Data driving control method based on sliding window and feature point selection in time-varying scene

Abstract

The invention discloses a data driving control method based on a sliding window and feature point selection in a time-varying scene, and relates to the technical field of adaptive control of data driving, and the method comprises the steps: collecting input and output signals of a time-varying system in real time through a sensor; dynamically updating an input / output track data structure based on a sliding window mechanism to ensure that the data structure always reflects the current dynamic characteristics of the system; feature point selection is carried out on the trajectory data structure through singular value decomposition, data information is quantized, dimension reduction processing is carried out, and a dimension reduction data structure is generated; constructing a control optimization objective function based on the dimension reduction data structure, solving a decision variable and generating a control instruction; and balancing the data structure dimension and the information completeness through an adaptive order adjustment mechanism. According to the method, the data is updated in real time through a sliding window mechanism, and the data dimension is selected and optimized based on the feature points of singular value decomposition, so that the timeliness, the calculation efficiency and the control precision of time-varying system data driving control are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of data-driven adaptive control technology, and in particular to a data-driven control method based on sliding window and feature point selection in time-varying scenarios. Background Technology

[0002] With the development of technology, the complexity of industrial systems is increasing, and they generally exhibit strong coupling, nonlinearity, or unknown parameters. Traditional system modeling and control paradigms, which rely on mechanistic analysis or system identification to obtain accurate mathematical models, are extremely difficult and costly, and their feasibility faces significant challenges. Even with tremendous efforts, it is difficult to construct accurate and reliable models.

[0003] Faced with this dilemma, data-driven control technology has emerged. The core advantage of this technology lies in its ability to construct a controller directly from offline or online acquired input and output data, without relying on a mathematical model of the system. This significantly reduces the dependence on prior knowledge, enabling effective system control even in the absence of a precise model. A typical implementation of this technology uses the specific matrix structure Hankel as the core data carrier. By recombining the input data sequence u(t) and the output data sequence y(t), an input matrix H is constructed. u and output matrix H y And optimize the linear combination coefficients of the Hankel matrix. This method is used to reproduce the dynamic characteristics of the system and generate control inputs. It has been validated in fixed-parameter systems, i.e., linear time-invariant (LTI) systems, and simplifies the control design process and improves the control accuracy of complex systems, such as in the simulation control of autonomous vehicles (its principle flowchart is shown in...). Figure 1 As shown in the figure, it can quickly and accurately generate control commands to ensure driving safety.

[0004] However, in real-world control scenarios, the characteristic of system parameters changing over time is prevalent. Taking autonomous vehicles driving on different road surfaces as an example, the tire friction coefficient dynamically changes with the environment. This not only breaks the assumption of a fixed-parameter system (LTI) but also exposes significant limitations of existing data-driven control technologies. Specifically, this manifests in the following three aspects: First, statically constructed Hankel matrices cannot adapt to dynamically changing system characteristics. Existing technologies use fixed offline datasets to construct the Hankel matrix. For systems with changing parameters, the data collected in the early stages only reflects historical states and cannot capture the system characteristics at the current moment. This leads to a continuous increase in the deviation between the predicted input / output and the actual system response, resulting in a continuous decrease in control accuracy and even causing system instability.

[0005] Second, excessively increasing the dimensionality of the Hankel matrix leads to a decline in real-time performance. Traditional methods tend to increase the amount of offline data to cover more system dynamics, but the number of columns in the Hankel matrix increases linearly with the amount of data, while the computational complexity of the control optimization problem is positively correlated with the square of the number of columns. In scenarios with changing parameters, adding unlimited amounts of data will cause the optimization solution time to exceed the upper limit of the control cycle, failing to meet real-time control requirements.

[0006] Third, the lack of key dynamic information may make it difficult to maintain the incentive conditions. The effectiveness of data-driven control depends on the input data satisfying the continuous incentive (PE) condition, i.e., the Hankel matrix H... u The rank is not less than the product of the system input dimension and the prediction window length: ; Where m is the input dimension and L is the prediction window length, to ensure that any input trajectory can be reproduced through linear combination. As the dynamic characteristics of the system evolve over time due to parameter changes, the static dataset may contain repetitive and invalid trajectories or omit input-output patterns under new operating conditions, resulting in the loss of key dynamic information and disrupting the excitation (PE) conditions, ultimately causing the control model to lose its ability to express the system dynamics.

[0007] While existing technologies have undergone some modifications to address the aforementioned challenges, significant shortcomings remain. Some methods dynamically update the Hankel matrix using a sliding window mechanism, which to some extent ensures data timeliness. However, the dimension of the Hankel matrix is ​​determined by both the window length and the amount of data. Its number of rows is determined by the window length L and is independent of the data volume, but its number of columns equals the number of valid data segments—that is, the total number of non-overlapping trajectory segments generated when the sliding window slides across the data sequence. If the data volume continues to increase, the number of columns in the Hankel matrix will grow linearly over time because its column count is calculated as N–L + 1, where N is the data volume and L is the window length. Therefore, simply updating the matrix by incorporating new data without actively limiting the number of columns will cause the number of columns in the Hankel matrix to grow indefinitely with system runtime, eventually exceeding the capacity of computing resources.

[0008] Existing technologies also employ dimensionality reduction methods based on global data statistical characteristics to compress the matrix, but they fail to consider the timeliness of the system's dynamic characteristics under parameter changes. In new operating conditions, key trajectories may be misjudged as noise and discarded, resulting in the compressed matrix failing to reflect the current system dynamics and thus reducing control accuracy. Furthermore, some adaptive control methods adjust the controller online by identifying changing parameters, but this relies on the premise that the system structure is known, and the identification process is susceptible to noise interference. Existing data-driven control is not deeply integrated with adaptive mechanisms, making it difficult to balance control accuracy and system stability in scenarios with changing parameters. Summary of the Invention

[0009] In view of this, the present invention proposes a data-driven control method based on sliding window and feature point selection in time-varying scenarios to solve the problem of insufficient timeliness of data-driven control in time-varying scenarios in the prior art.

[0010] The specific technical solution of this invention is as follows: Data-driven control methods based on sliding window and feature point selection in time-varying scenarios include: The input and output signals of the time-varying system are collected in real time by sensors, and the signals are normalized. The input and output trajectory data structure is dynamically updated based on the sliding window mechanism to include the latest data and remove outdated data, ensuring that the data structure always reflects the current dynamic characteristics of the system. Singular value decomposition is used to select feature points in the trajectory data structure, quantify the data information, and perform dimensionality reduction to generate a dimensionality-reduced data structure. A control optimization objective function is constructed based on a dimensionality-reduced data structure, the decision variables are solved, and control commands are generated. By balancing the data structure dimensionality and information completeness through an adaptive order adjustment mechanism, dynamic optimization of control precision and computational efficiency can be achieved.

[0011] Specifically, the sliding window mechanism includes setting the window length, sliding step size, and fixed column limit; the window length limits the time step coverage of a single trajectory segment, the sliding step size controls the data update frequency, and the fixed column limit constrains the growth of the column dimension of the trajectory data structure, so as to avoid the data structure from accumulating data volume and causing the computational burden to exceed the real-time control requirements.

[0012] Specifically, feature point selection is based on informational quantization using singular value decomposition; this includes: calculating the minimum non-zero singular value of the trajectory data structure to evaluate the data's informational value; when the minimum non-zero singular value is greater than a preset threshold, the data is determined to possess key dynamic information and dimensionality reduction is performed; when the minimum non-zero singular value is less than a preset threshold, the data is determined to be redundant and a sliding window is triggered to update the data.

[0013] Specifically, the singular value decomposition dimensionality reduction process adopts an update algorithm based on historical decomposition results. This algorithm uses the singular value decomposition results of the previous time step to calculate the new dimensionality reduction solution of the current data structure through incremental update, avoiding repeated calculation of the complete singular value decomposition, and significantly reducing the computational complexity to meet the real-time control time limit.

[0014] Specifically, the incremental update algorithm includes: using the singular value decomposition results of the previous time step, decomposing the new data into orthogonal components of the historical subspace, constructing a low-dimensional incremental matrix and diagonalizing it, and finally fusing them to generate a reduced-order matrix for the current time step.

[0015] Specifically, the incremental update algorithm reduces computational complexity to linear levels by performing the following steps: calculating the orthogonal components of the new data based on the historical left / right singular vector subspace; constructing a low-dimensional incremental matrix and performing fast diagonalization; and fusing the historical singular value decomposition results with the incremental decomposition results.

[0016] Specifically, the adaptive order adjustment mechanism is implemented by iteratively traversing the sequence of singular values; starting from the minimum order, the order is gradually increased until the current singular value is less than a preset threshold, at which point the traversal stops and the previous order is locked as the final dimensionality reduction order.

[0017] Specifically, the adaptive order dynamically adapts to time-varying systems: when the system is dynamically volatile, the order is increased to enhance the retention of dynamic information; when the system is dynamically stable, the order is maintained at the minimum to reduce computational redundancy.

[0018] Specifically, the control optimization objective function is constructed based on a dimensionality-reduced data structure, and the decision variables are solved through rolling optimization. The decision variables are used to generate control commands and are fed back to the data acquisition step in a closed loop, forming a continuously updated control loop to adapt to the dynamic changes of the time-varying system.

[0019] Specifically, the method is applicable to autonomous driving scenarios for unmanned vehicles; the input signal corresponds to the vehicle steering angle control quantity, and the output signal corresponds to the vehicle position and heading angle state quantity. The robustness of trajectory tracking is improved through control commands driven by dynamic data.

[0020] The beneficial effects of this invention are as follows: 1. By dynamically updating the trajectory data structure through a sliding window mechanism, the latest system operation data is continuously incorporated, effectively solving the problem of insufficient timeliness of static data, improving the real-time tracking capability of the dynamic characteristics of time-varying systems, and reducing the decline in control performance caused by model lag.

[0021] 2. The update algorithm based on historical singular value decomposition results uses the decomposition results of the previous moment for incremental updates, replacing the repeated singular value decomposition of the complete trajectory data structure, which significantly reduces the computational burden in the control process and better meets the real-time requirements of industrial control scenarios.

[0022] 3. By quantitatively evaluating the key characteristics of the trajectory data structure, the structural dimension is adaptively determined to ensure that information reflecting the core dynamics of the system is preserved during dimensionality reduction, while balancing the computational burden. This is beneficial to enhancing the adaptability and overall performance of the control system in time-varying environments.

[0023] 4. The trajectory operation is based entirely on system input and output data. It does not rely on internal system models or parameter knowledge. It can automatically filter out key information fragments that characterize the current dynamics of the system based on real-time running data, thereby improving the adaptability and control effect of the method on time-varying systems with unknown or changing parameters.

[0024] 5. By quantifying key characteristics and using dimensional adaptive constraints to maintain data excitation characteristics, the reduced trajectory data structure is guaranteed to always have good information completeness, which is conducive to the stable operation of the closed-loop control system and provides support for the reliability of the control scheme. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a schematic diagram of the simulation control process for unmanned vehicles based on data-driven model predictive control in the existing technology; Figure 2 This is a flowchart illustrating the data-driven control method based on sliding window and feature point selection in time-varying scenarios according to the present invention. Detailed Implementation

[0027] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0028] This invention proposes a data-driven control method based on sliding window and feature point selection for time-varying scenarios, such as... Figure 2As shown, this is a data-driven control strategy based on a fusion of "sliding window + Hankel matrix reduction based on singular value decomposition (SVD)". The core of this method lies in updating data in real time through a sliding window mechanism and optimizing data dimensionality using feature point selection techniques, thereby achieving precise data-driven control of time-varying systems. Specifically, this method aims to achieve the following objectives: First, by dynamically updating the Hankel matrix using a sliding window, it ensures that the data used reflects the latest dynamics of the time-varying system in real time, effectively solving the problem of insufficient timeliness of static data; second, by combining informative evaluation with numerical singular value decomposition (SVD) optimization algorithms, the Hankel matrix is ​​reduced in dimensionality, and the results are verified through simulations of autonomous driving scenarios for unmanned vehicles, ultimately providing an efficient and robust data-driven control scheme for time-varying scenarios such as chemical engineering, robotics, and autonomous driving.

[0029] This invention uses data-driven control of unmanned vehicles as a typical application scenario. Its overall architecture revolves around a closed-loop process: data acquisition and processing, data construction, time-varying scenario optimization, and data-driven control and verification. The specific implementation steps are executed by computer equipment, encompassing a data acquisition module, a matrix construction module, an optimization solution module, and a control execution module. The specific process is as follows: First, the time-varying system data is acquired and preprocessed. Computer equipment acquires the input and output signals of the time-varying system in real time through sensors. Taking data-driven control of an autonomous vehicle as an example, the input data... Dimensions Where m is the number of system input channels, in this scenario m = 2, representing the steering angles of the left and right front wheels of the autonomous vehicle, respectively; t is the discrete time, and the sampling period Ts is determined by the dynamic characteristics of the system, which is taken here as... Output data Dimensions Where p is the number of system output channels, in this scenario p = 3, representing the autonomous vehicle's position information x, y, and heading angle yaw in the Cartesian coordinate system, respectively. To avoid instability in subsequent matrix calculations due to differences in the magnitude of input and output data, the collected data needs to be normalized, and the normalized data needs to be concatenated into a sequence in chronological order: ; This step ensures that the data volume based on the normalized data sequence is uniform, eliminates the interference of numerical differences on the calculation and optimization of the Hankel matrix, and provides standardized input for subsequent steps.

[0030] Next, a dynamic Hankel matrix is ​​constructed based on the sliding window mechanism. The core parameters of the sliding window include the window length L (representing the number of time steps contained in a single trajectory segment, which needs to cover the minimum dynamic cycle of the system), the window sliding step size S (representing the number of new data points updated in each round of control; the smaller the step size, the more frequent the data updates and the stronger the timeliness), and the fixed number of columns M (representing the maximum number of columns in the Hankel matrix to avoid unlimited growth of the number of columns).

[0031] The formula for constructing the Hankel matrix is ​​as follows: ; In time-varying scenarios, the Hankel matrix needs to be dynamically updated. In each control cycle Ts, the computer executes the following update strategy: New data is added by incorporating the S most recently acquired normalized data points into the input matrix U and output matrix Y; old data is removed by deleting the S oldest data points in U and Y, maintaining the sequence length at N; and matrix reconstruction is performed by re-extracting trajectory segments based on the updated U and Y to generate new input and output Hankel matrices, keeping their dimensions unchanged. This sliding window update mechanism ensures that the Hankel matrix always contains the latest data, reflecting the current dynamics of the time-varying system in real time, thus addressing the problem of insufficient timeliness of static Hankel matrices.

[0032] Then, Hankel matrix dimensionality reduction optimization combining informative evaluation and numerical SVD is performed. The computer device quantifies the informativeness of real-time data using the smallest non-zero singular value of the Hankel matrix. Since singular values ​​reflect the linear independence of the matrix column vectors, the larger the smallest non-zero singular value, the more critical the system dynamic information contained in the data. The specific implementation principle is as follows: For the initial Hankel matrix updated by the sliding window, the input and output Hankel matrices are constructed, with the following expression: ; And obtain the matrix The rank of the matrix is ​​used to reflect the number of linearly independent columns. Singular value decomposition is then performed, and the singular values ​​are sorted in descending order to obtain the smallest non-zero singular value. Next, an information threshold is set. If the smallest non-zero singular value of the current Hankel matrix is ​​greater than this threshold, i.e. If the result is positive, it means that the current Hankel matrix contains enough key dynamic information to be directly used for order reduction processing; otherwise, discard the S data points added in the latest sliding window (considered as redundant information) and update the Hankel matrix again.

[0033] However, traditional SVD methods require processing of the matrix. Recalculate, its complexity This approach fails to meet real-time requirements. Therefore, this invention considers employing a rank modification algorithm based on historical SVD, which uses the SVD result from the previous time step to quickly update the current reduced-order matrix, thereby reducing computational complexity to [missing information]. The operation steps are as follows: First, initialize the SVD parameters, using the initial matrix. Performing a full SVD decomposition yields the effective subspace of the left singular vectors. A diagonal matrix composed of non-zero singular values and the effective subspace of right singular vectors : ; Then, a fast SVD method based on sliding window updates is adopted. When new data is added to the sliding window, it can be... Represented as a history matrix with rank increments: ; Where A represents the flattened vector of the newly added input / output (I / O) trajectory segment, and B represents the incremental column index vector.

[0034] Based on historical SVD results Perform the following quick update steps: Calculate the orthogonal components of A in the history left singular subspace: ; Calculate the orthogonal components of B in the history right singular subspace: ; Construct an incremental SVD matrix and diagonalize it: ; Update the current SVD results: .

[0035] Subsequently, a reduced-order joint Hankel matrix is ​​generated based on the SVD results. The former Column (of which) Construct a reduced-order matrix for the adaptive order: ; And The final reduced-order matrix is ​​obtained by splitting by rows: .

[0036] The above method avoids redundant calculations of the complete SVD, thereby improving computational efficiency and meeting the real-time control requirements of time-varying systems. Regarding the adaptive order... The determination can be based on Implement filtering. Specifically, from Begin, traverse ,examine The i-th singular value If the singular value is greater than a certain preset threshold, that is: ,but If it is less than this threshold, that is: If the condition is met, then stop iterating and take the previous i that satisfies the condition as the first i. The implementation of adaptive order can increase when the time-varying system is dynamically volatile. To retain more information during dynamic balancing Keep it to a minimum to reduce computational load.

[0037] The data construction module combines data volume, prediction step size, and continuous excitation judgment to generate input-output Hankel matrices. These matrices structurally store the dynamic trajectory segments of the input and output, becoming the core carrier of data-driven control and realizing the transformation from raw data to control model input. The offline data acquisition module integrates unmanned vehicle parameters and sensor data, providing raw input for subsequent analysis; the data processing module performs operations such as missing value imputation, data normalization, and dynamic feature extraction on the acquired data to ensure data quality and information validity, laying the data foundation for subsequent matrix construction and optimization.

[0038] The time-varying scenario optimization module is crucial for the device's adaptation to time-varying characteristics. This module generates a joint reduced-order matrix through informative evaluation, numerical SVD order reduction, sliding window updates, and adaptive order adjustment. While ensuring data timeliness, this module significantly reduces the matrix dimension, resolving the contradiction between computational complexity and dynamic tracking in time-varying scenarios, enabling the control model to adapt to the dynamic changes of the autonomous vehicle in real time. The optimal algorithm design module constructs an optimization objective function based on the reduced-order matrix and solves for the decision variable g through rolling optimization, thereby driving the optimal control decision. This process iterates continuously, implementing rolling optimization logic and ensuring the control accuracy of the autonomous vehicle under time-varying conditions. The data verification and analysis module quantitatively evaluates the effectiveness of the entire process, comprehensively verifying multiple dimensions such as data quality, matrix order reduction effect, and control error, providing a basis for continuous device optimization. Ultimately, precise data-driven control of the autonomous vehicle is achieved, fully covering the entire link from data acquisition to control execution, demonstrating efficient and robust control performance in time-varying scenarios.

[0039] The "Data-Driven Control Method Based on Sliding Window and SVD Informative Evaluation in Time-Varying Scenarios" of this invention achieves multi-dimensional improvements in timeliness and computational efficiency in time-varying system control through innovative designs such as dynamically updating the Hankel matrix using a sliding window, informative quantization based on singular values ​​and efficient SVD order reduction, and adaptive order adjustment. Specific technical effects are as follows: First, the timeliness of time-varying dynamic tracking is significantly improved. By updating the Hankel matrix in real time through a sliding window, this invention can continuously incorporate the latest input and output data of the time-varying system into the control model, ensuring that the Hankel matrix always reflects the current dynamic characteristics of the system. The sliding window can integrate newly acquired input and output data into the Hankel matrix within one control cycle. Compared with the traditional static Hankel matrix method that relies on offline fixed data, the tracking lag time of time-varying dynamic tracking is significantly shortened, enabling the controller to respond promptly to changes in system parameters and avoiding the accumulation of control deviations caused by model lag.

[0040] Secondly, the computational efficiency is optimized by orders of magnitude, meeting the requirements of real-time control. This invention uses a history-based SVD update algorithm to replace the traditional complete SVD or subjective feature point scoring method, significantly reducing computational complexity while ensuring reduced-order accuracy. For dimensions of... For the joint Hankel matrix, the computational efficiency of the SVD update algorithm based on historical data is significantly improved compared with the traditional SVD, meeting the real-time requirements of 10-100ms in industrial control.

[0041] Third, the control accuracy is on par with traditional methods, while exhibiting superior robustness. This invention, through informative quantization combined with an adaptive order design, ensures that the reduced-order Hankel matrix retains key dynamic information while avoiding redundant computation. In time-varying trajectory tracking scenarios for autonomous vehicles, compared to traditional data-driven control methods, this invention reduces output tracking error and improves control robustness. This effect stems from the adaptive order's ability to automatically increase to retain more information when the system is dynamically volatile, and maintain a minimum order to reduce redundant computation when the system is dynamically stable, achieving a dynamic balance between accuracy and computational efficiency.

[0042] Fourth, it has strong universality for time-varying systems and does not rely on prior knowledge of the system. Unlike adaptive control methods that rely on system models, this invention is entirely based on Hankel matrix operations on input and output data, without requiring prior knowledge of the mathematical model of the time-varying system. In time-varying system scenarios with unknown parameters, through sliding window updates and SVD informative evaluation, this invention can automatically filter out key data segments reflecting the real-time dynamics of the system, thereby improving control accuracy.

[0043] Fifth, the continuous excitation (PE) condition is maintained online, and the control stability is theoretically guaranteed. This invention ensures that the reduced-order Hankel matrix always satisfies the extended condition of the continuous excitation through informative quantization of minimum non-zero singular values ​​and adaptive order constraints. In time-varying scenarios of autonomous vehicle control, it guarantees the boundedness of the input and output of the closed-loop system, providing both theoretical and experimental guarantees for control stability.

[0044] In summary, this invention achieves a comprehensive improvement in timeliness, computational efficiency, control accuracy, and robustness in data-driven control of time-varying systems through a collaborative design of sliding window dynamic updates, SVD informative order reduction, and adaptive order adjustment. This method can be widely applied to time-varying scenarios such as autonomous driving and robot control, possessing significant industrial application value and academic innovation significance.

[0045] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A data-driven control method based on sliding window and feature point selection under time-varying scenarios, characterized in that, The application relates to a control method for a time-varying system. Real-time acquisition of input and output signals of the time-varying system is performed through a sensor, and the signals are normalized; A sliding window mechanism is used to dynamically update an input and output trajectory data structure, so that the latest data is included and the old data is removed, and the data structure always reflects the current dynamic characteristics of the system; Characteristic points are selected from the trajectory data structure through singular value decomposition, the data information is quantified, and dimension reduction processing is performed to generate a dimension-reduced data structure; A control optimization objective function is constructed based on the dimension-reduced data structure, and a decision variable is solved to generate a control instruction; An adaptive order adjustment mechanism is used to balance the dimension of the data structure and the information completeness, and the dynamic optimization of control accuracy and calculation efficiency is realized.

2. The data drive control method of claim 1, wherein, The sliding window mechanism includes setting a window length, a sliding step and a fixed column number upper limit; the time step coverage range of a single trajectory segment is limited by the window length, the data update frequency is controlled by the sliding step, and the column dimension growth of the trajectory data structure is restricted by the fixed column number upper limit, so that the calculation burden caused by the accumulation of data volume does not exceed the real-time control requirement.

3. The data drive control method of claim 1, wherein, The characteristic point selection is realized based on the information quantification of singular value decomposition; specifically, the minimum non-zero singular value of the trajectory data structure is calculated to evaluate the data information; when the minimum non-zero singular value is greater than a preset threshold, it is determined that the data has key dynamic information and dimension reduction is performed; when the minimum non-zero singular value is less than the preset threshold, it is determined that the data is redundant and the sliding window is triggered to update the data.

4. The data drive control method of claim 1, wherein, The singular value decomposition dimension reduction processing adopts an updating algorithm based on historical decomposition results; the algorithm uses the singular value decomposition results at the previous moment to calculate the dimension-reduced new solution of the current data structure in an incremental updating mode, avoids repeated calculation of complete singular value decomposition, significantly reduces the calculation complexity to meet the real-time control time limit.

5. The data drive control method of claim 4, wherein, The incremental updating algorithm includes: using the singular value decomposition results at the previous moment, decomposing the new data into orthogonal components of the historical subspace, constructing a low-dimensional incremental matrix and diagonalizing, and finally fusing to generate the reduced-order matrix at the current moment.

6. The data drive control method of claim 5, wherein, The incremental updating algorithm reduces the calculation complexity to a linear level, which is realized by the following steps: calculating the orthogonal components of the new data based on the historical left / right singular vector subspace; constructing a low-dimensional incremental matrix and performing fast diagonalization; fusing the historical singular value decomposition results and the incremental decomposition results.

7. The data drive control method of claim 1, wherein, The adaptive order adjustment mechanism is realized by iteratively traversing the singular value sequence; the order is increased step by step from the minimum order, and the traversal is stopped when the current singular value is less than a preset threshold, and the previous order is locked as the final dimension reduction order.

8. The data drive control method of claim 7, wherein, The adaptive order dynamically adapts to the time-varying system: the order is increased to enhance the retention of system dynamic information when the system dynamics is intense; the minimum order is maintained to reduce calculation redundancy when the system dynamics is stable.

9. The data drive control method of claim 1, wherein, The control optimization objective function is constructed based on the dimension-reduced data structure, and the decision variable is solved through rolling optimization; the decision variable acts on the generated control instruction, and is fed back to the data acquisition step to form a continuous updating control loop to adapt to the dynamic changes of the time-varying system.

10. The data drive control method of claim 1, wherein, The method is suitable for an unmanned vehicle automatic driving scene; wherein an input signal corresponds to a vehicle steering angle control quantity, and an output signal corresponds to a vehicle position and a heading angle state quantity, and trajectory tracking robustness is improved through dynamic data-driven control instructions.

Images