Neural network-based method and system for controlling motion state of tracked vehicle, and medium

Abstract

A neural network-based method and system for controlling a motion state of a tracked vehicle, and a medium. The method comprises the following steps: constructing a neural network prediction model driven on the basis of multivariate data (S100); on the basis of the multivariate data, performing model optimization on the neural network prediction model (S200); and, on the basis of the optimized neural network prediction model, constructing a motion state prediction model, acquiring a reference trajectory of a tracked vehicle, on the basis of the reference trajectory and the motion state prediction model, calculating an optimal control parameter, on the basis of the optimal control parameter, performing motion control on the tracked vehicle, and returning the optimal control parameter to the neural network prediction model for model parameter updating (S300). Thus, the neural network model driven on the basis of data can be used as a prediction model, the characteristics of a traditional physical model, such as parameter uncertainty, poor adaptability, and difficulty of analysis are overcome, and, on the basis of an online learning mechanism, real-time prediction of the motion state of the tracked vehicle and real-time dynamic adjustment of parameters of the prediction model are performed.

Description

Neural Network-Based Motion State Control Method, System, and Medium for Tracked Vehicles Technical Field

[0001] This invention relates to the field of predictive control technology for tracked vehicles, and in particular to a method, system, and medium for controlling the motion state of tracked vehicles based on neural networks. Background Technology

[0002] Unmanned tracked vehicles, as a highly adaptable and multifunctional mobile platform, have demonstrated irreplaceable advantages in various high-risk and complex environments such as military reconnaissance, engineering operations, emergency rescue, and geological exploration. These vehicles can traverse complex terrains, including steep slopes, ravines, and muddy areas, providing powerful support for various missions. However, the stability and control precision of tracked vehicles in complex terrain have always been a key research focus and challenge.

[0003] Tracked vehicles are significantly affected by the irregularity and instantaneous changes of road surfaces when traversing complex terrain. Especially under extreme conditions, such as steep slopes and ditches, if the vehicle's driving state cannot be accurately predicted, the system will struggle to formulate targeted control strategies based on future motion. This leads to dynamic response lag when dealing with varying terrain, increasing handling instability. Particularly during high-speed travel or cornering, the inability to promptly control the tracked drive system increases the risk of skidding and rollover, seriously threatening mission safety and success rates.

[0004] Currently, some progress has been made in the research of state prediction control methods for tracked vehicles, but there are still many shortcomings, as follows:

[0005] (1) Chinese Patent CN116861784A discloses a motion control method for tracked vehicles using a combined Gaussian regression model predictive control. This method decomposes the dynamics of the tracked vehicle into modelable and unmodelable parts, and uses a Gaussian regression model to predict the unmodelable parts. However, this type of method is more suitable for relatively simple and regular terrain conditions. In complex and varied terrain, the neglect of important parameters such as ground pressure leads to insufficient dynamic consideration, which may result in prediction bias and reduce the control accuracy and stability of the vehicle.

[0006] (2) Chinese Patent CN117784610A discloses a control method for a dual-motor driven tracked vehicle based on a neural network prediction model. This method uses a multibody dynamics model to generate motion state data of the vehicle under various working conditions and combines it with an LSTM neural network model for prediction. However, this type of method mainly relies on simulation data rather than real environment data, and therefore has poor adaptability to dynamic scenarios. At the same time, the LSTM model may not be able to respond in time when dealing with sudden changes in terrain or rapid changes in vehicle state, resulting in a decrease in prediction accuracy.

[0007] (3) Chinese Patent CN117864156A discloses a method and system for predicting the state of a tracked vehicle under extreme conditions. This method constructs a comprehensive control model containing multiple dynamic models and uses a multiplexed rolling time-domain estimation method to optimize the model parameters for prediction. However, this type of method relies on an idealized physical model, while the strong coupling of dynamics in actual tracked vehicle operation and the complexity of the interaction between the tracks and the ground make the constructed dynamic model inaccurate. In addition, the factors considered by the model cannot fully cover all actual situations, and due to the lack of online data acquisition and real-time update capabilities, the model ultimately struggles to cope with changes in terrain during actual driving.

[0008] In summary, the existing technical solutions have the following shortcomings in predictive control of the motion state of tracked vehicles:

[0009] (1) The existing technical solution does not take into account the influence of ground pressure on vehicle motion in complex terrain environment, resulting in large prediction error and low control accuracy.

[0010] (2) Most existing technical solutions rely on static model parameters trained offline, which are difficult to cope with the dynamic changes in vehicle status and environmental conditions under complex working conditions.

[0011] (3) Existing technical solutions generally analyze the motion state of tracked vehicles based on physical models. However, in complex terrain, the nonlinearity and strong coupling of tracked vehicle dynamics make it difficult for traditional physical models to effectively describe the motion, resulting in poor prediction and control accuracy.

[0012] Therefore, in order to overcome the shortcomings of existing technologies, there is an urgent need to develop a data-driven predictive control method for the motion state of tracked vehicles that can adapt to complex terrain. Summary of the Invention

[0013] The purpose of this invention is to provide a method, system, and medium for controlling the motion state of tracked vehicles based on neural networks, thereby solving all or one of the aforementioned problems in the prior art.

[0014] To solve the above-mentioned technical problems, the specific technical solution of the present invention is as follows:

[0015] On one hand, the present invention provides a neural network-based method for controlling the motion state of a tracked vehicle, comprising the following steps:

[0016] Construct a neural network prediction model based on multivariate data;

[0017] The neural network prediction model is optimized based on multivariate data.

[0018] A motion state prediction model is constructed based on the optimized neural network prediction model to obtain the reference trajectory of the tracked vehicle. Optimal control parameters are calculated based on the reference trajectory and the motion state prediction model. Motion control of the tracked vehicle is performed based on the optimal control parameters, and the optimal control parameters are returned to the neural network prediction model for model parameter updates.

[0019] As an improved approach, the construction of a neural network prediction model based on multivariate data-driven methods includes:

[0020] A three-layer backpropagation neural network containing a single hidden layer is used as the neural network prediction model.

[0021] The suspension pressure value of the tracked vehicle is collected as a set of feedback state parameters;

[0022] The road surface type information of the tracked vehicle is collected as a road surface type parameter set;

[0023] The three-dimensional position, three-dimensional velocity, vehicle heading angle, vehicle roll angle, vehicle pitch angle, and vehicle yaw rate of the tracked vehicle are collected as a set of motion state parameters.

[0024] The driving torques on both sides of the tracked vehicle are collected as a torque parameter set;

[0025] The feedback state parameter set, the road surface type parameter set, the motion state parameter set, and the torque parameter set are used as the multivariate data for training the neural network prediction model.

[0026] As an improved approach, the optimization of the neural network prediction model based on multivariate data includes:

[0027] The multivariate data is normalized.

[0028] The weight parameters and bias parameters of the neural network prediction model are initialized;

[0029] The loss function is a function that measures the deviation between the predicted value and the actual data at each time step based on the mean square error. The normalized multivariate data is input into the initialized neural network prediction model, and the neural network prediction model is trained based on the loss function.

[0030] Calculate the gradient parameters of the neural network prediction model, and update the model parameters of the neural network prediction model based on the gradient parameters.

[0031] As an improved approach, the construction of a motion state prediction model based on the optimized neural network prediction model includes:

[0032] The state variable prediction equation and the model output equation are determined based on the prediction equation of the optimized neural network prediction model.

[0033] By combining the state variable prediction equations and the model output equations, the motion state prediction model is obtained.

[0034] As an improved approach, the calculation of optimal control parameters based on the reference trajectory and the motion state prediction model includes:

[0035] The optimization objectives are set as follows: the position tracking deviation in the three directions should be small, the deviations in the heading angle, roll angle and pitch angle should be small, and the rate of change of the input control signal should be small.

[0036] Construct a target optimization function based on the aforementioned optimization objective;

[0037] Set state constraints to control the driving stability and compatibility of tracked vehicles under complex working conditions.

[0038] Set control constraints related to control smoothness and the boundedness of drive torque assignment and rate of change;

[0039] The reference trajectory is input into the motion state prediction model, and the motion state prediction model is invoked to calculate the optimal control sequence based on the objective optimization function, the state quantity constraints, and the control quantity constraints.

[0040] Based on the optimal output torque increment sequence of the optimal control sequence, the optimal control quantity of the tracked vehicle at the first moment is determined; the optimal control quantity at the first moment is used as the optimal control parameter of the tracked vehicle at the first moment.

[0041] As an improved solution, obtaining the reference trajectory of the tracked vehicle includes:

[0042] The three-dimensional position, heading angle, roll angle, and pitch angle of the tracked vehicle are selected from the set of motion state parameters as the trajectory tracking target quantities of the reference trajectory.

[0043] As an improved approach, the construction of a neural network prediction model based on multivariate data-driven methods further includes:

[0044] The multivariate data is preprocessed, and the preprocessed multivariate data is used as the input to the neural network prediction model;

[0045] The number of nodes in the output layer of the neural network prediction model is determined based on the multivariate data.

[0046] Set tanh as the activation function between the input layer and the hidden layer of the neural network prediction model;

[0047] Set Purelin as the activation function between the hidden layer and the output layer of the neural network prediction model;

[0048] The number of hidden layer nodes in the neural network prediction model is determined using an empirical formula.

[0049] As an improved approach, calculating the gradient parameters of the neural network prediction model includes:

[0050] Calculate the output layer weight gradient, output layer threshold gradient, hidden layer weight gradient, and hidden layer threshold gradient;

[0051] The output layer weight gradient, the output layer threshold gradient, the hidden layer weight gradient, and the hidden layer threshold gradient are used as the gradient parameters.

[0052] On the other hand, the present invention also provides a neural network-based tracked vehicle motion state control system, comprising:

[0053] The initialization module is used to: build a neural network prediction model based on multivariate data;

[0054] The training module is used to optimize the neural network prediction model based on multivariate data.

[0055] The model deployment module is used to: construct a motion state prediction model based on the optimized neural network prediction model, obtain the reference trajectory of the tracked vehicle, calculate the optimal control parameters based on the reference trajectory and the motion state prediction model, perform motion control on the tracked vehicle based on the optimal control parameters, and return the optimal control parameters to the neural network prediction model for model parameter update.

[0056] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of the tracked vehicle motion state control method of the neural network.

[0057] The beneficial effects of the technical solution of this invention are:

[0058] 1. The neural network-based tracked vehicle motion state control method of the present invention can take ground pressure as a key parameter affecting the motion stability of tracked vehicles on different terrains, and use the pressure value measured by the suspension unit pressure sensor as the input feature of the neural network, so that the model can more comprehensively obtain the influence of complex terrain changes on the vehicle motion state.

[0059] 2. The neural network-based motion state control method for tracked vehicles described in this invention can use a data-driven neural network model as a prediction model, which solves the problems of uncertain parameters, poor adaptability, and difficult analysis of traditional physical models.

[0060] 3. The tracked vehicle motion state control method based on neural networks described in this invention can predict the motion state of tracked vehicles in real time based on BP neural networks and online learning mechanisms, and supports real-time dynamic adjustment of prediction model parameters, enabling the prediction model to adapt to complex terrain and road surface changes, and achieve stable trajectory tracking control of tracked vehicles.

[0061] 4. The tracked vehicle motion state control system based on the neural network described in this invention can achieve the tracked vehicle motion state control method based on the neural network described in this invention through the mutual cooperation of system modules.

[0062] 5. The computer-readable storage medium of the present invention can enable the guidance system module to cooperate in order to realize the neural network tracked vehicle motion state control method of the present invention, and the computer-readable storage medium of the present invention also effectively improves the operability of the neural network tracked vehicle motion state control method. Attached Figure Description

[0063] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0064] Figure 1 is a flowchart illustrating the neural network-based motion state control method for tracked vehicles according to Embodiment 1 of the present invention.

[0065] Figure 2 is a schematic diagram of the vehicle structure of the tracked vehicle in the neural network motion state control method for tracked vehicles described in Embodiment 1 of the present invention.

[0066] Figure 3 is a schematic diagram of the contact between the track of the tracked vehicle and the ground in the neural network-based tracked vehicle motion state control method described in Embodiment 1 of the present invention.

[0067] Figure 4 is a schematic diagram of the network structure of the BP neural network in the tracked vehicle motion state control method of the neural network in Embodiment 1 of the present invention;

[0068] Figure 5 is a schematic diagram of the logical architecture of the neural network-based tracked vehicle motion state control method described in Embodiment 1 of the present invention;

[0069] Figure 6 is a schematic diagram of the architecture of the tracked vehicle motion state control system based on the neural network described in Embodiment 2 of the present invention;

[0070] The markings in the attached diagram are as follows:

[0071] 1. Inertial navigation system; 2. GPS; 3. LiDAR; 4. Camera; 5. Suspension unit; 6. Pressure sensor. Detailed Implementation

[0072] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0073] In the description of this invention, it should be noted that the embodiments described in this invention are only some embodiments of this invention, not all embodiments; based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0074] The terms "first," "second," etc., used in this specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0075] Example 1: This example provides a neural network-based method for controlling the motion state of a tracked vehicle, as shown in Figures 1-5, including the following steps:

[0076] S100. Construct a neural network prediction model based on multivariate data-driven methods, including:

[0077] S101. Construct the feedback state parameter set for the tracked vehicle, as follows:

[0078] In this embodiment, a tracked vehicle with four road wheels on each side is selected as a specific example for illustration. Pressure sensors are installed in the suspension device of each road wheel. These sensors measure the load on the suspension unit, accurately capturing the load characteristics borne by each road wheel at each moment, thereby indirectly mapping the force between the road wheel and the ground. Since there is an approximately linear relationship between the compression of the suspension system and the force borne by the road wheel, the suspension pressure data collected by the sensors can intuitively reflect the ground pressure distribution of the tracked vehicle at each road wheel position. Finally, the eight suspension pressure values ​​are expressed as follows:

[0079] ;

[0080] It should be noted that the above pressure values ​​are integrated into a feedback state parameter set P, which is then input into a neural network model to predict the vehicle's dynamic response under various working conditions, thereby providing strong data support for optimizing vehicle control strategies. Given that the ground pressure of tracked vehicles in complex terrain is affected by a variety of factors, making it difficult to obtain accurate measured values, and considering the significant impact of the load distribution of each road wheel on the vehicle's dynamic characteristics during operation, the above data is used as a substitute. Although the suspension pressure and the road wheel ground pressure are not completely consistent in value, the conversion relationship between them is linear enough, and this relationship will be naturally captured and represented during the learning process of the neural network.

[0081] It should also be noted that the schematic diagram of the tracked vehicle structure in this application is shown in Figure 2. The lidar is generally installed on the upper front side of the tracked vehicle and is used to provide the terrain point cloud information ahead; the camera is installed in front of the tracked vehicle and is used to collect road surface information as input for the subsequent neural network; the positioning system is installed at the center of mass of the tracked vehicle and is used to provide the spatial coordinate information of the vehicle at the current moment, including the position in three directions, the turning angle in three directions, and the yaw rate; the pressure sensor is installed on the suspension arm of the road wheel and is used to detect the load distribution information of the road wheel of the tracked vehicle; in addition, the on-board sensors are also used to provide the motion state information of the tracked vehicle at each moment.

[0082] S102. Construct the road surface type parameter set for the tracked vehicle, as follows:

[0083] In this embodiment, a camera mounted on the tracked vehicle captures real-time images of the road surface in front of the vehicle. Simultaneously, data collected from the state detection unit and suspension unit is combined with multi-source information fusion technology to comprehensively analyze the collected information. To ensure temporal and spatial consistency between image data and dynamic data, the system uses timestamp synchronization technology for data alignment. The processed data is input into an optimized GCN (Graph Convolutional Neural Network). This network can deeply analyze sensor data to estimate key physical characteristics of the road surface, such as smoothness, bumpiness, and slipperiness. These characteristics provide the network with sufficient information to identify more accurate road surface types. The road surface type identification results, as the output of the deep learning network, provide valuable reference information for the intelligent control and navigation of the tracked vehicle. Finally, the road surface type parameter set is defined as follows:

[0084] ;

[0085] It should be noted that, for the aforementioned set N containing 10 road surface feature parameters, the discrete classification features are converted into binary vector form using one-hot encoding technology. In subsequent steps, these features are input into the BP neural network model along with other numerical features to further predict the motion state of the tracked vehicle. Given the significant differences in resistance under different road surface conditions, the above steps are performed to obtain accurate road surface type identification results in order to more accurately evaluate the kinematic performance of the tracked vehicle on complex terrain. Furthermore, the high-resolution camera equipped in this invention, with its excellent focusing ability and adaptability, can clearly capture road surface details at closer range.

[0086] S103. Construct the motion state parameter set for the tracked vehicle, as detailed below:

[0087] In this embodiment, the three-dimensional position (x, y, z) and three-dimensional velocity of the tracked vehicle are selected. Vehicle heading angle Roll angle Pitch angle and yaw rate These parameters, together, constitute the motion state parameter set X. These parameters are collected in real time by high-precision sensors, providing a solid data foundation for accurately predicting the motion state of tracked vehicles, and can be expressed as:

[0088] ;

[0089] In this embodiment, the driving torque on both sides of the tracked vehicle is used as the control variable for trajectory tracking control rolling optimization and denoted as:

[0090] ;

[0091] It should be noted that the motion state of tracked vehicles is complex and varied, covering a variety of working conditions such as acceleration and deceleration, steering, climbing, and obstacle crossing. Especially when facing extremely complex terrains such as trenches, steep slopes, side slopes, and vertical walls, it is necessary to select parameters that can comprehensively reflect the motion characteristics of tracked vehicles under various scenarios. Therefore, the current motion state of the vehicle is also used as an important reference value for prediction.

[0092] S104. Parameter set integration, as detailed below:

[0093] In this embodiment, since the motion state prediction of a tracked vehicle is a process that is influenced by multiple factors, P, N, and X, determined in the aforementioned steps, constitute the key elements of the prediction. In order to make accurate predictions of the motion state at the next moment, P, N, and X are combined with the tracked vehicle drive torque U obtained by time-by-time optimization in the predictive control model, and used together as the input information of the neural network.

[0094] S105. Determine the BP neural network prediction model, as follows:

[0095] In this embodiment, a backpropagation (BP) neural network is used to construct the prediction model. Specifically, a three-layer BP neural network containing a single hidden layer (where the input layer is considered layer 0, the hidden layer is layer 1, and the output layer is layer 2) can effectively approximate any nonlinear function when the number of hidden layer nodes is sufficiently large. Therefore, this embodiment selects the above-mentioned three-layer BP neural network, and its prediction model is expressed as follows:

[0096] = ;

[0097] It should be noted that, considering the high-frequency sampling rate of the sensors, the motion characteristics of the tracked vehicle, and the computing power of the model, a time step of 0.1 seconds is set as the prediction time step. This means that the BP neural network model will use the current state variable as input to predict the motion state 0.1 seconds later. This design aims to ensure that the prediction is both forward-looking and can make full use of real-time data, thereby improving the accuracy and practicality of the prediction.

[0098] S106. Neural network pre-configuration, as detailed below:

[0099] (1) The above multidimensional variable data is preprocessed, and the preprocessed data is used as the input of the BP neural network, as shown below: Since there are 21 input variables in the input layer, the input layer is determined to have 21 nodes.

[0100] (2) The tracked vehicle motion state vector output by the neural network prediction model at the next moment is represented as: Based on this, the number of nodes in the output layer is set to 10, as follows: Among them, the 10 nodes represent the horizontal positions at time t+1. Longitudinal position Vertical position , heading angle α, roll angle β, pitch angle Velocity in three directions and yaw rate ;

[0101] (3) In the connection from the input layer to the hidden layer, tanh is set as the activation function to perform a smooth nonlinear mapping of the input, specifically expressed as tanh( )= ; among them, tanh( The output after the tanh activation function is denoted as in subsequent steps. The input values ​​for the hidden layer;

[0102] (4) A linear activation function Purelin is set between the hidden layer and the output layer. Since the output state prediction value of the tracked vehicle often exhibits an approximately linear trend in a short period during its movement, the linear activation function Purelin can accurately capture and maintain this linear relationship, ensuring the accuracy of the output prediction result. Specifically, it is expressed as purelin( )= Among them, purelin ( The output of the purelin activation function is denoted as in subsequent steps. The input values ​​for the output layer;

[0103] (5) The number of hidden layer nodes in the BP neural network is determined using an empirical formula. This method fully considers the impact of network structure on prediction performance and aims to find a configuration of the number of hidden layer nodes that can effectively capture data features while avoiding overfitting or underfitting. The empirical formula is as follows: λ; where m is the number of input layer nodes and n is the number of output layer nodes. Adjusting constants based on experience, ∈[1,10];

[0104] (6) Given that the real-time driving environment of the tracked vehicle is a dynamic online learning scenario and lacks a comprehensive and static dataset, 8 hidden layer nodes were selected as the starting point in the initial model construction stage. After subsequent experiments, it was confirmed that this configuration of the number of nodes can ensure that the model reaches the optimal state in terms of prediction performance and generalization ability. The final training parameter table is as follows:

[0105] Layer output size parameters (number of layers): Input layer 210, Hidden layer 8176, Output layer 1090

[0106] S200. Optimize / train the neural network prediction model based on multivariate data, including:

[0107] S201. Data normalization processing, as detailed below:

[0108] In this embodiment, an experience-based maximum and minimum value setting strategy is used for input and output normalization. Based on the physical properties of the tracked vehicle's state parameters, its actual operating range, and accumulated experience data, an initial maximum and minimum value are set for each input feature. The formula is expressed as follows:

[0109] =2 ;

[0110] Where x is the original data, For the normalized data, To be the minimum value, It is the maximum value;

[0111] It should be noted that since this invention operates in an online learning mode, it cannot rely on a pre-set dataset to directly calculate the values ​​required for normalization. Therefore, an experience-based maximum and minimum value setting strategy is adopted. This strategy not only conforms to the actual operation of tracked vehicles, but also ensures the smooth progress of the normalization process, laying a solid foundation for subsequent model training and prediction.

[0112] S202, Parameter initialization, as follows:

[0113] ;

[0114] in, Let be the weights from the h-th neuron to the k-th neuron in the l-th layer of the neural network; , Let H and K be random variables whose values ​​are in the range [-1, 1]; H is the number of neurons in layer l-1; and K is the number of neurons in layer l. G is the bias of the k-th neuron in the l-th layer; G is the scaling factor for adjusting the weights and the range of the bias.

[0115] Among them, the weights from the input layer to the hidden layer after initialization and bias and the weights from the hidden layer to the output layer and bias The following are examples:

[0116] ;

[0117] ;

[0118] The hidden layer input is: The hidden layer output is: ;

[0119] The input to the output layer is: The output layer output is: .

[0120] S203. Set the loss function as follows:

[0121] In this embodiment, a function based on mean squared error to measure the deviation between the predicted value and the actual data at each time step is used as the loss function. The function formula is as follows: Where n is the number of output variables; This is the predicted value for the c-th output; This is the actual value of the c-th output.

[0122] S204, Gradient calculation, as follows:

[0123] (1) Calculate the output layer weight gradient: First, Since there is only the j-th output It is about The function, therefore, for other outputs, is: ; so the original form ;in, .

[0124] (2) Calculate the output layer threshold gradient: For the threshold (Threshold of the j-th output node), according to the above formula The following modifications have been made: .

[0125] (3) Calculate the gradient of the hidden layer weights:

[0126] ;

[0127] Since the above expression only has the j-th element... It is about Since it is a function, the above expression can be simplified to: ;

[0128] Based on the differentiation of the composite function, the original expression =

[0129] Formula ②;

[0130] Where, equation ② = ;Depend on have to: .

[0131] (4) Calculate the threshold gradient of the hidden layer:

[0132] Based on the logic of solving the hidden layer weight gradient, and according to the formula have:

[0133]

[0134] .

[0135] S205. Update the model parameters based on the above gradients, as follows:

[0136] ;

[0137] in, The learning rate is fixed at 0.01. This represents the old values ​​of the weights from neuron j to i in the weight matrix from the previous time step; This is the old value of the bias of neuron i; , These are the gradient values ​​of the aforementioned loss function with respect to each parameter;

[0138] It should be noted that whenever the weights and biases of the BP neural network are updated, the BP neural network immediately uses the latest parameters to predict the state at the next moment. In this process, each acquisition and processing of new data will activate the model prediction and self-update. As long as the tracked vehicle is in continuous motion, the system will continuously receive new state data, thereby driving the model parameters to be continuously adjusted. Finally, the BP neural network forms a closed-loop prediction process: "real-time data acquisition → state prediction → error calculation → parameter update". This process ensures that the model can keep up with the dynamic changes of the tracked vehicle and achieve accurate real-time prediction.

[0139] S300. Construct a motion state prediction model based on the optimized neural network prediction model, obtain the reference trajectory of the tracked vehicle, calculate optimal control parameters based on the reference trajectory and the motion state prediction model, perform motion control on the tracked vehicle based on the optimal control parameters, and return the optimal control parameters to the neural network prediction model for model parameter update, including:

[0140] S301. Receive the reference trajectory from the navigation system, as detailed below:

[0141] In this embodiment, considering the trajectory tracking control requirements of tracked vehicles traveling in complex terrain such as trenches and slopes, x, y, z, α, β, and γ are selected from the motion state variable set X as the target quantities for trajectory tracking, expressed as: .

[0142] S302. Construct the prediction model, as follows:

[0143] (1) Based on the aforementioned data-driven neural network model as the prediction model, considering the state prediction at time t with multiple time steps k, the prediction equation in the prediction control model is: ;

[0144] (2) Define state variables Control quantity The formula is then approximated and linearized.

[0145] (3) In order to achieve smooth processing of the control signal, it is necessary to process the change in the control input. After optimization, and taking into account the cumulative impact of all control input changes on the system state in the future prediction time domain, the final state variable prediction equation is derived as follows: Where A is the state transition matrix, B is the control input matrix, and H is the sum of external disturbances and constant terms. To control the time domain;

[0146] (4) The model output equation of the predictive control model is ;

[0147] (5) By simultaneously solving the state variable prediction equations and the output equations of the control model, the final prediction model is obtained, which is expressed as: ;

[0148] (6) Based on the recursive relationship between the system state and system output in the future time domain according to the above prediction model, the system prediction output model is obtained by comprehensive analysis as follows: ;in, The coefficient matrix of the state variables. Let be the coefficient matrix of the input; then, the output at time t... Changes in control quantity at time t Reference values ​​in each prediction time domain ;

[0149] S303. Set the multi-objective optimization function, as follows:

[0150] In this embodiment, to enable the tracked vehicle to stably track the reference trajectory, three optimization indicators and optimization directions are set, specifically:

[0151] (1) The position tracking deviation in the three directions should be small;

[0152] (2) The deviations in heading angle, roll angle and pitch angle should be small;

[0153] (3) The rate of change of the input control signal should be small;

[0154] In this embodiment, the objective optimization function is: ;in, For the prediction time domain of the system, To track the weighting coefficients, As a relaxation factor, , These are the weight matrices, To predict the output of the control model, The reference trajectory output by the controller. To control the increment.

[0155] S304. Set constraints as follows:

[0156] In this embodiment, considering that the tracked vehicle needs to operate in complex terrain and variable working conditions, while also meeting physical limitations and ensuring control stability, the following constraints are set for the state variables:

[0157] ;

[0158] In this embodiment, based on the smoothness of control and the bounded characteristics of the magnitude and rate of change of the driving torque, the constraints on the control quantity are set as follows:

[0159] .

[0160] S305. Calculate the optimal control quantity, as follows:

[0161] In this embodiment, by combining the above-mentioned multi-objective optimization function and various constraints, the optimal control value is calculated, and the optimal control sequence is obtained as follows: ;

[0162] In this embodiment, the optimal output torque increment sequence at time t is calculated, and the first term is added to the control quantity at time t-1 to obtain the optimal control quantity at time t: ;

[0163] In this embodiment, the optimal control quantity obtained at time t is transmitted to the tracked vehicle and used as the input of the neural network model at the current time to perform rolling prediction of the motion state in the next prediction time domain. At this point, the control flow of one motion control cycle of the tracked vehicle ends, and the tracked vehicle will move under the control action. After that, the method will return to step S301 to start the control flow of the next cycle. Finally, the cyclic control is performed based on the above flow until the tracked vehicle successfully completes the tracking task.

[0164] In summary, this application integrates an online learning mechanism. By inputting the multi-dimensional state information of the tracked vehicle at time k into a BP neural network, and based on the forward propagation characteristics of the neural network, it predicts the motion state variables of the tracked vehicle at time k+1. When the vehicle reaches time k+1, the error between the real-time state information fed back by the sensors and the previously predicted state information is calculated and compared using this method. The weights and biases of the neural network are dynamically adjusted using a gradient descent algorithm, thereby endowing the model with the ability to adapt to complex and varied terrain. Subsequently, the tracked vehicle is precisely controlled based on the motion state predicted by this method. A neural network is used to replace the traditional dynamic equations as the prediction model. In each prediction time domain, based on the predicted state of the tracked vehicle, the objective function is constructed by comprehensively considering the reference trajectory and the physical constraints of the system, and the optimal control sequence is solved to continuously optimize and adjust the driving torque of the tracked vehicle, thereby achieving stable trajectory tracking control of the tracked vehicle in complex terrain.

[0165] It should be noted that the above examples are merely for explaining the present invention and should not be construed as limiting the scope of protection of the present invention.

[0166] Example 2: This example, based on the same inventive concept as the neural network-based tracked vehicle motion state control method described in Example 1, provides a neural network-based tracked vehicle motion state control system, as shown in Figure 6, including:

[0167] The initialization module is used to: build a neural network prediction model based on multivariate data;

[0168] The training module is used to optimize the neural network prediction model based on multivariate data.

[0169] The model deployment module is used to: construct a motion state prediction model based on the optimized neural network prediction model, obtain the reference trajectory of the tracked vehicle, calculate the optimal control parameters based on the reference trajectory and the motion state prediction model, perform motion control on the tracked vehicle based on the optimal control parameters, and return the optimal control parameters to the neural network prediction model for model parameter update.

[0170] Example 3: This example provides a computer-readable storage medium, including:

[0171] The storage medium is used to store computer software instructions used to implement the tracked vehicle motion state control method of the neural network described in Embodiment 1. It includes a program for executing the above-described tracked vehicle motion state control method of the neural network. Specifically, the executable program can be built into the tracked vehicle motion state control system of the neural network described in Embodiment 2. In this way, the tracked vehicle motion state control system of the neural network can implement the tracked vehicle motion state control method of the neural network described in Embodiment 1 by executing the built-in executable program.

[0172] Furthermore, the computer-readable storage medium in this embodiment can be any combination of one or more readable storage media, wherein the readable storage medium includes an electrical, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof.

[0173] Unlike existing technologies, the method, system and medium for controlling the motion state of tracked vehicles using a neural network proposed in this application can use a data-driven neural network model as the prediction model, which solves the problems of uncertain parameters, poor adaptability and difficult analysis of traditional physical models. Based on an online learning mechanism, the motion state of tracked vehicles can be predicted in real time and the parameters of the prediction model can be dynamically adjusted in real time. This enables the prediction model to adapt to complex terrain and road surface changes, and achieve stable trajectory tracking control of tracked vehicles.

[0174] It should be understood that in the various embodiments of this document, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this document.

[0175] It should also be understood that, in the embodiments herein, the term "and / or" is merely a description of the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following associated objects have an "or" relationship.

[0176] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this document.

[0177] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0178] In the embodiments provided herein, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the couplings or direct couplings or communication connections shown or discussed may be indirect couplings or communication connections through some interfaces, devices, or units, or they may be electrical, mechanical, or other forms of connection.

[0179] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments described herein, depending on actual needs.

[0180] Furthermore, the functional units in the various embodiments of this document can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0181] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this paper, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this paper. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0182] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for controlling the motion state of a tracked vehicle based on a neural network, characterized in that, Includes the following steps: Construct a neural network prediction model based on multivariate data; The neural network prediction model is optimized based on multivariate data. A motion state prediction model is constructed based on the optimized neural network prediction model to obtain the reference trajectory of the tracked vehicle. Optimal control parameters are calculated based on the reference trajectory and the motion state prediction model. Motion control of the tracked vehicle is performed based on the optimal control parameters, and the optimal control parameters are returned to the neural network prediction model for model parameter updates.

2. The method for controlling the motion state of tracked vehicles based on neural networks according to claim 1, characterized in that: The construction of a neural network prediction model based on multivariate data-driven methods includes: A three-layer backpropagation neural network containing a single hidden layer is used as the neural network prediction model. The suspension pressure value of the tracked vehicle is collected as a set of feedback state parameters; The road surface type information of the tracked vehicle is collected as a road surface type parameter set; The three-dimensional position, three-dimensional velocity, vehicle heading angle, vehicle roll angle, vehicle pitch angle, and vehicle yaw rate of the tracked vehicle are collected as a set of motion state parameters. The driving torques on both sides of the tracked vehicle are collected as a torque parameter set; The feedback state parameter set, the road surface type parameter set, the motion state parameter set, and the torque parameter set are used as the multivariate data for training the neural network prediction model.

3. The method for controlling the motion state of tracked vehicles based on neural networks according to claim 1, characterized in that: The optimization of the neural network prediction model based on multivariate data includes: The multivariate data is normalized. The weight parameters and bias parameters of the neural network prediction model are initialized; The loss function is a function that measures the deviation between the predicted value and the actual data at each time step based on the mean square error. The normalized multivariate data is input into the initialized neural network prediction model, and the neural network prediction model is trained based on the loss function. Calculate the gradient parameters of the neural network prediction model, and update the model parameters of the neural network prediction model based on the gradient parameters.

4. The method for controlling the motion state of tracked vehicles based on neural networks according to claim 1, characterized in that: The construction of the motion state prediction model based on the optimized neural network prediction model includes: The state variable prediction equation and the model output equation are determined based on the prediction equation of the optimized neural network prediction model. By combining the state variable prediction equations and the model output equations, the motion state prediction model is obtained.

5. The method for controlling the motion state of tracked vehicles based on neural networks according to claim 1, characterized in that: The calculation of optimal control parameters based on the reference trajectory and the motion state prediction model includes: The optimization objectives are set as follows: the position tracking deviation in the three directions should be small, the deviations in the heading angle, roll angle and pitch angle should be small, and the rate of change of the input control signal should be small. Construct a target optimization function based on the aforementioned optimization objective; Set state constraints to control the driving stability and compatibility of tracked vehicles under complex working conditions. Set control constraints related to control smoothness and the boundedness of drive torque assignment and rate of change; The reference trajectory is input into the motion state prediction model, and the motion state prediction model is invoked to calculate the optimal control sequence based on the objective optimization function, the state quantity constraints, and the control quantity constraints. Based on the optimal output torque increment sequence of the optimal control sequence, the optimal control quantity of the tracked vehicle at the first moment is determined; the optimal control quantity at the first moment is used as the optimal control parameter of the tracked vehicle at the first moment.

6. The method for controlling the motion state of tracked vehicles based on neural networks according to claim 2, characterized in that: The acquisition of the reference trajectory of the tracked vehicle includes: The three-dimensional position, heading angle, roll angle, and pitch angle of the tracked vehicle are selected from the set of motion state parameters as the trajectory tracking target quantities of the reference trajectory.

7. The method for controlling the motion state of a tracked vehicle based on a neural network according to claim 2, characterized in that: The construction of the neural network prediction model based on multivariate data further includes: The multivariate data is preprocessed, and the preprocessed multivariate data is used as the input to the neural network prediction model; The number of nodes in the output layer of the neural network prediction model is determined based on the multivariate data. Set tanh as the activation function between the input layer and the hidden layer of the neural network prediction model; Set Purelin as the activation function between the hidden layer and the output layer of the neural network prediction model; The number of hidden layer nodes in the neural network prediction model is determined using an empirical formula.

8. The method for controlling the motion state of tracked vehicles based on neural networks according to claim 3, characterized in that: The calculation of the gradient parameters of the neural network prediction model includes: Calculate the output layer weight gradient, output layer threshold gradient, hidden layer weight gradient, and hidden layer threshold gradient; The output layer weight gradient, the output layer threshold gradient, the hidden layer weight gradient, and the hidden layer threshold gradient are used as the gradient parameters.

9. A tracked vehicle motion state control system based on neural networks, characterized in that, include: The initialization module is used to: build a neural network prediction model based on multivariate data; The training module is used to optimize the neural network prediction model based on multivariate data. The model deployment module is used to: construct a motion state prediction model based on the optimized neural network prediction model, obtain the reference trajectory of the tracked vehicle, calculate the optimal control parameters based on the reference trajectory and the motion state prediction model, perform motion control on the tracked vehicle based on the optimal control parameters, and return the optimal control parameters to the neural network prediction model for model parameter update.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the neural network-based motion state control method according to any one of claims 1 to 8.

Images