Aircraft brain neural network decoupling control method

Abstract

The application discloses an aircraft brain neural network decoupling control method, which comprises the following steps: acquiring multi-dimensional information of each channel of an aircraft related to different postures; constructing a decoupling neural network, including a feature coding layer, a dynamic decoupling layer and a control action mapping layer, wherein the input is a multi-channel control error fusion signal, and the output is independent control instructions of each channel after decoupling; the feature coding layer firstly outputs a weighted feature vector by using an attention mechanism module as a whole, and then further optimizes the weighted feature vector by using a contrast representation learning module to generate a strengthened feature vector; the decoupling neural network is connected with a controlled object model, and the decoupling neural network is trained online; and a trained aircraft posture decoupling network controller is used for decoupling control of the aircraft. The decoupling process is optimized by using orthogonality and low rank constraints, so that the model complexity is effectively reduced, and the redundant dimension interference is effectively inhibited.

Description

Technical Field

[0001] This invention belongs to the field of aircraft control technology, specifically relating to a brain-like neural network decoupling control method for aircraft. Background Technology

[0002] In the field of aircraft attitude control, traditional methods typically rely on precise dynamic models to design controllers, such as PID control, LQR control, or sliding mode control. These methods perform well in scenarios with simple structures and controllable environments, but their control performance is significantly limited in complex flight missions involving high maneuverability, nonlinearity, and strong coupling. In recent years, neural network-based control methods have gained increasing attention, especially reinforcement learning strategies, which are widely used in aircraft attitude control to continuously optimize control strategies through interaction with the environment. However, most existing methods still suffer from insufficient generalization ability and poor real-time performance, particularly struggling to maintain stable control performance when faced with parameter variations or external disturbances.

[0003] Furthermore, Xu Min et al. proposed an adaptive decoupling control law in their paper "Design of Adaptive Decoupling Control Law for Small Ducted Fan UAVs." By introducing a model reference adaptive control method, the attitude and angular velocity control loop of the UAV can stably follow the output of the reference model under the influence of uncertain disturbances and parameter perturbations, and achieve a certain degree of channel decoupling. Although this method alleviates the channel coupling problem to some extent, its control law design does not consider the modularity and low-rank characteristics of biological neural systems, resulting in high model complexity, poor generalization ability, and insufficient robustness.

[0004] Therefore, there is an urgent need for a novel controller representation model construction method that can balance the advantages of brain-like neural network structures with the multi-channel decoupled control requirements of aircraft. This method should be able to effectively separate the dynamic coupling relationships of each control channel in complex and variable flight environments, improve the response speed and anti-interference capability of the control system, and thus enhance the adaptability and safety of the aircraft attitude control system. Summary of the Invention

[0005] The purpose of this invention is to provide a brain-like neural network decoupling control method for aircraft. This method constructs a brain-like neural network decoupling controller representation model for aircraft, trains the network to capture the dynamic coupling relationships between various control channels, and achieves effective integration and separation of multimodal error information, thereby improving the adaptive capability and safety of the aircraft attitude control system. By introducing an attention mechanism and contrastive learning to dynamically adjust feature weights and optimize feature discriminability, the ability to extract key features is improved. This not only enhances the model's ability to suppress redundant information but also provides more accurate global perception input for subsequent decoupling. Orthogonality and low-rank constraints are used to optimize the decoupling process, effectively reducing model complexity and suppressing interference from redundant dimensions.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] A decoupling control method for a brain-like neural network in an aircraft, the method comprising the following steps:

[0008] Acquire multi-dimensional information about the aircraft and its different attitudes from various channels;

[0009] A decoupling neural network is constructed, which includes a feature encoding layer, a dynamic decoupling layer, and a control action mapping layer. The input of the decoupling neural network is a multi-channel control error fusion signal, and the output of the decoupling neural network is the independent control command of each channel after decoupling.

[0010] The feature encoding layer first uses the attention mechanism module to output a weighted feature vector, and then further optimizes it through the contrastive representation learning module to generate an enhanced feature vector F, which is then converted into a temporal encoding suitable for the dynamic decoupling layer of the decoupled neural network and input into the dynamic decoupling layer.

[0011] The dynamic decoupling layer consists of N coupled oscillating neurons, and its characterization constraint L, which possesses orthogonality and low-rank properties, is:

[0012] L=p1L o +p2L l +p3L s +p4L d

[0013]

[0014] L l =||ε|| nuc

[0015]

[0016] Where p1, p2, p3, and p4 are weighting coefficients, corresponding to the priorities of different constraint terms; L o For orthogonality constraints, The parameterized orthogonal projection matrix is ​​θ, which is the parameter matrix of the decoupled neural network, and d represents the dimension of the features; I3 is a 3×3 identity matrix; L l For low-rank constraints, ε is the residual tensor; L s For independence constraints; M(f) i ,f j ) represents the decoupling characteristic f i ,f j Mutual information; L d For continuity constraints, t represents the time step; Φ(θ) t+1 ), Φ(θ) tf represents the parameterized orthogonal projection matrices for time steps t+1 and t, respectively; i ,f j These are the decoupling features of the i-th neuron and the j-th neuron, respectively;

[0017] The decoupled neural network is connected to the controlled object model, and the decoupled neural network is trained online. The loss function is defined as:

[0018]

[0019] Where m is the dimension of the aircraft control channel. Let n be the real-time control error vector of the i-th channel. i Let L be the dimension of the error vector of the i-th control channel, and L be a characterization constraint term with orthogonality and low rank characteristics.

[0020] A trained aircraft attitude decoupling network controller is used for the decoupling control of the aircraft.

[0021] Furthermore, the online training method is either a reinforcement learning mode or an online incremental learning mode.

[0022] Furthermore, each coupled oscillatory neuron follows a nonlinear coupled oscillatory differential equation:

[0023]

[0024] in, The intrinsic dynamic matrix of the coupled oscillating neuron determines its fundamental oscillation frequency ω. i The intrinsic dynamic matrix A of the coupled oscillating neuron i With the fundamental oscillation frequency ω i There is a direct mathematical connection between them, A i The eigenvalues ​​of A determine the oscillatory behavior of the system. By adjusting A... i To control and optimize ω i B i f is the input coupling matrix; f(·) is the nonlinear activation function; Let be the coupling coefficient matrix between coupled oscillating neurons i and j, which satisfies the regularization constraint. j represents other coupled oscillating neurons connected to coupled oscillating neuron i; γ is a preset coupling strength threshold; n represents the number of dimensions of the state variables of each coupled oscillating neuron; t represents the time variable. z is the first derivative of the state variable; i (t), z j (t) is the state variable.

[0025] Furthermore, the multi-dimensional information from each channel related to different attitudes includes attitude angle error, angular rate error, three-axis control signals, wind speed, wind direction, and atmospheric pressure.

[0026] Furthermore, in the feature encoding layer, feature fusion of the data is achieved through the synergistic effect of attention mechanism and contrastive representation learning, specifically as follows:

[0027] Preprocessing is performed on multi-dimensional information from each channel related to different poses, and the features of each pose are combined to construct a feature vector set X = {x pitch ,x roll ,x yaw},

[0028] Where, x pitch This represents the feature vector related to pitch, including pitch angle error, pitch rate error, pitch angle control signal, pitch-related wind speed and direction, and atmospheric pressure; x roll This represents the feature vector related to roll, including roll angle error, roll rate, roll angle control signal error, roll-related wind speed and direction, and atmospheric pressure; x yaw The feature vectors related to yaw include yaw angle error, yaw rate error, yaw angle control signal, yaw-related wind speed and direction, and atmospheric pressure.

[0029] In the attention mechanism module, internal weights are first assigned to the features of each pose, and then internal weights are assigned to the features of each pose. Weighting is performed, with internal weights used... This indicates that it was calculated using the Softmax function;

[0030]

[0031] in It is the importance score of the i-th feature under pose ο. ) is the importance score of the j-th feature under pose ο;

[0032] Then the local weighted feature of the i-th feature in the o-th pose is: Summing all local weighted features under the same pose yields the local weighted features under the current pose.

[0033] Then, we introduce the global attention vector β = [β pitch ,β roll ,β yaw Distribute global attention weights among the roll, pitch, and yaw attitudes, satisfying the following:

[0034] β ο ≥0.

[0035] Where, β ο Let X represent the global attention weights for the οth pose, and let X represent the global attention weights for pitch, roll, and yaw, respectively. The final output is the feature vector X of the attention mechanism module. t :

[0036]

[0037] X t The input contrastive representation learning module processes the state features of the same pose at different time points. Construct positive sample pairs and combine features between different poses. Construct negative sample pairs ο≠ο′, where t, t+Δt represent different times, t′ represents a time different from t, Δt represents the time interval, and ο, ο′ represent different postures.

[0038] By constructing positive and negative sample pairs and optimizing feature representations based on the contrastive loss function, the distinguishability between different features is enhanced. The enhanced feature vector output by the contrastive representation learning module is represented as F.

[0039] The contrast loss function L contrastive for:

[0040]

[0041] Where s(·,·) represents feature similarity.

[0042] Furthermore, the decoupling neural network also includes pulse sequence time length T calculation and data processing, converting the enhanced feature vector F into a timing code suitable for the dynamic decoupling layer of the decoupling neural network, which is then used as the input signal sequence to the dynamic decoupling layer.

[0043] Furthermore, the controlled object model is a mathematical model of a controlled object in the fields of hypersonic aircraft and unmanned aerial vehicles.

[0044] The present invention also protects a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, can implement the contents of the method.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] (1) This invention addresses the challenges of heterogeneity, real-time performance, and dynamic adaptability in processing aircraft data (such as attitude angle error, angular rate error, etc.) by introducing the synergistic effect of attention mechanisms and contrastive learning. This method significantly improves the extraction capability of key features and the suppression capability of redundant information by dynamically allocating feature weights, optimizing feature discriminability, and combining contrastive learning of positive and negative sample pairs, providing more accurate global perception input for decoupled control. It not only improves the suppression capability of redundant information but also provides more accurate global perception input for decoupled control, significantly enhancing the system's adaptability and robustness, and achieving efficient processing of aircraft data (such as attitude angle error, angular rate error, wind speed and direction, etc.).

[0047] (2) This invention constructs a closed-loop online training framework based on real-time error backpropagation and adaptive momentum stochastic gradient descent algorithm. By combining matrix orthogonality and low-rank constraints, it realizes dynamic optimization and model interpretability improvement of the decoupled controller of the brain-like neural network for aircraft. By suppressing redundant dimension interference through orthogonality constraints and reducing model complexity through low-rank constraints, the robustness and safety of the system are significantly enhanced while ensuring high decoupling accuracy.

[0048] (3) This invention is specifically designed to address the strong coupling problem between channels (such as pitch, roll, and yaw) in the attitude control of aircraft, and solves the problems of severe cross-interference, slow response speed, and limited control accuracy that still exist in the existing technology in multi-channel cooperative control. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the brain-like neural network decoupling control method for aircraft of the present invention.

[0050] Figure 2 This is a schematic diagram of the feature coding layer and the coupled oscillating neuron structure in one embodiment of the present invention. Detailed Implementation

[0051] The present invention will be further explained below with reference to the embodiments and accompanying drawings, but this is not intended to limit the scope of protection of this application.

[0052] The present invention provides a brain-like neural network decoupling control method for aircraft, the method comprising the following steps:

[0053] Acquire multi-dimensional information of the aircraft, including attitude angle error, angular rate error, three-axis control signals, wind speed, wind direction, atmospheric pressure, etc.

[0054] A decoupling neural network is constructed, comprising a feature encoding layer, a dynamic decoupling layer, and a control action mapping layer. The input to the decoupling neural network is a multi-channel control error fusion signal. This fusion signal considers the influence of various factors, encompassing attitude angle error, angular rate error, three-axis control signals, and multi-dimensional information such as wind speed, wind direction, and atmospheric pressure, enabling global perception capability for decoupling control. The output of the decoupling neural network consists of independent control commands for each decoupled channel, corresponding to the desired control quantity for each channel (e.g., roll, pitch, yaw), achieving physical separation of multi-channel control.

[0055] The feature encoding layer first uses an attention mechanism module to output a weighted feature vector, and then further optimizes it through a contrastive representation learning module to generate a strengthened feature vector F = [f pitch ;f roll ;f yaw This is converted into a timing-coded input suitable for the dynamic decoupling layer of a decoupled neural network.

[0056] The dynamic decoupling layer consists of N coupled oscillating neurons, and its representational constraint L, which possesses orthogonality and low-rank properties, is:

[0057] L=p1L o +p2L l +p3L s +p4L d

[0058]

[0059] L l =||ε|| nuc

[0060]

[0061]

[0062] Where p1, p2, p3, and p4 are weighting coefficients, corresponding to the priorities of different constraint terms; L o For orthogonality constraints, The parameterized orthogonal projection matrix is ​​Φ(θ), where θ is the parameter matrix of the decoupled neural network, d represents the dimension of the features, and determines the number of rows in the Φ(θ) matrix; I3 is a 3×3 identity matrix; L l For low-rank constraints, ε is the residual tensor; L s For independence constraints; M(f) i ,f j ) represents the decoupling characteristic f i ,f j Mutual information; L d For continuity constraints, t represents the time step; Φ(θ) t+1), Φ(θ) t f represents the parameterized orthogonal projection matrices for time steps t+1 and t, respectively; i ,f j These are the decoupling features of the i-th neuron and the j-th neuron, respectively;

[0063] Coupled oscillators combine with neurons in a neural network to form coupled oscillatory neurons. Figure 2 In this context, E and I represent excitatory neurons (outputting excitatory connections) and inhibitory neurons (outputting inhibitory connections), respectively. It can contain positive and negative weights, corresponding to excitatory and inhibitory connections, respectively. Each oscillator can exhibit periodic behavior individually, but when these oscillators interact with each other, their overall behavior is influenced by each other, which may lead to new dynamic properties. A neuron can be regarded as an oscillator, and neurons communicate and influence each other through interactions. These interactions enable the activities of neurons to be synchronized or interact in other complex ways, similar to coupled oscillators in physical systems.

[0064] The state variables of each coupled oscillating neuron are represented as follows:

[0065]

[0066] Adding a coupling matrix to the differential equations of the state variables can adapt to the decoupling requirements of the aircraft control channel. Each coupled oscillating neuron then follows a nonlinear coupled oscillating differential equation:

[0067]

[0068] Where n represents the dimension of the state variable of each coupled oscillating neuron; t represents the time variable, reflecting the change of the state variable over time; The intrinsic dynamic matrix of the coupled oscillating neuron determines its fundamental oscillation frequency ω. i The intrinsic dynamic matrix A of the coupled oscillating neuron i With the fundamental oscillation frequency ω i There is a direct mathematical connection between them, A i The eigenvalues ​​of A determine the oscillatory behavior of the system. By adjusting A... i To control and optimize ω i ; is the input coupling matrix, p is the dimension of the input signal of the dynamic decoupling layer; f(·) is the nonlinear activation function used to map the external control error signal to the oscillation amplitude adjustment; Let be the coupling coefficient matrix between coupled oscillating neurons i and j, which satisfies the regularization constraint. j represents other coupled oscillating neurons connected to coupled oscillating neuron i, and γ is a preset coupling strength threshold. γ is used to balance decoupling robustness and information interaction capability. = is the first derivative of the state variable.

[0069] The optimization algorithm aims to find a suitable set of parameters (i.e., the elements of the matrix) such that the fundamental oscillation frequency of each coupled oscillating neuron can meet the decoupling objective of the system. In this embodiment, the optimization algorithm adopts the gradient descent method.

[0070] The decoupled neural network is connected to the aircraft dynamics model, and trained using reinforcement learning or online incremental learning. The loss function is defined as follows:

[0071]

[0072] Where m is the dimension of the aircraft control channel (in this application, it is divided into three channels: roll, pitch, and yaw, each calculated independently). Let n be the real-time control error vector of the i-th channel (including attitude angle deviation and angular rate deviation). i Let L be the error vector dimension of the i-th control channel, and L be a representation constraint term with orthogonality and low rank, used to guide the network to adaptively learn the independent control features between channels.

[0073] Reinforcement learning enhances the model's adaptability to changing environments.

[0074] During training, a closed-loop online training framework is constructed based on real-time error backpropagation and adaptive momentum stochastic gradient descent algorithms. This framework dynamically matches the temporal features required for decoupling. Each coupled oscillating neuron not only possesses its own oscillatory characteristics but also interacts with other coupled oscillating neurons through a coupling matrix. This interaction leads to complex collective oscillatory behavior, where the relative positional relationships between the coupled oscillating neurons (oscillation phase difference Δφ) are considered. ij ) is the key factor. The system adjusts {ω i W ij The state evolution of coupled oscillating neurons should satisfy the decoupling objective, i.e., when the system reaches steady state, the state difference between adjacent coupled oscillating neurons should be ||z||. i (t)-z j (t)‖2 decays to within the preset error range, and the oscillation phase difference Δφ output by each coupled oscillating neuron ij Corresponding one-to-one with the decoupling requirements of the aircraft control channels, the decay of the state difference ‖zi(t)-zj(t)‖2 implies an increase in the synchronization between the coupled oscillating neurons, thereby leading to an oscillation phase difference Δφ ij The stability and controllability of the oscillation phase difference Δφ ij From the state variable ||zi (t)-z j (t)‖2 and z j (t) is obtained through calculation.

[0075] The adaptive momentum stochastic gradient descent algorithm described in this invention adjusts the learning rate based on the historical sum of squared gradients for each parameter, assigning different learning rates to different parameters.

[0076] Furthermore, in the feature encoding layer, feature fusion of the data is achieved through the synergistic effect of attention mechanism and contrastive representation learning. The specific technical implementation is as follows:

[0077] The input multimodal aircraft data is preprocessed, and the features of each attitude are combined to construct a feature vector set X = {x pitch ,x roll ,x yaw},

[0078] Where, x pitch This represents the characteristic vector related to pitch, including pitch angle error, pitch rate error, pitch angle control signal, pitch-related wind speed and direction, and atmospheric pressure; x roll The eigenvectors related to roll include roll angle error, roll rate, roll angle control signal error, roll-related wind speed and direction, and atmospheric pressure; x yaw The characteristic vectors related to yaw include yaw angle error, yaw rate error, yaw angle control signal, yaw-related wind speed and direction, and atmospheric pressure.

[0079] In the attention mechanism module, internal weights are first assigned to the features of each pose, and then internal weights are assigned to the features of each pose. Weighting is performed, with internal weights used... This means that the Softmax function is used to calculate and highlight the importance of different dimensional features under this pose, that is, to highlight its key dimensions.

[0080]

[0081] in It is the importance score of the i-th feature under pose ο. ) is the importance score of the j-th feature under pose ο;

[0082] Then the local weighted feature of the i-th feature in the o-th pose is: Summing all local weighted features under the same pose yields the local weighted features under the current pose.

[0083] Then, we introduce the global attention vector β = [βpitch ,β roll ,β yaw Global attention weights are allocated among the roll, pitch, and yaw attitudes, with a higher global attention weight assigned to the attitude that has the greatest impact on the current flight state, satisfying the following:

[0084] β ο ≥0.

[0085] Where, β ο Let X represent the global attention weights for the οth pose, and let X represent the global attention weights for pitch, roll, and yaw, respectively. The final output is the feature vector X of the attention mechanism module. t = Integrates information from all postures:

[0086]

[0087] Where, β ο Represents the global attention weights for the o-th pose. This represents the internal weight assigned to the i-th feature within the o-th pose. This represents the i-th feature within the ο-th attitude. The attention mechanism module highlights feature components strongly correlated with the aircraft state and suppresses redundant information;

[0088] X t The input contrastive representation learning module processes the state features of the same pose at different time points. Construct positive sample pairs and combine features between different poses. Construct negative sample pairs ο≠ο′, where t, t+Δt represent different times, t′ represents a time different from t, +Δt represents the time interval, and ο, ο′ represent different postures;

[0089] By constructing positive and negative sample pairs and optimizing feature representations based on the contrastive loss function, the discriminability between different features is enhanced. The enhanced feature vector output by the contrastive representation learning module is then represented as F = [f...]. pitch ;f roll ;f yaw ];[f pitch To compare the pitch-related enhancement features processed by the learning module, f roll To contrast the enhanced features related to rollover after processing by the learning module; f yaw To compare the enhanced features related to yaw after processing by the learning module;

[0090] The contrast loss function L contrastive for:

[0091]

[0092] Where s(·,·) represents feature similarity.

[0093] Enhance feature discriminativeness by optimizing the contrastive loss function.

[0094] Furthermore, the decoupling neural network also includes pulse sequence duration T calculation and data processing. It models the input signal sequence temporally using an event-driven approach: when the difference between adjacent data points exceeds a set threshold, the pulse sequence duration T is updated, allowing the time window to dynamically adapt to signal changes. Next, a phase encoding method is used to convert signal intensity into the relative position of the pulse within the time period. This generates a pulse sequence with a temporal structure through phase mapping, thereby converting the enhanced feature vector F into a temporal code suitable for the dynamic decoupling layer of the decoupling neural network. This code serves as the input signal sequence I to the dynamic decoupling layer.

[0095] Furthermore, the controlled object model is a mathematical model of a controlled object in fields such as hypersonic aircraft and drones.

[0096] The present invention provides an online-trained decoupling network controller for aircraft attitude (also known as a representation model of an aircraft brain-like neural network decoupling controller). Figure 1 This is a flowchart illustrating a brain-like neural network decoupling control method for aircraft. The attitude decoupling network controller is connected to the controlled object model. The entire system is set to input commands related to the three-axis attitude (pitch, yaw, and roll) of the aircraft. The set input commands are compared with the state feedback values ​​from the aircraft model to obtain the error of that state (target control quantity). This error is then combined with other feedback quantities (attitude angular velocity, control signals, etc.) to form raw data X, which is input into the attitude decoupling network controller. After data preprocessing, feature encoding layer processing, and online training, the control output of the aircraft attitude decoupling network controller is obtained, namely the three-axis rudder angles of the aircraft, including pitch, yaw, and roll rudder angles. This invention avoids inter-channel interference, ensuring dynamic decoupling of the pitch, yaw, and roll channels, making the target output of the aircraft attitude control model close to the set input commands.

[0097] This invention improves the dynamic performance and interpretability of a brain-like neural network decoupling controller for aircraft by combining matrix orthogonality and low-rank constraints in its online training framework. Introducing orthogonality constraints faces challenges such as high computational complexity and limited model representation, while low-rank constraints require a trade-off between dimensionality reduction and accuracy. By introducing orthogonal regularization terms into the representation constraints and combining them with efficient matrix factorization, orthogonality and learning ability are balanced. Cross-validation is used to determine the optimal rank value, and tensor decomposition techniques are combined to achieve effective dimensionality reduction. Ultimately, while suppressing redundant interference and reducing model complexity, the decoupling accuracy, robustness, and safety of the system are significantly improved.

[0098] Example 1

[0099] This implementation discloses a brain-like neural network decoupling control method for hypersonic vehicles, addressing the complex nonlinear and strongly coupled dynamic characteristics of hypersonic vehicles. The decoupling neural network includes a feature encoding layer for sensing key vehicle state information and converting it into enhanced features using an attention mechanism module and a contrastive representation learning module; a dynamic decoupling layer for processing enhanced features and achieving dynamic decoupling between roll, pitch, and yaw channels; a control action mapping layer mapping the decoupled features to actual control outputs, generating corresponding three-axis control surface yaw angles, thereby achieving attitude adjustment of the vehicle; the dynamic decoupling layer consists of N... o = It consists of 100 coupled oscillating neurons, and the action mapping layer consists of 3 neurons;

[0100] A decoupled network controller for aircraft attitude is constructed. The decoupled neural network is trained to effectively represent and decouple the aircraft attitude control signals. The controlled object model is the mathematical model of aircraft attitude, the control system is attitude closed-loop control, and the target control quantity is three attitude angles, so as to achieve efficient decoupling and independent control between multiple channels.

[0101] The steps of this method are:

[0102] The first step is to construct a decoupled neural network:

[0103] The input to the decoupled neural network is the aircraft state signal sequence X = [x1, x2, ... x...]. n ], n=12 represents the dimensions of the signal sequence, namely the three-axis angular errors (pitch, yaw, roll), the three-axis angular velocity errors (pitch, yaw, roll), the three-axis control signals (pitch, yaw, roll), wind speed, wind direction, and atmospheric pressure; the output is the aircraft's three-axis rudder angle signal sequence Y=[y1,y2,…y m ], m=3 control signal sequence dimensions, namely pitch rudder angle, yaw rudder angle and roll rudder angle.

[0104] By collecting data from multiple sensors, and after preprocessing such as time alignment, noise reduction, error calculation, and parameter estimation, key features such as pitch angle error, angular rate error, wind speed and direction, atmospheric pressure, body parameter deviation, and aerodynamic moment deviation are combined to form a structured feature vector for each time step. Feature vectors related to each attitude are then extracted to construct a feature vector set X = {x}. pitch ,x roll ,x yaw The attention mechanism module assigns weights to each feature vector, and performs weighted summation on the feature vectors within each pose. Calculate the weights using Softmax:

[0105]

[0106] in It is the importance score of the i-th feature under pose o.

[0107] Standardization and adaptive weight allocation are used to integrate multi-source data to improve the ability to extract key features.

[0108] Then, weights are assigned among the three attitudes, emphasizing the attitude that has the greatest impact on the current flight state. A global attention vector β = [β...] is introduced. pitch ,β roll ,β yaw ],satisfy:

[0109] β ο ≥0.

[0110] The final output feature of the attention mechanism module is the weighted feature X. t :

[0111]

[0112] Highlight the feature components that are strongly correlated with the aircraft's state and suppress redundant information.

[0113] Next, the weighted features are input into the contrastive representation learning module, which constructs positive sample pairs (different states of the same pose).

[0114] ) and negative sample pairs (different pose features) ),ο≠ο′), based on the contrastive loss function to optimize feature representation:

[0115]

[0116] Where s(·,·) represents feature similarity.

[0117] The input signal sequence is modeled temporally using an event-driven approach: when the difference between adjacent data points exceeds a set threshold, the pulse sequence duration T is updated, allowing the time window to dynamically adapt to signal changes. Next, a phase encoding method is employed to convert signal intensity into the relative position of pulses within a time period. This involves generating a pulse sequence with a temporal structure through phase mapping, thereby transforming the enhanced feature vector into a temporally encoded input dynamic decoupling layer suitable for coupled neural networks.

[0118] These time-encoded pulse sequences are processed using a coupled oscillating neuron model. The state variables of each coupled oscillating neuron are... Following the nonlinear coupled oscillatory differential equation:

[0119]

[0120] in The intrinsic dynamic matrix of the coupled oscillating neuron determines its fundamental oscillation frequency ω. i By optimizing the algorithm design to match the temporal characteristics required for decoupling; Let f(·) be the input coupling matrix, and f(·) be the nonlinear activation function used to map the external control error signal to the oscillation amplitude adjustment. Let be the coupling coefficient matrix between coupled oscillating neurons i and j, which satisfies the regularization constraint. (j represents other coupled oscillating neurons connected to coupled oscillating neuron i, and γ is a preset coupling strength threshold), to balance decoupling robustness and information exchange capability; the system adjusts {ω i W ij The state evolution of coupled oscillating neurons should satisfy the decoupling objective, i.e., when the system reaches steady state, the state difference between adjacent coupled oscillating neurons should be ||z||. i (t)-z j (t)‖2 decays to within the preset error range, and the oscillation phase difference Δφ output by each coupled oscillating neuron ij It corresponds one-to-one with the decoupling requirements of the aircraft control channel.

[0121] The representation constraint L with orthogonality and low rank properties is constructed as follows:

[0122] L=p1L o +p2L l +p3L s +p4L d

[0123]

[0124] L l =||ε|| nuc

[0125]

[0126] By adjusting the topology of the decoupled neural network using representational constraints L, the dynamic decoupling process between the attitude control channels of the aircraft is optimized.

[0127] The second step is to train the decoupling neural network.

[0128] A decoupled neural network is trained using online incremental learning, with network weights continuously updated through interaction with the aircraft attitude model. The error E of the s-th sample is... s As shown below:

[0129] E s =[e1,e2,e3]

[0130] Among them, e i This is the real-time control error vector of the i-th channel (including attitude angle deviation and angular rate deviation), where i takes the values ​​1, 2, and 3.

[0131] The output of the decoupled neural network is the rudder angle Y = [y1, y2, y3] of the three rudder surfaces. Calculate the loss of the decoupled neural network. e represents the angle and angular velocity errors of the three attitude channels. In this embodiment, the number of error signals is six (the number of error signals in the input depends on the number of target control variables; for example, if the control system is attitude control, then it consists of three attitude angles and three attitude angular velocities). The gradient of this loss function with respect to the network parameters is calculated using the error backpropagation algorithm, and the network weights are adjusted accordingly. The learning rate and momentum factor are dynamically adjusted using optimization algorithms such as stochastic gradient descent to gradually optimize the network parameters, thereby obtaining a representation model of the decoupled controller of the aircraft's neuromorphic neural network.

[0132] The third step is to use the trained representation model of the aircraft's brain-like neural network decoupling controller for the decoupling control of the aircraft.

[0133] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. An aircraft brain-like neural network decoupling control method, characterized in that, The method comprises the following steps: Obtaining multi-dimensional information of each channel related to different attitudes of the aircraft; Building a decoupling neural network, which comprises a feature encoding layer, a dynamic decoupling layer and a control action mapping layer, the input of the decoupling neural network being a multi-channel control error fusion signal, and the output of the decoupling neural network being independent control instructions of each channel after decoupling; The feature encoding layer firstly outputs a weighted feature vector by using an attention mechanism module as a whole, and then further optimizes the weighted feature vector by using a contrastive representation learning module to generate a strengthened feature vector F, which is converted into a time sequence encoding suitable for the dynamic decoupling layer of the decoupling neural network and input into the dynamic decoupling layer; The dynamic decoupling layer is composed of N coupled oscillation neurons and has orthogonal characteristics and a representation constraint L with low rank characteristics, and the constraint L is: L = p1L o + p2L l + p3L s + p4L d L l =||ε| nuc where p1, p2, p3, p4 are weight coefficients corresponding to the priority of different constraint terms; L o is the orthogonality constraint, is the parameterized orthogonal projection matrix, θ is the parameter matrix of the decoupling neural network, d represents the dimension of the feature; I3 is a 3x3 unit matrix; L l is the low-rank constraint, and ε is the residual tensor; L s is the independence constraint; M(f i ,f j ) is the mutual information of the decoupled features f i ,f j ; L d is the continuity constraint, t represents the time step; Φ(θ t+1 ), Φ(θ t ) respectively represent the parameterized orthogonal projection matrix at the t+1 time step and the t time step; f i ,f j are the decoupled features of the i-th neuron and the j-th neuron respectively; Connecting the decoupling neural network with a controlled object model, and training the decoupling neural network online, wherein the loss function is defined as: wherein m is the dimension of the aircraft control channel, is the real-time control error vector of the i-th channel, n i is the dimension of the error vector of the i-th control channel, L is a constraint term with orthogonal and low-rank characteristics Using the trained aircraft attitude decoupling network controller for decoupling control of the aircraft.

2. The method of claim 1, wherein, The online training mode is a reinforcement learning mode or an online incremental learning mode.

3. The method of claim 1, wherein, Each coupled oscillation neuron follows a nonlinear coupled oscillation differential equation: wherein, is the intrinsic dynamic matrix of the coupled oscillatory neuron, determines its base oscillation frequency ω i i is directly mathematically related to the base oscillation frequency ω i , the eigenvalues of A i determine the oscillation behavior of the system, ω i is controlled and optimized by adjusting A i ; B i is the input coupling matrix; f(·) is a nonlinear activation function; is the coupling coefficient matrix between coupled oscillatory neurons i and j, which satisfies the regularization constraint j represents other coupled oscillatory neurons connected to coupled oscillatory neuron i, γ is a preset coupling strength threshold; n represents the dimension number of the state variable of each coupled oscillatory neuron; t represents the time variable; is the first-order derivative of the state variable; z i (t), z j (t) is the state variable.​ 4. The method of claim 1, wherein, The multi-dimensional information of each channel related to different attitudes includes attitude angle error, angular rate error, three-axis control signal, wind speed, wind direction and atmospheric pressure.

5. The method of claim 1, wherein, In the feature encoding layer, the feature fusion of data is realized by the synergistic effect of attention mechanism and contrastive representation learning, specifically as follows: The multi-dimensional information of each channel related to different postures is preprocessed, and the feature vectors of each posture are combined to construct a feature vector set X = {x pitch ,x roll ,x yaw} wherein x pitch represents a feature vector related to pitch, including pitch angle error, pitch angle rate error, pitch angle control signal, pitch related wind speed and wind direction, atmospheric pressure; x roll represents a feature vector related to roll, including roll angle error, roll angle rate, roll angle control signal error, roll related wind speed and wind direction, atmospheric pressure; x yaw represents a feature vector related to yaw, including yaw angle error, yaw angle rate error, yaw angle control signal, yaw related wind speed and wind direction, atmospheric pressure; In the attention mechanism module, first, the features of each posture are assigned an internal weight, and the features inside each posture are weighted , and the internal weight is represented by , which is calculated using a Softmax function. wherein is the importance score of the i-th feature at pose o, ) is the importance score of the j-th feature at pose o; The local weighted feature of the i-th feature in the 0-th pose is The local weighted features of all the features in the same pose are summed to obtain the local weighted feature in the current pose Then introduce the global attention vector β = [β pitch ,β roll ,β yaw ] to distribute the global attention weight among the three attitudes of roll, pitch and yaw, satisfying: β ο ≥0. wherein β ο is the global attention weight in the i-th pose, respectively represents the global attention weight of the pitch, roll, and yaw, and the output feature vector X of the attention mechanism module is finally output t : X t inputting the contrastive representation learning module, state features of the same gesture at different time points constructing a positive sample pair, features between different gestures o≠o' constructing a negative sample pair, where t, t+Δt represent different times, t' represents a time different from t, Δt represents a time interval, o, o' represent different gestures; By constructing positive sample pairs and negative sample pairs, the feature representation is optimized based on a contrastive loss function to enhance the distinguishability between different features, and the strengthened feature vector output by the contrastive representation learning module is represented as F. The contrastive loss function L contrastive is: Where s(·,·) represents the feature similarity.

6. The method of claim 5, wherein, The decoupling neural network further comprises pulse sequence time length T calculation and data processing, which converts the strengthened feature vector F into a time sequence encoding suitable for the dynamic decoupling layer of the decoupling neural network, and uses the time sequence encoding as an input signal sequence input into the dynamic decoupling layer.

7. The method according to any of claims 1 to 6, characterized in that The controlled object model is a mathematical model of a controlled object in the field of hypersonic aircraft or unmanned aerial vehicle. 8.A computer readable storage medium having stored thereon a computer program, which, when executed by a processor, can implement the content of the method of any one of claims 1-7.

Images