A method for anti-skid agricultural vehicle navigation control based on driver historical data

Abstract

This invention discloses an anti-skid navigation control method for agricultural machinery based on historical driver data, comprising: S1, collecting operational data of the driver operating the agricultural machinery on slippery road sections, using the Pearson correlation coefficient method to screen key dependent variables affecting steering control, and constructing training samples; S2, constructing an attention-BP-LSTM neural network regression prediction model for desired wheel angle and steering wheel torque; S3, model training; S4, evaluating and verifying the model's prediction results; S5, deploying the model on the terminal controller of the agricultural machinery to acquire real-time operational data of the agricultural machinery. When slippage is detected, the model outputs the desired wheel angle and steering wheel torque, and, combined with the safety protection mechanism of output limiting and gradual control, automatically adjusts the steering mechanism of the agricultural machinery to control the agricultural machinery to escape slippage. This invention generates precise and real-time control commands through deep learning of the driver's historical sideslip correction data, solving the problems of traditional methods relying on physical models and lacking experience utilization.

Description

Technical Field

[0001] This invention belongs to the field of intelligent agricultural machinery technology, specifically relating to an anti-skid agricultural machinery navigation control method based on driver historical data. Background Technology

[0002] In the unmanned operation of paddy field direct seeders, slippage is a crucial factor affecting their safety and stability. Traditional control methods for direct seeders are primarily based on physical models, using sensors to collect vehicle motion parameters in real time, such as heading angle, speed, and yaw rate, and then calculating corresponding control commands according to preset physical formulas. However, this physical model-based method has several limitations. On the one hand, the physical model has limitations due to the difficulty in determining the soil adhesion coefficient in paddy fields, leading to insufficient control accuracy. On the other hand, traditional methods lack effective utilization of driver experience, while the experience accumulated by drivers in long-term operations for correcting sideslip is invaluable in dealing with complex sideslip situations. Currently, although some research attempts to introduce data-driven methods into the field of unmanned driving, there is still no mature technical solution for anti-slip navigation control of agricultural machinery using historical driver data. Therefore, a new control method is urgently needed that can fully utilize historical driver sideslip data to improve the control performance of paddy field direct seeders under sideslip conditions. Summary of the Invention

[0003] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and propose an anti-skid agricultural machinery navigation control method based on driver historical data.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] A navigation control method for anti-skid agricultural machinery based on driver history data includes the following steps:

[0006] S1. Collect operational data of drivers operating agricultural machinery on slippery road sections, use the Pearson correlation coefficient method to screen key dependent variables affecting steering control, and construct training samples of the expected wheel angle and steering wheel motor torque of agricultural machinery.

[0007] S2. Based on the constructed training samples, construct an attention-BP-LSTM neural network regression prediction model for the desired wheel angle and steering wheel torque;

[0008] S3. Training and constructing an attention-BP-LSTM neural network regression prediction model;

[0009] S4. Evaluate and verify the prediction results of the attention-BP-LSTM neural network regression prediction model to obtain the trained attention-BP-LSTM neural network regression prediction model.

[0010] S5. Deploy the attention-BP-LSTM neural network regression prediction model on the terminal controller of the unmanned agricultural machinery to obtain real-time operation data of the unmanned agricultural machinery. When the agricultural machinery is detected to be slipping, the trained attention-BP-LSTM neural network regression prediction model is used to obtain the expected wheel angle and steering wheel torque. Combined with the safety protection mechanism of output limiting and gradual control, the steering mechanism of the agricultural machinery is automatically adjusted to control the agricultural machinery to get out of the slippage.

[0011] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0012] 1. This invention uses GNSS antennas and IMUs on unmanned agricultural machinery to acquire information such as speed, heading deviation, yaw rate, and position deviation during the machinery's movement. An attention-BP-LSTM controller is constructed using a data-driven approach, eliminating the need for complex and precise dynamic models of the machinery, thus avoiding modeling errors and enhancing the system's practicality. By introducing a spatiotemporal dual attention module into the model, key sideslip characteristics (position deviation, yaw rate) and timing nodes are dynamically focused based on the machinery's slip state. Simultaneously, the Adam algorithm, with momentum optimization and adaptive learning rate adjustment, is used to train the model. Finally, the physical constraint of the steering wheel motor torque is introduced as a safety optimization target, improving control accuracy and eliminating the risk of motor overload. Through mean absolute error verification, this invention significantly solves the sideslip problem of agricultural machinery on nonlinear surfaces such as paddy fields, ensuring the quality of operations. The method of this invention has strong adaptability and can effectively improve the intelligence and practicality of unmanned agricultural machinery. Attached Figure Description

[0013] Figure 1 This is an overall flowchart of the method of the present invention.

[0014] Figure 2 This is a mapping diagram of the attention-BP-LSTM neural network regression prediction model of the present invention.

[0015] Figure 3 This is a flowchart of the attention-BP-LSTM neural network regression prediction model of the present invention. Detailed Implementation

[0016] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0017] Examples; such as Figure 1 As shown, an anti-skid navigation control method for agricultural machinery based on driver historical data includes the following steps:

[0018] S1. Collect operational data of drivers operating agricultural machinery on slippery road sections (including dry fields, wet fields, and flooded fields), use the Pearson correlation coefficient method to screen key dependent variables affecting steering control, and construct training samples of the expected wheel angle and steering wheel motor torque of the agricultural machinery.

[0019] The Pearson correlation coefficient method was used to screen the key dependent variables affecting steering control, specifically:

[0020] There are multiple dependent variables affecting the expected wheel angle and steering wheel motor torque of unmanned agricultural machinery. The Pearson correlation coefficient method is used to quantify the influence of the input variables on the output. The correlation coefficient calculation formula is as follows:

[0021]

[0022] in, For the sample values ​​of the independent variable, The mean of the independent variable is the sample value. For the sample values ​​of the dependent variable, The mean of the dependent variable. The correlation coefficient, The closer to 1, the stronger the correlation, and the more significant the impact on the expected wheel angle and steering wheel torque of unmanned agricultural machinery;

[0023] Based on the Pearson correlation coefficient method, the five most significant factors affecting the expected wheel angle and steering wheel motor torque of unmanned agricultural machinery were identified: the speed of the agricultural machinery. Heading deviation Positional deviation yaw rate and heading angle And form a 5-dimensional input vector x:

[0024]

[0025] The output vector is:

[0026]

[0027] in, For the desired wheel angle, This refers to the steering torque of the steering wheel motor;

[0028] Collect several sets of input and output vectors to establish training samples D for the desired wheel angle and steering wheel motor torque:

[0029]

[0030] S2. Based on the constructed training samples, construct an attention-BP-LSTM neural network regression prediction model for the desired wheel angle and steering wheel torque;

[0031] The attention-BP-LSTM neural network regression prediction model includes an LSTM layer, a spatiotemporal dual attention module, and a BP neural network layer. The LSTM layer extracts temporal features, the spatiotemporal dual attention module dynamically assigns feature weights and temporal weights, and the BP neural network layer is used for feature fusion and output prediction. Figure 2 The image shows the mapping diagram of the attention-BP-LSTM neural network regression prediction model.

[0032] The construction of the attention-backpropagation-LSTM neural network regression prediction model specifically includes:

[0033] The input layer is set to have 5 neurons to receive... Components; set the time step T;

[0034] The hidden layers include LSTM layers, spatiotemporal dual attention modules, and BP neural network layers;

[0035] Let the number of units in the LSTM layer be... The number of nodes in the BP layer is The formula for calculating the number of cells in an LSTM layer is:

[0036]

[0037] in, Where T is the number of training samples, and T is the time step. Input feature number;

[0038] The formula for calculating the number of nodes in the BP layer is:

[0039]

[0040] These two calculation formulas enable the model to adapt to the amount of data being processed, prevent overfitting and underfitting, and provide powerful feature extraction capabilities.

[0041] The time step is set to T=5, and the agricultural machinery state data from five consecutive time points are organized into an input sequence. This aims to provide the model with sufficient temporal context information, enabling it to learn the complete dynamic process from the start of slippage to its continuation and recovery. The input time series is as follows:

[0042]

[0043] Configure the activation functions for the hidden and output layers, including:

[0044] The hidden layer has the following tanh activation function:

[0045]

[0046] The output layer uses the purelin linear activation function as follows:

[0047]

[0048] By using different activation functions in formulas (8) and (9), the feature values ​​output by the LSTM layer can converge faster with zero as the center, effectively capturing the positive and negative trends of the agricultural machinery's motion state; and allowing the output layer network to output arbitrary values, directly corresponding to the control quantity.

[0049] The computation process of the LSTM layer includes:

[0050] Input gate calculation is used to control the amount of new information written:

[0051]

[0052] The system determines how much of the current state information of the agricultural machinery needs to be incorporated into memory. When a sudden sideslip of the agricultural machinery is detected, the input gate increases the weight of the new state information.

[0053] Forget gate calculation, used to measure historical memory retention rate:

[0054]

[0055] The forgetting gate primarily controls the degree to which historical state information is forgotten. The forgetting gate can learn and reduce the importance of historical memories that are unrelated to slippage.

[0056] Candidate state calculation and cell state update:

[0057]

[0058]

[0059] This is the memory core of LSTM, which integrates long-term memory filtered by the forget gate and new information filtered by the input gate, thus enabling it to learn the dynamic process of agricultural machinery sliding, rather than just the state at a certain moment.

[0060] The calculation of the output gate and the output of the hidden state are as follows:

[0061]

[0062] ;

[0063] Output gate Control the current memory cell state How much needs to be output? The final hidden state. It contains all the temporal feature information extracted for decision-making at the current moment. It will be fed into the spatiotemporal dual attention module for weighted optimization, and used to finally predict the desired wheel angle and torque.

[0064] The backpropagation calculation for a BP neural network layer is as follows:

[0065] Backpropagation at time step:

[0066] (16);

[0067] Output layer error calculation:

[0068] (17);

[0069] in, is the derivative of the linear activation function, with a value of 1;

[0070] The formula for error propagation in hidden layers is:

[0071] (18);

[0072] in, The derivative of the tanh function is expressed as:

[0073] .

[0074] The spatiotemporal dual attention module is connected in series between the LSTM layer and the BP neural network layer, including a feature attention layer and a temporal attention layer, with the following specific configuration:

[0075] The core of the feature attention layer is to determine the current slip state of the agricultural machinery and dynamically assign weights to five input features. The input is the temporal features output from the LSTM layer, and the output is five core features: speed, etc. Heading deviation Positional deviation yaw rate and heading angle Dynamic weights The weights are linked to the slip criterion for agricultural machinery slippage, satisfying... ;

[0076] The formula for calculating the weights of the feature attention layer is:

[0077]

[0078] in, The dynamic weight of the i-th core feature. This is the weight matrix. Let be the temporal feature vector output by the LSTM layer, 'slip' be the slip criterion, and λ be the sideslip weight gain coefficient. It is the bias vector;

[0079] In this embodiment, the following settings are provided: for The matrix, It is a 5×1 bias vector.

[0080] When slip=0 (normal operation), Uniform distribution (approximately 0.2 each) achieves balanced feature extraction;

[0081] When slip=1 (slip condition), the position deviation and yaw rate change rapidly, therefore (Position deviation weight) increased to 0.35~0.4. The weight of yaw rate is increased to 0.25~0.3, and the total weight of other features is less than 0.35, ensuring that the model focuses on key slip features.

[0082] Temporal Attention Layer: The temporal attention layer assigns weights to the features of the LSTM layer across 5 time steps (t-4 to t); the input is the weighted features output by the feature attention layer, and the output is the dynamic weights for the 5 time steps t-4 to t. Weights at times t and t-1 under sideslip conditions , Automatic upgrade.

[0083] The formula for calculating the weights of the temporal attention layer is:

[0084]

[0085] in, It is a weight matrix. For bias vectors, For the output of the feature attention layer, k=0 and k=1 correspond to time t and t-1, respectively. This is the timing gain coefficient, which is used when the agricultural machinery is in a slip state. and The sum is greater than 0.6.

[0086] In this embodiment, the settings are as follows: for The matrix, It is a 5×1 bias vector;

[0087] When slip=1, at time t The weights are increased to 0.4~0.45 at time t-1. The weights range from 0.25 to 0.3, and the total weights from time t-2 to t-4 are less than 0.3, focusing on key time nodes after the slip occurs.

[0088] When slip=1, the feature attention layer prioritizes positional deviation and yaw rate, simulating how a driver perceives sudden skidding through vision (positional deviation) and tactile sensation (yaw rate) during sideslip; while the temporal attention layer focuses on the most recent moment, simulating the driver's rapid response to sudden situations.

[0089] The output of the spatiotemporal dual attention module is:

[0090]

[0091] in, It is the final output vector of the attention module, which serves as the input to the BP neural network layer. for The input feature at time i.

[0092] During training, the weights of the attention module The algorithm updates via backpropagation through time (BPTT), and the gradient calculation is incorporated into the overall loss function (the sum of the original MSE loss and the safety constraint loss), specifically:

[0093] ;

[0094] ;

[0095] in, The total loss function is the result of accumulating gradients over 5 time steps, which ensures that the attention weights are optimized in tandem with the overall model.

[0096] S3. Training and constructing an attention-BP-LSTM neural network regression prediction model;

[0097] An attention-BP-LSTM neural network regression prediction model trained using the Adam optimizer and a learning rate decay strategy; such as Figure 3 As shown, the training process is an iterative loop that includes forward propagation, loss calculation, backpropagation, and parameter updates, until the model meets the accuracy requirements. The specific steps are as follows:

[0098] Forward propagation: Input the training samples into the network and calculate the output according to the structure described in step S2.

[0099] Loss calculation: The mean square error between the predicted value and the actual value is used as the loss function;

[0100]

[0101] Backpropagation: Calculating the gradient of the loss function with respect to all network weight parameters. To address the time-series data characteristics processed by the LSTM layer, the Backpropagation Through Time (BPTT) algorithm is employed, specifically as follows:

[0102] Since the LSTM layer processes time-series data of length 5, the BPTT algorithm is needed to propagate the error backward along the time axis from the final time t to the initial time t-4 in order to calculate the gradient of the loss function with respect to the parameters of the LSTM layer. The core steps are as follows:

[0103] By calculating the error term between the predicted value and the true value of the output layer:

[0104]

[0105] in, For Hadama accumulation, This is the derivative of the output layer activation function. Since the output layer is a linear activation function, the derivative value is 1.

[0106] The error of the output layer is backpropagated to the feature fusion layer of the attention-BP-LSTM model, and further propagated to the hidden state of the LSTM layer at time t:

[0107]

[0108] Expanding along the time axis: Starting from time t, the gradients at times t-1, t-2, ..., t-4 are calculated sequentially. This process involves the LSTM's forget gate, input gate, output gate, and backpropagation of cell state errors. The gradients of each parameter at each time step are calculated using the chain rule.

[0109] Parameter gradient summarization and update: Accumulate the gradients of all time steps (t to t-4), and then use the Adam optimizer to update all weights of the LSTM layer and BP layer at once.

[0110] The parameter update formula for the Adam optimizer is:

[0111]

[0112] in, and These are the bias correction estimates for the first and second moments, respectively. For smoothing terms;

[0113] The learning rate decay is represented as:

[0114]

[0115] Among them, the initial learning rate attenuation coefficient decay period .

[0116] S4. Evaluate and verify the prediction results of the attention-BP-LSTM neural network regression prediction model to obtain the trained attention-BP-LSTM neural network regression prediction model; specifically:

[0117] The model accuracy is evaluated by calculating the mean absolute error of wheel angle prediction and steering wheel motor torque prediction. The calculation formula is as follows:

[0118]

[0119]

[0120] in, and These are predicted values ​​for wheel angle and torque. and The MAE represents the true values ​​of wheel angle and torque. The smaller the MAE value, the better the model's predictive performance.

[0121] The model training is determined by the MAE value calculated by formulas (30) and (31). If the training is unsuccessful, return to step S3 to perform BPTT backpropagation to update the weights. Otherwise, proceed to step S5.

[0122] S5. Deploy the attention-BP-LSTM neural network regression prediction model on the terminal controller of the unmanned agricultural machinery to obtain real-time operation data of the unmanned agricultural machinery. When the agricultural machinery is detected to be slipping, the trained attention-BP-LSTM neural network regression prediction model is used to obtain the expected wheel angle and steering wheel torque. Combined with the safety protection mechanism of output limiting and gradual control, the steering mechanism of the agricultural machinery is automatically adjusted to control the agricultural machinery to get out of the slippage.

[0123] The criterion for detecting whether agricultural machinery has slipped is:

[0124]

[0125] in, , and To set the threshold for judging composite slip; when A value greater than 1 indicates that the overall movement of the agricultural machinery has significantly deviated from the normal range, and is therefore determined to be slip. This method uses the square root of the sum of squares to fuse the deviations of the three dimensions into a single scalar. The system determines slip (slip=1) if and only if this comprehensive deviation is greater than 1. Furthermore, this method is superior to single-parameter judgment and has stronger anti-interference capabilities.

[0126] During operation, the agricultural machinery's current status is monitored in real time. If the machinery's movement significantly deviates from the normal range, the system determines it to be slippage and automatically switches to an attention-backpropagation-LSTM neural network control mode. The predicted values ​​for the next moment are used as the target decision for the desired wheel angle and steering wheel motor torque. Once correction is complete, the system detects... When this occurs, the system determines that the operation is normal and restores the original control mode.

[0127] The trained attention-BP-LSTM neural network regression prediction model is deployed on the terminal controller of the unmanned agricultural machinery. Data from the unmanned agricultural machinery's operation over a period of time is used as test data sample x. When slippage is detected, the attention-BP-LSTM neural network regression prediction model is used to obtain the corresponding predicted expected wheel angle of the unmanned agricultural machinery. and steering wheel motor torque And use this as the target decision quantity.

[0128] In step S5, the security protection mechanism is as follows:

[0129] The expected wheel angle is based on the output of the attention-BP-LSTM neural network regression prediction model. The gradual control and steering wheel motor torque T control steering mechanism, wherein the torque output is set to meet the following constraints:

[0130]

[0131] Gradual control of wheel angle:

[0132]

[0133] in, It is the new wheel angle that is executed. Wheel angle gradual control is used to achieve a smooth transition of steering commands and avoid mechanical shock.

[0134] The described maximum torque The torque of the steering wheel motor and the steering mechanism of the agricultural machinery of the CES-T model selected in this embodiment are set. The gradient coefficient of 0.3 is the optimal balance point between command response speed and mechanical system stability obtained through field testing.

[0135] It should also be noted that, in this specification, terms such as "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0136] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for anti-skid agricultural vehicle navigation control based on driver historical data, characterized in that, Includes the following steps: S1. Collect operational data of drivers operating agricultural machinery on slippery road sections, use the Pearson correlation coefficient method to screen key dependent variables affecting steering control, and construct training samples of the expected wheel angle and steering wheel motor torque of agricultural machinery. S2. Based on the constructed training samples, construct an attention-BP-LSTM neural network regression prediction model for the desired wheel angle and steering wheel torque; S3. Training and constructing an attention-BP-LSTM neural network regression prediction model; S4. Evaluate and verify the prediction results of the attention-BP-LSTM neural network regression prediction model to obtain the trained attention-BP-LSTM neural network regression prediction model. S5. Deploy the attention-BP-LSTM neural network regression prediction model on the terminal controller of the unmanned agricultural machinery to obtain real-time operation data of the unmanned agricultural machinery. When the agricultural machinery is detected to be slipping, the trained attention-BP-LSTM neural network regression prediction model is used to obtain the expected wheel angle and steering wheel torque. Combined with the safety protection mechanism of output limiting and gradual control, the steering mechanism of the agricultural machinery is automatically adjusted to control the agricultural machinery to get out of the slippage.

2. The anti-skid agricultural vehicle navigation control method based on driver history data according to claim 1, characterized in that, In step S1, the Pearson correlation coefficient method is used to screen the key dependent variables affecting steering control, specifically: There are multiple dependent variables affecting the expected wheel angle and steering wheel motor torque of unmanned agricultural machinery. The Pearson correlation coefficient method is used to quantify the influence of the input variables on the output. The correlation coefficient calculation formula is as follows: wherein, is a sample value of the independent variable, is the mean of the sample of the independent variable, is a sample value of the dependent variable, is the mean of the sample of the dependent variable, is the correlation coefficient, the closer to 1 represents the stronger correlation, the more significant the influence of the expected wheel angle and steering wheel torque on the unmanned agricultural machine; According to the Pearson correlation coefficient method, the most significant 5 factors affecting the expected wheel angle and steering motor torque of the unmanned agricultural machine are screened out: the speed , heading deviation , position deviation , yaw rate and heading angle , and then a 5-dimensional input vector x is composed The output vector is: wherein, is the desired wheel angle, is the steering torque of the steering motor; Collect several sets of input and output vectors to establish training samples D for the desired wheel angle and steering wheel motor torque:

3. The anti-skid agricultural vehicle navigation control method based on driver history data according to claim 1, characterized in that, In step S2, the attention-BP-LSTM neural network regression prediction model includes an LSTM layer, a spatiotemporal dual attention module, and a BP neural network layer; wherein the LSTM layer is used to extract temporal features, the spatiotemporal dual attention module is used to dynamically allocate feature weights and temporal weights, and the BP neural network layer is used for feature fusion and output prediction. The construction of the attention-backpropagation-LSTM neural network regression prediction model specifically includes: The input layer is set to 5 neurons, which receive components; a time step T is set; The hidden layers include LSTM layers, spatiotemporal dual attention modules, and BP neural network layers; The number of LSTM layer units is set as The number of BP layer nodes is The calculation formula of the number of LSTM layer units is: wherein, T is the number of training samples, T is the time step, is the number of input features; The formula for calculating the number of nodes in the BP layer is: With a time step of T=5, the input time series is: Configure the activation functions for the hidden and output layers, including: The hidden layer has the following tanh activation function: The output layer uses the purelin linear activation function as follows: The computation process of the LSTM layer includes: Input gate calculation is used to control the amount of new information written: Forget gate calculation, used to measure historical memory retention rate: Candidate state calculation: Cell state update calculation: The calculation of the output gate and the output of the hidden state are as follows: ； The backpropagation calculation for a BP neural network layer is as follows: Backpropagation at time step: ； Output layer error calculation: ； wherein, is the derivative of the linear activation function, which has a value of 1; The formula for error propagation in hidden layers is: ； wherein is the derivative of the tanh function, denoted as: 。 4. The anti-skid agricultural vehicle navigation control method based on driver history data according to claim 1, wherein, The spatiotemporal dual attention module is connected in series between the LSTM layer and the BP neural network layer, including a feature attention layer and a temporal attention layer, with the following specific configuration: The feature attention layer takes temporal features from the LSTM layer as input and outputs five core features. (Speed) Heading deviation Positional deviation yaw rate and heading angle Dynamic weights The weights are linked to the slip criterion for agricultural machinery slippage, satisfying... ; The temporal attention layer takes as input the weighted features output from the feature attention layer and outputs dynamic weights for five time steps t-4 to t. Weights at times t and t-1 under sideslip conditions , Automatic upgrade.

5. The anti-skid agricultural vehicle navigation control method based on driver history data according to claim 4, wherein, The formula for calculating the weights of the feature attention layer is: wherein, is a dynamic weight for the i-th core feature, is a weight matrix, is a time series feature vector output by the LSTM layer, slip is a slip criterion, and λ is a side slip weight gain coefficient, is a bias vector. The formula for calculating the weights of the temporal attention layer is: wherein, is a weight matrix, is a bias vector, is a feature attention layer output, k = 0 and k = 1 correspond to t and t-1 moments, is a timing gain coefficient, when the agricultural machine is in a side slip state, and the sum is greater than 0.

6.

6. The anti-skid agricultural vehicle navigation control method based on driver history data according to claim 5, wherein, The output of the spatiotemporal dual attention module is: wherein, is the attention module final output vector as input to the BP neural network layer, is the i-th input feature at time t. is the i-th input feature at time t.

7. The anti-skid agricultural vehicle navigation control method based on driver history data according to claim 1, wherein, In step S3, the attention-BP-LSTM neural network regression prediction model is trained using the Adam optimizer and the learning rate decay strategy. The parameter update formula for the Adam optimizer is: wherein, and are bias-corrected estimates of the first and second moments, respectively, is a smoothing term; The learning rate decay is represented as: wherein the initial learning rate the decay coefficient the decay period .

8. The anti-skid agricultural vehicle navigation control method based on driver history data according to claim 1, wherein, In step S4, the evaluation and verification of the prediction results of the attention-BP-LSTM neural network regression prediction model are specifically carried out as follows: The model accuracy is evaluated by calculating the mean absolute error of wheel angle prediction and steering wheel motor torque prediction. The calculation formula is as follows: wherein, and are the predicted values of wheel angle and torque, and are the true values of wheel angle and torque, the smaller the value of MAE, the better the model prediction performance; The model training success is determined based on the calculated MAE value. If the training is unsuccessful, return to step S3; otherwise, proceed to step S5.

9. The anti-skid agricultural vehicle navigation control method based on driver history data according to claim 5, wherein, In step S5, the criterion slip for detecting whether the agricultural machinery has slipped is as follows: wherein, , and is a judgment threshold value for setting composite slip; when the value is greater than 1, it indicates that the overall motion state of the agricultural machine has deviated significantly from the normal range, and it is determined that slip has occurred.

10. The anti-skid agricultural machinery navigation control method based on driver historical data according to claim 1, characterized in that, In step S5, the security protection mechanism is as follows: The expected wheel angle is based on the output of the attention-BP-LSTM neural network regression prediction model. The gradual control and steering wheel motor torque T control steering mechanism, wherein the torque output is set to meet the following constraints: Gradual control of wheel angle: in, It is the new wheel angle that is executed. Wheel angle gradual control is used to achieve a smooth transition of steering commands and avoid mechanical shock.

Images