Intelligent agriculture integrated control method and system based on edge computing

Abstract

The invention discloses an intelligent agriculture integrated control method and system based on edge calculation, and belongs to the field of mechanical automation control, and the method comprises the steps: building a dynamic twin body, quantifying a response deficit between an expected vehicle body attitude angle change rate and an actual attitude angle change rate under a driving instruction, and defining the response deficit as an observation signal; a control target is redefined based on the observation signal, the control target is set to track a stable manifold, the stable manifold is defined by a preset observation signal amplitude boundary, and the state of the target agricultural machine is controlled to be maintained in the stable manifold; monitoring the steering jerk of the target agricultural machine; according to the method, dynamic twinborn quantitative response deficit characters are constructed, observation signals are defined, absolute trajectory accurate tracking is replaced with stable manifold as a control target, physical disturbance in a dynamic farmland environment is effectively compensated, agricultural machine path tracking errors are reduced, and the operation problems of seedling pressing, seedling damage and the like are reduced.

Description

Smart Agriculture Integrated Control Method and System Based on Edge Computing Technical Field

[0001] This invention relates to the field of mechanical automation control technology, and more specifically, to an integrated control method and system for smart agriculture based on edge computing. Background Technology

[0002] In agricultural operations, agricultural machinery faces complex operating conditions. Existing agricultural machinery operation control systems mainly rely on serial timing logic based on perception, decision-making, and control to make operational decisions. When faced with dynamically changing agricultural environments, such as soft foundations in paddy fields, the soil shear force changes continuously, causing the machinery to skid during movement. Traditional trajectory tracking algorithms based on fixed motion models cannot compensate for such physical disturbances in real time, resulting in large path tracking errors and causing problems such as crushed seedlings and seedling damage. Secondly, the common interferences in the agricultural environment, such as dust, water mist, and crop leaf obstruction, seriously affect the perception stability of visual sensors and lidar, leading to unreliable environmental recognition results and consequently affecting the accuracy of subsequent decision-making and control.

[0003] The common cause of the above technical problems is that the existing control architecture treats dynamically changing physical parameters in the farmland environment, such as soil rheological properties and transient visual occlusion, as external disturbances and attempts to overcome these problems through complex models and higher-frequency corrections. This approach leads the system to continuously pursue an absolutely accurate but not real-time environmental representation and generate control commands based on it. In scenarios where agricultural machinery operates at high speed and ground physical parameters change dynamically, the excessive pursuit of absolute accuracy in trajectory tracking will force the control system to perform high-frequency and high-load compensation calculations, exacerbate command delays, and ultimately damage the instantaneous stability of the system under dynamic disturbances. Summary of the Invention

[0004] To address the problems existing in the prior art, the present invention aims to provide an integrated control method and system for smart agriculture based on edge computing. This method can quantify the response deficit and define the observation signal by constructing a dynamic twin, and use a stable manifold as the control target to replace the absolute trajectory for precise tracking. This effectively compensates for physical disturbances in the dynamic farmland environment, reduces the tracking error of agricultural machinery, and reduces operational problems such as seedling damage.

[0005] To address the aforementioned problems, the present invention adopts the following technical solution: Firstly, a smart agriculture integrated control method based on edge computing, comprising: establishing a dynamic twin; quantifying the response deficit between the expected rate of change of the vehicle's attitude angle under a driving command and the actual rate of change of the attitude angle, and defining this response deficit as an observation signal; redefining the control target based on the observation signal, setting the target to track a stable manifold, which is defined by a preset amplitude boundary of the observation signal, and controlling the target agricultural machinery to maintain its state within the stable manifold; monitoring the steering abruptness of the target agricultural machinery, and when the steering abruptness exceeds a threshold calculated in real time from the amplitude of the observation signal, indicating a strong physical interaction event. The system triggers a time window; within the trigger time window, it performs directional and parameter-based collaborative sampling, determines the overlap between the physical envelope determined by the target agricultural machinery's state and the set of inaccessible points in the sampling point cloud, and outputs an instantaneous passability flag; it couples the instantaneous passability flag with a strong physical interaction event type, and performs operations based on a preset response primitive mapping relationship; it asynchronously stores the event point coordinates, observation signal features, and event type into a circular queue; when the target agricultural machinery is determined to have entered a stable state based on the observation signal, it extracts the marker points and constructs a physical interaction feature layer; it uses the physical interaction feature layer to generate a reference path and a priority for subsequent directional and parameter-based collaborative sampling in the corresponding area.

[0006] Furthermore, the generation of the observation signal includes: establishing a basic mapping table based on the driving command and the expected steady-state vehicle attitude angle change rate, introducing virtual soil mechanical parameters, and expanding the basic mapping table into response curve templates with different parameter values ​​through a parameterized nomograph to form a pre-set response template library; acquiring the instantaneous vehicle attitude angle change rate sequence, extracting segments of the sequence within a sliding window, performing shape matching between the segments and the response curve templates of the corresponding driving commands in the pre-set response template library, selecting a virtual working condition reference from the response curve templates based on the consistency rate of change trend and the closeness of energy amplitude, and calculating the residual vector between the segments and the virtual working condition reference; performing short-time window energy integration on the residual vector to obtain the energy envelope, performing piecewise linearization approximation on the energy envelope to identify its rising edge, plateau period, and falling edge, and encoding and setting the slope of the rising edge, the average energy of the plateau period, and the duration of the falling edge as the observation signal.

[0007] Furthermore, the definition of the preset observation signal amplitude boundary includes: cutting historical observation signals into fixed-duration signal segments, calculating the statistical characteristics of each segment and, based on density clustering, back-mapping the clustering region boundary to generate offline primitive boundary clusters; fragmenting and calculating statistical characteristics of real-time observation signals, matching the obtained feature vectors with the feature centers of the offline primitive boundary clusters, selecting the top K primitive boundaries with the highest confidence values ​​based on the matching confidence, and taking their upper and lower envelopes as dynamic boundary envelopes; monitoring the state trend of the target agricultural machinery, querying the preset elastic adjustment function according to the trend to adjust the width of the dynamic boundary envelope, and generating the final stable manifold boundary.

[0008] Furthermore, maintaining the target agricultural machinery state within the stable manifold includes: constructing a virtual dynamic attractor on the stable manifold, the attractor moving along the tangent direction of the stable manifold, its normal position being determined by solving a dynamic equilibrium equation in real time; wherein, the dynamic equilibrium equation is the balance between the current amplitude of the observed signal and a nonlinear restoring force, the nonlinear restoring force being a function of the distance of the attractor from the center of the manifold; constructing an asymmetric virtual potential energy field between the target agricultural machinery state and the virtual dynamic attractor, calculating the virtual potential energy gradient of the target agricultural machinery state in the potential energy field, and generating the change in driving command based on the negative component of the virtual potential energy gradient and the sine of the angle between it and the state motion direction.

[0009] Furthermore, based on strong physical interaction events and triggering time windows, the process includes: performing adaptive multi-scale decomposition on the steering abruptness signal to obtain intrinsic mode components and trend remainders; extracting the rising slope and overshoot peak value of the energy envelope from the first intrinsic mode component; and extracting the sign and linear fitting slope from the trend remainder; when the rising slope exceeds a dynamic threshold of exponential scaling of the observed signal amplitude, and the sign of the trend remainder is different from the sign of the overshoot peak value, it is determined to be a strong physical interaction event, and the length of the time window corresponding to the strong physical interaction event is set according to the ratio between the absolute value of the overshoot peak value and the absolute value of the linear fitting slope.

[0010] Furthermore, the instantaneous accessibility flag is output, including: generating a hierarchical scanning sequence based on the state and motion manifold prediction of the target agricultural machinery; the hierarchical scanning sequence is used to define the focal area and risk buffer layer, and to allocate scanning resource ratios to the focal area and risk buffer layer; based on the predicted spatiotemporal region covered by the physical envelope, an evidence field is initialized, and scanning is performed according to the hierarchical scanning sequence. The spatial points obtained in the scan are classified into collision evidence or risk field strength values ​​according to their positional relationship with the physical envelope and mapped to the evidence field; when the scanning resources in the focal area are exhausted, the instantaneous accessibility flag is output according to whether there is collision evidence in the evidence field or whether its average risk field strength is lower than a preset risk threshold.

[0011] Furthermore, operations are performed based on the pre-defined primitive mapping relationship, including: indexing the main strategy fragment and auxiliary strategy fragment from the strategy fragment library according to the type of strong physical interaction event and the instantaneous accessibility flag, and determining their initial weights based on the historical performance scores of the strategy fragments; adjusting the action intensity parameters of the strategy fragments online based on impedance sensing, and mixing the adjusted instructions of the main strategy fragment and auxiliary strategy fragment according to the initial weights to generate composite control instructions.

[0012] Furthermore, the priorities for performing directional and parameter-based collaborative sampling include: dynamic clustering based on event point coordinates, occurrence timestamps, and observed signal characteristics to generate scene objects that include spatial boundaries, time windows, and observed signal characteristics, as well as statistical distribution and statistics of major event types; mining causal relationships based on the spatiotemporal order and spatial adjacency of scene objects; generating virtual repulsion fields for path optimization based on the determination results of causal relationships; and assigning higher perception priority to the corresponding regions.

[0013] Furthermore, the priority for performing directional and parameter-based collaborative sampling also includes: adjusting the field amplitude of the virtual repulsion field and giving it a higher perception priority based on the comparison results between actual events and expectations in subsequent operations, and integrating new event points into scene objects to update their boundaries and feature distributions.

[0014] Secondly, this invention also provides an integrated smart agriculture control system based on edge computing, comprising: an observation generation module, which establishes a dynamic twin, quantifies the response deficit between the expected rate of change of vehicle attitude angle under driving commands and the actual rate of change of attitude angle, and defines the response deficit as an observation signal; a manifold tracking module, which redefines the control target based on the observation signal, sets the control target to track a stable manifold, which is defined by a preset amplitude boundary of the observation signal, and controls the target agricultural machinery to maintain its state within the stable manifold; and an event detection module, which monitors the steering abruptness of the target agricultural machinery, and when the steering abruptness exceeds a threshold calculated in real time from the amplitude of the observation signal, it marks it as a strong physical interaction event and triggers a response. The system comprises the following modules: a time window; a passage assessment module, which performs directional and parameter-based collaborative sampling within the trigger time window, determines the overlap between the physical envelope determined by the target agricultural machinery's state and the set of traversable points in the sampling point cloud, and outputs an instantaneous passageability flag; a response mapping module, which couples the instantaneous passageability flag with strong physical interaction event types and performs operations based on preset response primitive mapping relationships; and a path planning module, which asynchronously stores event point coordinates, observation signal characteristics, and event types into a circular queue. When the target agricultural machinery is determined to have entered a stable state based on the observation signal, it extracts marker points and constructs a physical interaction feature layer. The physical interaction feature layer is used to generate a reference path and a priority for subsequent directional and parameter-based collaborative sampling in the corresponding area.

[0015] Compared with the prior art, the beneficial effects of the present invention are: (1) This solution quantifies the response deficit by constructing a dynamic twin and defines the observation signal, and uses a stable manifold as the control target to replace the absolute trajectory for precise tracking, effectively compensating for physical disturbances in the dynamic farmland environment, reducing the tracking error of agricultural machinery path, and reducing operational problems such as seedling crushing and seedling damage.

[0016] (2) This scheme combines hierarchical scanning sequences and evidence field evaluation to output instantaneous accessibility flags, and with the accurate judgment of strong physical interaction events and time window triggering mechanism, improves the reliability of sensing data under complex interference environment.

[0017] (3) Based on the pre-set response primitive mapping relationship, this scheme can quickly respond to strong physical interaction events by dynamically allocating the weights of the main and auxiliary strategy segments and adjusting the action intensity online. It does not require high-frequency and high-load compensation calculations, thus ensuring the instantaneous stability and operation continuity of agricultural machinery under dynamic disturbances.

[0018] (4) This scheme generates scene objects by dynamic clustering and mines causal relationships, constructs a physical interaction feature layer for reference path generation and sampling priority allocation, realizes early avoidance of high-risk areas and reasonable tilting of perception resources, and improves the overall operation efficiency and environmental adaptability of agricultural machinery. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0020] Figure 1 is a flowchart of the integrated control method for smart agriculture based on edge computing according to the present invention; Figure 2 is a flowchart of the various modules in the integrated control system for smart agriculture based on edge computing according to the present invention. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0022] Example 1: Please refer to Figure 1. The integrated control method for smart agriculture based on edge computing includes the following steps: Step 1, establish a dynamic twin, quantify the response deficit between the expected rate of change of vehicle body attitude angle under the drive command and the actual rate of change of attitude angle, and define this response deficit as the observation signal. The specific operation is as follows: When establishing the dynamic twin, firstly, based on the physical structural parameters of the target agricultural machinery, including the body frame size, suspension system stiffness parameters, wheel diameter and ground contact area parameters, etc., and the power transmission characteristics, including engine output torque characteristics, gearbox transmission ratio range, drive axle response delay characteristics, etc., combined with the dynamic response law of agricultural machinery in farmland operation scenarios, a digital twin model that accurately maps to the actual agricultural machinery attitude motion is constructed. This dynamic twin is constructed around the response relationship between the drive command and the rate of change of vehicle body attitude angle, and can completely reproduce the expected evolution process of the rate of change of agricultural machinery body attitude under different drive command inputs, such as roll angle and pitch angle. The system uses dynamic twins to simulate and calculate the actual motion state of agricultural machinery, including steering angle and heading angle. For each drive command received during operation, including steering angle and speed adjustment commands, a dynamic twin performs real-time simulation calculations, outputting the time series of the expected vehicle attitude angle change rate corresponding to the command. Simultaneously, attitude sensors deployed on the agricultural machinery, such as inertial measurement units, can collect data on the vehicle attitude angle change rate during actual operation in real time, ensuring that the data acquisition frequency is synchronized with the twin simulation calculation frequency. The expected vehicle attitude angle change rate output by the dynamic twin is compared with the actual vehicle attitude angle change rate collected by the sensors on a time-by-time basis. By calculating the numerical difference and cumulative deviation at the same time point, the response deficit generated during the drive command response is quantified. This response deficit essentially reflects the degree of deviation between the actual agricultural machinery motion state and the ideal response state. This response deficit is formally defined as an observation signal used for state assessment in subsequent control processes.

[0023] The generation of the observed signal also includes the following steps: Step 11, establishing a basic mapping table based on the driving command and the expected steady-state vehicle attitude angle change rate, introducing virtual soil mechanics parameters, and expanding the basic mapping table into response curve templates with different parameter values ​​through a parameterized nomograph to form a pre-set response template library. The specific operation is as follows: First, sort out the various common driving commands of the target agricultural machinery, including steering commands, driving speed adjustment commands, etc. For each type of driving command, obtain the expected vehicle attitude angle change rate data corresponding to the command in the steady-state operation stage through multiple standard working condition tests. Based on these data, establish a basic mapping table between the driving command and the expected steady-state vehicle attitude angle change rate. This mapping table can directly reflect the target... The correspondence between the two under quasi-conditions is established. Considering the impact of dynamic changes in soil mechanical properties on the attitude response of agricultural machinery during farmland operations, virtual soil mechanical parameters are introduced. These parameters cover indicators such as soil cohesion, internal friction angle, and shear modulus, and are used to simulate the effects of different soil conditions on the motion response of agricultural machinery. Through the parametric nomograph, an engineering chart tool, each set of correspondences in the basic mapping table is expanded according to different values ​​of the virtual soil mechanical parameters to generate multiple sets of response curve templates with differentiated parameter configurations. Each curve template corresponds to the response law of the vehicle attitude angle change rate under a specific driving command and a specific combination of virtual soil mechanical parameters. All generated response curve templates are sorted and classified to form a complete pre-set response template library.

[0024] Step 12: Obtain the instantaneous vehicle attitude angle change rate sequence, extract segments of the sequence within a sliding window, perform shape matching between the segments and the response curve templates of the corresponding drive commands in the preset response template library, select a virtual working condition reference from the response curve templates based on the consistency rate of change trend and the closeness of energy amplitude, and calculate the residual vector between the segment and the virtual working condition reference. The specific operations are as follows: Using attitude sensing devices deployed on the agricultural machinery, continuously collect instantaneous vehicle attitude angle change rate data during operation to form a continuous instantaneous vehicle attitude angle change rate sequence. This sequence records the fluctuations in the attitude angle change rate of the agricultural machinery over time during operation. Set a sliding window of fixed duration, and extract segments of the instantaneous vehicle attitude angle change rate sequence according to the length of the sliding window. The sliding step size of the sliding window is preset according to the dynamic characteristics of the working scenario to ensure that the extracted signal segments can completely reflect the short-term... The response characteristics of the agricultural machinery's attitude within the interval are analyzed. For each captured signal segment, all response curve templates corresponding to the driving command received by the agricultural machinery are selected from the preset response template library. The signal segment is then matched with these curve templates. During the shape matching process, the consistency rate of the change trend and the proximity of the energy amplitude between the signal segment and each template are calculated. The consistency rate of the change trend is obtained by comparing the slope change patterns of the two curves, and the proximity of the energy amplitude is determined by calculating the proportion of the difference between the two signal energies. Based on the calculation results, the response curve template with the highest consistency rate of the change trend and the best proximity of the energy amplitude is selected as the virtual working condition reference for the current working condition. Subsequently, the difference between the value of each moment in the signal segment and the corresponding moment value of the virtual working condition reference is calculated. These differences are arranged in chronological order to form a residual vector, which intuitively reflects the instantaneous deviation between the actual response and the reference response.

[0025] Step 13: Perform short-window energy integration on the residual vector to obtain the energy envelope. Approximate the energy envelope using piecewise linearization to identify its rising edge, plateau period, and falling edge. Encode the slope of the rising edge, the average energy of the plateau period, and the duration of the falling edge as the observed signal. Specifically, the following steps are performed: For the residual vector obtained in Step 12, use the short-window energy integration method for signal processing. Set a fixed integration window duration, and perform square integration on the values ​​of the residual vector within each short window to obtain the energy value corresponding to each short window. Connect the energy values ​​of all short windows in chronological order to form the energy envelope of the residual vector. This envelope highlights the energy variation characteristics of the residual signal and weakens the influence of instantaneous interference noise. Subsequently, perform piecewise linearization approximation on the energy envelope. The least squares method is used to fit the energy envelope piecewise. Based on the slope changes and energy value fluctuation range of the fitted line, three characteristic stages of the energy envelope are identified: the rising edge, the plateau period, and the falling edge. The rising edge stage is the interval where the slope of the energy envelope is positive and continuously increasing; the plateau period stage is the interval where the slope of the energy envelope is close to zero and the energy value is stable within a certain range; and the falling edge stage is the interval where the slope of the energy envelope is negative and continuously decreasing. For the three identified characteristic stages, key characteristic parameters are extracted, namely the slope of the rising edge, the average energy value of the plateau period, and the duration of the falling edge. After normalizing these three characteristic parameters, they are combined and encoded according to a preset encoding rule to form the final observation signal. This encoded observation signal can comprehensively and concisely reflect the variation law of residual energy.

[0026] In a preferred embodiment of the present invention, step 2 is further included: redefining the control target based on the observed signal, setting the control target to track a stable manifold, which is defined by a preset amplitude boundary of the observed signal, and controlling the target agricultural machinery state to remain within the stable manifold. The specific operation is as follows: after acquiring the defined observed signal, the traditional absolute trajectory precise tracking is no longer used as the control target. Instead, the control target is redefined based on the dynamic characteristics of the observed signal. The observed signal can reflect the changing pattern of the drive command response deficit, and its amplitude fluctuation is directly related to the degree of deviation between the agricultural machinery state and the ideal response, as well as the influence intensity of environmental disturbances. Setting the control target to track a stable manifold, the boundary of which is determined by the preset amplitude boundary of the observed signal, essentially defines the stable operating range of the agricultural machinery state by defining the reasonable fluctuation range of the observed signal. The entire control process revolves around maintaining the target agricultural machinery state within the stable manifold. By monitoring the relative position of the observed signal amplitude and the manifold boundary in real time, the control strategy is dynamically adjusted to ensure that the agricultural machinery always maintains a stable operating state in the dynamic farmland environment, avoiding control delay and system instability caused by excessive pursuit of absolute precision.

[0027] The definition of the preset observation signal amplitude boundary also includes the following steps: Step 21, cut the historical observation signal into fixed-duration signal segments, calculate the statistical characteristics of each segment, and based on density clustering, reverse map the clustering region boundary to generate offline primitive boundary clusters. The specific operation is as follows: First, collect the observation signal data generated by the target agricultural machinery under various historical operation scenarios. These scenarios include typical working conditions such as different soil conditions, crop types, and operation speeds to ensure the comprehensiveness and representativeness of the historical data. Cut the collected historical observation signals evenly according to a fixed duration to form several signal segments of the same length. The fixed duration is determined based on the dynamic response cycle of agricultural machinery operation to ensure that each segment can completely reflect a continuous operation. The system observes the changing characteristics of signals during operation; it calculates the statistical characteristics of each signal segment, including key parameters such as the mean, variance, peak value, valley value, and slope of the observed signal within the segment. These parameters can quantify the overall characteristics and fluctuation patterns of the segment; it uses a density clustering algorithm to perform cluster analysis on the statistical feature vectors of all signal segments, and divides different clustering regions according to the spatial distribution density of the feature vectors. Each clustering region corresponds to a type of operation with similar observed signal characteristics. The boundaries of each clustering region are reverse-mapped according to the dimension of the observed signal amplitude, that is, the clustering boundary of the feature space is transformed into the value boundary of the observed signal amplitude. After all the mapped amplitude boundaries are sorted and classified, offline primitive boundary clusters are formed.

[0028] Step 22: Fragment the real-time observed signal and calculate statistical features. Match the obtained feature vector with the feature centers of the offline primitive boundary clusters. Based on the matching confidence, select the top K primitive boundaries with the highest confidence values, and take their upper and lower envelopes as the dynamic boundary envelopes. The specific operation is as follows: During the real-time operation of agricultural machinery, the continuously generated observed signal is fragmented according to the same fixed duration and sliding step size as in Step 21 to ensure that the time scale of the real-time signal fragments is consistent with that of the historical signal fragments, thus ensuring the comparability of feature calculations. For each fragmented real-time signal, the same calculation method as in Step 21 is used to extract its mean, variance, peak value, and other statistical features to form a real-time feature vector. Extract the feature center corresponding to each primitive boundary from the offline primitive boundary cluster. The feature center is the mean of the statistical feature vectors of all historical signal fragments within the cluster to which the primitive boundary belongs. The vector represents the typical characteristics of the observed signal under this type of working condition. The real-time feature vector is matched with the feature center of each primitive boundary for similarity calculation. The matching process uses common matching algorithms such as cosine similarity or Euclidean distance. The matching confidence of each primitive boundary is calculated based on the matching results. The higher the confidence value, the closer the working condition corresponding to the primitive boundary is to the current real-time working condition. The primitive boundaries are sorted from high to low according to the matching confidence value. The top K primitive boundaries with the highest confidence are selected. The value of K is preset according to the complexity of the working scenario. Usually, 3 to 5 are selected to balance the comprehensiveness of the reference and the computational efficiency. The upper and lower limits of the amplitude of the selected top K primitive boundaries are respectively enveloped. That is, the maximum value of the upper limit of the amplitude of all primitive boundaries is taken as the overall upper limit, and the minimum value of the lower limit of the amplitude of all primitive boundaries is taken as the overall lower limit, forming a dynamic boundary envelope. This envelope can adapt to the fluctuation characteristics of the observed signal under the current real-time working condition.

[0029] Step 23: Monitor the state trend of the target agricultural machinery. Based on the trend, query the preset elastic adjustment function to adjust the width of the dynamic boundary envelope and generate the final stable manifold boundary. The specific operation is as follows: Through the inertial measurement unit, speed sensor, steering angle sensor and other equipment deployed on the agricultural machinery, collect state parameters such as the agricultural machinery's driving speed, steering angle, roll angle and pitch angle in real time. The collection frequency is consistent with the observation signal processing frequency to ensure the time synchronization of state data and signal data. Based on the collected state parameters, construct a state trend identification index system. This system includes three indicators: the rate of change of state parameters, the amplitude of change and the number of consecutive fluctuations. The rate of change is calculated by the ratio of the difference between parameters at adjacent times to the time interval. The amplitude of change is calculated by the difference between the maximum and minimum values ​​of the parameters. The number of consecutive fluctuations is determined by the continuous duration of the statistical parameters deviating from the steady-state threshold. According to the above indicators, the state trend of the agricultural machinery is divided into three typical types: stable operation, slow fluctuation and drastic change. Stable operation corresponds to all indicators being in the preset steady-state range. Slow fluctuation corresponds to the small rate and amplitude of change of indicators and the few fluctuations. Dramatic change corresponds to the rate or amplitude of change of indicators exceeding the threshold and the number of consecutive fluctuations meeting the standard.

[0030] A pre-defined elastic adjustment function library is provided, which configures dedicated elastic adjustment functions for three types of state trends. The function's purpose is to quantitatively adjust the width of the dynamic boundary envelope based on the dynamic characteristics of the state trend. The adjustment principle is that the more stable the state, the narrower the boundary width to ensure control accuracy; the more volatile the state, the wider the boundary width to improve anti-interference capability. All elastic adjustment functions adopt a piecewise function form. For stable operating trends, a linear decay adjustment function is configured, expressed as W=Wo-Ks*α, where W is the boundary width adjustment amount, Wo is the initial adjustment reference value, Ks is the stable trend adjustment coefficient, and α is the state stability coefficient, obtained by normalizing the fluctuation amplitude of the state parameters, with a value range of 0 to 1. For slow fluctuation trends, a linear increasing adjustment function is configured, expressed as W=Wo-Km*β, where Km is the slow fluctuation adjustment coefficient, and β is the fluctuation intensity coefficient, obtained by normalizing the change rate of the state parameters, with a value range of 0 to 1. For volatile changing trends, an exponential growth adjustment function is configured, expressed as W equals Wo multiplied by e. Kv*rWhere Kv is the drastic change adjustment coefficient, and r is the fluctuation severity coefficient, which is obtained by normalizing the product of the number of consecutive fluctuations and the change amplitude, and the value ranges from 0 to 1; in the above function, Wo, Ks, Km, and Kv are preset parameters, which are calibrated through multiple sets of farmland operation experiments. The calibration process aims at the stability and control accuracy of agricultural machinery operation under different state trends. The optimal parameter values ​​are determined by orthogonal experiment method and stored in the function library; during real-time monitoring, the state trend type of the current agricultural machinery is first determined by the state trend identification index system, and then the corresponding adjustment function is called from the elastic adjustment function library. The real-time calculated state trend characteristic parameters α, β, or r are substituted into the function to calculate the specific boundary width adjustment amount; based on the adjustment amount, the width of the dynamic boundary envelope generated in step 22 is corrected. During the correction process, the center position of the boundary envelope is kept consistent with the dynamic boundary envelope. The width optimization is achieved only by adjusting the amplitude span of the upper and lower boundaries. After the correction is completed, the final stable manifold boundary is generated. This boundary can dynamically adapt to the real-time operating status of the agricultural machinery and can also control the accuracy and stability of the system.

[0031] The control of the target agricultural machinery's state to remain within the stable manifold also includes the following steps: Step 24, constructing a virtual dynamic attractor on the stable manifold. This attractor moves along the tangent direction of the stable manifold, and its normal position is determined by solving a dynamic equilibrium equation in real time. The dynamic equilibrium equation is the balance between the current amplitude of the observed signal and a nonlinear restoring force, where the nonlinear restoring force is a function of the attractor's distance from the manifold center. Specifically, a virtual dynamic attractor is constructed on the stable manifold generated in step 23. This attractor is not a fixed point but moves along the tangent direction of the stable manifold over time. The moving speed matches the agricultural machinery's travel speed and operating rhythm, ensuring that the attractor is always on a reasonable operating trajectory within the stable manifold. The normal position of the attractor, i.e., the position perpendicular to the manifold tangent direction, is determined by... The dynamic equilibrium equation is determined through real-time solution. This equation balances the current amplitude of the observed signal with the nonlinear restoring force. The derivation logic of the dynamic equilibrium equation is based on the principle of force balance. The current amplitude of the observed signal reflects the disturbance intensity of the agricultural machinery's state deviating from the manifold center, and the nonlinear restoring force is used to pull the agricultural machinery back to the manifold center. When the two are balanced, the normal position of the attractor can keep the agricultural machinery stable. This dynamic equilibrium equation can be expressed as: A(t) = F(d), where A(t) represents the current amplitude of the observed signal, which is a real-time parameter that changes with time t; d represents the normal distance of the virtual dynamic attractor from the stable manifold center; F(d) represents the nonlinear restoring force function. Combining the requirements of stable control of the agricultural machinery state, an exponentially saturated nonlinear function is adopted. The specific expression is: F(d) equals k1 multiplied by 1 in parentheses minus e. -K2*dThis expression, based on the constraints of agricultural machinery conditions, adopts an exponential growth and saturation form to avoid abrupt changes in the initial stage of the restoring force and overload in the later stage. This ensures both a smooth adaptation of the restoring force when the deviation is small and a rapid increase in the restoring force to stabilize the state when the deviation is large, while also conforming to the logic of force balance. Here, k1 is the maximum saturation amplitude of the nonlinear restoring force, calibrated by the rated output capacity of the agricultural machinery power system; K2 is the restoring force growth rate coefficient, preset according to the disturbance intensity of the farmland operation scenario. This function satisfies the characteristic of monotonically increasing with the input parameter d. The farther the attractor deviates from the manifold center, the greater the nonlinear restoring force, effectively pulling the attractor back to a reasonable position. The current amplitude A(t) of the observed signal is collected in real time and substituted into the dynamic equilibrium equation A(t) composed of the above specific expression, equals 1 minus e (k1). -K2*d By inversely solving the equation, we can obtain that d is equal to the negative K2 numerator multiplied by the natural logarithm function, where the parameter of the logarithm function is 1 minus A(t) divided by k1, thus determining the real-time normal position of the attractor.

[0032] Step 25: Construct an asymmetric virtual potential energy field between the target agricultural machinery state and the virtual dynamic attractor. Calculate the virtual potential energy gradient of the target agricultural machinery state in the potential energy field. Based on the negative component of the virtual potential energy gradient and the product of its angle with the direction of motion, generate the change in driving command. The specific operation is as follows: Based on the real-time state parameters of the target agricultural machinery, including vehicle attitude angle, driving speed, steering angle, etc., and the real-time position parameters of the virtual dynamic attractor, construct a virtual potential energy field. The potential energy value of this field is related to the distance between the agricultural machinery state and the attractor, and exhibits asymmetric characteristics. That is, when the agricultural machinery state deviates from different sides of the attractor, the rate of change of the potential energy field is different, to adapt to the dynamic characteristics and control requirements of the agricultural machinery motion. When the agricultural machinery state deviates towards the stable manifold boundary, the potential energy value of the field increases rapidly, forming a strong constraint effect. As the center of the shape approaches, the potential energy value of the potential energy field changes slowly, ensuring the stability of the motion. Based on the constructed asymmetric virtual potential energy field, the gradient calculation method is used to solve the virtual potential energy gradient of the target agricultural machinery state in the potential energy field. The direction of the potential energy gradient points to the direction in which the potential energy value decreases the fastest, and its amplitude reflects the rate of change of potential energy. At the same time, the motion direction of the target agricultural machinery state is obtained in real time through data from the agricultural machinery attitude sensor and motion state sensor. The angle between the negative component of the virtual potential energy gradient and the motion direction of the agricultural machinery state is calculated. This angle can reflect the degree of fit between the negative gradient direction and the motion direction of the agricultural machinery. The amplitude of the negative component of the virtual potential energy gradient is multiplied by the sine value of the angle. The calculation result is used as the core calculation basis for the change of the driving command. This change can dynamically adjust the driving command of the agricultural machinery, so that the agricultural machinery state gradually moves closer to the virtual dynamic attractor and is finally maintained in the stable manifold.

[0033] In a preferred embodiment of the present invention, step 3 is further included: monitoring the steering abruptness of the target agricultural machinery. When the steering abruptness exceeds a threshold calculated in real time from the amplitude of the observed signal, it is marked as a strong physical interaction event and a time window is triggered. The specific operation is as follows: during the operation of the target agricultural machinery, its steering abruptness parameter is continuously monitored. This parameter reflects the rate of change of the steering angular velocity and is directly related to the dynamic load borne by the agricultural machinery steering system and the intensity of physical interference from the external environment. The steering abruptness is collected in real time by an angular velocity sensor deployed on the agricultural machinery steering system. First, the steering angular velocity signal is acquired, and then the first-order time derivative operation is performed on the signal to obtain a continuous steering abruptness signal. The acquisition and operation frequencies are both... The frequency of the observed signal processing is kept synchronized to ensure data timing consistency. At the same time, based on the real-time amplitude of the observed signal generated in step 1, a dynamic threshold is calculated using a preset exponential scaling algorithm. This threshold is dynamically adjusted as the amplitude of the observed signal changes to adapt to the judgment requirements under different disturbance intensities. The real-time calculated steering abruptness value is compared with the dynamic threshold every hour. When the steering abruptness value exceeds the dynamic threshold, it indicates that the agricultural machinery is currently encountering strong external physical interference and is judged as a strong physical interaction event. Once a strong physical interaction event is judged, the corresponding time window is immediately triggered. This time window is used to centrally collect and analyze the agricultural machinery status data and environmental perception data before and after the event.

[0034] The method based on strong physical interaction events and triggering time windows includes the following steps: Step 31, performing adaptive multi-scale decomposition on the steering abruptness signal to obtain intrinsic mode components and trend remainders. The rising slope and overshoot peak of the energy envelope are extracted from the first intrinsic mode component, and the sign and linear fitting slope are extracted from the trend remainder. Specifically, the steering abruptness signal acquired and calculated in Step 3 is processed using an adaptive multi-scale decomposition algorithm. This algorithm does not require a preset decomposition scale and can adaptively divide the scale levels according to the frequency characteristics of the steering abruptness signal, effectively separating useful feature components and noise interference components in the signal. Through adaptive multi-scale decomposition, the steering abruptness signal... The signal is decomposed into several intrinsic mode components and a trend remainder term. The intrinsic mode components reflect the fluctuation characteristics of the signal at different frequency scales, and the trend remainder term reflects the overall trend of the signal. Among the multiple intrinsic mode components obtained by decomposition, the first intrinsic mode component is selected as the analysis object. This component contains the high-frequency fluctuation characteristics in the turning speed signal and best reflects the instantaneous change law when a strong physical interaction event occurs. The envelope extraction method in Hilbert-Huang transform is used to perform energy envelope analysis on the first intrinsic mode component. First, the upper and lower envelopes are fitted to the local maxima and minima of this component by cubic spline interpolation algorithm, respectively. Then, the mean of the upper and lower envelopes is calculated to obtain the energy envelope of the first intrinsic mode component. Two key feature parameters are extracted from the energy envelope: the rising slope and the overshoot peak. The rising slope is calculated by linearly fitting the rising phase of the energy envelope, with the fitting interval being the complete range from the initial value to the peak value. The slope value reflects the steepness of the energy envelope's rise. The overshoot peak is the maximum amplitude of the energy envelope exceeding the steady-state value, obtained by finding the difference between the peak value and the steady-state value. Simultaneously, feature extraction is performed on the decomposed trend remainder. First, the overall sign of the trend remainder is determined by a sign function, i.e., whether the overall trend remainder value is positive or negative. Then, the least squares method is used to linearly fit the trend remainder to obtain the linear fitting slope of the trend remainder. This slope reflects the steepness of the overall trend of the turning abruptness signal.

[0035] Step 32: When the rising edge slope exceeds the dynamic threshold of exponential scaling of the observed signal amplitude, and the sign of the trend remainder term differs from the sign of the overshoot peak, a strong physical interaction event is determined. The length of the time window corresponding to the strong physical interaction event is set based on the ratio between the absolute value of the overshoot peak and the absolute value of the linear fitting slope. The specific operation is as follows: Based on the feature parameters extracted in step 31, a dual determination condition for strong physical interaction events is constructed to ensure the accuracy and reliability of the determination results. The first determination condition is the rising edge slope threshold determination. The dynamic threshold is calculated using the exponential scaling algorithm mentioned in step 3, and its specific expression is T. s equals ke Multiply by e kf*A(t) The derivation of this formula is based on the correlation between the amplitude of the observed signal and the rising edge slope. When a strong physical interaction event occurs, the larger the amplitude of the observed signal, the higher the corresponding rising edge slope threshold should be. An exponential function is used to achieve a non-linear correlation between the two, ensuring both sensitivity in low-disturbance scenarios and avoiding misjudgments in high-disturbance scenarios. In the formula, T... s The rising edge slope dynamic threshold; k e The threshold reference coefficient is calibrated through farmland operation tests, with a value range of 0.8 to 1.2; kf is the exponential scaling factor, preset according to signal characteristics, with a value range of 0.5 to 1.0; the rising edge slope extracted in step 31 is compared with this dynamic threshold to determine whether the rising edge slope exceeds the dynamic threshold; the second judgment condition is the sign consistency judgment, which judges whether the sign of the trend remainder term is different from the sign of the overshoot peak, that is, when the sign of the trend remainder term is positive, the sign of the overshoot peak is negative, or when the sign of the trend remainder term is negative, the sign of the overshoot peak is positive. This condition is used to distinguish strong physical interaction events from normal operation fluctuations, ensuring that only external sudden physical interference is judged; when the above two judgment conditions are met at the same time, it is formally judged as strong. For physical interaction events, the ratio of the absolute values ​​of the overshoot peak and the linear fitting slope extracted in step 31 is calculated. This ratio reflects the instantaneous intensity and sustained impact of a strong physical interaction event. A larger ratio indicates a stronger instantaneous intensity and a shorter sustained impact, while a smaller ratio indicates a weaker instantaneous intensity and a longer sustained impact. Based on this ratio, the time window length corresponding to the strong physical interaction event is set through a preset linear mapping relationship. The mapping relationship is calibrated through multiple sets of experimental data to ensure that the time window length can fully cover the entire process of the event's occurrence, development, and decay. Specifically, the mapping logic is that the ratio and the time window length are negatively correlated, that is, the larger the ratio, the shorter the time window length, and the smaller the ratio, the longer the time window length.

[0036] In a preferred embodiment of the present invention, step 4 is further included: within the trigger time window, directional and parameter-determined collaborative sampling is performed to determine the overlap between the physical envelope determined by the target agricultural machinery status and the set of inaccessible points in the sampling point cloud, and an instantaneous passability flag is output. The specific operation is as follows: after step 3 triggers the time window corresponding to the strong physical interaction event, the directional and parameter-determined collaborative sampling process is immediately started. This process is to sample the agricultural machinery's operating environment and its own status. In the directional and parameter-determined collaborative sampling, the orientation is determined based on the real-time movement trend of the target agricultural machinery and the type of strong physical interaction event, that is, focusing sampling in the direction where there may be environmental interference or physical constraints. For example, when the agricultural machinery sideslips, the focus is on sampling in the side-slip direction and the working area in front. The parameter determination is based on parameters such as the amplitude of the observed signal and the speed of steering, and sets parameters such as the sampling frequency and sampling resolution. The higher the event intensity, the higher the sampling frequency and the finer the resolution to ensure the capture of key environmental and status information. The sampling process is carried out through the visual sensors and laser radars on the agricultural machinery. Simultaneously, sensing devices collect spatial point cloud data of the work area. This point cloud data includes spatial information such as environmental obstacles, soil undulations, and crop distribution. Simultaneously, based on the real-time status parameters of the target agricultural machinery, including vehicle attitude angle, speed, and steering angle, a physical envelope of the machinery is constructed. This physical envelope is a three-dimensional spatial region that conforms to the outline of the machinery and its working range, representing the spatial area that the machinery may occupy during movement. From the collected spatial point cloud, an insurmountable point set is selected. This set includes spatial points corresponding to areas where the machinery cannot pass normally, such as obstacles and soft ground. The selection criteria are preset point cloud feature thresholds, such as reflection intensity thresholds and spatial position height thresholds. The overlap between the physical envelope and the insurmountable point set is calculated. The overlap is the ratio of the number of points in the insurmountable point set that fall within the physical envelope to the total number of points within the physical envelope. This ratio directly reflects the risk of conflict between the machinery and insurmountable areas during passage. Based on the overlap determination result, an instantaneous passability indicator is output.

[0037] The output of instantaneous traffic clearance flags also includes the following steps: Step 41, based on the state and motion manifold prediction of the target agricultural machinery, a hierarchical scanning sequence is generated; the hierarchical scanning sequence is used to define the focal area and risk buffer layer, and to allocate scanning resource ratios to the focal area and risk buffer layer. The specific operation is as follows: First, the real-time state parameters of the target agricultural machinery are obtained, including vehicle position, driving speed, steering angular velocity, vehicle attitude angle, etc. These parameters are synchronously collected through positioning modules, inertial measurement units, speed sensors and other devices deployed on the agricultural machinery to ensure the real-time performance and accuracy of the data. At the same time, based on... The stable manifold constructed in step 2, combined with the current state of the agricultural machinery, is used to predict the kinematic manifold of the machinery in the near future using a linear prediction model. The kinematic manifold characterizes the expected trajectory and state change patterns of the agricultural machinery under conditions of no external interference. The prediction duration is consistent with the time window length to ensure coverage of the complete impact cycle of strong physical interaction events. Based on the real-time state of the agricultural machinery and the kinematic manifold prediction results, a hierarchical scanning sequence is generated. This sequence is used to define the scanning range, scanning order, and resource allocation rules of the sensing equipment. The hierarchical scanning sequence divides the scanning area into two levels: a focal zone and a risk buffer layer. The core area to be covered by the physical envelope of the agricultural machinery is the area most directly affected by strong physical interaction events. It is usually set within a range of 1 to 3 meters in front of the agricultural machinery and 0.5 meters to the left and right. The scanning results in this area directly determine whether the agricultural machinery can pass safely. The risk buffer layer is a transitional area outside the focal area, distributed in a ring around the focal area, with a preset width of 0.5 to 1 meter. It is used to detect potential environmental risks in advance and reserve reaction time for the agricultural machinery to adjust its status. The hierarchical scanning sequence allocates scanning resources to both areas simultaneously. Scanning resources include scanning frequency, sampling point density, and data processing priority. Based on risk level, resource allocation is set, with 70% to 80% of scanning resources allocated to the focal area, using a higher scanning frequency, such as 100Hz, and a higher sampling point density to ensure the accuracy of information collection in the core area; the risk buffer layer is allocated 20% to 30% of scanning resources, using a relatively lower scanning frequency, such as 50Hz, to reduce the system's data processing load while ensuring risk perception. During the generation of the scanning sequence, the scanning order is determined simultaneously, prioritizing the scanning of the central area of ​​the focal area, and then gradually expanding to the edge and the risk buffer layer to ensure that information in the core area is acquired and processed first.

[0038] Step 42: Based on the predicted spatiotemporal region covered by the physical envelope, an evidence field is initialized, and scanning is performed according to the hierarchical scanning sequence. Spatial points acquired during the scan are classified into collision evidence or risk field strength values ​​based on their positional relationship with the physical envelope and mapped to the evidence field. When scanning resources in the focal area are exhausted, an instantaneous passability flag is output based on whether there is collision evidence in the evidence field or whether its average risk field strength is lower than a preset risk threshold. The specific operation is as follows: Based on the agricultural machinery physical envelope constructed in Step 4, combined with the motion manifold prediction results, the predicted spatiotemporal region covered by the physical envelope within the time window is determined. This region includes the spatial position of the physical envelope at various future times, with the time dimension consistent with the time window length and the spatial dimension matching the physical envelope and motion trajectory. An evidence field is initialized based on this predicted spatiotemporal region. The evidence field is a three-dimensional grid field with the same size as the predicted spatiotemporal region, and the grid resolution is set according to the scanning accuracy, for example, 0.1 m × 0.1 m × 0.1 m. Each grid cell stores access risk evidence information for that location. During evidence field initialization, the initial risk field strength of all grid cells is set to 0, with no collision evidence markers. Subsequently, the sensing device executes scanning operations according to the hierarchical scanning sequence generated in step 41. The scanning process strictly follows the preset scanning range, order, and resource allocation rules, prioritizing the scanning and data acquisition of the focal area. For each spatial point acquired during the scanning process, its positional relationship with the physical envelope is determined through spatial coordinate comparison. If the coordinates of a spatial point fall within the spatial range of the physical envelope, the point is marked as collision evidence, and a collision evidence marker is added to the corresponding grid cell in the evidence field. If the coordinates of a spatial point fall within the risk buffer layer but do not enter the physical envelope, the risk field strength is calculated based on the straight-line distance between the point and the physical envelope. The closer the distance, the greater the risk field strength. The calculation method uses a linear decay model, specifically: E equals E0 multiplied by D (within parentheses). maX -d m Then divide by D maX The derivation of this formula is based on the distance correlation between a spatial point and its physical envelope. The closer the distance, the greater the impact on agricultural machinery passage, and the higher the risk field strength should be. A linear relationship is used to simplify the calculation, ensuring real-time performance and operability, while also meeting the risk characterization requirements of the evidence field. In the formula, E is the risk field strength value corresponding to the spatial point; E0 is the maximum risk field strength benchmark value, calibrated experimentally, with a range of 0 to 1; D... maX The maximum effective distance of the risk buffer layer, i.e., the farthest distance within the risk buffer layer from the physical envelope, is equal to the width of the risk buffer layer; d mThe actual straight-line distance between the spatial point and the physical envelope is used. The calculated risk field strength value is mapped to the corresponding grid cell in the evidence field, and the risk field strength information of the grid cell is updated. The scanning operation is continuously performed until the scanning resources in the focal area are exhausted. The criterion for exhausting the scanning resources is that the preset number of scans in the focal area or the number of sampling points has been completed. For example, the preset number of scans in the focal area is 10. Once this number of scans is completed, it is determined that the resources are exhausted. At this time, the evidence field is evaluated as a whole, and the output judgment condition of the instantaneous passability flag is constructed: if any grid cell in the evidence field is marked with collision evidence, the agricultural machinery passability risk is determined to be extremely high, and a low instantaneous passability flag is output; if there is no collision evidence in the evidence field, the average risk field strength value of all grid cells is calculated, that is, the average risk field strength. The average risk field strength is compared with the preset risk threshold. If the average risk field strength is lower than the threshold, a high instantaneous passability flag is output; if it is higher than or equal to the threshold, a medium instantaneous passability flag is output. The preset risk threshold is calibrated through multiple sets of farmland operation experiments to ensure that it is adapted to the passability judgment requirements of different operation scenarios.

[0039] In a preferred embodiment of the present invention, step 5 is further included, which couples the instantaneous accessibility flag with the type of strong physical interaction event and performs an operation based on a preset response primitive mapping relationship. The specific operation is as follows: after completing the strong physical interaction event judgment and triggering the time window in step 3 and completing the output of the instantaneous accessibility flag in step 4, the instantaneous accessibility flag is coupled with the type of strong physical interaction event to form a unique combination of operating conditions. This combination can accurately characterize the type of physical interference encountered by the agricultural machinery and the corresponding level of access risk. Based on the response primitive mapping relationship established in advance through farmland operation experiments and simulation verification, this mapping relationship uses the combination of operating conditions as the index key. The corresponding preset response operation logic is to match the appropriate control strategy for different types of strong physical interaction events. At the same time, the aggressiveness of the strategy is adjusted in combination with the instantaneous accessibility flag to avoid aggravating the instability of the agricultural machinery state or affecting the operation efficiency due to improper control actions. Through this coupling and mapping mechanism, the results of environmental perception and state judgment are transformed into executable control operations to realize real-time response to strong physical interaction events and ensure that the agricultural machinery can quickly recover to a stable operating state under dynamic disturbances.

[0040] The operation based on the pre-set response primitive mapping relationship also includes the following steps: Step 51, according to the strong physical interaction event type and instantaneous trafficability indicator, index the main strategy fragment and auxiliary strategy fragment from the strategy fragment library, and determine its initial weight based on the historical performance score dynamics of the strategy fragment. The specific operation is as follows: First, the strong physical interaction events are classified into types based on the physical causes and manifestations of the events, including sideslip disturbance, collision warning, soft foundation subsidence, crop shading, etc. Each type is uniquely identified by the observed signal characteristics, steering abruptness characteristics, and physical envelope overlap characteristics when the event is triggered. The instantaneous trafficability indicator is divided into three levels: high, medium, and low according to the judgment result of step 4, corresponding to the working conditions of no obvious traffic risk, potential traffic risk, and serious traffic risk, respectively. The strategy fragment library is pre-built and stores several independent strategy fragments. Each strategy fragment corresponds to a specific control action, including steering angle correction, speed reduction, vehicle posture fine adjustment, and working depth adjustment. Each strategy fragment is labeled with the applicable event type and trafficability indicator. The system uses a combination of strong physical interaction event types and instantaneous traffic visibility flags as index keys to retrieve matching primary and secondary strategy fragments from the strategy fragment library. The primary strategy fragment represents the core control action for the current event, while the secondary strategy fragment represents supplementary control actions to optimize the primary strategy's effectiveness. For example, when a sideslip disturbance event is combined with a high traffic visibility flag, the primary strategy fragment corrects the steering angle, and the secondary strategy fragment fine-tunes the driving speed. Each strategy fragment is associated with a historical performance score, which is dynamically updated based on historical operational data. The scoring dimensions include disturbance suppression effectiveness, trajectory tracking accuracy, and operational efficiency loss. A higher score indicates better performance of the strategy fragment under the corresponding working conditions. Based on the dynamic values ​​of the historical performance scores, initial weights are assigned to the indexed primary and secondary strategy fragments. The assignment rule is that strategy fragments with higher performance scores have larger initial weights. The weight calculation uses normalization to ensure that the sum of the initial weights of the primary and secondary strategy fragments is 1. For example, if the performance score of the primary strategy fragment is 80 and the secondary strategy fragment is 60, the initial weights are 0.57 and 0.43, respectively.

[0041] Step 52: Based on impedance sensing, the action intensity parameters of the strategy segments are adjusted online. The adjusted instructions of the main strategy segment and the auxiliary strategy segment are then mixed according to the initial weights to generate a composite control instruction. The specific operation is as follows: After determining the main and auxiliary strategy segments and the initial weights, the real-time operating impedance is obtained through the impedance sensing module deployed on the agricultural machinery. The impedance sensing module includes force sensors, torque sensors, and attitude sensors, which respectively collect soil shear impedance, steering system resistance, driving system resistance, and inertial impedance caused by changes in vehicle posture. The various impedance signals are fused and processed to obtain a comprehensive operating impedance value, which reflects the intensity of physical resistance experienced by the agricultural machinery during the current operation. Based on the comprehensive operating impedance value, the action intensity parameters of the main strategy segment and the auxiliary strategy segment are adjusted online. The action intensity parameters include the amplitude of the steering correction angle, the proportion of speed reduction, and the magnitude of attitude adjustment. The adjustment logic is that the larger the comprehensive operating impedance value, the larger the adjustment magnitude of the action intensity parameters, that is, the more aggressive the control action, and vice versa, the smaller the adjustment magnitude and the smoother the control action. The adjustment process adopts a linear mapping relationship, specifically expressed as S equals S0 multiplied by 1+k (within parentheses). r *Z, this formula is based on the correlation between operating impedance and control action strength. Higher impedance requires stronger control actions to counteract disturbances. A linear relationship is used to simplify calculations and ensure real-time performance, while also adapting to the response characteristics of agricultural machinery actuators. In the formula, S is the adjusted action strength parameter; S0 is the preset baseline action strength parameter for the strategy segment; k r Z is the impedance adjustment coefficient, calibrated through farmland operation tests, with a value range of 0.2 to 0.8; Z is the normalized comprehensive operating impedance value, with a value range of 0 to 1. The control commands of the main and auxiliary strategy segments after online adjustment are multiplied by their respective initial weights, and the product results are added together to obtain the composite control command. This command integrates the core response capability of the main strategy and the optimization capability of the auxiliary strategy, while adapting to the dynamic changes of real-time operating impedance. Finally, the composite control command is output to the steering system, power system and other actuators of the agricultural machinery to complete the response operation to strong physical interaction events.

[0042] In a preferred embodiment of the present invention, step 6 is further included: asynchronously storing the event point coordinates, observation signal characteristics, and event type into a circular queue; when the target agricultural machinery is determined to have entered a stable state based on the observation signal, the marker points are extracted and a physical interaction feature layer is constructed; the physical interaction feature layer is used to generate a reference path and a priority for subsequent directional and parameter-based collaborative sampling in the corresponding area; the specific operation is as follows: after completing the response operation for the strong physical interaction event, the event point coordinates, observation signal characteristics, and event type corresponding to the event are asynchronously stored in a preset circular queue. The asynchronous storage mode can complete data caching without occupying the main control process's computing resources; the circular queue achieves orderly data management and capacity control through a first-in-first-out storage rule, avoiding storage overload caused by continuous data accumulation; continuously monitoring the change in the amplitude of the observation signal generated in step 1; when the amplitude of the observation signal is continuously within the stable manifold boundary determined in step 2 and reaches the preset stable duration threshold, the target agricultural machinery is determined to have entered a stable operating state. This determination result indicates that the impact of this strong physical interaction event has been completed. Complete elimination; at this point, extract the marker points corresponding to all strong physical interaction events stored in the current and historical records from the circular queue. The marker points are the correlation data of the event point coordinates at the time of each event and the corresponding observed signal characteristics and event types. These marker points are then mapped into a grid according to the spatial coordinates of farmland operations to construct a physical interaction feature layer. This layer uses a two-dimensional or three-dimensional spatial grid as its carrier, storing the type of physical interaction event occurring at that location, the statistical value of the observed signal characteristics, and the frequency of event occurrence in the corresponding grid cells, thus realizing the spatialized and structured expression of discrete event data. Based on the physical interaction feature layer, subsequent operation planning is carried out. On the one hand, the spatial distribution characteristics of high-risk physical interaction events in the layer are extracted, and an optimized reference path is generated in combination with the basic path requirements of farmland operations, so that subsequent operations of agricultural machinery can avoid high-risk interaction areas. On the other hand, according to the frequency of event occurrence and feature intensity of each region in the layer, sampling priorities are assigned for subsequent directional and parameter-based collaborative sampling in the corresponding regions. The more frequently events occur and the higher the feature intensity, the higher the sampling priority is assigned, ensuring that sensing resources are tilted towards high-risk areas.

[0043] Step 6 further includes the following steps: Step 61, dynamic clustering is performed based on event point coordinates, occurrence timestamps, and observed signal features to generate scene objects containing spatial boundaries, time windows, and observed signal features, as well as statistical distribution and statistics of major event types. The specific operation is as follows: The event point coordinates, occurrence timestamps, and observed signal features of all strong physical interaction events stored in the circular queue are used as clustering input data. These three types of data together constitute a multi-dimensional feature vector, where event point coordinates are the spatial dimension feature, occurrence timestamps are the temporal dimension feature, and observed signal features are the signal dimension feature. The multi-dimensional features can comprehensively characterize the spatiotemporal attributes and signal features of physical interaction events. The observed signal features of each event consist of three core parameters: rising edge slope, plateau period energy mean, and falling edge duration, forming a three-dimensional signal feature vector. A dynamic clustering algorithm is used to perform clustering analysis on this multi-dimensional feature vector. This algorithm can adaptively adjust the number and range of clusters according to the spatial distribution density and temporal correlation of the feature vector, without needing to preset fixed cluster categories, thus adapting to the non-uniform and random distribution characteristics of physical interaction events in farmland operation scenarios. During clustering, feature similarity is used as the core criterion. The spatial distance, time difference, and correlation coefficient of signal features between any two feature vectors are calculated. The correlation coefficient of signal features can be calculated using existing formulas such as the Pearson correlation coefficient. By calculating the ratio of the covariance of two signal feature vectors to their respective standard deviations, the linear correlation between the two in terms of numerical change trends is quantified. The value range is [-1, 1]. The closer the absolute value is to 1, the more consistent the signal feature trends are, which meets the requirements for feature similarity determination. When the combined similarity of spatial distance, time difference, and correlation coefficient of signal features exceeds a preset threshold, the two corresponding physical interaction events are grouped into the same cluster. After clustering all feature vectors, feature extraction and structured encapsulation are performed on each cluster. First, the spatial boundary of the cluster is calculated based on the coordinates of all event points within the cluster. The minimum bounding box method is used to determine the spatial start and end range of the region, accurately representing the spatial area where physical interaction events occur in a concentrated manner. Then, the time window of the cluster is determined based on the timestamps of all events within the cluster, i.e., the time interval between the first and last events within the cluster, reflecting the temporal distribution characteristics of physical interaction events in the region. At the same time, the statistical distribution of observed signal features within the cluster is calculated, including the mean, variance, and peak value of each characteristic parameter of the observed signal, quantifying the signal characteristic patterns of physical interaction events in the region. Finally, the frequency of occurrence of each type of physical interaction event within the cluster is counted, and the event type with the highest frequency is determined as the main event type of the cluster. The spatial boundary, time window, statistical distribution of observed signal features, and statistics of main event types of each cluster are integrated to generate a structured scene object, realizing the regionalization and structured aggregation of discrete physical interaction event points, enabling high-risk interaction areas in farmland operation scenarios to be clearly identified and represented.

[0044] Step 62: Based on the spatiotemporal order and spatial adjacency of scene objects, causal relationships are mined. Based on the determination of causal relationships, a virtual repulsion field is generated for path optimization, and corresponding regions are assigned higher perception priority. The specific operations are as follows: Spatiotemporal correlation analysis is performed on all scene objects generated in Step 61. First, the time window and spatial boundary features of each scene object are extracted. Then, it is determined whether there is a spatiotemporal correlation between any two scene objects. The criteria for determining spatiotemporal correlation are that the time windows of the two scene objects are sequential, and the time windows of the later-occurring scene object and the earlier-occurring scene object overlap or the time interval is less than a preset threshold. Simultaneously, the spatial boundaries of the two scene objects are spatially adjacent, meaning the minimum straight-line distance between the spatial boundaries is less than a preset spatial correlation threshold. The values ​​may have some overlap; based on the determination that there are spatiotemporal correlations between scene objects, further explore their causal relationships, calculate the correlation coefficient between the statistical values ​​of the observed signal features of the scene object that occurred first and the statistical values ​​of the observed signal features of the scene object that occurred later. The statistical values ​​of the observed signal features of each scene object are composed of three statistics: the mean, variance, and peak value of the observed signal features of all events in the cluster, forming a three-dimensional statistical value vector. The correlation coefficient is also calculated using the existing formula Pearson correlation coefficient. It is consistent with the signal feature correlation coefficient in step 61. The linear correlation between the two statistical value vectors is quantified by the ratio of covariance to standard deviation. The value range is [-1, 1]. The closer the absolute value is to 1, the more similar the statistical patterns of the observed signal features of the two scene objects are, and the more likely there is a causal relationship, which is suitable for the determination requirements of causal relationship mining.

[0045] When the correlation coefficient exceeds a preset correlation threshold, a causal relationship is determined between the two scene objects, meaning that the earlier physical interaction event is the cause of the later physical interaction event. For scene objects with causal relationships and high-risk single scene objects, the risk level is calculated based on their main event types and statistical values ​​of observed signal characteristics. The risk level is positively correlated with the peak value of the observed signal characteristics and the frequency of event occurrence. A virtual repulsion field is generated based on the risk level. The field amplitude of the virtual repulsion field is positively correlated with the risk level of the scene object. The higher the risk level of the scene object, the larger the field amplitude of the virtual repulsion field in the corresponding area. This repulsion field will be integrated into the path planning process of agricultural machinery for subsequent optimization of reference paths, so that the planned reference paths can extend along the direction of decreasing field amplitude of the virtual repulsion field, away from high-risk causal relationship areas. At the same time, all spatial areas corresponding to scene objects with causal relationships are given higher perception priority, which is higher than high-risk and low-risk areas without causal relationships. This ensures that when performing directional and parameter-based collaborative sampling in this area, scanning resources can be allocated first, achieving accurate perception and real-time monitoring of high-risk areas with causal relationships.

[0046] Step 63: Based on the comparison between actual events and expectations in subsequent operations, adjust the field amplitude of the virtual repulsion field and its perception priority to a higher level. Integrate new event points into scene objects to update their boundaries and feature distributions. The specific operations are as follows: After completing reference path planning and sampling priority allocation based on the physical interaction feature layer and scene objects, when the agricultural machinery subsequently travels to the spatial area corresponding to the generated scene object, predict the types and intensities of physical interaction events that may occur in the area based on the main event types, statistical distribution of observed signal features, and risk level of the scene object, forming expected event features. Simultaneously, monitor the actual physical interaction events occurring in the area in real time, extract the type and intensity features of the actual events, compare and analyze the actual event features with the expected event features, and calculate the feature deviation between the two. The deviation includes the matching degree of event types and the difference in event intensity. The matching degree of event types is the degree of consistency between the actual event type and the expected event type. The difference in event intensity is the absolute difference between the peak value of the observed signal features of the actual event and the statistical peak value of the observed signal features of the expected event. Dynamically adjust the field amplitude and perception priority of the virtual repulsion field corresponding to the area according to the magnitude of the feature deviation. When the feature deviation is large, it indicates that the physical repulsion of the area is in a critical state. The characteristics of the interactive events have changed, requiring an increase in the amplitude of the virtual repulsion field and a higher perception priority for this area. This ensures that the reference path further avoids this area and that perception resources are further tilted towards it. When the characteristic deviation is small, it indicates that the physical interactive event characteristics in this area are basically consistent with expectations. The amplitude of the virtual repulsion field can be slightly adjusted or kept unchanged depending on the actual situation, and the perception priority should also remain at the original level. The event point coordinates, timestamps, and observation signal characteristics corresponding to the newly occurring physical interactive events in this area should be used as new event points and integrated into the corresponding scene object. The space of the scene object should then be recalculated. If a new event point exceeds the original spatial boundary, the minimum bounding box is expanded to update the spatial boundary. The statistical distribution of the observed signal characteristics of the scene object is recalculated, and the observed signal characteristic parameters of the new event point are incorporated into the statistical calculation to update the mean, variance, and other statistical values. At the same time, the occurrence frequency of each type of physical interaction event within the scene object is recalculated, and the main event types are adjusted according to the new frequency proportion. Through the incorporation of new event points and the recalculation of scene object characteristics, the scene object is dynamically updated, enabling it to reflect the latest characteristics of physical interaction events in the corresponding area in real time and adapt to the dynamic changes in the farmland operation environment.

[0047] Example 2: Referring to Figure 2, based on Example 1, this example provides an integrated smart agriculture control system based on edge computing, including: an observation generation module, which establishes a dynamic twin, quantifies the response deficit between the expected rate of change of the vehicle's attitude angle under the driving command and the actual rate of change of the attitude angle, and defines the response deficit as an observation signal; a manifold tracking module, which redefines the control target based on the observation signal, sets the control target to track a stable manifold, which is defined by a preset amplitude boundary of the observation signal, and controls the target agricultural machinery to maintain its state within the stable manifold; and an event detection module, which monitors the steering abruptness of the target agricultural machinery, and when the steering abruptness exceeds a threshold calculated in real time from the amplitude of the observation signal, it is marked as a strong physical interaction. The system triggers events within a time window. The accessibility assessment module performs directional and parameter-based collaborative sampling within the trigger time window, determining the overlap between the physical envelope determined by the target agricultural machinery's state and the set of traversable points in the sampling point cloud, and outputs an instantaneous accessibility flag. The response mapping module couples the instantaneous accessibility flag with strong physical interaction event types, performing operations based on preset response primitive mapping relationships. The path planning module asynchronously stores event point coordinates, observation signal characteristics, and event types into a circular queue. When the target agricultural machinery is determined to have entered a stable state based on the observation signal, it extracts marker points and constructs a physical interaction feature layer. This physical interaction feature layer is used to generate a reference path and a priority for subsequent directional and parameter-based collaborative sampling in the corresponding area.

[0048] The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concepts, should be covered within the scope of protection of the present invention.

Claims

1. A smart agriculture integrated control method based on edge computing, characterized in that, include: A dynamic twin is established to quantify the response deficit between the expected rate of change of vehicle body attitude angle under the driving command and the actual rate of change of attitude angle, and this response deficit is defined as the observation signal; The control target is redefined based on the observed signal. The control target is set to track a stable manifold, which is defined by a preset amplitude boundary of the observed signal, and the control target agricultural machinery state is maintained within the stable manifold. Monitor the steering abruptness of the target agricultural machinery. When the steering abruptness exceeds the threshold calculated in real time from the amplitude of the observed signal, it is marked as a strong physical interaction event and a time window is triggered. Within the trigger time window, directional and parameter-based collaborative sampling is performed to determine the overlap between the physical envelope determined by the target agricultural machinery's state and the set of inaccessible points in the sampling point cloud, and an instantaneous passability flag is output. The instantaneous passability flag is coupled with a strong physical interaction event type, and the operation is performed based on a preset response primitive mapping relationship. The event point coordinates, observation signal features, and event type are asynchronously stored in a circular queue. When the target agricultural machinery is determined to have entered a stable state based on the observation signal, the marked points are extracted and a physical interaction feature layer is constructed. The physical interaction feature layer is used to generate a reference path and a priority for subsequent directional and parameter-based collaborative sampling in the corresponding area.

2. The integrated control method for smart agriculture based on edge computing according to claim 1, characterized in that, The generation of the observation signal includes: establishing a basic mapping table based on the driving command and the expected steady-state vehicle attitude angle change rate; introducing virtual soil mechanical parameters; and expanding the basic mapping table into response curve templates with different parameter values ​​through a parameterized nomograph to form a preset response template library; acquiring the instantaneous vehicle attitude angle change rate sequence; extracting segments of the sequence within a sliding window; performing shape matching between the segments and the response curve templates of the corresponding driving commands in the preset response template library; selecting a virtual working condition reference from the response curve templates based on the consistency rate of change trend and the closeness of energy amplitude; and calculating the residual vector between the segments and the virtual working condition reference; performing short-time window energy integration on the residual vector to obtain the energy envelope; performing piecewise linearization approximation on the energy envelope to identify its rising edge, plateau period, and falling edge; and encoding and setting the slope of the rising edge, the average energy of the plateau period, and the duration of the falling edge as the observation signal.

3. The integrated control method for smart agriculture based on edge computing according to claim 1, characterized in that, The definition of the preset observation signal amplitude boundary includes: cutting historical observation signals into fixed-duration signal segments, calculating the statistical characteristics of each segment and performing density clustering, and back-mapping the cluster region boundaries to generate offline primitive boundary clusters; fragmenting real-time observation signals and calculating statistical characteristics, matching the obtained feature vectors with the feature centers of the offline primitive boundary clusters, selecting the top K primitive boundaries with the highest confidence values ​​based on the matching confidence, and taking their upper and lower envelopes as dynamic boundary envelopes; monitoring the state trend of the target agricultural machinery, querying the preset elastic adjustment function according to the trend to adjust the width of the dynamic boundary envelope, and generating the final stable manifold boundary.

4. The integrated control method for smart agriculture based on edge computing according to claim 3, characterized in that, Maintaining the target agricultural machinery's state within the stable manifold includes: constructing a virtual dynamic attractor on the stable manifold, which moves along the tangent direction of the stable manifold, and its normal position is determined by solving a dynamic equilibrium equation in real time; wherein, the dynamic equilibrium equation is the balance between the current amplitude of the observed signal and a nonlinear restoring force, the nonlinear restoring force being a function of the attractor's distance from the manifold center; constructing an asymmetric virtual potential energy field between the target agricultural machinery's state and the virtual dynamic attractor, calculating the virtual potential energy gradient of the target agricultural machinery's state in the potential energy field, and generating a change in the driving command based on the product of the negative component of the virtual potential energy gradient and the sine of the angle between it and the state's motion direction.

5. The integrated control method for smart agriculture based on edge computing according to claim 1, characterized in that, Based on strong physical interaction events and triggering time windows, the process includes: performing adaptive multi-scale decomposition on the steering abruptness signal to obtain intrinsic mode components and trend remainders; extracting the rising slope and overshoot peak value of the energy envelope from the first intrinsic mode component; and extracting the sign and linear fitting slope from the trend remainder; when the rising slope exceeds a dynamic threshold that is exponentially scaled by the amplitude of the observed signal, and the sign of the trend remainder is different from the sign of the overshoot peak value, it is determined to be a strong physical interaction event, and the length of the time window corresponding to the strong physical interaction event is set according to the ratio between the absolute value of the overshoot peak value and the absolute value of the linear fitting slope.

6. The integrated control method for smart agriculture based on edge computing according to claim 1, characterized in that, The instantaneous accessibility flag is output, including: generating a hierarchical scanning sequence based on the state and motion manifold prediction of the target agricultural machinery; the hierarchical scanning sequence is used to define the focal area and risk buffer layer, and allocate scanning resource ratios to the focal area and risk buffer layer; based on the predicted spatiotemporal region covered by the physical envelope, an evidence field is initialized, and scanning is performed according to the hierarchical scanning sequence. The spatial points obtained in the scan are classified into collision evidence or risk field strength values ​​according to their positional relationship with the physical envelope, and mapped to the evidence field; when the scanning resources in the focal area are exhausted, the instantaneous accessibility flag is output according to whether there is collision evidence in the evidence field or whether its average risk field strength is lower than a preset risk threshold.

7. The integrated control method for smart agriculture based on edge computing according to claim 1, characterized in that, The operation is performed based on the pre-defined primitive mapping relationship, including: indexing the main strategy fragment and auxiliary strategy fragment from the strategy fragment library according to the type of strong physical interaction event and the instantaneous accessibility flag, and dynamically determining the initial weight of the strategy fragment based on the historical performance score of the strategy fragment; adjusting the action intensity parameters of the strategy fragment online based on impedance sensing, and mixing the adjusted instructions of the main strategy fragment and auxiliary strategy fragment according to the initial weight to generate a composite control instruction.

8. The integrated control method for smart agriculture based on edge computing according to claim 1, characterized in that, The priorities for performing directional and parameter-based collaborative sampling include: dynamic clustering based on event point coordinates, occurrence timestamps, and observed signal characteristics to generate scene objects that include spatial boundaries, time windows, and observed signal characteristics, as well as statistical distribution and statistics of major event types; mining causal relationships based on the spatiotemporal order and spatial adjacency of scene objects; generating virtual repulsion fields for path optimization based on the determination results of causal relationships; and assigning higher perception priority to the corresponding areas.

9. The integrated control method for smart agriculture based on edge computing according to claim 8, characterized in that, Priorities for performing directional and parameter-based collaborative sampling also include: adjusting the field amplitude of the virtual repulsion field and giving it a higher perception priority based on the comparison results between actual events and expectations in subsequent operations, and integrating new event points into scene objects to update their boundaries and feature distributions.

10. An edge computing-based integrated control system for smart agriculture, applied to any one of the edge computing-based integrated control methods for smart agriculture as described in claims 1-9, characterized in that, include: The observation generation module establishes a dynamic twin, quantifies the response deficit between the expected rate of change of the vehicle body attitude angle under the driving command and the actual rate of change of the attitude angle, and defines the response deficit as the observation signal; the manifold tracking module redefines the control target based on the observation signal, sets the control target to track a stable manifold, which is defined by the preset amplitude boundary of the observation signal, and controls the target agricultural machinery state to remain within the stable manifold; The event detection module monitors the turning abruptness of the target agricultural machinery. When the turning abruptness exceeds the threshold calculated in real time from the amplitude of the observed signal, it is marked as a strong physical interaction event and a time window is triggered. The passage assessment module performs directional and parameter-based collaborative sampling within the triggered time window, determines the overlap between the physical envelope determined by the state of the target agricultural machinery and the set of inaccessible points in the sampling point cloud, and outputs an instantaneous passage indicator. The response mapping module couples the instantaneous accessibility flag with the type of strong physical interaction event and performs operations based on the preset response primitive mapping relationship; The path planning module asynchronously stores event point coordinates, observation signal characteristics, and event types into a circular queue. When the target agricultural machinery is determined to have entered a stable state based on the observation signal, it extracts the marker points and constructs a physical interaction feature layer. The physical interaction feature layer is used to generate a reference path and a priority for subsequent directional and parameter-based collaborative sampling in the corresponding area.

Images