Optical fiber flexible body morphology inversion method based on wavelength residual neural network

Abstract

The embodiment of the present invention discloses a method for inverting the morphology of an optical fiber flexible body based on a wavelength residual neural network, which relates to the technical field of a refueling hose sensor in an aerial refueling device, and can reconstruct the morphology of an aerial refueling hose, thereby improving the accuracy of the sensor sensing the deformation of the hose during the refueling process. The present invention includes: collecting the shape data of the aerial refueling hose and using it as a data label, and recording the correspondence between the data label and the central wavelength offset of the distributed optical fiber sensor according to the user operation. Constructing a wavelength residual neural network block based on the wavelength data measurement of the distributed optical fiber sensor. Using the feature extraction layer of the wavelength residual neural network as a feature extractor and training it. Constructing a neural network of a multi-layer perceptron, and using the trained multi-layer perceptron to invert the actual deformation of the aerial refueling hose.

Description

Technical Field

[0001] The invention relates to the technical field of a refueling hose sensor in aerial refueling equipment, and in particular to an optical fiber flexible body morphology inversion method based on a wavelength residual neural network. Background Art

[0002] With the continuous advancement of aviation technology and drone technology, refueling machines are needed in many practical scenarios. Unlike manned refueling machines, which can be manually adjusted by pilots and manually respond to various emergencies, once the refueling hose of an unmanned refueling machine is extended, its shape change cannot be effectively sensed. In addition, the aerial refueling hose is affected by complex aerodynamic factors during the refueling process and is prone to whip deformation, affecting docking efficiency and flight safety.

[0003] The existing aerial refueling visual monitoring system can only obtain the position information of the cone sleeve at the end of the refueling hose, but cannot accurately perceive the deformation of the hose during the refueling process, which affects the drone operator's judgment on the actual state of the refueling hose, thereby threatening the flight safety of the drone.

[0004] Therefore, it is necessary to integrate sensors into the aerial refueling hose to obtain the hose shape. However, since the outer layer of the hose is a superelastic rubber structure, its strain response is nonlinear. The strain information sensed by the sensor is difficult to correlate and map with the hose shape characteristics, making it difficult to reconstruct the shape of the aerial refueling hose, resulting in difficulties in related ground testing and a lack of data for iterating the UAV flight control. For the aerial refueling drones that have been put into operation, the sensors on the drones cannot provide accurate sensing information due to the above reasons, resulting in inaccurate simulation images fed back to the drone operating system interface. Therefore, experienced operators are still required to complete the oil pipe docking manually, so that the execution efficiency of aerial refueling tasks is generally low. Summary of the invention

[0005] The embodiment of the present invention provides a fiber optic flexible body morphology inversion method based on wavelength residual neural network, which can reconstruct the shape of an aerial refueling hose, thereby improving the accuracy of the sensor sensing the deformation of the hose during the refueling process.

[0006] To achieve the above object, the embodiments of the present invention adopt the following technical solutions:

[0007] A method for optical fiber flexible body morphology inversion based on wavelength residual neural network, comprising:

[0008] S1. Install distributed fiber optic sensors in the aerial refueling hose and record the position of each sensor on the aerial refueling hose.

[0009] S2, collecting the shape data of the aerial refueling hose and using it as a data tag, and recording the corresponding relationship between the data tag and the central wavelength offset of the distributed optical fiber sensor according to the user operation;

[0010] S3, constructing a wavelength residual neural network block based on wavelength data measurement of a distributed optical fiber sensor, and inputting the training set and the test set into the wavelength residual neural network block for training;

[0011] S4, using the feature extraction layer of the wavelength residual neural network as a feature extractor and training it;

[0012] S5, constructing a neural network of a multi-layer perceptron, inputting the features of the training set and the test set extracted by the feature extractor into the multi-layer perceptron and training the multi-layer perceptron;

[0013] S6. Use the trained multi-layer perceptron to invert the actual deformation of the aerial refueling hose.

[0014] Specifically, S1 includes: on the outer rubber surface of the aerial refueling hose, such as Figure 1 As shown, three distributed optical fiber sensors are arranged along the circumferential directions of 0°, 120° and 240°, respectively, that is, each distributed optical fiber sensor serves as a morphology sensing path.

[0015] Specifically, S2 includes: setting the aerial refueling hose to different shapes, and collecting shape data of the aerial refueling hose under each shape and central wavelength offset data of the distributed optical fiber sensor, wherein the deformation shape can be manually set by researchers. The shape data includes a deformation curve corresponding to a shape C=k 1 x 3 +k 2 x 2 +k 3 x+ k 4 , the label value of the data label is expressed as K =[ k 1 ，k 2 ，k 3 ，k 4 ], k 1 ~k 4 They represent the four adjustment factors in the aerial refueling hose shape curve C respectively;

[0016] The collected central wavelength offset data of the distributed optical fiber sensor include:

[0017] △ l 0 o =[△ l 0 o ,1 , △ l 0 o ,2 , △ l 0 o ,3 , ..., △ l 0 o ,n ]、

[0018] △ l 120 o =[△ l 120 o ,1 , △ l 120 o ,2 , △ l 120 o ,3 , ..., △ l 120 o ,n ]、

[0019] △ l 240 o =[△ l 240 o ,1 , △ l 240 o ,2 , △ l 240 o ,3 , ..., △ l 240 o ,n ],

[0020] Among them, △ l 0 oIndicates the central wavelength offset of the distributed optical fiber sensor arranged along the 0° direction of the hose circumference, △ l 120 o It represents the central wavelength offset of the distributed optical fiber sensor arranged along the circumference of the hose at 120°, △ l 240 o It represents the central wavelength offset of the distributed optical fiber sensor arranged along the 240° direction of the hose circumference, △ l 0 o ,n represents the nth data collection point in the distributed optical fiber sensor arranged along the 0° direction of the hose circumference, △ l 120 o ,n represents the nth data collection point in the distributed optical fiber sensor arranged along the circumference of the hose at 120°, △ l 240 o ,n represents the nth data collection point in the distributed optical fiber sensor arranged along the 240° direction of the hose circumference. n Represents the number of data collection points in the distributed optical fiber sensor; unfolds the aerial refueling hose surface into a two-dimensional plane, and maps the central wavelength offset data of the distributed optical fiber sensor to the two-dimensional plane to obtain a two-dimensional central wavelength offset numerical matrix, and then continues to obtain a three-dimensional central wavelength offset matrix and stores it in the sample database.

[0021] Furthermore, before S3, it includes: in the ResNet18 neural network, the three-dimensional center wavelength offset matrix of the distributed optical fiber sensor is passed through a convolution layer to obtain a center wavelength offset numerical matrix △ of size 64×125×125 l [1] , where the convolution kernel size of the convolution layer is 3×7×7, the output channel is 64, the step size is 2, and the padding is 3; △ l [1] After the PReLU activation function, it is input into the maximum pooling layer to obtain a central wavelength offset numerical matrix of size 64×63×63 △ l [2] , where the pooling kernel size of the maximum pooling layer is 3×3, the stride is 2, and the padding is 1.

[0022] Specifically, S3 includes: constructing two wavelength residual neural network blocks, which can be distinguished by whether they contain convolution downsampling. The first wavelength residual neural network block includes: two convolution layers with the same structure and one residual calculation layer. The structure of the two convolution layers with the same structure is: the convolution kernel size is 3×3, the step size is 1, and the padding is 1; the structure of the second wavelength residual neural network block includes: three convolution layers, wherein the convolution kernel size of the first convolution layer is 1×1, the output channel is 64, the step size is 2, and the padding is 1; the structure of the second convolution layer and the third convolution layer is the same, that is, the convolution kernel size is 3×3, the step size is 1, and the padding is 1.

[0023] The process of inputting the training set and the test set into the wavelength residual neural network block for training includes: l [2] The first wavelength residual neural network block is input, and passes through two convolutional layers of the first wavelength residual neural network block, wherein the output of the second convolutional layer is represented as F [1] (△ l ), F [1] (△ l ) is the element in row i and column j. , , among which , The numerical matrix △ represents the central wavelength offset of the output of the first convolutional layer in the first wavelength residual neural network block l [3] The i-th row and j-th column element in is the central wavelength offset numerical matrix △ l [2] The i-th row and j-th column element in ; k 1 , p 1 , s 1 is the convolution kernel size, padding and stride of the first convolutional layer in the first wavelength residual neural network block, k 2 , p 2 , s 2 is the convolution kernel size, padding and stride of the second convolutional layer in the first wavelength residual neural network block; l [2] is sent to the end of the first wavelength residual neural network block and combined with F [1] (△ l ) to obtain the output H of the first wavelength residual neural network block [1] (△ l ), H [1] (△ l ) = F [1] (△ l )+△ l [2] ; H [1] (△ l ) inputs the second wavelength residual neural network block, and the output of the second wavelength residual neural network block is H [2] (△ l );Then H [2] (△ l ) Input the third wavelength residual neural network block, the output of the third wavelength residual neural network block is H [3] (△ l ), H [3] (△ l ) = F [3] (△ l )+W*{ H [2] (△ l )},F [3] (△ l ) is H [2] (△ l ) is a numerical matrix of center wavelength offset obtained by two convolutional layers in the third wavelength residual neural network block, where W*{} represents a convolution operation.

[0024] Specifically, in S4, the cross entropy loss function is used in the process of training the feature extractor L ( k , y ) is used as a performance optimization function to optimize the effect of feature extraction by the feature extractor, where: , k is the adjustment factor parameter, k 1 ~k 4 Equal to the above k 1 ~k 4 , y is the adjustment factor parameter, k i Adjustment factor label for aerial refueling hose, yi The probability that the input central wavelength offset numerical matrix belongs to the i-th adjustment factor label of the aerial refueling hose is predicted by the ResNet18 neural network model, where i is the number of the adjustment factor label and C is the maximum value of the number of the adjustment factor label; the cross entropy loss function is used, and the Adam gradient descent back propagation optimization algorithm is used to update the network parameters, where the first-order moment of the gradient is estimated as: w t =β 1 w t-1 + (1 -β 1 ) ▽J ( W t ), the second-order moment of the gradient is estimated as: v t =β 2 v t-1 + (1 -β 2 ) ( ▽J ( W t )) 2 , w t is the first-order moment estimate of the t-th iteration, ▽J ( W t ) is the gradient of the tth iteration, t is the number of iterations, β 1 is the decay rate of the first-order moment estimate, W t represents the network parameters of the tth iteration, v t is the second-order moment estimate of the t-th iteration, β 2 is the decay rate of the second-order moment estimate. The Adam gradient descent back-propagation optimization algorithm is used to update the network parameters, including: ,in, represents the bias correction of the first-order moment estimate, represents the bias correction of the second-order moment estimate, W t+1 It is represented as the network parameters of t+1 iterations, a is the learning rate, and e is a constant.

[0025] Specifically, in S5, the neural network architecture of the constructed multilayer perceptron includes: 1 input layer, 2 hidden layers and 1 output layer; the connection between the two hidden layers and the connection between the second hidden layer and the output layer all use the tansig activation function; the number of neuron nodes in the first hidden layer is 256, and the number of neuron nodes in the second hidden layer is 128; the output layer contains 4 neuron nodes, and the 4 neuron nodes correspond to the four adjustment factors in the shape curve C of the aerial refueling hose. k 1 ，k 2 ，k 3 ，k 4 .

[0026] The optical fiber flexible body morphology inversion method based on wavelength residual neural network provided in the embodiment of the present invention realizes the reconstruction analysis of the aerial refueling hose morphology, improves the accuracy of the sensor's perception of the deformation of the hose during the refueling process, and can provide reliable sensing information for UAV operators to perform refueling tasks, thereby increasing the execution efficiency of aerial refueling tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0028] Figure 1 A schematic diagram of arranging a distributed optical fiber sensor on an aerial refueling hose provided by an embodiment of the present invention;

[0029] Figure 2 A schematic diagram of a method flow chart provided by an embodiment of the present invention;

[0030] Figure 3 A schematic diagram of a residual neural network in front of a multi-layer perceptron provided in an embodiment of the present invention;

[0031] Figure 4 A schematic diagram of a multilayer perceptron provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0032] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The embodiments of the present invention will be 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 with the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention. It can be understood by those skilled in the art that, unless specifically stated, the singular forms "one", "one", "said" and "the" used herein may also include plural forms. It should be further understood that the term "including" used in the specification of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or their groups. It should be understood that when we say that an element is "connected" or "coupled" to another element, it can be directly connected or coupled to other elements, or there may also be intermediate elements. In addition, the "connection" or "coupling" used here may include wireless connection or coupling. The term "and / or" used herein includes any unit and all combinations of one or more associated listed items. It can be understood by those skilled in the art that, unless otherwise defined, all terms (including technical terms and scientific terms) used herein have the same meaning as generally understood by those skilled in the art in the field to which the present invention belongs. It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless defined as herein.

[0033] The embodiment of the present invention provides a method for inverting the morphology of an optical fiber flexible body based on a wavelength residual neural network. Figure 2 As shown, including:

[0034] S1. Install distributed fiber optic sensors in the aerial refueling hose and record the position of each sensor on the aerial refueling hose.

[0035] S2. Collect the shape data of the aerial refueling hose and use it as a data label. Record the corresponding relationship between the data label and the central wavelength offset of the distributed optical fiber sensor according to the user operation. The corresponding relationship between the shape data and the central wavelength offset can be manually set in advance. Use the recorded corresponding relationship to establish a sample database, which includes: a training set, a test set, and a validation set. Use the shape data of the aerial refueling hose and the central wavelength offset of the distributed optical fiber sensor to establish a database, and use the shape data as a data label, where the label value is expressed as K =[ k 1，k 2 ，k 3 ， k 4 ], determine the correspondence between the aerial refueling hose shape data label and the central wavelength offset of the distributed optical fiber sensor, and randomly divide each correspondence into training set, test set and validation set according to proportion.

[0036] S3, construct a wavelength residual neural network block based on the wavelength data measurement of the distributed optical fiber sensor, and input the training set and the test set into the wavelength residual neural network block for training. Among them, based on the ResNet18 residual neural network, a wavelength residual neural network block based on the wavelength data measurement of the distributed optical fiber sensor can be constructed, and the training set and the test set can be input into the ResNet18 neural network for training and learning, and the trained parameter model can be saved.

[0037] S4. Using the feature extraction layer of the wavelength residual neural network as a feature extractor and training it, wherein the trained feature extractor is used for feature extraction of the aerial refueling hose morphology.

[0038] S5. Construct a neural network of a multi-layer perceptron, input the features of the training set and the test set extracted by the feature extractor into the multi-layer perceptron and train the multi-layer perceptron, and the verification set is used to verify the trained multi-layer perceptron. Among them, based on the ResNet18 residual neural network, a wavelength residual neural network block based on the wavelength data measurement of the distributed optical fiber sensor is constructed, and the training set and the test set are input into the ResNet18 neural network for training and learning, and the trained parameter model is saved. The ResNet18 neural network feature extraction layer is selected as the feature extractor for the morphological perception of the aerial refueling hose, and the training set and the test set are put into the feature extractor to extract features. Construct a multi-layer perceptron neural network, input the training set and the test set features extracted by the ResNet18 neural network into the multi-layer perceptron for training, and after the training is completed, the verification set is used to verify the morphological perception of the aerial refueling hose.

[0039] S6. Use the trained multi-layer perceptron to invert the actual deformation of the aerial refueling hose.

[0040] In this embodiment, the processing steps S2 to S5 are designed to realize the all-fiberized flexible body deformation inversion based on the wavelength residual neural network. Specifically:

[0041] S2: Use the shape data of the aerial refueling hose and the central wavelength offset of the distributed optical fiber sensor to establish a database, and use the shape data as the data label, where the label value is expressed as K =[ k1 ，k 2 ，k 3 ，k 4 ], determine the correspondence between the aerial refueling hose shape data label and the central wavelength offset of the distributed optical fiber sensor, and randomly divide each correspondence into a training set, a test set, and a validation set according to the proportion;

[0042] S3: Based on the ResNet18 residual neural network, a wavelength residual neural network block based on the wavelength data measurement of the distributed optical fiber sensor is constructed, and the training set and the test set are input into the ResNet18 neural network for training and learning, and the trained parameter model is saved;

[0043] S4: Select the ResNet18 neural network feature extraction layer as the feature extractor for aerial refueling hose morphology perception, and put the training set and test set into the feature extractor to extract features;

[0044] S5: Construct a multi-layer perceptron neural network, input the training set and test set features extracted by the ResNet18 neural network into the multi-layer perceptron for training, and after the training is completed, use the verification set to perform aerial refueling hose morphology perception verification.

[0045] The specific steps of establishing the database of the aerial refueling hose shape data and the center wavelength offset measured by the distributed optical fiber sensor on the outer rubber surface in S2 and dividing the corresponding relationship data set are:

[0046] S2.1), such as Figure 1 As shown in the figure, three distributed optical fiber sensor shape sensing paths are arranged on the outer rubber surface of the aerial refueling hose along the circumferential direction of 0°, 120°, and 240°. The deformation curve C of the aerial refueling hose with different shapes can be expressed by a cubic polynomial. C=k 1 x 3 +k 2 x 2 +k 3 x+k 4 Approximately, k 1 , k 2 , k 3 , k 4 Various adjustment factors constitute the aerial refueling hose shape data label matrix K =[ k 1 ，k 2 ，k3 ，k 4 ]. After the aerial refueling hose changes shape, the central wavelength offset data of the distributed optical fiber sensor is recorded: △ l 0 o =[△ l 0 o ,1 , △ l 0 o ,2 , △ l 0 o ,3 , …, △ l 0 o ,n ], △ l 120 o =[△ l 120 o ,1 , △ l 120 o ,2 , △ l 120 o ,3 , …, △ l 120 o ,n ], △ l 240 o =[△ l 240 o ,1 , △ l 240 o ,2 , △ l 240 o ,3 , …, △ l 240 o ,n ].

[0047] S2.2), according to the layout path of the distributed optical fiber sensor, the aerial refueling hose surface is unfolded into a two-dimensional plane, so that the central wavelength offset data measured by the three distributed optical fiber sensor paths are mapped to the corresponding positions in the two-dimensional plane, and the central wavelength offset data are interpolated into a 251×251 two-dimensional central wavelength offset numerical matrix by using cubic spline interpolation fitting. , and then copy the two-dimensional center wavelength offset numerical matrix into a 3×251×251 three-dimensional center wavelength offset numerical matrix As input data for the ResNet18 neural network.

[0048] S2.3), perform data enhancement operations of rotation, translation and scaling on the three-dimensional central wavelength offset numerical matrix, and under different curvatures, correspond to unique three-dimensional central wavelength offset numerical matrices to form a three-dimensional central wavelength offset numerical matrix database.

[0049] S2.4), randomly extract 80% training samples, 10% test samples and 10% validation samples from the distributed optical fiber sensor three-dimensional central wavelength offset numerical matrix database to create training set sample data, test set sample data and validation set sample data.

[0050] The specific steps of constructing the wavelength residual neural network block and modifying the number of connections of the fully connected layer of the ResNet18 neural network in S3 are:

[0051] S3.1), such as Figure 3 As shown in the figure, a ResNet18 residual neural network is constructed. The specific method is as follows: inside the ResNet18 neural network, the numerical matrix of the three-dimensional central wavelength offset of the distributed optical fiber sensor First, it passes through a convolutional layer, in which the convolution kernel size is 3×7×7, the output channel is 64, the stride is 2, and the padding is 3. After this convolutional layer, a 64×125×125 central wavelength offset numerical matrix △ is obtained. l [1] .

[0052] The PReLU activation function is included between the convolutional layers. , △λ represents the elements in the numerical matrix of the central wavelength offset of the distributed optical fiber sensor, where a is usually a number between 0 and 1 and can be adjusted through network training data. This activation function introduces nonlinearity into the neural network and improves the ability of the neural network to learn complex function mapping from data. The numerical matrix of the central wavelength offset obtained by the first convolutional layer △ l [1]After the PReLU activation function, it is input into the maximum pooling layer. The pooling kernel (Maxpooling Kernel) in the maximum pooling layer is 3×3, the step size is 2, and the padding is 1, and the output is 64×63×63. The numerical matrix of the center wavelength offset obtained by the maximum pooling layer is recorded as △ l [2] .

[0053] S3.2), the central wavelength offset numerical matrix △ obtained by the maximum pooling layer l [2] Input wavelength residual neural network block, such as Figure 3 As shown, the wavelength residual neural network blocks between the maximum pooling layer and the average pooling layer are divided into two types, which are distinguished by whether they contain convolution downsampling. For example, the first wavelength residual neural network block built after the maximum pooling layer and the second wavelength residual neural network block thereafter are residual blocks without convolution downsampling. The residual block without convolution downsampling includes two convolution layers and one residual calculation layer. It includes two convolution layers and one residual calculation layer. The first convolution layer has a convolution kernel size of 3×3, a step size of 1, and a padding of 1. The second convolution layer has the same structure as the first convolution layer, and the numerical matrix of the center wavelength offset output by the second layer is 64×63×63. The numerical matrix of the center wavelength offset output by the second convolution layer is denoted as F [1] (△λ), ,in , is the central wavelength offset numerical matrix F [1] The i-th row and j-th column element in (△λ); △ l [2] The i-th row and j-th column element in 1 , p 1 ,s 1 is the kernel size, padding and stride of the first convolutional layer in the wavelength residual neural network block; k 2 , p 2 ,s 2 are the convolution kernel size, padding and stride of the second convolution layer in the wavelength residual neural network block. The convolution kernel size, padding and stride of the first and second convolution layers of the three wavelength residual neural network blocks are the same.

[0054] S3.3), the output H of the first wavelength residual neural network block [1] (△λ), H [1] (△ l ) = F [1] (△ l )+△ l[2] The feature obtained by the convolution kernel in the wavelength residual neural network block is F [1] (△ l ) contains the features, namely the real output of the neural network and the numerical matrix of the central wavelength offset △ l [2] The difference, F [1] (△ l ) =H [1] (△ l )-△ l [2] The second wavelength residual neural network block has the same structure as the first wavelength residual neural network block. The central wavelength offset numerical matrix output by the second wavelength residual neural network block is denoted as H [2] (△λ). Among them, H [1] (△λ), H [2] (△λ), H [3] (△λ) represents the output of the first wavelength residual neural network block, the output of the second wavelength residual neural network block, and the output of the third wavelength residual neural network block, respectively. l [1] ,△ l [2] ,△ l [3] They represent the numerical matrix of the central wavelength offset after the first convolutional layer in the entire neural network (see Figure 3 ), the numerical matrix of the central wavelength offset after the first maximum pooling layer, the numerical matrix of the central wavelength offset after the first convolutional layer in the first wavelength residual neural network block, F [1] (△ l ) 、F [2] (△ l ) 、F [3] (△ l ) respectively represent the center wavelength offset numerical matrix output by the second convolution layer in the first wavelength residual neural network block, the center wavelength offset numerical matrix output by the second convolution layer in the second wavelength residual neural network block, and the center wavelength offset numerical matrix output by the second convolution layer in the third wavelength residual neural network block.

[0055] S3.4), such as Figure 3As shown, the third wavelength residual neural network block is constructed, in which the downsampled wavelength residual neural network block is introduced. Different from other wavelength residual neural network blocks, the central wavelength offset numerical matrix H [2] (△λ) passes through a convolutional layer with a kernel size of 1×1, a step size of 2, and a padding of 1, and is then combined with the outputs of the two convolutional layers in the residual neural network block to obtain the output of the third wavelength residual neural network block H [3] (△λ), H [3] (△ l ) = F [3] (△ l )+W*{ H [2] (△ l )}, where F [3] (△ l ) is H [2] (△ l )The numerical matrix of the center wavelength offset obtained by two convolutional layers in the third wavelength residual neural network block; W*{} represents the convolution operation.

[0056] In S4, the ResNet18 neural network feature extraction layer is selected as the feature extractor for aerial refueling hose morphology perception. The specific steps of putting the training set and the test set into the feature extractor to extract features are as follows:

[0057] S4.1), select the cross entropy loss function L ( k , y ) as a ResNet18 neural network performance optimization function, , where k i Label the curvature of the aerial refueling hose; i The ResNet18 neural network model predicts the probability that the input central wavelength offset numerical matrix belongs to the i-th curvature label of the aerial refueling hose.

[0058] S4.2), by finding the gradient of the cross entropy loss function , and back-propagate the gradient value of the loss function to update the parameters of the ResNet18 neural network. L represents the cross entropy loss function in S4.1). L ( k , y ). Taking the first wavelength residual neural network block as an example, according to the chain derivation rule, we get , where W represents the residual block convolution kernel parameter. From the formula in S3, we can know that H[1] (△ l )=F [1] (△ l )+△ l [2] , , , Substituting the three equations obtained from S3 into the chain rule, we can get ,in represents the gradient of the first convolutional layer, represents the gradient of the second convolutional layer. Thus, the loss function gradient can be expressed as .

[0059] S4.3), the Adam gradient descent back-propagation optimization algorithm is used to update the network parameters, that is, the parameter value W in the convolution kernel.

[0060] S4.3.1), compute the first-order moment estimate of the gradient, w t =β 1 w t-1 + (1 -β 1 ) ▽J ( W t ),in w t is the first-order moment estimate of the t-th iteration, ▽J ( W t ) is the gradient of the tth iteration, t is the number of iterations, β 1 is the decay rate of the first-order moment estimate (usually set to 0.9).

[0061] S4.3.2), compute the second-order moment estimate of the gradient, v t =β 2 v t-1 + (1 -β 2 ) ( ▽J ( W t )) 2 ,in v t is the second-order moment estimate of the t-th iteration, β 2 is the decay rate of the second-order moment estimate (usually set to 0.999).

[0062] S4.3.3), compute the bias correction for the first moment estimate, ; and bias correction of the second-order moment estimates, .

[0063] S4.3.4), update the network parameters by gradient descent back propagation, ,in, represents the bias correction of the first-order moment estimate, represents the bias correction of the second-order moment estimate, W t+1 It is expressed as the network parameters of t+1 iterations, a is the learning rate, and e is a small constant used to prevent the denominator from being zero (usually set to 10 -8 ).

[0064] During the training process, if the performance of the model on the test set (such as accuracy or loss value) does not improve significantly within several consecutive iterations, this may indicate that the model has approached or reached the optimal state, and training can be stopped at this time.

[0065] In S5, a multi-layer perceptron neural network is constructed, and the extracted training set and test set features are input into the multi-layer perceptron for training. After the training is completed, the specific steps of using the verification set to perform the aerial refueling hose morphology perception verification are as follows:

[0066] S5.1), there is a nonlinear mapping relationship between the aerial refueling hose shape curve C and the central wavelength offset D1 of the distributed optical fiber sensor. Since the multilayer perceptron is a multilayer feedforward neural network that uses a loss function back propagation algorithm and can learn the nonlinear mapping relationship between input and output, the multilayer perceptron is used to solve the nonlinear mapping problem.

[0067] like Figure 4 As shown in the figure, a multilayer perceptron neural network is constructed. The multilayer perceptron neural network consists of four layers: 1 input layer, 2 hidden layers, and 1 output layer. The two hidden layers and the second hidden layer and the output layer are connected using the tansig activation function. The tansig activation function is expressed as , where f (x') represents the tansig activation function and x' represents the element in the feature length vector input to the multilayer perceptron. The size is 1×512, so the input layer of the multilayer perceptron has 512 neuron nodes; the number of neuron nodes in the first hidden layer is 256; the number of neuron nodes in the second hidden layer is 128; the output layer contains 4 neuron nodes, corresponding to the four adjustment factors in the shape curve C of the aerial refueling hose.

[0068] S5.2), in the multi-layer perceptron, the input feature long vector Output after the first hidden layer , p’ 1 Output after the second hidden layer p’ 2 =tan sig [ oh 2 p’ 1 + b 2 ], p’ 2 After passing through the output layer, the final output K = oh 3 p’ 2 + b 3 , oh 1 represents the weights in the first hidden layer, oh 2 represents the weights in the second hidden layer, oh 3 represents the weights in the output layer, b 1 represents the bias in the first hidden layer, b 2 represents the bias in the second hidden layer, b 3 represents the bias in the output layer.

[0069] Flatten the training set and test set features learned by the last convolutional layer in the ResNet18 neural network into a long feature vector And the corresponding shape data label matrix K =[ k 1 ，k 2 ，k 3 ，k 4 ] is input into the multi-layer perceptron for training the multi-layer perceptron neural network.

[0070] S5.3), input the feature vector of the aerial refueling hose validation set into the trained multi-layer perceptron neural network model, and obtain the shape data label matrix output by the neural network K’ =[ k’ 1 ，k’ 2 ，k’ 3 ，k’ 4 ]Thus, the deformation curve of the refueling hose obtained by neural network inversion is obtained.C = k’ 1 x 3 + k’ 2 x 2 + k’ 3 x + k’ 4 , x Represents the transverse coordinate value in the deformation curve.

[0071] The shape data label matrix obtained by the neural network K’ Label matrix with known actual hose deformation shape data K Find the relative error , which is used to verify the aerial refueling hose morphology inversion method based on distributed fiber optic sensors and wavelength residual neural network.

[0072] The main advantages of this embodiment are: first, it is used to solve the problem that the existing aerial refueling visual monitoring system is difficult to capture the deformation of the refueling hose and is greatly affected by the weather environment; second, it is used to solve the problem that the sensor-perceived strain caused by the superelasticity of the outer rubber of the refueling hose is difficult to correlate and map with the hose morphological characteristics. In practical applications, the deformation of the aerial refueling hose during the refueling process can be fed back to the UAV operating system interface in real time, and the aerial refueling hose morphology can be visualized, which can not only provide reliable information for UAV operators to perform refueling tasks, increase the execution efficiency of aerial refueling tasks, and ensure the flight safety of UAVs, but also provide key technical support for the establishment of a "new generation of digital twin systems for refueling aircraft" and the realization of intelligent control of the refueling process.

[0073] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The above is only a specific implementation method of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or replacements that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for optical fiber flexible body morphology inversion based on wavelength residual neural network, characterized in that: include: S1. Install distributed fiber optic sensors in the aerial refueling hose and record the position of each sensor on the aerial refueling hose; S2, collecting shape data of the aerial refueling hose and using it as a data label, and recording the correspondence between the data label and the central wavelength offset of the distributed optical fiber sensor according to the user operation, wherein a sample database is established using the recorded correspondence, and the sample database includes: a training set, a test set, and a validation set; S3, constructing a wavelength residual neural network block based on wavelength data measurement of a distributed optical fiber sensor, and inputting the training set and the test set into the wavelength residual neural network block for training; S4, using the feature extraction layer of the wavelength residual neural network as a feature extractor and training it, wherein the trained feature extractor is used to extract features of the numerical matrix of the central wavelength offset of the distributed optical fiber sensor; S5, constructing a neural network of a multi-layer perceptron, inputting the features of the training set and the test set extracted by the feature extractor into the multi-layer perceptron and training the multi-layer perceptron, and the verification set is used to verify the trained multi-layer perceptron; S6. Use the trained multi-layer perceptron to invert the actual deformation of the aerial refueling hose.

2. The method according to claim 1, characterized in that S1 includes: On the outer rubber surface of the aerial refueling hose, three distributed optical fiber sensors are arranged along the circumferential angles of 0°, 120° and 240° respectively.

3. The method according to claim 1 or 2, characterized in that S2 include: The aerial refueling hose is set to different shapes, and the shape data of the aerial refueling hose under each shape and the central wavelength offset data of the distributed optical fiber sensor are collected, wherein the shape data includes a deformation curve corresponding to one shape C=k 1 x 3 +k 2 x 2 +k 3 x+k 4 , the label value of the data label is expressed as K =[ k 1 ，k 2 ，k 3 ，k 4 ], k 1 ~k 4 They represent the four adjustment factors in the aerial refueling hose shape curve C respectively; the central wavelength offset data of the distributed optical fiber sensor collected include: λ 0 o =[△ λ 0 o ,1 , △ λ 0 o ,2 , △ λ 0 o ,3 , …, △ λ 0 o ,n ], △ λ 120 o =[△ λ 120 o ,1 , △ λ 120 o ,2 , △ λ 120 o ,3 , …, △ λ 120 o ,n ], △ λ 240 o =[△ λ 240 o ,1 , △ λ 240 o ,2 , △ λ 240 o ,3 , …, △ λ 240 o ,n ], where △ λ 0 o It represents the central wavelength offset of the distributed optical fiber sensor arranged along the 0° direction of the hose circumference, △ λ 120 o It represents the central wavelength offset of the distributed optical fiber sensor arranged along the circumference of the hose at 120°, △ λ 240 o It represents the central wavelength offset of the distributed optical fiber sensor arranged along the 240° direction of the hose circumference, △ λ 0 o ,n represents the nth data collection point in the distributed optical fiber sensor arranged along the 0° direction of the hose circumference, △ λ 120 o ,n represents the nth data collection point in the distributed optical fiber sensor arranged along the circumference of the hose at 120°, △ λ 240 o ,n represents the nth data collection point in the distributed optical fiber sensor arranged along the 240° direction of the hose circumference. n Indicates the number of data collection points in the distributed optical fiber sensor; The aerial refueling hose curved surface is unfolded into a two-dimensional plane, and the central wavelength offset data of the distributed optical fiber sensor is mapped to the two-dimensional plane to obtain a two-dimensional central wavelength offset numerical matrix, and then a three-dimensional central wavelength offset matrix is ​​continuously obtained and stored in the sample database.

4. The method according to claim 1, characterized in that: Prior to S3, this included: In the ResNet18 neural network, the three-dimensional center wavelength offset matrix of the distributed optical fiber sensor is passed through a convolution layer to obtain a center wavelength offset numerical matrix of size 64×125×125△ λ [1] , where the convolution kernel size of the convolution layer is 3×7×7, the output channel is 64, the stride is 2, and the padding is 3; △ λ [1] After the PReLU activation function, it is input into the maximum pooling layer to obtain a central wavelength offset numerical matrix of size 64×63×63 △ λ [2] , where the pooling kernel size of the maximum pooling layer is 3×3, the stride is 2, and the padding is 1.

5. The method according to claim 4, characterized in that S3 includes: Two types of wavelength residual neural network blocks are constructed, wherein the first type of wavelength residual neural network block includes: two convolutional layers with the same structure and one residual calculation layer, wherein the structures of the two convolutional layers with the same structure are: the convolution kernel size is 3×3, the step size is 1, and the padding is 1; The structure of the second wavelength residual neural network block includes: three convolutional layers, where the convolution kernel size of the first convolutional layer is 1×1, the output channel is 64, the stride is 2, and the padding is 1; the structures of the second and third convolutional layers are the same: the convolution kernel size is 3×3, the stride is 1, and the padding is 1.

6. The method according to claim 5, characterized in that The process of inputting the training set and the test set into the wavelength residual neural network block for training includes: △ λ [2] Input the first wavelength residual neural network block, and pass through two convolutional layers of the first wavelength residual neural network block, where the output of the second convolutional layer is represented by F [1] (△ λ ), F [1] (△ λ ) is the element in row i and column j. , , among which , The numerical matrix △ represents the central wavelength offset of the output of the first convolutional layer in the first wavelength residual neural network block λ [3] The i-th row and j-th column element in is the central wavelength offset numerical matrix △ λ [2] The i-th row and j-th column element in ; k 1, p 1, s 1 is the convolution kernel size, padding and stride of the first convolution layer in the first wavelength residual neural network block, k 2, p 2, s 2 is the convolution kernel size, padding and stride of the second convolution layer in the first wavelength residual neural network block; the output H of the first wavelength residual neural network block [1] (△ λ ) = F [1] (△ λ )+△ λ [2] ; H [1] (△ λ ) inputs the second wavelength residual neural network block, the output of the second wavelength residual neural network block is H [2] (△ λ ); Then H [2] (△ λ ) Input the third wavelength residual neural network block, the output of the third wavelength residual neural network block is H [3] (△ λ ), H [3] (△ λ ) = F [3] (△ λ )+W*{ H [2] (△ λ )},F [3] (△ λ ) is H [2] (△ λ ) is a numerical matrix of center wavelength offset obtained by two convolutional layers in the third wavelength residual neural network block, where W*{} represents a convolution operation.

7. The method according to claim 1, characterized in that In S4, the cross entropy loss function is used in the process of training the feature extractor L ( k , y ) as the performance optimization function, where , k is the regulating factor, y is the probability parameter, k i The adjustment factor label corresponding to the aerial refueling hose, y i The probability that the input central wavelength offset numerical matrix belongs to the i-th adjustment factor label of the aerial refueling hose is predicted by the ResNet18 neural network model, where i is the number of the adjustment factor label and C is the maximum value of the number of the adjustment factor label; The cross entropy loss function is used and the Adam gradient descent back propagation optimization algorithm is adopted to update the network parameters, where the first-order moment of the gradient is estimated as: w t =β 1 w t-1 + (1 -β 1) ▽J ( W t ), the second-order moment of the gradient is estimated as: v t =β 2 v t-1 + (1 - β 2) ( ▽J ( W t )) 2 , w t is the first-order moment estimate of the t-th iteration, ▽J ( W t ) is the gradient of the tth iteration, t is the number of iterations, β 1 is the decay rate of the first-order moment estimate, W t represents the network parameters of the tth iteration, v t is the second-order moment estimate of the t-th iteration, β 2 is the decay rate of the second-order moment estimate.

8. The method according to claim 7, characterized in that The Adam gradient descent back propagation optimization algorithm is used to update the network parameters, including: ,in, represents the bias correction of the first-order moment estimate, represents the bias correction of the second-order moment estimate, W t+1 It is represented as the network parameters of t+1 iterations, a is the learning rate, and e is a constant.

9. The method according to claim 3, characterized in that: In S5, the architecture of the neural network of the constructed multi-layer perceptron includes: 1 input layer, 2 hidden layers and 1 output layer; The connection between the two hidden layers and the connection between the second hidden layer and the output layer all use the tansig activation function; The number of neuron nodes in the first hidden layer is 256, and the number of neuron nodes in the second hidden layer is 128; the output layer contains 4 neuron nodes, which correspond to the four adjustment factors in the shape curve C of the aerial refueling hose. k 1 ， k 2 ，k 3 ，k 4 .

Images