A Memory-Based Invasion-Tolerant Control Method for Glucose Regulation Systems Based on Radial Basis Function Neural Network Attack Compensation

Abstract

This invention discloses a memory-based attack-tolerant control method for a blood glucose regulation system based on radial basis function neural network attack compensation. First, the T-S fuzzy modeling method and a T-S fuzzy observer based on sampled output are used to handle the system's nonlinearity and estimate the overall system state, respectively. Second, to maintain system security in the event of malicious attacks in the communication network, an estimator based on a radial basis function neural network is introduced to approximate the actual attack signal. Then, using the estimated state, the memory state related to the sampling period, and the estimated attack signal, a memory-based attack-tolerant controller based on T-S fuzzy logic is constructed. Finally, using Lyapunov stability theory, a sufficient condition for ensuring system stability is derived. Based on this condition, the controller and observer gains are solved, and blood glucose is regulated to restore the blood glucose level to a safe range. Compared to existing control methods, this invention proposes a memory-based attack-tolerant control method based on attack compensation, effectively reducing the disruption of the blood glucose regulation process by malicious attack signals, improving the stability and security of blood glucose control, and possessing certain engineering application value.

Description

Technical Field

[0001] This invention relates to a memory-based intrusion-tolerant control method for a blood glucose regulation system based on radial basis function neural network attack compensation, and more particularly to a security control method for a networked blood glucose regulation system where malicious attacks exist on the control signal. Background Technology

[0002] In recent years, blood glucose regulation systems based on artificial pancreas have become a promising technology for maintaining blood glucose levels in diabetic patients. These systems employ a closed-loop control mechanism, continuously monitoring blood glucose levels and administering insulin as needed, mimicking the glucose regulation function of a healthy pancreas. With advancements in the Internet of Things (IoT) and network communication technologies, blood glucose regulation systems have evolved into networked and intelligent systems, known as networked artificial pancreas systems. Through network-connected devices, networked artificial pancreas systems can achieve real-time data exchange and more flexible blood glucose control.

[0003] However, with the convenience of network applications, their vulnerability to malicious attacks is an unavoidable issue. Networked artificial pancreas systems are susceptible to threats and risks from cyberattacks, such as denial-of-service attacks and spoofed data injection attacks, which may lead to decreased blood regulation performance or even hyperglycemia or hypoglycemia. Therefore, the security control of networked systems has attracted increasing interest and attention from researchers. At the same time, from the perspective of attack tolerance, designing resilient mechanisms is also a focus of many researchers. Summary of the Invention

[0004] Objectives of the invention: 1) To address the nonlinear dynamics and unmeasurable full-state problems of networked artificial pancreas systems, a Time-varying and time-delayed TS fuzzy system model is constructed using both the TS fuzzy modeling method and a state observer based on sampled data. 2) A memory-based attack-tolerant control scheme based on radial basis function neural networks is proposed. This scheme utilizes a memory observer state based on historical data and attack estimation signals to improve control performance.

[0005] The specific steps of this invention are as follows: Step 1: Establish a state-space model of the blood glucose regulation system in patients with type 1 diabetes; Step 2: Design an attack estimator based on a radial basis function neural network; Step 3: Construct a memory-based invasion-tolerant controller for the observed state and attack estimation using the estimated state, the memory state related to the sampling period, and the attack estimation signal; Step 4: Export the closed-loop control system model and stability analysis conditions, use these conditions to determine the controller and observer parameters, and perform blood glucose control on patients with type 1 diabetes. Technical solution:

[0006] The memory-type invasion-tolerant control method for a blood glucose regulation system based on radial basis function neural network attack compensation is characterized by the following state-space model: Where η1(t) is blood glucose concentration, η2(t) is remote insulin concentration, η3(t) is serum insulin concentration, η4(t) is dietary disturbance, and G b I is the baseline value for blood glucose concentration. b κ1 represents the baseline value of insulin concentration, κ3 / κ2 represents insulin-independent glucose utilization, κ4 represents insulin degradation rate, and κ5 represents the effect of external blood glucose fluctuations on the time to peak blood glucose concentration. Based on the above networked artificial pancreas system model, its state-space model is established under deviation conditions, specifically including: The dynamics of the system can be rewritten as its deviation state: [η1(t) η2(t) η3(t) η4(t)] T =[η 10 η 20 η 30 η 40 ] T +[x l (t) x2(t) x3(t) x4(t)] T Where u(t) represents the control input (insulin infusion rate), and x(t) = [x1(t) x2(t) x3(t) x4(t)] T It concerns the equilibrium point η0 = [η 10 η 20 η 30 η 40 ] T =[G b 0 I b 0] T The deviation state. The blood glucose regulation system model based on TS fuzzy logic is as follows: in, Since the continuous glucose monitoring system only measures blood glucose concentration, the following TS fuzzy observer was designed to estimate the state of the entire system: in: For the state of the fuzzy observer, L is the output of the fuzzy observer. j Let u be the gain matrix of the fuzzy observer to be designed. a (t) represents the real control signal for the fake data injection attack, and The system (2) constrained by the injection of attack signals into unknown false data is described as follows: Furthermore, it is assumed Define the observer error as The error system can be derived from the following formula: Among them, h(t)=t-κh, t∈[kh, (κ+1)h].

[0007] To reconstruct unknown spoofing attack signals, an estimator based on radial basis function neural networks is designed, taking advantage of the fact that radial basis function neural networks can approximate any continuous function with arbitrary precision. Fake data injection attack signal It can be described as: in, Indicates the approximation error. The Gaussian function of the hidden layer. This represents the ideal weight vector. In a radial basis function neural network structure, For the network's input vector, Where c d and b is The center and width. The ideal weights and output of a radial basis function neural network are shown below: in, These are the output layer weights to be designed. The estimation error of the weight vector, The estimated attack signal is obtained through a radial basis function neural network.

[0008] To control blood glucose levels in diabetic patients within a safe range, a memory-based invasion tolerance controller based on TS fuzzy logic is constructed as follows: In the formula K j and F j The gain of the state feedback controller that needs to be designed. If F j =0, This can then be further simplified to a memoryless conventional controller: By definition The augmented closed-loop time-varying TS fuzzy system is obtained: Where z(t) is the system performance output, and

[0009] Find the derivative of the Lyapunov function and require it to be negative definite. Derive the corresponding linear matrix inequality conditions for solving the controller, specifically including: 1) Select the Lyapunov-Krasovskii functional as shown below: V(t)=V1(t)+V2(t)+V3(t), (12)where: V1(t)=<P|δ(t)> , 2) Based on the previous step, calculate the derivative of the Lyapunov-Krasovskii functional: in: To find the update rules of radial basis function neural networks, based on (6), (8), and (10), we have According to the above formula, the update law can be solved as follows: Since the system state e(t) in δ(t) is unmeasurable, the choice matrix is... Then (14) equals: It can be implemented using all measurable signals and matrices. According to Wirtinger's inequality, we have: Further, there are: To simplify the expression, we define an augmented vector: in: For system (10) and the constructed matrix P, we can obtain: 3) Further, we can obtain the following matrix inequality: according to The result is: According to (21) and (22), and taking into account and Θ ij -Γ i <0, can be obtained 4) The sufficiency conditions for the asymptotic stability of the TS fuzzy networked closed-loop system are further obtained as follows: For a given scalar h, α1, α2, ν1, ν2, Augmented systems (10) constrained by spoofed data injection attacks are uniformly eventually bounded if there exist P > 0, Q > 0, R > 0 and matrix Γ. i (i = 1, 2), H such that: Θ ij -Γ i <0, (25) F ii <0, (26) F ij +F ji ≤0(i <j), (27) in: F ij =μ j (Θ ij -Γ i )+Γ i , 5) Based on the above sufficient conditions, matrix X = P -1 Constructed as By employing left-multiplication and right-multiplication techniques along with Schur's complement, the conditions are further transformed into optimization conditions, yielding the following linear matrix inequality: in: 6) The controller and observer gains that meet the system requirements can be obtained using the mincx solver in MATLAB's LMI toolbox:

[0010] The beneficial effects of this invention: This invention proposes a novel memory-based attack-tolerant control scheme based on radial basis function neural networks. Compared with traditional memoryless and uncompensated control schemes, this control scheme uses radial basis function neural networks to estimate and compensate for unknown attacks, offsetting the performance degradation caused by the attacks. By utilizing the memory observer state based on historical data, it further improves the control accuracy of the controller, enhances control performance, and improves the stability and safety of blood glucose control in diabetic patients. Attached Figure Description

[0011] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0012] Figure 2 This is a comparison of the attack estimation signal and the actual attack signal over time based on a radial basis function neural network.

[0013] Figure 3 The study shows the blood glucose concentration trends in type 1 diabetic patients subjected to spoofing attacks under both memory-based and memoryless uncompensated controller scenarios. Detailed Implementation

[0014] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0015] like Figure 1 As shown, a memory-based security control method for a blood glucose regulation system based on attack estimation and compensation using a radial neural network includes the following steps: Step 1: The established blood glucose regulation system model based on TS fuzzy logic is as follows: in, Step 2: Design a state observer based on TS fuzziness and an attack estimator based on radial basis function neural network. The specific steps are as follows: 1) Since the continuous glucose monitoring system only measures blood glucose concentration, a TS fuzzy observer was designed to estimate the state of the entire system, which is given as follows: in: For the state of the fuzzy observer, L is the output of the fuzzy observer. j Let u be the gain matrix of the fuzzy observer to be designed. a(t) represents the real control signal for the fake data injection attack, and The system (2) constrained by the injection of attack signals into unknown false data is described as follows: Furthermore, it is assumed Define the observer error as The error system can be derived from the following formula: Where h(t) = t - kh, t ∈ [kh, (k+1)h]. 2) To reconstruct unknown spoofed data injection attack signals, an attack estimator based on radial basis function neural networks (RBNs) was designed, leveraging the RBN's ability to approximate any continuous function with arbitrary precision: in, These are the output layer weights to be designed. The estimation error of the weight vector, The estimated attack signal is obtained through a radial basis function neural network. Step 3: Using the estimated state, the memory state related to the sampling period, and the attack estimation signal, construct a memory-based invasion-tolerant controller based on the observed state and attack estimation, as follows: In the formula K j and F j The gain of the state feedback controller that needs to be designed. Step 4: By defining The augmented closed-loop time-varying TS fuzzy system is obtained: Where z(t) is the system performance output, and By selecting a novel Lyapunov-Krasovskii functional and a weighting matrix for the radial basis function neural network, some sufficient linear matrix inequalities are derived for calculating the gains of the fuzzy controller and observer, as follows: in: 6) The controller that meets the system requirements can be obtained using the mincx solver in MATLAB's LMI toolbox:

[0016] Step 3 of this invention proposes a novel memory-based intrusion-tolerant control scheme based on observation state and attack estimation. This scheme uses the current and historical observer states as well as the estimated attack signals, which helps to improve control performance and enhance the security and stability of the system.

[0017] An embodiment of the present invention is described below: The parameters for the blood glucose regulation system model are selected as shown in the table below: Table 1. Parameter values ​​for the networked artificial pancreas system With sampling time h = 1 min, α1 = 0.01, α2 = 1.5, and b = 0.001, solve the linear matrix inequality. The corresponding controller and observer gains are: K1=[0.3432 -982.1270 -0.2891 3.5906], K2=[0.3470 -950.9119 -0.2953 3.5120], F1=[0.0165 3.0766 -0.0262 0.0148], F2=[0.0299 -75.1159 -0.0380 0.2408], Without considering memory delay signals, the corresponding memoryless controller and observer gains in [13, 9, 10] are: K1=[0.3419 -881.6173 -0.3062 3.3846], K2=[0.3578 -926.5317 -0.3226 3.5694], In the simulation, assume x(0) = [100 0.001 0 10] T , Based on G b＝ 80, I b =10, so we can get η(0) = [180 0.001 10 10] T . Figure 2 This is a comparison graph showing the changes of the attack signal and its estimate over time, where curve a represents the attack signal and curve b represents the estimated attack signal. Figure 3The graph shows a comparison between the memory-type invasive controller and the conventional controller without memory, where curve a represents the trend of blood glucose concentration over time under the memory-type invasive controller and curve b represents the trend of blood glucose concentration over time under the conventional controller without memory. Figure 2 This demonstrates that the constructed radial basis function neural network can accurately approximate the spoofed data injection attack signal. Figure 3 The results show that the duration of hyperglycemia in diabetic patients using a conventional, memoryless controller was approximately 210 minutes, while the duration using a memory-based attack-tolerant controller was drastically reduced to approximately 100 minutes. This comparison validates the advantages of incorporating historical data to improve control effectiveness and using estimated attack signals to compensate for the negative impact of actual malicious attacks, thus helping to reduce the risk of hyperglycemia in diabetic patients.

Claims

1. A memory-based attack-tolerant control method for a blood glucose regulation system based on radial basis function neural network attack compensation, comprising the following steps: Step 1: Establish a state-space model of the blood glucose regulation system in patients with type 1 diabetes; Step 2: Design a state observer based on TS fuzzy logic and an attack estimator based on radial basis function neural network; Step 3: Construct a memory-based intrusion-tolerant controller based on the observed state and attack estimation using the estimated state, the memory state related to the sampling period, and the attack estimation signal; Step 4: Export the closed-loop control system model and stability analysis conditions, use these conditions to determine the controller and observer parameters, and perform blood glucose control on patients with type 1 diabetes. The memory-type invasion-tolerant control method for a blood glucose regulation system based on radial basis function neural network attack compensation is characterized in that, A blood glucose regulation system for type 1 diabetes patients based on the Bergman model was established as follows: Where η1(t) is blood glucose concentration, η2(t) is remote insulin concentration, η3(t) is serum insulin concentration, η4(t) is dietary disturbance, and G b I is the baseline value for blood glucose concentration. b κ1 represents the baseline value of insulin concentration, κ3 / κ2 represents insulin-independent glucose utilization, κ4 represents insulin degradation rate, and κ5 represents the effect of external blood glucose fluctuations on the time to peak blood glucose concentration. Based on the above blood glucose regulation system model, its state-space model is established under the deviation state, specifically including: The dynamics of the system can be rewritten as its deviation state: [η1(t) η2(t) η3(t) η4(t)] T =[the 10 or 20 or 30 or 40 ] T +[x1(t) x2(t) x3(t) x4(t)] T , Where u(t) represents the control input (insulin infusion rate), It concerns the equilibrium point η0 = [η 10 η 20 η 30 η 40 ] T =[G b 0 I b 0] T The deviation state. The blood glucose regulation system model based on TS fuzzy logic is as follows: in, ν2(t)=1-ν1(t). Step 2: Design an attack estimator based on a radial basis function neural network. Since the dynamic blood glucose monitoring system only measures blood glucose concentration, a TS fuzzy observer is designed to estimate the state of the entire system, which is given as follows: in: For the state of the fuzzy observer, L is the output of the fuzzy observer. j Let u be the gain matrix of the fuzzy observer to be designed. a (t) represents the real control signal for the fake data injection attack, and The system (2) constrained by the injection of attack signals into unknown false data is described as follows: Furthermore, it is assumed Define the observer error as The error system can be derived from the following formula: Where h(t) = t - kh, t ∈ [kh, (k+1)h]. To reconstruct unknown spoofing attack signals, an estimator based on radial basis function neural networks was designed, taking advantage of the fact that radial basis function neural networks can approximate any continuous function with arbitrary precision. Fake data injection attack signal It can be described as: in, Indicates the approximation error. The Gaussian function of the hidden layer. This represents the ideal weight vector. In a radial basis function neural network structure, For the network's input vector, Where c d and b is The center and width. The ideal weights and output of a radial basis function neural network are shown below: in, These are the output layer weights to be designed. The estimation error of the weight vector, The estimated attack signal is obtained through a radial basis function neural network.

2. The memory-type invasion-tolerant control method for a blood glucose regulation system based on radial basis function neural network attack compensation according to claim 1, characterized in that: Step 3: Using the estimated state, the memory state related to the sampling period, and the attack estimation signal, construct the following memory-based invasion-tolerant controller based on the observation state and attack estimation: In the formula K j and F j The gain of the state feedback controller that needs to be designed. If F j =0, This can then be further simplified to a memoryless conventional controller: Step 4: Define A closed-loop blood glucose regulation system based on TS fuzzy logic was obtained: Where z(t) is the system performance output, and 3) Select the Lyapunov-Krasovskii functional as shown below: V(t)=V1(t)+V2(t)+V3(t), (12) in: V1(t)=<P|δ(t)> , 4) Based on the previous step, calculate the derivative of the Lyapunov-Krasovskii functional: in: To find the update rules of radial basis function neural networks, based on (6), (8), and (10), we can obtain... According to the above formula, the update law can be solved as follows: Since the system state e(t) in δ(t) is unmeasurable, the choice matrix is... Then (14) equals: It can be implemented using all measurable signals and matrices. According to Wirtinger's inequality, we have: Further, there are: To simplify the expression, we define an augmented vector: in: For system (10) and the constructed matrix P, we can obtain: 5) Further, we can obtain the following matrix inequality: according to The result is: According to (21) and (22), and taking into account and Θ ij -Γ i <0, we can get 6) Further, the sufficiency conditions for the asymptotic stability of the TS fuzzy networked closed-loop system are as follows: For a given scalar h, α1, α2, ν1, ν2, Augmented systems (10) constrained by spoofed data injection attacks are uniformly eventually bounded if there exist P > 0, Q > 0, R > 0 and matrix Γ. i (i = 1, 2), H such that: I ij -C i <0, (25) F ii <0, (26)F ij +F ji ≤0(i<j), (27) in: F ij =μ j (I ij -C i )+C i , n δ =2n x a=1,...,6。 7) Based on the above sufficient conditions, matrix X = P -1 Constructed as By employing left-multiplication and right-multiplication techniques along with Schur's complement, the conditions are further transformed into optimization conditions, yielding the following linear matrix inequality: in: 8) The controller and observer gains that meet the system requirements can be obtained using the mincx solver in MATLAB's LMI toolbox: 9) Incorporate the controller and observer gain into the closed-loop blood glucose regulation system to control blood glucose in diabetic patients, enabling the control system to maintain a safe blood glucose level even under attack.

Images