Complex system control method, device and equipment based on diffusion model
By using a denoising network based on a diffusion model and an inverse dynamics model, the problem of poor control performance caused by uncertainties in the dynamics model in complex systems is solved, and efficient finite-time control is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TSINGHUA UNIVERSITY
- Filing Date
- 2026-01-06
- Publication Date
- 2026-05-05
AI Technical Summary
In existing technologies, the finite-time control performance of complex systems is poor, mainly because the uncertainty of the system dynamics model makes it difficult to obtain accurate system equations and parameters, especially in nonlinear dynamic systems, resulting in poor control performance.
A denoising network based on a diffusion model and an inverse dynamics model are used. The denoising network denoises the initial system noise information, generates multiple target prediction system states, and uses the inverse dynamics model to estimate the states of adjacent systems, generating a control signal sequence to achieve finite-time control.
It improves the control efficiency and effectiveness of complex systems, has good generalization ability and robustness, and can generate a complete control signal sequence that conforms to the basic laws of the system within a finite time.
Smart Images
Figure CN121978918A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of control system technology, and in particular to a method, apparatus and device for controlling complex systems based on a diffusion model. Background Technology
[0002] Complex systems consist of multiple interacting components and exhibit nonlinear dynamics and emergent behavior. The purpose of finite-time control of complex systems is to guide them to reach a target system state within a finite time.
[0003] In existing technologies, precise system dynamics models are primarily relied upon to achieve finite-time control of complex systems. However, in practical applications, given that accurate system equations and parameters are often difficult to obtain from system dynamics models, especially for complex systems with nonlinear dynamics, the uncertainty of the system dynamics model severely affects the control performance, resulting in poor control performance. Summary of the Invention
[0004] This application provides a control method, apparatus, and device for complex systems based on a diffusion model, which addresses the shortcomings of existing technologies where poor control performance is caused by uncertainties in the system dynamics model, thereby improving the control performance of complex systems.
[0005] This application provides a control method for complex systems based on a diffusion model, including: Acquire the current system state, target system state, and initial system noise information sampled within a preset time period for a complex system; The current system state, the target system state, and the initial system noise information are all input into the denoising network in the diffusion model. The initial system noise information is denoised by the denoising network to obtain multiple target predicted system states of the complex system within the preset time period. The prediction system state sequence, which consists of the multiple target prediction system states, is input into the inverse dynamics model. The inverse dynamics model is used to estimate the states of two adjacent target prediction systems in the prediction system state sequence, and generate the corresponding control signal sequence. Finite-time control of the complex system is performed based on the control signal sequence.
[0006] According to the complex system control method based on a diffusion model provided in this application, the current system state, the target system state, and the initial system noise information are all input into a denoising network of the diffusion model. The initial system noise information is denoised through the denoising network to obtain multiple target predicted system states of the complex system within a preset time period, including: The current system state, the target system state, and the initial system noise information are all input into the denoising network. The initial system noise information is denoised through the denoising network to obtain the denoised system noise information. If the number of denoising attempts does not reach the preset number of denoising attempts, the denoised system noise information is re-determined as the new initial system noise information, and the new initial system noise information, the current system state, and the target system state are input into the diffusion model again. The diffusion model is used to denoise the new initial system noise information until the number of denoising attempts reaches the preset number of denoising attempts. When the number of denoising attempts reaches the preset number of denoising attempts, the multiple prediction system states obtained when the preset number of denoising attempts is reached are determined as the multiple target prediction system states.
[0007] According to the complex system control method based on a diffusion model provided in this application, the denoising network includes a double U-shaped network module and a residual connection module. The current system state, the target system state, and the initial system noise information are all input into the denoising network. The initial system noise information is denoised through the denoising network to obtain denoised system noise information, including: The current system state, the target system state, and the initial system noise information are all input to the dual-U network module. The initial system noise information is denoised by the dual-U network module to obtain the first system noise information and the second system noise information. The first system noise information and the second system noise information are input to the residual connection module. The residual connection module then sums the first system noise information and the second system noise information to determine the denoised system noise information.
[0008] According to the complex system control method based on a diffusion model provided in this application, the dual-U-shaped network module includes a first U-shaped network unit and a second U-shaped network unit. The current system state, the target system state, and the initial system noise information are all input into the dual-U-shaped network module. The initial system noise information is then denoised by the dual-U-shaped network module to obtain first system noise information and second system noise information, including: The current system state, the target system state, and the initial system noise information are all input to the first U-shaped network unit. The initial system noise information is denoised by the first U-shaped network unit to obtain the first system noise characteristics and the first system noise information. The current system state, the target system state, the initial system noise information, and the first system noise characteristics are all input to the second U-shaped network unit, and the second system noise information is obtained through the second U-shaped network unit.
[0009] According to the complex system control method based on a diffusion model provided in this application, the first U-shaped network unit includes a first one-dimensional U-shaped network and a first-order expanded system estimation network. The current system state, the target system state, and the initial system noise information are all input into the first U-shaped network unit. The initial system noise information is denoised by the first U-shaped network unit to obtain first system noise features and first system noise information, including: The current system state, the target system state, and the initial system noise information are all input into the first one-dimensional U-shaped network. The initial system noise information is then denoised using the first one-dimensional U-shaped network to obtain the first system noise features. The noise features of the first system are input into the first-order expanded system estimation network, and the first-order coefficients of the noise features of the first system are estimated by the first-order expanded system estimation network to obtain the noise information of the first system.
[0010] According to the complex system control method based on a diffusion model provided in this application, the second U-shaped network unit includes a second one-dimensional U-shaped network and a second-order expanded system estimation network. The step of inputting the current system state, the target system state, the initial system noise information, and the first system noise characteristics into the second U-shaped network unit, and obtaining the second system noise information through the second U-shaped network unit, includes: The current system state, the target system state, the initial system noise information, and the first system noise feature are all input into the second one-dimensional U-shaped network, and the second system noise feature is obtained through the second one-dimensional U-shaped network. The second system noise features are input into the second-order expanded system estimation network, and the second system noise features are estimated by first-order coefficients through the second-order expanded system estimation network to obtain the second system noise information.
[0011] This application also provides a control device for complex systems based on a diffusion model, comprising: The acquisition unit is used to acquire the current system state of the complex system, the target system state, and the initial system noise information sampled within a preset time period; The denoising unit is used to input the current system state, the target system state and the initial system noise information into the denoising network in the diffusion model, and to denoise the initial system noise information through the denoising network to obtain multiple target predicted system states of the complex system within the preset time period. The generation unit is used to input the prediction system state sequence composed of the multiple target prediction system states into the inverse dynamics model, and to estimate the states of two adjacent target prediction systems in the prediction system state sequence through the inverse dynamics model to generate the corresponding control signal sequence. A control unit for performing finite-time control of the complex system based on the control signal sequence.
[0012] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the complex system control method based on the diffusion model as described in any of the preceding claims.
[0013] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the complex system control method based on the diffusion model as described in any of the preceding claims.
[0014] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the complex system control method based on the diffusion model as described in any of the preceding claims.
[0015] The diffusion model-based control method, apparatus, and device for complex systems provided in this application, when performing finite-time control of a complex system, first acquire the current system state, the target system state, and initial system noise information sampled within a preset time period. Then, the current system state, the target system state, and the initial system noise information are all input into the denoising network of the diffusion model. The denoising network denoises the initial system noise information, obtaining multiple target predicted system states of the complex system within the preset time period. Next, the predicted system state sequence composed of these multiple target predicted system states is input into an inverse dynamics model. The inverse dynamics model estimates the states of two adjacent target predicted systems, generating corresponding control signal sequences. Finite-time control of the complex system is then performed based on these control signal sequences. In this way, by denoising the initial system noise information through the diffusion model's denoising network and obtaining multiple target predicted system states of the complex system within the preset time period, and by estimating the states of two adjacent target predicted systems through the inverse dynamics model, a complete control signal sequence conforming to the basic laws of the complex system can be generated in one go. This not only improves the control efficiency of the complex system but also enhances the control effect within a finite time period. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 This is a schematic diagram of the framework of a complex system control method based on a diffusion model, provided for an embodiment of this application.
[0018] Figure 2 This is a flowchart illustrating a complex system control method based on a diffusion model, provided as an embodiment of this application.
[0019] Figure 3 This is a schematic diagram illustrating a process for obtaining target system noise information of a complex system within a preset time period using a denoising network, as provided in an embodiment of this application.
[0020] Figure 4 This is a schematic diagram of the structure of a denoising network provided in an embodiment of this application.
[0021] Figure 5 This is a schematic diagram of the structure of a complex system control device based on a diffusion model, provided as an embodiment of this application.
[0022] Figure 6 This is a schematic diagram of the physical structure of an electronic device provided in an embodiment of this application. Detailed Implementation
[0023] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0024] In the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone, where A and B can be singular or plural. In the textual description of this application, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.
[0025] Complex systems have a wide-ranging impact on fields such as intelligent manufacturing, power systems, and biological systems. The control requirements for complex systems in these fields are also becoming increasingly stringent. Finite-time control of complex systems aims to ensure that the system reaches a target state within a finite timeframe.
[0026] For example, complex systems can be ecosystems, climate systems, intelligent manufacturing systems, transportation systems, financial systems, multi-agent systems, or urban systems, etc., and can be specifically configured according to actual needs.
[0027] In existing technologies, precise system dynamics models are primarily used to achieve finite-time control of complex systems. However, considering that accurate system equations and parameters are often difficult to obtain from system dynamics models, especially for complex systems with nonlinear dynamics, the uncertainty of the system dynamics model severely affects the control effect, resulting in poor control performance.
[0028] To address the shortcomings of existing technologies in terms of poor control performance due to uncertainties in system dynamics models, and thus improve the control performance of complex systems, this application embodiment pre-constructs a diffusion model and an inverse dynamics model, including a denoising network. The diffusion model models the joint distribution of control signals and the current system state of the complex system, while the inverse dynamics model addresses issues such as overfitting and insufficient generalization ability in generating predicted system states. Furthermore, it explicitly models the mapping relationship between system states and control signals, improving the applicability of the control signals and ensuring that the generated control signal sequence conforms to the fundamental laws of complex systems. This combination of the diffusion model and inverse dynamics model allows for the generation of a complete control signal sequence for the complex system in a single step, improving not only the control efficiency of complex systems but also the control performance within a finite timeframe, demonstrating good generalization ability and robustness.
[0029] Based on this, embodiments of this application provide a control method for complex systems based on a diffusion model, which can be found in [reference needed]. Figure 1 As shown, Figure 1This document provides a schematic diagram of the framework for a complex system control method based on a diffusion model, as illustrated in an embodiment of this application. First, the current system state, the target system state, and initial system noise information sampled within a preset time period are acquired. These are then input into a denoising network of the diffusion model to denoise the initial system noise, yielding multiple predicted system states for the complex system within the preset time period. Next, a sequence of predicted system states is input into an inverse dynamics model to estimate the states of adjacent predicted systems, generating corresponding control signal sequences. Finally, the complex system is controlled within a finite time period based on these control signal sequences. This method, by denoising the initial system noise through the diffusion model's denoising network and obtaining the predicted system state sequence within the preset time period, and by estimating the states of adjacent predicted systems within the predicted system state sequence using the inverse dynamics model, allows for the generation of a complete control signal sequence conforming to the fundamental laws of the complex system. This not only improves the control efficiency of the complex system but also enhances its control effect within a finite time period.
[0030] It is understood that the execution subject of the diffusion model-based complex system control method provided in this application can be a computer, server, or a specially configured diffusion model-based complex system control device or other electronic equipment, or a diffusion model-based complex system control device installed in such electronic equipment. The diffusion model-based complex system control device can be implemented by software, hardware, or a combination of both, and can be configured according to actual needs.
[0031] The diffusion-based complex system control method provided in this application will be described in detail below through several specific embodiments. It is understood that these specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.
[0032] Figure 2 A flowchart illustrating a complex system control method based on a diffusion model, provided as an embodiment of this application, is shown below. Figure 2 As shown, this control method for complex systems based on the diffusion model can include: S201. Obtain the current system state of the complex system, the target system state, and the initial system noise information sampled within a preset time period.
[0033] The target system state can be understood as the state that a complex system is expected to reach after finite-time control.
[0034] For example, when obtaining the current system state of a complex system, it can be obtained by observing the current system state through a state observer, or by measuring with sensors, etc. The specific settings can be configured according to actual needs.
[0035] For example, when obtaining the target system state of a complex system, the system state preset by the user can be determined as the target system state, or the target system state of the complex system can be estimated through methods such as distributed observers. The specific settings can be made according to actual needs.
[0036] For example, when obtaining the initial system noise information of a complex system sampled within a preset time period, the noise can be randomly sampled and statistically analyzed during the startup or idle phase of the complex system to obtain the initial system noise information of the complex system sampled within the preset time period. Alternatively, external disturbances and noise can be estimated using a disturbance observer to obtain the initial system noise information of the complex system sampled within the preset time period. The specific settings can be configured according to actual needs.
[0037] The preset time period can be set according to actual needs.
[0038] After obtaining the current system state, the target system state, and the initial system noise information, the following step S202 can be executed: S202. Input the current system state, the target system state, and the initial system noise information into the denoising network of the diffusion model. Denoise the initial system noise information through the denoising network to obtain the predicted system states of multiple targets of the complex system within a preset time period.
[0039] For example, in the embodiments of this application, in addition to obtaining multiple target prediction system states, multiple prediction control signals can also be obtained through the denoising network, which can be set according to actual needs.
[0040] For example, the current system state of a complex system can be denoted as: The target system state can be denoted as The initial system noise information can be denoted as: The preset time period can be recorded as Preset time period The states of multiple target prediction systems within can be denoted as: , , , ..., , .
[0041] Among them, multiple target prediction system states , , , ..., , The resulting predicted system state sequence can be understood as the state sequence of a complex system predicted by the diffusion model within a preset time period. The predicted state trajectory within.
[0042] It is understood that, in the embodiments of this application, by pre-constructing a diffusion model including a denoising network, the denoising network can denoise the initial system noise information with reference to the current system state and the target system state, so as to model the joint distribution of the control signal and the current system state of the complex system, thereby obtaining multiple target predicted system states of the complex system within a preset time period.
[0043] After obtaining the predicted system states of multiple targets within a preset time period for a complex system through a denoising network, the corresponding control signal sequence can be generated with the help of a pre-built inverse dynamics model, as shown in S203 below: S203. Input the prediction system state sequence consisting of multiple target prediction system states into the inverse dynamics model. Estimate the states of two adjacent target prediction systems in the prediction system state sequence through the inverse dynamics model to generate the corresponding control signal sequence.
[0044] For example, in the embodiments of this application, an autoregressive multilayer perceptron can be used as an inverse dynamic model, which can be denoted as: .
[0045] A sequence of predicted system states, consisting of multiple target predicted system states, is input into an inverse dynamics model. This inverse dynamics model can estimate the states of two adjacent target predicted system states within the predicted system state sequence. Its mathematical expression is as follows: This allows for the generation of the corresponding control signal sequence. By fully utilizing the Markov property of time-invariant systems and considering only the relationship between the predicted system states at adjacent times, the complexity of the control signal sequence generation process is greatly simplified. Indicates the time period within the preset time frame Predicting the system state at any given time and Predicting the system state at any given time Estimate the corresponding control signal to obtain it.
[0046] The prediction system state sequence, consisting of multiple target prediction system states, is input into the inverse dynamics model. It can predict the system state. and predict system state The corresponding control signal is obtained by estimation, and can be denoted as: ; For predicting the system state and predict system state The corresponding control signal is obtained by estimation, and can be denoted as: ; For predicting the system state and predict system state The corresponding control signal is obtained by estimation, and can be denoted as: This generates the corresponding control signal sequence, which can be denoted as... , , ..., .
[0047] It is understood that, in the embodiments of this application, by pre-constructing an inverse dynamics model, not only can the problems of overfitting and insufficient generalization ability of the diffusion model in generating and predicting system states be solved, but also the mapping relationship between system states and control signals can be explicitly modeled, improving the applicability of control signals, thereby enabling the generated control signal sequence to conform to the basic laws of complex systems.
[0048] S204. Finite-time control of complex systems based on control signal sequences.
[0049] As can be seen, in this embodiment, when performing finite-time control on a complex system, the current system state, the target system state, and the initial system noise information sampled within a preset time period can be obtained first. The current system state, the target system state, and the initial system noise information are all input into the denoising network of the diffusion model. The denoising network denoises the initial system noise information, obtaining multiple target predicted system states of the complex system within the preset time period. Then, the predicted system state sequence composed of these multiple target predicted system states is input into the inverse dynamics model. The inverse dynamics model estimates the states of two adjacent target predicted systems, generating corresponding control signal sequences. Finite-time control is then performed on the complex system based on these control signal sequences. In this way, by denoising the initial system noise information through the diffusion model's denoising network, multiple target predicted system states of the complex system within a preset time period are obtained. Furthermore, by estimating the states of two adjacent target predicted systems in the predicted system state sequence through the inverse dynamics model, a complete control signal sequence conforming to the basic laws of the complex system can be generated at once. This not only improves the control efficiency of the complex system but also enhances the control effect within a finite time period.
[0050] Based on the above Figure 1 In the illustrated embodiment, for example, in S202 above, the current system state, the target system state, and the initial system noise information are all input into the denoising network of the diffusion model. The specific implementation of obtaining the target system noise information of the complex system within a preset time period through the denoising network can be found below. Figure 3 The example shown.
[0051] Figure 3 This application provides a flowchart illustrating a process for obtaining target system noise information of a complex system within a preset time period using a denoising network, as illustrated in the embodiments of this application. For example, please refer to [link to relevant documentation]. Figure 3 As shown, the method may include: S301. Input the current system state, the target system state, and the initial system noise information into the denoising network. Use the denoising network to denoise the initial system noise information to obtain the denoised system noise information.
[0052] For example, in the embodiments of this application, see... Figure 4 As shown, Figure 4 This is a schematic diagram of a denoising network provided in an embodiment of this application. The denoising network may include a dual-U-shaped network module and a residual connection module. The current system state, the target system state, and the initial system noise information are all input into the denoising network. When denoising the initial system noise information through the denoising network, the current system state, the target system state, and the initial system noise information can be input into the dual-U-shaped network module first. The initial system noise information is denoised through the dual-U-shaped network module to obtain the first system noise information and the second system noise information. Then, the first system noise information and the second system noise information are input into the residual connection module. The residual connection module sums the first system noise information and the second system noise information to determine the denoised system noise information.
[0053] For example, the noise information of the first system can be denoted as: The noise information of the second system can be denoted as: By analyzing the noise information of the first system With the second system noise information Summing yields the denoised system noise information. If the initial system noise information can be denoted as... The system noise information obtained after this denoising can be denoted as: .
[0054] For example, in the embodiments of this application, it can be combined with Figure 4As shown, the dual-U-shaped network module includes a first U-shaped network unit and a second U-shaped network unit. The current system state, the target system state, and the initial system noise information are all input into the dual-U-shaped network module. The initial system noise information is denoised by the dual-U-shaped network module to obtain the first system noise information and the second system noise information. Alternatively, the current system state, the target system state, and the initial system noise information can all be input into the first U-shaped network unit. The initial system noise information is denoised by the first U-shaped network unit to obtain the first system noise characteristics and the first system noise information. The current system state, the target system state, the initial system noise information, and the first system noise characteristics are all input into the second U-shaped network unit to obtain the second system noise information.
[0055] For example, the noise characteristics of the first system can be denoted as: The noise of the second system can be denoted as .
[0056] For example, in an embodiment of this application, the first U-shaped network unit includes a first one-dimensional U-shaped network and a first-order expanded system estimation network. The current system state, the target system state, and the initial system noise information are all input into the first U-shaped network unit. The initial system noise information is denoised by the first U-shaped network unit to obtain the first system noise features and the first system noise information. This includes: inputting the current system state, the target system state, and the initial system noise information into the first one-dimensional U-shaped network, denoising the initial system noise information by the first one-dimensional U-shaped network to obtain the first system noise features; inputting the first system noise features into the first-order expanded system estimation network, and estimating the first system noise features by the first-order expanded system estimation network to obtain the first system noise information.
[0057] If the initial system noise information can be denoted as Given time step encoding and conditional tag encoding The specific implementation of obtaining the noise information of the first system can be found in the following formula: in, This represents the first one-dimensional U-shaped network. For batch size, The length of the system state sequence. For feature dimension, for Dimension From the current system state through a single linear layer To the target system state Calculated.
[0058] For example, in an embodiment of this application, the second U-shaped network unit includes a second one-dimensional U-shaped network and a second-order expanded system estimation network. The current system state, the target system state, the initial system noise information, and the first system noise feature are all input into the second U-shaped network unit. The second system noise information is obtained through the second U-shaped network unit, including: inputting the current system state, the target system state, the initial system noise information, and the first system noise feature into the second one-dimensional U-shaped network, and obtaining the second system noise feature through the second one-dimensional U-shaped network; inputting the second system noise feature into the second-order expanded system estimation network, and performing first-order coefficient estimation on the second system noise feature through the second-order expanded system estimation network to obtain the second system noise information.
[0059] If the initial system noise information can be denoted as Given time step encoding and conditional tag encoding The specific implementation of obtaining the noise information of the second system can be found in the following formula: in, This represents the second one-dimensional U-shaped network. For batch size, The length of the system state sequence. For feature dimension, for Dimension From the current system state through a single linear layer To the target system state Calculated.
[0060] After obtaining the noise information of the first system respectively Second system noise information Then, the noise information of the denoised system can be determined through the residual connection module, as shown in the following formula: It is understood that, in the embodiments of this application, the advantages of using a denoising network to denoise the initial system noise information are as follows: by combining a first-order expansion coefficient estimation network and a second-order expansion coefficient estimation network, the dynamic characteristics of the nonlinear system are captured, which not only preserves the theoretical basis of optimal control of linear systems, but also enhances the modeling ability of nonlinear dynamics; at the same time, the design of the residual connection module ensures the training stability of the deep network and the effective transfer of features.
[0061] Considering that the main idea of the diffusion model is to denoise the initial system noise information multiple times, after denoising the initial system noise information through the denoising network and obtaining the denoised system noise information, it can be determined whether the number of denoising times has reached the preset number of denoising times. If the number of denoising times has not reached the preset number of denoising times, the following S302 is executed; if the number of denoising times has reached the preset number of denoising times, the following S303 is executed.
[0062] S302. If the number of denoising attempts has not reached the preset number of denoising attempts, the denoised system noise information is redefined as the new initial system noise information. The new initial system noise information, the current system state, and the target system state are then input into the diffusion model again. The diffusion model is used to denoise the new initial system noise information until the number of denoising attempts reaches the preset number of denoising attempts.
[0063] The preset number of noise reduction attempts can be set according to actual needs.
[0064] S303. When the number of denoising attempts reaches the preset number of denoising attempts, the multiple prediction system states obtained when the preset number of denoising attempts is reached are determined as multiple target prediction system states.
[0065] The predicted system state sequence, consisting of multiple target predicted system states, can be understood as the complex system predicted by the diffusion model within a preset time period. The predicted state trajectory within.
[0066] As can be seen from the embodiments of this application, the denoising network with a dual U-shaped network architecture is used to gradually optimize the noisy input, and finally generate multiple target prediction system states, that is, the complex system predicted by the diffusion model within a preset time period. The predicted state trajectory within the range. Furthermore, during the generation process, conditional label encoding is used. The method guides the optimization direction and uses a classifier-free guidance technique to make the generated predicted state trajectory deviate from the distribution of random training data to obtain better control effect. This allows the subsequent estimation of the states of two adjacent target predicted systems in the predicted system state sequence through the inverse dynamics model. It can generate a complete control signal sequence that conforms to the basic laws of complex systems at once, which not only improves the control efficiency of complex systems, but also improves the control effect of complex systems within a limited time.
[0067] For example, the training of the aforementioned denoising network and inverse dynamics model can be carried out in an end-to-end manner. Simultaneous optimization is performed by minimizing the combined objective function of denoising loss and inverse dynamics loss. An inverse dynamics prediction error term is specifically added to the loss function to ensure that the inverse dynamics model can accurately predict the required control signal. This ensures both the quality of the generated control signal sequence and the consistency between the generated result and the actual system dynamics. It explicitly models the local relationship between the system state and the control signal, avoiding the need for a complex diffusion model to fit this relatively simple mapping relationship. Furthermore, by separating the tasks of system state generation and control signal calculation, each module can better perform its specific function.
[0068] For diffusion models, in order to achieve finite-time control of complex systems, the finite-time optimal control problem can be refactored into a condition generation task. Given multiple sets of control data samples... ,in, For control signal samples, The labels are assigned to the corresponding system state samples, and the goal is to construct a conditional distribution. To generate optimal control signal sequence samples. Here, θ is a learnable parameter, and the optimization objective can be expressed as: in, Indicates the optimization objective. This represents the initial system state sample. This represents a sample of the target system's state. System dynamic constraints. Obtained through implicit learning of data.
[0069] In this embodiment, a conditional diffusion process is used to construct an initial diffusion model for training the diffusion model. During the forward pass of the initial diffusion model, the control data samples... By gradually adding noise over K time steps, the system noise information with gradually added noise can be obtained. The noise addition process at each step can follow the following Gaussian distribution: in, This represents the forward distribution of the noisy data at step k, conditional upon the noisy data at step (k-1). The variance scheduling represents the control noise level, and N() represents a Gaussian distribution. Represents the identity matrix.
[0070] In the reverse process of the initial diffusion model, the model learns how to progressively denoise from pure system noise information. Begin reconstructing the original control data sample Each step of the denoising process can follow the following Gaussian distribution: in, This represents a set of learnable parameters conditioned on the (k-1)th step denoised data and the kth step denoised data. The inverse distribution, This represents the mean of the conditional distribution. The variance of the conditional distribution is usually fixed as a constant during diffusion.
[0071] To generate optimal control signal sequences with lower energy consumption, this application employs a classifier-free guided technique. Specifically, during training, the conditional label r (corresponding to the optimization objective J) is set to null with probability ω following a Bernoulli distribution with parameter β, allowing the diffusion model to learn both conditional and unconditional generation simultaneously. Therefore, the training objective of the diffusion model is a weighted combination of conditional and unconditional losses, and its diffusion model loss... Please refer to the following formula: in, Indicates conditional loss. Indicates unconditional loss. Indicates noise label, This indicates conditional noise prediction. This indicates unconditional noise prediction.
[0072] During inference, in order to guide the inverse dynamics model to generate a low-energy-consumption control signal sequence, a guiding coefficient w is introduced and sampled as follows: in, This represents unclassifier-guided weighted noise, which is the predicted noise actually used in the denoising process. The intensity of conditionally generated noise is controlled by the guiding coefficient. The label is proportional to the optimization objective J. By adjusting the guiding coefficient w, control energy consumption and other performance indicators can be balanced while maintaining control effectiveness. The label is input when generating the control signal sequence. =0 and employ weighted sampling, sacrificing diversity for control effectiveness. This classifier-free approach avoids explicit classifier training, thus providing a more efficient condition generation mechanism.
[0073] For the inverse dynamics model, when training the inverse dynamics model, an initial control signal sequence can be generated first through the initial inverse dynamics model. Input it into the actual system or simulation environment and record the corresponding system status labels. The generated control signal-system status label is combined into a new sample. It is added to the retraining data pool for training the initial inverse dynamics model.
[0074] The core objectives of training the initial inverse dynamics model are: minimizing the error between the target system state and the actual reached system state, minimizing control energy consumption, and ensuring that the generated control signal conforms to physical constraints. To this end, a comprehensive loss function is designed, comprising two main parts: diffusion model loss and inverse dynamics loss. The diffusion model loss is described above. Inverse dynamic loss It can be represented as: in, Represents the parameters of the inverse dynamics model. express Predicting the system state at any given time. express Predicting the system state at any given time. express Control signal label at any time This represents the control signal generated through the initial inverse dynamics model.
[0075] By adopting the above training method, the inverse dynamics model can continuously learn control strategies that are closer to the optimal solution. Furthermore, the retraining process keeps the original training objectives and loss functions unchanged, and only gradually improves the model performance by expanding the dataset. This ensures the stability of the training and effectively enhances the model's ability to explore optimal control.
[0076] Understandably, in the training process of the diffusion model and the inverse dynamics model described above, the network parameters are first randomly initialized and pre-trained using an initial dataset. Then, an iterative optimization phase begins, alternately optimizing the diffusion network and the inverse dynamics model, while periodically executing a retraining strategy to expand the dataset. The entire training process judges convergence by monitoring the stability of the target loss and control effect on the validation set, stopping training when a preset convergence condition is met. This end-to-end training scheme ensures that the model can learn effective control policies while maintaining good generalization ability.
[0077] To verify the effectiveness of the diffusion-based complex system control method provided in this application, a Kuramoto oscillator network system will be used as an example. The Kuramoto oscillator network system is a typical nonlinear complex system, widely used in the study of synchronization phenomena such as biological rhythms and neural networks. A ring-shaped Kuramoto oscillator network system consisting of 8 nodes will be constructed, and its system dynamic equations are as follows: in, Let ω represent the phase state of the i-th oscillator at time t, and ω be the natural frequency (set to 0 in this experiment). To control the input.
[0078] Regarding experimental data, 20,000 training samples and 1,000 test samples were obtained, with the initial phase derived from a Gaussian distribution. Sampling, random intervention control signal from The sampling is performed in the middle, and the system simulation is based on a preset time period T=16 time steps.
[0079] To evaluate the performance of the diffusion model-based complex system control method provided in this application, it can be compared with four existing typical baseline methods. Two main evaluation metrics are set for the comparison: target loss and control energy. Target loss measures the distance between the final system state and the target system state, while control energy reflects the cumulative energy consumption of the control signal.
[0080] Table 1 Model Target loss Controlling energy PID controller 7.91e-1 7.399 SAC algorithm 7.99e-1 0.733 DiffPhyCon model 1.79e-2 1.108 Baggio algorithm 1.56e-5 1.016 The technical solution provided in this application 1.14e-5 0.737 As can be seen from Table 1 above, compared with the four existing typical baseline methods, the complex system control method based on the diffusion model provided in this application has achieved significant improvements in both target loss and control energy, demonstrating superiority in the finite-time optimal control problem of complex nonlinear systems.
[0081] The following describes the control device for complex systems based on the diffusion model provided in this application. The control device for complex systems based on the diffusion model described below can be referred to in correspondence with the control method for complex systems based on the diffusion model described above.
[0082] Figure 5 A schematic diagram of a complex system control device based on a diffusion model is provided as an embodiment of this application. For example, please refer to [link to relevant documentation]. Figure 5 As shown, the complex system control device 50 based on the diffusion model may include: The acquisition unit 501 is used to acquire the current system state, the target system state, and the initial system noise information sampled within a preset time period of the complex system. The denoising unit 502 is used to input the current system state, the target system state and the initial system noise information into the denoising network in the diffusion model, and to denoise the initial system noise information through the denoising network to obtain multiple target predicted system states of the complex system within the preset time period. The generation unit 503 is used to input the prediction system state sequence composed of the multiple target prediction system states into the inverse dynamics model, and to estimate the states of two adjacent target prediction systems in the prediction system state sequence through the inverse dynamics model to generate the corresponding control signal sequence. Control unit 504 is used to perform finite-time control on the complex system based on the control signal sequence.
[0083] For example, in this embodiment of the application, the denoising unit 502 is used to input the current system state, the target system state, and the initial system noise information into the denoising network of the diffusion model, and to denoise the initial system noise information through the denoising network to obtain multiple target predicted system states of the complex system within the preset time period, including: The current system state, the target system state, and the initial system noise information are all input into the denoising network. The initial system noise information is denoised through the denoising network to obtain the denoised system noise information. If the number of denoising attempts does not reach the preset number of denoising attempts, the denoised system noise information is re-determined as the new initial system noise information, and the new initial system noise information, the current system state, and the target system state are input into the diffusion model again. The diffusion model is used to denoise the new initial system noise information until the number of denoising attempts reaches the preset number of denoising attempts. When the number of denoising attempts reaches the preset number of denoising attempts, the multiple prediction system states obtained when the preset number of denoising attempts is reached are determined as the multiple target prediction system states.
[0084] For example, in this embodiment of the application, the denoising network includes a dual U-shaped network module and a residual connection module. The denoising unit 502 is used to input the current system state, the target system state, and the initial system noise information into the denoising network, and to denoise the initial system noise information through the denoising network to obtain denoised system noise information, including: The current system state, the target system state, and the initial system noise information are all input to the dual-U network module. The initial system noise information is denoised by the dual-U network module to obtain the first system noise information and the second system noise information. The first system noise information and the second system noise information are input to the residual connection module. The residual connection module then sums the first system noise information and the second system noise information to determine the denoised system noise information.
[0085] For example, in this embodiment of the application, the dual-U-shaped network module includes a first U-shaped network unit and a second U-shaped network unit. The denoising unit 502 is used to input the current system state, the target system state, and the initial system noise information into the dual-U-shaped network module, and to denoise the initial system noise information through the dual-U-shaped network module to obtain first system noise information and second system noise information, including: The current system state, the target system state, and the initial system noise information are all input to the first U-shaped network unit. The initial system noise information is denoised by the first U-shaped network unit to obtain the first system noise characteristics and the first system noise information. The current system state, the target system state, the initial system noise information, and the first system noise characteristics are all input to the second U-shaped network unit, and the second system noise information is obtained through the second U-shaped network unit.
[0086] For example, in this embodiment of the application, the first U-shaped network unit includes a first one-dimensional U-shaped network and a first-order expanded system estimation network. The denoising unit 502 is used to input the current system state, the target system state, and the initial system noise information into the first U-shaped network unit, and to denoise the initial system noise information through the first U-shaped network unit to obtain the first system noise features and the first system noise information, including: The current system state, the target system state, and the initial system noise information are all input into the first one-dimensional U-shaped network. The initial system noise information is then denoised using the first one-dimensional U-shaped network to obtain the first system noise features. The noise features of the first system are input into the first-order expanded system estimation network, and the first-order coefficients of the noise features of the first system are estimated by the first-order expanded system estimation network to obtain the noise information of the first system.
[0087] For example, in this embodiment of the application, the second U-shaped network unit includes a second one-dimensional U-shaped network and a second-order expanded system estimation network. The denoising unit 502 is used to input the current system state, the target system state, the initial system noise information, and the first system noise features into the second U-shaped network unit, and obtain the second system noise information through the second U-shaped network unit, including: The current system state, the target system state, the initial system noise information, and the first system noise feature are all input into the second one-dimensional U-shaped network, and the second system noise feature is obtained through the second one-dimensional U-shaped network. The second system noise features are input into the second-order expanded system estimation network, and the second system noise features are estimated by first-order coefficients through the second-order expanded system estimation network to obtain the second system noise information.
[0088] The diffusion-based complex system control device 50 provided in this application embodiment can execute the technical solution of the diffusion-based complex system control method in any of the above embodiments. Its implementation principle and beneficial effects are similar to those of the diffusion-based complex system control method. Please refer to the implementation principle and beneficial effects of the diffusion-based complex system control method, which will not be repeated here.
[0089] Figure 6 This is a schematic diagram of the physical structure of an electronic device provided in an embodiment of this application, such as... Figure 6 As shown, the electronic device may include: a processor 610, a communication interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communication interface 620, and the memory 630 communicate with each other through the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute a complex system control method based on a diffusion model. The method includes: acquiring the current system state, the target system state, and initial system noise information sampled within a preset time period of the complex system; inputting the current system state, the target system state, and the initial system noise information into a denoising network in the diffusion model, and denoising the initial system noise information through the denoising network to obtain multiple target predicted system states of the complex system within the preset time period; inputting the predicted system state sequence composed of the multiple target predicted system states into an inverse dynamics model, and estimating two adjacent target predicted system states in the predicted system state sequence through the inverse dynamics model to generate corresponding control signal sequences; and performing finite-time control of the complex system based on the control signal sequences.
[0090] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0091] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the complex system control method based on the diffusion model provided by the above methods. The method includes: acquiring the current system state, the target system state, and initial system noise information sampled within a preset time period of the complex system; inputting the current system state, the target system state, and the initial system noise information into a denoising network in the diffusion model, and denoising the initial system noise information through the denoising network to obtain multiple target predicted system states of the complex system within the preset time period; inputting the predicted system state sequence composed of the multiple target predicted system states into an inverse dynamics model, and estimating two adjacent target predicted system states in the predicted system state sequence through the inverse dynamics model to generate a corresponding control signal sequence; and performing finite-time control of the complex system based on the control signal sequence.
[0092] In another aspect, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a complex system control method based on a diffusion model provided by the above methods. The method includes: acquiring the current system state, the target system state, and initial system noise information sampled within a preset time period of the complex system; inputting the current system state, the target system state, and the initial system noise information into a denoising network in the diffusion model, and denoising the initial system noise information through the denoising network to obtain multiple target predicted system states of the complex system within the preset time period; inputting a predicted system state sequence composed of the multiple target predicted system states into an inverse dynamics model, and estimating two adjacent target predicted system states in the predicted system state sequence through the inverse dynamics model to generate a corresponding control signal sequence; and performing finite-time control of the complex system based on the control signal sequence.
[0093] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0094] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A control method for complex systems based on a diffusion model, characterized in that, include: Acquire the current system state, target system state, and initial system noise information sampled within a preset time period for a complex system; The current system state, the target system state, and the initial system noise information are all input into the denoising network in the diffusion model. The initial system noise information is denoised by the denoising network to obtain multiple target predicted system states of the complex system within the preset time period. The prediction system state sequence, which consists of the multiple target prediction system states, is input into the inverse dynamics model. The inverse dynamics model is used to estimate the states of two adjacent target prediction systems in the prediction system state sequence, and generate the corresponding control signal sequence. Finite-time control of the complex system is performed based on the control signal sequence.
2. The method according to claim 1, characterized in that, The process involves inputting the current system state, the target system state, and the initial system noise information into a denoising network of a diffusion model. The denoising network then denoises the initial system noise information to obtain multiple target predicted system states of the complex system within the preset time period, including: The current system state, the target system state, and the initial system noise information are all input into the denoising network. The initial system noise information is denoised through the denoising network to obtain the denoised system noise information. If the number of denoising attempts does not reach the preset number of denoising attempts, the denoised system noise information is re-determined as the new initial system noise information, and the new initial system noise information, the current system state, and the target system state are input into the diffusion model again. The diffusion model is used to denoise the new initial system noise information until the number of denoising attempts reaches the preset number of denoising attempts. When the number of denoising attempts reaches the preset number of denoising attempts, the multiple prediction system states obtained when the preset number of denoising attempts is reached are determined as the multiple target prediction system states.
3. The method according to claim 2, characterized in that, The denoising network includes a dual U-shaped network module and a residual connection module. The current system state, the target system state, and the initial system noise information are all input into the denoising network. The denoising network denoises the initial system noise information to obtain denoised system noise information, including: The current system state, the target system state, and the initial system noise information are all input to the dual-U network module. The initial system noise information is denoised by the dual-U network module to obtain the first system noise information and the second system noise information. The first system noise information and the second system noise information are input to the residual connection module. The residual connection module then sums the first system noise information and the second system noise information to determine the denoised system noise information.
4. The method according to claim 3, characterized in that, The dual-U-shaped network module includes a first U-shaped network unit and a second U-shaped network unit. The current system state, the target system state, and the initial system noise information are all input to the dual-U-shaped network module. The initial system noise information is then denoised by the dual-U-shaped network module to obtain first system noise information and second system noise information, including: The current system state, the target system state, and the initial system noise information are all input to the first U-shaped network unit. The initial system noise information is denoised by the first U-shaped network unit to obtain the first system noise characteristics and the first system noise information. The current system state, the target system state, the initial system noise information, and the first system noise characteristics are all input to the second U-shaped network unit, and the second system noise information is obtained through the second U-shaped network unit.
5. The method according to claim 4, characterized in that, The first U-shaped network unit includes a first one-dimensional U-shaped network and a first-order expanded system estimation network. The current system state, the target system state, and the initial system noise information are all input into the first U-shaped network unit. The first U-shaped network unit denoises the initial system noise information to obtain first system noise features and first system noise information, including: The current system state, the target system state, and the initial system noise information are all input into the first one-dimensional U-shaped network. The initial system noise information is then denoised using the first one-dimensional U-shaped network to obtain the first system noise features. The noise features of the first system are input into the first-order expanded system estimation network, and the first-order coefficients of the noise features of the first system are estimated by the first-order expanded system estimation network to obtain the noise information of the first system.
6. The method according to claim 4, characterized in that, The second U-shaped network unit includes a second one-dimensional U-shaped network and a second-order expanded system estimation network. The step of inputting the current system state, the target system state, the initial system noise information, and the first system noise features into the second U-shaped network unit, and obtaining the second system noise information through the second U-shaped network unit, includes: The current system state, the target system state, the initial system noise information, and the first system noise feature are all input into the second one-dimensional U-shaped network, and the second system noise feature is obtained through the second one-dimensional U-shaped network. The second system noise features are input into the second-order expanded system estimation network, and the second system noise features are estimated by first-order coefficients through the second-order expanded system estimation network to obtain the second system noise information.
7. A control device for complex systems based on a diffusion model, characterized in that, include: The acquisition unit is used to acquire the current system state of the complex system, the target system state, and the initial system noise information sampled within a preset time period; The denoising unit is used to input the current system state, the target system state and the initial system noise information into the denoising network in the diffusion model, and to denoise the initial system noise information through the denoising network to obtain multiple target predicted system states of the complex system within the preset time period. The generation unit is used to input the prediction system state sequence composed of the multiple target prediction system states into the inverse dynamics model, and to estimate the states of two adjacent target prediction systems in the prediction system state sequence through the inverse dynamics model to generate the corresponding control signal sequence. A control unit for performing finite-time control of the complex system based on the control signal sequence.
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the complex system control method based on the diffusion model as described in any one of claims 1 to 6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the complex system control method based on the diffusion model as described in any one of claims 1 to 6.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the complex system control method based on the diffusion model as described in any one of claims 1 to 6.