Sensing, storage and computing integrated device for posture control

Abstract

The present invention relates to a sensing, storage and computing integrated device for posture control, including a hardware neural network formed by a memristor array, the hardware neural network accepts posture analog signal input from a posture sensor and an external posture control instruction, and the hardware neural network outputs an analog control signal to a controlled object driver; the network structure of the hardware neural network corresponds to the network structure of a posture control network trained by software, and the conductance value of each memristor in the hardware neural network is determined based on the weight value of the corresponding neuron in the posture control network. The present invention can achieve a full-angle range convergence effect by integrating the algorithm design of a high-performance posture controller with the design of an end-to-end hardware neural network, and has a faster finite time convergence capability, and the analog signal system has a higher response bandwidth.

Description

Technical Field

[0001] The present invention relates to the technical field of attitude control algorithms and devices, and in particular to a sensing, storage and computing integrated device for attitude control. Background Art

[0002] The attitude control of a suspended rigid body is an important issue in the field of system motion control and is widely used in aerospace and underwater autonomous systems. Fixed-wing and multi-rotor aircraft, satellites, submarines, etc. all involve attitude control of a suspended rigid body. Usually, the attitude control loop of a suspended rigid body is located in the innermost layer of the overall control framework, which requires as large a control range and high-speed response capability as possible.

[0003] Attitude control of suspended rigid bodies remains challenging because their configuration space is not Euclidean. The rigid body configuration space consists of a set of 3×3 rotation matrices, which is a compact manifold that is not globally differentiable from any three-dimensional Euclidean space. In addition, on compact manifolds, global stability cannot be achieved with the help of smooth static controllers due to topological obstacles. In terms of the convergence region, the best smooth and static controllers can achieve approximate global convergence. Approximate global convergence means that the set of points in the state space, from which the vector field of the closed-loop system cannot converge to the desired equilibrium point, is "small" in the sense of zero Lebesgue measure. It should be emphasized that although approximate global convergence can meet general attitude control requirements in a theoretical sense, in practical applications, the global convergence of the controller cannot be guaranteed due to the influence of measurement errors, etc.

[0004] In most cases, to design, analyze and implement geometric controllers, users need to understand the core concepts of Lie groups and differentiable manifolds. Designing a controller to ensure stronger convergence results, such as finite-time convergence (asymptotic convergence without time uncertainty), is not easy to achieve on manifolds. In addition, the current attitude inner-loop controller architecture is still based on traditional embedded systems. Its running speed is limited by communication delays and iterative calculation time, while high-performance computing units bring high energy consumption. The problems faced by traditional embedded attitude control systems, such as low response bandwidth, long convergence time, limited angle range, and high power consumption, make it difficult to meet stable applications in special environments. For example, aerospace devices with high requirements for response speed and accuracy are not suitable for the application of traditional embedded attitude control systems. At least targeted optimization of single performance is required. Summary of the invention

[0005] The attitude control controller in the prior art has problems such as low response bandwidth, long convergence time, and limited angle range. It is difficult to apply to fields such as aerospace that have high requirements for response speed and accuracy. In this regard, the present invention provides an integrated sensing, storage, and computing device for attitude control.

[0006] The technical solution of the present invention provides a sensing, storage and computing integrated device for posture control, including a hardware neural network formed by a memristor array, the hardware neural network accepts posture analog signal input from a posture sensor and an external posture control instruction, and the hardware neural network outputs an analog control signal to a controlled object driver;

[0007] The network structure of the hardware neural network corresponds to the network structure of the posture control network trained by the software, and the conductance value of each memristor in the hardware neural network is determined based on the weight value of the corresponding neuron in the posture control network.

[0008] Preferably, the posture control network is obtained independently from the hardware neural network training, the memristor neurons of the hardware neural network correspond one-to-one to the node neurons in the posture control network, and the conductance values ​​of the memristor neurons directly correspond one-to-one to the trained weight values ​​of the corresponding node neurons.

[0009] Preferably, the posture control network is obtained independently of the hardware neural network training, the memristor neurons of the hardware neural network correspond one-to-one to the node neurons in the posture control network, and the initial values ​​of the conductance values ​​of the memristor neurons directly correspond one-to-one to the trained weight values ​​of the corresponding node neurons; and enhanced training is performed based on the integrated sensing, storage and computing device system for posture control to fine-tune the conductance values ​​of the memristor neurons.

[0010] Preferably, the training data set of the attitude control network is obtained based on a finite time convergence attitude controller:

[0011] Constructing a supervised training dataset based on a finite time convergence attitude controller ,in, is the state quantity of the controlled system, is the analog control signal of the corresponding system, is the inertial navigation acceleration signal, is the inertial navigation angular velocity signal, As observation quantities, the three express the posture state of the controlled object; is the external attitude control command, As network input; is a simulated control signal used as a training label.

[0012] Preferably, first, based on a finite time convergence attitude controller, a supervised training data set is obtained and constructed ,

[0013] Then construct a complex reasoning network model ,in, is the current drive output, To drive the output at the next moment, is the current system state, is the inference network; is the Gaussian distribution function, where is the mean network, A variance network; training the complex reasoning network model based on the supervised training data set; generating simulation training data that can be used for the posture control network training based on the trained complex reasoning network model;

[0014] Finally, at least the simulation data set generated by the complex reasoning network model is used as a training data set to train the posture control network.

[0015] Preferably, in the step of generating a simulation data set for training the posture control network based on the trained complex reasoning network model, random interference simulating the noise characteristics of the hardware neural network is added to the input data of the complex reasoning network model.

[0016] Preferably, at least the simulation data set generated by the complex reasoning network model is used as a training data set in the step of training the posture control network to mix the simulation data set and the supervised training data set to construct a mixed data set for training the posture control network.

[0017] Preferably, in the finite time convergence controller used to construct the supervised training data set, the dynamic equation of the controlled object is:

[0018]

[0019] in, is the estimated value of the rotational inertia of the controlled object; represents the posture matrix; Provides driving torque for the system; Represents the rotation speed vector Construct an antisymmetric matrix;

[0020] Angular velocity error:

[0021]

[0022] in, Specify rotation speed for three dimensions; To specify the current posture;

[0023] Construct the specified posture generation law of segmented changes such as:

[0024]

[0025] in, are the first, second and third columns of the desired posture matrix at time data, The orientation angle The constructed orientation vector, is the expected acceleration, is the acceleration due to gravity, is the segmented command acceleration, is the segmented command z-axis attitude angle, For estimation Time acceleration, For estimation Moment angle;

[0026] For angular velocity error and expected attitude , construct the angular velocity error change rate The expression is as follows:

[0027]

[0028]

[0029] Among them, the system angular acceleration Take as control quantity ;

[0030] For control volume , construct control quantity The expression is as follows,

[0031]

[0032] in, Driving torque for physical systems; The components are expressed as follows:

[0033]

[0034] in, , , are the attitude angle errors of the x-axis, y-axis and z-axis respectively, , , are the angular velocity errors of the x-axis, y-axis, and z-axis respectively; for The feedback gain matrix of for An exponential gain matrix, and each element of which is a positive number less than 1; There are the following expressions:

[0035] .

[0036] Preferably, the real system mathematical expression of the controlled object and the corresponding observer mathematical expression are respectively:

[0037]

[0038]

[0039] in, is the system state quantity; is the attitude quaternion; 3D velocity vector; is the three-dimensional displacement vector; is the three-dimensional angular velocity The corresponding offset; for The random walk driving force of is the three-dimensional acceleration The corresponding offset; for The random walk driving force of is the gravity acceleration bias; Represents the estimated state quantity corresponding to the system state quantity;

[0040] Constructing angular velocity estimates and speed estimation as follows,

[0041]

[0042] in, is the angular velocity sensing noise, For acceleration sensing noise, is the original angular velocity output of the inertial navigation sensor; It is the original acceleration output of the inertial navigation sensor;

[0043] Among them, the attitude transfer matrix and the corresponding rotation matrix Expressed as:

[0044]

[0045]

[0046] in, , is the constraint between the components of the quaternion, , , They are , , Axis estimated angular velocity;

[0047] Construct the state error of the inertial navigation sensor system:

[0048] ,

[0049] The state transfer matrix corresponding to the state error of the inertial navigation sensor system is:

[0050]

[0051]

[0052] in, Represents a vector Construct the antisymmetric matrix, for The identity matrix, for All zero matrix;

[0053] The covariance matrix of the inertial navigation sensor system is as follows:

[0054]

[0055] According to the state error of the inertial navigation sensor system and external correction Constructing Correction Error :

[0056]

[0057] in, is an estimated value; is the transfer matrix from state to observation error.

[0058] Preferably, in the hardware neural network, the multiplication operation is performed at each row and column intersection in the memristor array according to Ohm's law; the intersection of the i-th row and the j-th column is connected through a memristor, and the conductance value of the memristor is expressed as , the input of the i-th row of the array is , the sum of row and column currents is enforced by Kirchhoff's law, and the sum of the currents read by the devices along the columns , to achieve neuronal excitation The forward propagation required ; The error term of the integral realization along the line The sum required for back propagation .

[0059] The present invention provides a specialized attitude control chip for full-angle range and high-bandwidth attitude control requirements through the design of an end-to-end hardware neural network, making the attitude control system structure more compact; compared with the embedded neural network implementation method, the present invention maps the parameters of the attitude control network to the conductance value of the non-volatile memristor, thereby being able to maintain the parameters of the neural network after power failure. At the same time, the analog circuit solution avoids the calculation delay and power consumption overhead caused by data transfer between the memory and the calculator, thereby improving the overall execution speed from perception to control.

[0060] The technical solution of the present invention can achieve a full-angle range convergence effect and has a faster finite-time convergence capability by integrating the algorithm design of a high-performance attitude controller with the design of an end-to-end hardware neural network. At the same time, the analog signal system has a higher response bandwidth. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a structural schematic diagram of the sensing, storage and computing integrated device for posture control of the present invention;

[0062] Figure 2 It is a schematic diagram of the structure of the hardware neural network of the sensing, storage and computing integrated device for posture control of the present invention;

[0063] Figure 3 It is a schematic diagram of a posture control network training process for determining a hardware neural network in the posture control-oriented sensing, storage and computing integrated device of the present invention;

[0064] Figure 4 Convergence effect comparison of the finite time convergence controller of the present invention in a static target scenario;

[0065] Figure 5 Convergence effect comparison of the finite time convergence controller of the present invention in a dynamic target scenario;

[0066] Figure 6 It is a schematic diagram of the posture control network training process for determining the hardware neural network in the posture control-oriented sensing, storage and computing integrated device of the present invention. DETAILED DESCRIPTION

[0067] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. In this specification, the size ratios in the drawings do not represent the actual size ratios, but are only used to reflect the relative position relationship and connection relationship between the components. Components with the same name or the same number represent similar or identical structures and are only for illustrative purposes.

[0068] The technical solution of the present application provides an integrated sensing, storage and computing device for posture control, which provides an end-to-end posture control neural network device based on a memristor array, and can implement the entire process of posture control based on on-chip hardware. Since it is not necessary to complete the calculation tasks of the corresponding algorithm based on a general-purpose computing unit, the utilization rate of its computing resources is higher, and low-power operation of the system can be achieved. In addition, due to the improvement of the robustness and stability of the neural network algorithm, the corresponding hardware neural network system also has the characteristics of short convergence time and high control accuracy.

[0069] Figure 1 This is a schematic diagram of the structure of the integrated sensing, storage and computing device for posture control of the present application. The integrated sensing, storage and computing device for posture control includes a hardware neural network formed by a memristor array. The schematic diagram of the hardware neural network structure is shown in FIG. Figure 2 The hardware neural network receives the attitude analog signal input from the attitude sensor and the external attitude control instruction, and outputs the analog control signal to the controlled object driver. The network structure of the hardware neural network corresponds to the network structure of the attitude control network trained by the software, and the conductance value of each memristor in the hardware neural network is determined based on the weight value of the corresponding neuron in the attitude control network.

[0070] The hardware neural network actually replicates the posture control network trained by software in hardware. In the hardware neural network composed of memristors, multiplication operations are performed at each row and column intersection in the memristor array according to Ohm's law. The intersection of the i-th row and j-th column is connected by a memristor, and the conductance value of the memristor is expressed as , the input of the i-th row of the array is , the sum of row and column currents is enforced by Kirchhoff's law, and the sum of the currents read by the devices along the columns , which is equivalent to realizing neuron excitation The forward propagation required Similarly, the integration along the line realizes the error term The sum required for back propagation . That is, the following equation is equivalent between the two models.

[0071]

[0072]

[0073] Although the posture control network is trained independently of the hardware neural network, in order to be transplanted into the memristor array technology, its input and output, network structure, activation function, etc. still need to be improved in a targeted manner according to the actual characteristics of the hardware neural network, rather than directly copying the existing neural network learning algorithm. Based on the given data training and verification, the posture control network can naturally obtain a neural network model with a good generalization degree, and directly mapping the neural network model to the hardware neural network can also expect to obtain the expected basically consistent prediction output. The mapping here means not only ensuring that the hardware neural network and the posture control model have the same network structure, but also that the memristor neurons and the node neurons in the posture control model can correspond one to one, and that the conductance value of the memristor is directly corresponding to the trained weight value of the corresponding node neuron. The memristor neuron is generally the smallest node unit connecting the row line and the column line in the hardware neural network, which usually includes at least one memristor device, and in the prior art may also include other electrical devices connected in series with the memristor device, such as transistors, gate tubes, etc., and it is not ruled out that the smallest unit formed by two memristor devices is formed.

[0074] Generally, the hardware neural network obtained under the mapping of the trained posture control network is sufficient, but in practical applications, there is also a need to optimize the learning effect of the hardware neural network based on specific usage scenarios. After all, when training the posture control neural network, it is usually based on a more generalized application scenario. For the consideration of input-output ratio, it will not be directly trained for segmented applications, especially specific scenarios, in the early stage. Therefore, in order to improve the control effect of the hardware neural network in the business scenario, the network weights obtained by the initial mapping can be updated and fine-tuned based on the business scenario on the basis of the mapped hardware neural network through retraining based on the business scenario to make it more in line with the current business scenario, providing better control accuracy for the specific object. Therefore, from this perspective, the conductance value of each memristor in the hardware neural network referred to in the previous article is determined based on the weight value of the corresponding neuron in the posture control network, which not only refers to the direct mapping of the conductance value of each memristor to the weight value of the corresponding neuron in the posture control network, but also refers to the result of updating and fine-tuning the process based on the business scenario after each memristor obtains the conductance value of the weight value of the corresponding neuron directly mapped in the posture control network. The updating and fine-tuning process is based on the applied business scenario, the business system, and the controlled object. The analog control signal and its corresponding attitude simulation signal and the data set of the external attitude control instruction are trained again through the hardware neural network. During the training process, the fine-tuning update of the memristor conductance value is achieved through the back propagation of the hardware neural network.

[0075] like Figure 2As shown, the analog control signal output by the hardware neural network is a current signal. A power amplifier can be set before outputting it to the controlled object driver to amplify the output current signal to meet the input requirement of the controlled object. The power amplifier generally provides an amplified voltage control signal according to the needs of the controlled object driver, which is used to drive the controlled object driver to adjust the posture of the controlled object.

[0076] The training data set of the attitude control network is obtained based on the finite time convergence attitude controller. Simply put, a set of output data corresponding to the input data and attitude can be obtained through the operation of the finite time convergence attitude controller. That is, for a given attitude and the input of the external attitude control instruction, the preset finite time convergence controller can always obtain a corresponding set of analog control signal outputs. Therefore, a sufficient supervised training data set can be constructed based on the finite time convergence attitude controller. ,in, is the state quantity of the controlled system, is the analog control signal of the corresponding system, is the inertial navigation acceleration signal, is the inertial navigation angular velocity signal, As observation quantities, the three express the posture state of the controlled object; is the external attitude control command, and the combination of the two is used as the network input. is the simulated control signal of the corresponding system, which serves as the training label.

[0077] In the case where the training data set is determined, even if the specific training process of the posture control network based on the established structure is not disclosed in the present application, the technical personnel in this field are actually sufficient to obtain the posture control network with updated network weights after training through conventional training means based on the known content in the field.

[0078] Naturally, the obtained supervised training dataset can also be It is a common practice for those skilled in the art to directly train the posture control network to obtain the weight values ​​of each node in the model. However, in this application, since the posture control network needs to be mapped to the hardware neural network, the results of direct training are often significantly different from the expected control effect according to experiments. In other words, the control stability of the trained posture control network after being applied to the hardware neural network cannot reach the original control accuracy of the posture controller.

[0079] Therefore, the technical solution of the present application also provides a network training step for optimizing the effect. The training step of the posture control network of the present application is carried out in two steps. The training steps are as follows: Figure 3 Based on the finite time convergence attitude controller, a supervised training data set is obtained and constructed through physical testing and other methods. ,in, is the state quantity of the controlled system, is the analog control signal of the corresponding system, is the inertial navigation acceleration signal, is the inertial navigation angular velocity signal, As observation quantities, the three express the posture state of the controlled object; is the external attitude control command, and the combination of the two is used as the network input. is the simulated control signal of the corresponding system, which serves as the training label.

[0080] In order to generate more generalized and expanded training data beyond the physical method, we first design a complex reasoning network model ,in, is the current drive output, To drive the output at the next moment, is the current system state, The inference network may be an existing neural network model or other deep learning models, such as a Transformer architecture, LSTM architecture, RNN architecture, and other neural network models. is the Gaussian distribution function, where is the mean network, The invention is a variance network, and its purpose is to expand the training data set in a targeted manner. The invention is to expand the training data set in a targeted manner. The complex reasoning network model is trained based on the initially obtained supervised training data set. Then a large amount of simulation training data is generated through the complex reasoning network model. The simulation training data can be adjusted and constructed in a targeted manner according to the supervised training data set, and the attitude range of the controlled object and the range of the external attitude control instructions can be determined according to the specific situation of the application scenario of the finite time convergence controller, and the possible input of the complex reasoning network model is determined within this range, and the corresponding output, that is, the corresponding simulation control signal, is obtained after calculation by the complex reasoning network model.

[0081] The loss function used during training is as follows:

[0082]

[0083] in, is the loss function, is the variance network, is the observed data at time n, Output data for n moments, Output drive for the next moment, is the inverse solution of the variance network, and N is the number of all data.

[0084] The reason why it is necessary to design and construct the process of complex reasoning network model in the preparation stage is not only to expand the range of data sets used for the posture control model, but also to provide the possibility of targeted transformation of the generated simulation data set through the complex reasoning network model. In this case, the noise characteristics and other characteristics of the hardware neural network can be simulated by artificially randomly interfering with the data in the process of generating the simulation data set, such as the small random attenuation of the voltage value representing the input information in the memristor neural network, so that it may be closer to the real state of the hardware neural network when used for posture control training, so that after the node weights of the posture control network are sent to the hardware neural network, the conductance value matrix corresponding to the memristor array in the hardware neural network can more realistically match the actual characteristics of the memristor array when used, providing a more stable and accurate analog control signal output. It is preferred to directly apply random noise to the input data for generating the simulation data set, rather than directly adding random information to the input and output data after generation, because the random noise added in this way does not have a correlation between the input information and the output information, and the subsequent posture control training cannot reflect the mutual influence between the hardware neural network and the input, so it cannot be ensured whether the control accuracy of the analog control signal of the posture control network after training can be effectively improved.

[0085] Then, the posture control network is trained based on the simulation data set generated by the complex reasoning network model. Training the posture control network based on the simulation data set means at least using the simulation data set to train the posture control network. In this application, it can be considered to rely solely on the simulation data set to train the posture control network posture control strategy network On the other hand, training Figure 6 As shown in the reinforcement learning diagram, it is more preferable to mix the simulation data set and the original supervised training data set to construct a balanced distribution mixed data set for the training of the posture control network. Its purpose is to provide both original data that is faithful to the system control strategy and randomized simulation data that represents the analog circuit characteristics of the memristor neural network, thereby achieving the purpose of taking both into account. Compared with using only the simulation data set, the effect is closer to the actual control effect and is less prone to underfitting.

[0086] The following loss function can be used in the training process:

[0087]

[0088]

[0089] in, About obey The mean of the distribution, for About mixed datasets The mean of is the reward function (optional, such as system posture tracking error, filtered tracking error, etc.); is the temperature weight of the training process; For the forgetting factor; is the value network of the training process, is the corresponding parameter, After smoothing (Optional, smoothing method such as exponential decay filtering method, etc.); The decision network adapted for the final memristor, is the training parameter.

[0090] The system dynamics model used in the process of determining the supervised training data set on which the attitude control network relies is a full-angle range finite-time convergence controller based on an iterative Kalman observer. In this finite-time convergence controller, the dynamic equation of the controlled object is:

[0091]

[0092] in, is the estimated value of the rotational inertia of the controlled object; represents the posture matrix; Provides driving torque for the system; Represents the rotation speed vector Construct an antisymmetric matrix.

[0093] Construct the angular velocity error for the controlled object as follows,

[0094]

[0095] in, Specify rotation speed for three dimensions; To specify the current posture.

[0096] In order to achieve full-angle range control capability, for the above specified current posture , the generation law of the specified posture for segmented changes is constructed as follows,

[0097]

[0098] in, are the first column of the desired posture matrix ( axis), the second column ( axis), and the third column ( axis) at time data, The orientation angle The constructed orientation vector, is the expected acceleration, is the acceleration due to gravity, is the segmented command acceleration, The z-axis attitude angle of the segment command is determined by the command attitude get, For estimation Time acceleration, For estimation Moment angle.

[0099] For angular velocity error and expected attitude , construct the angular velocity error change rate The expression is as follows:

[0100]

[0101]

[0102] Among them, the system angular acceleration Take as control quantity .

[0103] For control volume , construct control quantity The expression is as follows,

[0104]

[0105] in, Driving torque for physical systems; The components are expressed as follows:

[0106]

[0107] in, , , Respectively 1, 2, 3 axis ( , , Axis) attitude angle error, , , Respectively 1, 2, 3 axis ( , , axis) angular velocity error, for The feedback gain matrix of for An exponential gain matrix with each element being a positive number less than 1; There are the following expressions:

[0108]

[0109] In order to establish the kinematic model of the controlled object based on the iterative Kalman observer, it is necessary to establish the corresponding forward kinematic expression of the inertial navigation system state observer according to the system attitude input, including the real system mathematical expression of the controlled object and the corresponding observer mathematical expression, which are:

[0110]

[0111]

[0112] in, is the system state quantity; is the attitude quaternion; 3D velocity vector; is the three-dimensional displacement vector; is the three-dimensional angular velocity The corresponding offset; for The random walk driving force of is the three-dimensional acceleration The corresponding offset; for The random walk driving force of is the gravity acceleration bias; Represents the estimated state quantity corresponding to the system state quantity.

[0113] Constructing angular velocity estimates and speed estimation as follows,

[0114]

[0115] in, is the angular velocity sensing noise, For acceleration sensing noise, is the original angular velocity output of the inertial navigation sensor; It is the raw acceleration output of the inertial navigation sensor.

[0116] Among them, the attitude transfer matrix and the corresponding rotation matrix Expressed as:

[0117]

[0118]

[0119] in, , is the constraint between the components of the quaternion, , , They are , , Axis estimated angular velocity.

[0120] Construct the state error of the inertial navigation sensor system:

[0121]

[0122] in, is the state quantity With the state estimator The difference vector of is the state quantity The covariance matrix of .

[0123] The state transfer matrix corresponding to the state error of the inertial navigation sensor system is:

[0124]

[0125]

[0126] in, Represents a vector Construct the antisymmetric matrix, for The identity matrix, for All-zero matrix.

[0127] The covariance matrix of the inertial navigation sensor system is as follows:

[0128]

[0129] in, is the estimated state error The covariance matrix of is the state quantity The covariance matrix of the inertial navigation sensor system is and external correction (Optional include vision / compass / radar / GPS, etc.) Build correction error :

[0130]

[0131] in, is an estimated value, is the observed noise, is the transfer matrix from state to observation error.

[0132] Figure 4 , Figure 5The effect of the finite time convergence controller constructed based on the kinematic model of the iterative Kalman observer on the digital embedded system, where the control instructions used are static pitch, roll, and yaw angles. It can be seen that the proposed finite time convergence controller has a faster convergence speed and a higher response bandwidth.

[0133] The above content only describes the preferred implementation mode of the present invention, and does not limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solution of the present invention by ordinary technicians in this field should fall within the protection scope determined by the claims of the present invention.

Claims

1. A sensing, storage and computing integrated device for posture control, characterized in that: The hardware neural network includes a hardware neural network formed by a memristor array, the hardware neural network receives a posture analog signal input from a posture sensor and an external posture control instruction, and the hardware neural network outputs an analog control signal to a controlled object driver; The network structure of the hardware neural network corresponds to the network structure of the posture control network trained by the software, and the conductance value of each memristor in the hardware neural network is determined based on the weight value of the corresponding neuron in the posture control network; The training data set of the attitude control network is obtained based on the finite time convergence attitude controller: Constructing a supervised training dataset based on a finite time convergence attitude controller ,in, is the state quantity of the controlled system, is the analog control signal of the corresponding system, is the inertial navigation acceleration signal, is the inertial navigation angular velocity signal, As observation quantities, the three express the posture state of the controlled object; is the external attitude control command, As network input; To simulate control signals, used as training labels; First, based on the finite time convergence attitude controller, a supervised training dataset is obtained and constructed. , Then construct a complex reasoning network model ,in, is the current drive output, To drive the output at the next moment, is the current system state, is the inference network; is the Gaussian distribution function, where is the mean network, A variance network; training the complex reasoning network model based on the supervised training data set; generating simulation training data that can be used for the posture control network training based on the trained complex reasoning network model; Finally, at least the simulation data set generated by the complex reasoning network model is used as a training data set to train the posture control network; In the finite time convergence controller used to construct a supervised training data set, the dynamic equation of the controlled object is: in, is the estimated value of the rotational inertia of the controlled object; represents the posture matrix; Provides driving torque for the system; Represents the rotation speed vector Construct an antisymmetric matrix; Angular velocity error: in, Specify rotation speed for three dimensions; To specify the current posture; Construct the specified posture generation law of segmented changes such as: in, are the first, second and third columns of the desired posture matrix at time data, The orientation angle The constructed orientation vector, is the expected acceleration, is the acceleration due to gravity, is the segmented command acceleration, is the segmented command z-axis attitude angle, For estimation Time acceleration, For estimation Moment angle; For angular velocity error and expected attitude , construct the angular velocity error change rate The expression is as follows: Among them, the system angular acceleration Take as control quantity ; For control volume , construct control quantity The expression is as follows, in, Driving torque for physical systems; The components are expressed as follows: in, , , are the attitude angle errors of the x-axis, y-axis and z-axis respectively, , , are the angular velocity errors of the x-axis, y-axis, and z-axis respectively; for The feedback gain matrix of for An exponential gain matrix, and each element of which is a positive number less than 1; There are the following expressions: 。 2. The sensing, storage and computing integrated device for posture control according to claim 1, characterized in that: The posture control network is obtained independently from the hardware neural network training, the memristor neurons of the hardware neural network correspond one-to-one with the node neurons in the posture control network, and the conductance values ​​of the memristor neurons directly correspond one-to-one with the trained weight values ​​of the corresponding node neurons.

3. The sensing, storage and computing integrated device for posture control according to claim 1, characterized in that: The posture control network is obtained independently from the hardware neural network training. The memristor neurons of the hardware neural network correspond one-to-one to the node neurons in the posture control network. The initial value of the conductance value of the memristor neuron directly corresponds one-to-one to the trained weight value of the corresponding node neuron; and enhanced training is performed based on the integrated sensing, storage and computing device system for posture control to fine-tune the conductance value of the memristor neuron.

4. The sensing, storage and computing integrated device for posture control according to claim 1, characterized in that: In the step of generating a simulation data set for training the posture control network based on the trained complex reasoning network model, random interference simulating the noise characteristics of the hardware neural network is added to the input data of the complex reasoning network model.

5. The sensing, storage and computing integrated device for posture control according to claim 1, characterized in that: At least the simulation data set generated by the complex reasoning network model is used as a training data set in the step of training the posture control network, and the simulation data set and the supervised training data set are mixed to construct a mixed data set for training the posture control network.

6. The sensing, storage and computing integrated device for posture control according to claim 1, characterized in that: The real system mathematical expression of the controlled object and the corresponding observer mathematical expression are: in, is the system state quantity; is the attitude quaternion; 3D velocity vector; is the three-dimensional displacement vector; is the three-dimensional angular velocity The corresponding offset; for The random walk driving force of is the three-dimensional acceleration The corresponding offset; for The random walk driving force of is the gravity acceleration bias; Represents the estimated state quantity corresponding to the system state quantity; Constructing angular velocity estimates and speed estimation as follows, in, is the angular velocity sensing noise, For acceleration sensing noise, is the original angular velocity output of the inertial navigation sensor; It is the original acceleration output of the inertial navigation sensor; Among them, the attitude transfer matrix and the corresponding rotation matrix Expressed as: in, , is the constraint between the components of the quaternion, , , They are , , Axis estimated angular velocity; Construct the state error of the inertial navigation sensor system: ， The state transfer matrix corresponding to the state error of the inertial navigation sensor system is: in, Represents a vector Construct the antisymmetric matrix, for The identity matrix, for All zero matrix; The covariance matrix of the inertial navigation sensor system is as follows: According to the state error of the inertial navigation sensor system and external correction Constructing Correction Error : in, is an estimated value; is the transfer matrix from state to observation error.

7. The sensing, storage and computing integrated device for posture control according to any one of claims 1 to 3, characterized in that: In the hardware neural network, the multiplication operation is performed at each row and column intersection in the memristor array according to Ohm's law; the intersection of the i-th row and the j-th column is connected through a memristor, and the conductance value of the memristor is expressed as , the input of the i-th row of the array is , the sum of row and column currents is enforced by Kirchhoff's law, and the sum of the currents read by the devices along the columns , to achieve neuronal excitation The forward propagation required ; The integral along the line realizes the error term The sum required for back propagation .

Images