pseudo physical field data.

[0154] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or apparatus.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A flow field prediction method based on active transfer learning and incomplete supervision, characterized in that: include: Constructing the initial transfer model, initial student model and teacher model; Obtain a first labeled dataset, a second labeled dataset, and an unlabeled dataset; The initial migration model is trained using the first labeled dataset, and the control parameters of the initial migration model are optimized based on the deviation of the migration loss function analysis until the model converges, thereby obtaining the control parameters of the migration model. Initializing the control parameters of the initial student model and the teacher model according to the migration model control parameters; Preheat training the initial student model using a second label data set to obtain a pretrained student model; Actively annotating part of the unlabeled data in the unlabeled dataset through a pre-trained student model based on an uncertainty query strategy to obtain corresponding sparsely annotated physical field data; Obtain unlabeled sub-data in the unlabeled dataset that lacks corresponding sparsely annotated physical field data, and add corresponding pseudo physical field data to the unlabeled sub-data using the teacher model; Constructing a real physical field dataset according to the second real physical field data, the sparsely annotated physical field data, and the pseudo physical field data corresponding to the second labeled data in the second labeled dataset; The pre-trained student model performs prediction based on the second labeled data in the second labeled data set, the unlabeled data and the unlabeled sub-data in the unlabeled data set to obtain a predicted physical field data set, and trains the pre-trained student model by analyzing the deviation between the real physical field data set and the predicted physical field data set through a comprehensive loss function until the pre-trained student model converges to obtain a student model; The obtained student model is used to predict the physical data of the flow field around the wing.

2. The flow field prediction method based on active transfer learning and incomplete supervision according to claim 1 is characterized in that: The step of training the initial migration model using the first label data set and optimizing the control parameters of the initial migration model based on the deviation of the migration loss function analysis until the model converges to obtain the migration model control parameters includes: Obtaining labeled data from the first labeled data set and splicing the quantile level based on the channel dimension, inputting the data into the initial migration model for processing, and obtaining predicted labeled physical field data; Analyzing migration deviation data between the predicted label physical field data and the real label physical field data corresponding to the label data by using a migration loss function; Among them, L sqr is the migration loss function, and the first label dataset is is the i-th labeled data in the first labeled data set, For the initial migration model based on labeled data The quantile level during the ep-th training, For the initial migration model, the control parameter θ is based on the migration model according to the labeled data and quantile levels Output predicted label physical field data, There is labeled data in the first labeled dataset The corresponding real label physical field data, l τ Quantile loss function, τ r The quantile level set when training the initial migration model obeys the uniform distribution U[0,1], and E is the quantile loss function l τ Expected values obtained from conducting deviation assessments; The control parameters of the initial migration model are optimized by minimizing the migration deviation data until a preset maximum number of iterations is reached or the model converges, thereby obtaining the migration model control parameters.

3. The flow field prediction method based on active transfer learning and incomplete supervision according to claim 1 is characterized in that: The step of actively annotating part of the unlabeled data in the unlabeled data set through a pre-trained student model based on an uncertainty query strategy to obtain corresponding sparsely annotated physical field data includes: Analyzing the prediction uncertainty of the pre-trained student model based on the sample data; Where u(x) is the pre-trained student model f s The uncertainty of the prediction of sample data x, α is the significance level; Annotating the unlabeled data at high uncertainty positions in the unlabeled dataset based on the prediction uncertainty to obtain sparsely annotated physical field data; Where M is a binary mask, To identify the unlabeled data in the unlabeled dataset Uncertainty prediction, and sort all unlabeled data in the unlabeled dataset in descending order of uncertainty, select the first m% of unlabeled data in the sequence for sparse physical annotation, and Q(m%) is the m% quantile of the uncertainty function u.

4. The flow field prediction method based on active transfer learning and incomplete supervision according to claim 1 is characterized in that: The step of obtaining unlabeled sub-data in the unlabeled data set that lacks corresponding sparsely annotated physical field data, and adding corresponding pseudo physical field data to the unlabeled sub-data through the teacher model comprises: Obtaining unlabeled sub-data in the unlabeled data set, inputting the data into the teacher model for processing, and outputting corresponding pseudo physical field data; in, is pseudo-physical field data, For the teacher model f s Based on the teacher model control parameter θ t For unlabeled sub-data With a fixed quantile level τ c Concatenate the obtained data values.

5. The flow field prediction method based on active transfer learning and incomplete supervision according to claim 1, characterized in that: The step of: the pre-trained student model performs prediction based on the second labeled data in the second labeled data set, the unlabeled data and the unlabeled sub-data in the unlabeled data set to obtain a predicted physical field data set, and the pre-trained student model is trained by analyzing the deviation between the real physical field data set and the predicted physical field data set through a comprehensive loss function until the pre-trained student model converges to obtain a student model; including: Obtaining second labeled data in the second labeled data set and unlabeled data in the unlabeled data set and inputting them into the pre-trained student model to obtain corresponding second predicted physical field data and sparse predicted physical field data; Analyze the deviation between the second predicted physical field data and the second real physical field data, and between the sparse predicted physical field data and the sparse labeled physical field data through a supervised loss function to obtain supervised deviation data; Among them, L sup is the supervision loss function, D t is a supervised dataset constructed based on the second labeled dataset, the unlabeled data in the unlabeled dataset, and the corresponding sparsely annotated physical field data, x t is the second labeled data and unlabeled data in the supervised dataset, Ω is the area where the physical place is located, For the second real physical field data and sparsely labeled physical field data in the supervised dataset, The second predicted physical field data and sparse predicted physical field data output by the pre-trained student model; Obtaining unlabeled sub-data in the unlabeled data set, inputting the data into the pre-trained student model, and obtaining corresponding predicted physical field data; Analyze the deviation between pseudo physical field data and predicted physical field data through a pseudo loss function to obtain pseudo deviation data; Among them, L cr is a pseudo loss function, To use the unlabeled sub-data in the unlabeled dataset and the corresponding pseudo-physical field data The pseudo dataset constructed, M is the binary mask, ⊙ is the Hadamard product, Predicted physics data output by the pre-trained student model; Construct a comprehensive loss function based on the supervised loss function and the pseudo loss function; L=L sup +λL cr Among them, L is the comprehensive loss function, λ is the weight value; The pre-trained student model is trained by analyzing the deviation between the real physical field data set and the predicted physical field data set through a comprehensive loss function to obtain a student model.

6. The flow field prediction method based on active transfer learning and incomplete supervision according to claim 5, characterized in that: The step of obtaining the second labeled data in the second labeled data set and the unlabeled data in the unlabeled data set and inputting them into the pre-trained student model to obtain the corresponding second predicted physical field data and sparse predicted physical field data includes: Obtaining second labeled data and unlabeled data in the supervised data set, inputting them into the pre-trained student model for processing, and outputting second predicted physical field data and sparse predicted physical field data; in, For the second predicted physical field data and sparse predicted physical field data, f s (x t ,τ s θ s ) is the student model f s Based on the student model control parameter θ s For the second labeled data and unlabeled data x t with random quantile level τ s Concatenate the obtained data values.

7. The flow field prediction method based on active transfer learning and incomplete supervision according to claim 5, characterized in that: The step of obtaining unlabeled sub-data in the unlabeled data set and inputting the data into the pre-trained student model to obtain corresponding predicted physical field data includes: Obtaining unlabeled sub-data in the unlabeled data set, inputting the data into the pre-trained student model for processing, and outputting predicted physical field data; in, To predict physical field data, f s (x t ,τ s θ s ) is the student model f s Based on the student model control parameter θ s For unlabeled sub-data with random quantile level τ s Concatenate the obtained data values.

8. The flow field prediction method based on active transfer learning and incomplete supervision according to claim 1 is characterized in that: The first label data set includes label data and corresponding real label physical field data; The second labeled data set includes second labeled data and second real physical field data; The unlabeled dataset includes unlabeled data and unlabeled sub-data.

Images