Control method for bilateral motor-driven tracked vehicle based on neural network prediction model

Abstract

This invention provides a control method for dual-motor driven tracked vehicles based on a neural network prediction model. By introducing a Long Short-Term Memory (LSTM) neural network framework, it trains a neural network prediction model for the tracked unmanned platform, fully leveraging the LSTM's ability to process sequential data. This allows for the extraction of vehicle dynamics features from past driving data to predict the state at the next moment. This method can comprehensively control lateral displacement deviation, heading angle deviation, longitudinal velocity, and yaw rate, as well as track the desired trajectory position, while also ensuring steering controllability. Furthermore, this invention provides a trajectory tracking control method based on Neural Network Model Predictive Control (NN-MPC), simultaneously considering the vehicle's kinematics and dynamics, effectively overcoming the problem of poor trajectory tracking performance caused by the strong coupling of lateral and longitudinal dynamics in the physical model MPC method.

Description

Technical Field

[0001] This invention belongs to the field of tracked vehicle control technology, specifically relating to a comprehensive control method for a tracked vehicle driven by dual motors based on an LSTM neural network prediction model, used to achieve comprehensive control of handling stability and trajectory tracking under high-speed conditions. Background Technology

[0002] Trajectory tracking control, as one of the core technologies of tracked unmanned vehicles (UAVs), ensures that the UAVs stably and accurately follow the desired trajectory determined by the planning layer. In some recent existing technologies involving the control of tracked UAVs, Model Predictive Control (MPC) is employed to achieve trajectory tracking control. This approach considers nonlinear dynamic models and predicts the system's output over a future period. By solving constrained multi-objective optimization problems in a rolling manner, the system satisfies both multi-objective tracking accuracy and constraint conditions. The predictive model, as the core component of MPC, plays a decisive role in the accuracy and stability of trajectory tracking. However, these existing technologies, especially in terms of predictive models, still have many problems. For example, many solutions only control the speed of the tracks on both sides based on position tracking using kinematic models, neglecting dynamic constraints and control, and failing to consider dynamic control for handling stability, thus failing to guarantee the controllability of the steering trajectory during tracking. Chinese patent CN111812974B proposes a comprehensive control method based on a dynamic prediction model to directly control the drive torque of the dual motors. While this improves handling stability and trajectory tracking, its overly complex model, with strong nonlinearity and coupling, results in low computational efficiency and poor stability. Furthermore, it cannot guarantee applicability to specific conditions such as high-speed steering and large-angle steering. Therefore, providing a comprehensive control method for tracked vehicles that is more applicable to various working conditions and has higher accuracy and robustness is an urgent technical problem to be solved in this field. Summary of the Invention

[0003] In view of this, and to address the technical problems existing in this field, the present invention provides a control method for a dual-motor driven tracked vehicle based on a neural network prediction model, specifically including the following steps:

[0004] Step 1: Establish an LSTM neural network model and set the LSTM neural network structure and parameters; the predictive control model is used to determine the vehicle's lateral position X, longitudinal position Y, and heading angle at each time step, including global coordinates. Yaw rate ω and longitudinal speed v x The motion status information, and the input drive torque T of the motors on both sides of the track. R T L It can predict the vehicle's motion state at the next moment.

[0005] Step 2: Based on the multibody dynamics model of the RecurDyn unmanned tracked platform, simulated driving torque is provided to the vehicle to obtain the pose, speed and yaw rate simulation data of the tracked vehicle under multiple working conditions. The original dataset composed of these simulation data is preprocessed accordingly to form a training dataset for training the LSTM neural network.

[0006] Step 3: Train the LSTM neural network model using the training dataset and verify the training effect; the trained LSTM neural network model is used to provide state input for the predictive control model;

[0007] Step 4: Select For state variables, u(t) = [T] R (t),T L (t)] T To control the amount, The output variable is defined as t, where t represents time and the subscript e represents the expected value of the corresponding parameter. Mechanical limitations, control smoothness, and handling stability are considered, and corresponding constraints are set. A multi-objective optimization function is established based on the predicted output of the neural network and the reference value in the predicted time domain. Thus, a trajectory tracking predictive control model based on neural network model predictive control (NN-MPC) is constructed.

[0008] Step 5: Combine the state variables corresponding to the real-time motion state of the tracked vehicle to predict the future motion state variables of the tracked vehicle. Then, use the quadratic programming algorithm in combination with the set constraints to solve the multi-objective optimization problem and obtain the optimal control quantity of the driving torque of the two tracked sides through real-time rolling optimization.

[0009] Furthermore, the network structure of the LSTM neural network model established in step one specifically consists of an input layer, an LSTM layer, a fully connected layer (Dense layer), and a Dropout layer;

[0010] The input layer consists of a sequence of time_steps * input_dim, where time_steps represents the length of the historical data sequence and input_dim represents the number of selected data features. In the trajectory prediction of tracked vehicles, the former indicates how much past information is expected to be used to predict the vehicle state at the next moment, while the latter represents the dimension of the described vehicle state, i.e., x. LSTM =[X global ,U], where U = [T] R ,T LTherefore, input_dim = 7, and time_steps = 20, indicating that the vehicle state X at the next moment is predicted using historical data of the past 20 * 0.1 = 2 seconds. global (k+1);

[0011] The number of hidden units in the LSTM layer is set to num_units_LSTM = 32, and the network parameters for the forget gate include:

[0012] Weight parameters:

[0013] num_units_LSTM*(num_units_LSTM+input_dim)=32*(32+7)=2348

[0014] Bias parameters:

[0015] num_units_LSTM*1 = 32

[0016] That is, the forget gate has a total of num = 2347 + 7 = 2354 parameters;

[0017] The input gate contains 2*num network parameters; the output gate contains num network parameters.

[0018] The total number of parameters that need to be trained in the LSTM layer is:

[0019] params = 4 * num = 9416

[0020] The fully connected layer is specifically used to integrate and reduce the dimensionality of the LSTM output. It uses the ReLU function as the activation function and sets the number of hidden units in the fully connected layer to 16.

[0021] The Dropout layer is specifically used to randomly disable half of the units in the fully connected layer to avoid overfitting during training, and its number of nodes is also set to 16.

[0022] The output layer is used to transform the network parameters into vehicle motion state information to be predicted.

[0023]

[0024] Furthermore, the preprocessing of the original dataset in step two includes sequential data cleaning, data normalization, serialization processing with fixed time lengths, and data partitioning.

[0025] Furthermore, in the trajectory tracking predictive control model based on neural network model predictive control (NN-MPC) described in step four, the LSTM network is specifically equivalent to a single-layer LSTM neural network that receives time series data and a fully connected neural network that receives one-dimensional data. The single-layer LSTM neural network is independent of the model predictive control and does not participate in the state iteration in the prediction time domain.

[0026] The prediction of state variables is specifically expressed in the following form:

[0027]

[0028] Among them, F h and F C This represents the process of solving for predicted values ​​and cell states within a single-layer LSTM neural network.

[0029] The specific form of the prediction equation in the predictive control model is as follows:

[0030]

[0031] To improve solution efficiency, the formula is approximated as linearized, resulting in:

[0032]

[0033] in,

[0034]

[0035]

[0036]

[0037]

[0038]

[0039] Furthermore, define the state variable ξ = [h, C, X] global ],u=[T R ,T L As the control variable, the above system of equations can be written as:

[0040]

[0041] in,

[0042]

[0043] The specific form of the output equation of the predictive control model is as follows:

[0044]

[0045] in,

[0046] Combining the above equations, we get:

[0047]

[0048] Furthermore, the output of the trajectory tracking predictive control model at time t in step five is expressed as:

[0049]

[0050] in,

[0051]

[0052]

[0053] Y(t)=[η(t+1|t) T ,…,η(t+N P |t) T ] T ,

[0054] ΔU(t)=[Δu(t|t) T ,…,Δu(t+N C -1|t) T ] T

[0055] Where N p For prediction in the time domain, N c To control the time domain; after discretizing the continuous state space, we have, A kt =e AT , C kt =C.

[0056] Furthermore, to improve computational efficiency, the above parameters are simplified:

[0057]

[0058] The reference value for each prediction time domain is Y. ref (t)=[η ref (t+1|t) T ,…,η ref (t+N P |t) T ] T ,

[0059] Define the following multi-objective optimization function:

[0060]

[0061] Where σ is the weighting coefficient and ε is the relaxation factor σε 2 The relaxation term can speed up the convergence process and prevent situations where the objective function cannot be solved due to the absence of an optimal solution in the quadratic programming (in which case the suboptimal feasible solution will be output).

[0062] The constraints on the state variables include:

[0063]

[0064] The constraints on the output include:

[0065]

[0066]

[0067] Control constraints include:

[0068]

[0069] in, i∈(R,L),T m P is the maximum output torque of the motor. m is the maximum output power of the motor, and n is the motor speed;

[0070] Combining the above multi-objective optimization function and constraints, we obtain the following simplified standard form of quadratic programming:

[0071]

[0072] The optimal output torque increment sequence at time t is obtained by solving the problem. The first term is taken and added to the control quantity at time t-1 to obtain the optimal control quantity at time t.

[0073] The method provided by the present invention has at least the following advantages over the prior art:

[0074] (1) By introducing the Long Short-Term Memory Neural Network (LSTM) framework, the neural network prediction model of the tracked unmanned platform was trained, giving full play to the LSTM's ability to process sequential data, and realizing the extraction of vehicle dynamic features from past driving data to predict the state of the next moment.

[0075] (2) It achieves comprehensive control of lateral displacement deviation, heading angle deviation, longitudinal velocity and yaw rate, as well as tracking of the desired trajectory position point, while also ensuring the controllability of steering.

[0076] (3) A trajectory tracking control method based on neural network model predictive control (NN-MPC) is provided, which simultaneously considers the vehicle's kinematics and dynamics characteristics, effectively overcoming the problem of poor trajectory tracking performance caused by strong coupling of lateral and longitudinal dynamics in the physical model MPC method. Attached Figure Description

[0077] Figure 1 The diagram shows the LSTM neural network structure provided in this invention.

[0078] Figure 2 This refers to the acquisition of raw training data in the method provided by this invention.

[0079] Figure 3 This is a schematic diagram illustrating the serialization process of the original dataset.

[0080] Figure 4 Images showing the iterative cycle variation to verify the accuracy of the neural network;

[0081] Figure 5 This is a diagram of the trajectory tracking predictive control strategy based on NN-NMPC proposed in this invention;

[0082] Figure 6 A comparison of the Dyn-NMPC and NN-NMPC trajectory tracking control strategies of the model predictive controller designed for the method of this invention at medium and high speeds. Detailed Implementation

[0083] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0084] The control method for dual-motor driven tracked vehicles based on a neural network prediction model provided by this invention specifically includes the following steps:

[0085] Step 1: Establish an LSTM neural network model and set the LSTM neural network structure and parameters, such as... Figure 1 As shown; the predictive control model is used to determine the vehicle's lateral position X, longitudinal position Y, and heading angle at each moment, including the global coordinates. Yaw rate ω and longitudinal speed v x The motion status information, and the input drive torque T of the motors on both sides of the track. R T L It can predict the vehicle's motion state at the next moment.

[0086] Step 2: Based on the RecurDyn unmanned tracked platform multibody dynamics model, simulated driving torque is provided to the vehicle to obtain simulation data of the tracked vehicle's pose, speed, and yaw rate under various working conditions. The data acquisition is as follows: Figure 2 As shown; perform corresponding preprocessing on the original dataset composed of these simulation data to form a training dataset for training the LSTM neural network;

[0087] Step 3: Train the LSTM neural network model using the training dataset and verify the training effect; the trained LSTM neural network model is used to provide state input for the predictive control model;

[0088] Step 4: Select For state variables, u(t) = [T] R (t),T L (t)] T To control the amount, The output variable is defined as t, where t represents time and the subscript e represents the expected value of the corresponding parameter. Mechanical limitations, control smoothness, and handling stability are considered, and corresponding constraints are set. A multi-objective optimization function is established based on the predicted output of the neural network and the reference value in the predicted time domain. Thus, a trajectory tracking predictive control model based on neural network model predictive control (NN-MPC) is constructed.

[0089] Step 5: Combine the state variables corresponding to the real-time motion state of the tracked vehicle to predict the future motion state variables of the tracked vehicle. Then, use the quadratic programming algorithm in combination with the set constraints to solve the multi-objective optimization problem and obtain the optimal control quantity of the driving torque of the two tracked sides through real-time rolling optimization.

[0090] In a preferred embodiment of the present invention, the network structure of the LSTM neural network model established in step one specifically consists of an input layer, an LSTM layer, a fully connected layer (Dense layer), and a Dropout layer; the main elements of the LSTM neural network unit structure include:

[0091] (1) Activation functions: Commonly used activation functions include Sigmoid, Tanh, ReLU, Softmax, etc.

[0092] (2) Cell state: Cell state is the core of LSTM, capable of retaining important information from historical sequence data and flowing throughout the network. The update of cell state can be represented as:

[0093]

[0094] Updated cell state C t Compared with the network output h at the previous time step t-1 and the current input information xt Together they determine the network's output h t , is represented as:

[0095]

[0096] The updating and output of the cell state described above are achieved through the forget gate, update gate, and output gate.

[0097] (1) Forgotten Gate

[0098] LSTM networks filter information by using a forget gate, selectively discarding or retaining data. The input to the forget gate is the network output h from the previous time step. t-1 and the current input information x t After Sigmoid nonlinearization, the output f t Used to screen the state of primitive cells, as shown in the figure, f t The value ranges from 0 to 1. Values ​​closer to 0 indicate a tendency to forget, while values ​​closer to 1 indicate that the data is more important and should be retained.

[0099] (2) Input Gate

[0100] The input gate is the core of updating the cell state, and it also updates the network output h from the previous time step. t-1 and the current input information x t Input to the Sigmoid function:

[0101] i t =σ(W i ·[x t ,h t-1 ]+b t )

[0102] Meanwhile, under Tanh's influence, the historical network output and the current input are compressed into a value of [-1, 1]. Used to characterize data sequence features:

[0103]

[0104] (3) Output gate

[0105] The output of the output gate is the output h of the LSTM network. t , by the updated cell state C t After o t Filter and retrieve:

[0106]

[0107] in,

[0108] o t =σ(W o·[x t ,h t-1 ]+b o )

[0109] The input to an LSTM is a sequence containing data information over a period of time. Each set of data in the sequence is fed into the network with the same network parameters until the sequence ends and the prediction output is completed.

[0110] In each layer of a neural network, the input layer is a sequence of time_steps * input_dim, where time_steps represents the length of the historical data sequence, and input_dim represents the number of selected data features. In the trajectory prediction of tracked vehicles, the former indicates how much past information is expected to be used to predict the vehicle state at the next moment, while the latter represents the dimension of the described vehicle state, i.e., x. LS TM = [X globa l,U], where U = [T] R ,T L Therefore, input_dim = 7, and time_steps = 20, indicating that the vehicle state X at the next moment is predicted using historical data of the past 20 * 0.1 = 2 seconds. global (k+1);

[0111] The LSTM layer is the core layer, and the number of hidden units it contains directly affects the model's accuracy and training convergence speed. Too few hidden units lead to insufficient feature extraction and an inaccurate model, while too many not only severely impact training speed but also increase the risk of overfitting. This invention sets the number of hidden units to num_units_LSTM = 32, and the network parameters for the forget gate include:

[0112] Weight parameters:

[0113] num_units_LSTM*(num_units_LSTM+input_dim)=32*(32+7)=2348

[0114] Bias parameters:

[0115] num_units_LSTM*1 = 32

[0116] That is, the forget gate has a total of num = 2347 + 7 = 2354 parameters;

[0117] The input gate contains 2*num network parameters; the output gate contains num network parameters.

[0118] The total number of parameters that need to be trained in the LSTM layer is:

[0119] params = 4 * num = 9416

[0120] The fully connected layer is specifically used to integrate and reduce the dimensionality of the LSTM output. It uses the ReLU function as the activation function and sets the number of hidden units in the fully connected layer to 16.

[0121] The Dropout layer is specifically used to randomly disable half of the units in the fully connected layer to avoid overfitting during training, and its number of nodes is also set to 16.

[0122] The output layer is used to transform the network parameters into vehicle motion state information to be predicted.

[0123]

[0124] The network parameters that need to be obtained during training in this embodiment are shown in Table 1:

[0125] Table 1 Training Parameters

[0126]

[0127] Step two: First, set the initial driving torque in RecurDyn. Then, at each subsequent moment, apply a random torque increment to both sides, and record the driving torque, pose information, vehicle speed, and yaw rate received by the tracked unmanned platform at each moment. Collect approximately 1000 seconds of simulation data, forming approximately 50,000 information elements to create the original dataset, such as... Figure 2 As shown.

[0128] The preprocessing of the original dataset includes the following steps in sequence: data cleaning, data normalization, serialization processing according to fixed time length, and data partitioning:

[0129] (1) Data Cleaning: Data anomalies sometimes occur during the acquisition and recording process (especially in real vehicle data acquisition; this invention, based on multibody dynamics software with built-in sensors, has fewer such anomalies). Invalid data, outliers, and missing values ​​need to be deleted or processed from the dataset. The nearest neighbor imputation method is used to fill in missing data, and the 3σ principle is used to process outliers. Outliers that exceed the proximity metric should be deleted. At the same time, since vehicle information is continuous, filtering methods can be used to eliminate noise and make the data smoother.

[0130] (2) Feature Normalization: Although the network structure already scales the data using the sigmoid or tanh function, to improve efficiency and ensure the extraction of features with smaller values, it is customary to map the value ranges of different features to the same scale to avoid certain features having a significant impact on model training. The specific method used is Min-Max normalization.

[0131]

[0132] By processing the data using the above formula, the original data is mapped to the interval [-1, 1], thus achieving standardization.

[0133] (3) Serialization processing: For the LSTM-based model of this invention, the time series data is divided into multiple samples according to a fixed length (time_steps = 20), and a corresponding target value is provided for each sample, such as... Figure 3 As shown.

[0134] (4) Data partitioning: The dataset is divided into a training set, a validation set, and a test set for model training, optimization, and evaluation. This invention partitions the dataset from the original dataset in a 7:3 ratio. Research also shows that further partitioning the training set into smaller batch sizes is beneficial for model sequencing and avoiding overfitting. Therefore, this invention sets the batch size to 2000 for multiple training iterations.

[0135] Step three involves training an LSTM neural network prediction model and verifying its accuracy using mean squared error (MSE). For example, the most commonly used MSE can be selected as the loss function, and the relationship between the loss value and the iteration cycle can be generated using matplotlib in the TensorFlow framework. Figure 4 As can be seen, the loss value meets the requirements, and the value is eventually less than 0.1 and infinitely close to 0, which satisfies the accuracy requirements of the LSTM neural network vehicle model.

[0136] In step four, the prediction equation for the state, the input, output, and dimensional changes of the neural network can be expressed as follows:

[0137]

[0138] The final output of the NN network As can be seen, only the input to the LSTM layer is a multi-dimensional matrix; subsequent inputs and outputs are all one-dimensional vectors. As mentioned earlier, the output of the LSTM layer is obtained by sequentially feeding the sequence data into 20 LSTM networks sharing the same network parameters (time_steps). Each network receives a past predicted output and the corresponding sequence state information of the original cell state, and outputs a new predicted output and an updated cell state. This corresponds to the LSTM unit state, and its mathematical expression is:

[0139] [h t C t ] T =f LSTM (h t-1 C t-1 ,[Xglobal (t), U(t)] T )

[0140] This means that the predicted output h of the LSTM at time t t It can be directly from h t-1 C t-1 ,X global (t),U(t) can be obtained, and h can also be output. t and C t Used for the next iteration. Based on this idea, the LSTM layer with time_steps = 20 in the formula is decomposed into an LSTM' layer with time_steps = 19 and a Dense' layer. Their network parameters are the same, and the changes in the input, output, and dimensions of the neural network can be further expressed as:

[0141]

[0142] Therefore, in the trajectory tracking predictive control model based on neural network model predictive control (NN-MPC), the LSTM network can be equivalent to a single-layer LSTM neural network that receives time series data and a fully connected neural network that receives one-dimensional data. The single-layer LSTM neural network is independent of the model predictive control and does not participate in the state iteration in the prediction time domain.

[0143] The prediction of state variables is specifically expressed in the following form:

[0144]

[0145] Among them, F h and F C This represents the process of solving for predicted values ​​and cell states within a single-layer LSTM neural network.

[0146] The specific form of the prediction equation in the predictive control model is as follows:

[0147]

[0148] To improve solution efficiency, the formula is approximated as linearized, resulting in:

[0149]

[0150] in,

[0151]

[0152]

[0153]

[0154]

[0155]

[0156] Furthermore, define the state variable ξ = [h, C, X] global ],u=[T R ,T L As the control variable, the above system of equations can be written as:

[0157]

[0158] in,

[0159]

[0160] The specific form of the output equation of the predictive control model is as follows:

[0161]

[0162] in,

[0163] Combining the above equations, we get:

[0164]

[0165] This completes the construction of the NN-NMPC prediction model, and the complete control strategy framework is as follows: Figure 5 As shown.

[0166] Establish state equations The system predictive output expression is obtained by expanding, discretizing, and deriving at the reference point; the objective function is set according to the values ​​of the predicted output and the expected output, and the objective function is solved using a quadratic programming algorithm to finally output the relevant control quantity.

[0167] In a preferred embodiment of the present invention, the output of the trajectory tracking prediction control model at time t in step five is expressed as:

[0168]

[0169] in,

[0170]

[0171]

[0172] Y(t)=[η(t+1|t) T ,…,η(t+N P |t) T ] T ,

[0173] ΔU(t)=[Δu(t|t) T ,…,Δu(t+N C -1|t) T ] T

[0174] Where N p For prediction in the time domain, N c To control the time domain; after discretizing the continuous state space, we have, A kt =e AT , C kt =C.

[0175] Furthermore, to improve computational efficiency, the above parameters are simplified:

[0176]

[0177] The reference value for each prediction time domain is Y. ref (t)=[η ref (t+1|t) T ,…,η ref (t+N P |t) T ] T ,

[0178] Define the following multi-objective optimization function:

[0179]

[0180] Where σ is the weighting coefficient and ε is the relaxation factor σε 2 As a relaxation term, it can speed up the convergence of the solution and prevent the situation where the objective function cannot be solved due to the lack of an optimal solution in the quadratic programming (in this case, the suboptimal feasible solution will be output);

[0181] The constraints on the state variables include:

[0182]

[0183] The constraints on the output include:

[0184]

[0185]

[0186] Control constraints include:

[0187]

[0188] in, i∈(R,L),T mP is the maximum output torque of the motor. m is the maximum output power of the motor, and n is the motor speed.

[0189] Combining the above multi-objective optimization function and constraints, we obtain the following simplified standard form of quadratic programming:

[0190]

[0191] The optimal output torque increment sequence at time t is obtained by solving the problem. The first term is taken and added to the control quantity at time t-1 to obtain the optimal control quantity at time t. This quantity is then output to the controlled object to achieve control.

[0192] Figure 6 A specific example based on the method provided in this invention is shown. A Matlab / Simulink-Recurdyn simulation experiment was conducted, using a trajectory tracking strategy based on a dynamic prediction model as a control. Under high adhesion conditions (μ = 0.85), the double-track trajectory was tracked at a desired speed of 72 km / h. The weight matrix was set empirically. The simulation results are as follows. Figure 6 As shown in (a) to (e).

[0193] Through simulation Figure 6 Figures (a) to (c) depict the motion trajectories of a tracked unmanned platform under a double lane change maneuver at a speed of 72 km / h, employing two different control strategies: one based on a dynamic prediction model and the other on a neural network prediction model. The figures show that during high-speed turning, the trajectory tracking control strategy using the original dynamic prediction model deviates significantly from the desired trajectory, resulting in a large lateral deviation, although the lateral deviation remains within a reasonable range. In contrast, the trajectory tracking control strategy based on the neural network prediction model ensures that the vehicle can effectively track the double lane change maneuver without instability at the desired speed of 72 km / h, with the lateral and lateral deviations constrained to within 0.56 m and 0.52 rad, respectively.

[0194] Figure 6 (d) and (e) respectively describe the yaw rate and vehicle speed of the tracked unmanned platform under two different control strategies: one based on a dynamic prediction model and the other on a neural network prediction model. It can be seen that the yaw rate and speed changes are relatively stable under both strategies, but the yaw rate is generally smaller under the control strategy using the neural network prediction model. It is worth noting that due to the high desired vehicle speed, but limited by the unmanned platform's motor power and ground adhesion conditions, the acceleration distance is insufficient, causing the vehicle to begin turning before reaching its maximum speed. This leads to… Figure 6 (e) The overall vehicle speed was slightly lower than the expected speed.

[0195] It should be understood that the sequence number of each step in the embodiments of the present invention 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 the present invention.

[0196] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A control method for dual-motor driven tracked vehicles based on a neural network prediction model, characterized in that: Specifically, the following steps are included: Step 1: Establish an LSTM neural network model and set the LSTM neural network structure and parameters; the predictive control model is used to determine the vehicle's lateral position at each time step, including in global coordinates. Longitudinal position Heading angle yaw rate and longitudinal speed The motion status information, and the input drive torque of the motors on both sides of the track. , It can predict the vehicle's motion state at the next moment; Step 2: Based on the multibody dynamics model of the RecurDyn unmanned tracked platform, simulated driving torque is provided to the vehicle to obtain the pose, speed and yaw rate simulation data of the tracked vehicle under multiple working conditions. The original dataset composed of these simulation data is preprocessed accordingly to form a training dataset for training the LSTM neural network. Step 3: Train the LSTM neural network model using the training dataset and verify the training effect; the trained LSTM neural network model is used to provide state input for the predictive control model; Step 4: Select For state variables, To control the amount, For output variables, where t Indicates time, subscript e The expected value of the corresponding parameter is represented; and corresponding constraints are set for mechanical limitations, control smoothness, and handling stability respectively; a multi-objective optimization function is established based on the predicted output of the neural network and the reference value in the prediction time domain; thus, a trajectory tracking predictive control model based on neural network model predictive control is constructed. Step 5: Combine the state variables corresponding to the real-time motion state of the tracked vehicle to predict the future motion state variables of the tracked vehicle. Then, use the quadratic programming algorithm in combination with the set constraints to solve the multi-objective optimization problem and obtain the optimal control quantity of the driving torque of the two tracked sides through real-time rolling optimization. The network structure of the LSTM neural network model established in step one specifically consists of an input layer and a... Layer, a fully connected layer and a Layer composition; The input layer is a The sequence, where This indicates the length of the historical data sequence, while This indicates the number of data features selected; in tracked vehicle trajectory prediction, the former indicates how much past information is expected to be used to predict the vehicle state at the next moment, while the latter indicates the dimension of the described vehicle state, i.e. ,in , ,therefore At the same time, set , indicating past tense Historical data is used to predict the vehicle status at the next moment. ; The number of hidden units in the layer is set to The network parameters for the forget gate include: Weight parameters: Bias parameters: That is, the total number of parameters in the forget gate indivual; Input gates include One network parameter; the output gate contains One network parameter; but The total number of parameters that need to be trained in the layer is: Fully connected layers are specifically used for... The output is integrated and dimensionality reduced using... The function is used as the activation function, and the number of hidden units in the fully connected layer is set to 16; The layer is specifically used to randomly disable half of the units in the fully connected layer to avoid overfitting during training, and its number of nodes is also set to 16. The output layer is used to transform the network parameters into vehicle motion state information to be predicted. ； In the trajectory tracking predictive control model based on neural network model predictive control described in step four, the LSTM network is specifically equivalent to a single-layer LSTM neural network that receives time series data and a fully connected neural network that receives one-dimensional data. The single-layer LSTM neural network is independent of the model predictive control and does not participate in the state iteration in the prediction time domain. The prediction of state variables is specifically expressed in the following form: in, F h and F C This represents the process of solving for predicted values ​​and cell states within a single-layer LSTM neural network. The specific form of the prediction equation in the predictive control model is as follows: To improve solution efficiency, the formula is approximated as linearized, resulting in: in, ， ， ， ， ， ， Define state variables , To control the variables, the above system of equations can be written as: in, ， ， The specific form of the output equation of the predictive control model is as follows: in, ; Combining the above equations, we get: ； In step five, the trajectory tracking predictive control model is in t The output at time t is represented as: in, ， , ， in N p To predict the time domain, N c To control the time domain; after discretizing the continuous state space, we have , , , ; To improve computational efficiency, the above parameters are simplified: The reference value for each prediction time domain is , ; Define the following multi-objective optimization function: in, These are the weighting coefficients. As a relaxation factor, As a relaxation term, it can speed up the convergence process and prevent situations where the objective function cannot be solved due to the absence of an optimal solution in the quadratic programming. The constraints on the state variables include: The constraints on the output include: Control constraints include: in, , i∈ ( R , L ), It is the maximum output torque of the motor. This is the maximum output power of the motor. n It refers to the motor speed; Combining the above multi-objective optimization function and constraints, we obtain the following simplified standard form of quadratic programming: The solution is obtained by solving the problem. t The optimal output torque increment sequence at any given time is taken, and the first term is compared with... t- The sum of the control values ​​at time 1 gives the result. t The optimal control quantity at any given time.

2. The method as described in claim 1, characterized in that: The preprocessing of the original dataset in step two includes data cleaning, data normalization, serialization processing with fixed time lengths, and data partitioning.

Images