Efficient sampling adaptive data fusion method

Abstract

The invention discloses an efficient sampling adaptive data fusion method. The method comprises the following steps: S100, generating an initial low-precision sample set; s200, obtaining a high-precision sample set by an infinitesimal method; s300, mixing the low-precision and high-precision sample sets to obtain a new sample set; s400, calculating the real distance between the low-precision sample set and the nearest high-precision sample; s500, establishing an adaptive RBF neural network model, and processing the new sample set; s600, substituting the real distance into an adaptive RBF neural network model to update parameters and search an optimal function value; according to the method, a mixed precision pneumatic data fusion model of different precision data is associated, and an infinitesimal division sampling method and a self-adaptive data fusion method are provided, so that efficient initialization and high-precision prediction of an RBF variable credibility model are realized.

Description

Technical Field

[0001] The invention relates to the field of data processing, and in particular to an efficient sampling and adaptive data fusion method. Background Art

[0002] Aerodynamic analysis in the propeller design stage requires a large amount of high-precision aerodynamic data to improve design performance, but the acquisition cost is high. In recent years, with the complexity of the flight environment and the diversification of flight missions, modern aerospace vehicles need to be jointly guaranteed by multidisciplinary responses and a large number of design criteria. As far as the design process of the aircraft is concerned, it is necessary to seek the optimal method under the coupling of multiple disciplines such as flight mechanics, aerodynamics, structure and flight control, so the design scheme needs to be repeatedly iterated and continuously optimized. Evaluating aircraft performance indicators with data close to the actual flight state can effectively reduce the design iteration cycle and improve the optimization efficiency. For example, Airbus conducted about 18,000 hours of wind tunnel tests to achieve the development of A130, and the development of the US F-22 "Raptor" fighter went through about 44,000 hours of high and low speed wind tunnel tests. The contradiction between the high cost of obtaining aerodynamic performance data and the rapid design of aircraft forces researchers to seek more efficient means of obtaining aerodynamic forces. Summary of the invention

[0003] Therefore, in order to solve the above-mentioned shortcomings, the present invention provides an efficient sampling adaptive data fusion method. The present invention associates a mixed-precision aerodynamic data fusion model of data with different precisions, and proposes a microelement partitioning sampling method and an adaptive data fusion method, so as to achieve efficient initialization and high-precision prediction of the RBF variable credibility model.

[0004] Specifically, an efficient sampling adaptive data fusion method includes the following steps:

[0005] S100, generating an initial low-precision sample set;

[0006] S200, microelement method to obtain high-precision sample sets;

[0007] S300, mixing low-precision sample sets and high-precision sample sets to obtain new sample sets;

[0008] S400, calculating the true distance between the low-precision sample set and the nearest high-precision sample;

[0009] S500, establishing an adaptive RBF neural network model to process the new sample set;

[0010] S600, bringing the real distance into the adaptive RBF neural network model to update parameters and find the optimal function value.

[0011] Optionally, the method of step S100 is: adopting a uniform sampling method to obtain a low-fidelity sample set, and using the low-fidelity to construct a priori model to obtain a response value.

[0012] Optionally, the method of step S200 is:

[0013] Based on the accurate low-fidelity response surface, a uniformly distributed low-fidelity sample set is generated. According to the characteristics of the data domain and based on the microelement method, the sampling area is divided into microelements according to the principle that the greater the variance, the greater the uncertainty. The variance of each one is calculated, and the area with the largest variance is selected for sampling to obtain a high-precision sample set.

[0014] Optionally, the method of step S300 is:

[0015] The obtained low-precision sample set is mixed and sorted with the high-precision sample set, and the low-precision sample data that overlaps with the high-precision is deleted to obtain a mixed sample set, and the mixed sample set is normalized to form a new sample set.

[0016] Optionally, the method of step S400 is: since the new sample set is composed of low-precision and high-precision samples, the Euclidean distance between each low-precision sample in the new sample set and its nearest high-precision sample is calculated respectively to prepare data for the gradient descent in step S600.

[0017] Optionally, in the step S500, the adaptive RBF neural network model is established according to the RBF neural network model;

[0018] The way to process the new sample set is: cluster the new data set to obtain the class center, adaptively obtain the number of hidden layers of the optimal neural network based on the silhouette coefficient, and obtain the loss value through forward prediction.

[0019] Optionally, the method of step S600 is:

[0020] According to the actual distance in step S400, the samples are divided into two categories according to the distance, and the deviation value of the lower-precision samples with a closer distance is substituted into the sigmoid function for correction. The parameters of the neural network are updated by the gradient descent method, and the optimal value of the objective function is found. When the model converges, the predicted value is output and the predicted result is denormalized.

[0021] The present invention has the following advantages:

[0022] Based on the RBF neural network model, the present invention proposes an efficient sampling adaptive data fusion method for improving the utilization efficiency of high-precision samples and reducing the interference of the prior model of the low-precision sample set on the high-precision samples. The invention adopts a sampling method of data domain micro-element division, and redistributes the high-precision sample data points according to the size of the data uncertainty contained in the data domain micro-element, thereby improving the global accuracy of the model under the premise of limited high-precision data. At the same time, in view of the problem that low-precision data interferes with high-precision data in the data fusion method, the relative distance between the low-precision data and the corresponding high-precision data point is calculated during the training process, and an activation function is introduced in the link of calculating the error to correct the loss coefficient. The present invention can avoid overfitting of high-precision data in the data space where the amount of high-precision data is insufficient, and effectively improve the prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a method flow chart of an efficient sampling adaptive data fusion method. DETAILED DESCRIPTION

[0024] Embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0025] In this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such 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.

[0026] As described in the background technology, aerodynamic analysis in the propeller design stage requires a large amount of high-precision aerodynamic data to improve design performance, but its acquisition cost is high. In recent years, with the complexity of the flight environment and the diversification of flight missions, modern aerospace aircraft need to be jointly guaranteed by multidisciplinary responses and a large number of design criteria. As far as the design process of the aircraft is concerned, it is necessary to seek the optimal method under the multidisciplinary coupling of flight mechanics, aerodynamics, structure and flight control, so the design scheme needs to be repeatedly iterated and continuously optimized. Evaluating aircraft performance indicators with data close to the actual flight state can effectively reduce the design iteration cycle and improve the optimization efficiency. For example, Airbus conducted about 18,000 hours of wind tunnel tests to achieve the development of A130, and the development of the US F-22 "Raptor" fighter went through about 44,000 hours of high and low speed wind tunnel tests. The contradiction between the high cost of obtaining aerodynamic performance data and the rapid aircraft design forced researchers to seek more efficient aerodynamic acquisition methods.

[0027] Based on the above reasons, the present invention provides an efficient sampling adaptive data fusion method for improving the utilization efficiency of high-precision samples and reducing the interference of low-precision sample set prior models on high-precision samples; Figure 1 As shown, the efficient sampling adaptive data fusion method includes the following steps:

[0028] Exemplarily, the forward propagation process of the RBF neural network is as follows:

[0029] Step S100, generating an initial low-precision sample set;

[0030] Step S200, using the microelement method to obtain a high-precision sample set;

[0031] Step S300, mixing the low-precision sample set and the high-precision sample set to obtain a new sample set;

[0032] Step S400, calculating the true distance between the low-precision sample set and the nearest high-precision sample;

[0033] Step S500, establishing an adaptive RBF neural network model to process the new sample set;

[0034] Step S600: Bring the real distance into the adaptive RBF neural network model to update parameters and find the optimal function value.

[0035] The above technical features, based on the RBF neural network model, propose an efficient sampling adaptive data fusion method to improve the utilization efficiency of high-precision samples and reduce the interference of the low-precision sample set prior model on high-precision samples. The invention adopts a sampling method of data domain micro-element division, and redistributes high-precision sample data points according to the size of the data uncertainty contained in the data domain micro-element, thereby improving the global accuracy of the model under the premise of limited high-precision data volume.

[0036] In one embodiment, the method of step S100 is: using a uniform sampling method to obtain a low-fidelity sample set, and using the low-fidelity to construct a priori model to obtain a response value.

[0037] In one embodiment, the method of step S200 is:

[0038] Based on the accurate low-fidelity response surface, a uniformly distributed low-fidelity sample set is generated. According to the characteristics of the data domain and based on the microelement method, the sampling area is divided into microelements according to the principle that the greater the variance, the greater the uncertainty. The variance of each one is calculated, and the area with the largest variance is selected for sampling to obtain a high-precision sample set.

[0039] For example, within a complete data domain, the uncertainty of data is not necessarily uniformly distributed. The uncertainty of data in a certain area may be large, while the uncertainty of data in other areas is small. Therefore, it is necessary to divide the uncertainty of the data domain to allocate sampling points. In order to reflect the uncertainty distribution of data in a data domain, the sampling method of micro-element division is used to divide a complete data domain into multiple data domain micro-elements and then calculate the variance of each to reflect the uncertainty of the current data micro-element. The expression is as follows: Q is the total spatial uncertainty, q i is the uncertainty element:

[0040] Q=q1+q2+…q n

[0041] Divide the complete data domain into multiple data domain elements q i , calculate the variance of each data domain element separately The variance can reflect the uncertainty of the data domain micro-element, so the high-precision sampling points can be allocated according to the uncertainty of each data domain micro-element. The data domain micro-element with higher uncertainty can be allocated more high-precision data sampling points, while the data domain micro-element with lower uncertainty can be allocated fewer high-precision data sampling points. The allocation method is as follows:

[0042]

[0043] In the formula, A is the total number of sample points, a i The number of sample points collected in each data domain element.

[0044] By uniformly sampling each data domain element according to the assigned high-precision sampling points, the sampling efficiency can be maximized as much as possible.

[0045] The number of high-precision data samples is set according to the following formula:

[0046]

[0047] Where n is the total number of sample points collected, and k is the dimension of the design space.

[0048] In one embodiment, the method of step S300 is:

[0049] The obtained low-precision sample set is mixed and sorted with the high-precision sample set, and the low-precision sample data that overlaps with the high-precision is deleted to obtain a mixed sample set, and the mixed sample set is normalized to form a new sample set.

[0050] In one embodiment, the method of step S400 is: since the new sample set is composed of low-precision and high-precision samples, the Euclidean distance between each low-precision sample in the new sample set and its nearest high-precision sample is calculated respectively to prepare data for the gradient descent in step S600.

[0051] In one embodiment, in step S500, the adaptive RBF neural network model is established according to the RBF neural network model;

[0052] The way to process the new sample set is: cluster the new data set to obtain the class center, adaptively obtain the number of hidden layers of the optimal neural network based on the silhouette coefficient, and obtain the loss value through forward prediction.

[0053] For example, the RBF (Radical Basis Function) neural network is a forward three-layer network, which includes an input layer, a hidden layer, and an output layer, and there is no loop for feedback between them. The function of the input layer is value transfer, the hidden layer and the output layer are operation units, the activation function of the hidden layer neurons is RBF, and the activation function of the output layer neurons is a linear summation function.

[0054] Let h j is the output of the jth neuron in the hidden layer, where h j The output calculation formula is as follows:

[0055]

[0056] in As the kernel function, the Gaussian radial basis kernel function is used, and the formula is as follows:

[0057]

[0058] in is the coordinate vector of the center point of the Gaussian basis function of the jth neuron in the hidden layer, i = 1…n, j = 1…m, b = [b1,…,b m ] T , b j is the width of the Gaussian basis function of the jth neuron in the hidden layer.

[0059] The RBF network weights are:

[0060] w=[w1,…w m ] T ;

[0061] RBF final network output formula:

[0062] y(x)=w T h=w1h1+w2h2+…+w m h m ;

[0063] Given an input sample x, the activation function Gaussian basis function is given by the formula Indicates the calculation of the output h of the hidden layer i , and then the final output y(x) is obtained by weighting the output layer, which is expressed by the following formula:

[0064]

[0065] Where n is the number of hidden layer nodes of the RBF neural network, w i is the weight connecting the hidden layer nodes and the output layer nodes, c i is the center of the hidden layer node, σ i is the width parameter of the hidden layer nodes.

[0066] After calculating y(x), we need to calculate the error of the output layer. Given the training data set {(x i ,y i )}, the loss value of the model is defined as the square loss expressed by the following formula, where m is the number of samples:

[0067]

[0068] Finally, the minimum value is solved iteratively along the direction of gradient descent (gradient descent method), and the model parameters w are updated during the iteration process. i 、c i , σ i , so that the loss function is minimized. The parameter update rule is expressed by the following formula, where η is the learning rate:

[0069]

[0070] The above is the forward propagation process of the RBF neural network model. Based on the above RBF neural network model, an adaptive adjustment module is added to it to form an adaptive RBF neural network. The specific steps are as follows:

[0071] Known true value y(x) and RBF network prediction output The loss value Loss can be obtained.

[0072] 1) The sigmoid activation function is introduced in the process of calculating the loss Loss, where the sigmoid function is as follows:

[0073]

[0074] 2) Perform data preprocessing on the low-precision data set, traverse the low-precision data set and calculate the distance x between each data point and the nearest high-precision data point dis .

[0075] 3) When calculating the Loss of this point, the point x dis Substitute the sigmoid function to obtain the deviation correction coefficient to correct the Loss, that is, Among them, ω, The selection is made based on the data distribution in the sample set.

[0076] In one embodiment, the method of step S600 is:

[0077] According to the actual distance in step S400, the samples are divided into two categories according to the distance, and the deviation value of the lower-precision samples with a closer distance is substituted into the sigmoid function for correction. The parameters of the neural network are updated by the gradient descent method, and the optimal value of the objective function is found. When the model converges, the predicted value is output and the predicted result is denormalized.

[0078] For example, the low-precision data set is preprocessed, and the low-precision data set is traversed to calculate the distance x between each data point and the nearest high-precision data point dis .

[0079] When calculating the Loss of this point, the point x dis Substitute the sigmoid function to obtain the deviation correction coefficient to correct the Loss, that is, Where ω=-130.89,

[0080] The present invention adopts a sampling method of data domain micro-element division, and redistributes high-precision sample data points according to the size of the data uncertainty contained in the data domain micro-element, thereby improving the global accuracy of the model under the premise of limited high-precision data. At the same time, in view of the problem that low-precision data interferes with high-precision data in the data fusion method, the relative distance between low-precision data and corresponding high-precision data points is calculated during the training process, and an activation function is introduced in the link of calculating errors to correct the loss coefficient. This method can avoid overfitting of high-precision data in the data space where the amount of high-precision data is insufficient, and effectively improve the prediction accuracy of the model.

[0081] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An efficient sampling adaptive data fusion method, comprising the following steps: S100, generating an initial low-precision sample set; S200, microelement method to obtain high-precision sample sets; S300, mixing low-precision sample sets and high-precision sample sets to obtain new sample sets; S400, calculating the true distance between the low-precision sample set and the nearest high-precision sample; S500, establishing an adaptive RBF neural network model to process the new sample set; S600, bringing the real distance into the adaptive RBF neural network model to update parameters and find the optimal function value.

2. According to claim 1, a highly efficient sampling adaptive data fusion method is characterized in that: The method of step S100 is: using a uniform sampling method to obtain a low-fidelity sample set, and using the low-fidelity to build a priori model to obtain a response value.

3. The efficient sampling adaptive data fusion method according to claim 1, characterized in that: The method of step S200 is: Based on the accurate low-fidelity response surface, a uniformly distributed low-fidelity sample set is generated. According to the characteristics of the data domain and based on the microelement method, the sampling area is divided into microelements according to the principle that the greater the variance, the greater the uncertainty. The variance of each one is calculated, and the area with the largest variance is selected for sampling to obtain a high-precision sample set.

4. The efficient sampling adaptive data fusion method according to claim 1, characterized in that: The method of step S300 is: The obtained low-precision sample set is mixed and sorted with the high-precision sample set, and the low-precision sample data that overlaps with the high-precision is deleted to obtain a mixed sample set, and the mixed sample set is normalized to form a new sample set.

5. The efficient sampling adaptive data fusion method according to claim 1, characterized in that: The method of step S400 is: respectively calculating the Euclidean distance between each low-precision sample in the new sample set and its nearest high-precision sample.

6. The efficient sampling adaptive data fusion method according to claim 1, characterized in that: In step S500, the adaptive RBF neural network model is established according to the RBF neural network model; the new sample set is processed by clustering the new data set to obtain the class center, adaptively obtaining the number of hidden layers of the optimal neural network based on the silhouette coefficient, and obtaining the loss value through forward prediction.

7. The efficient sampling adaptive data fusion method according to claim 1, characterized in that: The method of step S600 is: According to the actual distance in step S400, the high-precision and low-precision sample data points are divided into two categories according to the distance, and the deviation value of the lower-precision sample with a closer distance is substituted into the function for correction. The parameters of the neural network are updated by the gradient descent method, and the optimal value of the objective function is found. When the adaptive RBF neural network model converges, the predicted value is output and the predicted result is denormalized.

Images