Data-driven distributed learning control method for multi-vehicle network system

Abstract

The invention provides a data-driven distributed learning control method for a multi-vehicle network system, and the method comprises the steps: firstly, designing different test inputs for each individual vehicle, executing test iteration, and collecting corresponding output data; each individual vehicle determines a gain matrix of a distributed learning control algorithm by using output data collected by the individual vehicle; according to the topology of the multi-vehicle network system and the expected target output receiving condition, each individual vehicle determines an adjacent weight, and a gain parameter is determined based on the adjacent weight; and in an adjustment iteration stage, each individual vehicle establishes a distributed learning control algorithm by using a gain matrix, a gain parameter, collected input data, an adjacent weight and other data, updates control input during next iteration, and controls the multi-vehicle network system to execute the next iteration. According to the invention, all individual vehicles in a multi-vehicle network system with unknown model information can be ensured to realize high-precision real-time tracking of expected target output within the preset operation time.

Description

Technical Field

[0001] This invention relates to the field of intelligent control of unmanned systems, and specifically to a data-driven distributed learning control method for multi-vehicle network systems. Background Technology

[0002] Distributed learning control, a smart control method combining distributed control and iterative learning control, has gained widespread attention in the field of cooperative control of complex network systems in recent years. Distributed control updates control strategies in real time through local information interaction between individual systems and their neighbors, ensuring the network system achieves global cooperative control goals. It possesses good applicability, robustness, and flexibility, and is particularly suitable for complex scenarios where the desired goal is known only to a subset of systems. Iterative learning control, on the other hand, focuses on repetitive tasks, collecting relevant experience information during the iterative operation of the system. Through experience accumulation and learning, iterative optimization of control inputs is achieved, gradually bringing the system output closer to the target output to achieve high-precision control. Distributed learning control combines the advantages of both methods, enabling complex network systems to gradually improve their global cooperative performance through neighbor information interaction and experience learning. Therefore, it is widely used in fields such as UAV swarms, intelligent transportation, and distributed energy management. Currently, the design of most distributed learning control algorithms relies on the model information of the network system. However, the increasing complexity of network system structures and the presence of various disturbances make it difficult to obtain accurate models, rendering traditional distributed learning control methods inapplicable. Furthermore, the selection of gain parameters in distributed learning algorithms often depends on the global topology information of the network system, which is typically not known to the individuals within the network. These issues limit the practical application of existing distributed learning control algorithms.

[0003] Currently, for multi-vehicle network systems, distributed learning control methods are generally only applicable to systems where the individual vehicle model information is known. Furthermore, the design of distributed learning control algorithms requires knowledge of global topology information, thus having significant limitations and failing to meet the practical needs of production and daily life. Summary of the Invention

[0004] In view of this, the present invention provides a data-driven distributed learning control method for multi-vehicle network systems. This method does not rely on the model information of the multi-vehicle network system, but only uses the collected input and output data to design a distributed learning control algorithm. The selection of the gain parameter depends only on the local information of the multi-vehicle network system. It aims to provide a generally applicable distributed control design method for the autonomous high-precision cooperative problem of multi-vehicle network systems with completely unknown model information.

[0005] To solve the above-mentioned technical problems, the present invention is implemented as follows.

[0006] A data-driven distributed learning control method for a multi-vehicle network system includes:

[0007] Step A: Design different test input data for each individual vehicle, execute test iterations, and collect corresponding test output data;

[0008] Step B: Each individual vehicle uses its collected test output data to determine the gain matrix K of the distributed learning control algorithm. i Construct the input data matrix U using the test input data. i Embedded in a distributed learning control algorithm, and by selecting test input data, the input data matrix U is optimized. i Non-singular;

[0009] Step C: Based on the topology of the multi-vehicle network system and the desired output reception, each individual vehicle determines the first adjacency weight α that indicates whether it can directly communicate with its neighbor j. ij The second adjacency weight d, and whether it can receive the expected target output instruction. j The gain parameter μ of the distributed learning control algorithm is determined based on two adjacency weights. i ;

[0010] Step D: Each individual vehicle utilizes the gain matrix K i Gain parameter μ i Input data matrix U i A distributed learning control algorithm is established based on two adjacency weights, the control input and output of the vehicle under its current adjustment iteration, the output of the neighboring vehicle under its current adjustment iteration, and the desired target output. This algorithm updates the control input of the individual vehicle for the next iteration and implements control.

[0011] Step E: The multi-vehicle network system collaborates to perform multiple iterations until all individual vehicles achieve target tracking.

[0012] Preferably, in step A, designing different test input data for each individual vehicle, performing test iterations, and collecting corresponding test output data involves:

[0013] For individual vehicle i, execute n i +1 test iterations; each test iteration has N runtimes, and the test input data for all N runtimes in the first test iteration is 0; the subsequent n... i In this test iteration, the test input data at time 0 of the h-th test iteration is 0, and the test input data at times 1 to N-1 are not 0, denoted as u. i,h h∈{2,3,…,n i +1};

[0014] After ni The input data matrix is ​​composed of test input data at non-zero time points of each test iteration. By selecting test input data, U i Non-singular;

[0015] Based on the designed test input data, perform test iterations and collect the corresponding test output data. This refers to the test output data generated by individual vehicle i at time t during the h-th test iteration.

[0016] Preferably, the step of selecting test input data to make U i Non-singular means: in the last n i In each test iteration, the test input for the h-th test iteration is selected based on the test inputs chosen in the previous h-1 test iterations to ensure the input data matrix U is... i The non-singularity of.

[0017] Preferably, in step B, each individual vehicle uses the test output data it has collected to determine the gain matrix K of the distributed learning control algorithm. i The steps include:

[0018] Step b1: Individual vehicle i constructs output data matrix Y based on the collected test output data. i :

[0019]

[0020] Among them, y i,l Construct a method using the difference between the test output data of each test iteration and the test output data of the first test iteration, and let... For individual vehicle i, execute n i +1 test iteration, This refers to the test output data generated by individual vehicle i at time t during the h-th test iteration; h∈{2,3,…,n} i +1};

[0021] Step b2: Individual vehicle i selects the transformation matrix Makes span(H) i,1 = span(Y) i ), where m i For Y i The rank of Nn o For Y i The number of rows; span(Y) i ) represents Y i The generation space;

[0022] Step b3: Select matrix Make H i =[H i,1 H i,2 It is not singular;

[0023] Step b4: H i The inverse matrix is in,

[0024] Step b5: Select the gain matrix K for the distributed learning control algorithm corresponding to individual vehicle i. i It meets the following conditions:

[0025]

[0026] Where, 0 < γ i ≤1 represents any arbitrarily chosen design parameter. It is a unit vector.

[0027] Preferably, the gain matrix K is selected. i At that time, further F i,1 Y i Full rank property and H i,1 Full rank property, guaranteeing that the gain matrix K satisfies the selection condition for the gain matrix. i If it definitely exists, then the gain matrix is:

[0028]

[0029] Preferably, in step C, the first adjacency weight α ij Second Adjacency Weight d j The determination is as follows:

[0030] If individual vehicle i can receive information from individual vehicle j, then individual vehicle j is called a neighboring vehicle of individual vehicle i. Let α ij >0; otherwise, α ij =0;

[0031] If individual vehicle i can directly receive the desired target output, then d i >0; otherwise, let d i =0.

[0032] Preferably, in step C, the gain parameter μ of the distributed learning control algorithm is determined based on two adjacency weights. i for:

[0033]

[0034] in, Let v represent the set of neighbors of vehicle i. jLet j represent a neighbor in the neighbor set.

[0035] Preferably, step D includes:

[0036] Individual vehicle i obtains its output y under the current adjustment iteration k. i,k (t), t∈{0,1,…,N} and the output of neighboring vehicle j in the current iteration. Let N represent the set of neighbors of individual vehicle i, and let N represent the N running times contained in one iteration.

[0037] The desired target output y is obtained. d (t), t∈{1,2,…,N};

[0038] Individual vehicle i, based on the control input u under the current adjustment iteration i,k (t), Selected gain parameter μ i Input data matrix U i The selected gain matrix K i The two adjacent weights α ij and d j The output y of individual vehicle i under the current adjustment iteration i,k (t), Output y under the current adjustment iteration of neighboring vehicles j,k (t) and the desired target output y d (t), the distributed learning control algorithm is established as follows:

[0039]

[0040] in, The gain matrix K i The submatrix that satisfies Let v represent the set of neighbors of individual vehicle i. j Let j represent a neighbor in the neighbor set.

[0041] Beneficial effects:

[0042] 1) The method of this invention is applicable to multi-vehicle network systems where model information is completely unknown. First, different test inputs are designed for each individual vehicle. Test iterations are performed under these test inputs to collect output data that reflects vehicle characteristics. The collected input and output data can then be used to design a distributed learning control algorithm, where the input data is directly embedded in the algorithm, and the output data is used for selecting the gain matrix. Therefore, the design of the distributed learning control algorithm no longer requires known model information of the multi-vehicle network system, ensuring that all individual vehicles in the system can achieve high-precision tracking of the desired target output. This reduces the dependence of most current distributed learning control algorithm designs on multi-vehicle network system model information.

[0043] 2) The gain matrix K of the distributed learning control algorithm in the method of this invention i The choice is flexible. The selection of the gain matrix for each individual vehicle depends only on the output data Y collected by that vehicle itself. i Therefore, each individual vehicle has a different gain matrix in the distributed learning control algorithm, and it is not required to satisfy the full row / column rank condition. This is more flexible than the design of traditional distributed learning control algorithms that require each individual vehicle to have the same gain matrix.

[0044] 3) In a preferred embodiment, the present invention provides a specific expression for the gain matrix when the row rank is full or the column rank is full, which makes the selection of the gain matrix simpler and eliminates the need for complex calculations or design.

[0045] 4) The method of this invention no longer requires vehicles in a multi-vehicle network system to have quasi-regularity. This is achieved by selecting a suitable transformation matrix H. i,1 Effectively mining the essential dynamic characteristics of vehicles reflected in the data allows for the selection of the gain matrix based on these characteristics. Compared with traditional distributed learning control methods that are only applicable to quasi-regular multi-vehicle network systems, this method has greater general applicability.

[0046] 5) The method of this invention no longer requires knowledge of the global topology information of the multi-vehicle network system. The selection of the gain parameter of each individual vehicle depends only on its adjacency weight with respect to neighboring vehicles. Therefore, only the neighbor information is needed, and the distributed learning control algorithm can be designed using only local information. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the architecture of a data-driven distributed learning control method for a multi-vehicle network system provided in an embodiment of the present invention.

[0048] Figure 2 A flowchart illustrating the data-driven distributed learning control method for a multi-vehicle network system provided in an embodiment of the present invention.

[0049] Figure 3 This is a schematic diagram of the topology of a multi-vehicle network system provided in an embodiment of the present invention.

[0050] Figure 4 This is a schematic diagram illustrating the dynamic evolution of the tracking error along the iteration axis in a multi-vehicle network system provided in an embodiment of the present invention.

[0051] Figure 5 This is a schematic diagram illustrating the dynamic evolution of the outputs of all individual vehicles and the desired target output along the time axis in a multi-vehicle network system provided in an embodiment of the present invention. Detailed Implementation

[0052] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0053] This invention provides a data-driven distributed learning control method for multi-vehicle network systems. It eliminates the need for prior knowledge of the precise model information of the multi-vehicle network system. Each individual vehicle can establish a distributed learning control algorithm based on its own input and output data collected during test iterations, ensuring the completion of high-precision collaborative tracking tasks in the multi-vehicle network system. In this embodiment, the method includes the following steps:

[0054] Step S1: Design different test inputs for each individual vehicle, execute test iterations, and collect the corresponding output data.

[0055] In some embodiments, step S1 specifically includes the following steps:

[0056] A multi-vehicle network system consists of n individual vehicles, each of which has the following dynamics:

[0057]

[0058] Where t∈{0,1,…,N} is the running time, N is the running period, and each iteration takes input data at time 0-(N-1). The input at time N-1 determines the output at time N. This represents the number of iterations. and Let n be the state, control input, and output of the i-th vehicle at time t in the k-th iteration. s n i and n o Let A, B, and C be the dimensions of the state, control input, and output, respectively. Let A, B, and C be the system matrices corresponding to each vehicle. Without loss of generality, let the initial state of the i-th vehicle be iteratively invariant, i.e. Meanwhile, its corresponding relative degree matrix satisfies CB≠0. The system matrices A, B, and C, and the initial state x of the individual vehicle... i,0 , All are unknown.

[0059] Perform n operations on the i-th vehicle i +1 test iterations, where the h-th test iteration is denoted as its T-th iteration. h This is the Nth iteration. Each test iteration has N runtimes. The test input is chosen as:

[0060] ● First test iteration: Each test iteration has N runtimes. Represented as...

[0061] ·after n i Test iterations: after n i In this test iteration, the test input data at time step 0 of the h-th test iteration is 0, while the test input data at times 1 to N-1 are not 0. This is represented as...

[0062] {2,3,…,n i +1}. Where, the last n i The input data matrix is ​​composed of test input data at non-zero time points of each test iteration. By selecting test input data, U i It is a non-singular matrix.

[0063] Given the test input, executing the corresponding test iterations allows us to collect the corresponding test output data. This refers to the test output data generated by individual vehicle i at time t during the h-th test iteration.

[0064] In a preferred solution, in the last n i In each test iteration, the test input for the h-th test iteration can be selected based on the test inputs chosen in the previous h-1 test iterations, to ensure that the input data matrix U i The non-singularity of the input data matrix U. i The non-singularity of this ensures that the collected output data can reflect the complete system matrix information of individual vehicles, and that U i Embedded control algorithms can effectively overcome the impact of different control designs for different individual vehicles using different data.

[0065] Step S2: Each individual vehicle uses the test output data it has collected to determine the gain matrix K of the distributed learning control algorithm. i .

[0066] In this step, for the gain matrix K i The selection method specifically includes the following steps:

[0067] S21: Individual vehicle i constructs an output data matrix based on the collected test output data:

[0068]

[0069] Among them, y i,l Construct a method using the difference between the test output data of each test iteration and the test output data of the first test iteration, and let... For individual vehicle i, execute n i +1 test iteration, This refers to the test output data generated by individual vehicle i at time t during the h-th test iteration; h∈{2,3,…,n} i +1}.

[0070] S22: Individual vehicle i selects the transformation matrix Makes span(H) i,1 = span(Y) i ), where m i For Y i The rank of span(Y). i ) represents Y i The generation space.

[0071] S23: Corresponding selection matrix Make H i =[H i,1 H i,2 ] is non-singular. Additionally, let H be a denoted H. i The inverse matrix is in,

[0072] S25: Select a gain matrix K for the distributed learning control algorithm corresponding to individual vehicle i that satisfies the following conditions. i :

[0073]

[0074] Where, 0 < γ i ≤1 represents any arbitrarily chosen design parameter. The gain matrix is ​​a unit vector. The above-described gain matrix design conditions effectively ensure that, even when each individual vehicle selects a different gain matrix using its own collected output data, all vehicles can still effectively track the same desired target output.

[0075] In a preferred embodiment, the gain matrix K i The choice is usually not unique, F i,1 Y i full rank of rows and H i,1The full rank property of the column matrix can guarantee that the gain matrix K satisfies the above conditions. i It must exist, and the gain matrix can be specifically chosen as... This scheme provides a specific expression for the gain matrix when the row rank is full or the column rank is full, which simplifies the selection of the gain matrix and eliminates the need for multiple complex calculations or designs.

[0076] Step S3: Each individual vehicle obtains its own output in the current iteration, the outputs of its neighboring vehicles in the current iteration, and the desired target output.

[0077] In some embodiments, step S3 specifically includes the following steps:

[0078] The i-th vehicle obtains its output y in the current iteration. i,k (t), t∈{0,1,…,N} and the output of neighboring vehicles in the current iteration. At the same time, obtain the desired target output. If the i-th vehicle can receive the desired target output information, then let d i >0, otherwise, let d i =0. Each vehicle can receive the desired target output directly or indirectly through neighboring vehicles. However, only vehicles d that directly receive the desired target output will receive the output. i >0.

[0079] Step S4: Obtain the topology of the multi-vehicle network system. Each individual vehicle, based on its adjacency weights in the topology, combines them with another adjacency weight d, which represents the reception status of the desired target output information. i Determine the gain parameter μ of the distributed learning control algorithm. i .

[0080] In some embodiments, step S4 specifically includes the following steps:

[0081] S41: The topology of a multi-vehicle network system consists of a directed graph with n nodes. To describe, among which, For a set of nodes, Let be the set of edges. Let α be the adjacency weight matrix. ij Values ​​≥0 represent adjacency weights. α ij >0 indicates that individual vehicle i can receive information from individual vehicle j, in which case individual vehicle j is called individual vehicle i's neighbor vehicle; otherwise, α ij =0. Therefore, the i-th vehicle can obtain its own initial state information x. i,0 and the initial status information of its neighboring vehicles. in, Let i be the set of all neighboring vehicles of the i-th vehicle.

[0082] S42: Individual vehicle i is determined by its adjacency weight α ij and d i Determine the gain parameter μ of the distributed learning control algorithm. i as follows:

[0083]

[0084] Make I n -ML is a random matrix, where I n It is an identity matrix of dimension n, M = diag{μ1,μ2,…,μ n Let} be a diagonal matrix, and L be a directed graph. The Laplace matrix is ​​used to ensure the convergence of the vehicle system.

[0085] Since each vehicle can receive the desired target output information directly or indirectly through neighboring vehicles. It holds true, therefore μ must exist. i A value greater than 0 makes the above selection condition for the gain parameter true.

[0086] Gain parameter μ for each individual vehicle i The selection depends only on its adjacency weights relative to neighboring vehicles. Therefore, the design of distributed learning control algorithms can be carried out using only local information. It is no longer necessary to know the global topology information of the multi-vehicle network system; only the neighbor information is required.

[0087] Step S5: Each individual vehicle establishes a distributed learning control algorithm based on the control input of the current iteration, the selected gain parameters, the collected input data, the selected gain matrix, the adjacency weights, the output of the individual vehicle in the current iteration, the output of the neighboring vehicles in the current iteration, and the desired target output. It then updates the control input for the next iteration and controls the multi-vehicle network system to execute the next iteration.

[0088] In some embodiments, step S5 specifically includes the following steps:

[0089] S51: Individual vehicle i is determined by the control input u in the current iteration. i,k (t), Selected gain parameter μ i The collected input data U i The selected gain matrix K i Adjacency weight α ij and d j The output y of an individual vehicle in the current iteration i,k (t), the output y of the neighboring vehicle in the current iteration j,k (t) and the desired target output y d(t), establish a distributed learning control algorithm, update the control input for the next iteration, and control the multi-vehicle network system to execute the next iteration. The specific distributed learning control algorithm is as follows:

[0090]

[0091] in, The gain matrix K i The submatrix that satisfies

[0092] Step S6: Return to step S5 for the next iteration. The multi-vehicle network system iterates multiple times until all individual vehicles complete high-precision tracking of the desired target output.

[0093] It should be noted that in the control method provided in the embodiments of the present invention, the execution order of steps S2, S3 and S4 is not fixed and is not limited here.

[0094] The following describes in detail a data-driven distributed learning control method for a multi-vehicle network system according to the present invention with reference to embodiments, but it should not be construed as limiting the scope of protection of the present invention.

[0095] Example 1

[0096] Consider a multi-vehicle network system consisting of 6 individual vehicles, each with the following dynamics:

[0097]

[0098] Wherein, system matrices A, B, and C are respectively:

[0099]

[0100] The operating cycle of the multi-vehicle network system is N=200. The initial state of each individual vehicle is:

[0101] x 1,0 =[1,0] T x 2,0 =[0,1] T x 3,0 =[0,-1] T ;

[0102] x 4,0 =[3,0] T x 5,0 =[0,3] T x 6,0 =[0,-3] T .

[0103] S1: Perform two test iterations on the i-th vehicle, selecting the test input as:

[0104] ●First test iteration:

[0105] ●Second test iteration: in,

[0106]

[0107] Therefore, U i =[u i,2 ] is a non-singular matrix.

[0108] Given the test input, perform two corresponding test iterations. This allows us to collect the corresponding test output.

[0109] S2: Each individual vehicle uses its collected output data to determine the gain matrix of the distributed learning control algorithm. This includes the following steps:

[0110] S21: The i-th vehicle constructs an output data matrix based on the collected output data:

[0111]

[0112] Among them, y i,l satisfy

[0113] S22: The transformation matrix H for the i-th vehicle selection i,1 =Y i Makes span(H) i,1 = span(Y) i );

[0114] S23: Select the gain matrix for the distributed learning control algorithm corresponding to the i-th vehicle. as well as thus as well as Established.

[0115] S3: Each individual vehicle obtains its own output in the current iteration. Output of neighboring vehicles in the current iteration And the expected target output:

[0116]

[0117] Figure 3 It also reflects the individual vehicle's reception of the desired target output, by Figure 3It can be seen that each individual vehicle can obtain the information of the desired target output directly or indirectly from its neighboring vehicles.

[0118] S4: Obtain the topology of the multi-vehicle network system. Each individual vehicle determines the gain parameters of the distributed learning control algorithm based on the adjacency weights in the topology. This includes the following steps:

[0119] S41: The topology of the multi-vehicle network system is... Figure 3 The diagram shown consists of a directed graph with 6 nodes. describe.

[0120] S42: The i-th vehicle determines the gain parameter of the distributed learning control algorithm based on the adjacency weights in the topology. Therefore, the following is satisfied:

[0121]

[0122] S5: Each individual vehicle establishes a distributed learning control algorithm based on the control input of the current iteration, the selected gain parameters, the collected input data, the selected gain matrix, the adjacency weights, the output of the individual vehicle in the current iteration, the output of neighboring vehicles in the current iteration, and the desired target output. This algorithm updates the control input for the next iteration and controls the multi-vehicle network system to execute the next iteration. Specifically, this includes the following steps:

[0123] S51: The i-th vehicle establishes a distributed learning control algorithm based on the control input of the current iteration, the selected gain parameters, the collected input data, the selected gain matrix, the adjacency weights, the output of the individual vehicle in the current iteration, the output of the neighboring vehicles in the current iteration, and the desired target output. This algorithm updates the control input for the next iteration and controls the multi-vehicle network system to execute the next iteration. The specific distributed learning control algorithm is as follows:

[0124]

[0125] in, The gain matrix K i The submatrix that satisfies The initial control input for each individual vehicle in the distributed learning control algorithm is:

[0126] u 1,0 (t)=1, u 2,0 (t)=-1,u 3,0 (t) = 3,

[0127] u 4,0 (t)=-2,u 5,0 (t)=2, u 6,0 (t) = -3.

[0128] S52: Apply the updated control input to the multi-vehicle network system to control the multi-vehicle network system to execute the next iteration process.

[0129] S6: Return to S5 for the next iteration. The multi-vehicle network system iterates multiple times until all individual vehicles achieve high-precision tracking of the desired target output.

[0130] Figure 4 The figure illustrates the dynamic evolution of the tracking error of a multi-vehicle network system along the iteration axis. It shows that the tracking error of all individual vehicles' outputs relative to the desired target output gradually decreases along the iteration axis, eventually converging to zero. Furthermore, Figure 5 This figure illustrates the dynamic evolution of the outputs of all individual vehicles in a multi-vehicle network system along the time axis after 195 iterations. It demonstrates that the outputs of all individual vehicles in the multi-vehicle network system consistently achieve high-precision tracking of the desired output at any given time. Therefore, the data-driven distributed learning control method for multi-vehicle network systems provided by this invention can overcome the influence of unknown models in multi-vehicle network systems. By designing the distributed learning control algorithm using only input and output data, it ensures that each individual vehicle in a multi-vehicle network system with non-regular dynamics achieves high-precision real-time tracking of the desired output within a preset running time.

[0131] The specific embodiments described above only illustrate the design principles of the present invention. The shapes and names of the components in this description may differ and are not limited. Therefore, those skilled in the art can modify or make equivalent substitutions to the technical solutions described in the foregoing embodiments; and these modifications and substitutions do not depart from the inventive spirit and technical solutions of the present invention, and should all fall within the protection scope of the present invention.

Claims

1. A data-driven distributed learning control method for a multi-vehicle network system, characterized in that, include: Step A: Design different test input data for each individual vehicle, execute test iterations, and collect corresponding test output data; Step B: Each individual vehicle uses its collected test output data to determine the gain matrix K of the distributed learning control algorithm. i Construct the input data matrix U using the test input data. i Embedded in a distributed learning control algorithm, and by selecting test input data, the input data matrix U is optimized. i Non-singular; Step C: Based on the topology of the multi-vehicle network system and the desired output reception, each individual vehicle determines the first adjacency weight α that indicates whether it can directly communicate with its neighbor j. ij The second adjacent weight d, indicating whether it can receive the desired target output instruction. j The gain parameter μ of the distributed learning control algorithm is determined based on two adjacency weights. i ; Step D: Each individual vehicle utilizes the gain matrix K i Gain parameter μ i Input data matrix U i A distributed learning control algorithm is established based on two adjacency weights, the control input and output of the vehicle under its current adjustment iteration, the output of the neighboring vehicle under its current adjustment iteration, and the desired target output. This algorithm updates the control input of the individual vehicle for the next iteration and implements control. Step E: The multi-vehicle network system collaborates to perform multiple iterations until all individual vehicles achieve target tracking.

2. The method as described in claim 1, characterized in that, In step A, designing different test input data for each individual vehicle, executing test iterations, and collecting corresponding test output data are as follows: For individual vehicle i, execute n i +1 test iteration; Each test iteration has N runtimes. In the first test iteration, all N runtimes correspond to 0 as the test input data; in the subsequent n iterations... i In this test iteration, the test input data at time 0 of the h-th test iteration is 0, and the test input data at times 1 to N-1 are not 0, denoted as u. i,h h∈{2,3,…,n i +1}; After n i The input data matrix is ​​composed of test input data at non-zero time points of each test iteration. By selecting test input data, U i Non-singular; Based on the designed test input data, perform test iterations and collect the corresponding test output data. This refers to the test output data generated by individual vehicle i at time t during the h-th test iteration.

3. The method as described in claim 2, characterized in that, The method of selecting test input data makes U i Non-singular means: in the last n i In each test iteration, the test input for the h-th test iteration is selected based on the test inputs chosen in the previous h-1 test iterations to ensure the input data matrix U is... i The non-singularity of.

4. The method as described in claim 1, characterized in that, In step B, each individual vehicle uses the test output data it has collected to determine the gain matrix K of the distributed learning control algorithm. i The steps include: Step b1: Individual vehicle i constructs output data matrix Y based on the collected test output data. i : Among them, y i,l Construct a method using the difference between the test output data of each test iteration and the test output data of the first test iteration, and let... For individual vehicle i, execute n i +1 test iteration, This refers to the test output data generated by individual vehicle i at time t during the h-th test iteration; h∈{2,3,…,n} i +1}; Step b2: Individual vehicle i selects the transformation matrix Makes span(H) i,1 = span(Y) i ), where m i For Y i The rank of Nn o For Y i The number of rows; span(Y) i ) represents Y i The generation space; Step b3: Select matrix Make H i =[H i,1 H i,2 It is not singular; Step b4: H i The inverse matrix is in, Step b5: Select the gain matrix K for the distributed learning control algorithm corresponding to individual vehicle i. i It meets the following conditions: Where 0 < γi ≤ 1 are arbitrarily chosen design parameters. It is a unit vector.

5. The method as described in claim 4, characterized in that, Choose the gain matrix K i At that time, further F i,1 Y i Full rank property and H i,1 Full rank property, guaranteeing that the gain matrix K satisfies the selection condition for the gain matrix. i If it definitely exists, then the gain matrix is:

6. The method as described in claim 1, characterized in that, In step C, the first adjacency weight α ij Second adjacent weight d j The determination is as follows: If individual vehicle i can receive information from individual vehicle j, then individual vehicle j is called a neighboring vehicle of individual vehicle i. Let α ij >0; otherwise ，α ij ＝0； If individual vehicle i can directly receive the desired target output, then d i >0; otherwise, let d i =0.

7. The method as described in claim 1 or 6, characterized in that, In step C, the gain parameter μ of the distributed learning control algorithm is determined based on two adjacency weights. i for: in, Let v represent the set of neighbors of vehicle i. j Let j represent a neighbor in the neighbor set.

8. The method as described in claim 1, characterized in that, Step D includes: Individual vehicle i obtains its output y under the current adjustment iteration k. i,k (t), t∈{0,1,…,N} and the output of neighboring vehicle j in the current iteration. Let N represent the set of neighbors of individual vehicle i, and let N represent the N running times contained in one iteration. The desired target output y is obtained. d (t), t∈{1,2,…,N}; Individual vehicle i, based on the control input u under the current adjustment iteration i,k (t), Selected gain parameter μ i Input data matrix U i The selected gain matrix K i The two adjacent weights α ij and d j The output y of individual vehicle i under the current adjustment iteration i,k (t), Output y under the current adjustment iteration of neighboring vehicles j,k (t) and the desired target output y d (t), the distributed learning control algorithm is established as follows: in, The gain matrix K i The submatrix that satisfies Let v represent the set of neighbors of individual vehicle i. j Let j represent a neighbor in the neighbor set.

Images