Agricultural machinery automatic driving operation method applied to automatic driving harvester

Abstract

The invention discloses an agricultural machinery automatic driving operation method applied to an automatic driving harvester. The method belongs to the technical field of automatic driving of agricultural machinery, and comprises the following specific operation steps: firstly, acquiring internal environment information of a greenhouse by using a millimeter wave radar, and constructing a two-dimensional grid map by using a Cartograph algorithm; secondly, according to the drawn grid map, an improved PSA algorithm is adopted to complete global path planning of the automatic driving harvester; and finally, the automatic driving harvester carries a ZED camera to collect image information within 5-6 m around the vehicle, an xLSTM model is used to identify the collected image information, whether there is an obstacle around the vehicle is determined, and if there is an obstacle, a DCS algorithm is used to complete local path planning of the operation vehicle. According to the invention, farmland crops can be harvested more quickly and accurately, the labor cost is reduced, and the harvesting efficiency is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of automatic driving of agricultural machinery, and relates to an automatic driving operation method of agricultural machinery applied to an automatic driving harvester. Background Art

[0002] In recent years, the technology of autonomous driving for agricultural machinery has developed rapidly, showing explosive growth and expanding scope of application, which is injecting new vitality into agricultural modernization. The application of autonomous driving technology in agricultural machinery enables tractors, harvesters, etc. to complete farmland operations autonomously, improving operation efficiency and precision, and helping to improve agricultural production efficiency and sustainable development. Agricultural production is becoming increasingly sophisticated, thanks to the gradual development of agricultural industrialization technology.

[0003] Farmland image acquisition technology and path planning technology are important technical means for autonomous harvesters to achieve automatic harvesting tasks; there are often some obstacles in real farmlands, which brings challenges to the autonomous driving of harvesters; relying on sensors to perceive the surrounding specific environment, obtain image data information, and draw two-dimensional maps, the harvester can avoid inaccessible areas, harvest crops smoothly, and improve accuracy and reliability; since most of the current large-scale agricultural greenhouses cover a large area, reasonable path planning can allow the harvester to travel along the best path, shorten driving time, improve harvesting efficiency, and reduce costs in all aspects; therefore, farmland image acquisition technology and path planning technology are extremely important for autonomous harvesters. Summary of the invention

[0004] In view of the above problems, the purpose of the present invention is to solve the existing problems. The present invention proposes an agricultural machinery automatic driving operation method applied to an automatic driving harvester, aiming to improve the efficiency and accuracy of agricultural machinery harvesting.

[0005] The technical solution of the present invention is: an agricultural machinery automatic driving operation method applied to an automatic driving harvester according to the present invention, and its operation steps are as follows:

[0006] Step 1: Use millimeter-wave radar to collect information about the greenhouse's internal environment, and use the Cartographer algorithm to build a two-dimensional grid map;

[0007] Step 2: Based on the grid map in step 1, the improved PSA algorithm is used to complete the global path planning of the automatic driving harvester;

[0008] Step 3: The autonomous harvester is equipped with a ZED camera to collect image information within 5-6 meters around the vehicle, and uses the xLSTM model to identify the collected image information to determine whether there are obstacles (ground piles, large rocks, etc.) around the vehicle. If there are obstacles, the DCS algorithm is used to complete the local path planning of the operating vehicle.

[0009] Furthermore, in step 1, the specific method of using the millimeter wave radar to collect the internal environment information of the greenhouse is: the UAV is equipped with the millimeter wave radar to complete the collection of greenhouse environment point cloud data and complete the harvester driving environment modeling;

[0010] Point cloud data is a set of three-dimensional points used to describe the nonlinear surface shape of an object in three-dimensional space. Each point in the point set is represented by orthogonal length, width and height three-dimensional coordinates in space.

[0011] The point cloud data set collected by the millimeter wave radar is S = {s i}, where i∈N, s i =[x i ,y i, z i ] T ∈R 3 ;

[0012] Divide the 3D point set area into two parts with imaging side length d s A finite voxel set V with side length i , to ensure the integrity of the surface image captured by the viewpoint, the center point of the voxel is regarded as the voxel coordinate, and the calculation formula for selecting the viewpoint is:

[0013]

[0014] In the formula, p o is the point that the UAV track needs to pass through; P o Point cloud set of greenhouse environment collected by UAV millimeter wave radar.

[0015] Furthermore, in step 1, the specific method of using the Cartographer algorithm to construct a two-dimensional grid map is:

[0016] (1) Local mapping: The pose of a point in the point cloud data obtained by the radar in its own coordinate system is set to ξ = (ξ x ,ξ y ,ξ θ ), where ξ x , y Respectively represent the x-axis and y-axis translation of the point relative to the scanning origin, ξ θ Indicates the amount of rotation;

[0017] Before constructing the sub-map, it is necessary to transform the poses of all points in the point cloud data from the radar coordinate system to the corresponding sub-map coordinate system. Assume that the transformed pose is ξ ′ , then:

[0018]

[0019] Where: Tξ Represents the corresponding pose transformation, which is represented by the rotation matrix and translation vector Composition, that is

[0020] The Cartographer algorithm uses a probabilistic grid to represent the submap, inserting the point cloud data after coordinate transformation into the corresponding grid points in the probabilistic grid. Each insertion will update the probability value of the grid point. Before inserting the point cloud data, the Ceres library needs to be used to optimize the position of the scan point relative to the current submap to obtain the position with the highest probability of the scan point corresponding to the point cloud data in the submap. This process can be converted into a nonlinear least squares problem, and the formula is:

[0021]

[0022] Where: M smooth It is a bicubic interpolation function, which smoothes the probability values ​​of the corresponding grid points in the sub-map and obtains the maximum value by reversely solving the minimum value;

[0023] (2) Loop detection: The core essence of loop detection is a matching analysis process, which mainly analyzes the latest scan frame data captured in real time, and uses it as a reference to systematically explore and confirm the best matching frame that meets the predetermined similarity standard in the preset search area around the current frame, so as to successfully construct an effective loop closure; as shown in the formula:

[0024]

[0025] In the search space W, the above formula is used to filter the best frame among the matching frames;

[0026] (3) Global Mapping: After loop closure detection, global pose optimization is performed. The problem description can be expressed as:

[0027]

[0028] In the above formula, Represents the pose of the subgraph; represents the scanned poses, which are also relative to the world coordinate system; ξ ij Describes the relative pose relationship between the sub-image and the scan, that is, how to align the scan data to the sub-image;∑ ij Represents the covariance matrix between poses, which describes the uncertainty or error range of pose estimation;

[0029] The calculation formula of residual E is as follows:

[0030]

[0031] Furthermore, in step 2, the specific method of using the improved PSA algorithm to complete the global path planning of the automatic driving harvester according to the two-dimensional grid map is:

[0032] According to the drawn two-dimensional grid image, the decision variable is the area of ​​a single crop, the constraint is the area of ​​greenhouse farmland, and the objective function is the mileage of the autonomous harvester.

[0033] Assuming that the number of decision variables in the group is d, the upper and lower bounds of the variables are u and I respectively; the control parameters of PSA include the maximum number of iterations T and the population size n, then the initial population can be expressed as:

[0034] x ij =(u j -I j )·r1+I j ,i=1,2,…,n; j=1,2,…,d

[0035] In the formula, x ij represents the jth dimension of the ith individual; u j and I j are the upper and lower bounds of the jth variable (dimension), respectively; r1 is a random number between 0 and 1;

[0036] Assume that the objective function is F, and the objective function value of the i-th individual is f i for:

[0037] f i =F(x i )

[0038] For the problem of minimizing the mileage of the harvester, the best individual x at iteration t is * (t) is the individual corresponding to the historical minimum of the population; the deviation cycle of the population under several iterations t is e k-i (t) The overall deviation of the previous i iterations:

[0039] e k-i (t) = e k (ti)+x * (t)-x * (ti),i=1,2,…,t

[0040] When the number of iterations is t, the PID adjustment output value Δu(t) is expressed as:

[0041] Δu(t)=K p ·r2·[e k (t)-e k-1 (t)]+K i ·r3·e k (t)+Kd ·r4

[0042] ·[e k (t)-2e k-1 (t)+e k-2 (t)]

[0043] Where r2, r3, and r4 are random number vectors with n rows and 1 column from 0 to 1; K p , K i , K d They are the adjustment coefficients for proportion, integration and differentiation, and are adjusted according to the specific conditions of the farmland;

[0044] In addition to the original Δu(t), PSA also adds a regulation factor called zero output to prevent the algorithm from falling into the local optimum; zero output is defined as:

[0045] o(t)=(cos(1-t / T)+λr5·L)·e k (t)

[0046]

[0047] λ=[ln(T-t+2) / ln(T)] 2

[0048] Where r5 is a random number vector with n rows and d columns ranging from 0 to 1; L is the Lévy flight function, u and v are random number matrices with n rows and d columns that conform to the standard normal distribution, β is a factor set to 1.5; λ is an adjustment coefficient that decreases slowly with the increase of t, which helps the algorithm to fully explore; in the later stage, λ decreases rapidly, which helps the algorithm to switch from exploration to exploitation;

[0049] The update of all individuals is related to Δu(t) and o(t); the population update formula is defined as:

[0050] x(t+1)=x(t)+η·Δu(t)+(1-η)·o(t)

[0051] In the formula, η is a matrix with n rows and 1 column, expressed as η=r6cos(t / T), and r6 is a matrix with n rows and 1 column of 0 to 1 random numbers.

[0052] Furthermore, in step 2, an improved adaptive particle swarm optimization algorithm (PSO) will be introduced later to update the sampled particle swarm, thereby ensuring the diversity of correct particles; the specific steps are:

[0053] In the particle update process, dynamic differential mutation, crossover and selection operations are introduced to obtain all optimal solutions;

[0054] (1) Differential mutation: To obtain new individual particles, two particles in the group are processed by differential method, and the initial value of ζ is a constant real number;

[0055] In order to ensure the reliability of the value during the differential mutation operation, it is necessary to ensure the diversity of the initial population and improve the global search performance of the algorithm. Therefore, the adaptive mutation operator is introduced, and the formula is:

[0056]

[0057] Where: λ represents the difference; e represents the constant; G m and G represent the current number of iterations and the maximum number of iterations respectively. As the number of iterations increases, λ decreases gradually.

[0058] (2) Differential cross processing: The differential cross processing formula is:

[0059] u i,G+1 =(u 1i,G+1 ,u 2i,G+1 ,…,u Di,G+1 )

[0060]

[0061] Where: CR∈[0,1] represents the crossover probability; rnbr(i)∈[1,2,…,D] represents a random integer; u represents a newborn individual; i represents the number of particles; j represents the number of crossovers; v ji,G+1 、x ji,G Represents different cross population particles of output;

[0062] (3) Elite retention strategy: Solve the fitness value of the temporary population obtained in the above two steps, and at the same time, use a one-to-one method to complete the new population particle u i (t) and the original population particle x i (t) to compare the fitness values ​​and retain the elite; get the minimum value as the new individual of the next generation of particles, and use x i (t+1) indicates that the calculation formula is:

[0063]

[0064] Where: f(·) represents the fitness function.

[0065] Furthermore, in step 3, the xLSTM model is used to identify the collected image information to determine whether there are obstacles around the vehicle. The specific workflow is as follows:

[0066] When new image acquisition data enters the xLSTM network, the forget gate determines which data should be forgotten or retained, and the input gate determines which data should be retained and updates the cell state, controlling the weight of the new input at the current time step;

[0067] As the core of the entire xLSTM network, the cell state can store and transmit information, and can also control the flow and update of information; when the information of the current time step needs to be passed to the next layer or output layer, the output gate controls which data should be output;

[0068] sLSTM adds a scalar update mechanism to the traditional LSTM and optimizes the gating mechanism, making it more suitable for processing sequence data with subtle time changes:

[0069] c t =f t c t-1 +i t z t

[0070] n t =f t n t-1 +i t

[0071]

[0072] mLSTM greatly enhances the model's memory capacity and parallel processing capabilities by extending the vector operations in traditional LSTM to matrix operations:

[0073]

[0074] n t =f t n t-1 +i t k t

[0075]

[0076] q t =W q x t +b q

[0077]

[0078] v t =W v x t +b v

[0079]

[0080] In the formula, i, f and o represent the input, forget and output gates respectively, i t is the input gate output at time t, f t is the output of the forget gate at time t, o t is the output gate output at time t, c t is the memory unit, x t is the input vector, h t is the output vector, σ is the activation function, h t-1 is the hidden state of the previous time step, b z , b i , b f , b o , b q , b k , b v is the bias term, and W is the weight matrix of the forget gate.

[0081] Furthermore, in step 3, when there is an obstacle, the DCS algorithm is used to complete the local path planning of the working vehicle. The specific workflow is as follows:

[0082] DCS is a model for iteratively optimizing team performance that divides team members into three categories and assigns them responsibilities based on performance: (1) High achievers: These creative creators excel at divergent thinking and self-driven learning; they shine when creating a variety of innovative solutions; (2) Average performers: They are idea refiners who use convergent thinking to shape various ideas into viable solutions; (3) Low performers: These members are like newly recruited soldiers; they bring diversity to the team, promote exploration, and increase the pool of ideas; high achievers are good at creating different solutions and exploring different areas to prevent premature identification of suboptimal solutions; in contrast, average performers are good at improving solutions in promising areas.

[0083] The beneficial effects of the present invention are as follows: 1. The present invention combines the UAV and the millimeter-wave radar, giving full play to the respective functions of the UAV and the millimeter-wave radar, and improving the reliability and robustness of detection; 2. The present invention adopts the Cartographer algorithm to construct a two-dimensional grid map, and through local mapping, loop detection and global mapping, the two-dimensional grid map is constructed. Figure 3Steps to achieve more precise positioning and map construction; 3. The present invention introduces an improved adaptive particle swarm optimization algorithm (PSO) to update the sampled particle swarm to ensure the diversity of correct particles; in the particle update process, dynamic differential mutation, hybridization and selection operations are introduced to obtain all optimal solutions; 4. The present invention adopts a meta-heuristic algorithm PSA to perform global path planning for the harvester, so that the automatic driving harvester completes the harvesting of crops according to the optimal path; the algorithm is based on incremental PID control, and the entire population converges to the optimal state by adjusting the system deviation; 5. The present invention uses the xLSTM algorithm to test the fit of the obtained data, and finally obtains a complete two-dimensional grid map; xLSTM is an extension of the traditional LSTM, which improves LSTM by introducing a new gating mechanism and memory structure, aiming to improve the performance and scalability of LSTM when processing large-scale data; sLSTM is in the traditional A scalar update mechanism is added to the LSTM, and the gating mechanism is optimized to make it more suitable for processing sequence data with subtle time changes; mLSTM greatly enhances the model's memory capacity and parallel processing capabilities by extending the vector operations in traditional LSTM to matrix operations; each state of mLSTM is no longer a single vector, but a matrix, which allows it to capture more complex data relationships and patterns within a single time step; 6. The present invention adopts the DCS algorithm to complete the local path planning of the operating vehicle; the differentiated creative search (DCS) algorithm is a highly innovative meta-heuristic optimization method that uses differentiated knowledge acquisition and creative realism strategies to solve complex optimization problems; DCS effectively optimizes in a broad search space through mathematical modeling and implementation, and has high efficiency in solving complex optimization problems, strong evolutionary capabilities, fast search speed and excellent optimization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 is a system flow chart of an embodiment of the present invention;

[0085] Figure 2 is a flow chart of an improved PSA algorithm optimized by an adaptive particle swarm optimization algorithm in an embodiment of the present invention;

[0086] Figure 3 It is a flow chart of path planning of an automatic driving harvester in an embodiment of the present invention. DETAILED DESCRIPTION

[0087] The specific technical scheme of the present invention is further described in detail below with reference to specific examples.

[0088] As shown in the figure, the agricultural machinery automatic driving operation method applied to the automatic driving harvester described in the present invention takes the agricultural greenhouse as an example and is implemented by using a drone equipped with a millimeter-wave radar. The global path planning of the harvester is completed by the drone equipped with a millimeter-wave radar, and the identification and classification of obstacles inside the greenhouse are completed by the high-definition camera carried. The specific implementation steps are as follows:

[0089] Step 1: The drone is equipped with a millimeter-wave radar to collect information about the internal environment of the greenhouse, and the Cartographer algorithm is used to construct a two-dimensional grid map;

[0090] Specifically: The UAV is equipped with millimeter-wave radar to complete the collection of greenhouse environment point cloud data and complete the harvester driving environment modeling;

[0091] Point cloud data is a set of three-dimensional points used to describe the nonlinear surface shape of an object in three-dimensional space. Each point in the point set is represented by orthogonal length, width and height three-dimensional coordinates in space.

[0092] The point cloud data set collected by the millimeter wave radar is S = {s i}, where i∈N, s i =[x i ,y i, z i ] T ∈R 3 ;

[0093] Divide the 3D point set area into two parts with imaging side length d s A finite voxel set V with side length i , to ensure the integrity of the surface image captured by the viewpoint, the center point of the voxel is regarded as the voxel coordinate, and the calculation formula for selecting the viewpoint is:

[0094]

[0095] In the formula, p o is the point that the UAV track needs to pass through; P o Point cloud set of greenhouse environment collected by UAV millimeter wave radar;

[0096] Before obtaining a complete two-dimensional grid map, the data collected by the millimeter-wave radar needs to be optimized through the Cartographer algorithm to finally achieve the composition;

[0097] The Cartographer algorithm is used to build a two-dimensional grid map. Cartographer is a system that can provide 2D and 3D real-time simultaneous localization and mapping (SLAM) across multiple platforms and sensor configurations, using a grid map representation. The algorithm is mainly divided into two parts. The first part is called Local SLAM, which builds and maintains a series of submaps through laser scans frame by frame, and the so-called submap is a series of grid maps. When there is a new LaserScan, it will be inserted into the best position in the submap through the Ceres Scan Matching method. However, submaps will have the problem of error accumulation. Therefore, the second part of the algorithm, called Global SLAM, is to perform closed-loop detection through LoopClosure to eliminate accumulated errors. That is, when a submap is built, that is, no new laser scans will be inserted into the submap, the algorithm will add the submap to the closed-loop detection. The essence of closed-loop detection is also an optimization problem, which is expressed in the form of a pixel-accurate match. The method to solve the optimization problem is Branch-and-Bound Approach.

[0098] (1) Local mapping: The pose of a point in the point cloud data obtained by the radar in its own coordinate system is set to ξ = (ξ x ,ξ y ,ξ θ ), where ξ x , y Respectively represent the x-axis and y-axis translation of the point relative to the scanning origin, ξ θ Indicates the amount of rotation;

[0099] Before constructing the sub-map, it is necessary to transform the poses of all points in the point cloud data from the radar coordinate system to the corresponding sub-map coordinate system. Assume that the transformed pose is ξ ′ , then:

[0100]

[0101] Where: T ξ Represents the corresponding pose transformation, which is represented by the rotation matrix and translation vector Composition, that is

[0102] The Cartographer algorithm uses a probabilistic grid to represent the submap, inserting the point cloud data after coordinate transformation into the corresponding grid points in the probabilistic grid. Each insertion will update the probability value of the grid point. Before inserting the point cloud data, the Ceres library needs to be used to optimize the position of the scan point relative to the current submap to obtain the position with the highest probability of the scan point corresponding to the point cloud data in the submap. This process can be converted into a nonlinear least squares problem, and the formula is:

[0103]

[0104] Where: M smooth It is a bicubic interpolation function, which smoothes the probability values ​​of the corresponding grid points in the sub-map and obtains the maximum value by reversely solving the minimum value;

[0105] (2) Loop closure detection: Back-end loop closure detection is mainly used to offset the accumulated positioning error generated over time. In the Cartographer system, a series of small local sub-graphs are constructed by the front-end and gradually spliced ​​to construct a grid map. Although the positioning deviation of the sub-graph constructed with a small amount of scanning data is relatively small, as the running time increases, the small error will be superimposed and amplified. The Cartographer system design specifically introduces loop closure detection technology to correct the pose estimation in the sub-graph, thereby eliminating the negative impact of the accumulated error. In this process, the system records the specific pose information of each sub-graph in the world coordinate system and the actual pose state of the drone each time the radar scan frame is obtained. This information is subsequently used for loop closure detection.

[0106] In order to improve the accuracy and consistency of the map, a loop detection mechanism is introduced to determine whether there is any overlap between the sub-images at all previous moments and the radar information at the current moment, that is, whether a loop occurs; once a loop is found, the system will start the global optimization mapping process, adjusting and optimizing the relative poses between the sub-images to eliminate the accumulated errors, thereby building a more accurate and coherent map.

[0107] The core essence of loop detection is a matching analysis process, which mainly analyzes the latest scan frame data captured in real time, and uses it as a reference to systematically explore and confirm the best matching frame that meets the predetermined similarity standard in the preset search area around the current frame, so as to successfully construct an effective loop closure; as shown in the formula:

[0108]

[0109] In the search space W, the above formula is used to filter the best frame among the matching frames;

[0110] (3) Global Mapping: After loop closure detection, global pose optimization is performed. The problem description can be expressed as:

[0111]

[0112] In the above formula, Represents the pose of the subgraph; represents the scanned poses, which are also relative to the world coordinate system; ξ ij Describes the relative pose relationship between the sub-image and the scan, that is, how to align the scan data to the sub-image;∑ ij Represents the covariance matrix between poses, which describes the uncertainty or error range of pose estimation; the calculation formula of the residual E is as follows:

[0113]

[0114] Step 2: According to the two-dimensional grid map drawn in step 1, the improved PSA algorithm is used to complete the global path planning of the automatic driving harvester. At the same time, the improved adaptive particle swarm optimization algorithm (PSO) is introduced to update the sampled particle swarm to ensure the diversity of correct particles;

[0115] Specifically, according to the grid map drawn by the Cartographer algorithm, the improved PSA algorithm is used to complete the global path planning of the harvester; the PSA algorithm is a meta-heuristic algorithm based on incremental PID control, which consists of a set of decision variables, constraints and an objective function;

[0116] According to the drawn two-dimensional grid image, the decision variable is the area of ​​a single crop, the constraint is the area of ​​greenhouse farmland, and the objective function is the mileage of the autonomous harvester.

[0117] Assuming that the number of decision variables in the group is d, the upper and lower bounds of the variables are u and I respectively; the control parameters of PSA include the maximum number of iterations T and the population size n, then the initial population can be expressed as:

[0118] x ij =(u j -I j )·r1+I j ,i=1,2,…,n; j=1,2,…,d

[0119] In the formula, x ij represents the jth dimension of the ith individual; u j and I j are the upper and lower bounds of the jth variable (dimension); r1 is a random number between 0 and 1; assuming the objective function is F, the objective function value f of the i-th individual i for:

[0120] f i =F(x i)

[0121] For the problem of minimizing the mileage of the harvester, the best individual x at iteration t is * (t) is the individual corresponding to the historical minimum of the population; the deviation cycle of the population under several iterations t is e k-i (t) The overall deviation of the previous i iterations:

[0122] e k-i (t) = e k (ti)+x * (t)-x * (ti),i=1,2,…,t

[0123] When the number of iterations is t, the PID adjustment output value Δu(t) is expressed as:

[0124] Δu(t)=K p ·r2·[e k (t)-e k-1 (t)]+K i ·r3·e k (t)+K d ·r4

[0125] ·[e k (t)-2e k-1 (t)+e k-2 (t)]

[0126] Where r2, r3, and r4 are random number vectors with n rows and 1 column from 0 to 1; K p , K i , K d They are the adjustment coefficients of proportion, integration and differentiation, which are adjusted according to the specific conditions of the farmland. In addition to the original Δu(t), PSA also adds a regulation factor called zero output to prevent the algorithm from falling into the local optimum. Zero output is defined as:

[0127] o(t)=(cos(1-t / T)+λr5·L)·e k (t)

[0128]

[0129] λ=[ln(T-t+2) / ln(T)] 2

[0130] Where r5 is a random number vector with n rows and d columns ranging from 0 to 1; L is the lévy flight function, u and v are random number matrices with n rows and d columns that conform to the standard normal distribution, β is a factor set to 1.5; λ is an adjustment coefficient that decreases slowly with the increase of t, which helps the algorithm to fully explore; in the later stage, λ decreases rapidly, which helps the algorithm to switch from exploration to exploitation;

[0131] The update of all individuals is related to Δu(t) and o(t); the population update formula is defined as:

[0132] x(t+1)=x(t)+η·Δu(t)+(1-η)·o(t)

[0133] Where η is a matrix with n rows and 1 column, expressed as η = r6cos(t / T), where r6 is a matrix with n rows and 1 column containing 0 to 1 random numbers;

[0134] However, in the PSA algorithm, the continuous iteration and resampling of the particle swarm will lead to the degradation of particle diversity, and the correct particles will still be discarded, making it impossible to achieve a better global optimal solution search.

[0135] Furthermore, in order to ensure that the correct particles are not lost, the present invention introduces an improved adaptive particle swarm optimization algorithm (PSO) to update the sampled particle swarm to ensure the diversity of correct particles; in the particle update process, dynamic differential mutation, hybridization and selection operations are introduced to obtain all optimal solutions;

[0136] Particle swarm optimization is a heuristic optimization algorithm that simulates the foraging and other production behaviors of groups such as bird flocks or fish schools in nature:

[0137]

[0138] In the above formula, ω is the inertia weight used to control the local and global search capabilities of the equilibrium particle; c1 and c2 are learning factors; r1 and r2 are randomly generated random numbers in the [0,1] interval used to update the particle's speed and position; V i t and V i t+1 is the velocity of particle i at time t and time t+1; and is the position of particle i at time t and time t+1;

[0139] (1) Differential mutation: To obtain new individual particles, two particles in the group are processed by differential method, and the initial value of ζ is a constant real number;

[0140] In order to ensure the reliability of the value during the differential mutation operation, it is necessary to ensure the diversity of the initial population and improve the global search performance of the algorithm. Therefore, the adaptive mutation operator is introduced, and the formula is:

[0141]

[0142] Where: λ represents the difference; e represents the constant; G m and G represent the current number of iterations and the maximum number of iterations respectively. As the number of iterations increases, λ decreases gradually.

[0143] (2) Differential cross processing: The differential cross processing formula is:

[0144] u i,G+1 =(u 1i,G+1 ,u 2i,G+1 ,…,u Di,G+1 )

[0145]

[0146] Where: CR∈[0,1] represents the crossover probability; rnbr(i)∈[1,2,…,D] represents a random integer; u represents a newborn individual; i represents the number of particles; j represents the number of crossovers; v ji,G+1 、x ji,G Represents different cross population particles of output;

[0147] (3) Elite retention strategy: Solve the fitness value of the temporary population obtained in the above two steps, and at the same time, use a one-to-one method to complete the new population particle u i (t) and the original population particle x i (t) to compare the fitness values ​​and retain the elite; get the minimum value as the new individual of the next generation of particles, and use x i (t+1) indicates that the calculation formula is:

[0148]

[0149] Where: f(·) represents the fitness function;

[0150] For the problem of minimizing the mileage of the harvester, the best individual x at iteration t is * (t) is the individual corresponding to the historical minimum of the population; the deviation cycle of the population under several iterations t is e k-i (t) The overall deviation of the previous i iterations:

[0151] e k-i (t) = e k (ti)+x * (t)-x *(ti),i=1,2,…,t

[0152] When the number of iterations is t, the PID adjustment output value Δu(t) is expressed as:

[0153] Δu(t)=K p ·r2·[e k (t)-e k-1 (t)]+K i ·r3·e k (t)+K d ·r4

[0154] ·[e k (t)-2e k-1 (t)+e k-2 (t)]

[0155] Where r2, r3, and r4 are random number vectors with n rows and 1 column from 0 to 1; K p , K i , K d They are the adjustment coefficients for proportion, integration and differentiation, and are adjusted according to the specific conditions of the farmland;

[0156] In addition to the original Δu(t), PSA also adds a regulation factor called zero output to prevent the algorithm from falling into the local optimum; zero output is defined as:

[0157] o(t)=(cos(1-t / T)+λr5·L)·e k (t)

[0158]

[0159] λ=[ln(T-t+2) / ln(T)] 2

[0160] Where r5 is a random number vector with n rows and d columns ranging from 0 to 1; L is the Lévy flight function, u and v are random number matrices with n rows and d columns that conform to the standard normal distribution, β is a factor set to 1.5; λ is an adjustment coefficient that decreases slowly with the increase of t, which helps the algorithm to fully explore; in the later stage, λ decreases rapidly, which helps the algorithm to switch from exploration to exploitation;

[0161] The update of all individuals is related to Δu(t) and o(t); the population update formula is defined as:

[0162] x(t+1)=x(t)+η·Δu(t)+(1-η)·o(t)

[0163] Where η is a matrix with n rows and 1 column, expressed as η = r6cos(t / T), where r6 is a matrix with n rows and 1 column containing 0 to 1 random numbers;

[0164] After multiple iterations, the global optimal path planning of the autonomous harvester was obtained based on the improved PSA algorithm.

[0165] Step 3: Considering that in addition to crops, there are various obstacles (such as stones, ground piles, etc.) in the greenhouse that hinder the passage of the harvester; the harvester is equipped with a ZED binocular camera to collect image information within 5-6m around the vehicle, and the xLSTM model is used to identify the collected image information to determine whether there are obstacles around the vehicle; if there are obstacles, the DCS algorithm is used to complete the local path planning of the operating vehicle.

[0166] ZED is a binocular camera produced by STEREOLABS, which is widely used in unmanned vehicles and collaborative robotic arms of scientific research institutions. Its 3D resolution can reach RGB resolution in Ultra mode. Its physical size is 175×30×33mm, the frame rate can reach 100FPS, the depth range is 0.3-25m, the maximum field of view is 90°(H)×60°(V)×100°(D), and the interface is USB3.0.

[0167] xLSTM is an extension of traditional LSTM. It improves LSTM by introducing new gating mechanisms and memory structures, aiming to improve the performance and scalability of LSTM when processing large-scale data. There are five main differences between xLSTM and original LSTM: First, xLSTM introduces an exponential gating mechanism, which is a new gating technology. Unlike traditional sigmoid gating, exponential gating can provide more dynamic information filtering capabilities, which helps to improve the memory and forgetting process. Second, a new memory hybrid technology is added to the sLSTM single memory architecture, which still maintains scalar updates, but improves the storage and utilization efficiency of information through improved hybrid methods. mLSTM introduces matrix memory, which allows parallel processing and improves storage capacity. It uses covariance update rules, which is suitable for processing large-scale parallel data and solves the limitations of LSTM in parallelization. system; 3. In order to prevent numerical stability problems caused by exponential gating, xLSTM introduces additional normalization and stabilization steps in the gating calculation, such as using the maximum value recording method to maintain stability; 4. xLSTM integrates these improved LSTM units into residual blocks, which are further stacked to form a complete network architecture. This design enables xLSTM to process complex sequence data more effectively; 5. xLSTM is comparable to the latest Transformer and state-space models in performance, especially in large-scale applications and long sequence processing. Compared with LSTM, xLSTM expands LSTM into multiple LSTM variants: sLSTM (scalar LSTM) and mLSTM (matrix LSTM), each variant is optimized for specific performance and functions to handle various complex sequence data problems;

[0168] sLSTM adds a scalar update mechanism to the traditional LSTM and optimizes the gating mechanism, making it more suitable for processing sequence data with subtle time changes; namely:

[0169] c t =f t c t-1 +i t z t

[0170] n t =f t n t-1 +i t

[0171]

[0172] mLSTM greatly enhances the model's memory and parallel processing capabilities by extending the vector operations in traditional LSTM to matrix operations. Each state of mLSTM is no longer a single vector, but a matrix, which allows it to capture more complex data relationships and patterns in a single time step; namely:

[0173]

[0174] n t =f t n t-1 +i t k t

[0175]

[0176] q t =W q x t +b q

[0177]

[0178] v t =W v x t +b v

[0179]

[0180] In the formula, i, f and o represent the input, forget and output gates respectively, i t is the input gate output at time t, f t is the output of the forget gate at time t, o t is the output gate output at time t, c t is the memory unit, x t is the input vector, h t is the output vector, σ is the activation function, h t-1 is the hidden state of the previous time step, b z 、b i 、b f 、b o 、b q 、b k 、b v is the bias term, and W is the weight matrix of the forget gate.

[0181] The workflow of xLSTM is as follows: when new input enters the xLSTM network, the forget gate determines which information should be forgotten or retained, and the input gate determines which information should be retained and updates the cell state, controlling the weight of the new input at the current time step; as the core of the entire xLSTM network, the cell state can store and transmit information, and can also control the flow and update of information; when the information of the current time step needs to be passed to the next layer or output layer, the output gate controls which information should be output.

[0182] The xLSTM model is used to calculate the weights of the image data collected by the ZED camera, then collect and calculate again, continuously update the particle position, and repeat the process. Finally, obstacles around the harvester are identified, and the DCS algorithm is used to complete the local path planning of the harvester.

[0183] The Differentiated Creative Search (DCS) algorithm is an extremely innovative meta-heuristic optimization method that uses differentiated knowledge acquisition and creative realism strategies to solve complex optimization problems. DCS effectively optimizes in a broad search space through mathematical modeling and implementation. It has high efficiency in solving complex optimization problems, strong evolutionary capabilities, fast search speed and excellent optimization capabilities.

[0184] Using the concepts of differentiated knowledge acquisition and creative realism, an innovation optimization model called differentiated creative search (DCS) is proposed; DCS is a model for iteratively optimizing team performance that divides team members into three categories and assigns them responsibilities based on their performance; (1) High achievers: These creative creators excel in divergent thinking and self-driven learning; they shine when creating a variety of innovative solutions; (2) Average performers: idea refiners who use convergent thinking to shape various ideas into viable solutions; (3) Low performers: These members are like newly recruited soldiers; they bring diversity to the team, promote exploration, and increase the pool of ideas; the role of divergent thinking and convergent thinking is shown in the figure; high achievers are good at creating different solutions and exploring different areas to prevent premature identification of suboptimal solutions; in contrast, average performers are good at improving solutions in promising areas.

[0185] In DCS, the optimization process starts with a set of candidate solutions (team members) X, which are randomly generated between the upper bound U and the lower bound L of the optimization problem; that is:

[0186] x i,d =LB d +U(0,1)×(UB d -LB d )

[0187] In the formula, U(0,1) represents the uniform distribution on the interval (0,1), LBd and UB d denote the lower and upper bounds of the d-th dimension specified in the optimization problem, respectively.

[0188] The difference between DKA and DE lies mainly in the basic concept, which is more similar to natural process rather than the random principle embodied in DE. Differentiated knowledge acquisition (DKA) focuses on the speed of new knowledge acquisition, which has many effects on individuals. This change can be seen from how many attributes or dimensions of an individual's existing knowledge are changed by the new knowledge. In the mathematical formula, the parameter η,t is the quantitative knowledge acquisition rate (q KR ); the calculation formula is as follows:

[0189]

[0190] In the formula, the symbol [·] represents a symbol indicating that a given value is rounded to the nearest integer, where is the individual's variable at the tth iteration The coefficient value is calculated as follows:

[0191]

[0192] In the formula, is the individual variable at the tth iteration The value of R i,t is the ranking (order) of the i-th individual at the beginning of the t-th iteration, The coefficient is a quantitative determinant of the degree of imperfection of a person's knowledge and varies with the size of the knowledge gap; a higher value means a larger knowledge gap, indicating that the individual needs to learn, absorb and integrate new knowledge or experience; conversely, The smaller the value, the less imperfect the knowledge is, indicating that the individual's knowledge base is more comprehensive and solid;

[0193] The effect of the DKA process on each individual X can be performed using the following formula:

[0194] j rand =randint(1,D)

[0195]

[0196] In the formula, v i,d is the element at the dth position (dimension) of the test member, v i,t is the trial member v of the tth iteration i , where x i,d Represents x i The element at the dth position (dimension) in ; U(0,1) represents the uniform distribution on the interval (0,1), η i,trepresents the quantitative knowledge acquisition rate of the individual at the tth iteration (q KR ), j rand is an integer randomly chosen from 1 to D and generated once for each sample.

[0197] The DCS / Xbest / current-to-rand strategy relies on the knowledge base of the top performers and combines the random contributions from two different team members, which are brought by the current individual; therefore, diversity and convergence are promoted in the generated solutions; this approach reflects the interaction between different teams and is formulated as follows:

[0198] v i,d =w×X best +λ t ×(X r2 -X i )+ω i,t ×(X r1 -X i )

[0199] In the formula, λ t is the coefficient value λ of the individual at the tth iteration; the coefficient expression is:

[0200]

[0201] In the above example, X best is the dth position of the best performing individual in the current iteration, w represents the cognitive weight of the best performance, and the default value is 1; ω i,t is the ω coefficient value of individual i at the tth iteration; where X r1 is the dth position of an individual randomly selected from {1,2,…,NP} and X r1 ≠X≠X best represents three different individuals (or indices); t is the coefficient value λ at the tth iteration; X r2 is the dth position of an individual randomly selected from {ngs+1,…,NP} and X r1 ≠X r2 ≠X≠X best ; The parameter ngs represents the number of high achievers and is determined by 100p%; For larger groups, covering 20% ​​of the population, 'p' is set to 0.2; For groups with a population size of 30 or less, the minimum value of 'ngs' is 6.

[0202] The new strategy DCS / Xrand / Linnik (α, σ) formula is as follows:

[0203] v i,d =X r1 +Lk(α,σ)

[0204] In the formula, v i,d The element at the dth position (dimension) of the vector (called member) is the experimental member of X, where represents the set of experimental members of team X; X r1 is the dth position of an individual randomly selected from {1,2,…,NP} and X r1 ≠X≠X best indicates three different individuals (or indices), these individuals must be different; X best represents the team member with the best target value in the tth iteration, and Lk(α,σ) represents the Linnik distribution random number generator with control parameters α and σ.

[0205] A team with a constantly changing membership produces more diverse ideas, so the algorithm systematically replaces underperforming members with new members; the formula for generating new members is as follows:

[0206] V NP =LB+U(0,1)×(UB-LB)

[0207] Where V NP represents the NP-th vector, U(0,1) represents the uniform distribution on the interval (0,1), LB and UB are the lower and upper bounds of the optimization problem;

[0208] Retrospective Assessment (RA) is a key tool for team development. It establishes evaluation criteria as a benchmark to measure success and then analyzes past performance data to identify trends and areas for improvement. This reflection allows the team to recognize its strengths, weaknesses, decision-making, individual contributions, and communication dynamics. RA fosters a culture of continuous improvement, enabling teams to learn from experience, make informed decisions, and optimize collaboration. Therefore, the systematic implementation of RA can improve team dynamics and achieve better results. The study simplified the RA framework to focus on selection and best performance tracking, which is a basic but essential process in metaheuristic algorithms. The equation for the selection process is as follows:

[0209]

[0210] Where, X i,t+1 represents the i-th individual in X at the t+1 iteration.

[0211] After multiple iterations, the local optimal path planning of the autonomous harvester was obtained based on the DCS algorithm, so as to avoid obstacles and complete the harvesting of crops.

[0212] Global path planning requires obtaining environmental information inside the greenhouse and collecting various data through environmental maps to perform path planning; while local path planning only requires the camera to collect surrounding environmental signals in real time, and determine the orientation and obstacle distribution by understanding the surrounding environmental map, so as to select the best path from the current node to the target node; combining global path planning and local path planning methods, the autonomous driving harvester can efficiently complete crop harvesting in the greenhouse, while performing dynamic obstacle avoidance through local path planning.

Claims

1. An agricultural machinery automatic driving operation method applied to an automatic driving harvester, characterized in that: The operation steps are as follows: Step (1): Use millimeter wave radar to collect greenhouse internal environment information and use Cartographer algorithm to construct a two-dimensional grid map; Step (2): Based on the grid map in step (1), the improved PSA algorithm is used to complete the global path planning of the automatic driving harvester; Step (3): The autonomous harvester is equipped with a ZED camera to collect image information around the vehicle, and the xLSTM model is used to identify the collected image information to determine whether there are obstacles around the vehicle. If there are obstacles, the DCS algorithm is used to complete the local path planning of the operating vehicle.

2. The method for automatic driving of agricultural machinery applied to an automatic driving harvester according to claim 1, characterized in that: In step (1), the specific method of using millimeter wave radar to collect greenhouse internal environment information is: The UAV is equipped with millimeter-wave radar to collect point cloud data of the greenhouse environment and model the harvester driving environment; Among them, point cloud data is a set of three-dimensional points that describe the nonlinear surface shape of an object in three-dimensional space, and each point in the point set is represented by orthogonal length, width and height three-dimensional coordinates in space; The point cloud data set collected by the millimeter wave radar is S = {s i }, where i∈N, s i =[x i ,y i, z i ] T ∈R 3 ; The three-dimensional point set area is divided into two parts with imaging side length d s A finite voxel set V with side length i , to ensure the integrity of the surface image captured by the viewpoint, the center point of the voxel is regarded as the voxel coordinate, and the calculation formula for selecting the viewpoint is: In the formula, p o is the point that the UAV track needs to pass through; P o Point cloud set of greenhouse environment collected by UAV millimeter wave radar.

3. The method for automatic driving of agricultural machinery applied to an automatic driving harvester according to claim 1, characterized in that: In step (1), the specific method of using the Cartographer algorithm to construct a two-dimensional grid map is: (1.1) Local mapping: The pose of a point in the point cloud data obtained by the radar in its own coordinate system is set to ξ = (ξ x ,ξ y ,ξ θ ), In the formula, ξ x , y Respectively represent the x-axis and y-axis translation of the point relative to the scanning origin, ξ θ Indicates the amount of rotation; Before constructing the sub-map, the poses of all points in the point cloud data are converted from the radar coordinate system to the corresponding sub-map coordinate system. Assume that the pose after conversion is ξ ′ , then: Where: T ξ Represents the corresponding pose transformation, which is represented by the rotation matrix and translation vectors Composition, that is (1.2) Loop detection: Loop detection is an in-depth analysis of the latest scan frame data captured in real time, and using it as a reference to systematically explore and confirm the best matching frame that meets the predetermined similarity standard in the preset search area around the current frame, thereby constructing a loop closure; as shown in the following formula: In the search space W, the above formula is used to filter the best frame among the matching frames; (1.3) Global Mapping: After loop closure detection, global pose optimization is performed. The problem description is expressed as: In the formula, Represents the pose of the subgraph; represents the scanned pose, ξ ij Describes the relative pose relationship between the sub-image and the scan, that is, how to align the scan data to the sub-image;∑ ij Represents the covariance matrix between poses, which describes the uncertainty or error range of pose estimation; The calculation formula of residual E is as follows:

4. The method for automatic driving of agricultural machinery applied to an automatic driving harvester according to claim 3, characterized in that: In step (1.1), the Cartographer algorithm uses a probability grid to represent the submap, inserting the point cloud data after coordinate transformation into the corresponding grid points in the probability grid. Each insertion will update the probability value of the grid point. Before inserting the point cloud data, the Ceres library is used to optimize the position of the scan point relative to the current sub-map to obtain the position with the highest probability of the scan point corresponding to the point cloud data in the sub-map. This process is converted into a nonlinear least squares problem, and the formula is: Where: M smooth It is a bicubic interpolation function, which smoothes the probability values ​​of the corresponding grid points in the submap and obtains the maximum value by reversely solving the minimum value.

5. The method for automatic driving of agricultural machinery applied to an automatic driving harvester according to claim 1, characterized in that: In step (2), the specific method of using the improved PSA algorithm to complete the global path planning of the automatic driving harvester according to the two-dimensional grid map is: According to the drawn two-dimensional grid image, the decision variable is the area of ​​a single crop, the constraint is the area of ​​greenhouse farmland, and the objective function is the mileage of the autonomous harvester. Assume that the number of decision variables in this group is d, the upper and lower bounds of the variables are u and I respectively; the control parameters of PSA include the maximum number of iterations T and the population size n, then the initial population is expressed as: x ij =(u j -I j )·r1+I j ,i=1,2,…,n;j=1,2,…,d In the formula, x ij represents the jth dimension of the ith individual; u j and I j are the upper and lower bounds of the jth variable respectively; r1 is a random number between 0 and 1; Assume that the objective function is F, and the objective function value of the i-th individual is f i for: f i =F(x i ) For the problem of minimizing the mileage of the harvester, the best individual x at iteration t is * (t) is the individual corresponding to the historical minimum of the population; the deviation cycle of the population under several iterations t is e k-i (t) The overall deviation of the previous i iterations: e k-i (t)=e k (t-i)+x * (t)-x * (t-i),i=1,2,…,t When the number of iterations is t, the PID adjustment output value Δu(t) is expressed as: Δu(t)=K p ·r2·[e k (t)-e k-1 (t)]+K i ·r3·e k (t)+Kd·r4·[e k (t)-2e k-1 (t)+e k-2 (t)] Where r2, r3, and r4 are random number vectors with n rows and 1 column from 0 to 1; K p , K i , K d They are the adjustment coefficients for proportion, integration and differentiation, and are adjusted according to the specific conditions of the farmland; On the original Δu(t), PSA adds a regulation factor called zero output to prevent the algorithm from falling into the local optimum; zero output is defined as: o(t)=(cos(1-t / T)+λr5·L)·e k (t) λ=[ln(T-t+2) / ln(T)] 2 Where r5 is a random number vector with n rows and d columns ranging from 0 to 1; L is the Lévy flight function, u and v are random number matrices with n rows and d columns that conform to the standard normal distribution, and β is a factor set to 1.5; λ is the adjustment coefficient, and it decreases slowly as t increases; In the later stage, λ decreases rapidly, which helps the algorithm shift from exploration to exploitation; The update of all individuals is related to Δu(t) and o(t); the population update formula is defined as: x(t+1)=x(t)+η·Δu(t)+(1-η)·o(t) In the formula, η is a matrix with n rows and 1 column, expressed as η=r6cos(t / T), and r6 is a matrix with n rows and 1 column of 0 to 1 random numbers.

6. The method for automatic driving of agricultural machinery applied to an automatic driving harvester according to claim 1, characterized in that: In step (2), the improved adaptive particle swarm optimization algorithm is finally introduced to update the sampled particle swarm. The specific steps are as follows: (1) Differential mutation: To obtain new individual particles, two particles in the group are processed by differential method, and the initial value of ζ is a constant real number; In order to ensure the reliability of the value during the differential mutation operation, ensure the diversity of the initial population and improve the global search performance of the algorithm, the adaptive mutation operator is introduced. The formula is: Where: λ represents the difference; e represents the constant; G m and G represent the current number of iterations and the maximum number of iterations respectively. As the number of iterations increases, λ decreases gradually. (2) Differential cross processing: The differential cross processing formula is: in i,G+1 =(in 1i,G+1 ,in 2i,G+1 ,…,in Di,G+1 ) Where: CR∈[0,1] represents the crossover probability; rnbr(i)∈[1,2,…,D] represents a random integer; u represents a newborn individual; i represents the number of particles; j represents the number of crossovers; v ji,G+1 、x ji,G Represents different cross population particles of output; (3) Elite retention strategy: Solve the fitness value of the temporary population obtained in the above two steps, and at the same time, use a one-to-one method to complete the new population particle u i (t) and the original population particle x i (t) fitness value comparison, elite retention, get the minimum value as the new individual of the next generation of particles, and use x i (t+1) indicates that the calculation formula is: Where: f(·) represents the fitness function.

7. The method for automatic driving of agricultural machinery applied to an automatic driving harvester according to claim 1, characterized in that: In step (3), the autonomous harvester is equipped with a ZED camera to collect image information within 5-6 m around the vehicle.

8. The method for automatic driving of agricultural machinery applied to an automatic driving harvester according to claim 1, characterized in that: In step (3), the workflow of using the xLSTM model to identify the collected image information and determine whether there are obstacles around the vehicle is as follows: When new image acquisition data enters the xLSTM network, the forget gate determines which data should be forgotten or retained, and the input gate determines which data should be retained and updates the cell state, controlling the weight of the new input at the current time step; As the core of the entire xLSTM network, the cell state can store and transmit information while controlling the flow and update of information; When the information of the current time step needs to be passed to the next layer or output layer, the output gate controls which data should be output; sLSTM adds a scalar update mechanism to the traditional LSTM and optimizes the gating mechanism, making it more suitable for processing sequence data with subtle time changes: c t =f t c t-1 +i t z t n t =f t n t-1 +i t mLSTM extends the vector operations in traditional LSTM to matrix operations, enhancing the model's memory capacity and parallel processing capabilities: n t =f t n t-1 +i t k t q t =W q x t +b q v t =W v x t +b v In the formula, i, f and o represent the input, forget and output gates respectively, i t is the input gate output at time t, f t is the output of the forget gate at time t, o t is the output gate output at time t, c t is the memory unit, x t is the input vector, h t is the output vector, σ is the activation function, h t-1 is the hidden state of the previous time step, b z 、b i 、b f 、b o 、b q 、b k 、b v is the bias term, and W is the weight matrix of the forget gate.

9. The method for automatic driving of agricultural machinery applied to an automatic driving harvester according to claim 1, characterized in that: In step (3), the obstacles include but are not limited to ground piles and large rocks.

10. The method for automatic driving of agricultural machinery applied to an automatic driving harvester according to claim 1, characterized in that: In step (3), if there is an obstacle, the workflow of using the DCS algorithm to complete the local path planning of the working vehicle is as follows: DCS is a model for iteratively optimizing team performance that divides team members into three categories: (1) High achievers: They excel in divergent thinking and self-driven learning; (2) Average performers: They are idea refiners who use convergent thinking to shape various ideas into viable solutions; (3) Low performers: They bring diversity to the team, promote exploration, and increase the idea pool.

Images