Iterative learning robust safety prediction control method based on spoofing attack data optimization

By constructing a Rosser model for deception attacks and designing an iterative learning robust safety predictive control law, the problems of deception attack defense and low-latency stability in intermittent process control are solved. The system achieves stability and robustness under low latency and periodic disturbances, thereby improving control performance and energy efficiency.

CN118939909APending Publication Date: 2024-11-12UNIV OF SCI & TECH LIAONING
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410987825.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-23
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

Existing technologies have failed to effectively defend against network link spoofing attacks from controller to actuator in intermittent process control, and traditional stability conditions are computationally intensive and conservative, making it difficult to cope with the effects of low latency and periodic interference.

Method used

Within a two-dimensional system framework, a Rosser model for deception attacks is constructed, and an iterative learning robust security predictive control law is designed. The controller is optimized using historical batch deception attack data. By combining the Lyapunov-Razumikhin function and robust security invariant set, the control law gain K is solved online to ensure system stability.

Benefits of technology

It effectively resists deception attacks, maintains stable and efficient system operation, improves the system's robustness and control performance under low latency and periodic interference, and enhances energy utilization efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118939909A_ABST
    Figure CN118939909A_ABST
Patent Text Reader

Abstract

An iterative learning robust safety prediction control method based on spoofing attack data optimization belongs to the field of advanced control of industrial processes, and comprises the following steps: step 1, establishing a state space model with low network delay and a spoofing attack intermittent process under a two-dimensional system framework; 2, constructing a two-dimensional Rosser model with batch direction state deviation and time direction output error; step 3, designing a controller based on the established two-dimensional Rosser model and the spoofing attack data of historical batches; step 4, on the basis of a Lyapunov-Razumikhin function, constructing a Lyapunov energy function with low network delay and a spoofing attack intermittent process; 5, on the basis of the robust safety invariant set and a given stability condition, online solving is carried out to obtain a control law gain K; according to the method, the gain of the controller is updated in real time in an online solving mode, and the controller is continuously improved and optimized by using the spoofing attack data of historical batches so as to ensure that the system can effectively resist the spoofing attack.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of advanced control of industrial processes and relates to an iterative learning robust security predictive control method based on deception attack data optimization. Background Art

[0002] Products produced by intermittent process production lines have the characteristics of high added value and small batches, and their related theories and applications have also made certain progress. With the popularization of industrial IoT and the breakthrough of 5G technology, network-based intermittent process control systems have received more attention. Thanks to high-speed network channels, comprehensive information exchange can be carried out between production units, which can greatly improve the working efficiency of the control system.

[0003] However, due to the heavy reliance on the network for information interaction and real-time control, network control systems are vulnerable to malicious network attacks from opponents, including denial of service (DoS) and deception attacks. The damage to the stability of the entire control system is fatal. The secure predictive control method has received widespread attention from researchers. However, related research mainly focuses on the defense of deception attacks on the link from the sensor to the central controller, and rarely considers the deception attack problem in the network link from the controller to the actuator, especially in the control of intermittent processes. Related research has been in a vacuum. Moreover, affected by external factors, the intensity or frequency of the interference or uncertainty factors to which the intermittent process is subjected will show periodic changes. If the random interference range of the current production batch is the same or similar to the deception attack energy range, the deception attack may be masked by the interference. Therefore, the repeatability and periodicity of the intermittent process will make the deception attack more concealed.

[0004] In addition, as industrial IoT and 5G technologies have significantly reduced network latency, the problems of high computational complexity and conservatism of stability conditions derived from traditional time-delay-dependent stability methods have become increasingly prominent, and it is urgent to seek a low-latency stability condition design method. Summary of the invention

[0005] In order to solve the above problems, this paper proposes an iterative learning robust safety predictive control method based on deception attack data optimization for intermittent processes with low latency, unknown interference and deception attacks in a two-dimensional system framework. This method constructs a two-dimensional deception attack Rosser model, including output errors along the time direction and state deviations along the batch direction. Then, based on this model and the deception attack records of historical batches, an iterative learning robust safety predictive control law is designed. Based on the Lyapunov-Razumikhin function, the Lyapunov energy function with low network latency and intermittent processes of deception attacks is constructed, and the stability conditions of the system are derived. By solving the problem online, the real-time control law gain K is obtained to ensure that the system can effectively resist deception attacks and maintain stable and efficient operation.

[0006] The present invention is achieved by the following method:

[0007] The iterative learning robust safety predictive control method based on deception attack data optimization is characterized by: the specific steps are as follows:

[0008] Step 1: Under the two-dimensional system framework, establish a state space model with low network latency and intermittent process of deception attack:

[0009] The state space model with uncertainty, interval time-varying delay and external unknown interference is as follows:

[0010]

[0011] In the formula, Indicates that the discrete The system status at the moment, the input and output under deception attack, and unknown external interference, d m represents the upper bound of the time delay, g represents the number of time delay steps, C represents the discrete The uncertain state matrix, uncertain control input matrix and output matrix of the system at the moment satisfy A g , B represents the state constant matrix and input constant matrix respectively, N, F a and F b Represent the known constant matrices, Indicates the κth batch of discrete The uncertain perturbation at time satisfies

[0012]

[0013] Where I represents the unit matrix of the corresponding dimension;

[0014] The deception attack control input can be expressed as:

[0015]

[0016] In the formula, Indicates that the discrete The control input of the training using historical batch deception attack data at all times, Indicates a deception attack signal, satisfying in, Indicates the κth batch of discrete Control input at all times, The κth batch of discrete Always deceive the attack energy, Indicates that the discrete Always deceiving attack indicators, satisfying:

[0017]

[0018] Step 2: Construct a two-dimensional Rosser model with batch-wise state deviation and time-wise output error:

[0019] For the intermittent process (1) with low network delay and spoofing attack, the iterative learning control law is given as follows:

[0020]

[0021] In the formula, Indicates that the discrete The iterative learning control input at time, u0 represents the initial value;

[0022] The output error along the time direction is defined as:

[0023]

[0024] In the formula, Indicates that the discrete The output error at time y r represents the set value; then define the state deviation along the batch direction as:

[0025]

[0026] In the formula, Δ represents the increment along the batch direction, and f can represent the state, input or interference. According to formulas (5) and (6), the state deviation and output error of the system can be obtained as:

[0027]

[0028] In the formula, Indicates that the discrete The state deviation at the moment, Indicates that the discrete The state deviation at the moment, Indicates that the discrete The expansion of the moment, satisfying Indicates that the discrete The moment interference increment;

[0029] Based on formula (4), the historical batch deception attack data is introduced, and the iterative learning robust security predictive control law is obtained as follows:

[0030]

[0031] In the formula, According to (8), (9) and (10), the two-dimensional Rosser model with deception attack is constructed as follows:

[0032]

[0033] In the formula, represents the extended state constant matrix, represents the extended input constant matrix, D represents the extended interference matrix, and represents the extended constant matrix, represents the expanded output matrix, Indicates that the discrete The state matrix expanded at every moment, Indicates that the discrete The input matrix expanded at each moment, Indicates that the discrete The ever-increasing system state, Indicates that the discrete The ever-increasing system state, Indicates that the discrete The ever-increasing system state, Indicates that the discrete System output of time increments;

[0034] In addition, the probability of a deception attack satisfies the following Bernoulli distribution:

[0035]

[0036] In the formula, Indicates the known constant; incremental deception attack energy is bounded, satisfies Where δ is a constant;

[0037] Step 3: Design a controller based on the established two-dimensional Rosser model and historical batches of deception attack data:

[0038] because The incremental formula is:

[0039]

[0040] Furthermore, the deception attack control law can be designed as:

[0041]

[0042] Where K represents the control law gain, m and n represent the number of prediction steps along the time and batch directions respectively; therefore, the closed-loop two-dimensional deception attack Rosser model can be obtained:

[0043]

[0044] Based on the closed-loop system model, the optimization problem is given as follows:

[0045]

[0046] In the formula, Ω x represents an invariant set, represents the performance index function, Q and R represent the state weight matrix and input weight matrix respectively;

[0047] Step 4: Based on the Lyapunov-Razumikhin function, construct the Lyapunov energy function with low network latency and intermittent process of deception attack:

[0048] The Lyapunov-Razumikhin function with time direction and batch direction is given as follows:

[0049]

[0050] In the formula, represents the Lyapunov-Razumikhin energy function, and represents the intermediate variable, Indicates that the discrete The transpose of the state of the system that is augmented at a moment in time, Indicates that the discrete The transpose of the system state augmented at each moment, P is a positive definite matrix, ξ represents the upper bound of the energy function, and is a scalar; in addition, the invariant set Ω of the system x satisfy:

[0051]

[0052] In the formula, represents mathematical expectation;

[0053] Step 5: Based on the robust safety invariant set, the stability condition is given and solved online to obtain the control law gain K:

[0054] For system (15), if there exists a scalar represents the set of real numbers such that the following inequality holds

[0055]

[0056] Then the invariant set Ω x is a robust and secure invariant set;

[0057] Based on the robust safety invariant set, the following stability conditions are solved online to solve the control law gain K:

[0058]

[0059] in,

[0060]

[0061] and θ g represents the intermediate variable, γ g represents the time delay constant, G, P1 and Y represent unknown positive definite matrices, υ, ε1, ε2 are unknown scalars, 0 represents the 0th-order matrix of the corresponding dimension, and * represents the transpose of the corresponding unknown; represents the inverse matrix of matrix P1, G -1 represents the inverse matrix of matrix G, r max represents the upper bound of the iterative learning control input;

[0062] By solving the linear matrix inequality conditions (20) and (21) online, the system control law gain K is obtained to ensure that the system can effectively resist deception attacks and maintain stable and efficient operation. Compared with the existing methods, the advantages of the present invention are as follows:

[0063] Based on the network control framework, the present invention develops an iterative learning robust safety predictive control method based on deception attack data optimization. By incorporating historical batches of deception attack data into the controller design, the controller is continuously improved and optimized, thereby ensuring that the system can effectively resist deception attacks. At the same time, in the derivation of stability conditions, a robust safety invariant set is introduced to enhance the robustness of the system by limiting the system state within a safe range, which can effectively deal with the impact of deception attacks on system stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1Range diagram for random deception attack and jamming;

[0065] Figure 2 The method proposed in the present invention is Output response diagram when ;

[0066] Figure 3 The method proposed in the present invention is Control input diagram when ;

[0067] Figure 4 The method proposed in the present invention is Time 1 Image trajectory diagram;

[0068] Figure 5 This is a comparison diagram of the output and control input of the method proposed by the present invention and the traditional ILC method;

[0069] Figure 6 Comparison of control input after stable tracking between the proposed method and the traditional ILC method;

[0070] Figure 7 Simplified illustration of the injection molding process: (a) injection phase, (b) holding phase, (c) cooling phase, (d) mold opening and part ejection phase;

[0071] Figure 8 It is a flow chart of the steps of the present invention. DETAILED DESCRIPTION

[0072] The present invention will be further explained below in conjunction with the accompanying drawings and embodiments.

[0073] The iterative learning robust safety predictive control method based on deception attack data optimization is characterized by: the specific steps are as follows:

[0074] Step 1: Under the two-dimensional system framework, establish a state space model with low network latency and intermittent process of deception attack:

[0075] The state space model with uncertainty, interval time-varying delay and external unknown interference is as follows:

[0076]

[0077] In the formula, Indicates that the discrete The system status at the moment, the input and output under deception attack, and unknown external interference, d m represents the upper bound of the time delay, g represents the number of time delay steps, C represents the discrete The uncertain state matrix, uncertain control input matrix and output matrix of the system at the moment satisfy A g , B represents the state constant matrix and input constant matrix respectively, N, F a and F b Represent the known constant matrices, Indicates the κth batch of discrete The uncertain perturbation at time satisfies

[0078]

[0079] Where I represents the unit matrix of the corresponding dimension;

[0080] The deception attack control input can be expressed as:

[0081]

[0082] In the formula, Indicates that the discrete The control input of the training using historical batch deception attack data at all times, Indicates a deception attack signal, satisfying in, Indicates the κth batch of discrete Control input at all times, The κth batch of discrete Always deceive the attack energy, Indicates that the discrete Always deceiving attack indicators, satisfying:

[0083]

[0084] Step 2: Construct a two-dimensional Rosser model with batch-wise state deviation and time-wise output error:

[0085] For the intermittent process (1) with low network delay and spoofing attack, the iterative learning control law is given as follows:

[0086]

[0087] In the formula, Indicates that the discrete The iterative learning control input at time, u0 represents the initial value;

[0088] The output error along the time direction is defined as:

[0089]

[0090] In the formula, Indicates that the discrete The output error at time y rrepresents the set value; then define the state deviation along the batch direction as:

[0091]

[0092] In the formula, Δ represents the increment along the batch direction, and f can represent the state, input or interference. According to formulas (5) and (6), the state deviation and output error of the system can be obtained as:

[0093]

[0094] In the formula, Indicates that the discrete The state deviation at the moment, Indicates that the discrete The state deviation at the moment, Indicates that the discrete The expansion of the moment, satisfying Indicates that the discrete The moment interference increment;

[0095] Based on formula (4), the historical batch deception attack data is introduced, and the iterative learning robust security predictive control law is obtained as follows:

[0096]

[0097] In the formula, According to (8), (9) and (10), the two-dimensional Rosser model with deception attack is constructed as follows:

[0098]

[0099] In the formula, represents the extended state constant matrix, represents the extended input constant matrix, D represents the extended interference matrix, and represents the extended constant matrix, represents the expanded output matrix, Indicates that the discrete The state matrix expanded at every moment, Indicates that the discrete The input matrix expanded at each moment, Indicates that the discrete The ever-increasing system state, Indicates that the discrete The ever-increasing system state, Indicates that the discrete The ever-increasing system state, Indicates that the discrete System output of time increments;

[0100] In addition, the probability of a deception attack satisfies the following Bernoulli distribution:

[0101]

[0102] In the formula, Indicates the known constant; incremental deception attack energy is bounded, satisfies Where δ is a constant;

[0103] Step 3: Design a controller based on the established two-dimensional Rosser model and historical batches of deception attack data:

[0104] because The incremental formula is:

[0105]

[0106] Furthermore, the deception attack control law can be designed as:

[0107]

[0108] Where K represents the control law gain, m and n represent the number of prediction steps along the time and batch directions respectively; therefore, the closed-loop two-dimensional deception attack Rosser model can be obtained:

[0109]

[0110] Based on the closed-loop system model, the optimization problem is given as follows:

[0111]

[0112] In the formula, Ω x represents an invariant set, represents the performance index function, Q and R represent the state weight matrix and input weight matrix respectively;

[0113] Step 4: Based on the Lyapunov-Razumikhin function, construct the Lyapunov energy function with low network latency and intermittent process of deception attack:

[0114] The Lyapunov-Razumikhin function with time direction and batch direction is given as follows:

[0115]

[0116] In the formula, represents the Lyapunov-Razumikhin energy function, and represents the intermediate variable, Indicates that the discrete The transpose of the state of the system that is augmented at a moment in time, Indicates that the discrete The transpose of the system state augmented at each moment, P is a positive definite matrix, ξ represents the upper bound of the energy function, and is a scalar; in addition, the invariant set Ω of the system x satisfy:

[0117]

[0118] In the formula, represents mathematical expectation;

[0119] Step 5: Based on the robust safety invariant set, the stability condition is given and solved online to obtain the control law gain K:

[0120] For system (15), if there exists a scalar represents the set of real numbers such that the following inequality holds

[0121]

[0122] Then the invariant set Ω x is a robust and secure invariant set;

[0123] Based on the robust safety invariant set, the following stability conditions are solved online to solve the control law gain K:

[0124]

[0125] in,

[0126]

[0127] and θ g represents the intermediate variable, γ g represents the time delay constant, G, P1 and Y represent unknown positive definite matrices, υ, ε1, ε2 are unknown scalars, 0 represents the 0th-order matrix of the corresponding dimension, and * represents the corresponding unknown transpose; P1 -1 represents the inverse matrix of matrix P1, G -1 represents the inverse matrix of matrix G, r max represents the upper bound of the iterative learning control input;

[0128] By solving the linear matrix inequality conditions (20) and (21) online, the system control law gain K is obtained to ensure that the system can effectively resist deception attacks and maintain stable and efficient operation.

[0129] Embodiment 1:

[0130] The present invention proposes an iterative learning robust safety predictive control method based on deception attack data optimization, which can effectively solve the problems of network delay and deception attack.

[0131] The input and output model of the injection molding process with uncertainty is as follows:

[0132] The injection velocity (IV) model corresponding to the valve opening (VO) during the injection stage is:

[0133]

[0134] The model of nozzle pressure (NP) corresponding to injection velocity (IV) is:

[0135]

[0136] Where IV is the injection rate, VO is the valve opening, and NP is the intracavitary pressure. Definition:

[0137]

[0138] The system state space model is as follows:

[0139]

[0140] Based on the above formula, considering unknown bounded interference, the state space model is as follows:

[0141]

[0142] In the formula, Set value y r =80, υ=0.1, Q=diag{10,3,10}, R=0.5,

[0143] Figure 1 The random spoofing attack and jamming in the system are shown. Figure 1 In , we set the deception attack energy and interference to be random and bounded, and the upper and lower limits of the deception attack and interference both satisfy [-0.5 0.5].

[0144] Figure 2 The figure shows part of the system output from the first batch to the twentieth batch. It can be seen that the output tracking of the system in the first batch is slightly poor, and the output after the third batch can achieve stable tracking of the set point.

[0145] Figure 3Shows the input situation of some batches between Batch 1 and Batch 20. From Figure 3 it can be seen that the controller fluctuates around 1 bar. The lower the energy input to the controller, the lower the energy consumption. The proposed method can produce high-quality products with lower energy consumption in a shorter operation cycle.

[0146] Figure 4 Shows the phase trajectory of the system state (system output). It can be seen that the proposed method only requires one operation cycle for the image trajectory to converge near the singularity. This indicates that the proposed method converges faster in the batch direction, that is, it generates high-quality products in fewer operation cycles.

[0147] Figure 5 Shows the comparison results of the output and input between the proposed method and the traditional ILC method. From Figure 5 the output response, the traditional ILC method (red curve) requires 20 cycles to achieve stable tracking of the setpoint, while the proposed method only requires 3 operation cycles. This shows that the proposed method has superior convergence performance in the batch direction. In addition, under the same random disturbance conditions, the output fluctuation is significantly lower than that of the comparison method. This indicates that the method has excellent tracking performance, thus improving the product quality. In the comparison of the control input, in the initial operation cycle, the traditional ILC method cannot optimize or improve the control performance due to the lack of historical control input data, resulting in the inability to achieve stable tracking of the setpoint in the initial batch. In addition, the gain of the offline control method cannot effectively solve the real-time dynamic characteristics of the system, resulting in large fluctuations in the control input and reducing the energy utilization efficiency.

[0148] Figure 6 Compares the control inputs of the two methods after stable tracking of the setpoint, including Batches 21, 22, and 23. The gray area represents the low energy input fluctuation range (0.8 < OD < 1.2). This intuitively reflects the control input fluctuations of the two methods. It can be seen that the control input fluctuation range of the proposed method is basically within the low energy input fluctuation range. In contrast, at the same sampling time, the control input adjustment range of the comparison method is larger.

[0149] The above results show that the method proposed in the present invention can ensure the stable operation of the system under deception attacks, and compared with the traditional ILC method, it has a faster convergence speed in the batch direction and smaller fluctuations in the control input, which means that the present invention has higher energy utilization efficiency.

Claims

1. An iterative learning robust safety predictive control method based on deception attack data optimization, characterized by: The specific steps are as follows: Step 1: Under the two-dimensional system framework, establish a state space model with low network latency and intermittent process of deception attack: The state space model with uncertainty, interval time-varying delay and external unknown interference is as follows: In the formula, Indicates that the discrete The system status at the moment, the input and output under deception attack, and unknown external interference, d m represents the upper bound of the time delay, g represents the number of time delay steps, They represent the discrete The uncertain state matrix, uncertain control input matrix and output matrix of the system at the moment satisfy A g , B represents the state constant matrix and input constant matrix respectively, N, F a and F b Represent the known constant matrices, Indicates the κth batch of discrete The uncertain perturbation at time satisfies Where I represents the unit matrix of the corresponding dimension; The deception attack control input can be expressed as: In the formula, Indicates that the discrete The control input of the training using historical batch deception attack data at all times, Indicates a deception attack signal, satisfying in, Indicates the κth batch of discrete Control input at all times, The κth batch of discrete Always deceive the attack energy, Indicates that the discrete Always deceiving attack indicators, satisfying: Step 2: Construct a two-dimensional Rosser model with batch-wise state deviation and time-wise output error: For the intermittent process (1) with low network delay and spoofing attack, the iterative learning control law is given as follows: In the formula, Indicates that the discrete The iterative learning control input at time, u0 represents the initial value; The output error along the time direction is defined as: In the formula, Indicates that the discrete The output error at time y r represents the set value; then define the state deviation along the batch direction as: In the formula, Δ represents the increment along the batch direction, and f can represent the state, input or interference. According to formulas (5) and (6), the state deviation and output error of the system can be obtained as: In the formula, Indicates that the discrete The state deviation at the moment, Indicates that the discrete The state deviation at the moment, Indicates that the discrete The expansion of the moment, satisfying Indicates that the discrete The moment interference increment; Based on formula (4), the historical batch deception attack data is introduced, and the iterative learning robust security predictive control law is obtained as follows: In the formula, According to (8), (9) and (10), the two-dimensional Rosser model with deception attack is constructed as follows: In the formula, represents the extended state constant matrix, represents the extended input constant matrix, D represents the extended interference matrix, and represents the extended constant matrix, represents the expanded output matrix, Indicates that the discrete The state matrix expanded at every moment, Indicates that the discrete The input matrix expanded at each moment, Indicates that the discrete The ever-increasing system state, Indicates that the discrete The ever-increasing system state, Indicates that the discrete The ever-increasing system state, Indicates that the discrete System output of time increments; In addition, the probability of a deception attack satisfies the following Bernoulli distribution: In the formula, Indicates the known constant; incremental deception attack energy is bounded, satisfies Where δ is a constant; Step 3: Design a controller based on the established two-dimensional Rosser model and historical batches of deception attack data: because The incremental formula is: Furthermore, the deception attack control law can be designed as: Where K represents the control law gain, m and n represent the number of prediction steps along the time and batch directions respectively; therefore, the closed-loop two-dimensional deception attack Rosser model can be obtained: Based on the closed-loop system model, the optimization problem is given as follows: In the formula, Ω x represents an invariant set, represents the performance index function, Q and R represent the state weight matrix and input weight matrix respectively; Step 4: Based on the Lyapunov-Razumikhin function, construct the Lyapunov energy function with low network latency and intermittent process of deception attack: The Lyapunov-Razumikhin function with time direction and batch direction is given as follows: In the formula, represents the Lyapunov-Razumikhin energy function, and represents the intermediate variable, Indicates that the discrete The transpose of the state of the system that is augmented at a moment in time, Indicates that the discrete The transpose of the system state augmented at each moment, P is a positive definite matrix, ξ represents the upper bound of the energy function, and is a scalar; in addition, the invariant set Ω of the system x satisfy: In the formula, represents mathematical expectation; Step 5: Based on the robust safety invariant set, the stability condition is given and solved online to obtain the control law gain K: For system (15), if there exists a scalar represents the set of real numbers such that the following inequality holds Then the invariant set Ω x is a robust and secure invariant set; Based on the robust safety invariant set, the following stability conditions are solved online to solve the control law gain K: in, and θ g represents the intermediate variable, γ g represents the time delay constant, G, P1 and Y represent unknown positive definite matrices, υ, ε1, ε2 are unknown scalars, 0 represents the 0th-order matrix of the corresponding dimension, and * represents the transpose of the corresponding unknown; represents the inverse matrix of matrix P1, G -1 represents the inverse matrix of matrix G, r max represents the upper bound of the iterative learning control input; By solving the linear matrix inequality conditions (20) and (21) online, the system control law gain K is obtained to ensure that the system can effectively resist deception attacks and maintain stable and efficient operation.

Citation Information

Cited By

  • Anti-spoofing-attack robust safety prediction control method in injection molding process

    CN122131673A