Unmanned navigation method and system applied to pasture harvesting

Abstract

The invention provides an unmanned navigation method and system applied to pasture harvesting, and relates to the technical field of agricultural machinery automation, and the method comprises the steps: firstly capturing a pasture growth state signal, a terrain interaction feedback signal and an operation track residual signal in an operation area, and constructing a track residual dynamic linkage model; generating a trajectory evolution rule including trajectory bending amplitude adjustment, elevation compensation and offset avoidance modes based on the trajectory residual dynamic linkage model; real-time signals and current operation track data of equipment in the operation process are collected through a global sensing assembly; inputting the parameter into a track residual dynamic linkage model to calculate and generate a track deviation correction parameter; and according to the trajectory deviation correction parameter and the trajectory evolution rule, generating a dynamic trajectory adjustment instruction to drive the equipment to update the operation trajectory, and reversely inputting the updated data into the model to complete loop optimization. The precision and adaptability of unmanned navigation for pasture harvesting are improved, and the harvesting efficiency and quality are improved.

Description

Technical Field

[0001] This invention relates to the field of agricultural machinery automation technology, and more specifically, to an unmanned driving navigation method and system for forage harvesting. Background Technology

[0002] In traditional forage harvesting operations, autonomous vehicle navigation primarily relies on pre-set fixed paths or local adjustments based on simple environmental perception. However, the environment of forage planting areas is complex and variable, with uneven growth patterns, varying height and density across different areas, and diverse terrain including undulations and depressions. Existing navigation methods do not fully consider the impact of forage growth status signals on the trajectory of autonomous equipment. For example, in densely growing areas, the equipment may need to adjust the trajectory curvature for better harvesting. Regarding terrain feedback signals, traditional methods struggle to accurately adjust trajectory elevation based on real-time terrain changes, potentially leading to incomplete harvesting or collisions with terrain when encountering slopes or low-lying areas. Furthermore, residual trajectory signals are not effectively utilized; remnants from previous harvests may affect subsequent operations, but traditional navigation cannot use this information to avoid trajectory deviations. These shortcomings result in poor navigation accuracy and adaptability for autonomous equipment during forage harvesting, hindering efficient and precise harvesting operations. Summary of the Invention

[0003] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide an unmanned navigation method for forage harvesting, the method comprising: The system captures signals of pasture growth status, terrain interaction feedback, and residual operation trajectory within the work area, and constructs a dynamic linkage model for trajectory residue. This model associates the interaction paths between pasture growth status signals, terrain interaction feedback signals, residual operation trajectory signals, and the subsequent operation trajectory of the unmanned vehicle. Based on the trajectory residue dynamic linkage model, the trajectory evolution rules of the unmanned driving equipment are generated by signal feature matching. The trajectory evolution rules include the trajectory curvature adjustment method corresponding to the pasture growth status signal, the trajectory elevation compensation method corresponding to the terrain interaction feedback signal, and the trajectory deviation avoidance method corresponding to the operation trajectory residue signal. The unmanned vehicle's all-domain perception components collect real-time signals of pasture growth, real-time terrain interaction feedback, real-time residual signals of the operation trajectory, and current operation trajectory data of the equipment during the operation process. The real-time signal of pasture growth status, the real-time feedback signal of terrain interaction, the real-time residual signal of operation trajectory, and the current operation trajectory data of the equipment are input into the trajectory residual dynamic linkage model to perform trajectory adaptation calculation and generate trajectory deviation correction parameters. Based on the trajectory deviation correction parameters and the trajectory evolution rules, a dynamic trajectory adjustment command is generated to drive the unmanned driving equipment to update its operating trajectory. At the same time, the updated equipment operating trajectory data and the newly generated operating trajectory residual signal are input back into the trajectory residual dynamic linkage model to complete the cyclic optimization of the trajectory residual dynamic linkage model.

[0004] Furthermore, embodiments of the present invention also provide an unmanned navigation system for forage harvesting, characterized in that it includes: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the above-described unmanned navigation method for forage harvesting by executing the machine-executable instructions.

[0005] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions stored in a computer-readable storage medium, a processor of an unmanned navigation system for hay harvesting reads the machine-executable instructions from the computer-readable storage medium, the processor executes the machine-executable instructions, causing the unmanned navigation system for hay harvesting to perform the above-described unmanned navigation method for hay harvesting.

[0006] Based on the above, a dynamic linkage model for trajectory residue is constructed to link the interaction paths of pasture growth status signals, terrain interaction feedback signals, operation trajectory residue signals, and subsequent operation trajectories of unmanned vehicles. Based on the trajectory evolution rules generated by this dynamic linkage model, the trajectory can be flexibly adjusted according to different signal characteristics. For example, the trajectory curvature can be adjusted according to pasture growth status signals, trajectory elevation compensation can be performed according to terrain interaction feedback signals, and trajectory deviation avoidance can be achieved according to operation trajectory residue signals. This greatly improves the equipment's adaptability to complex environments. Real-time signals and the equipment's current operation trajectory data are collected by a global perception component and input into the dynamic linkage model for trajectory residue to calculate trajectory adaptability and generate trajectory deviation correction parameters. This achieves real-time and accurate correction of the operation trajectory. Finally, dynamic trajectory adjustment commands are generated based on the trajectory deviation correction parameters and trajectory evolution rules, driving the equipment to update its operation trajectory. The updated data is then input back into the dynamic linkage model for trajectory residue to complete iterative optimization, enabling the navigation system to continuously improve itself and maintain high-precision navigation performance. This significantly improves the efficiency and quality of pasture harvesting and reduces operating costs. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the execution flow of an unmanned navigation method for forage harvesting provided in an embodiment of the present invention.

[0008] Figure 2 This is a schematic diagram of exemplary hardware and software components of an unmanned navigation system for forage harvesting provided in an embodiment of the present invention. Detailed Implementation

[0009] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating an embodiment of the unmanned navigation method for hay harvesting provided by the present invention. The unmanned navigation method for hay harvesting will be described in detail below.

[0010] Step S110: Capture the pasture growth status signal, terrain interaction feedback signal, and operation trajectory residue signal within the operation area, and construct a trajectory residue dynamic linkage model. The trajectory residue dynamic linkage model associates the interaction path between the pasture growth status signal, terrain interaction feedback signal, operation trajectory residue signal, and the subsequent operation trajectory of the unmanned driving equipment.

[0011] This embodiment uses a typical grassland pasture operation area as the application scenario. The terrain of this pasture operation area alternates between gentle slopes and flat land, with the main forage grasses being sheepgrass and needlegrass. There are also localized differences in forage grass density due to low-lying terrain. In this scenario, the unmanned vehicle needs to complete a large-scale forage harvesting operation, and the accuracy of its operation trajectory directly affects harvesting efficiency and forage utilization. To achieve precise navigation, a dynamic linkage model of trajectory residue needs to be established first, which can comprehensively reflect the correlation between multi-source signals and the operation trajectory. The construction of the aforementioned dynamic linkage model of trajectory residue requires multiple stages, including signal acquisition, feature extraction, correlation modeling, and optimization verification. In this embodiment, a hybrid architecture combining deep learning and traditional machine learning is adopted in the construction of the dynamic linkage model of trajectory residue. The dynamic linkage model of trajectory residue is divided into four layers, from bottom to top: a ternary signal input layer, a feature correlation operation layer, a trajectory mapping layer, and a result output layer. Each layer achieves signal interaction through a fully connected data link. The ternary signal input layer contains three parallel feature input channels, each corresponding to a type of signal feature vector. A dimension matching layer is set inside the channel to uniformly map the input feature vectors to a 256-dimensional feature space. The ReLU activation function is used for nonlinear transformation. The pasture growth status feature channel receives the concatenation result of three sub-feature vectors: plant distribution, stem toughness, and leaf density. The terrain interaction feedback feature channel processes contact reaction force, deformation rebound, and particle friction sub-vectors. The operation trajectory residue feature channel receives surface compaction marks, pasture residue distribution, and trajectory path imprint sub-vectors. The outputs of the three channels are recalibrated through a channel attention mechanism and then enter the feature association operation layer. The feature association operation layer consists of four residual network blocks connected in series. Each residual block contains two 3×3 convolutional layers (256 and 512 channels respectively), a batch normalization layer, and a LeakyReLU activation function. Gradients are passed between blocks via skip connections. A spatial attention module is inserted after the third residual block to enhance key features by calculating the channel weights of the feature maps. This layer ultimately outputs a 1024-dimensional fused feature vector, which contains high-order correlation information of the three types of input signals. The trajectory mapping layer adopts a bidirectional long short-term memory (BiLSTM) network structure, containing two LSTM units, one forward and one backward. Each unit has 256 memory nodes. The input is the fused feature vector sequence output by the feature association operation layer (with a time step of 5, corresponding to feature data from 5 consecutive acquisition cycles). The temporal dependencies of the feature sequence are captured through a gating mechanism. The network output is reduced to a 32-dimensional intermediate feature vector through a fully connected layer. The output layer contains three parallel fully connected branches, corresponding to the three adjustment dimensions of trajectory curvature amplitude, elevation compensation, and offset avoidance, respectively. Each branch consists of two hidden layers (128 neurons + 64 neurons) and one output neuron, and uses a linear activation function to output specific adjustment values. The branches achieve feature reuse by sharing hidden layer parameters.

[0012] The model training process is divided into two stages: pre-training and fine-tuning. The pre-training stage uses 100,000 sets of historical operational data, covering operational scenarios with different seasons, terrains (mountains, plains, depressions), and pasture types (sheepgrass, needlegrass, alfalfa). The input data consists of grid-based triplet signal sets containing three types of feature vectors (each sample has a dimension of 256×3), and the output is the corresponding trajectory adjustment parameters (continuous values ​​for curvature amplitude, elevation compensation, and offset avoidance). Before training, the input features are preprocessed: Min-Max normalization is used to scale all feature values ​​to the [0,1] interval; missing values ​​are filled using K-nearest neighbor interpolation; and outlier samples other than 3σ are identified and removed using the Z-Score method. Pre-training used the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, and 100 training epochs. A cosine annealing learning rate strategy was employed (decreasing to 0.8 times the current value every 10 epochs). The loss function was a weighted sum of mean squared error (MSE) and an L2 regularization term (weight decay coefficient of 0.0005). Five-fold cross-validation was used during training. An early stopping mechanism was triggered when the validation set loss did not decrease for 15 consecutive epochs, and the model parameters with the minimum validation set loss were saved. In the fine-tuning phase, 5000 sets of real-time collected data from the current working region were used. The parameters of the first two residual blocks of the feature association operation layer were frozen, and only the remaining network layers were trained. The learning rate was reduced to 0.0001, the batch size was 16, and the training epochs were 20 to adapt to the signal feature distribution of the specific working environment.

[0013] During the model application phase, the input data processing flow is as follows: First, real-time signals collected by the global perception component are received. The real-time signal of pasture growth status is obtained by acquiring one frame every 30 seconds using a multispectral camera (1024×768 resolution). After extracting plant contours using an image segmentation algorithm, the distribution density gradient (using the Sobel operator to calculate the x / y direction gradient) and uniformity (based on the information entropy formula) are calculated. The terrain interaction feedback signal is obtained by collecting contact reaction force from pressure sensors (sampling frequency 100Hz) installed on the wheel suspension system. High-frequency noise is removed using a low-pass filter (cutoff frequency 5Hz). The residual signal of the operation trajectory is obtained by capturing surface images after operation using a rear high-definition camera (2048×1536 resolution). The compaction trace contour is extracted using an edge detection algorithm. All real-time signals are processed to form a feature vector (256 dimensions) with the same dimensions as the training data. After undergoing the same Min-Max normalization process as during pre-training, the vector is input into the model. During model inference, a sliding window mechanism (window size 5, step size 1) is used to process continuously acquired feature sequences. The intermediate feature vectors output by the BiLSTM layer are converted into three-dimensional trajectory adjustment parameters by a fully connected network in the trajectory mapping layer. The output results need to be checked against equipment kinematic constraints (e.g., bending angle range -30° to 30°, elevation compensation range -0.5m to 0.5m), and truncation is performed when the range is exceeded. After the inference results are superimposed with the basic adjustment parameters in the trajectory evolution rules, the final dynamic trajectory adjustment command is generated and sent to the actuator controller via the CAN bus at a frequency of 10Hz.

[0014] To ensure the reliability of the model in actual operations, the main model adopts the aforementioned deep learning architecture, while the backup model uses a traditional machine learning model based on random forests (containing 500 decision trees, a maximum depth of 20, and a minimum number of leaf node samples of 5). When the main model's inference time exceeds 100ms or its output parameters exceed the normal range, the system automatically switches to the backup model. After each day's operations, the system automatically collects real-time signal data, trajectory adjustment commands, and actual operational results (forage harvesting rate, fuel consumption). Model parameters are updated through incremental learning, and the update process uses knowledge distillation technology (with a temperature parameter set to 5) to transfer knowledge from the main model to the backup model, ensuring synchronized performance improvement for both models. Regarding privacy protection, all collected image data is treated with Gaussian noise (noise standard deviation 0.01) using a differential privacy algorithm before storage. Location information is obtained using coordinate offset technology (offsets are randomly generated within a 5m range) to ensure that the original data cannot be used to reverse-locate specific work sites.

[0015] Step S111: Divide the work area into multiple signal acquisition grids of equal area, with the boundary of each signal acquisition grid seamlessly connected to the adjacent grid.

[0016] For the aforementioned grassland and pasture operation area, a grid-based division method was adopted. First, the planar coordinate data of the operation area was acquired using satellite remote sensing imagery. Based on the total area of ​​the area and preset grid size parameters, the operation area was divided into M rows and N columns of equal-area rectangular grids. The side length of each grid was determined according to the uniformity of pasture growth and the accuracy requirements of sensor data acquisition, ensuring that the grid boundaries completely coincided with adjacent grids without overlap or gaps. For example, during grid division, the upper left corner of the operation area was used as the origin, the east-west direction was designated as the X-axis, and the north-south direction as the Y-axis. Grids were divided sequentially according to the set side lengths, and each grid was assigned a unique two-dimensional coordinate code (i, j), where i is the row index and j is the column index. After division, a grid division vector map was generated as a spatial reference for subsequent sensor deployment and data acquisition.

[0017] Step S112: Deploy pasture growth sensing devices within each signal acquisition grid. The pasture growth sensing devices capture plant distribution signals, stem toughness signals, and leaf density signals of pasture within the grid using vegetation sensing technology, and continuously record the continuous change data of each signal.

[0018] A forage growth sensing device is deployed at the center of each signal acquisition grid. This device integrates a multispectral camera, a stem mechanical sensor, and a laser leaf density meter. The multispectral camera generates a two-dimensional image signal of plant distribution by acquiring reflectance spectral data in the 400-900nm band. The image resolution is set according to the grid size to ensure coverage of the entire grid area. The stem toughness signal is acquired by a contact mechanical probe installed at the bottom of the device. The probe performs puncture tests on forage stems at different locations within the grid at a set frequency, recording the resistance change curve during the puncture process. The leaf density signal is obtained through a laser scanner. The laser beam emitted by the scanner covers the grid in a fan-shaped angle, and the number of leaves per unit volume is calculated by receiving the intensity and frequency of the reflected beam. The device has a built-in data storage module that continuously records the raw data of the above three signals at a set sampling interval (e.g., once every 30 seconds), and adds a collection timestamp and grid coordinate code.

[0019] Step S113: Deploy terrain interactive detection devices within each signal acquisition grid. The terrain interactive detection devices capture contact reaction force signals, deformation rebound signals, and particle friction signals of the grid surface through pressure sensing technology, and continuously record the continuous change data of each signal.

[0020] The terrain interactive detection device employs an embedded installation method, selecting three evenly distributed sampling points within each grid, with an integrated terrain sensor installed at each sampling point. The sensor probe is buried at a predetermined depth below the ground surface, internally containing a pressure strain gauge, a displacement sensor, and a friction coefficient detection module. The contact reaction force signal is obtained by collecting the normal force exerted by the ground surface on the probe through the pressure strain gauge, with the sampling frequency consistent with that of the pasture growth sensing device. The deformation rebound signal is obtained by recording the probe's sinking depth under pressure and its recovery displacement curve after the pressure is removed through the displacement sensor. The particle friction signal is obtained by detecting the relative motion resistance between the probe and surrounding soil particles, reflecting the looseness of the surface soil. The device also features timestamp and coordinate encoding functions, storing continuous change data of the three signals in real time.

[0021] Step S114: Deploy trajectory residue monitoring equipment in each signal acquisition grid. The trajectory residue monitoring equipment uses image sensing technology to capture surface compaction traces, pasture residue distribution signals, and trajectory path imprint signals in the already operated areas within the grid, and continuously records the continuous change data of each signal.

[0022] The track residue monitoring equipment uses high-definition industrial cameras mounted above the grid. The camera installation height is adjusted according to the grid size to ensure that a single image can completely cover the entire grid. Surface compaction trace signals are obtained by acquiring grayscale images of the surface under different lighting conditions. Compacted and uncompacted areas exhibit different grayscale values ​​in the image due to differences in reflectivity. Forage residue distribution signals are obtained by using image segmentation algorithms to identify unharvested forage areas in the image and calculate their area proportion and distribution pattern. Track path imprint signals are obtained by detecting wheel tracks left by equipment on the ground and extracting the width, depth, and continuity features of the imprints. The camera captures images at set time intervals, and the raw image data, along with the corresponding timestamps and grid coordinates, are stored together.

[0023] Step S115: Set the signal synchronization acquisition cycle. According to the set cycle, the plant distribution signal, stem toughness signal, leaf density signal, contact reaction force signal, deformation rebound signal, particle friction signal, surface compaction mark signal, pasture residue distribution signal, and trajectory path imprint signal in each signal acquisition grid are synchronously bound to form a grid ternary signal group.

[0024] A unified signal synchronization acquisition cycle is established, which is determined based on a combination of the pasture growth rate and the equipment operating speed. At the beginning of each acquisition cycle, the system sends a synchronization acquisition command to all three types of sensor devices within the grid. Upon receiving the command, each device simultaneously starts data acquisition. After acquisition, each device packages the signal data for that cycle and sends it to the regional data processing center via a wireless transmission module. The data processing center associates and binds the received pasture growth status signals (plant distribution, stem toughness, leaf density), terrain interaction feedback signals (contact reaction force, deformation rebound, particle friction), and operation trajectory residual signals (surface compaction marks, pasture residual distribution, trajectory path imprints) within the same grid and the same cycle, generating a grid ternary signal group containing nine types of signal data. Each signal group is assigned a unique identifier, including grid coordinates (i,j) and the acquisition cycle number.

[0025] Step S116: Extract feature vectors of pasture growth status related signals in each grid ternary signal group. The feature vectors are constructed based on the spatial distribution of plant distribution signals, the amplitude variation trend of stem toughness signals, and the density fluctuation of leaf density signals. Extract feature vectors of terrain interaction feedback related signals in each grid ternary signal group. The feature vectors are constructed based on the response peak variation of contact reaction force signals, the recovery time variation of deformation rebound signals, and the intensity fluctuation of particle friction signals. Extract feature vectors of operation trajectory residue related signals in each grid ternary signal group. The feature vectors are constructed based on the area expansion trend of surface compaction trace signals, the uniformity variation of pasture residue distribution signals, and the depth fluctuation of trajectory path imprint signals.

[0026] Feature extraction is performed on the three types of signals in each grid ternary signal group. For the forage growth status signal, the spatial distribution of the plant distribution signal is characterized by calculating the centroid coordinates, distribution entropy, and clustering coefficient of the plant pixels in the image; the amplitude variation trend of the stem toughness signal is achieved by extracting the peak value, mean, and variance of the resistance curve; the density fluctuation of the leaf density signal is obtained by calculating the standard deviation of the number of leaves per unit area and the spatial autocorrelation coefficient. The extracted feature values ​​are arranged in a preset order to form a forage growth status feature vector with dimension D1.

[0027] In the feature extraction of terrain interaction feedback signals, the change in the peak response of the contact reaction force signal is achieved by detecting the maximum and second-largest peak values ​​and their time difference in the output curve of the pressure strain gauge; the change in the recovery time of the deformation rebound signal is characterized by calculating the time it takes for the subsidence depth recorded by the displacement sensor to reach 50% and 90% of its stable value; and the intensity fluctuation of the particle friction signal is obtained by calculating the root mean square error of the output value of the friction coefficient detection module. These feature values ​​are combined into a terrain interaction feedback feature vector with dimension D2.

[0028] In terms of feature extraction of residual signals of operation trajectory, the area expansion trend of surface compaction trace signals is realized by calculating the growth rate of compacted area and morphological complexity index within the continuous acquisition period; the uniformity change of pasture residue distribution signals is characterized by calculating the spatial distribution standard deviation and information entropy of the residue area; the depth fluctuation of trajectory path imprint signals is obtained by extracting the gradient change of gray values ​​in the imprint image and the variance of the depth estimate. The above feature values ​​constitute a D3 dimension operation trajectory residue feature vector.

[0029] Step S117: Collect historical operation trajectory data of multiple unmanned driving devices in different work areas. The historical operation trajectory data includes trajectory curvature records, trajectory elevation compensation records, and trajectory deviation avoidance records.

[0030] By integrating with the operational log system of unmanned vehicles, historical operational data from multiple devices operating in grassland and pasture environments similar to the current operational area over the past three years were collected. The historical operational trajectory data is stored as a GPS coordinate sequence, containing the device's position, speed, and attitude information at each sampling moment. Trajectory curvature records were extracted by calculating the rate of change of the tangent angle between adjacent path points; trajectory elevation compensation records were directly obtained from the output data of the device's elevation sensor, reflecting the device's height adjustment in undulating terrain; and trajectory deviation avoidance records were calculated by comparing the lateral deviation distance between the planned path and the actual path. The above data was cleaned to remove outliers caused by equipment malfunctions or extreme weather, forming a standardized historical operational trajectory dataset.

[0031] Step S118: Perform correlation calculations on the pasture growth status feature vector, terrain interaction feedback feature vector, and operation trajectory residual feature vector of each signal acquisition grid with the corresponding historical operation trajectory data to establish a correspondence table among the four.

[0032] Based on grid coordinates and operation timestamps, the three types of feature vectors for each signal acquisition grid are matched with trajectory records within the corresponding spatiotemporal range in historical operation trajectory data. For example, if an unmanned vehicle happens to pass through a grid during the operation period corresponding to the three-element signal group in a specific acquisition cycle, the feature vector of that grid is associated with the trajectory curvature, elevation compensation, and offset avoidance records of the vehicle within that grid area. By calculating the mutual information value between the feature vectors and trajectory records, sample pairs with strong correlations are selected, and a four-way correspondence table is constructed, containing grid coordinates, the three types of feature vectors, and the corresponding trajectory records. Each record in the table contains the values ​​of each dimension of the feature vector and the specific values ​​of the trajectory parameters.

[0033] Step S119: Build the basic architecture of the trajectory residual dynamic linkage model. The basic architecture includes a three-element signal input layer, a feature association operation layer, a trajectory mapping layer, and a result output layer. Each layer realizes continuous signal interaction through a data transmission link.

[0034] The trajectory residue dynamic linkage model adopts a deep learning architecture. The ternary signal input layer is designed with three parallel feature input channels, receiving feature vectors of pasture growth status, terrain interaction feedback, and operation trajectory residue, respectively. Each channel contains a fully connected layer that maps the input feature vectors to a feature space of a unified dimension. The feature association operation layer consists of multiple residual network blocks, each containing a convolutional layer, a batch normalization layer, and an activation function. It captures the nonlinear correlation between the three types of features through cross-channel feature fusion operations. The trajectory mapping layer adopts a long short-term memory network structure, mapping the associated feature sequence to predicted values ​​of trajectory parameters. The result output layer is a fully connected layer that outputs the predicted results of trajectory curvature amplitude, elevation compensation, and offset avoidance. All layers are connected through fully connected data transmission links to ensure continuous signal flow and interaction within the model.

[0035] Step S1110: Input the correspondence table of all signal acquisition grids into the feature association operation layer of the infrastructure, and generate an initial correlation coefficient matrix through the multi-signal interaction algorithm of the feature association operation layer. The initial correlation coefficient matrix reflects the strength of the interaction between the four.

[0036] The constructed correspondence table is input into the feature association operation layer of the model in grid number order. The multi-signal interaction algorithm of this feature association operation layer first standardizes the three types of input feature vectors to eliminate dimensional differences. Then, it calculates the association weights between each dimension of the feature vectors through an attention mechanism, multiplies the weight values ​​by the feature values, and performs cross-feature concatenation. Next, a multilayer perceptron performs a nonlinear transformation on the concatenated features, outputting association values ​​reflecting the strength of the interaction between features. These association values ​​are arranged according to the feature dimension and the trajectory parameter dimension to form an initial association coefficient matrix. The row indices of the matrix correspond to each dimension of the three types of feature vectors, the column indices correspond to each dimension of the trajectory parameters, and the matrix element values ​​represent the association strength between the corresponding feature dimension and the trajectory parameter dimension.

[0037] In another example, step S1110: input the correspondence table of all signal acquisition grids into the feature association operation layer of the infrastructure, and generate an initial correlation coefficient matrix through the multi-signal interaction algorithm of the feature association operation layer. The initial correlation coefficient matrix reflects the strength of the interaction between the four.

[0038] Step S11101: Import the correspondence table of all signal acquisition grids into the feature association operation layer of the trajectory residue dynamic linkage model infrastructure in order of grid number. Perform feature parsing on each correspondence table to extract the feature vector of pasture growth status, the feature vector of terrain interaction feedback, the feature vector of operation trajectory residue, and the core association features of historical operation trajectory data.

[0039] Arranged in ascending order by the grid's two-dimensional coordinate encoding (i,j), the correspondence tables of all signal acquisition grids are sequentially imported into the feature association operation layer. For each correspondence table, structured parsing is performed through the feature parsing module to extract four types of core data: pasture growth status feature vectors (including plant distribution, stem toughness, and leaf density sub-vectors), terrain interaction feedback feature vectors (including contact reaction force, deformation rebound, and particle friction sub-vectors), operation trajectory residue feature vectors (including surface compaction marks, pasture residue distribution, and trajectory path imprint sub-vectors), and historical operation trajectory data (including curvature amplitude, elevation compensation, and offset avoidance parameters). For each type of data, principal component analysis is used to extract the top K principal components with the highest variance contribution as core association features, ensuring that key information of the original data is preserved while reducing dimensionality.

[0040] Step S11102: Through the multi-signal interaction algorithm of the feature association operation layer, perform pairwise interaction operations on various feature vectors in the correspondence table of a single grid and historical operation trajectory data to generate local correlation coefficients within the grid. The local correlation coefficients reflect the interaction strength between different features and trajectory data within a single grid.

[0041] The multi-signal interaction algorithm adopts a feature association method based on mutual information, and performs pairwise combination operations on four core association features of a single grid: (1) pasture growth status feature vector and historical operation trajectory data; (2) terrain interaction feedback feature vector and historical operation trajectory data; (3) operation trajectory residual feature vector and historical operation trajectory data; (4) pairwise combination between the three types of feature vectors. For each pair of combinations, its mutual information value is calculated. This mutual information value quantifies the degree of dependence between the two feature sets. The larger the mutual information value, the higher the association strength. The calculated mutual information values ​​are arranged according to the feature dimensions to form a local association coefficient matrix within the grid with a dimension of (D total features × D trajectory parameters), where D total features is the total number of dimensions of the three types of feature vectors, and D trajectory parameters is the number of parameter dimensions of the historical operation trajectory data (such as the three dimensions of curvature amplitude, elevation compensation, and offset avoidance).

[0042] Step S11103: Based on the local correlation coefficients of all grids, construct a grid adjacency graph, where nodes in the grid adjacency graph represent grids and the weights of edges reflect the similarity of signal features between adjacent grids; use a graph propagation algorithm or a spatial interpolation algorithm to process the local correlation coefficients and generate a global correlation feature set.

[0043] An undirected grid adjacency graph is constructed with each signal acquisition grid as a node. An edge connects two nodes if and only if the corresponding grids are physically adjacent (e.g., sharing an edge or a vertex). The edge weight is obtained by calculating the cosine similarity of the three types of feature vectors of adjacent grids; the higher the similarity, the larger the weight. Based on this adjacency graph, a graph propagation algorithm (such as label propagation) is used to spatially diffuse the local correlation coefficients: the correlation coefficient of each grid is passed to adjacent grids according to the weight of the adjacent edges. After multiple iterations, the correlation coefficient of each grid contains the spatial correlation information of the surrounding grids. For non-adjacent grids, a Kriging spatial interpolation algorithm is used to fill in the spatial correlation gaps, ultimately forming a global correlation feature set covering all grids. This global correlation feature set contains the enhanced correlation coefficient of each grid after considering spatial correlation.

[0044] Step S11104: Based on the global correlation feature set, construct an initial correlation matrix framework. The row dimension of the initial correlation matrix framework corresponds to various feature vectors and historical operation trajectory data, and the column dimension corresponds to each signal acquisition grid.

[0045] The initial correlation matrix framework is designed as a two-dimensional matrix structure. The row dimension is divided into four sub-blocks based on data type: pasture growth status feature sub-block (containing D1 dimensions), terrain interaction feedback feature sub-block (containing D2 dimensions), operation trajectory residue feature sub-block (containing D3 dimensions), and historical operation trajectory data sub-block (containing D4 dimensions), for a total of D1+D2+D3+D4 rows. The column dimension corresponds to all signal acquisition grids, arranged in grid number order, with a total of M×N columns (M rows and N columns of grid). Storage space is reserved for each element position in the matrix framework to fill the correlation strength data obtained in subsequent calculations.

[0046] Step S11105: Fill the initial association matrix frame with the association strength data in the global association feature set, standardize the filled matrix, remove redundant coefficients from the standardized matrix, and generate an initial association coefficient matrix. Each element of the initial association coefficient matrix corresponds to the association strength between a certain type of feature and trajectory data in the relevant grid.

[0047] The enhanced correlation coefficients of each grid in the global correlation feature set are filled into the initial correlation matrix framework according to the row-dimensional correspondence. After filling, the matrix is ​​column-normalized: Z-Score normalization is applied to each column of data (i.e., the correlation strength between all features and trajectory data in each grid) to make the mean of each column data 0 and the standard deviation 1, eliminating the dimensional differences between different grids. Then, redundant coefficients are removed. The variance inflation factor (VIF) method is used to detect and remove multicollinear features. When the VIF value of a feature dimension is greater than a preset threshold (e.g., 10), it is deleted from the matrix. The final initial correlation coefficient matrix has the dimension of (D effective features + D trajectory parameters) × (M × N), where D effective features is the total dimension of the features after removing redundancy, and the matrix element values ​​are the normalized correlation strength between the corresponding features and trajectory data in the relevant grid.

[0048] Step S1111: Input the initial correlation coefficient matrix into the trajectory mapping layer, and construct the initial trajectory residual dynamic linkage model by combining the trajectory evolution algorithm.

[0049] After receiving the initial correlation coefficient matrix, the trajectory mapping layer uses it as the initial weight parameters for the Long Short-Term Memory (LSTM) network. The trajectory evolution algorithm sets the network's time step and number of memory units based on the time-series characteristics of historical trajectory data. During model training, the correlation coefficient matrix and network weights are adjusted using backpropagation to minimize the error between the model's predicted trajectory parameters and the actual values ​​in the historical data. After training, the model parameters are saved, forming the initial trajectory residual dynamic linkage model. This initial trajectory residual dynamic linkage model can receive new ternary signal feature vectors and output the corresponding trajectory parameter prediction results.

[0050] Step S1112: Collect the verification ternary signal group and the corresponding actual operation trajectory data under different operation scenarios. Input the verification ternary signal group into the initial trajectory residual dynamic linkage model, obtain the predicted operation trajectory data output by the initial trajectory residual dynamic linkage model, compare the predicted operation trajectory data with the actual operation trajectory data, extract the difference value between the two, and adjust the correlation coefficient matrix of the feature correlation operation layer and the trajectory evolution algorithm parameters of the trajectory mapping layer based on the difference value to form the final trajectory residual dynamic linkage model.

[0051] Operational scenarios with different seasons and terrain conditions than the training data were selected as the validation set. Triple signal sets and corresponding actual operational trajectory data were collected for these scenarios. The validation triple signal sets were input into the initial model to obtain predicted operational trajectory data. The mean square error between the predicted trajectory and the actual trajectory was calculated in three dimensions: curvature amplitude, elevation compensation, and offset avoidance. The difference value was extracted. If the difference value exceeded a preset threshold, the correlation coefficient matrix element values ​​of the feature association operation layer and the long short-term memory network parameters (such as forget gate and input gate weights) of the trajectory mapping layer were adjusted according to the error backpropagation path. The above validation and adjustment process was repeated until the prediction error of the model on the validation set met the preset requirements. The model obtained at this point is the final trajectory residual dynamic linkage model.

[0052] Step S120: Based on the trajectory residual dynamic linkage model, generate trajectory evolution rules for unmanned driving equipment through signal feature matching. The trajectory evolution rules include trajectory curvature adjustment method corresponding to pasture growth status signal, trajectory elevation compensation method corresponding to terrain interaction feedback signal, and trajectory offset avoidance method corresponding to operation trajectory residual signal.

[0053] In the aforementioned grassland and pasture operation scenario, the constructed trajectory residue dynamic linkage model includes the correlation between multi-source signals and the operation trajectory. To transform the model output into executable equipment control commands, trajectory evolution rules need to be generated. The generation process of these rules is based on the correlation coefficient matrix and trajectory mapping logic in the model. By matching signal features with trajectory adjustment parameters, the specific adjustment methods of trajectory parameters under different signal features are clarified. The trajectory evolution rules need to cover the trajectory adjustment dimensions corresponding to three types of signals: pasture growth status, terrain interaction feedback, and operation trajectory residue, forming a structured rule set.

[0054] Step S121: Analyze the correlation coefficient matrix and trajectory evolution algorithm parameters of the trajectory mapping layer in the trajectory residual dynamic linkage model, and extract the coefficient distribution corresponding to the pasture growth state feature vector, the coefficient distribution corresponding to the terrain interaction feedback feature vector, and the coefficient distribution corresponding to the operation trajectory residual feature vector.

[0055] Using a model parameter parsing tool, the correlation coefficient matrix of the feature association operation layer and the long short-term memory network parameters of the trajectory mapping layer in the trajectory residue dynamic linkage model were read. For the correlation coefficient matrix, the coefficient columns of each dimension of the corresponding pasture growth state feature vector were separated by row index, and the mean, standard deviation, and distribution entropy of each column coefficient were calculated to form the coefficient distribution corresponding to the pasture growth state feature vector. Using the same method, the coefficient distributions corresponding to the terrain interaction feedback feature vector and the operation trajectory residue feature vector were extracted. At the same time, the network weight parameters related to the processing of the three types of feature vectors in the trajectory mapping layer, such as the weight matrices of the input gate and output gate, were analyzed to further verify the accuracy of the coefficient distribution. The extracted coefficient distribution data was stored as a structured file as the basis for subsequent rule generation.

[0056] Step S122: Obtain the trajectory adjustment structure parameters of the unmanned driving device, which include the bending adjustment range of the steering mechanism, the lifting stroke range of the elevation compensation mechanism, and the lateral movement range of the trajectory offset mechanism.

[0057] By consulting the technical manual and hardware parameter tables of the autonomous driving equipment, the physical parameters of the trajectory adjustment mechanisms were obtained. The bending adjustment range of the steering mechanism was obtained by measuring the maximum rotation angle of the steering joint, including the left and right bending angle ranges. The lifting stroke range of the elevation compensation mechanism was determined by reading the maximum lifting height and minimum lowering height of the hydraulic lifting system. The lateral movement range of the trajectory offset mechanism was obtained by measuring the maximum lateral displacement distance of the guide wheels. These parameters were then converted into standardized numerical ranges, such as converting angle ranges to radians and length ranges to meters, and stored as a configuration file for the equipment's structural parameters.

[0058] Step S123: Based on the plant distribution signal features in the feature vector of pasture growth status, extract the plant distribution feature values ​​related to trajectory curvature; perform matching calculations between the plant distribution feature values ​​and the curvature adjustment range of the steering mechanism to generate trajectory curvature amplitude adjustment values ​​corresponding to different plant distributions.

[0059] Relevant dimensions of plant distribution signal features, such as plant center of gravity shift and distribution density gradient, are extracted from the feature vector of pasture growth status. These dimension values ​​are normalized to fall within the [0,1] interval. The normalized feature values ​​are then linearly mapped to the bending adjustment range of the steering mechanism; for example, a feature value of 0 corresponds to the minimum bending angle, and a feature value of 1 corresponds to the maximum bending angle. Through this matching operation, trajectory bending amplitude adjustment values ​​corresponding to different plant distribution feature values ​​are generated. For example, when the plant distribution center of gravity shifts significantly to the left, the corresponding bending amplitude adjustment value is the larger leftward bending angle.

[0060] Step S1231: Extract plant distribution signal features from the feature vector of pasture growth status, wherein the plant distribution signal features include distribution density gradient data and distribution uniformity data.

[0061] Dimensions related to plant distribution are selected from the feature vector of pasture growth status. Distribution density gradient data is obtained by calculating the gradient change rate of plant pixels in the X and Y axes within the grid, reflecting the trend of plant density variation in the horizontal direction. Distribution evenness data is obtained by calculating the spatial distribution entropy value of plant pixels; the higher the entropy value, the more uneven the distribution. These two types of data are extracted from the feature vector to form independent distribution density gradient sequences and distribution evenness sequences.

[0062] Step S1232: Obtain the bending adjustment range data of the steering mechanism, wherein the bending adjustment range data includes the minimum bending angle and the maximum bending angle.

[0063] The bending adjustment range data of the steering mechanism is read from the equipment structure parameter configuration file. The minimum bending angle is the minimum angle value of the equipment steering joint rotating left or right, and the maximum bending angle is the corresponding maximum rotation angle value. For example, the minimum bending angle to the left is -θ_max, and the maximum bending angle to the left is -θ_min (where θ_max>θ_min>0), and the same applies to the right. These angle values ​​are converted to radians and stored as bending adjustment range parameters.

[0064] Step S1233: Based on the distribution density gradient data, extract the distribution density feature value related to the trajectory curvature; perform correlation calculation between the distribution density feature value and the curvature adjustment range data of the steering mechanism to determine the adjustment ratio of the curvature angle.

[0065] Feature extraction is performed on the distribution density gradient data, calculating the maximum value of the gradient magnitude and the direction angle, with the gradient direction angle used as the distribution density feature value. Based on the gradient direction angle's range (0-360 degrees), it is divided into multiple intervals, each corresponding to a bending direction (left or right) and an adjustment ratio. For example, when the gradient direction angle is in the 90-180 degree range, a leftward bending requirement is determined, and the adjustment ratio is the ratio of the gradient magnitude to the maximum gradient magnitude. This adjustment ratio is multiplied by the bending adjustment range of the steering mechanism to obtain the bending angle adjustment component based on the distribution density gradient.

[0066] Step S1234: Based on the distribution uniformity data, extract the distribution uniformity feature value related to the trajectory curvature; perform correlation calculation between the distribution uniformity feature value and the curvature adjustment range data of the steering mechanism to determine the correction ratio of the curvature angle.

[0067] The uniformity characteristic value of the distribution uniformity data is obtained by calculating the difference between the distribution entropy value and the ideal uniform distribution entropy value. The larger the difference, the more uneven the distribution. This difference is then normalized to obtain the uniformity characteristic value. Different correction ratios are set according to the magnitude of the characteristic value; the larger the characteristic value (the more uneven the distribution), the higher the correction ratio. The correction ratio is multiplied by the bending angle adjustment component based on the distribution density gradient to obtain the corrected bending angle adjustment ratio.

[0068] Step S1235: Calculate the initial trajectory bending amplitude adjustment value by combining the adjustment ratio and correction ratio of the bending angle.

[0069] Add the adjustment ratio of the bending angle to the correction ratio to obtain the comprehensive adjustment coefficient. Multiply this coefficient by the bending adjustment range of the steering mechanism (the difference between the maximum bending angle and the minimum bending angle), and add the minimum bending angle to obtain the initial trajectory bending amplitude adjustment value. For example, the initial adjustment value = minimum bending angle + comprehensive adjustment coefficient × (maximum bending angle - minimum bending angle).

[0070] Step S1236: Match the initial trajectory curvature adjustment value with the correlation coefficient matrix in the trajectory residual dynamic linkage model to obtain the correlation coefficient corresponding to the plant distribution signal characteristics.

[0071] In the correlation coefficient matrix, find the coefficient columns corresponding to each dimension of the plant distribution signal features, and calculate the weighted average of these coefficients as the correlation coefficients corresponding to the plant distribution signal features. The weight values ​​are normalized values ​​of the feature values ​​of each dimension to ensure that the correlation coefficients can reflect the overall correlation strength of the feature vectors.

[0072] Step S1237: Perform a weighted calculation on the initial trajectory curvature adjustment value based on the correlation coefficient to generate the intermediate trajectory curvature adjustment value.

[0073] The initial trajectory curvature adjustment value is multiplied by the correlation coefficient to obtain the weighted intermediate trajectory curvature adjustment value. The larger the correlation coefficient, the stronger the influence of plant distribution signal characteristics on trajectory curvature, and the higher the weight of the adjustment value.

[0074] Step S1238: Compare the intermediate trajectory curvature adjustment value with the trajectory curvature record corresponding to the plant distribution in the historical operation trajectory data, and extract the difference ratio after comparison.

[0075] Samples with similar plant distribution characteristics to the historical operational trajectory data are selected, and the trajectory curvature amplitude records of these samples are extracted. The absolute difference between the intermediate trajectory curvature amplitude adjustment value and the mean of these records is calculated, and then the difference is divided by the mean of the records to obtain the difference ratio.

[0076] Step S1239: Based on the difference ratio, the intermediate trajectory curvature adjustment value is corrected a second time to generate the final trajectory curvature adjustment value corresponding to different plant distributions.

[0077] If the difference ratio is positive, it indicates that the intermediate adjustment value is greater than the historical average, and it needs to be adjusted downward by multiplying the difference ratio by the correction factor; if it is negative, it needs to be adjusted upward. The correction factor is dynamically adjusted according to the model's prediction error, ultimately yielding the adjustment value for the trajectory curvature amplitude corresponding to different plant distributions.

[0078] Step S12310: Store the final trajectory curvature adjustment values ​​according to the classification of plant distribution signal characteristics, forming a correspondence table of plant distribution-trajectory curvature adjustment values.

[0079] Plant distribution signal characteristics are classified according to a combination of distribution density gradient and distribution uniformity, with each category corresponding to a unique feature identifier. The final trajectory curvature adjustment value is associated with the feature identifier and stored to form a lookup table structure for easy real-time querying and matching.

[0080] Step S124: Based on the stem toughness signal features in the feature vector of pasture growth state, extract the stem toughness feature values ​​related to trajectory bending; perform matching calculations between the stem toughness feature values ​​and the bending adjustment range of the steering mechanism to generate trajectory bending amplitude correction values ​​corresponding to different stem toughnesses.

[0081] Relevant dimensions of stem toughness signal features, such as the peak mean and slope of the resistance curve, are extracted from the feature vector of pasture growth status. These dimensional values ​​are normalized and then nonlinearly mapped to the bending adjustment range of the steering mechanism. For example, an S-shaped function is used to map the stem toughness feature value to a bending amplitude correction coefficient; the higher the stem toughness (the larger the peak resistance), the smaller the correction coefficient, to avoid oversteering due to excessive stem stiffness. The correction coefficient is multiplied by the bending amplitude adjustment value based on plant distribution to obtain a trajectory bending amplitude correction value that incorporates the stem toughness factor.

[0082] Step S125: Based on the contact reaction signal features in the terrain interaction feedback feature vector, extract the contact reaction feature values ​​related to the trajectory elevation; perform matching calculations between the contact reaction feature values ​​and the lifting stroke range of the elevation compensation mechanism to generate trajectory elevation compensation values ​​corresponding to different contact reactions.

[0083] The peak value and mean value of the contact reaction force signal are extracted from the terrain interaction feedback feature vector to reflect the surface hardness. These feature values ​​are normalized and mapped to the lifting range of the elevation compensation mechanism. A larger contact reaction force indicates a harder surface and a smaller elevation compensation value (the equipment can be lowered); a smaller contact reaction force indicates a softer surface and a larger elevation compensation value (the equipment needs to be raised). The trajectory elevation compensation value corresponding to different contact reaction force feature values ​​is calculated using a linear interpolation method.

[0084] Step S126: Based on the deformation rebound signal features in the terrain interactive feedback feature vector, extract the deformation rebound feature values ​​related to the trajectory elevation; perform matching calculations between the deformation rebound feature values ​​and the lifting stroke range of the elevation compensation mechanism to generate trajectory elevation compensation correction values ​​corresponding to different deformation rebounds.

[0085] The deformation rebound signal characteristics are represented by recovery time and rebound displacement. A shorter recovery time and a larger rebound displacement indicate better surface elasticity. These feature values ​​are extracted, normalized, and then mapped to the lifting range of the elevation compensation mechanism. Better surface elasticity results in a larger elevation compensation correction value to prevent equipment from shaking due to surface rebound; conversely, a smaller correction value is applied to weaker surface elasticity. This correction value is then added to the elevation compensation value based on contact reaction force to obtain the final trajectory elevation compensation value.

[0086] Step S127: Based on the surface compaction trace signal features in the residual feature vector of the operation trajectory, extract the surface compaction feature values ​​related to the trajectory offset; perform matching calculations between the surface compaction feature values ​​and the lateral movement range of the trajectory offset mechanism to generate trajectory offset avoidance values ​​corresponding to different surface compaction traces.

[0087] The characteristics of surface compaction traces are represented by the rate of increase in compacted area and the complexity of its shape. A faster increase in compacted area and a more complex shape indicate that the area has been compacted multiple times and requires offsetting. These feature values ​​are extracted, normalized, and then mapped to the lateral movement range of the trajectory offset mechanism. Larger feature values ​​indicate a greater offset, meaning the equipment moves away from the compacted area. The lateral direction of the offset (left or right) is determined by calculating the centroid offset direction of the compacted area.

[0088] Step S128: Based on the residual grass distribution signal features in the residual feature vector of the operation trajectory, extract the residual distribution feature values ​​related to the trajectory offset; perform matching calculations between the residual distribution feature values ​​and the lateral movement range of the trajectory offset mechanism to generate trajectory offset avoidance correction values ​​corresponding to different residual grass distributions.

[0089] The distribution characteristics of residual forage are characterized by the proportion of residual area and the uniformity of distribution. A larger proportion of residual area and a more uneven distribution indicate poor harvesting in that area, requiring re-harvesting by offsetting the area. After extracting and normalizing these feature values, they are mapped to the lateral movement range of the trajectory offset mechanism. The larger the feature value, the larger the offset correction value. This correction value is added to the offset value based on surface compaction marks to obtain the final trajectory offset correction value.

[0090] Step S129: Integrate the trajectory curvature adjustment value, trajectory curvature correction value, trajectory elevation compensation value, trajectory elevation compensation correction value, trajectory deviation avoidance value, and trajectory deviation avoidance correction value to form a basic trajectory adjustment parameter set.

[0091] The generated trajectory adjustment values ​​are categorized and integrated according to three dimensions: trajectory curvature, elevation compensation, and offset avoidance. Each dimension includes a basic adjustment value and a correction value, which are combined according to a set priority. For example, the final parameter for the trajectory curvature dimension is the trajectory curvature amplitude adjustment value plus the trajectory curvature amplitude correction value; the same applies to the elevation compensation and offset avoidance dimensions. The combined parameters are stored in a basic trajectory adjustment parameter set, where each parameter includes its magnitude, direction of action, and activation conditions.

[0092] Step S1210: Based on the correlation coefficient matrix in the trajectory residual dynamic linkage model, assign evolution weights to each parameter in the basic trajectory adjustment parameter set, and logically sort the parameters in the basic trajectory adjustment parameter set according to the evolution weight order to generate a parameter execution order table. The evolution weights are consistent with the distribution of the corresponding feature vector coefficients. The parameter execution order table includes the order of action and numerical superposition of different parameters.

[0093] From the correlation coefficient matrix of the trajectory residual dynamic linkage model, the mean value of the eigenvector coefficient distribution corresponding to each basic trajectory adjustment parameter is extracted as the baseline value of the evolution weight. The baseline value is normalized to obtain the evolution weight of each parameter; a higher weight value indicates a more significant impact of the parameter on trajectory adjustment. The parameters in the basic trajectory adjustment parameter set are sorted according to their evolution weights from high to low. Simultaneously, the dependencies between parameters are considered; for example, the trajectory elevation compensation parameter must be executed after the trajectory curvature parameter to avoid interference with the mechanism's motion. Based on the sorting results and dependencies, a parameter execution order table is generated, specifying the execution order, numerical superposition method (e.g., addition, multiplication), and effective threshold of each parameter.

[0094] Step S12101: Analyze the correlation coefficient matrix in the trajectory residual dynamic linkage model, and extract the coefficient distribution data of the feature vector corresponding to each parameter in the basic trajectory adjustment parameter set. The coefficient distribution data reflects the correlation strength between the feature corresponding to each parameter and the operation trajectory.

[0095] Using model parameter analysis tools, we locate the coefficient columns in the correlation coefficient matrix that are related to each parameter in the set of basic trajectory adjustment parameters. For example, the trajectory curvature amplitude adjustment value corresponds to the plant distribution characteristic coefficient column in the pasture growth status feature vector. We extract all elements from these coefficient columns as the coefficient distribution data for the feature vector corresponding to that parameter. The statistical characteristics of the coefficient distribution data (such as mean and variance) directly reflect the correlation strength between the feature and the operational trajectory.

[0096] Step S12102: Calculate the mean correlation strength of each parameter based on the coefficient distribution data, and determine the evolution weight benchmark value of each parameter by combining the fluctuation of the coefficient distribution. The larger the mean correlation strength and the smaller the fluctuation, the higher the evolution weight benchmark value.

[0097] For each parameter, the arithmetic mean of the coefficient distribution data is calculated as the mean of the association strength, and the standard deviation is calculated as an indicator of the degree of volatility. The formula for calculating the evolution weight benchmark value is: Benchmark value = Mean of association strength / (1 + Volatility index). According to this formula, the parameter with a larger mean of association strength and a smaller degree of volatility has a higher evolution weight benchmark value, indicating that the association relationship of the parameter is more stable and the influence is stronger.

[0098] Step S12103: Normalize the evolution weight baseline value to obtain the final evolution weight of each parameter. The final evolution weight is positively correlated with the correlation strength of the corresponding parameter.

[0099] The baseline values ​​of the evolution weights of all parameters are summed, and then the baseline value of each parameter is divided by the sum to obtain the normalized final evolution weights. After normalization, the sum of the final evolution weights of all parameters is 1, and the weight value of each parameter is positively correlated with the mean of its correlation strength and negatively correlated with the degree of volatility.

[0100] Step S12104: Extract the function attributes of each parameter in the basic trajectory adjustment parameter set. The function attributes are divided into trajectory curvature adjustment, trajectory elevation compensation and trajectory offset avoidance. Parameters with the same function attribute are grouped into the same parameter group.

[0101] Based on their applicable objects and adjustment dimensions, the parameters in the basic trajectory adjustment parameter set are divided into three categories of attributes. Trajectory curvature adjustment parameters include trajectory curvature amplitude adjustment values ​​and correction values; trajectory elevation compensation parameters include trajectory elevation compensation values ​​and correction values; and trajectory offset avoidance parameters include trajectory offset avoidance values ​​and correction values. Parameters with the same attribute are grouped together for easier subsequent intra-group sorting and inter-group coordination.

[0102] Step S12105: Within each parameter group, perform preliminary sorting of parameters in descending order of final evolution weights to generate parameter sorting sequences within each parameter group, and record the mutual influence relationships between different parameter groups.

[0103] Within the trajectory curvature adjustment group, the final evolution weights of the trajectory curvature amplitude adjustment and correction values ​​are compared and sorted from highest to lowest; the same applies to the elevation compensation group and the offset avoidance group. Simultaneously, by analyzing the order of parameter adjustments in historical operational data and the equipment kinematic model, the interrelationships between parameter groups are determined. For example, trajectory curvature adjustment must be performed before trajectory offset avoidance to ensure that lateral offset is performed only after the steering action is completed.

[0104] Step S12106: Based on the mutual influence relationship between parameter groups, perform overall logical adjustment on the parameter sorting sequence within each parameter group to form a preliminary parameter execution order.

[0105] Based on the interrelationships between parameter groups, the execution order of the parameter groups is determined. For example, the trajectory curvature adjustment group is executed first, followed by the trajectory elevation compensation group, and finally the trajectory offset avoidance group. Within each group, parameters are arranged according to a preliminary sorting sequence, thus forming the overall preliminary parameter execution order.

[0106] Step S12107: Based on the preliminary parameter execution order, drive the unmanned vehicle model to perform trajectory adjustment in the trajectory simulation environment; if the simulation results show that the device kinematics is infeasible, the trajectory change exceeds the set threshold, or there is a collision with an obstacle, it is determined that there is an execution order conflict, and the preliminary parameter execution order is adjusted according to the conflict type by calling the preset sorting adjustment rules; repeat the simulation and adjustment process until the simulation results meet the preset feasibility conditions and generate a parameter execution order table.

[0107] A trajectory simulation environment is constructed, comprising a dynamic model of the unmanned vehicle and a model of the operating area environment. Preliminary parameter execution order is input into the simulation environment, driving the vehicle model to perform trajectory adjustments sequentially. Simulation monitoring is used to check whether the vehicle's kinematic parameters (such as velocity, acceleration, and joint angles) are within physical constraints, whether the trajectory curvature change exceeds a set abrupt change threshold, and whether collisions occur with virtual obstacles. If conflicts occur, preset rules are invoked to adjust the parameter execution order based on the conflict type, such as swapping the order of conflicting parameters or inserting delayed execution instructions. Simulation and adjustment are repeated until all conflicts are resolved and the simulation results meet feasibility requirements. The parameter execution order at this point is the final parameter execution order table.

[0108] Step S1211: Combine the parameter execution order table and the basic trajectory adjustment parameter set to form a trajectory evolution rule. The trajectory evolution rule includes the function logic, numerical calculation method and execution order requirements of all parameters.

[0109] The parameter execution order table is linked and integrated with the basic trajectory adjustment parameter set to clarify the position of each parameter in the execution order, its corresponding function logic (such as addition correction, multiplication correction), numerical calculation method (such as mapping formula based on feature values), and execution order requirements (such as pre-parameters, post-parameters). This information is then organized into a structured rule document containing fields such as rule ID, parameter type, target object, calculation logic, execution order, and effective conditions, forming a complete trajectory evolution rule.

[0110] Step S130: Collect real-time signals of pasture growth status, real-time feedback signals of terrain interaction, real-time residual signals of operation trajectory, and current operation trajectory data of the equipment through the all-domain perception component of the unmanned driving equipment during the operation process.

[0111] In the aforementioned grassland and pasture operation scenario, the unmanned vehicle is equipped with a global perception component to collect various dynamic signals and trajectory data in real time during the operation. This data forms the basis for subsequent trajectory adaptation calculations and dynamic adjustments, and the real-time performance, accuracy, and completeness of the data collection must be ensured. The global perception component comprises multiple sub-modules, each responsible for the collection and preprocessing of different types of signals, and achieves spatiotemporal alignment of the data through a unified time synchronization mechanism.

[0112] Step S131: Configure a real-time forage growth sensing module for the unmanned vehicle. This real-time forage growth sensing module maintains the same standard as the forage growth status signal acquisition in the trajectory residue dynamic linkage model, and captures the real-time signals of plant distribution, stem toughness, and leaf density at the current location of the vehicle operation.

[0113] The real-time forage growth sensing module is installed at the front end of the device, integrating a multispectral camera, a portable stem mechanics sensor, and a laser leaf density meter. The multispectral camera's shooting parameters (such as focal length and exposure time) are consistent with the acquisition standards used during model training, ensuring image signal consistency. The stem mechanics sensor contacts the forage via a robotic arm to collect puncture resistance signals in real time. The laser leaf density meter acquires leaf density data within a certain range around the device through rotational scanning. The module has a built-in preprocessing unit that filters and normalizes the raw signals, outputting real-time signals of plant distribution, stem toughness, and leaf density that match the model's input standards.

[0114] Step S132: Configure a terrain interaction real-time detection module for the unmanned vehicle. This terrain interaction real-time detection module maintains the same standard as the terrain interaction feedback signal acquisition in the trajectory residual dynamic linkage model, and captures the real-time contact reaction force signal, deformation rebound signal and particle friction signal at the current position of the vehicle operation.

[0115] The terrain interaction real-time detection module is installed on the device's walking mechanism and includes pressure sensors, displacement sensors, and friction coefficient sensors. The pressure sensors are installed at the wheel hubs to detect the normal reaction force of the ground on the wheels in real time; the displacement sensors measure the compression of the suspension system, indirectly reflecting surface deformation; and the friction coefficient sensors calculate the ground friction coefficient by detecting the slip rate of the drive wheels. The module's sampling frequency is consistent with the acquisition standard used during model training. After amplification and filtering, the raw signals output real-time signals of contact reaction force, deformation rebound, and particle friction.

[0116] Step S133: Configure a real-time trajectory residue monitoring module for the unmanned vehicle. This real-time trajectory residue monitoring module is consistent with the operation trajectory residue signal acquisition standard in the trajectory residue dynamic linkage model, and captures the real-time signals of ground compaction marks, pasture residue distribution, and trajectory path imprints at the current location after the equipment operation.

[0117] The real-time trajectory residue monitoring module consists of a high-definition camera and a laser profilometer installed at the rear of the equipment. The high-definition camera captures images of the ground surface after equipment operation, with image resolution consistent with the standard used during model training. The laser profilometer scans the three-dimensional contour of the ground surface to obtain depth information of the compaction marks. The module uses image recognition algorithms to extract the area and morphological features of the compaction marks in real time, calculates the uniformity of pasture residue distribution through contour analysis, and extracts the depth fluctuations of the trajectory path imprints through depth data. The output real-time signal is consistent with the feature extraction standard used during model training.

[0118] Step S134: Configure a real-time trajectory monitoring module for the unmanned vehicle. This real-time trajectory monitoring module is associated with the vehicle's steering mechanism, elevation compensation mechanism, and trajectory deviation mechanism to capture the vehicle's current trajectory curvature data, trajectory elevation compensation data, and trajectory deviation avoidance data.

[0119] The real-time trajectory monitoring module, through its interface with the equipment control system, directly reads the rotation angle of the steering mechanism, the extension and retraction of the hydraulic rod of the elevation compensation mechanism, and the lateral displacement of the trajectory deviation mechanism. It converts these physical quantities into data for trajectory bending amplitude (angle value), trajectory elevation compensation (height value), and trajectory deviation avoidance (lateral distance value). The module's sampling frequency is higher than the equipment's control frequency, ensuring that it can capture the dynamic changes in trajectory parameters.

[0120] Step S135: Start the real-time forage growth sensing module according to the set real-time data acquisition cycle, continuously collect and record real-time signals of plant distribution, stem toughness and leaf density, and form a real-time signal sequence of forage growth status.

[0121] The real-time data acquisition cycle is set as a function of the equipment's operating speed and the grid size to ensure that the equipment completes at least one signal acquisition within each grid. After the real-time pasture growth sensing module is activated, the module continues to operate according to the acquisition cycle. The real-time signals of plant distribution, stem toughness, and leaf density collected each time are arranged in chronological order to form a timestamped signal sequence. Each data point in the sequence includes information such as signal value, acquisition time, and equipment location coordinates.

[0122] Step S136: Following the same real-time data acquisition cycle, activate the terrain interaction real-time detection module to continuously acquire and record real-time signals of contact reaction force, deformation rebound, and particle friction, forming a terrain interaction real-time feedback signal sequence; and activate the trajectory residue real-time monitoring module to continuously acquire and record real-time signals of surface compaction marks, pasture residue distribution, and trajectory path imprints, forming a real-time residual signal sequence of the operating trajectory; and activate the trajectory real-time monitoring module to continuously acquire and record trajectory curvature amplitude data, trajectory elevation compensation data, and trajectory offset avoidance data, forming a data sequence of the equipment's current operating trajectory.

[0123] Using the same data collection cycle as the real-time pasture growth sensing module, the terrain interaction real-time detection module, trajectory residue real-time monitoring module, and trajectory real-time monitoring module are activated. Each module collects data periodically, forming a terrain interaction real-time feedback signal sequence, a work trajectory real-time residue signal sequence, and a current work trajectory data sequence, respectively. The timestamps of all sequences are generated based on the device's unified clock system to ensure time synchronization accuracy.

[0124] Step S137: Extract the real-time signal of pasture growth status, the real-time feedback signal of terrain interaction, the real-time residual signal of operation trajectory and the current operation trajectory data of the equipment at the same collection time point, assign a unique timestamp identifier to the collection time point, and associate the data with timestamp identifiers to form a real-time data combination unit.

[0125] Data from the same acquisition time point is extracted from four signal sequences using timestamp matching. A unique timestamp identifier is generated for this time point, in the format of "Device ID-Acquisition Cycle Number-Timestamp". The matched real-time pasture growth status signal, real-time terrain interaction feedback signal, real-time residual signal of the operation trajectory, and the current operation trajectory data of the equipment are packaged together and timestamp identifiers are added to form a real-time data combination unit. Each unit contains the current sampled value of the four types of signals and the corresponding equipment location coordinates.

[0126] Step S138: Construct a real-time data storage queue, store real-time data combination units from the most recent multiple acquisition cycles in chronological order of acquisition time, and synchronize the real-time data combination units in the real-time data storage queue to the global data interaction node of the work area through the device's wireless transmission module.

[0127] A fixed-length real-time data storage queue is built locally on the device, with the queue length set to store combinations of real-time data units from the most recent N acquisition cycles. When a new unit is generated, it is directly enqueued if the queue is not full; otherwise, the oldest unit is removed and enqueued. Simultaneously, the device's wireless transmission module sends the data units in the queue in batches to the global data interaction nodes according to a set synchronization period (e.g., every two acquisition cycles). An encryption protocol is used during transmission to ensure data security.

[0128] Step S139: Perform format unification processing on the real-time data combination units synchronized to the global data interaction nodes to generate standardized real-time data combination units.

[0129] The global data interaction nodes perform format verification and standardization on the received real-time data combination units. Verification includes data integrity, timestamp validity, and the reasonableness of signal value ranges. Standardization converts various signal values ​​into numerical ranges consistent with the input requirements of the trajectory residual dynamic linkage model (e.g., normalized to the [0,1] interval) and unifies the data format (e.g., JSON format). The processed standardized real-time data combination units are stored in the node's database for subsequent trajectory adaptation calculations.

[0130] Step S140: Input the real-time signal of pasture growth status, the real-time feedback signal of terrain interaction, the real-time residual signal of operation trajectory and the current operation trajectory data of the equipment into the trajectory residual dynamic linkage model, perform trajectory adaptation calculation, and generate trajectory deviation correction parameters.

[0131] In the aforementioned grassland and pasture operation scenario, standardized real-time data combination units are input into the trajectory residual dynamic linkage model. The model calculates the deviation between the current operational trajectory and the theoretically optimal trajectory by comparing real-time signal characteristics with the feature distribution during training, and generates corresponding correction parameters. Trajectory fit calculation is the core component for achieving dynamic navigation, requiring comprehensive consideration of the influence of multiple real-time signals to ensure the accuracy and real-time performance of the correction parameters.

[0132] Step S141: Receive standardized real-time data combination unit through the input layer of the trajectory residue dynamic linkage model, and extract real-time signals of pasture growth status, real-time feedback signals of terrain interaction, real-time residual signals of operation trajectory, and current operation trajectory data of equipment from the standardized real-time data combination unit.

[0133] The input layer of the trajectory residue dynamic linkage model contains four parallel data receiving channels, each receiving four types of data from a standardized real-time data combination unit. The input layer parses the received data, extracting real-time signals of pasture growth status (plant distribution, stem toughness, leaf density), real-time terrain interaction feedback signals (contact reaction force, deformation rebound, particle friction), real-time residual signals of the operational trajectory (surface compaction marks, pasture residue distribution, trajectory path imprints), and current equipment operational trajectory data (bending amplitude, elevation compensation, offset avoidance). The parsed data is then reorganized according to feature dimensions into a tensor format acceptable to the model.

[0134] Step S142: Extract features from the real-time signal of pasture growth status to generate a real-time feature vector of pasture growth status; input the real-time feature vector of pasture growth status into the feature association operation layer of the trajectory residual dynamic linkage model, and compare it with the pasture growth status feature vector stored in the feature association operation layer to generate a pasture growth status signal deviation vector.

[0135] The same feature extraction method used in the model training phase is employed to process the real-time signal of forage growth status. For example, the centroid offset and distribution entropy are calculated for the real-time signal of plant distribution; the peak and mean resistance values ​​are extracted from the stem toughness signal; and the number of leaves per unit volume and standard deviation are calculated from the leaf density signal. These feature values ​​are arranged in a preset order to generate a real-time feature vector of forage growth status. This vector is then compared dimension-by-dimensionally with the mean of the training set feature vectors stored in the feature association operation layer to obtain a deviation vector of forage growth status signal. Each element in the vector represents the degree of deviation between the real-time feature and the training mean.

[0136] Step S143: Extract features from the real-time terrain interaction feedback signal to generate a real-time terrain interaction feedback feature vector; input the real-time terrain interaction feedback feature vector into the feature association operation layer of the trajectory residual dynamic linkage model, and compare it with the terrain interaction feedback feature vector stored in the feature association operation layer to generate a terrain interaction feedback signal deviation vector.

[0137] Similarly, features are extracted from the real-time terrain interaction feedback signal, such as the peak value and response time of contact reaction force, the recovery time and displacement of deformation rebound, and the mean and variance of particle friction intensity, to form a real-time terrain interaction feedback feature vector. This vector is then compared with the training mean of the terrain interaction feedback feature vector stored in the feature association operation layer, and the difference is calculated dimension by dimension to generate a terrain interaction feedback signal deviation vector.

[0138] Step S144: Extract features from the real-time residual signal of the operation trajectory to generate a real-time residual feature vector of the operation trajectory; input the real-time residual feature vector of the operation trajectory into the feature association operation layer of the trajectory residual dynamic linkage model, and compare it with the residual feature vector of the operation trajectory stored in the feature association operation layer to generate a deviation vector of the residual signal of the operation trajectory.

[0139] Features are extracted from the real-time residual signals of the operation trajectory, including the area growth rate and morphological complexity of surface compaction marks, the uniformity and area proportion of pasture residue distribution, and the depth fluctuation and continuity of trajectory path imprints, forming a real-time residual feature vector of the operation trajectory. This vector is compared with the training mean of the residual feature vector of the operation trajectory stored in the feature association operation layer, and the difference is calculated to generate a deviation vector of the residual signal of the operation trajectory.

[0140] Step S145: Input the current operation trajectory data of the equipment into the trajectory mapping layer of the trajectory residual dynamic linkage model, and combine the pasture growth status signal deviation vector, terrain interaction feedback signal deviation vector and operation trajectory residual signal deviation vector to calculate the theoretical operation trajectory data through the trajectory evolution algorithm of the trajectory mapping layer.

[0141] Step S1451: Receive the current operation trajectory data of the equipment and transmit it to the trajectory mapping layer of the trajectory residue dynamic linkage model. Simultaneously import the pasture growth status signal deviation vector, terrain interaction feedback signal deviation vector and operation trajectory residue signal deviation vector for correlation calibration processing to obtain a set of calibrated deviation vectors.

[0142] The trajectory mapping layer receives the current operating trajectory data of the equipment (bending amplitude, elevation compensation, and offset avoidance) and imports three deviation vectors. The correlation calibration process calculates the covariance matrix between each deviation vector and orthogonals them, eliminating redundant information between features. The calibrated set of deviation vectors contains the three deviation vectors that have undergone decorrelation processing.

[0143] Step S1452: Based on the trajectory evolution algorithm of the trajectory mapping layer, decompose the core constituent dimensions of the current operation trajectory data of the equipment. The core constituent dimensions correspond to trajectory curvature amplitude data, trajectory elevation compensation data and trajectory offset avoidance data, and allocate an independent computing channel for each core constituent dimension.

[0144] The trajectory evolution algorithm decomposes the current operating trajectory data of the equipment into three core dimensions: trajectory curvature amplitude, trajectory elevation compensation, and trajectory offset avoidance. An independent long short-term memory network processing channel is allocated to each dimension. Each channel contains an input gate, a forget gate, an output gate, and a memory unit to process the time-series data for that dimension.

[0145] Step S1453: Input the pasture growth status signal deviation vector from the calibrated deviation vector set into the trajectory curvature amplitude calculation channel, and perform interactive calculation with the trajectory curvature amplitude data in the current operating trajectory data of the equipment to output the theoretical adjustment amount of trajectory curvature amplitude; input the terrain interactive feedback signal deviation vector into the trajectory elevation compensation calculation channel, and perform interactive calculation with the trajectory elevation compensation data in the current operating trajectory data of the equipment to output the theoretical adjustment amount of trajectory elevation compensation; input the residual signal deviation vector of the operating trajectory into the trajectory offset avoidance calculation channel, and perform interactive calculation with the trajectory offset avoidance data in the current operating trajectory data of the equipment to output the theoretical adjustment amount of trajectory offset avoidance.

[0146] In the trajectory curvature amplitude calculation channel, the pasture growth status signal deviation vector and the current trajectory curvature amplitude data interact through element-wise multiplication. The result is input into the Long Short-Term Memory (LSTM) network unit. The network learns the relationship between historical deviations and trajectory adjustments, and outputs the theoretical adjustment amount for the trajectory curvature amplitude. The elevation compensation and offset avoidance channels use the same interactive calculation method, and output the corresponding theoretical adjustment amounts respectively.

[0147] Step S1454: Extract the historical correlation parameters stored in the trajectory mapping layer of the trajectory residual dynamic linkage model. The historical correlation parameters reflect the interaction patterns between various deviation vectors and the various dimensions of the operation trajectory data. Import the historical correlation parameters into each calculation channel and correct the theoretical adjustment amount output by each calculation channel.

[0148] Historical correlation parameters include the weight matrix and bias term between the bias vector and trajectory adjustment amount learned during training. These parameters are imported into the corresponding computation channel and linearly combined with the theoretical adjustment amount, such as adjustment amount = theoretical adjustment amount × weight matrix + bias term. This corrects the theoretical adjustment amount, making it more consistent with historical interaction patterns.

[0149] Step S1455: Integrate the corrected theoretical adjustment amount of trajectory curvature amplitude, theoretical adjustment amount of trajectory elevation compensation, and theoretical adjustment amount of trajectory deviation avoidance, and combine them with the benchmark parameters of the current operation trajectory data of the equipment. Perform fusion calculation through trajectory evolution algorithm to generate preliminary theoretical operation trajectory data.

[0150] The corrected theoretical adjustments in the three dimensions are added to the baseline parameters of the equipment's current operating trajectory data (such as the current bending angle, elevation value, and offset) to obtain preliminary theoretical operating trajectory data. During the fusion process, the normalization layer in the trajectory evolution algorithm ensures that the parameters in each dimension are within the physical constraints.

[0151] Step S1456: Compare the theoretical parameters of each dimension in the preliminary theoretical operation trajectory data with the mechanical motion limit parameters of the unmanned driving equipment; if any theoretical parameter exceeds the corresponding mechanical motion limit, adjust the theoretical parameter according to the preset constraint rules, and feed the adjustment requirements back to the corresponding computing channel for iterative calculation until all theoretical parameters meet the mechanical motion limit constraints, and obtain the final theoretical operation trajectory data.

[0152] The mechanical motion limit parameters include the maximum bending angle of the steering mechanism, the maximum stroke of the elevation compensation mechanism, and the maximum displacement of the trajectory offset mechanism. Each parameter in the preliminary theoretical operating trajectory data is compared with these limit values. If the limit is exceeded, the excess is proportionally allocated to other dimensions or the adjustment amount is reduced, and the result is fed back to the corresponding calculation channel for recalculation. This process is iterated until all parameters are within the limit range, yielding the final theoretical operating trajectory data.

[0153] Step S146: Compare and calculate the theoretical operation trajectory data with the current operation trajectory data of the equipment, and extract the differences in trajectory curvature, trajectory elevation compensation and trajectory deviation avoidance between the two to form a trajectory difference dataset.

[0154] Calculate the differences between the final theoretical operation trajectory data and the current operation trajectory data of the equipment in three dimensions: trajectory curvature difference = theoretical curvature - current curvature; trajectory elevation compensation difference = theoretical elevation compensation - current elevation compensation; trajectory deviation avoidance difference = theoretical deviation avoidance - current deviation avoidance. Combine the above differences to form a trajectory difference dataset.

[0155] Step S147: Based on the correlation coefficient matrix in the trajectory residual dynamic linkage model, assign correction coefficients to the trajectory curvature amplitude difference, trajectory elevation compensation difference, and trajectory offset avoidance difference in the trajectory difference dataset. The correction coefficients are consistent with the corresponding feature vector coefficient distribution.

[0156] The coefficient columns related to trajectory curvature, elevation compensation, and offset avoidance dimensions are extracted from the correlation coefficient matrix, and the mean of each coefficient column is calculated as the correction coefficient for the corresponding difference. For example, the correction coefficient for the difference in trajectory curvature amplitude is the mean of the coefficient column corresponding to the pasture growth state feature vector, ensuring that the correction coefficient is consistent with the feature correlation strength.

[0157] Step S148: Perform weighted calculations on the differences in trajectory curvature amplitude, trajectory elevation compensation, and trajectory deviation avoidance according to the correction coefficients to generate correction values ​​for trajectory curvature amplitude, trajectory elevation compensation, and trajectory deviation avoidance.

[0158] Multiplying the difference in trajectory curvature amplitude by its correction factor yields the corrected value for trajectory curvature amplitude; the same applies to elevation compensation and offset avoidance dimensions. During the weighted calculation, the larger the correction factor, the greater the contribution of the corresponding difference to the corrected value.

[0159] Step S149: Match the trajectory curvature correction value, trajectory elevation compensation correction value, and trajectory offset avoidance correction value with the correspondence table in the trajectory residual dynamic linkage model. Based on the matching result, adjust the values ​​of the trajectory curvature correction value, trajectory elevation compensation correction value, and trajectory offset avoidance correction value to generate the final trajectory deviation correction parameters. The trajectory deviation correction parameters include the final trajectory curvature correction value, the final trajectory elevation compensation correction value, and the final trajectory offset avoidance correction value.

[0160] In the correspondence table of the trajectory residual dynamic linkage model, find the historical record closest to the current correction value and obtain the actual adjustment effect parameter of that record. If the historical adjustment effect parameter indicates that the original correction value was too large or too small, adjust the current correction value proportionally. For example, if a similar correction value in the historical record causes trajectory overshoot, multiply the current correction value by an adjustment coefficient less than 1. After matching and adjustment, the final trajectory deviation correction parameter is obtained.

[0161] Step S150: Based on the trajectory deviation correction parameters and the trajectory evolution rules, generate a dynamic trajectory adjustment command to drive the unmanned driving equipment to update the operation trajectory. At the same time, input the updated equipment operation trajectory data and the newly generated operation trajectory residual signal into the trajectory residual dynamic linkage model to complete the cyclic optimization of the trajectory residual dynamic linkage model.

[0162] In the aforementioned grassland and pasture operation scenario, trajectory deviation correction parameters are combined with trajectory evolution rules to transform into specific equipment control commands, enabling dynamic adjustment of the operation trajectory. Simultaneously, the adjusted trajectory data and newly generated residual trajectory signals are fed back to the model to optimize model parameters, improve the accuracy of subsequent trajectory predictions, and form a closed-loop optimization mechanism.

[0163] For example, step S151: parse the parameter execution order table, basic trajectory adjustment parameter set and evolution weight in the trajectory evolution rule, and extract the function logic and numerical calculation method of different trajectory adjustment parameters.

[0164] The rule parsing module reads the parameter execution order table from the trajectory evolution rules to clarify the execution order of each trajectory adjustment parameter; it extracts the baseline values ​​and correction methods of each parameter in the basic trajectory adjustment parameter set; and it obtains the evolution weights to determine the priority of the parameters. Simultaneously, it parses the action logic (such as addition and multiplication corrections) and numerical calculation methods (such as eigenvalue mapping formulas) defined in the rules.

[0165] Step S152: Extract the final trajectory curvature correction value, the final trajectory elevation compensation correction value, and the final trajectory offset avoidance correction value from the trajectory deviation correction parameters.

[0166] The correction values ​​for three dimensions are separated from the trajectory deviation correction parameters: final trajectory curvature correction value (Δθ), final trajectory elevation compensation correction value (Δh), and final trajectory offset avoidance correction value (Δd). These values ​​will be used to adjust the current trajectory parameters of the device.

[0167] Step S153: The final trajectory curvature correction value is superimposed with the trajectory curvature adjustment value and trajectory curvature correction value in the trajectory evolution rule to generate the target trajectory curvature parameter.

[0168] According to the operational logic in the trajectory evolution rules, the target trajectory curvature amplitude parameter = current trajectory curvature amplitude + trajectory curvature amplitude adjustment value + trajectory curvature amplitude correction value + final trajectory curvature amplitude correction value. During the superposition process, the parameter values ​​are accumulated sequentially according to the parameter execution order table.

[0169] Step S154: The final trajectory elevation compensation correction value is superimposed with the trajectory elevation compensation value and trajectory elevation compensation correction value in the trajectory evolution rule to generate the target trajectory elevation compensation parameter.

[0170] Similarly, the target trajectory elevation compensation parameter = current trajectory elevation compensation + trajectory elevation compensation value + trajectory elevation compensation correction value + final trajectory elevation compensation correction value, and the parameters are accumulated in the order of execution.

[0171] Step S155: The final trajectory offset avoidance correction value is superimposed with the trajectory offset avoidance value and trajectory offset avoidance correction value in the trajectory evolution rule to generate the target trajectory offset avoidance parameter.

[0172] Target trajectory offset avoidance parameter = current trajectory offset avoidance + trajectory offset avoidance value + trajectory offset avoidance correction value + final trajectory offset avoidance correction value, and the parameters are accumulated in the order of execution.

[0173] Step S156: Integrate the target trajectory curvature parameters, target trajectory elevation compensation parameters, and target trajectory offset avoidance parameters to generate a dynamic trajectory adjustment command. The dynamic trajectory adjustment command includes the specific values ​​of each target parameter and the execution timing requirements.

[0174] The three target parameters are packaged according to the device control protocol format to generate a dynamic trajectory adjustment command. The command includes the specific values ​​of each parameter (such as bending angle θ_target, elevation h_target, and offset d_target) and execution timing requirements (such as the effective time and duration of each parameter). The command format conforms to the device communication interface standard to ensure that the device can correctly parse and execute it.

[0175] Step S157: Transmit the dynamic trajectory adjustment command to the steering mechanism control module, elevation compensation mechanism control module, and trajectory offset mechanism control module of the unmanned driving equipment, drive each mechanism to act according to the values ​​and timing requirements in the command, and update the equipment's operating trajectory.

[0176] The device uses its internal CAN bus communication protocol to send dynamic trajectory adjustment commands to the control modules of the steering, elevation compensation, and trajectory offset mechanisms. Each control module drives the corresponding actuators (such as hydraulic valves and motors) to adjust the device's travel direction, height, and lateral position, thereby updating the work trajectory, based on the target parameters and execution sequence in the command.

[0177] Step S158: Collect updated equipment operation trajectory data through the real-time trajectory monitoring module of the unmanned vehicle. The updated equipment operation trajectory data includes updated trajectory curvature data, trajectory elevation compensation data, and trajectory deviation avoidance data.

[0178] After the device performs trajectory adjustments, the real-time trajectory monitoring module continuously collects updated trajectory parameters, including the adjusted trajectory curvature, elevation compensation, and offset avoidance data. This data is obtained through the feedback interface with the control module to ensure consistency with the actual actions.

[0179] Step S159: Collect the newly generated operation trajectory residual signal after updating the operation trajectory through the real-time monitoring module of the unmanned driving equipment. The newly generated operation trajectory residual signal includes new surface compaction trace signal, pasture residual distribution signal and trajectory path imprint signal.

[0180] After the equipment passes through a new work area, the real-time trajectory residue monitoring module collects signals of surface compaction marks, pasture residue distribution, and trajectory path imprints in that area. These signals reflect the actual operational effect of the updated trajectory.

[0181] Step S1510: Input the updated equipment operation trajectory data and the newly generated operation trajectory residual signal into the model through the input layer of the trajectory residual dynamic linkage model, and adjust the correlation coefficient matrix of the feature correlation operation layer and the trajectory evolution algorithm parameters of the trajectory mapping layer.

[0182] The updated trajectory data and the newly generated trajectory residual signal are used as feedback data and input into the trajectory residual dynamic linkage model through the model's backpropagation interface. The model employs an online learning mechanism, adjusting the element values ​​of the correlation coefficient matrix of the feature association operation layer and the long short-term memory network parameters (such as weights and biases) of the trajectory mapping layer based on the error between the feedback data and the theoretical predictions. The adjustment process uses a mini-batch gradient descent algorithm to ensure smooth updates of the model parameters.

[0183] Step S1511: Repeat the adjustment operation of the correlation coefficient matrix and trajectory evolution algorithm parameters until the difference between the theoretical operation trajectory data output by the trajectory residual dynamic linkage model and the actual updated equipment operation trajectory data meets the preset requirements, and complete the cyclic optimization of the trajectory residual dynamic linkage model.

[0184] The updated trajectory data and residual signals are continuously input into the model for parameter adjustment. After each adjustment, the difference between the theoretical trajectory data and the actual trajectory data output by the model (e.g., mean square error) is calculated. If the difference is greater than a preset threshold, adjustment continues; if it is less than or equal to the threshold, optimization stops. At this point, the model parameters have adapted to the latest operating environment, and the iterative optimization is complete.

[0185] In one exemplary embodiment, an unmanned navigation system for hay harvesting is provided. This unmanned navigation system for hay harvesting can be a terminal, server, etc., and its internal structure diagram can be as follows: Figure 2As shown, the unmanned navigation system for forage harvesting includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, near-field communication, or other technologies. When the computer program is executed by the processor, it implements an unmanned navigation method for forage harvesting. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, or it can be a button, trackball, or touchpad set on the shell of an unmanned navigation system used for hay harvesting, or it can be an external keyboard, touchpad, or mouse, etc.

[0186] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. An unmanned navigation method for forage harvesting, characterized in that, The method includes: The system captures signals of pasture growth status, terrain interaction feedback, and residual operation trajectory within the work area, and constructs a dynamic linkage model for trajectory residue. This model associates the interaction paths between pasture growth status signals, terrain interaction feedback signals, residual operation trajectory signals, and the subsequent operation trajectory of the unmanned vehicle. Based on the trajectory residue dynamic linkage model, the trajectory evolution rules of the unmanned driving equipment are generated by signal feature matching. The trajectory evolution rules include the trajectory curvature adjustment method corresponding to the pasture growth status signal, the trajectory elevation compensation method corresponding to the terrain interaction feedback signal, and the trajectory deviation avoidance method corresponding to the operation trajectory residue signal. The unmanned vehicle's all-domain perception components collect real-time signals of pasture growth, real-time terrain interaction feedback, real-time residual signals of the operation trajectory, and current operation trajectory data of the equipment during the operation process. The real-time signal of pasture growth status, the real-time feedback signal of terrain interaction, the real-time residual signal of operation trajectory, and the current operation trajectory data of the equipment are input into the trajectory residual dynamic linkage model to perform trajectory adaptation calculation and generate trajectory deviation correction parameters. Based on the trajectory deviation correction parameters and the trajectory evolution rules, a dynamic trajectory adjustment command is generated to drive the unmanned driving equipment to update its operating trajectory. At the same time, the updated equipment operating trajectory data and the newly generated operating trajectory residual signal are input back into the trajectory residual dynamic linkage model to complete the cyclic optimization of the trajectory residual dynamic linkage model.

2. The unmanned navigation method for forage harvesting according to claim 1, characterized in that, The capture of pasture growth status signals, terrain interaction feedback signals, and operation trajectory residue signals within the operation area is used to construct a dynamic linkage model of trajectory residue, including: The work area is divided into multiple signal acquisition grids of equal area, and the boundary of each signal acquisition grid is seamlessly connected with the adjacent grids; A forage growth sensing device is deployed in each signal acquisition grid. The forage growth sensing device captures the plant distribution signal, stem toughness signal and leaf density signal of the forage in the grid through vegetation sensing technology, and continuously records the continuous change data of each signal. Terrain interactive detection devices are deployed within each signal acquisition grid. These devices capture contact reaction force signals, deformation rebound signals, and particle friction signals of the grid surface using pressure sensing technology, and continuously record the continuous change data of each signal. Trajectory residue monitoring devices are deployed within each signal acquisition grid. These devices use image sensing technology to capture surface compaction marks, pasture residue distribution, and trajectory path imprints in the already operated areas within the grid, and continuously record the continuous changes in each signal. Set a signal synchronization acquisition cycle, and according to the set cycle, synchronously bind the plant distribution signal, stem toughness signal, leaf density signal, contact reaction force signal, deformation rebound signal, particle friction signal, surface compaction mark signal, pasture residue distribution signal, and trajectory path imprint signal within each signal acquisition grid to form a grid ternary signal group. Feature vectors related to pasture growth status are extracted from each grid ternary signal group. These feature vectors are constructed based on the spatial distribution of plant distribution signals, the amplitude variation trend of stem toughness signals, and the density fluctuation of leaf density signals. Feature vectors related to terrain interaction feedback are extracted from each grid ternary signal group. These feature vectors are constructed based on the response peak variation of contact reaction force signals, the recovery time variation of deformation rebound signals, and the intensity fluctuation of particle friction signals. Feature vectors related to operation trajectory residue are extracted from each grid ternary signal group. These feature vectors are constructed based on the area expansion trend of surface compaction trace signals, the uniformity variation of pasture residue distribution signals, and the depth fluctuation of trajectory path imprint signals. Collect historical operational trajectory data of multiple unmanned vehicles in different work areas. The historical operational trajectory data includes trajectory curvature records, trajectory elevation compensation records, and trajectory deviation avoidance records. The feature vectors of pasture growth status, terrain interaction feedback, and operation trajectory residue of each signal acquisition grid are correlated with the corresponding historical operation trajectory data to establish a correspondence table among the four. The basic architecture for building a dynamic linkage model of trajectory residue is provided. The basic architecture includes a three-element signal input layer, a feature association operation layer, a trajectory mapping layer, and a result output layer. Each layer achieves continuous signal interaction through a data transmission link. The correspondence table of all signal acquisition grids is input into the feature association operation layer of the infrastructure. An initial correlation coefficient matrix is ​​generated through the multi-signal interaction algorithm of the feature association operation layer. The initial correlation coefficient matrix reflects the strength of the interaction between the four. The initial correlation coefficient matrix is ​​input into the trajectory mapping layer, and the initial trajectory residual dynamic linkage model is constructed by combining the trajectory evolution algorithm. The verification ternary signal set and the corresponding actual operation trajectory data under different operation scenarios are collected. The verification ternary signal set is input into the initial trajectory residual dynamic linkage model to obtain the predicted operation trajectory data output by the initial trajectory residual dynamic linkage model. The predicted operation trajectory data is compared with the actual operation trajectory data, and the difference value between the two is extracted. Based on the difference value, the correlation coefficient matrix of the feature correlation operation layer and the trajectory evolution algorithm parameters of the trajectory mapping layer are adjusted to form the final trajectory residual dynamic linkage model.

3. The unmanned navigation method for forage harvesting according to claim 2, characterized in that, The step of inputting the correspondence table of all signal acquisition grids into the feature correlation operation layer of the infrastructure, and generating an initial correlation coefficient matrix through the multi-signal interaction algorithm of the feature correlation operation layer includes: All the correspondence tables of the signal acquisition grids are imported into the feature association operation layer of the trajectory residue dynamic linkage model infrastructure in order of grid number. Feature parsing is performed on each correspondence table to extract the feature vector of pasture growth status, the feature vector of terrain interaction feedback, the feature vector of operation trajectory residue, and the core association features of historical operation trajectory data. Through the multi-signal interaction algorithm of the feature association operation layer, pairwise interaction operations are performed on various feature vectors in the correspondence table of a single grid and historical operation trajectory data to generate local correlation coefficients within the grid. The local correlation coefficients reflect the interaction strength between different features and trajectory data within a single grid. Based on the local correlation coefficients of all grids, a grid adjacency graph is constructed, where nodes represent grids and edge weights reflect the similarity of signal features between adjacent grids. The local correlation coefficients are then processed using a graph propagation algorithm or a spatial interpolation algorithm to generate a global correlation feature set. Based on the global correlation feature set, an initial correlation matrix framework is constructed. The row dimension of the initial correlation matrix framework corresponds to various feature vectors and historical operation trajectory data, and the column dimension corresponds to each signal acquisition grid. The association strength data from the global association feature set is filled into the initial association matrix framework. The filled matrix is ​​then standardized, and redundant coefficients are removed from the standardized matrix to generate the initial association coefficient matrix. Each element of the initial association coefficient matrix corresponds to the association strength between a certain type of feature and trajectory data within the relevant grid.

4. The unmanned navigation method for forage harvesting according to claim 1, characterized in that, The method for generating trajectory evolution rules for autonomous driving equipment based on the trajectory residual dynamic linkage model through signal feature matching includes: The correlation coefficient matrix and trajectory evolution algorithm parameters of the trajectory mapping layer in the dynamic linkage model of trajectory residue are analyzed, and the coefficient distributions corresponding to the feature vectors of pasture growth status, terrain interaction feedback, and operation trajectory residue are extracted. The trajectory adjustment structure parameters of the unmanned driving device are obtained, including the bending adjustment range of the steering mechanism, the lifting stroke range of the elevation compensation mechanism, and the lateral movement range of the trajectory offset mechanism. Based on the plant distribution signal features in the feature vector of pasture growth status, plant distribution feature values ​​related to trajectory curvature are extracted; the plant distribution feature values ​​are matched with the curvature adjustment range of the steering mechanism to generate trajectory curvature amplitude adjustment values ​​corresponding to different plant distributions. Based on the stem toughness signal features in the feature vector of pasture growth state, stem toughness feature values ​​related to trajectory curvature are extracted; the stem toughness feature values ​​are matched with the curvature adjustment range of the steering mechanism to generate trajectory curvature amplitude correction values ​​corresponding to different stem toughnesses. Based on the contact reaction force signal features in the terrain interaction feedback feature vector, the contact reaction force feature value related to the trajectory elevation is extracted; the contact reaction force feature value is matched with the lifting stroke range of the elevation compensation mechanism to generate trajectory elevation compensation values ​​corresponding to different contact reactions. Based on the deformation rebound signal features in the terrain interaction feedback feature vector, the deformation rebound feature values ​​related to the trajectory elevation are extracted; the deformation rebound feature values ​​are matched with the lifting stroke range of the elevation compensation mechanism to generate trajectory elevation compensation correction values ​​corresponding to different deformation rebounds. Based on the surface compaction trace signal features in the residual feature vector of the operation trajectory, surface compaction feature values ​​related to trajectory offset are extracted; the surface compaction feature values ​​are matched with the lateral movement range of the trajectory offset mechanism to generate trajectory offset avoidance values ​​corresponding to different surface compaction traces. Based on the characteristics of the grass residue distribution signal in the residual feature vector of the operation trajectory, the residual distribution feature value related to the trajectory offset is extracted; the residual distribution feature value is matched with the lateral movement range of the trajectory offset mechanism to generate trajectory offset avoidance correction values ​​corresponding to different grass residue distributions. The values ​​of trajectory curvature adjustment, trajectory curvature correction, trajectory elevation compensation, trajectory elevation compensation correction, trajectory deviation avoidance, and trajectory deviation avoidance correction are integrated to form a basic trajectory adjustment parameter set. Based on the correlation coefficient matrix in the trajectory residual dynamic linkage model, an evolution weight is assigned to each parameter in the basic trajectory adjustment parameter set, and the parameters in the basic trajectory adjustment parameter set are logically sorted according to the evolution weight order to generate a parameter execution order table. The evolution weight is consistent with the distribution of the corresponding feature vector coefficients. The parameter execution order table includes the order of action and numerical superposition of different parameters. By combining the parameter execution order table and the basic trajectory adjustment parameter set, a trajectory evolution rule is formed. The trajectory evolution rule includes the function logic, numerical calculation method and execution order requirements of all parameters.

5. The unmanned navigation method for forage harvesting according to claim 4, characterized in that, The method involves extracting plant distribution feature values ​​related to trajectory curvature from the plant distribution signal features in the feature vector of pasture growth status; matching these plant distribution feature values ​​with the curvature adjustment range of the steering mechanism to generate trajectory curvature amplitude adjustment values ​​corresponding to different plant distributions, including: Extract plant distribution signal features from the feature vector of forage growth status, wherein the plant distribution signal features include distribution density gradient data and distribution uniformity data; Obtain the bending adjustment range data of the steering mechanism, wherein the bending adjustment range data includes the minimum bending angle and the maximum bending angle; Based on the distribution density gradient data, the distribution density feature value related to the trajectory curvature is extracted; the distribution density feature value is correlated with the curvature adjustment range data of the steering mechanism to determine the adjustment ratio of the curvature angle; Based on the distribution uniformity data, the distribution uniformity feature value related to the trajectory curvature is extracted; the distribution uniformity feature value is correlated with the curvature adjustment range data of the steering mechanism to determine the correction ratio of the curvature angle. Calculate the initial trajectory bending amplitude adjustment value by combining the bending angle adjustment ratio and the correction ratio; The initial trajectory curvature adjustment value is matched with the correlation coefficient matrix in the trajectory residual dynamic linkage model to obtain the correlation coefficient corresponding to the plant distribution signal characteristics. The initial trajectory curvature adjustment values ​​are weighted based on the correlation coefficient to generate intermediate trajectory curvature adjustment values. Compare the intermediate trajectory curvature adjustment value with the trajectory curvature record corresponding to the plant distribution in the historical operation trajectory data, and extract the difference ratio after comparison. Based on the difference ratio, the intermediate trajectory curvature adjustment value is corrected a second time to generate the final trajectory curvature adjustment value corresponding to different plant distributions. The final trajectory curvature adjustment values ​​are stored according to the classification of plant distribution signal characteristics, forming a correspondence table of plant distribution and trajectory curvature adjustment values.

6. The unmanned navigation method for forage harvesting according to claim 4, characterized in that, The correlation coefficient matrix in the trajectory residual dynamic linkage model is used to assign evolution weights to each parameter in the basic trajectory adjustment parameter set, and the parameters in the basic trajectory adjustment parameter set are logically sorted according to the evolution weight order to generate a parameter execution order table, including: The correlation coefficient matrix in the dynamic linkage model of trajectory residue is analyzed, and the coefficient distribution data of the feature vector corresponding to each parameter in the set of basic trajectory adjustment parameters is extracted. The coefficient distribution data reflects the correlation strength between the feature corresponding to each parameter and the operation trajectory. The mean correlation strength of each parameter is calculated based on the coefficient distribution data. Combined with the degree of fluctuation of the coefficient distribution, the baseline value of the evolution weight of each parameter is determined. The larger the mean correlation strength and the smaller the degree of fluctuation, the higher the baseline value of the evolution weight. The evolution weight baseline value is normalized to obtain the final evolution weight of each parameter. The final evolution weight is positively correlated with the correlation strength of the corresponding parameter. Extract the functional attributes of each parameter in the basic trajectory adjustment parameter set. The functional attributes are divided into trajectory curvature adjustment, trajectory elevation compensation, and trajectory offset avoidance. Parameters with the same functional attribute are grouped into the same parameter group. Within each parameter group, the parameters are initially sorted in descending order of their final evolution weights to generate a parameter sorting sequence within each parameter group, while also recording the mutual influence relationships between different parameter groups. Based on the mutual influence between parameter groups, the overall logical adjustment of the parameter sorting sequence within each parameter group is performed to form a preliminary parameter execution order; Based on the initial parameter execution order, the unmanned vehicle model is driven to perform trajectory adjustment in the trajectory simulation environment. If the simulation results show that the device kinematics is infeasible, the trajectory change exceeds the set threshold, or there is a collision with an obstacle, it is determined that there is an execution order conflict. The initial parameter execution order is adjusted according to the default sorting adjustment rules based on the conflict type. The simulation and adjustment process is repeated until the simulation results meet the default feasibility conditions and a parameter execution order table is generated.

7. The unmanned navigation method for forage harvesting according to claim 1, characterized in that, The process of collecting real-time signals of pasture growth status, real-time terrain interaction feedback signals, real-time residual signals of the operation trajectory, and current operation trajectory data of the equipment through the all-domain perception component of the unmanned driving equipment includes: A real-time forage growth sensing module is configured for unmanned driving equipment. This real-time forage growth sensing module maintains the same standard as the forage growth status signal acquisition in the trajectory residue dynamic linkage model, and captures real-time signals of plant distribution, stem toughness, and leaf density at the current location of the equipment operation. The unmanned vehicle is equipped with a terrain interaction real-time detection module. This terrain interaction real-time detection module is consistent with the terrain interaction feedback signal acquisition standard in the trajectory residue dynamic linkage model, and captures the real-time contact reaction force signal, deformation rebound signal and particle friction signal at the current position of the vehicle operation. A real-time trajectory residue monitoring module is configured for unmanned equipment. This real-time trajectory residue monitoring module is consistent with the operation trajectory residue signal acquisition standard in the trajectory residue dynamic linkage model. It captures the real-time signals of ground compaction marks, pasture residue distribution, and trajectory path imprints at the current location after the equipment operation. An unmanned driving device is equipped with a real-time trajectory monitoring module. This real-time trajectory monitoring module is associated with the device's steering mechanism, elevation compensation mechanism, and trajectory deviation mechanism to capture the device's current trajectory curvature data, trajectory elevation compensation data, and trajectory deviation avoidance data. The real-time sensing module for pasture growth is activated according to the set real-time data acquisition cycle, continuously collecting and recording real-time signals of plant distribution, stem toughness, and leaf density, forming a real-time signal sequence of pasture growth status. Following the same real-time data acquisition cycle, the terrain interaction real-time detection module is activated to continuously collect and record real-time signals of contact reaction force, deformation rebound, and particle friction, forming a terrain interaction real-time feedback signal sequence; and the trajectory residue real-time monitoring module is activated to continuously collect and record real-time signals of surface compaction marks, pasture residue distribution, and trajectory path imprints, forming a real-time residual signal sequence of the operating trajectory; and the trajectory real-time monitoring module is activated to continuously collect and record trajectory curvature amplitude data, trajectory elevation compensation data, and trajectory deviation avoidance data, forming a data sequence of the equipment's current operating trajectory. Extract real-time signals of pasture growth status, real-time feedback signals of terrain interaction, real-time residual signals of operation trajectory, and current operation trajectory data of equipment at the same collection time point, assign a unique timestamp identifier to the collection time point, and associate the data with timestamp identifiers to form a real-time data combination unit; Construct a real-time data storage queue to store real-time data combination units from the most recent collection cycles in chronological order of collection time. Then, through the device's wireless transmission module, synchronize the real-time data combination units in the real-time data storage queue to the global data interaction nodes in the work area. The real-time data combination units synchronized to the global data interaction nodes are processed to unify the format and generate standardized real-time data combination units.

8. The unmanned navigation method for forage harvesting according to claim 1, characterized in that, The process involves inputting the real-time signals of pasture growth status, terrain interaction, and operational trajectory residue into the trajectory residue dynamic linkage model to perform trajectory adaptation calculations and generate trajectory deviation correction parameters, including: The input layer of the trajectory residue dynamic linkage model receives standardized real-time data combination units and extracts real-time signals of pasture growth status, real-time feedback signals of terrain interaction, real-time residual signals of operation trajectory, and current operation trajectory data of equipment from the standardized real-time data combination units. Feature extraction is performed on the real-time signal of pasture growth status to generate a real-time feature vector of pasture growth status; the real-time feature vector of pasture growth status is input into the feature association operation layer of the trajectory residual dynamic linkage model, and compared with the pasture growth status feature vector stored in the feature association operation layer to generate a pasture growth status signal deviation vector. Features are extracted from the real-time terrain interaction feedback signal to generate a real-time terrain interaction feedback feature vector; the real-time terrain interaction feedback feature vector is input into the feature association operation layer of the trajectory residual dynamic linkage model, and compared with the terrain interaction feedback feature vector stored in the feature association operation layer to generate a terrain interaction feedback signal deviation vector. Feature extraction is performed on the real-time residual signal of the operation trajectory to generate a real-time residual feature vector of the operation trajectory; the real-time residual feature vector of the operation trajectory is input into the feature association operation layer of the trajectory residual dynamic linkage model, and compared with the residual feature vector of the operation trajectory stored in the feature association operation layer to generate a deviation vector of the residual signal of the operation trajectory. The current operating trajectory data of the equipment is input into the trajectory mapping layer of the trajectory residual dynamic linkage model. Combined with the pasture growth status signal deviation vector, terrain interaction feedback signal deviation vector and operating trajectory residual signal deviation vector, the theoretical operating trajectory data is calculated through the trajectory evolution algorithm of the trajectory mapping layer. The theoretical operation trajectory data is compared and calculated with the current operation trajectory data of the equipment. The differences in trajectory curvature, trajectory elevation compensation and trajectory deviation avoidance are extracted to form a trajectory difference dataset. Based on the correlation coefficient matrix in the trajectory residual dynamic linkage model, correction coefficients are assigned to the trajectory curvature amplitude difference, trajectory elevation compensation difference, and trajectory offset avoidance difference in the trajectory difference dataset. The correction coefficients are consistent with the distribution of the corresponding feature vector coefficients. The differences in trajectory curvature, trajectory elevation compensation, and trajectory deviation avoidance are weighted according to the correction coefficients to generate correction values ​​for trajectory curvature, trajectory elevation compensation, and trajectory deviation avoidance. The trajectory curvature correction value, trajectory elevation compensation correction value, and trajectory deviation avoidance correction value are matched with the correspondence table in the trajectory residual dynamic linkage model. Based on the matching result, the values ​​of the trajectory curvature correction value, trajectory elevation compensation correction value, and trajectory deviation avoidance correction value are adjusted to generate the final trajectory deviation correction parameters. The trajectory deviation correction parameters include the final trajectory curvature correction value, the final trajectory elevation compensation correction value, and the final trajectory deviation avoidance correction value.

9. The unmanned navigation method for forage harvesting according to claim 8, characterized in that, The process involves inputting the current operating trajectory data of the equipment into the trajectory mapping layer of the trajectory residual dynamic linkage model, combining the pasture growth status signal deviation vector, the terrain interaction feedback signal deviation vector, and the operating trajectory residual signal deviation vector, and calculating the theoretical operating trajectory data through the trajectory evolution algorithm of the trajectory mapping layer, including: The system receives the current operation trajectory data of the equipment and transmits it to the trajectory mapping layer of the trajectory residual dynamic linkage model. Simultaneously, it imports the pasture growth status signal deviation vector, terrain interaction feedback signal deviation vector, and operation trajectory residual signal deviation vector for correlation calibration processing to obtain a set of calibrated deviation vectors. Based on the trajectory evolution algorithm of the trajectory mapping layer, the core constituent dimensions of the current working trajectory data of the equipment are decomposed. The core constituent dimensions correspond to trajectory curvature amplitude data, trajectory elevation compensation data and trajectory offset avoidance data, and an independent computing channel is allocated to each core constituent dimension. The deviation vector of the pasture growth status signal from the calibrated deviation vector set is input into the trajectory curvature amplitude calculation channel, and interacts with the trajectory curvature amplitude data in the current operating trajectory data of the equipment to output the theoretical adjustment amount of trajectory curvature amplitude; the deviation vector of the terrain interactive feedback signal is input into the trajectory elevation compensation calculation channel, and interacts with the trajectory elevation compensation data in the current operating trajectory data of the equipment to output the theoretical adjustment amount of trajectory elevation compensation; the deviation vector of the residual signal of the operating trajectory is input into the trajectory offset avoidance calculation channel, and interacts with the trajectory offset avoidance data in the current operating trajectory data of the equipment to output the theoretical adjustment amount of trajectory offset avoidance. Historical correlation parameters stored in the trajectory mapping layer of the trajectory residual dynamic linkage model are extracted. These historical correlation parameters reflect the interaction patterns between various deviation vectors and different dimensions of the operation trajectory data. The historical correlation parameters are then imported into each calculation channel, and the theoretical adjustment amounts output by each calculation channel are corrected. The theoretical adjustment amounts for trajectory curvature amplitude, trajectory elevation compensation, and trajectory deviation avoidance are integrated and combined with the baseline parameters of the current operating trajectory data of the equipment. The fusion calculation is performed through the trajectory evolution algorithm to generate preliminary theoretical operating trajectory data. The theoretical parameters of each dimension in the preliminary theoretical operation trajectory data are compared with the mechanical motion limit parameters of the unmanned driving equipment. If any theoretical parameter exceeds the corresponding mechanical motion limit, the theoretical parameter is adjusted according to the preset constraint rules, and the adjustment requirement is fed back to the corresponding computing channel for iterative calculation until all theoretical parameters meet the mechanical motion limit constraints, and the final theoretical operation trajectory data is obtained.

10. An unmanned navigation system for forage harvesting, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the unmanned navigation method for forage harvesting according to any one of claims 1 to 9 by executing the machine-executable instructions.

Images