Helicopter system positive and inverse model establishment method based on reversible neural network

Abstract

The invention discloses a helicopter system positive and inverse model establishment method based on a reversible neural network, and the method comprises the steps: collecting the input data and output data of a helicopter nonlinear system according to the nonlinear characteristics of the helicopter system, and constructing a training set and a test set; respectively establishing a forward discrete model and a reverse discrete model of the nonlinear object, and designing a reversible neural network model; constructing a loss function according to the forward and reverse discrete models of the nonlinear object and a reversible neural network model; training a reversible neural network model according to the loss function, and obtaining a forward model and a reverse model of the system based on the trained reversible neural network model; and verifying the accuracy of the forward model and the reverse model of the helicopter nonlinear system. According to the method, the forward model and the reverse model of the nonlinear system of the helicopter can be directly and simply established at the same time through the reversible neural network.

Description

Technical Field

[0001] The invention belongs to the field of system modeling, in particular to a method for establishing a forward and inverse model of a helicopter system based on a reversible neural network. Background Art

[0002] System modeling is widely used in system control and process optimization in fields such as chemistry, mechanics, electrical engineering, and civil engineering. The models of some systems can be expressed in linear parameter form, and adaptive control can be applied to achieve output tracking, online identification and estimation of unknown systems. And system modeling and inverse modeling are important components in applications such as adaptive control, channel equalization and active noise control, disturbance elimination, etc. However, in practical engineering, the models of most systems are not linear, and their nonlinear characteristics must be considered when performing measurements or designing and calibrating measuring instruments. In the past few decades, a lot of research has focused on system modeling and identification of nonlinear systems.

[0003] However, the inverse modeling of complex nonlinear systems such as helicopters still faces great challenges. Currently, there are two possible methods for implementing inverse modeling of nonlinear systems: direct inverse modeling and indirect inverse modeling. Both direct inverse modeling and indirect inverse modeling methods can be implemented through neural networks. The implementation method of the direct inverse modeling method is to provide input-output samples to the neural network learning algorithm in reverse order, where the system output samples are used as network inputs and the system input samples are used as network expected outputs. Using neural networks for direct inverse modeling is fast because it approximates the inverse without actually inverting the nonlinear model. However, when the underlying mapping is not reversible, it may lead to incorrect learning. The implementation method of the indirect inverse modeling method is to first estimate the forward model and then build an inverse model of the estimated forward model. The indirect inverse modeling method provides greater flexibility and generally provides better performance. However, the indirect inverse modeling method needs to be carried out step by step, resulting in more time costs. Therefore, there is an urgent need in the art for a method for establishing a forward and inverse model of a helicopter system based on a reversible neural network.

[0004] The above information disclosed in this Background section is only for enhancement of understanding of the background of the invention and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention proposes a method for establishing a forward and inverse model of a helicopter system based on a reversible neural network. The present invention can model the nonlinear system of a helicopter through a reversible neural network, and can obtain the inverse model of the helicopter system while establishing the forward model of the helicopter system. Therefore, the method for establishing the forward and inverse model of a helicopter system based on a reversible neural network is direct and simple.

[0006] The purpose of the present invention is achieved through the following technical solutions.

[0007] A method for establishing a forward and inverse model of a helicopter system based on a reversible neural network comprises the following steps:

[0008] In the first step, according to the nonlinear characteristics of the helicopter system, the input and output data of the helicopter nonlinear system are collected to construct a training set and a test set;

[0009] In the second step, the forward and reverse discrete model expressions of the nonlinear object are established according to the collected input and output data of the helicopter nonlinear system, and the number and basic structure of the input and output variables of the reversible neural network model are designed;

[0010] In the third step, according to the forward and reverse discrete models and the reversible neural network model of the nonlinear object, the loss function is constructed with the actual sampling value as the reference value and the output value of the reversible neural network model as the estimated value;

[0011] In the fourth step, according to the loss function, the reversible neural network model is trained using the training set data using the Adam algorithm to obtain a trained reversible neural network model, a forward model of the helicopter nonlinear system is established based on the forward process of the trained reversible neural network model, and a reverse model of the helicopter nonlinear system is established based on the reverse process of the trained reversible neural network model;

[0012] In the fifth step, the data of the test set are respectively forward input into the helicopter nonlinear system model and reverse input into the helicopter nonlinear system inverse model to verify the accuracy of the helicopter nonlinear system forward model and the helicopter nonlinear system inverse model, so as to eliminate noise and disturbance in the system output.

[0013] In the method described, in the first step, according to the complex nonlinear characteristics of the helicopter system, acceleration sensors are installed at the input and output ends where the model needs to be established, the helicopter system is started or continuous vibration excitation is applied to the helicopter system, and the input and output vibration data of the helicopter nonlinear object are collected through the sensors installed at the input and output ends. Part of the collected data is used as a training set, and the remaining data is used as a training set.

[0014] In the method described above, in the second step, the forward and reverse discrete model expressions of the helicopter object are established according to the collected input and output data of the helicopter nonlinear system, and the number and basic structure of the input and output variables of the reversible neural network model are designed, as follows:

[0015] First, the forward discrete model expression of the helicopter nonlinear system is established:

[0016] y t =f(xt , ..., x t-a ;y t-1 , ..., y t-b )

[0017] Among them, y t represents the system output value at time t, f represents the forward function of the helicopter nonlinear system, x t , ..., x t-a represents the historical system input value at time t and before, y t-1 , ..., y t-b Represents the historical system output value at time t-1 and before.

[0018] Then, the inverse discrete model expression of the helicopter nonlinear object is established:

[0019] x t =g(x t-1 , ..., x t-a ;y t , ..., y t-b )

[0020] Among them, x t represents the system input value at time t, g represents the inverse function of the helicopter nonlinear system, x t-1 , ..., x t-a represents the historical system input value at time t-1 and before, y t , ..., y t-b Represents the historical system output value at time t and before.

[0021] Then design the number and basic structure of the input and output variables of the reversible neural network model. First, the basic unit of the reversible neural network is a reversible block, which consists of two complementary affine coupled layers. Here, the input vector u of the block is divided into two parts, u1 and u2, which are composed of a coefficient exp(s i ) and t i Affine function transformation of (i∈{1,2}) using element-wise multiplication (⊙) and addition:

[0022] v1=u1⊙exp(s2(u2))+t2(u2),

[0023] v2=u2⊙exp(s1(v1))+t1(v1)

[0024] Given an output v = [v1, v2], the expression is trivially reversible:

[0025] u2=(v2-t1(v1))⊙exp(-s1(v1)),

[0026] u1=(v1-t2(u2))⊙exp(-s2(u2))

[0027] Maps i and t i It can be any complex function of v1 and v2, which are not necessarily reversible. A reversible neural network is composed of a series of reversible blocks. Therefore, the network is reversible and can have both forward and reverse operations in the same network. The forward and reverse operations of a reversible neural network are expressed as follows:

[0028] v = for (u, s, t),

[0029] u=back(v,s,t).

[0030] Among them, v represents the forward output and reverse input of the deep reversible network, u represents the forward input and reverse output of the deep reversible network, s and t represent the set of all reversible block parameters in the deep reversible network, for represents the forward function of the deep reversible network, and back represents the reverse function of the deep reversible network.

[0031] In addition, to increase the model capacity, permutation layers are inserted between reversible blocks, which shuffle the elements of the subsequent layer input in a random but fixed way. This causes the split u = [u1, u2] to change between layers and enhances the interaction between individual variables. The number of reversible blocks and the structure of the reversible blocks are adjusted according to the complexity of the helicopter nonlinear system.

[0032] In the method described above, in the third step, the training set data is used to train the built reversible neural network model, and the actual sampling value is used as the reference value, and the output value of the reversible neural network model is used as the estimated value to construct the loss function. The reversible neural network is used to establish the forward and inverse models of the helicopter nonlinear system, so the loss function is composed of two parts, and its expression is as follows:

[0033] L total =L for +L back

[0034] Among them, L total Represents the total loss function of the reversible neural network model, L for and L back are the forward and reverse mean square error loss functions of the reversible neural network model respectively. L for It is the mean of the sum of squares of the corresponding point errors between the forward output result v of the reversible neural network and the actual system output sampling value y, and its expression is as follows:

[0035]

[0036] Where MSE(·) represents mean square error and n represents the number of samples. back The mean of the sum of squares of the corresponding point errors between the inverse output result u of the reversible neural network and the actual system input sampling value x is expressed as follows:

[0037]

[0038] In the method described above, in the fourth step, according to the loss function, the reversible neural network model is trained using the Adam algorithm using the training set data, and the forward model of the helicopter nonlinear system is established based on the forward process of the trained reversible neural network; the inverse model of the helicopter nonlinear system is established based on the inverse process of the trained reversible neural network, as follows:

[0039] The reversible network performs forward and backward iterations in an alternating manner, accumulating gradients from two directions before performing parameter updates, providing the opportunity to optimize both forward and reverse losses, thus enabling more efficient training. The network training and parameter updates are performed using the Adam algorithm, which is a stochastic optimization method with adaptive momentum. It is often used as an optimizer algorithm in deep learning and has great advantages over other types of stochastic optimization algorithms.

[0040] In the method described, in the fifth step, the data of the test set are respectively forward input into the helicopter nonlinear system model and reverse input into the helicopter nonlinear system inverse model to verify the accuracy of the helicopter nonlinear system forward model and the helicopter nonlinear system inverse model, so as to eliminate noise and disturbance in the system output.

[0041] The data of the test set is forward input into the reversible neural network. The forward model of the helicopter nonlinear system established based on the forward process of the reversible neural network is expressed as:

[0042]

[0043] in, represents the output value of the helicopter nonlinear system forward model established by the reversible neural network forward process at time t, The function representing the forward model of the helicopter nonlinear system established by the forward process of the reversible neural network, represents the historical system input value before and at time t in the test set, Represents the historical system output value at time t-1 and before in the test set;

[0044] The inverse model expression of the helicopter nonlinear system established by the inverse process of the reversible neural network is:

[0045]

[0046] in, represents the output value of the helicopter nonlinear system inverse model built by the reversible neural network inverse process at time t, The function representing the inverse model of the helicopter nonlinear system built by the inverse process of the reversible neural network, represents the historical system input value before and after time t-1 in the test set, Represents the historical system output value before and after time t in the test set.

[0047] In the fifth step (S5), the accuracy of the helicopter nonlinear system forward model established by the reversible neural network forward process and the helicopter nonlinear system inverse model established by the reverse process is represented by the error between the output of the respective models and the actual collected data;

[0048] Among them, the error of the helicopter nonlinear system forward model established by the reversible neural network forward process is expressed as:

[0049]

[0050] in, represents the error between the output value of the helicopter nonlinear system forward model established by the forward process in the test set and the true value of the system output, Represents the output value of the model built by the forward process of the reversible neural network, y test represents the true value of the system output in the test set;

[0051] The error of the inverse model of the helicopter nonlinear system established by the inverse process of the reversible neural network is expressed as:

[0052]

[0053] in, It represents the error between the output value of the inverse model of the helicopter nonlinear system built by the inverse process in the test set and the true value of the system input, represents the output value of the helicopter nonlinear system inverse model built by the inverse process of the reversible neural network, x test Represents the true value of the system input in the test set.

[0054] Compared with the prior art, the present invention has the following technical effects:

[0055] The method for establishing the forward and inverse model of the helicopter system based on the reversible neural network only needs to train a reversible neural network to achieve the corresponding inverse model establishment while establishing the nonlinear forward model of the helicopter. This method avoids the problem of the existing direct inverse modeling and indirect inverse modeling methods that require training two networks separately to establish two models in a step-by-step and time-consuming manner. Therefore, the method for establishing the forward and inverse model of the helicopter system based on the reversible neural network is direct and simple. Therefore, the method for establishing the forward and inverse model of the helicopter system based on the reversible neural network is direct and simple.

[0056] The above description is only an overview of the technical solution of the present invention. In order to make the technical means of the present invention clearer and to enable those skilled in the art to implement it according to the contents of the specification, and to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described below by way of example. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] By reading the detailed description of the preferred specific embodiments below, various other advantages and benefits of the present invention will become clear to those of ordinary skill in the art. The drawings in the specification are only for the purpose of illustrating the preferred embodiments and are not considered to be limitations of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without creative work. Moreover, the same reference numerals are used to represent the same components throughout the drawings.

[0058] In the attached picture:

[0059] Figure 1 It is a schematic diagram of the steps of a method for establishing a forward and inverse model of a helicopter system based on a reversible neural network according to an embodiment of the present invention;

[0060] Figure 2 is a data set schematic diagram of a method for establishing a forward and inverse model of a helicopter system based on a reversible neural network according to an embodiment of the present invention;

[0061] Figure 3(a) to Figure 3(c) A neural network reversible block structure and an overall schematic diagram of a reversible neural network of a method for establishing a forward and inverse model of a helicopter system based on a reversible neural network according to an embodiment of the present invention;

[0062] FIG. 4( a ) and FIG. 4( b ) are diagrams showing training results of a method for establishing a forward and inverse model of a helicopter system based on a reversible neural network according to an embodiment of the present invention;

[0063] FIG. 5( a ) and FIG. 5( b ) are test result diagrams of a method for establishing a forward and inverse model of a helicopter system based on a reversible neural network according to an embodiment of the present invention.

[0064] The present invention is further explained below in conjunction with the accompanying drawings and embodiments. DETAILED DESCRIPTION

[0065] The following will refer to the attached Figure 1 to Figure 5(b) Specific embodiments of the present invention are described in more detail. Although specific embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0066] It should be noted that certain words are used in the specification and claims to refer to specific components. Those skilled in the art should understand that technicians may use different nouns to refer to the same component. This specification and claims do not use the difference in nouns as a way to distinguish components, but use the difference in the functions of the components as the criterion for distinction. As mentioned throughout the specification and claims, "including" or "comprising" is an open term, so it should be interpreted as "including but not limited to". The subsequent description of the specification is a preferred embodiment of the present invention, but the description is based on the general principles of the specification and is not intended to limit the scope of the present invention. The scope of protection of the present invention shall be determined by the attached claims.

[0067] To facilitate understanding of the embodiments of the present invention, further explanation will be given below by taking specific embodiments as examples in conjunction with the accompanying drawings, and each of the accompanying drawings does not constitute a limitation on the embodiments of the present invention.

[0068] In one embodiment, Figure 1 The schematic diagram of the steps of the method for establishing the forward and inverse model of a helicopter system based on a reversible neural network is shown in FIG. The method for establishing the forward and inverse model of a helicopter system based on a reversible neural network comprises the following steps:

[0069] In the first step S1, according to the complex nonlinear characteristics of the helicopter system, acceleration sensors are installed at the input and output ends where the model needs to be established, the helicopter system is started or continuous vibration excitation is applied to the helicopter system, and the input and output vibration data of the helicopter nonlinear object are collected through the sensors installed at the input and output ends. Part of the collected data is used as a training set, and the remaining data is used as a training set.

[0070] The rotating parts of the helicopter generate centrifugal excitation force, forcing the body to vibrate. The main component of the helicopter's rotating parts is the main reducer, so acceleration sensors are installed on the main reducer and the body respectively, and the operation of the helicopter power system brings continuous excitation. The installed sensors collect the input and output vibration data of the helicopter's nonlinear objects respectively, and part of the collected data is used as a training set, and the remaining data is used as a test set.

[0071] In the second step S2, the forward and inverse discrete model expressions of the nonlinear objects between the main reducer and the body of the helicopter are established according to the collected input and output data of the helicopter nonlinear system, and the number and basic structure of the input and output variables of the reversible neural network model are designed, as follows:

[0072] First, the forward discrete model expression of the helicopter nonlinear object is established:

[0073] y t =f(x t , ..., x t-a ;y t-1 , ..., y t-b )

[0074] Among them, y t represents the system output value at time t, f represents the forward function of the helicopter nonlinear system, x t , ..., x t-a represents the historical system input value at time t and before, y t-1 , ..., y t-b represents the historical system output value at time t-1 and before, a and b represent the maximum order of input delay and output delay respectively.

[0075] Then, the inverse discrete model expression of the helicopter nonlinear object is established:

[0076] x t =g(x t-1 , ..., x t-a ;y t , ..., y t-b )

[0077] Among them, x t represents the system input value at time t, g represents the inverse function of the helicopter nonlinear system, x t-1 , ..., x t-a represents the historical system input value at time t-1 and before, y t , ..., y t-b Represents the historical system output value at time t and before.

[0078] According to the forward discrete model expression established above, the forward process of the reversible neural network model at time t can be determined, with x t , ..., x t-a ;y t-1 , ..., y t-b As the positive input of the reversible neural network model, y t As the forward output target value of the reversible neural network model. According to the above established reverse discrete model expression, the reverse process of the reversible neural network model at time t can be determined, with x t-1 , ..., x t-a ;y t , ..., y t-b As the reverse input of the reversible neural network model, x t As the reverse output target value of the reversible neural network model.

[0079] First, the basic unit of the reversible neural network is a reversible block, which consists of two complementary affine coupled layers. Here, the input vector u of the block is divided into two parts, u1 and u2, which are represented by a coefficient exp(s i ) and t i Affine function transformation of (i∈{1,2}) using element-wise multiplication (⊙) and addition:

[0080] v1=u1⊙exp(s2(u2))+t2(u2),

[0081] v2=u2⊙exp(s1(v1))+t1(v1)

[0082] Given an output v = [v1, v2], the expression is trivially reversible:

[0083] u2=(v2-t1(v1))⊙exp(-s1(v1)),

[0084] u1=(v1-t2(u2))⊙exp(-s2(u2))

[0085] Importantly, mappings i and t i It can be any complex function of v1 and v2, which are not necessarily reversible. A reversible neural network is composed of a series of reversible blocks. Therefore, the network is reversible and can have both forward and reverse operations in the same network. The forward and reverse operations of a reversible neural network are expressed as follows:

[0086] v = for (u, s, t),

[0087] u=back(v,s,t).

[0088] Among them, v represents the forward output and reverse input of the deep reversible network, u represents the forward input and reverse output of the deep reversible network, s and t represent the set of all reversible block parameters in the deep reversible network, for represents the forward function of the deep reversible network, and back represents the reverse function of the deep reversible network.

[0089] In addition, in order to increase the model capacity, a shuffle permutation layer is inserted between the reversible blocks, which shuffles the elements of the subsequent layer input in a random but fixed manner, and the input vector u = [u1, u2] changes between the permutation layers, and the interaction between the variables is enhanced. According to the number of reversible blocks designed for the helicopter nonlinear system and the structure of the reversible blocks, three reversible blocks are used in the example, and the nonlinear function in each reversible block uses two fully connected layers and a ReLU activation function.

[0090] In the third step S3, according to the discrete model expression and the reversible neural network model, the actual sampling value is used as the reference value, and the output value of the reversible neural network model is used as the estimated value to construct a loss function. The forward and inverse models of the helicopter nonlinear system are established using the reversible neural network, so the constructed loss function consists of two parts, L for and L back are the forward and reverse mean square error loss functions of the reversible neural network model respectively. L for It is the mean of the sum of squares of the corresponding point errors between the forward output result v of the reversible neural network and the actual system output sampling value y, and its expression is as follows:

[0091]

[0092] Where MSE(•) represents mean square error, n represents the number of samples, and i represents the i-th sample. back It is the mean of the sum of squares of the corresponding point errors between the inverse output result u of the reversible neural network and the actual system input sampling value x, and its expression is as follows:

[0093]

[0094] In the fourth step S4, according to the loss function, the reversible neural network model is trained using the training set data using the Adam algorithm to obtain a trained reversible neural network model, a helicopter nonlinear system forward model is established based on the forward process of the trained reversible neural network model, and a helicopter nonlinear system reverse model is established based on the reverse process of the trained reversible neural network model. The details are as follows:

[0095] The structure of the reversible neural network determines that it can forward fit the nonlinear function and reverse fit the inverse function of the nonlinear function through the same set of parameters. Therefore, the forward process of the same reversible neural network can be used to establish the helicopter nonlinear system model, and its reverse process can be used to establish the helicopter nonlinear system inverse model. The input true value in the training set data is forward input into the reversible neural network, that is, as the forward process input of the reversible neural network, and then the forward loss function is obtained according to the network forward process output and the output true value. According to the forward loss function, the error of the forward fitting of the reversible neural network is obtained, and the gradient of the reversible neural network in this direction is back-propagated with this error. Then the output true value in the training set data is reversely input into the reversible neural network, that is, as the reverse process input of the reversible neural network, and then the reverse loss function is obtained according to the network reverse process output and the input true value. According to the reverse loss function, the error of the reverse fitting of the reversible neural network is obtained, and the gradient of the reversible neural network in this direction is back-propagated with this error. Therefore, the reversible neural network performs forward and reverse iterations in an alternating manner. Accumulating gradients from two directions provides an opportunity to optimize the forward and reverse losses at the same time, and then the Adam algorithm is used to update the parameters. The training process is continued until the forward and reverse losses converge to obtain a trained reversible neural network model, wherein the forward process can fit the helicopter nonlinear system model, and the reverse process can fit the helicopter nonlinear system inverse model.

[0096] Adam is a stochastic optimization method with adaptive momentum, which is often used as an optimizer algorithm in deep learning. Compared with other types of stochastic optimization algorithms, it has better convergence advantages in reversible neural network training.

[0097] The data of the test set are forward input into the helicopter nonlinear system model and reverse input into the helicopter nonlinear system inverse model respectively to verify the accuracy of the forward model and inverse model of the helicopter nonlinear system established based on the reversible neural network.

[0098] In the method described, in the fifth step, the data of the test set are respectively forward input into the helicopter nonlinear system model and reverse input into the helicopter nonlinear system inverse model to verify the accuracy of the helicopter nonlinear system forward model and the helicopter nonlinear system inverse model, so as to eliminate noise and disturbance in the system output.

[0099] The data of the test set is forward input into the reversible neural network. The forward model of the helicopter nonlinear system established based on the forward process of the reversible neural network is expressed as:

[0100]

[0101] in, represents the output value of the helicopter nonlinear system forward model established by the reversible neural network forward process at time t, The function representing the forward model of the helicopter nonlinear system established by the forward process of the reversible neural network, represents the historical system input value before and at time t in the test set, Represents the historical system output value at time t-1 and before in the test set;

[0102] The inverse model expression of the helicopter nonlinear system established by the inverse process of the reversible neural network is:

[0103]

[0104] in, represents the output value of the helicopter nonlinear system inverse model built by the reversible neural network inverse process at time t, The function representing the inverse model of the helicopter nonlinear system built by the inverse process of the reversible neural network, represents the historical system input value before and after time t-1 in the test set, Represents the historical system output value before and after time t in the test set.

[0105] In the fifth step (S5), the accuracy of the helicopter nonlinear system forward model established by the reversible neural network forward process and the helicopter nonlinear system inverse model established by the reverse process is represented by the error between the output of each model and the actual collected data;

[0106] Among them, the error of the helicopter nonlinear system forward model established by the reversible neural network forward process is expressed as:

[0107]

[0108] in, represents the error between the output value of the helicopter nonlinear system forward model established by the forward process in the test set and the true value of the system output, Represents the output value of the model built by the forward process of the reversible neural network, y test represents the true value of the system output in the test set;

[0109] The error of the inverse model of the helicopter nonlinear system established by the inverse process of the reversible neural network is expressed as:

[0110]

[0111] in, It represents the error between the output value of the inverse model of the helicopter nonlinear system built by the inverse process in the test set and the true value of the system input, represents the output value of the helicopter nonlinear system inverse model built by the inverse process of the reversible neural network, x testRepresents the true value of the system input in the test set.

[0112] The method of the present invention is further described below in conjunction with the accompanying drawings.

[0113] Figure 1 The present invention is a flowchart of a method for establishing a forward and inverse model of a helicopter system based on a reversible neural network according to an embodiment of the present invention. First, according to the nonlinear characteristics of the helicopter system, the input and output data of the helicopter system are collected to construct a training set and a test set; then, the forward and inverse discrete model expressions of the helicopter system are established, and the number and basic structure of the input and output variables of the reversible neural network model are designed; the established reversible neural network model is trained using the training set data, and the loss function is constructed using the actual sampling value as a reference value and the output value of the reversible neural network model as an estimated value; according to the loss function, the reversible neural network model is trained using the training set data using the Adam algorithm to obtain the forward and inverse model of the helicopter nonlinear system; the test set is input into the reversible neural network model to verify the accuracy of the helicopter nonlinear forward and inverse model established by the reversible neural network, and a high-precision forward and inverse model of the helicopter nonlinear system is established.

[0114] Figure 2 is a data set schematic diagram of a method for establishing a forward and inverse model of a helicopter system based on a reversible neural network according to an embodiment of the present invention. The data set is composed of vibration data collected by sensors installed at the input and output ends of the helicopter nonlinear system, such as Figure 2 As shown, part of the collected data is used as a training set, and the remaining data is used as a training set.

[0115] Figure 3(a) to Figure 3(c) FIG3 is a neural network reversible block structure and an overall schematic diagram of a reversible neural network of a method for establishing a forward and inverse model of a helicopter system based on a reversible neural network according to an embodiment of the present invention. The structure of the neural network reversible block can realize forward and reverse operations with the same set of parameters. FIG3(a) is a process of realizing forward operation of the neural network reversible block, i.e., v1=u1⊙exp(s2(u2))+t2(u2), v2=u2⊙exp(s1(v1))+t1(v1); FIG3(b) is a process of realizing reverse operation of the neural network reversible block, i.e., u2=(v2-t1(v1))⊙exp(-s1(v1)), u1=(v1-t2(u2))⊙exp(-s2(u2)); FIG3(c) shows the overall structure of the reversible neural network, which is composed of three basic reversible blocks, and the mapping s in the basic reversible blocks is i and t i (i=1, 2) are all implemented through 2 fully connected layers and ReLU activation function.

[0116] Fig. 4(a) and Fig. 4(b) are training result diagrams of a method for establishing a forward and inverse model of a helicopter system based on a reversible neural network according to an embodiment of the present invention, wherein Fig. 4(a) represents the true output value of the helicopter nonlinear system and the output result of the model established by the forward process of the reversible neural network as well as the error between the two. The mean error (0.0248) between the output true value and the output result of the model established by the forward process of the reversible neural network is only 0.341% of the mean output true value (7.2666); Fig. 4(b) represents the true input value of the helicopter nonlinear system and the output result of the model established by the reverse process of the reversible neural network as well as the error between the two. The mean error (0.0146) between the input true value and the output result of the model established by the reverse process of the reversible neural network is only 0.169% of the mean input true value (8.6282); the mean errors in both figures are very close to 0, indicating that the forward and inverse models established by reversible neural network training have high accuracy. Fig. 5(a) and Fig. 5(b) are test result diagrams of a method for establishing a forward and inverse model of a helicopter system based on a reversible neural network according to an embodiment of the present invention, wherein Fig. 5(a) represents the true output value of the helicopter nonlinear system in the test set and the output result of the model established by the forward process of the reversible neural network as well as the error between the two. The mean error (0.0245) between the output true value and the output result of the model established by the forward process of the reversible neural network is only 0.337% of the mean output true value (7.2773); Fig. 5(b) represents the true input value of the helicopter nonlinear system in the test set and the output result of the model established by the reverse process of the reversible neural network as well as the error between the two. The mean error (0.0144) between the input true value and the output result of the model established by the reverse process of the reversible neural network is only 0.167% of the mean input true value (8.6196); the mean error values ​​in the two figures are very close to 0 and are not much different from the errors in the training set, indicating that the reversible neural network model obtained by training can still establish high-precision forward and inverse models of the helicopter nonlinear system in the test set.

[0117] In the preferred embodiment of the method for establishing the forward and inverse model of a helicopter system based on a reversible neural network of the present invention, in the first step S1: according to the complex nonlinear characteristics of the helicopter system, acceleration sensors are installed at the input and output ends where the model needs to be established, the helicopter system is started or continuous vibration excitation is applied to the helicopter system, and the input and output vibration data of the helicopter nonlinear object are collected through the sensors installed at the input and output ends, and part of the collected data is used as a training set, and the remaining data is used as a test set. In the preferred embodiment of the present invention, according to experience, 70% of the collected data is used as a training set, and the remaining 30% of the data is used as a test set, and the data is pre-processed by removing the mean. The training data only needs to reach a sufficient number, and the collected original data is simply processed and the data is removed from the mean.

[0118] In the preferred embodiment of the method for establishing the forward and inverse model of the helicopter system based on the reversible neural network of the present invention, in the second step S2: the forward and inverse discrete model expressions of the nonlinear object are established according to the collected input and output data of the helicopter nonlinear system, and the number and basic structure of the input and output variables of the reversible neural network model are designed. In the preferred embodiment of the method of the present invention, a and b in the forward and inverse discrete model expressions of the helicopter nonlinear object are set to 64 according to experience, and the mapping s in the basic reversible block is i and t i (i=1, 2) are implemented by two fully connected layers and ReLU activation function, where the fully connected layer contains 512 neurons. The deep reversible network consists of three basic reversible blocks, which not only ensures the accuracy of establishing the forward model and inverse model of the helicopter nonlinear system, but also avoids the problem of too many network parameters leading to long time consumption.

[0119] In a preferred embodiment of the method for establishing a forward and inverse model of a helicopter system based on a reversible neural network of the present invention, in the third step S3: according to the discrete model expression and the reversible neural network model, the actual sampling value is used as a reference value, and the output value of the reversible neural network model is used as an estimated value. The reversible neural network is used to establish the mean square error between the output values ​​of the forward and inverse models of the helicopter nonlinear system and the actual output and input sampling true values ​​to construct a loss function.

[0120] In a preferred embodiment of the method for establishing forward and inverse models of a helicopter system based on a reversible neural network of the present invention, in the fourth step S4: according to the loss function, the reversible neural network model is trained using the Adam algorithm using the training set data to obtain a trained reversible neural network model, whose forward process can fit the helicopter nonlinear system model, and the reverse process can fit the helicopter nonlinear system inverse model.

[0121] In the preferred embodiment of the method for establishing the forward and inverse model of the helicopter system based on the reversible neural network of the present invention, in the fifth step S5: the data of the test set are respectively forward input into the reversible neural network, that is, the helicopter nonlinear system model established, and reversely input into the reversible neural network, that is, the helicopter nonlinear system inverse model established, to verify the accuracy of the helicopter nonlinear system model and the inverse model established based on the reversible neural network. The accuracy of the helicopter nonlinear system model established by the forward process of the reversible neural network and the helicopter nonlinear system inverse model established by the reverse process is represented by the error between the output of the established model and the actual collected data.

[0122] In one embodiment, Figures 4(a) and 4(b) are training result diagrams of a method for establishing a forward and inverse model of a helicopter system based on a reversible neural network according to an embodiment of the present invention, wherein Figure 4(a) represents the true output value of the helicopter nonlinear system and the output result of the model established by the forward process of the reversible neural network, as well as the error between the two. The mean error (0.0248) between the output true value and the output result of the model established by the forward process of the reversible neural network is only 0.341% of the mean output true value (7.2666); Figure 4(b) represents the true input value of the helicopter nonlinear system and the output result of the model established by the reverse process of the reversible neural network, as well as the error between the two. The mean error (0.0146) between the input true value and the output result of the model established by the reverse process of the reversible neural network is only 0.169% of the mean input true value (8.6282). Figures 5(a) and 5(b) are test result diagrams of a method for establishing a forward and inverse model of a helicopter system based on a reversible neural network according to an embodiment of the present invention, wherein Figure 5(a) represents the true output value of the helicopter nonlinear system in the test set and the output result of the model built by the forward process of the reversible neural network, as well as the error between the two. The mean error (0.0245) between the output true value and the output result of the model built by the forward process of the reversible neural network is only 0.337% of the mean output true value (7.2773); Figure 5(b) represents the true input value of the helicopter nonlinear system in the test set and the output result of the model built by the reverse process of the reversible neural network, as well as the error between the two. The mean error (0.0144) between the input true value and the output result of the model built by the reverse process of the reversible neural network is only 0.167% of the mean input true value (8.6196). The closer the mean error is to 0, the higher the accuracy of the forward and inverse models of the helicopter nonlinear system established by the reversible neural network. As shown in Figure 4(a) and Figure 4(b), the mean error values ​​of the forward model and the inverse model are very close to 0. Therefore, the forward and inverse model establishment method of the helicopter system based on the reversible neural network can obtain high-precision system forward and inverse models in the training set after training. It can also be seen from Figure 5(a) and Figure 5(b) that the mean error values ​​of the forward model and the inverse model are very close to 0, and are not much different from the errors in the training set. Therefore, the forward and inverse models of the helicopter system established after training the reversible neural network can still maintain high accuracy of the system forward and inverse models in the test set.

[0123] In summary, the method can model and inverse model the helicopter nonlinear system through a reversible neural network, and can obtain the corresponding inverse model of the helicopter nonlinear system while establishing the forward model of the helicopter nonlinear system. Therefore, based on the reversible neural network, high-precision forward and inverse models of the helicopter nonlinear system can be directly and simply established. The established forward model of the helicopter nonlinear system can provide initial values ​​for the controller of the adaptive inverse control system. In addition, the forward and inverse models of the helicopter nonlinear system can be used for disturbance elimination.

[0124] In actual engineering, when a sensor is used to detect the system output, it may be accompanied by noise and disturbance. System models and inverse models are often used to eliminate disturbances. The established system model is connected in parallel with the system, that is, the system input drives both the system and the established model. The difference between the system output and the system model output is the noise and disturbance of the system. The sum of the system output noise and disturbance is used to drive the system inverse model to generate filtered noise and disturbance, which is subtracted from the system input. Using high-precision system models and inverse models can eliminate noise and disturbance in the system output.

[0125] Although the embodiments of the present invention are described above in conjunction with the accompanying drawings, the present invention is not limited to the above specific embodiments and application fields, and the above specific embodiments are only illustrative and instructive, rather than restrictive. A person of ordinary skill in the art can also make many forms under the guidance of this specification and without departing from the scope of protection of the claims of the present invention, all of which belong to the protection of the present invention.

Claims

1. A method for establishing a forward and inverse model of a helicopter system based on a reversible neural network, characterized in that: The method comprises the following steps: In the first step (S1), according to the nonlinear characteristics of the helicopter system, the input data and output data of the helicopter nonlinear system are collected, and a training set and a test set are constructed; In the second step (S2), the forward and reverse discrete models of the nonlinear object are respectively established according to the input data and output data of the helicopter nonlinear system, and a reversible neural network model is designed; In the third step (S3), a loss function is constructed based on the forward and reverse discrete models and the reversible neural network model of the nonlinear object, with the actual sampling value as a reference value and the output value of the reversible neural network model as an estimated value; In the fourth step (S4), according to the loss function, the reversible neural network model is trained using the data of the training set to obtain a trained reversible neural network model, a forward model of the helicopter nonlinear system is established based on the forward process of the trained reversible neural network model, and a reverse model of the helicopter nonlinear system is established based on the reverse process of the trained reversible neural network model; In the fifth step (S5), the data of the test set are respectively forward input into the helicopter nonlinear system model and reverse input into the helicopter nonlinear system inverse model to verify the accuracy of the helicopter nonlinear system forward model and the helicopter nonlinear system inverse model, so as to eliminate noise and disturbance in the system output.

2. The method according to claim 1, characterized in that Preferably, in the first step (S1), acceleration sensors are installed at the input and output ends where the model needs to be established, and input and output vibration data of the helicopter nonlinear object are collected. Part of the input and output vibration data is used as a training set, and the remaining data is used as a test set.

3. The method according to claim 1, characterized in that The second step (S2) includes: Establish the forward discrete model expression of the helicopter nonlinear object: and t =f(x t ,...,x t-a ;and t-1 ,...,and t-b ) Among them, y t represents the system output value at time t, f represents the forward function of the helicopter nonlinear system, x t , ..., x t-a represents the historical system input value at time t and before, y t-1 , ..., y t-b Represents the historical system output value at time t-1 and before; The inverse discrete model expression of the helicopter nonlinear object is established: x t =g(x t-1 ,...,x t-a ;y t ,...,y t-b ) Among them, x t represents the system input value at time t, g represents the inverse function of the helicopter nonlinear system, x t-1 , ..., x t-a represents the historical system input value at time t-1 and before, y t , ..., y t-b Represents the historical system output value at time t and before.

4. The method according to claim 3, characterized in that In the second step (S2), designing the reversible neural network model includes: designing the number and basic structure of the input and output variables of the reversible neural network model, and the forward and reverse operations of the reversible neural network are respectively expressed as: v = for (u, s, t) u=back(v,s,t) Among them, v represents the forward output and reverse input of the reversible neural network, u represents the forward input and reverse output of the reversible neural network, s and t represent the set of all reversible block parameters in the reversible network, for represents the forward function of the reversible neural network, and back represents the reverse function of the reversible network.

5. The method according to claim 4, characterized in that In the second step (S2), permutation layers are inserted between reversible blocks to increase the model capacity, and the number of reversible blocks and the structure of the reversible blocks are adjusted according to the complexity of the helicopter nonlinear system.

6. The method according to claim 4, characterized in that In the third step (S3), the loss function consists of two parts, and its expression is as follows: L total =L for +L back Among them, L total Represents the total loss function of the reversible neural network model, L for and L back They are the forward and reverse mean square error loss functions of the reversible neural network model, respectively.

7. The method according to claim 6, characterized in that In the third step (S3), L for It is the mean of the sum of squares of the corresponding point errors between the forward output result v of the reversible neural network and the actual system output sampling value y, and its expression is as follows: L back The mean of the sum of squares of the corresponding point errors between the inverse output result u of the reversible neural network and the actual system input sampling value x is expressed as follows: Where MSE(·) represents mean square error and n represents the number of samples.

8. The method according to claim 7, characterized in that In the fourth step (S4), the reversible neural network model is trained using the data of the training set according to the loss function, and a forward model of the helicopter nonlinear system is established based on the forward process of the trained reversible neural network; and a reverse model of the helicopter nonlinear system is established based on the reverse process of the trained reversible neural network.

9. The method according to claim 1, characterized in that: In the fifth step (S5), the accuracy of the helicopter nonlinear system forward model established by the reversible neural network forward process and the helicopter nonlinear system inverse model established by the reverse process is represented by the error between the output of each model and the actual collected data.

Images