Agricultural machine safety control method and device and computer readable storage medium

Abstract

The invention relates to an agricultural machine safety control method and device and a computer readable storage medium. The method comprises the following steps: synchronously acquiring agricultural machine operation data through multiple sensors, extracting terrain, earth surface and motion features, constructing feature vectors, and identifying a geographic position category label of an agricultural machine by using a pre-training model; loading a corresponding exclusive safety control rule according to the label, wherein the rule comprises a specific sensor combination, a risk mapping matrix and a basic operation parameter; screening real-time data according to a rule to calculate a risk state vector, and linearly transforming the risk state vector into a control instruction correction through a mapping matrix; and finally, a control instruction is generated in combination with the basic parameters and is issued to an actuator, so that safety control based on environment category accurate perception and dynamic rule self-adaption is realized, and the operation safety, the environment adaptability and the operation efficiency of the agricultural machinery under complex terrains are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of automation control, and in particular relates to a method and device for agricultural machinery safety control, and a computer-readable storage medium. Background Technology

[0002] In the current field of automated agricultural machinery operations, safety control systems primarily rely on preset fixed parameter thresholds for risk assessment and intervention, such as setting uniform tilt angle alarm limits, braking response curves, or obstacle detection distances. This general strategy is applicable in relatively regular plains, but it proves significantly inadequate in complex and varied geographical environments such as hilly terraces, riverbank wetlands, and orchards. Lacking the ability to identify the inherent category attributes of the operating environment, the system cannot distinguish the essential differences between flat land and slopes, hard soil and soft mud, open areas and narrow forest paths, leading to a rigid safety response mechanism. On steep terrain, fixed thresholds may be too lenient to prevent rollovers in time; on slippery surfaces, standard braking curves may be too aggressive, causing skidding. Furthermore, safety control systems often continuously utilize all sensors and employ the same fusion algorithm, which not only increases computational load and energy consumption but may also affect the timely perception of critical risks by focusing on irrelevant noise in specific environments.

[0003] Therefore, the existing technology is essentially a "one-size-fits-all" safety management model. Its core problem is that the safety control system of agricultural machinery is seriously disconnected from the semantic information of the geographical environment (i.e., "what type of terrain is this and its corresponding risk characteristics"), resulting in a lack of pertinence, foresight and adaptability in the control strategy. It may not only fail to react in dangerous terrain and cause accidents, but may also intervene excessively in safe environments and reduce operational efficiency. It is difficult to achieve synergistic optimization of safety and efficiency in complex agricultural scenarios. Summary of the Invention

[0004] Therefore, it is necessary to provide an agricultural machinery safety control method and device, and a computer-readable storage medium to address the above-mentioned technical problems.

[0005] Firstly, this application provides a method for controlling the safety of agricultural machinery, including:

[0006] S1. Collect synchronous multi-source data streams of agricultural machinery through multiple types of sensors; wherein, the synchronous multi-source data streams include the geographical location data, attitude motion data and environmental perception data of agricultural machinery;

[0007] S2. Based on synchronous multi-source data streams, extract terrain geometric features, surface visual features, and agricultural machinery movement features; construct a comprehensive feature vector based on the terrain geometric features, surface visual features, and agricultural machinery movement features.

[0008] S3. Input the comprehensive feature vector into the pre-trained geographic location classification model to determine the category of the current environment of the agricultural machinery and output the corresponding current geographic location category label.

[0009] S4. Based on the current geographic location category label, retrieve the exclusive safety control rule corresponding to the geographic location category label from the preset safety control rule base; wherein, the exclusive safety control rule includes sensor monitoring scheme, predefined mapping matrix and basic operating parameters; the sensor monitoring scheme defines the sensor combination selected for risk assessment under the current geographic location category corresponding to the current geographic location category label; the predefined mapping matrix defines the rules for mapping the real-time risk state vector of the agricultural machinery to the control command correction amount of the agricultural machinery under the current geographic location category;

[0010] S5. Based on the sensor monitoring scheme in the exclusive safety control rules, select target monitoring data from the synchronous multi-source data stream, and calculate the real-time risk status vector of the agricultural machinery based on the target monitoring data;

[0011] S6. Based on the predefined mapping matrix in the dedicated safety control rules, perform a linear transformation on the real-time risk state vector to generate control command correction quantities;

[0012] S7. Combine the control command correction amount with the basic operation parameters in the exclusive safety control rules to generate the final control command, and send the final control command to the actuator of the agricultural machinery to control the agricultural machinery to perform the corresponding operation.

[0013] Secondly, this application also provides an agricultural machinery safety control device for implementing the method described in the first aspect, the device comprising:

[0014] The multi-source data sensing module is used to collect synchronous multi-source data streams of agricultural machinery through multiple types of sensors; the synchronous multi-source data streams include the geographical location data, attitude motion data, and environmental perception data of the agricultural machinery.

[0015] The environmental feature fusion module is used to extract terrain geometric features, surface visual features, and agricultural machinery motion features based on synchronous multi-source data streams; and to construct a comprehensive feature vector based on the terrain geometric features, surface visual features, and agricultural machinery motion features.

[0016] The dynamic environment recognition module is used to input the comprehensive feature vector into the pre-trained geographic location classification model, determine the category of the current environment of the agricultural machinery, and output the corresponding current geographic location category label.

[0017] The safety policy matching module is used to retrieve the exclusive safety control rules corresponding to the current geographical location category label from the preset safety control rule library. The exclusive safety control rules include sensor monitoring schemes, predefined mapping matrices, and basic operating parameters. The sensor monitoring scheme defines the combination of sensors selected for risk assessment under the current geographical location category corresponding to the current geographical location category label. The predefined mapping matrix defines the rules for mapping the real-time risk state vector of the agricultural machinery to the control command correction amount of the agricultural machinery under the current geographical location category.

[0018] The risk status quantification module is used to select target monitoring data from synchronous multi-source data streams based on the sensor monitoring scheme in the dedicated safety control rules, and calculate the real-time risk status vector of agricultural machinery based on the target monitoring data;

[0019] The instruction correction module is used to perform a linear transformation on the real-time risk state vector according to the predefined mapping matrix in the dedicated safety control rules, and generate control instruction correction quantities.

[0020] The execution control module is used to combine the control command correction amount with the basic operation parameters in the dedicated safety control rules to generate the final control command, and then send the final control command to the actuator of the agricultural machinery to control the agricultural machinery to perform the corresponding operation.

[0021] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a semantic map-based field fire route planning method as described in the first aspect.

[0022] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a semantic map-based field fire route planning method as described in the first aspect.

[0023] The aforementioned agricultural machinery safety control method and device, along with a computer-readable storage medium, firstly collects the geographical location, attitude, and environmental data of the agricultural machinery simultaneously through multiple sensors, extracting comprehensive feature vectors of terrain, surface, and motion. Then, a pre-trained classification model is used to identify the specific geographical location category label of the agricultural machinery. Next, based on this category label, a dedicated safety control rule bound to it is intelligently retrieved from a rule base and loaded. This rule specifies the applicable sensor combination, the mapping matrix for risk quantification, and basic operating parameters. Then, real-time data is filtered according to the sensor scheme in the rule, and a state vector representing the immediate risk of the agricultural machinery is calculated. This risk vector is then linearly transformed into a precise control command correction amount using the pre-set mapping matrix in the rule. Finally, this correction amount is fused with the basic operating parameters in the rule to generate and issue the final control command to the agricultural machinery actuator. This achieves agricultural machinery safety control based on precise environmental category perception and dynamic rule adaptation, effectively improving the safety and efficiency of operations in complex agricultural terrain. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 A flowchart illustrating an agricultural machinery safety control method provided by the present invention;

[0026] Figure 2 This is a schematic diagram illustrating the process of obtaining the current geographic location category label in an optional embodiment of the present invention;

[0027] Figure 3 This is a schematic diagram of the structure of an agricultural machinery safety control device provided by the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0029] refer to Figure 1 The application presents a flowchart illustrating a method for controlling agricultural machinery safety, which includes the following steps:

[0030] S1. Collect synchronous multi-source data streams of agricultural machinery through multiple types of sensors; wherein, the synchronous multi-source data streams include the geographical location data, attitude motion data and environmental perception data of agricultural machinery.

[0031] Specifically, geographic location data is collected collaboratively by two types of sensors: a Global Navigation Satellite System (GNSS) receiver and an Inertial Measurement Unit (IMU). The GNSS receiver is used to acquire the absolute position information of the agricultural machinery, with key parameters including longitude. ,latitude Elevation The inertial measurement unit (IMU) incorporates a three-axis gyroscope and a three-axis accelerometer to output the angular rate of the relative position change of the agricultural machinery in real time. With linear acceleration The data compensates for the positioning shortcomings of the Global Navigation Satellite System in obstructed environments. Attitude and motion data are also collected via an onboard photoelectric odometer, and the agricultural machinery's speed is calculated by measuring wheel rotation speed and wheel diameter. With mileage This further enriches the dimensions of motion state data. Environmental perception data is collected through three types of sensors: LiDAR, industrial-grade machine vision cameras, and soil moisture sensors. LiDAR is used to collect distances to obstacles around the agricultural machinery. With direction Data: Machine vision cameras are used to collect surface image data of the work area, and soil moisture sensors are used to collect soil moisture content data of the work area. Data enables a comprehensive understanding of the work environment.

[0032] Synchronization of multi-source data streams employs timestamp synchronization technology, using the system clock of the agricultural machinery's onboard industrial control computer as a reference to add a uniformly formatted timestamp to the raw data output by each sensor. Linear interpolation is used to correct sampling time differences caused by variations in sampling frequencies between different sensors, eliminating invalid data with excessive timestamp deviations to ensure spatiotemporal consistency of data streams from different sensors at the same time dimension. The synchronized data is stored in a specified format on local storage devices, ensuring that storage latency is within a reasonable range and preventing data delays from affecting the timeliness of subsequent control responses.

[0033] S2. Based on synchronous multi-source data streams, extract terrain geometric features, surface visual features, and agricultural machinery movement features; construct a comprehensive feature vector based on terrain geometric features, surface visual features, and agricultural machinery movement features.

[0034] Specifically, the terrain geometry features are based on elevation data collected by the Global Navigation Satellite System. Data calculations were performed, with key parameters including terrain slope. Topographic slope With terrain curvature The terrain slope calculation employs a third-order inverse distance weighted method. This method improves the accuracy of elevation change rate calculation through weighted fitting. Specifically, a 3×3 elevation sampling grid is selected around the current location of the agricultural machinery. The grid contains 9 sampling points, and the spacing between adjacent sampling points is set to... Using the coordinates of the grid center point as... The eight surrounding sampling points are numbered sequentially in the order of top left, top, top right, right, bottom right, bottom, bottom left, and left, with coordinates as follows: ( First, calculate based on this grid. direction and The rate of change of elevation in a given direction is calculated using the following formula:

[0035]

[0036]

[0037] In the formula, to These are the elevation values ​​of eight sampling points in different directions within a 3×3 sampling grid. The sampling point spacing, for Rate of change of elevation in direction for The elevation change rate in the above two directions. The terrain slope is the angle between the tangent plane of the surface unit and the horizontal plane, calculated based on the elevation change rate in these two directions. The calculation formula is as follows:

[0038]

[0039] In the formula, For terrain slope, It is the arctangent function. Using pi (π), the calculation result is in degrees, and the slope range is between 0 and 90 degrees, which can intuitively reflect the steepness of the terrain.

[0040] The terrain aspect is the direction corresponding to the maximum slope, used to characterize the orientation of the terrain's tilt. North is defined as 0 degrees, and clockwise rotation is considered the positive direction. direction and The positive and negative relationships of the rate of change of elevation are calculated for different cases, and the calculation formula is as follows:

[0041]

[0042]

[0043]

[0044] In the formula, This is the slope aspect value, in degrees, ranging from 0 to 360 degrees, providing orientation characteristics for subsequent terrain category determination.

[0045] Topographic curvature reflects the degree of curvature of a surface unit and is used to distinguish between gentle terrain and complex undulating terrain. It is divided into planar curvature. With section curvature Planar curvature characterizes the degree of curvature of the terrain in the horizontal direction, while profile curvature characterizes the degree of curvature of the terrain in the slope direction. Both are calculated based on the second-order partial derivative of elevation, and the calculation formulas are as follows:

[0046]

[0047]

[0048] In the formula, For plane curvature, For the curvature of the profile, for Second-order partial derivative of elevation in direction, for Second-order partial derivative of elevation in direction, This is the mixed second-order partial derivative. The second-order partial derivative is calculated using the second-order difference method, which fits the curvature of a curve based on the elevation difference between adjacent sampling points. The unit of curvature is the reciprocal of the length; the larger the value, the more pronounced the terrain curvature.

[0049] The surface visual features are calculated based on surface image data acquired by machine vision cameras. To improve feature extraction accuracy, the surface images are first preprocessed using gray-level histogram equalization, then the CLAHE algorithm is used to enhance image contrast, followed by the Canny edge detection algorithm to enhance image contour recognition. Finally, a VGG16 convolutional neural network is used to extract three core features. Texture features are calculated based on the gray-level co-occurrence matrix (GLCM), which is used to statistically analyze the spatial distribution of gray values ​​in the image. A set distance is selected... With angle Construct a matrix and extract contrast. ,energy ,entropy Correlation The four parameters are calculated using the following formula:

[0050]

[0051]

[0052]

[0053]

[0054] In the formula, Contrast is used to reflect the degree of difference in the grayscale values ​​of an image; Energy reflects the uniformity of the grayscale distribution in an image; Entropy is a measure of the complexity of image information. Correlation reflects the degree of spatial correlation between image grayscale values; The elements of the gray-level co-occurrence matrix represent gray values. With grayscale value The probability of them occurring simultaneously at a given distance and angle; The gray level of the image. , These are the gray values ​​of two pixels in the gray-level co-occurrence matrix, and their values ​​range from [value range missing]. to Double summation covers the matrix elements corresponding to all combinations of gray values; for Mean gray level of orientation for Mean gray level of orientation for Standard deviation of directional grayscale for Directional grayscale standard deviation. Vegetation cover features were calculated using the supergreen feature (ExG) and the Otsu binarization segmentation algorithm. The supergreen feature was used to enhance the difference between vegetated and non-vegetated areas. The calculation formula is as follows:

[0055]

[0056] In the formula, The super-green characteristic value, The pixel value of the green channel of the image. The pixel value of the red channel of the image. This represents the pixel value of the blue channel in the image. The Otsu algorithm is used to binarize the ultragreen feature image, resulting in mask images of vegetated and non-vegetated areas, and the vegetation coverage is... This represents the ratio of the number of pixels in the vegetation area to the total number of pixels in the image. Obstacle contour features are obtained by extracting the minimum bounding rectangle of the obstacle using a contour detection algorithm, and the aspect ratio of the rectangle is calculated. Area proportion Aspect Ratio This represents the ratio of the rectangle's length to its width, and its area percentage. This is the ratio of the obstacle area to the image area, used to distinguish different types of obstacles.

[0057] The motion characteristics of agricultural machinery are obtained through the fusion calculation of data from the inertial measurement unit and the odometer. The core parameters include the roll angle. Pitch angle angular velocity of steering With driving acceleration Data fusion can improve the accuracy of motion parameter measurements. Roll angle With pitch angle The calculation is performed using a complementary filtering algorithm, which combines the high-frequency response characteristics of the inertial measurement unit with the long-term stability of the global navigation satellite system. The formula is as follows:

[0058]

[0059]

[0060] In the formula, The combined roll angle, The combined pitch angle. These are the filter coefficients, used to balance the weights of the two types of sensor data; The roll angle data is directly output by the inertial measurement unit. The pitch angle data is directly output by the inertial measurement unit; The roll angle data is derived from the elevation changes of the Global Navigation Satellite System. This is elevation angle data derived from elevation variations of the Global Navigation Satellite System. Turning angular velocity. The steering sensitivity of agricultural machinery is reflected by calculating the speed difference between the left and right wheels measured by the odometer, using the following formula:

[0061]

[0062] In the formula, For the steering angular velocity, The speed of the right wheel. The speed of the left wheel is measured by an odometer. The wheelbase is the distance between the centers of the left and right wheels of the agricultural machinery, and is an inherent parameter of the machinery. Acceleration during travel. The difference between the current speed and the speed at the previous moment is divided by the sampling time interval to reflect the trend of agricultural machinery speed change. The formula is as follows:

[0063]

[0064] In the formula, For driving acceleration, The speed at the current moment, The velocity at the previous moment, This is the sampling time interval, corresponding to the sensor's sampling frequency.

[0065] To eliminate the impact of differences in feature magnitude on subsequent model training, the extracted terrain geometric features, surface visual features, and agricultural machinery motion features need to be normalized, and then their dimensions need to be regularized in a preset order to construct a comprehensive feature vector. The mathematical expression is as follows:

[0066]

[0067] In the formula, For the comprehensive feature vector, For elevation, For vegetation coverage, The aspect ratio of the rectangle is... This represents the vector transpose operation, which converts a row vector into a column vector to meet the model input requirements.

[0068] S3. Input the comprehensive feature vector into the pre-trained geographic location classification model to determine the category of the current environment of the agricultural machinery and output the corresponding current geographic location category label.

[0069] Specifically, the geolocation classification model employs a network structure that integrates Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs). The CNN portion extracts spatial features from the comprehensive feature vector, capturing the spatial relationships between different features. The LSTM portion mines temporal correlation features from the comprehensive feature vector, adapting to the continuous data flow during agricultural machinery operations. The CNN portion includes a predetermined number of convolutional and pooling layers. The convolutional layers extract local features through convolutional kernels, while the pooling layers reduce dimensionality and retain key features. The LSTM portion includes a predetermined number of hidden layers, which memorize long-term temporal information through a gating mechanism. The output layer is a fully connected layer, with the number of neurons corresponding to the preset number of terrain categories, achieving feature-to-category mapping.

[0070] The model training process requires the pre-construction of a sample dataset. This dataset encompasses comprehensive feature vector samples from various typical geographical environments, with each sample labeled with a corresponding geographical location category. These labels guide model training. The model training employs the cross-entropy loss function, which is suitable for multi-class classification tasks and quantifies the difference between the model's predictions and the true labels. The specific formula is as follows:

[0071]

[0072] In the formula, This represents the cross-entropy loss value; the smaller the loss value, the higher the model's prediction accuracy. This refers to the number of terrain categories, i.e., the total number of preset geographical environment categories; This is a one-hot encoding of the real label. One-hot encoding is an encoding method used to represent category labels, where only the position corresponding to the real category is 1, and the rest are 0. The first output of the model The predicted probability for each terrain type is calculated, and the sum of the predicted probabilities for all types is 1. The Adam optimizer iteratively optimizes the model's weight parameters. The Adam optimizer combines momentum gradient descent with an adaptive learning rate strategy to improve training efficiency and stability. The learning rate is updated using an exponential decay strategy to avoid oscillations during later training stages. The formula is as follows:

[0073]

[0074] In the formula, For the first The learning rate for the next iteration. The initial learning rate, The attenuation coefficient is... The iteration count is specified. Training stops when the model's classification accuracy on the validation set reaches a preset threshold. The validation set is used to evaluate the model's generalization ability and avoid overfitting. The normalized composite feature vector is input into the trained model, and the class labels are output through feature mapping and a Softmax classifier. The Softmax classifier is an output layer function used for multi-class classification, which transforms the raw scores into a probability distribution, as shown in the following formula:

[0075]

[0076] In the formula, The first output of the fully connected layer The original score for terrain-like features, The first output of the fully connected layer The original scores of the terrain types are used to select the category with the highest predicted probability as the current geographic location category label, ensuring the reliability of the classification results.

[0077] S4. Based on the current geographic location category label, retrieve the exclusive safety control rule corresponding to the geographic location category label from the preset safety control rule library; wherein, the exclusive safety control rule includes sensor monitoring scheme, predefined mapping matrix and basic operating parameters; the sensor monitoring scheme defines the sensor combination selected for risk assessment under the current geographic location category corresponding to the current geographic location category label; the predefined mapping matrix defines the rules for mapping the real-time risk state vector of the agricultural machinery to the control command correction amount of the agricultural machinery under the current geographic location category.

[0078] Specifically, the dedicated safety control rules include sensor monitoring schemes, predefined mapping matrices, and basic operational parameters. These three types of rules work together to form a complete safety control scheme adapted to specific terrains. The safety control rule base is a structured database, stored in key-value pairs. The key is a category label, and the value is the corresponding dedicated rule. It is stored in the local directory of the agricultural machinery's onboard industrial control computer to ensure real-time rule retrieval.

[0079] The sensor monitoring scheme defines the sensor combinations selected for risk assessment under corresponding geographical location categories. These combinations are customized based on the core risk characteristics of the corresponding terrain. By activating core sensors, deactivating redundant sensors, or adjusting sensor operating parameters, energy consumption and computational load are reduced while maintaining risk monitoring accuracy. A predefined mapping matrix defines the rules for mapping the real-time risk state vector of agricultural machinery to control command correction quantities under corresponding geographical location categories. This is a fixed-dimensional matrix that achieves accurate conversion of risk factors into control adjustment quantities, in the following form:

[0080]

[0081] In the formula, For a predefined mapping matrix, To control the dimension of the command correction amount, corresponding to the speed correction amount Steering angle correction amount Braking pressure correction amount Adjustment amount of control parameters; The dimension of the real-time risk status vector varies with the terrain category, and the number of core risk indicators differs for different terrains. As a matrix element, representing the first... The first risk factor affects the second The influence weights of each control command correction are determined using the Analytic Hierarchy Process (AHP). AHP combines expert experience with objective data to determine weights, ensuring the rationality of the mapping relationship. Basic operating parameters include the rated travel speed of the agricultural machinery. Steering angle limit Braking pressure threshold Based on the operational specifications and safety standards set for different geographical environments, it provides basic benchmark values ​​for control commands.

[0082] S5. Based on the sensor monitoring scheme in the exclusive safety control rules, select target monitoring data from the synchronous multi-source data stream, and calculate the real-time risk status vector of the agricultural machinery based on the target monitoring data.

[0083] Specifically, the target monitoring data is the data stream corresponding to the sensor type specified in the sensor monitoring scheme. Redundant data that does not match the specified sensor type in the synchronous multi-source data stream is removed, and the filtered data is the target monitoring data. This target monitoring data directly corresponds to the core risk points of the current terrain. To ensure the comparability of different risk indicators, the original risk indicators in the target monitoring data need to be normalized. The min-max normalization method is used to map all indicators to the [0,1] interval, as shown in the following formula:

[0084]

[0085] In the formula, For the first The first sample The normalized value of each risk indicator. For the first The first sample The raw values ​​of each risk indicator For the first The minimum value of each risk indicator. For the first The maximum, minimum, and maximum values ​​of each risk indicator are determined based on historical operational data, covering the range of indicators under various operational scenarios.

[0086] The dimensions of the real-time risk status vector correspond to the terrain category. It is composed of normalized core risk indicators arranged in a preset order. The core risk indicators are set according to the core risk characteristics of the corresponding terrain, including types such as angle deviation, exceeding limit value, distance deviation, exceeding water content limit, wheel speed difference deviation, and area proportion. The mathematical expression is as follows:

[0087]

[0088] In the formula, This is a real-time risk state vector. to These are the normalized values ​​of each core risk indicator. The larger the vector value, the higher the corresponding risk level, providing a quantitative basis for subsequent control adjustments.

[0089] S6. Based on the predefined mapping matrix in the dedicated safety control rules, perform a linear transformation on the real-time risk state vector to generate control command correction quantities.

[0090] Specifically, the linear transformation is achieved through matrix multiplication, which converts each risk factor in the real-time risk state vector into the adjustment amount of the corresponding control parameter according to its weight. The calculation formula is as follows:

[0091]

[0092] In the formula, The control command correction vector is composed of the velocity correction amount. Steering angle correction amount Braking pressure correction amount Composed of elements such as rows and columns, matrix multiplication follows the row-column rule. By predefined weight allocation of the mapping matrix, the impact of different risk factors is transformed into corresponding control instruction adjustment amounts, achieving a precise correlation between risk factors and control strategies, and ensuring that the correction direction matches the risk type.

[0093] To prevent the correction amount from exceeding the actuator range and causing actuator malfunction, the control command correction amount needs to be constrained. Speed ​​correction is only allowed for deceleration, and its range is between the negative value and zero of the rated speed to avoid increasing the risk during acceleration. Steering angle correction is between the negative and positive values ​​of half the steering angle limit to ensure stable steering and prevent oversteering. Braking pressure correction is between zero and the maximum permissible increment, where the maximum permissible increment is the difference between the maximum braking system pressure and the reference braking pressure, to prevent excessive braking pressure from damaging braking components.

[0094] S7. Combine the control command correction amount with the basic operation parameters in the exclusive safety control rules to generate the final control command, and send the final control command to the actuator of the agricultural machinery to control the agricultural machinery to perform the corresponding operation.

[0095] Specifically, the basic control command vector corresponding to the basic operation parameters is the standard operation control parameter of the agricultural machinery under the corresponding terrain, and the mathematical expression is as follows:

[0096]

[0097] In the formula, Based on the control command vector, to These are the rated travel speeds of the agricultural machinery. Maximum steering angle Reference braking pressure These basic control parameters provide a baseline framework for the final control commands.

[0098] The final control command vector is generated by superimposing the basic control command vector with the constrained control command correction amount, and the sign function is used to determine whether correction is needed. The calculation formula is as follows:

[0099]

[0100] In the formula, For the final control command vector, This is a sign function used to determine whether there are any out-of-limit indicators in the real-time risk state vector. When an out-of-limit indicator exists, the sign function value is -1, corresponding to the correction direction of deceleration, reducing steering angle, and increasing braking pressure, thus realizing risk intervention; when there are no out-of-limit indicators, the sign function value is 0, no correction is made, the basic operating parameters are maintained, and operating efficiency is ensured.

[0101] The parameters in the final control command vector correspond to the driving speed. Steering angle Braking pressure Specifically, the calculation involves the sum of the products of the basic parameters, the corresponding correction values, and the sign function values. The final control commands are then sent to the actuators of the agricultural machinery via the CAN bus. These actuators include the drive motor controller, braking system controller, and steering system controller. Bus transmission ensures the real-time performance and stability of the command transmission.

[0102] The drive motor controller adjusts the motor speed based on the final driving speed parameter, achieving precise speed regulation through the conversion relationship between speed and driving speed. The calculation formula is as follows:

[0103]

[0104] In the formula, This refers to the motor speed. For the final driving speed, The transmission ratio is the ratio of the motor speed to the wheel speed. Here, is the wheel diameter; both are inherent parameters of agricultural machinery. The braking system controller adjusts the brake cylinder pressure based on the final braking pressure parameter, achieving linear pressure regulation through a hydraulic proportional valve to ensure adjustment accuracy and match braking performance with risk levels. The steering system controller adjusts the steering tie rod travel based on the final steering angle parameter, converting the tie rod travel into steering angle through mechanical transmission. The calculation formula is as follows:

[0105]

[0106] In the formula, For the steering tie rod travel, For the final steering angle, To control the wheelbase of agricultural machinery, the steering angle is precisely controlled by adjusting the travel of the steering tie rod, thereby completing an adaptive safety control closed loop.

[0107] This method enables dynamic adjustment of agricultural machinery safety control strategies based on the risk characteristics of different geographical environments, avoiding the rigidity of control strategies caused by fixed parameter thresholds. In dangerous terrain, the system can promptly trigger targeted safety intervention measures to reduce the incidence of agricultural machinery accidents; in safe environments, the system can reduce unnecessary intervention operations, ensuring the operational efficiency of agricultural machinery and achieving synergistic optimization of agricultural machinery safety and operational efficiency in complex agricultural scenarios.

[0108] The aforementioned agricultural machinery safety control method first collects the geographical location, attitude, and environmental data of the agricultural machinery simultaneously through multiple sensors, and extracts comprehensive feature vectors of terrain, surface, and motion from these data. Then, a pre-trained classification model is used to identify the specific geographical location category label of the agricultural machinery. Next, based on this category label, a dedicated safety control rule bound to it is intelligently retrieved from the rule base and loaded. This rule specifies the applicable sensor combination, the mapping matrix for risk quantification, and the basic operating parameters. Then, real-time data is filtered according to the sensor scheme in the rule, and a state vector representing the immediate risk of the agricultural machinery is calculated. This risk vector is then linearly transformed into a precise control command correction value using the pre-set mapping matrix in the rule. Finally, this correction value is fused with the basic operating parameters in the rule to generate and issue the final control command to the agricultural machinery actuator. This achieves agricultural machinery safety control based on precise environmental category perception and dynamic rule adaptation, effectively improving the safety and efficiency of operations in complex agricultural terrain.

[0109] refer to Figure 2 In an optional embodiment, S3 includes the following steps:

[0110] S11. Based on the comprehensive feature vector at the current moment, calculate the probability of belonging to each preset geographical location category through the geographical location classification model, and output the preset geographical location category with the highest probability as the main category label.

[0111] Specifically, the current comprehensive feature vector is the feature vector constructed and normalized in step S2. It contains full-dimensional information on terrain geometric features, surface visual features, and agricultural machinery movement features, and has already eliminated the influence of feature magnitude differences on the model output through preprocessing. The geographic location classification model adopts the aforementioned CNN-LSTM fusion network structure. This model, after prior training, has the ability to recognize and classify various terrain features, and can map the input comprehensive feature vector to the probability distribution of each preset geographic location category. The model output probability is calculated using a Softmax classifier, with the specific formula as follows:

[0112]

[0113] In the formula, The current comprehensive feature vector belongs to the first... The probability of a predefined geographic location category. The first output of the fully connected layer of the model The original score for terrain-like features, The first output of the fully connected layer The original score for terrain-like features, To ensure a reasonable probability distribution, the sum of the probabilities for all categories is set to 1, representing the total number of preset geographical location categories. After calculating the probability for each category, all probability values ​​are sorted, and the preset geographical location category corresponding to the highest probability is selected as the primary category label. The primary category label represents the most likely terrain category where the agricultural machinery is located at the current moment, providing a basis for subsequent decision-making. To further improve the reliability of the primary category label, the validity of the highest probability value is validated before output. If the highest probability value is lower than the preset minimum probability threshold, the current feature vector is deemed invalid, and the primary category label from the previous moment is used, avoiding classification bias caused by abnormal feature data.

[0114] S12. Based on a time-series queue containing historical comprehensive feature vectors of the most recent N consecutive historical moments, input the time-series queue into a pre-trained sequence prediction model to predict the feature vectors after the next M moments, and obtain the predicted feature vectors; wherein, the method of obtaining the historical comprehensive feature vectors is the same as that of the comprehensive feature vectors; N is a preset first positive integer, and M is a preset second positive integer.

[0115] Specifically, the construction of the time-series queue takes the current moment as the time node, traces back the comprehensive feature vectors of the most recent N consecutive historical moments, and stores them sequentially into the queue structure in chronological order, forming a dimension of The time-series feature matrix, where This refers to the dimension of the comprehensive feature vector. The method for obtaining the historical comprehensive feature vector is completely consistent with that for the current comprehensive feature vector. Both are generated through the data acquisition process in step S1, the feature extraction process in step S2, and the dimension normalization process. This ensures the consistency and comparability of all feature vectors in the time series queue and avoids introducing additional errors due to differences in feature generation methods.

[0116] N, the first preset positive integer, needs to be set in combination with the sensor sampling frequency and terrain change rate during agricultural machinery operation. It must ensure that the time-series queue contains sufficient historical feature information to reflect terrain change trends, while avoiding excessive queue length that would increase computational load and prolong prediction latency. The value can range from 5 to 20 and can be dynamically adjusted according to the specific operational scenario. M, the second preset positive integer, represents the prediction time step. Its value must match the agricultural machinery's speed and the magnitude of terrain change to ensure that the predicted feature vector's future timeframe covers the area the agricultural machinery is about to enter, providing effective support for terrain category prediction. The value can range from 1 to 5, balancing prediction accuracy and real-time performance.

[0117] The sequence prediction model employs a pre-trained Long Short-Term Memory (LSTM) network. This model possesses the ability to capture long-term dependencies in time-series data and predict future feature trends based on the changing patterns of historical feature vectors. The model input is the constructed time-series feature matrix, and the output is the predicted feature vectors for the next M time points. Each predicted feature vector has the same dimension as the current comprehensive feature vector and includes various terrain and motion feature parameters. The model training process uses historical time-series feature data as samples and employs a mean squared error loss function to optimize the model parameters. The specific formula is as follows:

[0118]

[0119] In the formula, This is the mean squared error loss value. For the sample size, For the feature vector dimension, For the first The first sample The true value of the dimensional feature. The prediction accuracy is improved by minimizing the loss value to obtain the predicted value of the corresponding feature. After training, the prediction error of the model must be controlled within a preset range.

[0120] S13. Input the predicted feature vector into the geographic location classification model to obtain the predicted geographic location category label and the corresponding predicted probability.

[0121] Specifically, the geographic location classification model used as the input for the predicted feature vector is completely identical to the model used in S11, requiring no additional training and ensuring the uniformity of classification standards and the comparability of results. Before input, the predicted feature vector needs to undergo the same normalization process as the current comprehensive feature vector, mapping each feature parameter to a preset interval to eliminate the impact of magnitude differences on the model's classification results.

[0122] The model's processing flow for predicted feature vectors is consistent with S11. First, spatial features are extracted from the predicted features using a convolutional neural network. Then, temporal correlation features are mined using a long short-term memory network. The raw scores for each preset geographic location category are output via a fully connected layer. Finally, the raw scores are converted into predicted probabilities using a softmax function. The output contains two core pieces of information: first, the predicted geographic location category label, i.e., the terrain category corresponding to the highest predicted probability; and second, the predicted probability corresponding to that category label, representing the model's confidence level in the future terrain category. Simultaneously, the model also outputs the predicted probability distribution for all preset categories, providing complete data support for subsequent multi-scenario decision-making.

[0123] S14. When the predicted geographic location category label is the same as the main category label, the main category label shall be used as the current geographic location category label.

[0124] Specifically, this step corresponds to the decision-making logic for stable terrain categories. When the predicted category matches the current main category, it indicates that the terrain where the agricultural machinery will be located in the future will remain consistent with the current terrain, with no trend of category change. In this case, the main category label is used as the final result, ensuring the stability of the category judgment. This scenario is common in areas with gentle terrain changes, such as plains, large orchards, where the agricultural machinery remains in the same terrain category for a relatively long period. The consistency between the predicted result and the current result verifies the reliability of the main category label, avoiding meaningless label switching that leads to frequent adjustments in the control strategy, thus balancing operational stability and efficiency. In actual execution, the duration of this consistent state needs to be recorded to provide auxiliary basis for subsequent judgments in terrain-changing scenarios.

[0125] S15. When the predicted geographic location category label is different from the main category label and the predicted probability is lower than the preset confidence threshold, the main category label shall be used as the current geographic location category label.

[0126] Specifically, the preset confidence threshold is the core parameter for determining the reliability of the prediction results. Its value is set based on the classification accuracy of the geographic location classification model and the requirements of the operational scenario. It is calibrated using historical test data to ensure that the accuracy of the prediction results reaches the preset standard when the value is above the threshold. When the predicted category is different from the main category but the predicted probability is lower than the threshold, it indicates that the model's confidence in predicting the future terrain category is insufficient. The prediction results may be affected by factors such as feature noise and model errors, and therefore lack reliability.

[0127] At this point, prioritizing the use of the primary category label as the final result can avoid misclassification caused by unreliable predictions and ensure that the current control strategy matches the actual terrain. For example, in the boundary area between hilly terraces and river wetlands, if the predicted feature vector is affected by edge features, resulting in a lower prediction probability, using the primary category label can avoid safety risks caused by switching control strategies due to misjudgment. Simultaneously, this low-confidence prediction event will be recorded. If such events occur frequently, a feature preprocessing optimization process will be triggered to improve subsequent prediction accuracy.

[0128] S16. When the predicted geographic location category label is different from the main category label, and the duration for which the predicted probability is higher than or equal to the preset confidence threshold does not exceed the preset time threshold, the main category label shall be used as the current geographic location category label.

[0129] Specifically, the preset time threshold is the time standard for determining whether the terrain category has undergone a stable change. Its value is set in combination with the agricultural machinery's driving speed and the width of the terrain transition area to ensure that the terrain is considered to be about to undergo a stable change only when the predicted category continuously meets the high confidence condition for a certain duration. The typical range is 0.5 to 2 seconds to avoid label switching errors caused by instantaneous high confidence predictions. The duration is calculated starting from the moment when the predicted probability is higher than or equal to the confidence threshold and the category is different. If the prediction result changes during the period (e.g., the category is the same, or the probability is lower than the threshold), the timer is reset.

[0130] This step is suitable for high-confidence prediction scenarios during the initial stage of terrain transition or due to transient disturbances, such as the edge area before agricultural machinery enters an orchard. The initial prediction of a category change is accurate and the probability meets the criteria, but the duration is insufficient. In this case, the primary category label is retained, allowing enough time to verify the stability of the prediction results. This design filters out false predictions caused by transient disturbances, ensuring that label switching only occurs in scenarios where the terrain has indeed undergone stable changes, balancing the foresight and stability of the control strategy.

[0131] S17. When the predicted geographic location category label is different from the main category label, and the duration for which the predicted probability is higher than or equal to the preset confidence threshold exceeds the preset time threshold, the predicted geographic location category label shall be used as the current geographic location category label.

[0132] Specifically, this step corresponds to scenarios where the terrain undergoes stable changes. When the predicted category differs from the main category, the predicted probability meets the standard, and the duration exceeds the time threshold, it indicates that the model's prediction of the future terrain category is highly reliable, and the terrain change trend is stable, meaning the agricultural machinery is about to enter a new terrain area. At this point, using the predicted geographic location category label as the final result allows for early switching of safety control rules to adapt to the new terrain, enabling proactive adjustments to the control strategy and preventing safety risks caused by lagging control strategies after the agricultural machinery enters the new terrain.

[0133] For example, when agricultural machinery moves from plains to hilly terraces, the predictive model detects the category change in advance, and the high-confidence prediction duration meets the standard. At this point, the system switches to the control rules corresponding to hilly terraces, adjusts the sensor monitoring scheme and safety parameter thresholds, and prepares for rollover protection in advance. After the label switch, the system synchronously updates the historical feature vectors in the time-series queue, and uses the new category as a basis for subsequent predictions and decisions, forming a closed-loop control. Meanwhile, to ensure a smooth switch, the control strategy adopts a gradual transition method to avoid affecting the stability of agricultural machinery operations due to sudden parameter changes.

[0134] In an optional embodiment, after S16, a category switching warning signal corresponding to the predicted geographic location category label is also generated.

[0135] Specifically, the core judgment result of step S16 is that the predicted geographic location category label is different from the main category label, and the predicted probability is higher than or equal to the preset confidence threshold but the duration does not exceed the preset time threshold. At this time, although the main category label is still used as the current geographic location category label, the prediction result has shown a clear category change trend, and an early warning signal can be generated to prepare for the switch of control rules in advance. The category switch early warning signal is a structured digital signal containing core information: predicted geographic location category label, predicted probability value, and early warning generation timestamp. Among them, the predicted geographic location category label is used to locate the corresponding preparatory rule, the predicted probability value helps to determine the weight benchmark for transition fusion, and the timestamp is used to calibrate the starting node of the subsequent transition time window.

[0136] The warning signal is transmitted to the rule matching module via the internal bus of the agricultural machinery vehicle control system. The transmission uses an encrypted format to prevent signal interference or tampering, and the transmission delay is controlled within a preset range to ensure real-time performance. Simultaneously, the system stores the warning signal in a local cache for a duration no less than the transition time window. This cache is used for signal verification and tracing during the transition process. If a similar, valid warning signal exists in the cache, the newly generated warning signal will overwrite the old signal, ensuring that the warning information is consistent with the latest prediction results.

[0137] Next, in this embodiment, S4 includes the following steps:

[0138] S21. Based on the current geographic location category label, match the first security control rule corresponding to the current geographic location category label in the security control rule base, and load the first security control rule into the active memory area as the current active rule.

[0139] Specifically, the current geographic location category label is the final label determined by the category judgment process in the previous implementation method. It may be the main category label (corresponding to scenarios S14, S15, and S16) or the predicted geographic location category label (corresponding to scenario S17). The security control rule base is a structured database that uses an index matching mechanism. It uses the geographic location category label as the index key to quickly retrieve the corresponding complete security control rules. The retrieval time is controlled within a preset range to meet real-time control requirements.

[0140] The first safety control rule is the rule that precisely matches the current tag, and includes three core components: sensor monitoring scheme, predefined mapping matrix, and basic operating parameters. The active memory area is a high-speed memory block specifically allocated in the agricultural machinery vehicle's industrial control computer, characterized by fast read / write speeds and high priority. After the rule is loaded, it can directly interact with the actuator control module, ensuring rapid issuance of control commands. During the loading process, the integrity of the rule is verified. If a rule has missing or incorrect parameters, a fault-tolerance mechanism is immediately triggered, using the previous version of the active rule already stored in memory, while simultaneously issuing a rule exception warning signal to ensure stable system operation.

[0141] S22. When no category switching warning signal is received, the currently active rule will be used as the exclusive security control rule.

[0142] Specifically, the absence of a warning signal means that the current predicted geographic location category label is consistent with the main category label, or the prediction result does not meet the high confidence condition, indicating that the terrain category is stable and no rule switching preparation is needed. In this case, the currently active rule in the active memory area is directly used as the dedicated safety control rule to ensure the stability of the control strategy and avoid meaningless rule adjustments that could affect the rhythm of agricultural machinery operations. This scenario is common in operation areas with long-term stable terrain, where the system maintains a single rule running continuously, reducing computational load and energy consumption.

[0143] S23. When a category switching warning signal is received, according to the predicted geographical location category label corresponding to the category switching warning signal, a second security control rule corresponding to the predicted geographical location category label is obtained from the security control rule base, and the second security control rule is loaded into the preparation memory area as a preparation rule.

[0144] Specifically, upon receiving an early warning signal, the rule preloading process is triggered. The predicted geographic location category label is parsed and extracted from the early warning signal. The rule matching method is the same as S21. The second safety control rule is retrieved from the safety control rule base through the label index. This rule is a special rule adapted to the terrain that may be entered in the future. Its structure is completely consistent with the first safety control rule, except that the parameters are adapted to the risk characteristics of the corresponding terrain.

[0145] The reserve memory region is a high-speed memory block independent of the active memory region. It runs in parallel with the active memory region without interfering with it. Loading the second security control rule into the reserve region instead of directly overwriting the active rule ensures that the current job control based on the first security control rule is not affected. It also completes the loading and verification of the new rule in advance, avoiding control interruptions due to rule loading delays during subsequent switching. After loading, rule integrity verification is also required. If the verification fails, the reserve rule loading process will be terminated, the current active rule will be maintained, and a warning signal indicating failure will be issued. Preloading will only be triggered again after the prediction result meets the conditions again.

[0146] S24. Starting from the generation time of the category switching warning signal, within the predetermined transition time window, the continuous parameters in the current active rules and the preparatory rules are fused based on the linear interpolation algorithm to generate transition safety control rules.

[0147] Specifically, the transition time window is a preset fixed duration, whose value is comprehensively calibrated based on the agricultural machinery's driving speed, the width of the terrain transition area, and the actuator's response speed. This ensures that the agricultural machinery can smoothly complete the transition from the current terrain to the predicted terrain within the window duration. The value range can be 1 to 3 seconds and can be dynamically adjusted according to the specific agricultural machinery model and operating scenario. The start point of the time window precisely corresponds to the timestamp generated by the warning signal, and the end point is the start point superimposed with the transition time window duration. The proportion of time within the window is calculated in real time for interpolation calculations.

[0148] Continuous parameters refer to parameters in safety control rules that can be adjusted by gradients, including rated driving speed and reference braking pressure in basic operating parameters, weight elements in predefined mapping matrices, and sampling frequency in sensor monitoring schemes. Discrete parameters (such as sensor combination selection) do not participate in interpolation fusion and continue to use the parameters in the current active rules. After the transition window ends, they will be directly switched to the discrete parameters in the preparatory rules.

[0149] The core of the linear interpolation algorithm is to achieve a smooth transition between the two rule parameters based on the time proportion, as shown in the following formula:

[0150]

[0151] In the formula, For the first time window The fused parameter values ​​at time points. The corresponding parameter value in the current active rule. For the corresponding parameter values ​​in the preparatory rules, For the first The percentage of time in a given moment. The calculation formula is as follows:

[0152]

[0153] In the formula, For the current moment, The time when the warning signal is generated (the start point of the window). This is the total duration of the transition time window. The value range is [0, 1]. When hour, , The fusion parameters are completely identical to the parameters of the currently active rules; as time goes on, Gradually increase The weight gradually decreases. The weight gradually increases; when hour, , The fusion parameters are completely equivalent to the pre-defined rule parameters.

[0154] After performing the above interpolation operation on all continuous parameters, the discrete parameters and the fused continuous parameters are integrated to form a complete transition safety control rule. The transition rule is dynamically updated over time, and the fused parameters are recalculated once in each sampling period to ensure the continuity of parameter adjustment.

[0155] S25. During the transition time window, the transition security control rules shall be used as exclusive security control rules; at the end of the transition time window, the preparatory rules shall be used as exclusive security control rules.

[0156] Specifically, within the transition time window, dynamically updated transition safety control rules are distributed to the actuator module as dedicated safety control rules to achieve gradient adjustments of control parameters. For example, when agricultural machinery transitions from plain fields (high rated speed) to hilly terraces (low rated speed), the rated driving speed is gradually reduced from plain parameters to hilly parameters through linear interpolation, and the braking pressure benchmark is gradually adjusted from the plain value to the hilly value. This avoids sudden parameter changes that could lead to problems such as rapid deceleration and steering jerking, ensuring smooth operation.

[0157] After the transition time window ends, the percentage of time Once linear interpolation and fusion are complete, the transition rule generation process automatically terminates. The prepared rules in the prepared memory area are directly used as dedicated safety control rules, while the old rules in the prepared memory area are cleared to prepare for the next possible rule switch. If the warning signal fails before the window ends (e.g., the predicted probability is lower than the confidence threshold), the system immediately terminates the transition process, maintains the currently active rules, and deletes the prepared rules, ensuring that rule switching is only performed when the terrain change trend is stable.

[0158] After the rule switch is completed, the currently active rule is updated to the original reserve rule and reloaded into the active memory area to form a closed-loop control. This ensures that subsequent operation control is based on the rule execution adapted to the new terrain, achieving precise matching between terrain changes and control strategies.

[0159] In one optional embodiment, the dedicated safety control rules also include a dynamic safety threshold set for the agricultural machinery, which includes roll angle alarm threshold, pitch angle alarm threshold, slip ratio alarm threshold, and obstacle warning distance threshold.

[0160] Specifically, the dynamic safety threshold set, together with the sensor monitoring scheme, predefined mapping matrix, and basic operating parameters, constitutes a complete and dedicated rule system. It sets quantified thresholds for core safety risks in agricultural machinery operations, enabling accurate risk assessment. The dynamic safety threshold set adapts and adjusts according to changes in geographical location, ensuring that the thresholds match the risk characteristics of the corresponding terrain. It specifically includes four core threshold categories: roll angle alarm threshold, pitch angle alarm threshold, slip ratio alarm threshold, and obstacle warning distance threshold.

[0161] The roll angle alarm threshold is the critical roll angle value at which the agricultural machinery is at risk of tipping over. It is set to adapt to the slope characteristics of different terrains; the steeper the slope, the more stringent the threshold setting, to avoid the risk of tipping over. The pitch angle alarm threshold is the critical angle at which the agricultural machinery loses its pitch direction balance. It is set for undulating terrain to prevent accidents caused by the agricultural machinery tipping forward or backward. The slip ratio alarm threshold is the safe critical value of the wheel slip ratio. It is adapted to the ground adhesion coefficient of different terrains. The lower the adhesion coefficient (such as wetlands), the smaller the threshold setting, to avoid wheel slippage and loss of control. The obstacle warning distance threshold is the critical safe distance between the agricultural machinery and surrounding obstacles. It is set according to the terrain obstruction and the speed of the agricultural machinery. When the obstruction is severe or the speed is high, the threshold setting is larger to allow sufficient time for avoidance. All four types of thresholds are based on historical operation data and safety standards to ensure the accuracy of risk assessment.

[0162] Next, in this implementation, S5 includes the following steps:

[0163] S31. Generate a sensor selection mask vector according to the sensor monitoring scheme defined in the dedicated security control rules; based on the sensor selection mask vector, select a subset of target sensor data from the synchronous multi-source data stream.

[0164] Specifically, the sensor monitoring scheme has clearly defined the core sensor combination used for risk assessment under the corresponding terrain. The sensor selection mask vector is a binary vector digitized from this scheme. The vector dimension is consistent with the number of sensors included in the synchronous multi-source data stream. Each element in the vector corresponds to one sensor. An element value of 1 indicates that the corresponding sensor is a core sensor and its data must be retained; an element value of 0 indicates that the corresponding sensor is a redundant sensor and its data must be discarded. The mask vector enables rapid data filtering, reduces the amount of redundant data processing, and improves computational efficiency.

[0165] For example, if a synchronous multi-source data stream includes five types of sensors: GNSS, IMU, lidar, soil moisture sensor, and odometer, and the corresponding terrain-specific sensor monitoring scheme selects only IMU, lidar, and odometer as core sensors, then the sensor selection mask vector is [0,1,1,0,1]. During the selection process, the mask vector is mapped element-by-element to the sensor data sequence, retaining only the sensor data corresponding to the element value of 1, forming a subset of target sensor data. This subset of data precisely matches the core risk parameter measurement requirements, laying the foundation for subsequent parameter calculations.

[0166] S32. Based on a subset of target sensor data, calculate the current roll angle measurement, current pitch angle measurement, current slip ratio measurement, and current nearest obstacle distance measurement of the agricultural machinery.

[0167] Specifically, the four types of measurement values ​​correspond to the four types of thresholds in the dynamic safety threshold set, and are calculated based on a subset of target sensor data to ensure that the measurement results correspond to the risk assessment requirements.

[0168] The current roll angle and pitch angle measurements are calculated based on the fusion of IMU data and GNSS data from a subset of target sensor data. A complementary filtering algorithm is used to eliminate errors from a single sensor. The calculation formula is as follows:

[0169]

[0170]

[0171] In the formula, This is the current roll angle measurement. This is the current pitch angle measurement value. These are the filter coefficients. The roll angle data is directly output by the IMU. The pitch angle data is directly output by the IMU. The roll angle data is derived from GNSS elevation changes. The elevation angle data is derived from GNSS elevation changes.

[0172] The current slip ratio measurement is calculated based on odometer data from a subset of the target sensor data. The slip ratio is the proportion of the difference between the actual and theoretical wheel speed to the theoretical speed, calculated using the following formula:

[0173]

[0174] In the formula, This is the current slip ratio measurement. This is the theoretical speed of the wheel (calculated from the wheel speed measured by the odometer and the wheel diameter). The actual speed of the wheels (the speed of the agricultural machinery calculated from GNSS data).

[0175] The current nearest obstacle distance measurement is calculated based on LiDAR data from a subset of the target sensor data. The obstacle distance data collected by the LiDAR is sorted, and the minimum value is selected as the current nearest obstacle distance measurement, denoted as . At the same time, abnormal data (such as invalid data that is outside the detection range of the lidar or is interfered with) are removed to ensure measurement accuracy.

[0176] S33. The ratio of the current roll angle measurement value to the roll angle alarm threshold is taken as the first risk ratio, the ratio of the current pitch angle measurement value to the pitch angle alarm threshold is taken as the second risk ratio, the ratio of the current slip ratio measurement value to the slip ratio alarm threshold is taken as the third risk ratio, and the ratio of the current nearest obstacle distance measurement value to the obstacle warning distance threshold is taken as the fourth risk ratio.

[0177] Specifically, risk parameters of different dimensions and magnitudes are standardized by using risk ratios, transforming them into ratio indicators of a uniform magnitude. This facilitates the subsequent construction of risk state vectors and the determination of risk levels. The calculation formulas for the four types of risk ratios are as follows:

[0178]

[0179]

[0180]

[0181]

[0182] In the formula, The first risk ratio, The roll angle alarm threshold; This is the second risk ratio. The pitch angle alarm threshold; This is the third risk ratio. This is the slip ratio alarm threshold; This is the fourth risk ratio. This is the obstacle warning distance threshold.

[0183] in, , , The larger the value, the closer the corresponding risk is to the critical state. A value greater than 1 indicates that the safety threshold has been exceeded and there is a safety risk. The smaller the value, the closer the obstacle is and the higher the risk. A value less than 1 indicates that the safe distance threshold has been exceeded, requiring a warning to be triggered. This method quantifies and normalizes risk, providing standardized indicators for subsequent vector construction.

[0184] S34. Construct a real-time risk state vector based on the first risk ratio, the second risk ratio, the third risk ratio, and the fourth risk ratio.

[0185] Specifically, the real-time risk state vector is constructed using four types of risk ratios as core elements, arranged in a preset order. The vector dimension is fixed at 4 dimensions to ensure compatibility with the dimensions of the subsequent predefined mapping matrix. During construction, the validity of the four types of risk ratios must first be verified, and abnormal ratios that exceed reasonable ranges (such as maximum or minimum values ​​caused by sensor malfunctions) must be removed. Abnormal ratios are replaced with the corresponding ratios from the previous time step to ensure the integrity and reliability of the vector.

[0186] The mathematical expression for the real-time risk state vector is as follows:

[0187]

[0188] In the formula, This is a real-time risk state vector. , , , These are the risk ratios for the first to fourth risks, For vector transpose operations, row vectors are converted into column vectors to meet the requirements of subsequent matrix operations. The constructed real-time risk state vector comprehensively quantifies the four core risks of the agricultural machinery: roll, pitch, slip, and obstacle collision, providing accurate risk input for subsequent generation of control command correction quantities through predefined mapping matrices.

[0189] In an optional embodiment, S6 includes the following steps:

[0190] S41. Read the predefined mapping matrix corresponding to the current geographic location category label from the configuration file associated with the exclusive security control rule; in the predefined mapping matrix, each row corresponds to a control instruction type to be corrected, each column corresponds to a risk ratio in the real-time risk status vector, and the value of the matrix element represents the adjustment weight coefficient of the corresponding risk ratio on the corresponding control instruction type.

[0191] Specifically, in addition to core parameters and thresholds, the dedicated safety control rules are also associated with a configuration file. This file stores predefined mapping matrices corresponding to various geographic location categories. The matrix parameters are customized for the risk characteristics and control requirements of different terrains, ensuring that the weight allocation fits the actual operational scenario. The configuration file is stored in a structured format, using the current geographic location category label as an index, which allows for quick location and retrieval of the corresponding matrix, with retrieval efficiency meeting the timing requirements of real-time control.

[0192] The predefined mapping matrix is ​​a two-dimensional matrix, whose dimensions strictly match the number of control command types and the dimension of the real-time risk state vector. The control command types to be corrected include core types such as speed correction, steering angle correction, and braking pressure correction, corresponding to each row of the matrix; each risk ratio in the real-time risk state vector (such as roll, pitch, slip, and obstacle distance correlation ratios) corresponds to each column of the matrix. Matrix elements are adjustment weight coefficients, whose values ​​are determined by historical operational data, the degree of risk impact, and expert experience. Positive numbers indicate that when the risk ratio increases, the correction magnitude of the corresponding control command needs to be increased; negative numbers indicate that the correction magnitude needs to be decreased; and zero indicates that the risk has no impact on the corresponding control command. For example, the weight coefficient of the roll angle risk ratio for speed correction and braking pressure correction is usually positive; the higher the risk, the greater the deceleration and the stronger the braking protection need to be.

[0193] After reading the matrix, a validity check is required. This check whether the matrix dimensions match the number of control instruction types and the number of risk ratios, and whether the matrix elements are within a preset reasonable range. If any abnormality is found, a fault tolerance mechanism is triggered, and the valid matrix read at the previous moment is reused. At the same time, the abnormal information is recorded to ensure the continuous operation of the process.

[0194] S42. Based on the real-time risk state vector, calculate the difference vector between the real-time risk state vector and the vector with all elements having a value of 1; where each element in the difference vector represents the degree to which a risk ratio exceeds the corresponding safety threshold, and the number of elements in the vector with all elements having a value of 1 is the same as the number of elements in the real-time risk state vector.

[0195] Specifically, the risk ratios in the real-time risk state vector have been standardized. As mentioned earlier, a risk ratio greater than 1 indicates that the corresponding risk exceeds the safety threshold, while a ratio less than or equal to 1 indicates that it is within the safe range. The vector with all elements equal to 1 (hereinafter referred to as the all-1 vector) has the same dimension as the real-time risk state vector. The difference between the two can accurately quantify the degree of exceedance of each type of risk, forming a difference vector. The formula for calculating the difference vector is as follows:

[0196]

[0197] In the formula, It is a difference vector. This is a real-time risk state vector. The difference vector is composed entirely of 1s. The physical meaning of each element in the vector is clear: a positive element value indicates that the corresponding risk ratio exceeds the safety threshold; the larger the value, the greater the degree of exceedance, requiring targeted strengthening of control corrections. A zero or negative element value indicates that the risk is within the safe range, requiring no additional adjustment of control instructions due to this risk, and the correction magnitude can even be appropriately weakened to ensure operational efficiency. This calculation quantifies the degree of risk exceedance, providing accurate input for subsequent weight fusion.

[0198] S43. Perform an inner product calculation on each row vector of the predefined mapping matrix and the difference vector to obtain the inner product calculation result; use the inner product calculation result as the element value of the corresponding row in the preliminary control command correction amount.

[0199] Specifically, each row vector of the predefined mapping matrix corresponds to the weight allocation of a type of control instruction. By performing an inner product operation with the difference vector, the degree of risk exceedance and its corresponding weight can be integrated to obtain the preliminary correction amount for that type of control instruction, reflecting the comprehensive impact of all risks on a single control instruction. The formula for calculating the inner product is as follows:

[0200]

[0201] In the formula, For the first The initial correction amount for class control instructions. For the predefined mapping matrix, the first Line 1 Column weighting coefficients The difference vector is the first One element, This represents the dimension of the difference vector (i.e., the number of risk ratios). After performing the inner product operation row by row, all results are arranged in order of control command type, forming a preliminary control command correction vector. Each element in the vector corresponds to the preliminary correction magnitude of a type of control command.

[0202] For example, for a speed correction command, the inner product of its corresponding row vector and difference vector is used to integrate the degree of exceedance and corresponding weights of four types of risks: roll, pitch, slip, and obstacle distance, to obtain the initial speed correction amount. If both slip risk and obstacle risk significantly exceed the threshold and the corresponding weights are positive, the initial speed correction amount is a large negative value, indicating that a significant deceleration is required.

[0203] S44. Input the initial control command correction amount into a preset nonlinear activation function for amplitude limiting and filtering to generate the control command correction amount; wherein, amplitude limiting means limiting the correction amount that exceeds the preset saturation upper limit, and filtering means filtering out small correction amounts whose absolute value is lower than the preset dead zone threshold.

[0204] Specifically, the initial control command correction may exceed the actuator's range or experience minor fluctuations, requiring optimization using a nonlinear activation function to balance correction effectiveness and actuator stability. Preset nonlinear activation functions can include ReLU variants, Sigmoid variants, and other functions adapted to the control scenario, achieving both amplitude limiting and filtering functions, with the output matching the actuator's control range.

[0205] Limiting processing addresses corrections exceeding a preset saturation upper limit. This upper limit corresponds to the actuator's physical range; for example, the saturation upper limit for speed correction is the maximum permissible deceleration at rated speed, and for braking pressure correction, it's the difference between the maximum braking system pressure and the reference pressure. When the initial correction exceeds the saturation upper limit, the activation function clamps it to that limit; when it falls below the lower limit (usually zero or a negative upper limit), it clamps it to the lower limit to prevent excessive corrections from causing actuator overload or malfunction.

[0206] The filtering process targets minute corrections whose absolute values ​​are below a preset dead-zone threshold. This dead-zone threshold is set based on the actuator's minimum response accuracy. When the absolute value of the correction is below this threshold, the actuator cannot respond effectively and may even cause frequent movements due to minute fluctuations, shortening the equipment's lifespan. The activation function filters out these minute corrections to zero, retaining only the effective corrections exceeding the dead-zone threshold, thus achieving smooth optimization of the correction signal.

[0207] After processing by the nonlinear activation function, the output control command correction amount accurately matches the risk state and adapts to the actuator performance. It can be directly used for subsequent fusion with basic operation parameters to generate the final control command, ensuring the accuracy, stability and safety of agricultural machinery control.

[0208] The aforementioned agricultural machinery safety control method first collects the geographical location, attitude, and environmental data of the agricultural machinery simultaneously through multiple sensors, and extracts comprehensive feature vectors of terrain, surface, and motion from these data. Then, a pre-trained classification model is used to identify the specific geographical location category label of the agricultural machinery. Next, based on this category label, a dedicated safety control rule bound to it is intelligently retrieved from the rule base and loaded. This rule specifies the applicable sensor combination, the mapping matrix for risk quantification, and the basic operating parameters. Then, real-time data is filtered according to the sensor scheme in the rule, and a state vector representing the immediate risk of the agricultural machinery is calculated. This risk vector is then linearly transformed into a precise control command correction value using the pre-set mapping matrix in the rule. Finally, this correction value is fused with the basic operating parameters in the rule to generate and issue the final control command to the agricultural machinery actuator. This achieves agricultural machinery safety control based on precise environmental category perception and dynamic rule adaptation, effectively improving the safety and efficiency of operations in complex agricultural terrain.

[0209] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0210] Based on the same inventive concept, this application also provides an apparatus for implementing the agricultural machinery safety control method described above. The solution provided by this apparatus is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the agricultural machinery safety control apparatus provided below can be found in the limitations of the agricultural machinery safety control method described above, and will not be repeated here.

[0211] In one exemplary embodiment, such as Figure 3 As shown, an agricultural machinery safety control device 30 is provided to implement the methods in the above-described method embodiments. The device includes:

[0212] The multi-source data sensing module 31 is used to collect synchronous multi-source data streams of agricultural machinery through multiple types of sensors; wherein, the synchronous multi-source data streams include the geographical location data, attitude motion data and environmental perception data of agricultural machinery.

[0213] The environmental feature fusion module 32 is used to extract terrain geometric features, surface visual features, and agricultural machinery motion features based on synchronous multi-source data streams; and to construct a comprehensive feature vector based on the terrain geometric features, surface visual features, and agricultural machinery motion features.

[0214] The dynamic environment recognition module 33 is used to input the comprehensive feature vector into the pre-trained geographic location classification model, judge the category of the current environment of the agricultural machinery, and output the corresponding current geographic location category label.

[0215] The safety policy matching module 34 is used to retrieve the exclusive safety control rules corresponding to the current geographical location category label from the preset safety control rule library. The exclusive safety control rules include sensor monitoring schemes, predefined mapping matrices, and basic operating parameters. The sensor monitoring schemes define the sensor combinations selected for risk assessment under the current geographical location category corresponding to the current geographical location category label. The predefined mapping matrix defines the rules for mapping the real-time risk state vector of the agricultural machinery to the control command correction amount of the agricultural machinery under the current geographical location category.

[0216] The risk status quantification module 35 is used to select target monitoring data from synchronous multi-source data streams based on the sensor monitoring scheme in the exclusive safety control rules, and calculate the real-time risk status vector of the agricultural machinery based on the target monitoring data.

[0217] The instruction correction module 36 is used to perform a linear transformation on the real-time risk state vector according to the predefined mapping matrix in the dedicated safety control rules, and generate the control instruction correction amount.

[0218] The execution control module 37 is used to combine the control command correction amount with the basic operation parameters in the exclusive safety control rules to generate the final control command, and then send the final control command to the actuator of the agricultural machinery to control the agricultural machinery to perform the corresponding operation.

[0219] Embodiments of this application also provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the aforementioned method embodiments.

[0220] Embodiments of this application also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0221] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0222] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for controlling the safety of agricultural machinery, characterized in that, The method includes: S1. Collect synchronous multi-source data streams of agricultural machinery through multiple types of sensors; wherein, the synchronous multi-source data streams include the geographical location data, attitude motion data, and environmental perception data of the agricultural machinery; S2. Based on the synchronous multi-source data stream, extract terrain geometric features, surface visual features, and agricultural machinery movement features; construct a comprehensive feature vector based on the terrain geometric features, surface visual features, and agricultural machinery movement features; S3. Input the comprehensive feature vector into the pre-trained geographic location classification model to determine the category of the current environment of the agricultural machinery and output the corresponding current geographic location category label; S4. Based on the current geographic location category label, retrieve the exclusive safety control rule corresponding to the geographic location category label from the preset safety control rule base; wherein, the exclusive safety control rule includes a sensor monitoring scheme, a predefined mapping matrix, and basic operating parameters; the sensor monitoring scheme defines the sensor combination selected for risk assessment under the current geographic location category corresponding to the current geographic location category label; the predefined mapping matrix defines the rules for mapping the real-time risk state vector of the agricultural machinery to the control command correction amount of the agricultural machinery under the current geographic location category; S5. Based on the sensor monitoring scheme in the exclusive safety control rules, select target monitoring data from the synchronous multi-source data stream, and calculate the real-time risk status vector of the agricultural machinery according to the target monitoring data; S6. Based on the predefined mapping matrix in the exclusive safety control rules, perform a linear transformation on the real-time risk state vector to generate the control command correction amount; S7. Combining the control command correction amount with the basic operation parameters in the exclusive safety control rules, generate a final control command and send the final control command to the actuator of the agricultural machinery to control the agricultural machinery to perform the corresponding operation.

2. The method according to claim 1, characterized in that, S3 includes: S11. Based on the comprehensive feature vector at the current moment, calculate the probability of belonging to each preset geographical location category through the geographical location classification model, and output the preset geographical location category with the highest probability as the main category label; S12. Based on a time-series queue containing historical comprehensive feature vectors of the most recent N consecutive historical moments, the time-series queue is input into a pre-trained sequence prediction model to predict the feature vectors after the next M moments, thus obtaining the predicted feature vectors; wherein, the method of obtaining the historical comprehensive feature vectors is the same as that of the comprehensive feature vectors; N is a preset first positive integer, and M is a preset second positive integer; S13. Input the predicted feature vector into the geographic location classification model to obtain the predicted geographic location category label and the corresponding predicted probability; S14. When the predicted geographic location category label is the same as the main category label, the main category label is used as the current geographic location category label; S15. When the predicted geographic location category label is different from the main category label, and the predicted probability is lower than a preset confidence threshold, the main category label is used as the current geographic location category label. S16. When the predicted geographic location category label is different from the main category label, and the duration for which the predicted probability is higher than or equal to the preset confidence threshold does not exceed a preset time threshold, the main category label is used as the current geographic location category label. S17. When the predicted geographic location category label is different from the main category label, and the duration for which the predicted probability is higher than or equal to the preset confidence threshold exceeds a preset time threshold, the predicted geographic location category label is used as the current geographic location category label.

3. The method according to claim 2, characterized in that, Following S16, a category switching warning signal corresponding to the predicted geographic location category label is also generated; S4 includes: S21. Based on the current geographic location category label, a first security control rule corresponding to the current geographic location category label is matched in the security control rule base, and the first security control rule is loaded into the active memory area as the current active rule. S22. When no category switching warning signal is received, the current active rule is used as the exclusive security control rule; S23. When the category switching warning signal is received, according to the predicted geographical location category label corresponding to the category switching warning signal, a second security control rule corresponding to the predicted geographical location category label is matched in the security control rule base, and the second security control rule is loaded into the preparation memory area as a preparation rule. S24. Taking the generation time of the category switching warning signal as the starting point, within a predetermined transition time window, the continuous parameters in the current active rule and the preliminary rule are fused based on a linear interpolation algorithm to generate a transition safety control rule. S25. During the transition time window, the transition security control rule is used as the exclusive security control rule; at the end of the transition time window, the preliminary rule is used as the exclusive security control rule.

4. The method according to claim 1, characterized in that, The exclusive safety control rules also include a dynamic safety threshold set for agricultural machinery, which includes roll angle alarm threshold, pitch angle alarm threshold, slip ratio alarm threshold, and obstacle warning distance threshold. S5 includes: S31. Generate a sensor selection mask vector according to the sensor monitoring scheme defined in the exclusive security control rules; based on the sensor selection mask vector, select a subset of target sensor data from the synchronous multi-source data stream; S32. Based on the target sensor data subset, calculate the current roll angle measurement value, current pitch angle measurement value, current slip ratio measurement value, and current nearest obstacle distance measurement value of the agricultural machinery; S33. The ratio of the current roll angle measurement value to the roll angle alarm threshold is used as the first risk ratio, the ratio of the current pitch angle measurement value to the pitch angle alarm threshold is used as the second risk ratio, the ratio of the current slip ratio measurement value to the slip ratio alarm threshold is used as the third risk ratio, and the ratio of the current nearest obstacle distance measurement value to the obstacle warning distance threshold is used as the fourth risk ratio. S34. Construct the real-time risk state vector based on the first risk ratio, the second risk ratio, the third risk ratio, and the fourth risk ratio.

5. The method according to claim 4, characterized in that, S6 includes: S41. Read the predefined mapping matrix corresponding to the current geographic location category label from the configuration file associated with the exclusive security control rule; wherein, in the predefined mapping matrix, each row corresponds to a control instruction type to be corrected, each column corresponds to a risk ratio in the real-time risk status vector, and the value of the matrix element represents the adjustment weight coefficient of the corresponding risk ratio on the corresponding control instruction type. S42. Based on the real-time risk state vector, calculate the difference vector between the real-time risk state vector and the vector with all elements having a value of 1; wherein, each element in the difference vector represents the degree to which a risk ratio exceeds the corresponding safety threshold, and the number of elements in the vector with all elements having a value of 1 is the same as the number of elements in the real-time risk state vector. S43. Perform an inner product calculation on each row vector of the predefined mapping matrix and the difference vector to obtain the inner product calculation result; use the inner product calculation result as the element value of the corresponding row in the preliminary control command correction amount; S44. The preliminary control command correction amount is input to a preset nonlinear activation function for amplitude limiting and filtering to generate the control command correction amount; wherein, the amplitude limiting means limiting the correction amount that exceeds the preset saturation upper limit, and the filtering means filtering out small correction amounts whose absolute value is lower than the preset dead zone threshold.

6. An agricultural machinery safety control device for implementing the method according to any one of claims 1 to 5, characterized in that, The device includes: The multi-source data sensing module is used to collect synchronous multi-source data streams of agricultural machinery through multiple types of sensors; wherein, the synchronous multi-source data streams include the geographical location data, attitude motion data, and environmental perception data of the agricultural machinery; The environmental feature fusion module is used to extract terrain geometric features, surface visual features, and agricultural machinery motion features based on the synchronous multi-source data stream; and to construct a comprehensive feature vector based on the terrain geometric features, surface visual features, and agricultural machinery motion features. The dynamic environment recognition module is used to input the comprehensive feature vector into the pre-trained geographic location classification model, determine the category of the current environment of the agricultural machinery, and output the corresponding current geographic location category label. The safety policy matching module is used to retrieve a specific safety control rule corresponding to the current geographical location category label from a preset safety control rule library. The specific safety control rule includes a sensor monitoring scheme, a predefined mapping matrix, and basic operating parameters. The sensor monitoring scheme defines the sensor combinations selected for risk assessment under the current geographical location category corresponding to the current geographical location category label. The predefined mapping matrix defines the rules for mapping the real-time risk state vector of the agricultural machinery to the control command correction amount of the agricultural machinery under the current geographical location category. The risk status quantification module is used to select target monitoring data from the synchronous multi-source data stream based on the sensor monitoring scheme in the exclusive safety control rules, and calculate the real-time risk status vector of the agricultural machinery based on the target monitoring data; The instruction correction module is used to perform a linear transformation on the real-time risk state vector according to the predefined mapping matrix in the exclusive safety control rules, and generate the control instruction correction amount; The execution control module is used to combine the control command correction amount with the basic operation parameters in the exclusive safety control rules to generate a final control command, and send the final control command to the actuator of the agricultural machinery to control the agricultural machinery to perform the corresponding operation.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 5.

Images