Mars power descending section disturbance estimation compensation method based on depth operator neural network

Abstract

The invention provides a Mars power descent section disturbance estimation compensation method based on a depth operator neural network, and the method comprises the steps: obtaining historical trajectory data of a target detector in a Mars power descent section, analyzing the historical trajectory data through a disturbance prediction model, and obtaining a disturbance prediction result of the target detector at a current moment; determining disturbance compensation of the target detector through a kinetic equation according to a disturbance prediction result and the position information at the current moment; executing landing control on the target detector according to the disturbance compensation; the training step of the disturbance prediction model is obtained by performing disturbance prediction training through a depth operator neural network according to historical control input. The method has the beneficial effects that the system resource consumption during guidance is reduced, and the real-time compensation precision of guidance is improved.

Description

Technical Field

[0001] The present invention relates to the field of Mars powered descent guidance technology, and in particular to a Mars powered descent disturbance estimation and compensation method, device, equipment and medium based on a deep operator neural network. Background Art

[0002] The powered descent of the probe on Mars is affected by many factors, including initial state uncertainty (mainly caused by navigation error, parachute opening point error and cumulative evolution error), model parameter uncertainty (aerodynamic parameters, engine parameters, etc.) and external disturbances (atmospheric circulation, gusts, sandstorms, etc.). It is necessary to fully consider the uncertainty and disturbance factors in the design of trajectory planning and guidance control. However, due to the variable atmospheric environment of Mars itself and the small number of Mars landing missions, it is difficult to establish a sufficiently accurate system model due to the lack of accurate model data in advance. Therefore, it is necessary to improve the existing methods in order to be competent for the guidance and control mission of the powered descent on Mars under uncertain conditions.

[0003] The current disturbance estimation and compensation technology solutions for the powered descent stage on Mars usually adopt disturbance observers and robust control strategies. The performance of the disturbance observer depends on the accuracy of the system model. In high-dimensional systems, the design and implementation of the disturbance observer are more complex. Robust control can handle disturbances within a certain range, but an upper limit of the disturbance needs to be set. If the upper limit is set to exceed the actual disturbance, the control strategy may be too conservative, thereby reducing the dynamic performance of the system; if the upper limit is lower than the actual disturbance, it may cause system instability. Summary of the invention

[0004] Aiming to solve at least one of the technical problems existing in the prior art, the present invention provides a disturbance estimation and compensation method for the Mars powered descent phase based on a deep operator neural network, which reduces the system resource consumption during guidance and improves the real-time compensation accuracy of the guidance.

[0005] One aspect of the present invention provides a method for estimating and compensating disturbances during a powered descent on Mars based on a deep operator neural network, comprising:

[0006] Acquire historical trajectory data of the target probe during the powered descent phase on Mars, analyze the historical trajectory data using a disturbance prediction model, and obtain a disturbance prediction result of the target probe at the current moment;

[0007] Determining disturbance compensation of the target detector through a dynamic equation according to the disturbance prediction result and the position information at the current moment;

[0008] performing landing control on the target probe according to the disturbance compensation;

[0009] The training steps of the disturbance prediction model include:

[0010] Acquire a data set, the data set comprising first simulation trajectory data, the first simulation trajectory data comprising a state trajectory of the detector, each state trajectory comprising historical control input and label data;

[0011] Performing disturbance prediction according to the historical control input by using a deep operator neural network to obtain a disturbance prediction value;

[0012] Iterative training is performed according to the disturbance prediction value and the label data to obtain the disturbance prediction model.

[0013] According to the Mars powered descent disturbance estimation and compensation method based on deep operator neural network, the data set is obtained by:

[0014] Determine the initial state of the probe in the simulation environment and use the ZEM / ZEV guidance law controller to control the probe to land;

[0015] The first simulation trajectory data is obtained by randomly setting the amplitude, phase, frequency and bias of the disturbance according to uniform distribution using Monte Carlo simulation.

[0016] According to the Mars powered descent disturbance estimation and compensation method based on deep operator neural network, the disturbance prediction is performed according to the historical control input by the deep operator neural network to obtain the disturbance prediction value, including:

[0017] The deep operator neural network includes a trunk network and a branch network, the current moment is input through the trunk network, and the historical control input is input through the branch network, wherein the historical control input includes the detector position, speed, acceleration and control amount at the historical moment;

[0018] Input a parameter of the historical control input through each branch network to obtain the branch network branch output vector, input the current moment through the trunk network and perform Fourier encoding processing to obtain the trunk output vector;

[0019] The branch output vectors are combined into a column vector and the inner product is performed with the trunk output vector to obtain the disturbance prediction value.

[0020] According to the Mars powered descent disturbance estimation and compensation method based on deep operator neural network, the Fourier encoding process includes:

[0021] The current moment of the backbone network input is encoded according to the frequency encoding number, including:

[0022]

[0023] in, is a learnable parameter in the network, and when the network is initialized

[0024] B~Norm(0 m ,α 2 1)

[0025] Among them, m is the information length of the historical control input to the branch network, and t is the current time of the trunk network input.

[0026] According to the Mars powered descent disturbance estimation and compensation method based on deep operator neural network, the method further includes:

[0027] The branch network and the trunk network include an input layer and a hidden layer, wherein the input layer adopts a fully connected layer or a convolutional layer, the activation function of the input layer adopts tanh, the input layer width of the branch network is determined according to the information length of the historical control input to the branch network, and the input layer width of the trunk network is 1; wherein the frequency coding number of Fourier coding adopts one of 32, 64, and 128; wherein the hidden layer adopts a fully connected layer or a convolutional layer, wherein the output dimension of the branch network is p / 4, and the output dimension of the trunk network is p, and p is a multiple of 4.

[0028] According to the Mars powered descent disturbance estimation and compensation method based on deep operator neural network, the disturbance prediction value and the label data are iteratively trained to obtain the disturbance prediction model, including:

[0029] Using a mean square error loss function to determine whether convergence occurs after N rounds of training based on the difference between the disturbance prediction value and the label data, and performing a loss function change trend on the training set and the validation set based on cross-validation;

[0030] Whether the iterative training stops is determined based on whether the deep operator neural network converges and the trend of the loss function change. The mean square error loss function L MSE for:

[0031]

[0032] The method for determining whether the deep operator neural network converges is:

[0033]

[0034] The cross validation is:

[0035]

[0036] in, is the true value of the disturbance, is the disturbance prediction value.

[0037] According to the Mars powered descent disturbance estimation and compensation method based on deep operator neural network, the landing control of the target probe is performed according to the disturbance compensation, including:

[0038] According to disturbance compensation, the dynamic equation and ZEM / ZEV feedback guidance law are used to control the lander to perform the landing control of the target probe, where the dynamic equation is:

[0039]

[0040] Where r is the detector position, υ is the detector velocity, is the time derivative of the detector position, is the time derivative of the detector velocity, a T is the acceleration provided by the reverse thrust engine, g is the acceleration of Mars gravity, d is the disturbance acceleration, T is the thrust output by the thruster, m is the mass of the probe, and c is the thruster discharge velocity;

[0041] The ZEM / ZEV feedback guidance law is:

[0042] t go =t f -t

[0043]

[0044] Among them, tg o is the remaining time, t f is the landing time, t is the current time, r f After feedback guidance, the detector is, v f is the velocity of the detector after feedback guidance, and a is the control acceleration.

[0045] Another aspect of the present invention provides a disturbance estimation and compensation device for a Mars powered descent phase, comprising:

[0046] The first module is used to obtain the historical trajectory data of the target probe during the powered descent phase on Mars, analyze the historical trajectory data through a disturbance prediction model, and obtain the disturbance prediction result of the target probe at the current moment;

[0047] The second module is used to determine the disturbance compensation of the target detector through a dynamic equation according to the disturbance prediction result and the position information at the current moment;

[0048] A third module is used to perform landing control on the target probe according to the disturbance compensation;

[0049] The training steps of the disturbance prediction model include:

[0050] A fourth module is used to obtain a data set, wherein the data set includes first simulation trajectory data, the first simulation trajectory data includes a state trajectory of the detector, and each state trajectory includes historical control input and label data;

[0051] A fifth module is used to perform disturbance prediction according to the historical control input through a deep operator neural network to obtain a disturbance prediction value;

[0052] The sixth module is used to perform iterative training according to the disturbance prediction value and the label data to obtain the disturbance prediction model.

[0053] Another aspect of an embodiment of the present invention provides an electronic device, including a processor and a memory;

[0054] The memory is used to store programs;

[0055] The processor executes the program to implement the method described above.

[0056] The embodiment of the present invention also discloses a computer program product or a computer program, which includes a computer instruction stored in a computer-readable storage medium. A processor of a computer device can read the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the method described above.

[0057] The beneficial effects of the present invention are as follows: the introduction of parameter uncertainty enhances the adaptability of the network, Fourier layer frequency coding improves the prediction capability of fast-changing and periodic disturbance signals, and meets the requirements of the rapidly changing environment during the Mars landing process; the network structure is optimized to adapt to the limitations of onboard computing resources, efficient real-time estimation and compensation are achieved, and the system response speed is improved; the dependence on high-precision models is reduced, and the adaptive capability in complex environments is enhanced through an active compensation mechanism, and effective learning is achieved on small data sets, thereby improving the feasibility of model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a schematic diagram of the standard deep operator neural network structure.

[0059] Figure 2 It is a flow chart of a method for estimating and compensating disturbances during powered descent on Mars based on a deep operator neural network according to an embodiment of the present invention.

[0060] Figure 3 It is a schematic diagram of the disturbance prediction process of the improved deep operator neural network according to an embodiment of the present invention.

[0061] Figure 4 It is an improved disturbance prediction deep operator neural network structure of an embodiment of the present invention.

[0062] Figure 5 It is a schematic diagram of the internal structure of the branch network and the trunk network of an embodiment of the present invention.

[0063] Figure 6 This is a diagram of network prediction effect without Fourier layer frequency coding according to an embodiment of the present invention.

[0064] Figure 7 This is the network prediction effect of introducing Fourier layer frequency coding in an embodiment of the present invention.

[0065] Figure 8 It is a comparison diagram of prediction errors with and without the Fourier layer in an embodiment of the present invention.

[0066] Fig. 9 4 is a diagram of the constant disturbance prediction result of an embodiment of the present invention.

[0067] Fig.10 This is a graph of the sinusoidal unbiased disturbance prediction results of an embodiment of the present invention.

[0068] Fig.11 It is the prediction result of the sinusoidal bias disturbance according to the embodiment of the present invention.

[0069] Fig.12 This is a feedforward compensation z-axis position control effect diagram of an embodiment of the present invention

[0070] Fig.13 This is a diagram showing the effect of feedforward compensation z-axis speed control according to an embodiment of the present invention.

[0071] Fig.14 This is a diagram showing the effect of feedforward compensation z-axis acceleration control according to an embodiment of the present invention.

[0072] Fig.15 It is a relationship diagram between the cumulative computation time and simulation time of the deep operator network in an embodiment of the present invention.

[0073] Fig.16 It is a schematic diagram of a disturbance estimation and compensation device for the Mars powered descent phase based on a deep operator neural network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0074] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. In the subsequent description, the use of suffixes such as "module", "component" or "unit" used to represent elements is only for the purpose of facilitating the description of the present invention, and has no specific meaning in itself. Therefore, "module", "component" or "unit" can be used in a mixed manner. "First", "second" and the like are only used for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features or implicitly indicating the order of the indicated technical features. In this subsequent description, the continuous numbering of the method steps is for the convenience of review and understanding. In combination with the overall technical solution of the present invention and the logical relationship between the various steps, adjusting the implementation order between the steps will not affect the technical effect achieved by the technical solution of the present invention. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0075] refer to Figure 1 Schematic diagram of the standard deep operator neural network structure. According to the universal approximation, a multilayer perceptron with at least one hidden layer can approximate any continuous function from a finite-dimensional input space to an output space with arbitrary accuracy as long as the activation function is nonlinear. By generalizing this theorem, we can obtain the universal operator approximation theorem: Assume that σ is a non-polynomial function, X is the branch space, are two compact sets, V is defined in K 1 The continuous function set C(K 1 ), G is a compact set that maps V to C(K 2 ), then for any ∈>0, there exist positive integers n, p, m and constants

[0076]

[0077] The universal operator approximation theorem shows that by constructing a trunk network and a branch network with activation functions, arbitrary precision approximation of nonlinear operators can be achieved. Assume that the state equation of a one-dimensional dynamic system affected by the time-varying control variable u(t) is:

[0078]

[0079] in is the system state quantity, represents a nonlinear function, and t is time. The standard deep operator neural network structure for solving this problem is as follows:

[0080] The standard deep operator neural network first converts the continuous control input u(t) into discrete sampling points through a fixed-position "sensor". The branch network receives the sampling point data as input and outputs the vector [b 1 , b 2 , …, b p ] T The backbone network takes time t as input and outputs the vector [t 1 , t 2 ,…,t p ] T The outputs of the two networks are vectorized by inner product operations to obtain the final output y(t). Similarly, when the system state is n-dimensional, the outputs of the branch network and the trunk network can be divided into n groups, and dot product operations are performed on each group to obtain outputs of corresponding dimensions.

[0081] Figure 2 The present invention is a flowchart of a disturbance estimation and compensation method for Mars powered descent based on a deep operator neural network, which includes a disturbance estimation and compensation step and a disturbance prediction model training step.

[0082] The disturbance estimation and compensation steps include but are not limited to steps S100 to S300:

[0083] S100, obtaining historical trajectory data of the target probe during the powered descent phase on Mars, analyzing the historical trajectory data through a disturbance prediction model, and obtaining a disturbance prediction result of the target probe at the current moment.

[0084] S200, determining the disturbance compensation of the target detector through a dynamic equation according to the disturbance prediction result and the position information at the current moment.

[0085] S300, performing landing control on the target probe according to disturbance compensation.

[0086] In some embodiments, the Mars powered descent phase refers to the process in which the lander uses the reverse thrust engine to adjust the probe's flight trajectory after the parachute deceleration phase to ensure that the lander lands safely and accurately at the target point. When performing trajectory optimization and guidance control design, the following dynamic equations are used:

[0087]

[0088] The variables in the formula are the detector position r, the detector velocity v, and the derivative of the detector position with respect to time The time derivative of the detector velocity The acceleration provided by the reverse thrust engine is a T, Mars gravity acceleration g, disturbance acceleration d, thrust output by thruster T, probe mass m, thruster discharge velocity c. In addition to the disturbance, there are also parameter uncertainties in the system, which are mainly reflected in the uncertainty of Mars gravity acceleration and thruster discharge velocity.

[0089] In some embodiments, the lander is controlled to perform landing control of the target probe using dynamic equations and ZEM / ZEV feedback guidance laws according to disturbance compensation.

[0090] In some embodiments, the ZEM / ZEV feedback guidance law is

[0091] t go =t f -t

[0092]

[0093] Among them, t go is the remaining time, t f is the landing time, t is the current time, r f After feedback guidance, the detector is, v f is the velocity of the detector after feedback guidance, and a is the control acceleration.

[0094] The disturbance prediction model training steps include but are not limited to steps S400 to S600:

[0095] S400 acquires a data set, the data set includes first simulation trajectory data, the first simulation trajectory data includes a state trajectory of the detector, and each state trajectory includes historical control input and label data.

[0096] In some embodiments, in a simulation environment, the initial state of the detector is fixed, and the ZEM / ZEV controller is used to control the landing of the detector. Each Monte Carlo simulation randomly sets the amplitude, phase, frequency, and bias of the disturbance according to a uniform distribution to simulate the change of the detector state under different disturbances. Each simulation generates a complete state trajectory. The data set consists of all state trajectories generated by the Monte Carlo simulation. Each trajectory contains two parts: network input and label data. The label data is the true value of the disturbance at each moment, and the network input is the position, velocity, acceleration, and control amount of the detector at each moment. The data set is divided into a training set and a validation set. Both the training set and the validation set contain several complete state trajectories. The validation set does not affect the network update and is used to verify the network performance.

[0097] S500, performing disturbance prediction according to historical control inputs through a deep operator neural network to obtain a disturbance prediction value.

[0098] In some embodiments, reference Figure 3The disturbance prediction process diagram of the improved deep operator neural network shown in FIG. 1 includes but is not limited to steps S510 to S530:

[0099] S510, the deep operator neural network includes a trunk network and a branch network, the current moment is input through the trunk network, and the historical control input is input through the branch network, wherein the historical control input includes the detector position, speed, acceleration and control amount at the historical moment.

[0100] S520, a parameter in the historical control input is inputted into each branch network to obtain a branch network branch output vector, and the current moment is inputted into the trunk network and Fourier encoding is performed to obtain a trunk output vector.

[0101] S530: merge the branch output vectors into a column vector and perform vector inner product processing with the trunk output vector to obtain a disturbance prediction value.

[0102] In some embodiments, reference Figure 4 The improved disturbance prediction deep operator neural network structure is shown. By inputting the state quantity of the detector, the system input is solved. Taking the z-axis as an example, the power descent section detector has three state quantities in the z-axis direction, namely, position r z , speed z and acceleration a z ; The system input of the z-axis consists of two parts: the z-axis acceleration u provided by the thruster z and the disturbance acceleration d about the z-axis z ,u z is a known system control input, so we only need to solve d z That's it.

[0103] It can be understood that the improved network of the embodiment of the present invention comprises four branch networks and one trunk network. The four branch networks respectively input the system control input (the acceleration provided by the reverse thrust engine in the z-axis direction) u at the past m moments z (t 1 ),u z (t 2 ), …, u z (t m ), position r z (t 1 ), r z (t 2 ), …, r z (t m ), speed z (t 1 ), z (t 2 ), …, υ z (t m) and acceleration a z (t 1 ), a z (t 2 ), …, a z (t m ), the outputs of the four branch networks are combined into a column vector The current time t is encoded by the Fourier coding layer and then input into the backbone network, and the backbone network outputs a column vector and Do the vector inner product to get the predicted value of the disturbance

[0104] In some embodiments, reference Figure 5 The internal structure diagram of the branch network and the main network is shown in the figure. According to the input history state information length m, the input layer width of the branch network is determined to be m. The hidden layer of the branch network can select common neural network layers such as fully connected layers or convolutional layers. To make the output smooth, the activation function uses tanh. The frequency coding number of the Fourier coding layer can be selected from 32, 64, 128, etc. The input of the main network is the current moment, and the input layer width of the main network is determined to be 1. The hidden layer of the main network can also select a fully connected layer or a convolutional layer, and the activation function uses tanh. The number and width of the hidden layers of the main network and the branch network can be adjusted according to the training effect. The output dimension of each branch network is p / 4, and the output is merged into a p-dimensional vector. The output dimension of the main network is p, and p is an integer multiple of 4.

[0105] The embodiment of the present invention uses the Adam adaptive momentum estimation algorithm. The Adam optimizer can accelerate convergence in the early stage of training and remain stable when the gradient is sparse. In terms of hyperparameters, the learning rate can be selected as 0.001, 0.0001, etc., and can be fine-tuned according to the training results; the number of training iterations can be set to 1000-3000, and the small batch data size is 512, which can be fine-tuned according to the training results.

[0106] S600, performing iterative training according to the disturbance prediction value and the label data to obtain a disturbance prediction model.

[0107] In some embodiments, the difference between the disturbance prediction value and the actual disturbance value is measured by using a mean square error loss function, where: is the true value of the disturbance, is the disturbance prediction value, and N is the number of samples in a single round of training. After a specified number of iterations of training, the network is judged to be converged by the loss function curves on the training set and the validation set. If the loss function on the validation set drops below 0.01 or the loss function fluctuation is less than 0.01 in 50 consecutive iterations, the network is considered to have converged. Through cross-validation, the change trend of the network's loss function on the training set and the validation set is observed to avoid overfitting of the network.

[0108]

[0109]

[0110] In some embodiments, the environment and lander parameters involved in the simulation are as shown in Table 1

[0111] Table 1 Mars environment and probe related parameter values

[0112] parameter name Value unit <![CDATA[g e ]]> Earth's gravitational acceleration 9.807 <![CDATA[m / s 2 ]]> g Mars gravity acceleration 3.7114 <![CDATA[m / s 2 ]]> <![CDATA[r z (0)]]> Detector z-axis initial height 1500 m <![CDATA[υ z (0)]]> Detector z-axis initial velocity -75 m / s <![CDATA[r z (t f )]]> Detector z-axis end height 0 m <![CDATA[υ z (t f )]]> Detector z-axis end speed 0 m / s <![CDATA[t f ]]> Simulation time 60 s m(0) Initial nominal mass of detector 1905 kg <![CDATA[I sp ]]> Specific impulse 225 s

[0113] In the embodiment of the present invention, the Mars powered descent stage adopts the following ZEM / ZEV feedback guidance law to control the lander, where t f =60s, r f =0,υ f = 0. The ZEM / ZEV feedback guidance law here has two parts:

[0114] 1) Generate a control signal dataset for training;

[0115] 2) Test the performance improvement of the feedback controller plus feedforward compensation compared to using only the feedback controller.

[0116] ZEM / ZEV feedback guidance is:

[0117] t go =t f -t

[0118]

[0119] The network structure used in training is the same as above. The four branch networks have the same structure, including two fully connected layers of [512, 256]. The activation function uses tanh. The input layer dimension is 50. The control amount, z-axis position, z-axis velocity, and z-axis acceleration of the past 50 sampling points are input respectively. Each branch network outputs a 20-dimensional feature vector, which is merged into an 80-dimensional feature vector. The backbone network contains 32 frequency-coded Fourier layers and a fully connected layer of [512, 256]. The activation function uses tanh. The input is the moment of the current sampling point, and the output is an 80-dimensional feature vector. The output of the backbone network and the branch network is cross-multiplied to obtain the predicted value of the z-axis disturbance at the current moment.

[0120] The hardware platform used for the simulation is as follows: CPU is 13th Gen Intel(R)Core(TM)i5-13400F2.50GHz, GPU is NVIDIA GeForce RTX 4060, and running memory is 32GB.

[0121] Fourier layer validity verification reference Figure 6 , Figure 7 and Figure 8 A sinusoidal acceleration disturbance with an amplitude of 1.7922m / s2, a phase of 1.2322rad, a frequency of 1.0357rad / s, and a bias of 0.3575m / s2 is added in the z-axis direction. The training data set is obtained by simulation. The deep operator network with Fourier layer added to the backbone network and the deep operator network without Fourier layer are trained for 2000 rounds. The test set is the same as the training set. The output of the two networks on the test set is as follows Figure 6 , Figure 7 and Figure 8 .

[0122] Figures 6 to 8 The horizontal axis is the simulation time. Figure 6 , Figure 7 The vertical axis is the disturbance estimate and the true value. Figure 8 is the disturbance estimation error. Figure 6 It can be seen that the original network structure is not suitable for periodic and long-term signals and can hardly predict disturbances correctly. Figure 7 It can be seen that after the introduction of Fourier layer frequency coding, the above problems are significantly improved, and the network can accurately predict the disturbance at the current moment. Figure 8 This shows that the perturbation estimation error of the network after the Fourier layer is introduced is much smaller than the estimation error of the original network. The simulation proves that the improved network structure proposed in the embodiment of the present invention can effectively improve the performance of the deep operator network in solving the inverse problem of the input fast-changing dynamic system.

[0123] In some embodiments, the prediction verification of a sinusoidal disturbance with parameter uncertainty is as follows:

[0124] The two important components of the disturbances to which the Mars lander is subjected during the landing process are sinusoidal disturbances and constant disturbances, which can be summarized as sinusoidal disturbances with bias. Due to the lack of Martian atmospheric data and the uncertainty of the prior information of the disturbance parameters, this section generates a data set under the action of sinusoidal disturbances with parameter uncertainty through simulation. 80% of the data set is used as a training set and 20% as a test set. There is no overlap between the training set and the test set. The training is repeated for 2000 rounds, and the distribution of disturbance parameters is shown in Table 2.

[0125] Table 2 Distribution of sinusoidal disturbance parameters

[0126] Parameter name Value unit Upper bound of the z-axis disturbance amplitude 2 <![CDATA[m / s 2 ]]> The lower bound of the z-axis disturbance amplitude 0 <![CDATA[m / s 2 ]]> Upper bound of z-axis perturbation phase π rad The lower bound of the z-axis perturbation phase 0 rad Upper limit of z-axis perturbation frequency 2 Hz Lower bound of z-axis perturbation frequency 1 Hz z-axis perturbation bias upper bound 0.5 <![CDATA[m / s 2 ]]> z-axis perturbation bias lower bound -0.5 <![CDATA[m / s 2 ]]>

[0127] Figures 9 to 11 The horizontal axis is the simulation time, and the vertical axis is the size of the true value of the disturbance and the predicted value. The prediction results for constant disturbance, sinusoidal unbiased disturbance, and sinusoidal biased disturbance are respectively shown. The error between the network prediction value and the true value is small, indicating that when the actual disturbance parameter is within the parameter distribution of the training set, the method described in the embodiment of the present invention can accurately estimate the disturbance after the training is completed.

[0128] In some embodiments, the feedforward compensation effect is verified as follows:

[0129] A sinusoidal disturbance with a constant bias is added to the z-axis direction. On the basis of the above ZEM / ZEV guidance law, an improved deep operator neural network predictive feedforward compensation is added to compare the control effects of no disturbance, no feedforward compensation and with feedforward compensation. The no disturbance case is used as the reference trajectory. The control effect is shown in Figures 12 to 15 As shown, the horizontal axis is the simulation time, and the vertical axes are the z-axis position, z-axis velocity, and z-axis acceleration. It can be seen that if only the feedback guidance law is used, due to the existence of disturbances, the z-axis velocity of the detector is about -5m / s and the height is about -100m at the termination time. In the actual mission, the detector has crashed under this condition. However, the feedforward compensation method proposed in the embodiment of the present invention is adopted, and compensation is added from the 10th second of the simulation time. The z-axis position, velocity, and acceleration of the detector are basically consistent with the undisturbed situation, which effectively improves the control accuracy of the powered descent stage.

[0130] In addition, to ensure that the network can be used for real-time computing of the task computer, the variation of the cumulative computing time with the simulation time is investigated. The results are as follows: Fig.15 As shown, the horizontal axis is the simulation time, and the vertical axis is the cumulative computing time of network reasoning. It can be seen that the computing time of the deep operator network increases linearly with the simulation time, that is, the single reasoning time remains unchanged, and the average single reasoning time is about 6.3ms, while the control period of the ZEM / ZEV guidance law is about 0.1s. It can be seen that the network described in the embodiment of the present invention can meet the requirements of real-time computing.

[0131] Fig.16 Schematic diagram of a Mars powered descent disturbance estimation and compensation device based on a deep operator neural network according to an embodiment of the present invention. The device includes a first module 1610, a second module 1620, a third module 1630, a fourth module 1640, a fifth module 1650, and a sixth module 1660.

[0132] Among them, the first module is used to obtain the historical trajectory data of the target probe in the powered descent phase of Mars, analyze the historical trajectory data through the disturbance prediction model, and obtain the disturbance prediction result of the target probe at the current moment; the second module is used to determine the disturbance compensation of the target probe through the dynamic equation according to the disturbance prediction result and the position information at the current moment; the third module is used to perform landing control on the target probe according to the disturbance compensation; the training steps of the disturbance prediction model include: the fourth module is used to obtain a data set, the data set includes first simulation trajectory data, the first simulation trajectory data includes the state trajectory of the probe, and each state trajectory includes historical control input and label data; the fifth module is used to perform disturbance prediction according to the historical control input through a deep operator neural network to obtain a disturbance prediction value; the sixth module is used to perform iterative training according to the disturbance prediction value and label data to obtain a disturbance prediction model.

[0133] Exemplarily, with the cooperation of the first module, the second module, the third module, the fourth module, the fifth module and the sixth module in the device, the embodiment device can implement any of the aforementioned disturbance estimation and compensation methods for the powered descent stage on Mars based on a deep operator neural network, that is, obtaining historical trajectory data of the target probe in the powered descent stage on Mars, analyzing the historical trajectory data through a disturbance prediction model, and obtaining a disturbance prediction result of the target probe at the current moment; determining the disturbance compensation of the target probe through a dynamic equation based on the disturbance prediction result and the position information at the current moment; performing landing control on the target probe based on the disturbance compensation; the training steps of the disturbance prediction model include: obtaining a data set, the data set includes first simulation trajectory data, the first simulation trajectory data includes a state trajectory of the probe, and each state trajectory includes historical control input and label data; performing disturbance prediction based on the historical control input through a deep operator neural network to obtain a disturbance prediction value; performing iterative training based on the disturbance prediction value and the label data to obtain a disturbance prediction model. The beneficial effects of the present invention are as follows: the introduction of parameter uncertainty enhances the adaptability of the network, Fourier layer frequency coding improves the prediction capability of fast-changing and periodic disturbance signals, and meets the requirements of the rapidly changing environment during the Mars landing process; the network structure is optimized to adapt to the limitations of onboard computing resources, efficient real-time estimation and compensation are achieved, and the system response speed is improved; the dependence on high-precision models is reduced, and the adaptive capability in complex environments is enhanced through an active compensation mechanism, and effective learning is achieved on small data sets, thereby improving the feasibility of model training.

[0134] An embodiment of the present invention further provides an electronic device, the electronic device comprising a processor and a memory;

[0135] The memory stores a program;

[0136] The processor executes the program to perform the aforementioned Mars powered descent disturbance estimation and compensation method based on deep operator neural network; the electronic device has the function of carrying and running the Mars powered descent disturbance estimation and compensation software system provided by an embodiment of the present invention, for example, a personal computer, a minicomputer, a main frame, a workstation, a network or distributed computing environment, a separate or integrated computer platform, or communicates with a charged particle tool or other imaging device, etc.

[0137] An embodiment of the present invention also provides a computer-readable storage medium, which stores a program, and the program is executed by a processor to implement the Mars powered descent disturbance estimation and compensation method based on a deep operator neural network as described above.

[0138] In some selectable embodiments, the function / operation mentioned in the block diagram may not occur in the order mentioned in the operation diagram. For example, depending on the function / operation involved, the two boxes shown in succession can actually be executed substantially simultaneously or the boxes can sometimes be executed in reverse order. In addition, the embodiment presented and described in the flow chart of the present invention is provided by way of example, for the purpose of providing a more comprehensive understanding of technology. The disclosed method is not limited to the operation and logic flow presented herein. Selectable embodiments are expected, wherein the order of various operations is changed and the sub-operation of a part for which is described as a larger operation is performed independently.

[0139] The embodiment of the present invention also discloses a computer program product or a computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the aforementioned Mars powered descent disturbance estimation and compensation method based on a deep operator neural network.

[0140] In addition, although the present invention is described in the context of functional modules, it should be understood that, unless otherwise specified, one or more of the functions and / or features described may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in separate physical devices or software modules. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the present invention. More specifically, in view of the properties, functions, and internal relationships of the various functional modules in the device disclosed herein, the actual implementation of the module will be understood within the conventional skills of the engineer. Therefore, those skilled in the art can implement the present invention set forth in the claims without excessive experimentation using ordinary techniques. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the present invention, which is determined by the full scope of the appended claims and their equivalents.

[0141] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0142] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0143] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0144] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0145] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0146] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

[0147] The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the described embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A method for estimating and compensating disturbances during Mars powered descent based on a deep operator neural network, characterized in that: include: Acquire historical trajectory data of the target probe during the powered descent phase on Mars, analyze the historical trajectory data using a disturbance prediction model, and obtain a disturbance prediction result of the target probe at the current moment; Determining disturbance compensation of the target detector through a dynamic equation according to the disturbance prediction result and the position information at the current moment; performing landing control on the target probe according to the disturbance compensation; The training steps of the disturbance prediction model include: Acquire a data set, the data set comprising first simulation trajectory data, the first simulation trajectory data comprising a state trajectory of the detector, each state trajectory comprising historical control input and label data; Performing disturbance prediction according to the historical control input by using a deep operator neural network to obtain a disturbance prediction value; Iterative training is performed according to the disturbance prediction value and the label data to obtain the disturbance prediction model.

2. The Mars powered descent disturbance estimation and compensation method based on deep operator neural network according to claim 1 is characterized in that: The dataset is obtained in the following way: Determine the initial state of the probe in the simulation environment and use the ZEM / ZEV guidance law controller to control the probe to land; The first simulation trajectory data is obtained by randomly setting the amplitude, phase, frequency and bias of the disturbance according to uniform distribution using Monte Carlo simulation.

3. The Mars powered descent disturbance estimation and compensation method based on deep operator neural network according to claim 2 is characterized in that: The performing disturbance prediction according to the historical control input by using a deep operator neural network to obtain a disturbance prediction value includes: The deep operator neural network includes a trunk network and a branch network, the current moment is input through the trunk network, and the historical control input is input through the branch network, wherein the historical control input includes the detector position, speed, acceleration and control amount at the historical moment; Input a parameter of the historical control input through each branch network to obtain the branch network branch output vector, input the current moment through the trunk network and perform Fourier encoding processing to obtain the trunk output vector; The branch output vectors are combined into a column vector and the inner product is performed with the trunk output vector to obtain the disturbance prediction value.

4. The Mars powered descent disturbance estimation and compensation method based on deep operator neural network according to claim 3 is characterized in that: The Fourier encoding process includes: The current moment of the backbone network input is encoded according to the frequency encoding number, including: in, is a learnable parameter in the network, and when the network is initialized B~Norm(0 m ,a 2 1) Among them, m is the information length of the historical control input to the branch network, and t is the current time of the trunk network input.

5. The Mars powered descent disturbance estimation and compensation method based on deep operator neural network according to claim 4 is characterized in that: The method further comprises: The branch network and the trunk network include an input layer and a hidden layer, wherein the input layer adopts a fully connected layer or a convolutional layer, the activation function of the input layer adopts tanh, the input layer width of the branch network is determined according to the information length of the historical control input to the branch network, and the input layer width of the trunk network is 1; wherein the frequency coding number of Fourier coding adopts one of 32, 64, and 128; wherein the hidden layer adopts a fully connected layer or a convolutional layer, wherein the output dimension of the branch network is p / 4, and the output dimension of the trunk network is p, and p is a multiple of 4.

6. The Mars powered descent disturbance estimation and compensation method based on deep operator neural network according to claim 1 is characterized in that: Iterative training is performed according to the disturbance prediction value and the label data to obtain the disturbance prediction model, including: Using a mean square error loss function to determine whether convergence occurs after N rounds of training based on the difference between the disturbance prediction value and the label data, and performing a loss function change trend on the training set and the validation set based on cross-validation; Whether the iterative training stops is determined based on whether the deep operator neural network converges and the trend of the loss function change. The mean square error loss function L MSE for: The method for determining whether the deep operator neural network converges is: The cross validation is: in, is the true value of the disturbance, is the predicted value of disturbance.

7. The Mars powered descent disturbance estimation and compensation method based on deep operator neural network according to claim 1 is characterized in that: The performing landing control on the target probe according to the disturbance compensation comprises: According to disturbance compensation, the dynamic equation and ZEM / ZEV feedback guidance law are used to control the lander to perform the landing control of the target probe, where the dynamic equation is: Where r is the detector position, v is the detector velocity, is the time derivative of the detector position, is the time derivative of the detector velocity, a T is the acceleration provided by the reverse thrust engine, g is the acceleration of Mars gravity, d is the disturbance acceleration, T is the thrust output by the thruster, m is the mass of the probe, and c is the thruster discharge velocity; The ZEM / ZEV feedback guidance law is: t go =t f -t Among them, t go is the remaining time, t f is the landing time, t is the current time, r f After feedback guidance, the detector is, v f is the velocity of the detector after feedback guidance, and a is the control acceleration.

8. A disturbance estimation and compensation device for a Mars powered descent, characterized in that: include: The first module is used to obtain the historical trajectory data of the target probe during the powered descent phase on Mars, analyze the historical trajectory data through a disturbance prediction model, and obtain the disturbance prediction result of the target probe at the current moment; The second module is used to determine the disturbance compensation of the target detector through a dynamic equation according to the disturbance prediction result and the position information at the current moment; A third module is used to perform landing control on the target probe according to the disturbance compensation; The training steps of the disturbance prediction model include: A fourth module is used to obtain a data set, wherein the data set includes first simulation trajectory data, the first simulation trajectory data includes a state trajectory of the detector, and each state trajectory includes historical control input and label data; A fifth module is used to perform disturbance prediction according to the historical control input through a deep operator neural network to obtain a disturbance prediction value; The sixth module is used to perform iterative training according to the disturbance prediction value and the label data to obtain the disturbance prediction model.

9. An electronic device, characterized in that: including a processor and a memory; The memory is used to store programs; The processor executes the program to implement the Mars powered descent disturbance estimation and compensation method based on a deep operator neural network as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: The storage medium stores a program, and the program is executed by a processor to implement the Mars powered descent disturbance estimation and compensation method based on a deep operator neural network as described in any one of claims 1-7.

Images