Safe dynamic positioning method of load switching unmanned ship based on spoofing attack detection

Abstract

The invention discloses a safe dynamic positioning control method for a load switching unmanned ship based on spoofing attack detection, and the method comprises the steps: constructing a switching rule based on an A2C algorithm according to an unmanned ship load switching model, so as to select an optimal action corresponding to the load mode switching of the unmanned ship; on the basis of an unmanned ship load mode corresponding to the optimal action, according to a constructed first event triggering mechanism, in combination with a constructed state observer and an A2C algorithm mode estimation / detection mechanism, whether a data signal is tampered by a spoofing attack or not is determined, and then the observation state of the unmanned ship is output; acquiring a control signal of the unmanned ship according to the observation state based on the constructed safety controller; and based on the constructed second event triggering mechanism, the control signal is transmitted to the actuator through a preset communication network, so that the safe dynamic positioning control of the load switching unmanned ship based on spoofing attack detection is realized. According to the method, the problem that the existing method cannot effectively realize load switching of the unmanned ship and cannot realize safe dynamic positioning control under the condition of considering spoofing attacks is solved.

Description

Technical Field

[0001] This invention relates to the field of ship safety dynamic positioning technology, and in particular to a safety dynamic positioning method for unmanned vessels with load switching based on deception attack detection. Background Technology

[0002] As the core carrier of maritime activities, ships play a fundamental role in many fields, including military defense, transportation, fisheries production, coastal tourism, environmental monitoring, oil development, pollution control, and scientific exploration. The motion control of manned ships mainly includes roll stabilization control, heading control, anchoring control, and dynamic positioning. With the rapid development of unmanned surface vessel (USV) technology and its widespread application in military, fisheries, tourism, and environmental monitoring, research on USVs with multimodal characteristics has become crucial. USVs exhibit different motion modes and behavioral characteristics during operation. These multimodal characteristics include changes in speed and direction, load changes, and position relative to the water surface, such as underwater, semi-submerged, or surface. Due to complex coupled nonlinearities and significant time lags, a single control strategy is insufficient to meet the performance requirements under all mission conditions. Specifically, different missions may require USVs to move at different speeds or directions, and in scenarios such as cargo loading and unloading, the load on the USV will change significantly. All of these require the control system to dynamically adjust according to the current mission to ensure optimal performance.

[0003] With the development of network technology, an increasing number of unmanned surface vessels (USVs) are relying on network communication to achieve remote control and collaborative operations among multiple USVs. For example, a single USV can interact with a remote land controller via a network, while a multi-USV system relies on the network for communication between neighbors. However, while network communication brings convenience, it also poses significant challenges to the security of control strategies, as it can expose transmitted data to various malicious attacks. On the one hand, attackers may tamper with the status or output information transmitted by USVs, turning it into insecure signals, causing the controller to generate insecure control inputs, thus leading to instability and performance degradation of the USVs. In particular, in multimodal USVs, the network output includes modal information, which may lead to asynchronous behavior between a single USV and the controller, further affecting control performance. On the other hand, for multi-USV systems, attackers may tamper with the neighbor information they receive, thereby altering the communication link and causing changes in the communication topology. Such attacks not only disrupt the normal operation of the system but also affect the collaborative operation capabilities among multiple USVs. Given these issues, researching how to defend against these potential network attacks and ensure the safe dynamic positioning of USVs is particularly important.

[0004] In collaborative operations involving multiple unmanned surface vessels (USVs), external interference and sensor malfunctions can lead to communication disruptions between some vessels. Employing a switching topology can more flexibly address these disruptions and reduce network resource requirements. By developing various switching model establishment technologies based on mass switching, heading angle switching, mission mode switching, and communication topology switching, it is possible to ensure that USVs select the most suitable model under different operating conditions, thereby accurately maintaining their position and navigation status. To address the need for different control strategies when vessels navigate in open and narrow waters, two types of controllers, one with 2 degrees of freedom and the other with 3 degrees of freedom, were designed, and switching models were established based on conditions such as speed and navigation mission. According to the control requirements of USVs in different operational modes such as navigation, stationary maintenance, and reversing, USVs equipped with different controllers are modeled as switching systems, and a performance-evaluation-based supervised switching control mechanism is used to achieve smooth switching between controllers.

[0005] Most conclusions focus on how cyber-physical systems can achieve adaptive security control through dynamic event triggering mechanisms when subjected to spoofing attacks. Spoofing attacks have been extensively studied not only in network control systems but also in handover systems. A handover Q-learning algorithm is used to detect modes and states separately. The estimated residuals of the system states and modes are used as environmental variables for the handover Q-learning algorithm to obtain the optimal evaluation threshold, which is then used for detection. For cumulative detectors, comparing the accumulated results with a set threshold can determine whether there are anomalies in the current data. A distributed attack detection algorithm combines distributed fixed-time observers and information from two end neighbors to detect attacked network links in general linear multi-agent systems under specific communication topologies.

[0006] Currently, existing solutions primarily rely on passive control strategies that appropriately limit the proportion of stable / unstable switching moments caused by attacks. While this approach can guarantee system stability to some extent, it is overly conservative, leading to significant impacts on system performance during attacks. Especially when the load changes of an unmanned surface vessel (USV) are modeled as a switching system, its modal and state information is vulnerable to tampering during network transmission, resulting in modal mismatch between the USV subsystem and the controller, severely affecting the USV's stability. In fact, when both the USV's state information and load modality are simultaneously tampered with, estimation methods relying solely on state dependencies may also lead to errors in the estimated load modality. Therefore, existing load modality tampering defense strategies are insufficient to describe situations where both state and load modality are tampered with. A deception attack defense strategy that considers both the USV's load modality and other transmitted data being tampered with is needed. A security framework that dynamically adjusts the control strategy accordingly is a critical issue that urgently requires resolution. Furthermore, most existing studies have overlooked the risk of handover signals being tampered with during network transmission. Since handover signals are one of the core characteristics of a handover system, their security in network transmission is crucial. Once the handover signals are attacked, it will lead to modal mismatch between the system and the controller, thereby triggering asynchronous handover behavior and seriously affecting the stability of the system. Summary of the Invention

[0007] This invention provides a secure dynamic positioning method for unmanned surface vessels with load switching based on deception attack detection, in order to overcome the above-mentioned technical problems.

[0008] To achieve the above objectives, the technical solution of this invention is: a safe dynamic positioning control method for unmanned surface vessels (USVs) with load switching based on deception attack detection, specifically including the following steps: S1: Constructing a uniformly segmented USV load switching model; S2: Constructing switching rules based on the A2C algorithm according to the USV load switching model; the method for constructing the switching rules based on the A2C algorithm includes: a training phase and a generation phase of the switching rules; the training phase of the switching rules: that is, constructing a reward function based on the USV load switching model, and training the network parameters of the Actor network based on the reward function to obtain an optimized Actor network; the generation phase of the switching rules: that is, based on the optimized Actor network, according to the environmental variables of the current USV environment, ... S3: Select the optimal action for the corresponding unmanned surface vessel (USV) load mode switching; S4: Based on the USV load mode corresponding to the optimal action, and according to the constructed first event triggering mechanism, transmit the USV load mode and status signal to the remote land control station through a pre-set communication network; S5: Through the constructed state observer and A2C algorithm mode estimation / detection mechanism, confirm whether the data signal transmitted to the remote land control station has been tampered with by a deception attack, and then output the observation status of the USV; and obtain the control signal of the USV based on the observation status according to the constructed safety controller; S6: Based on the constructed second event triggering mechanism, transmit the control signal to the actuator through the pre-set communication network, thereby realizing the safe dynamic positioning control of the USV load switching based on deception attack detection.

[0009] This invention provides a secure dynamic positioning method for unmanned surface vessels with load switching based on deception attack detection, with the following beneficial effects:

[0010] 1) Based on the uniformly segmented unmanned surface vessel (USV) load switching model, a switching rule based on the A2C algorithm is constructed. By defining the USV load mode as the action space, a reward function containing state convergence factors is constructed. Through the iterative learning process, the A2C algorithm can select the optimal load mode according to the current state information of the USV, thereby achieving efficient dynamic positioning of the USV load switching load switching.

[0011] 2) By constructing a state observer and an A2C algorithm modal estimation / detection mechanism, it is confirmed whether the data signals transmitted to the remote land control station have been tampered with by a deception attack, and then the observation state of the unmanned vessel is output. The mechanism design includes a reward function for the convergence of the observer state and the accuracy of the alarm information. By achieving accurate estimation of the unmanned vessel's load mode during the iterative learning process and detecting whether the modal information and state information have been tampered with, the error in the predicted load mode is effectively reduced. A security controller is designed based on the attack detection results, and watermark encryption technology is applied to the control signals transmitted back to the unmanned vessel to resist attacks and tampering. At the same time, combined with the constructed second event triggering mechanism, the control signals are transmitted to the actuator through a pre-set communication network, thereby effectively realizing the accuracy and efficiency of safe dynamic positioning control of the unmanned vessel for load switching based on deception attack detection. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0013] Figure 1 The flowchart shows the safe dynamic positioning control method for unmanned surface vessels with load switching based on deception attack detection according to the present invention.

[0014] Figure 2 This is a system block diagram of the unmanned surface vessel for load switching in this embodiment;

[0015] Figure 3 This is a flowchart of the A2C algorithm for generating switching rules in this embodiment;

[0016] Figure 4 This is a flowchart of the A2C algorithm for detecting spoofing attacks in this embodiment;

[0017] Figure 5 This is a simulation curve of the unmanned surface vessel data in this embodiment;

[0018] Figure 6 These are simulation diagrams of the unmanned vessel mode and the controller mode in this embodiment;

[0019] Figure 7 This is a comparison diagram of modalities under the A2C switching rule and the state-dependent switching rule in this embodiment;

[0020] Figure 8 The unmanned surface vessel error curves for the A2C switching rule and the state-dependent switching rule in this embodiment are shown.

[0021] Figure 9This is a comparison chart of the attack detection results of the A2C algorithm and the Q-learning algorithm in this embodiment;

[0022] Figure 10 This is a comparison chart of the control input u(t) with and without watermark in this embodiment. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] This embodiment provides a secure dynamic positioning control method for unmanned surface vessels with load switching based on deception attack detection, such as... Figure 1 As shown, the specific steps include:

[0025] S1: Construct a uniformly segmented unmanned vessel load switching model;

[0026] Specifically, the following steps are included:

[0027] S11: Obtain the unmanned vessel model, its expression is:

[0028]

[0029] In the formula: η(t) represents the position vector and r(t) represents the lateral and longitudinal positions; ψ(t) represents the heading angle; v(t) represents the velocity vector and Indicates oscillation velocity, Let γ(t) represent the sway velocity, γ(t) represent the bow angular velocity, and u(t) represent the control input vector. u1(t), u2(t), and u3(t) represent the sway force, lateral sway force, and yaw moment, respectively; This represents the disturbance vector describing the continuous impact of wind, waves, and ocean currents on unmanned vessels in the marine environment;

[0030] Among them, the interference vector The following external systems are used:

[0031]

[0032] The disturbance vector describes the continuous impact of wind, waves and ocean currents in the marine environment on the unmanned vessel (1). In order to reflect the dynamic characteristics of these environmental disturbances, this embodiment assumes that all eigenvalues ​​of the S matrix are non-negative.

[0033] M(m(t)), N, and R(ψ) represent the inertia matrix, damping matrix, and coordinate transformation matrix, respectively. Indicates time-varying load, Indicates the maximum load weight. m Indicates the unloaded weight; Indicates the additional hydrodynamic load, x g I represents the longitudinal component of the difference vector from the origin to the centroid in the preset coordinate system. Z Represents the inertia matrix. Y γ N v N γ R(ψ) represents the linear damping coefficient. In this embodiment, since the unmanned vessel can stop and anchor when performing scientific features and exploration tasks, the heading angle ψ is small enough, so R(ψ)≈I, where I represents the identity matrix.

[0034] S12: Set the dynamic positioning control target for the unmanned vessel:

[0035] η(t)→η * ,v(t)→0, (3)

[0036] Where η * Describe the desired position vector and Represents the desired position coordinates for dynamic positioning;

[0037] S13: In this embodiment, in order to characterize the load changes of the unmanned vessel, a uniformly segmented load switching model is constructed for the unmanned vessel, and the time-varying load is... Dividing it into L equal parts, then any load interval segment Can be represented as

[0038]

[0039] In the formula: θ q Indicates intermediate parameters and q represents the interval index; and when When m(t) is considered as m +θ q-1 And denote the interval index q as the subsystem mode, i.e., σ(t) = q. At this time, the inertia matrix M(m(t)) in the unmanned ship model can be written as:

[0040]

[0041] And according to M σ(t) Able to define intermediate variable A σ(t) B σ(t) C σ(t) for:

[0042]

[0043] Let the position vector x(t) = [η T (t)-η *T v T (t)] T Furthermore, in order to reduce the influence of w(t), the measured output y(t) can be constructed as y(t) = D. σ(t) x(t), thus, by (1)-(2), the uniformly segmented unmanned vessel load switching model can be constructed as

[0044]

[0045] In the formula: T represents the transpose; σ(t) represents the load mode of the unmanned vessel and D σ(t) y(t) represents the output matrix corresponding to the load mode; y(t) represents the measurement output of the unmanned vessel's dynamic positioning. Let x(t) represent the first derivative of x(t); x(t) represents the state signal of the unmanned vessel. At this point, the dynamic positioning problem of the unmanned vessel is further transformed into the stabilization and interference suppression problem of the unmanned vessel load switching model (5);

[0046] S2: Construct switching rules based on the A2C algorithm according to the unmanned vessel load switching model;

[0047] The method for constructing switching rules based on the A2C algorithm includes: a training phase for switching rules and a generation phase for switching rules.

[0048] The training phase of the switching rule involves constructing a reward function based on the unmanned vessel load switching model, training the network parameters of the Actor network based on the reward function, and obtaining an optimized Actor network. This specifically includes the following steps:

[0049] S21: Define the quintuple used for training the A2C algorithm. This stage will be modeled as a Markov decision process containing quintuples; where For the number of iterations, The quintuple is explained in detail below: First, Represents the set of environment variables and Let x(kh) represent the environment state x(kh) of the k-th sample in the j-th training round. Represents an action set and 'a' represents the set of load modes in the unmanned vessel load switching model. The value taken from the middle is the action; third, π j For strategy The abbreviation represents the environment variable at step k in the j-th round of training. and Using the parameters of the first Actor network Fourth, ε represents the discount factor used to balance long-term and short-term returns, and ε∈[0,1]. The larger ε is, the more long-term returns are pursued; the smaller ε is, the more short-term returns are pursued. Fifth... Let be the reward function, and let represent the environment variable at step k in the j-th round of training. Execute action The reward received later;

[0050] S22: In order to design a switching rule to stabilize the unmanned vessel load switching model (5), a reward function for stabilizing the model system is constructed, the expression of which is:

[0051]

[0052] In the formula: ρ1, ρ2, ρ3 represent design parameters and ρ1 < 0, ρ2 < 0, ρ3 < 0; Represents an asynchronous matrix; Describes the symmetric matrix under load mode σ(kh) and This represents a symmetric matrix under load mode σ(kh-h); Denotes a symmetric matrix; x j (kh-h) represents the environment state of the (k-1)th sample in the j-th training round; σ(kh) represents the load mode at time kh; σ(kh-h) represents the load mode at time kh-h; h represents the sampling step size;

[0053] Note 1: In this embodiment, to ensure that the load switching unmanned vessel model (5) selects the optimal state for convergent switching behavior, the reward function punishes state divergence and inhibits the state from remaining at a non-zero value for a long time, thereby guiding the A2C algorithm to learn the optimal switching rule. Specifically, when σ(kh) ≠ σ(kh-h) indicates that the unmanned vessel mode has switched in the k-th step, ρ1 < 0. express The smaller the value, The higher the reward, the more likely the A2C algorithm, after training, will prioritize selecting the option that yields the highest reward at step k. The smallest mode; when σ(kh) = σ(kh-h), the unmanned surface vessel mode does not switch. Ensure that the state of the unmanned vessel model (5) during load switching continues to approach zero. Ensure that the state cannot remain non-zero for an extended period;

[0054] S23: Initialize the network parameters of the A2C algorithm And make the experience buffer in, This represents the initial value of the action to be executed in the j-th training round; This represents the initial values ​​of the parameters of the first Actor network; This indicates the last step in round j-1. The corresponding first Actor network parameters; This represents the initial values ​​of the parameters of the first Critic network; The last step in round j-1 The corresponding first Critic network parameters;

[0055] S24: Using the first Actor network parameters in the A2C algorithm Calculate and obtain the environment variables at step k during the j-th training round. The action score for action a is expressed as follows:

[0056]

[0057] In the formula: Indicates action score; Represents the weight matrix The row vector corresponding to action 'a' in the data; Represents the weight matrix; h1 and h2 represent the bias vectors; h1 and h2 represent intermediate variables; the number of neurons in the input layer, hidden layer, and output layer of the Actor network are 6, n, and L, respectively.

[0058] S25: Scoring based on action Calculate the probability distribution of action a And obtain the environment variable at step k based on the probability distribution. The cumulative probability of the corresponding action is expressed as follows:

[0059]

[0060] In the formula: L represents the number of neurons in the output layer of the first Actor network; q represents the set of load modes in the unmanned vessel load switching model. Elements in; P q,k Represents the environment variable at step k. The cumulative probability of the actions of the first q elements;

[0061] S26: Set a random number μ and μ∈[0,1], and the cumulative probability of the action satisfies P. q-1,k <μ<P q,k When selecting an action And obtain the environment variable at step k based on the reward function. By performing actions Rewards received Will be determined by action Environment variables and rewards Combinations Store to experience buffer

[0062] S27: Determine current environment variables Does the set condition meet? and Where n represents the empirical buffer. The maximum number of samples that can be stored; s represents the first termination environment variable set and k represents the number of iterations; % represents the modulo operator;

[0063] If the condition is met, proceed to step S29; if the condition is not met, proceed to step S28.

[0064] S28: Based on the current experience buffer Combined with the first Critic network parameters Construct the reward function, whose expression is:

[0065]

[0066] In the formula: Indicates the j-th round of training. The return value of step environmental variables; Indicate design parameters and This represents the value function of the first Critic network; Represents the weight matrix; h1 represents the bias vector of the first Critic network; h4 and h5 represent intermediate variables; l represents the number of neurons in the output layer of the first Critic network. This represents the first discount factor; the number of neurons in the input layer, hidden layer, and output layer of the first Critic network are 6 each. 1.

[0067] Based on the return function and the current experience buffer Update the first Actor network parameters With the first Critic network parameters Its expression is

[0068]

[0069] In the formula: Represents the parameters of the first Actor network. The update gradient; α j Let α represent the learning rate of the first Actor network and α represent the learning rate of the first Actor network. j =1 / (1+e) j+α)∈[0,1], α represents the learning constant of the first Actor network; Represents the parameters of the first Actor network. The vector formed by taking the partial derivatives of each parameter in the vector; Represents the parameters of the first Actor network. Update gradient; β j Let β represent the learning rate of the first Critic network. j =1 / (1+e) j+β )∈[0,1]; β represents the learning constant of the first Critic network; express The vector formed by taking the partial derivatives of each parameter in the vector; These represent the updated parameters of the first Actor network and the first Critic network, respectively.

[0070] Let the current experience buffer If the set is empty, determine the current environment variables. Is it a terminating environment variable? If not, then set... Then proceed to step S29; if so, proceed to step S30.

[0071] S29: Determine if k satisfies in, This indicates the maximum number of steps the system can run; if this is met, then repeat step S24; otherwise, continue with step S30.

[0072] S30: Update the training epoch j = j + 1, and determine whether the current training epoch meets the termination condition j > Z, where Z represents the maximum training epoch; if it meets the condition, then the training ends and the optimized parameters of the first Actor network are output. With the first Critic network parameters Otherwise, repeat step S23;

[0073] The generation phase of the switching rules is as follows: based on the optimized first Actor network, the optimal action for switching the load mode of the unmanned vessel is selected according to the environmental variables currently in which the unmanned vessel is located.

[0074] Specifically, this stage is based on the current environmental variables s k And based on the trained Actor network parameters Select action a k Step S001, let k = 1; Step S002, observe the current environmental variable s k And based on the Actor network parameters Calculate the action probability (8) to select the optimal action.

[0075]

[0076] In the formula: Indicates the parameters of the corresponding first Actor network. The cumulative probability of the action obtained; kh represents the k-th sampling time; kh+h represents the (k+1)-th sampling time;

[0077] Step S003: Output the optimal action a corresponding to the unmanned vessel's load mode σ(t). k Update k = k + 1; Step S004, if Then proceed to step S002; otherwise, the system operation will terminate.

[0078] Note 2: Actor network parameters Learning iteratively through the reward function (10), since the reward function (10) encompasses the reward function (6), the switching rule (13) after training is essentially a maximization of the reward function (6).

[0079] like Figure 3 As shown, in this embodiment, the j-th round training phase and the switching rule generation phase of the above A2C algorithm are summarized into the following Algorithm 1:

[0080]

[0081] To ensure that Algorithm 1 can stably approximate the optimal switching rule, its convergence needs to be theoretically analyzed: When the number of multi-objective scenes is set to 1, the multi-objective A2C algorithm is the A2C algorithm in this embodiment. Based on Theorem IV.1 and Theorem IV.2 recorded in the literature (Amidzadeh M. A Scale-Independent Multi-Objective Reinforcement Learning with Convergence Analysis [C]. 2023 62nd IEEE Conference on Decision and Control (CDC). IEEE, 2023: 1326-1333.), it can be known that during the iterative training process of the A2C algorithm, as the number of iterations j becomes sufficiently large, Converging to θ * , Converging to φ * In this embodiment, to analyze the estimation error limit of Algorithm 1 during the action selection process, a theoretical derivation of the action error caused by the initial values ​​of the Actor network parameters is required: Algorithm 1 updates the Actor network parameters iteratively. Select the optimal mode However, during the iterative learning process, the Actor network parameters... Starting the iteration with arbitrary initial values ​​will make the estimated... and optimal There are errors between them; therefore, error analysis is performed on Algorithm 1. It is necessary, as can be obtained from formula (13).

[0082]

[0083] because and All from finite action sets The selected actions have a maximum difference of L-1 between any two actions, therefore there exists a constant. Make This indicates and The error is bounded. From the convergence of Algorithm 1, we know that as the number of iterations j becomes sufficiently large, Converging to θ * And according to (8)-(9), the estimated action is The probability gradually increases, which means The value approaches 0. Therefore, it is proven that Algorithm 1 can stably approximate the optimal switching rule;

[0084] S3: Based on the unmanned vessel load mode corresponding to the optimal action, and according to the constructed first event triggering mechanism, the unmanned vessel load mode and status signal are transmitted to the remote land control station through a pre-set communication network.

[0085] Specifically, the output y(t) and load mode σ(t) in system (5) are transmitted to the remote land control station via network I to generate control signals. In this embodiment, a first event triggering mechanism is designed to improve communication efficiency and reduce network resource consumption;

[0086] The first event triggering mechanism The construction formula is

[0087]

[0088] 1) Error detection conditions are:

[0089]

[0090] That is, in Update the data regularly;

[0091] 2) The modal matching condition is:

[0092]

[0093] That is, in Update the data regularly;

[0094] In the formula: Indicates the time of the l-th trigger; Indicates in The vth sampling time thereafter; Indicates intermediate variables; Indicates state error and express Real-time status signals of the unmanned vessel; express Real-time status signals of the unmanned vessel; Indicates the parameters for state error detection; Indicates the set trigger threshold and Describes a weighted matrix and These represent the load mode matching condition parameters for the unmanned surface vessel. Represents a set of integers; express The load mode corresponding to the time; express The load mode corresponding to the time; h represents the sampling step size.

[0095] In this embodiment, the unmanned surface vessel uses event triggering (14) to send a status signal. and modality The data is transmitted over a network to a remote land controller, a network susceptible to transmission delays and aperiodic spoofing attacks. For example... Figure 2 As shown, the transmission delay in the sensor-controller channel and the controller-actuator channel is... The range is from arrive in set up and at the same time So make therefore

[0096] In addition, data packets transmitted in the network can also be affected by spoofing attacks. The activation interval and dormancy interval of the ιth spoofing attack are defined as follows: and During the attack, the mode σ(t) and state signal x(t) transmitted in the sensor-controller channel were tampered with. and Control output in the controller-actuator channel Attacked and tampered with The tampered state signal, modal signal, and control output respectively satisfy... G x and G uGiven a matrix of appropriate dimension. Therefore, in this embodiment, the next step is to detect whether the controller receives... and To determine whether the system has been deceived or tampered with, a state observer and an A2C algorithm modal estimation and detection mechanism were designed.

[0097] S4: By constructing a state observer and using the A2C algorithm modal estimation / detection mechanism, it confirms whether the data signal transmitted to the remote land control station has been tampered with by a deception attack, and then outputs the observation status of the unmanned vessel; and based on the constructed safety controller, it obtains the control signal of the unmanned vessel according to the observation status.

[0098] Specifically, the following steps are included:

[0099] S41: Construct a state observer, and the expression of the state observer is:

[0100]

[0101] In the formula: The observer state represents the unmanned vessel's state signal; p(t) represents the observation residual. Represents the observation matrix; This represents the payload modes that the remote land control station expects to receive from the unmanned surface vessel; Γ l Indicates time parameter quantity and This indicates the actual status signal of the unmanned vessel;

[0102] S42: Next, in order to estimate the load mode and detect whether the data transmitted by the unmanned vessel and the load mode have been tampered with by a spoofing attack, an A2C algorithm mode estimation / detection mechanism is constructed to confirm whether the data signal transmitted to the remote land control station by the state observer has been tampered with by a spoofing attack, and then output the observation state of the unmanned vessel. The construction method of the A2C algorithm mode estimation / detection mechanism includes: mechanism training phase and attack detection / mode selection phase.

[0103] The training phase of the mechanism specifically includes the following steps:

[0104] S100: Set the quintuple used during the mechanism training phase. in, Indicates the number of iterations and Indicates the maximum number of iterations. The details are as follows: First, Represents the set of modes and alarm signals of the state observer and

[0105] This indicates that the mode is estimated according to the A2C algorithm. No status signal was detected. With load mode The parameter under attack; This indicates that the mode is estimated according to the A2C algorithm. And only status signals were detected. Parameters that have been manipulated by deception attacks; This indicates that the mode is estimated according to the A2C algorithm. And load mode was detected With status signals All parameters were manipulated by deception attacks; This indicates that the mode is estimated according to the A2C algorithm. And only load mode was detected Parameters that have been manipulated by deception attacks; L represents the total number of action types; L represents the total number of load modes; secondly, Represents the set of environment variables and Indicates the first The environment variables sampled in the kth round of training; col represents a column vector; Indicates the first The observer state at step k in the training round; Indicates the first Modal residual norm at step k in round training Indicates the first The modal residual at step k in the training round; This indicates the load mode received by the controller; This represents the estimated mode based on A2C. The observation residual norm of the state observer (15) The residuals represent the observations of the state observer; in this implementation, to reduce computational complexity, the modal estimation residual norm of the unmanned surface vessel is used. Norm of the residuals of the state estimation of unmanned ships The modal residuals are divided into finite discrete intervals based on their numerical values. Mutually exclusive intervals in satisfy The state estimation residuals are divided into O mutually exclusive intervals. o∈O, where satisfy Therefore, we can deduce that the environment variable in step k is represented as follows: third, Representation Strategy The abbreviation form is for the first time. In the k-th round of training, the environment variables Using the parameters of the second Actor network Execute action The probability; fourth, The discount factor is a parameter that balances long-term and short-term returns. The larger the discount factor, the more long-term returns are pursued, and the smaller the discount factor, the more short-term returns are valued. Indicates the first On the k-th step of training environment variables Execute action The reward received later;

[0106] S101: In order to estimate accurate modal signals And capable of detecting and Whether it is under attack, therefore, a system is built to detect load patterns. With status signals Reward function for whether it is attacked Its expression is

[0107]

[0108] In the formula: c1 and c2 represent given constants and c1 > 0, c2 < 0; Indicates the first The observer state residual in the k-th step of the training round; Indicates the first The observer state at step k in the training round; Indicates the first The observer state at step k-1 in the training round; This represents the estimation of the modal asynchronous matrix; Indicates load mode The positive definite matrix under and Indicates load mode The positive definite matrix under the given conditions; Represents a symmetric matrix; This represents the known load modes of the unmanned vessel during training; This indicates the probability of a known attack occurring during training; This represents the estimated load mode at the k-th sampling time. This represents the estimated load mode at the (k-1)th sampling time.

[0109] S102: Initialize the network parameters of the A2C algorithm And make the experience buffer in, Indicates the first The initial values ​​for the actions performed in each round of training; This represents the initial values ​​of the parameters of the second Actor network; Indicates the first The parameters of the second Actor network corresponding to the last step k in the round; This represents the initial values ​​of the parameters of the second Critic network; No. The final step in the wheel The corresponding second Critic network parameters;

[0110] S103: Through the second Actor network parameters in the A2C algorithm Calculate and obtain the first During the training round, the environment variables at step k. Corresponding action The action score is expressed as follows:

[0111]

[0112] In the formula: Represents the weight matrix; Represents the bias vector; Represents the weight matrix Corresponding actions in The row vector; where the number of neurons in the input layer, hidden layer, and output layer of this Actor network are 8, 4L;

[0113] S104: Scoring based on action Calculate actions probability distribution And according to the probability distribution Get the environment variables in step k The cumulative probability of the corresponding action is expressed as follows:

[0114]

[0115] In the formula: 4L represents the number of neurons in the output layer of the second Actor network; Indicating the index of the action type and Represents the environment variable at step k. Corresponding to the front The cumulative probability of each element's action;

[0116] S105: Set random number and And the cumulative probability of the action satisfies When selecting an action And obtain the environment variable at step k based on the reward function. By performing actions Rewards received Will be determined by action Environment variables and rewards Combinations Store to experience buffer

[0117] S106: Determine current environment variables Does the set condition meet: k%? n / 3)≠0 and in n Represents the experience buffer The maximum number of samples that can be stored; s This indicates the second termination environment variable set and s =[0 8×1 ], k represents the number of iterations; % represents the modulo operator;

[0118] If the condition is met, proceed to step S108; if the condition is not met, proceed to step S107.

[0119] S107: Based on the current experience buffer Combined with the second Critic network parameters Construct the reward function, whose expression is:

[0120]

[0121]

[0122] In the formula: Indicates the first During the first round of training The return value of step environmental variables; Indicate design parameters and This represents the value function of the second Critic network; Indicates intermediate variables; This represents the weight matrix of the second Critic network; represents the bias vector of the second Critic network; l represents the number of neurons in the output layer of the second Critic network; This represents the second discount factor; where the number of neurons in the input layer, hidden layer, and output layer of the second Critic network are 8, respectively. 1; if but It is 0 if it is 0, otherwise it is 1.

[0123] According to the return function With the current experience buffer Update the second Actor network parameters With the second Critic network parameters Its expression is

[0124]

[0125] In the formula: Indicates the parameters of the second Actor network. The update gradient; Indicates the learning rate of the second Actor network and This represents the learning constant of the second Actor network; express Each variable pair The vector formed by taking partial derivatives; Indicates the parameters of the second Actor network. The update gradient; Indicates the learning rate of the second Critic network and This represents the learning constant of the second Critic network; express The vector formed by taking the partial derivatives of each parameter in the vector; These represent the updated parameters of the second Actor network and the second Critic network, respectively.

[0126] Let the current experience buffer If the set is empty, determine the current environment variables. Is it a terminating environment variable? If not, then set... If k = k + 1, proceed to step S108; otherwise, proceed to step S109.

[0127] S108: Determine if k satisfies the condition. in, This indicates the maximum number of steps the system can run; if this is met, then repeat step S103; otherwise, continue with step S109.

[0128] S109: Update training rounds And determine whether the current training round meets the termination condition. This indicates the maximum number of training epochs; if satisfied, training ends and the optimized parameters of the second Actor network are output. With the second Critic network parameters Otherwise, repeat step S102;

[0129] S110: Based on optimizing the parameters of the second Actor network Based on the environmental variables observed so far for unmanned surface vessels Observation status Obtain the action probability distribution to select the optimal action for the corresponding unmanned surface vessel mode estimation / detection. Its expression is

[0130]

[0131] In the formula: Indicates the parameters of the corresponding second Actor network. The cumulative probability of the acquired action;

[0132] In this embodiment, the A2C algorithm detection and mode estimation stage: This stage is based on the current observation state. Modal and State Estimation Residual Norm and Determine environment variables And based on the trained Actor network parameters Select Step S100, let k = 1; Step S200, observe the current environmental variables. And based on the trained Actor parameters Calculate the action probability (18) and select the optimal action. Step S300, Output Update k = k + 1; Step S400, if If so, proceed to step S200; otherwise, the system will terminate. Figure 4 As shown, this embodiment will use the above-mentioned A2C algorithm's first... The training phase and the attack detection and modality estimation phase are summarized in Algorithm 2 as follows:

[0133]

[0134]

[0135] To ensure that Algorithm 2 can stably approximate the detection and modal values, its convergence needs to be theoretically analyzed: that is, when the number of multi-target scenes is set to 1, the multi-target A2C algorithm is the same as the A2C algorithm in this example. Similarly, according to Theorem IV.1 and Theorem IV.2, it can be known that during the iteration process of the A2C algorithm, when... When large enough, converged to converged to This embodiment analyzes the estimation error bound of Algorithm 2 in the action selection process, and requires a theoretical derivation of the action error caused by the initial values ​​of the Actor network parameters: Algorithm 2 updates the Actor network parameters iteratively. Select the optimal action However, during the update process, the algorithm's Actor network parameters... Starting the iteration with arbitrary initial values ​​will make the estimation action... With optimal action There are errors between them. Therefore, error analysis is performed on Algorithm 2. It is necessary. From (23), we can conclude that:

[0136]

[0137] because and All are from a finite set of actions Among the selected actions, the maximum difference between any two actions is... Therefore, there exists a constant. Make This indicates and The error is bounded. From the convergence of Algorithm 2, we can see that as the number of iterations increases... Big enough convergence to And according to (18)-(19), the estimated action is The probability gradually increases, which means The value approaches 0. Therefore, Algorithm 2 can stably approximate the convergence of the detection and modal values, thus concluding the theoretical analysis of the convergence.

[0138] S111: Based on the optimal action Select modal signals and state information;

[0139] Define it as in This represents the estimated load mode signal; v x (t) represents the attack detection result of the state; v σ (t) represents the result of modality attack detection;

[0140] And according to Build and obtain a detection model for detecting deception attacks, and a state output model for determining the observed state;

[0141] And the expression for the detection model is:

[0142]

[0143] In the formula: This represents the output of the detection model; This indicates the load mode signal transmitted through the communication network; This represents the estimated load mode signal; Indicates the time when the l-th event is triggered; Indicates the delay of the l-th transmission; This indicates the time when the (l+1)th event is triggered; This represents the (l+1)th transmission delay; in this embodiment, Sending to the controller means the current mode signal When deceived, the A2C algorithm is used to estimate the modal signal. When no spoofing attack is detected, use data transmitted over the network. At the same time, based on the alarm value υ x =υ x (t), based on the state output model, can... With observer output Choose between the output options;

[0144] The expression for the state output model is:

[0145]

[0146] In the formula: This represents the output of the state output model; This represents the estimated state signal; The status signal is transmitted through the communication network; this means that when a deception attack is detected, the observation status is used by the observer (13). When no spoofing attack is detected, use data transmitted over the network.

[0147] S112: Construct a safety controller based on the detection model and the state output model, i.e., formulas (24) and (25), and obtain the control signal of the unmanned vessel based on the observed state according to the constructed safety controller;

[0148] The expression for the security controller is:

[0149]

[0150] In the formula: u(t) represents the output of the safety controller; This represents the gain matrix of the safety controller.

[0151] S5: When the remote land controller sends the control signal u(t) to the actuator, a second event triggering mechanism is constructed to save network resources. Based on the constructed second event triggering mechanism, the control signal is transmitted to the actuator through a preset communication network, thereby realizing safe dynamic positioning control of the unmanned vessel for load switching based on deception attack detection. Specifically, it includes the following steps:

[0152] S51: Construct a second event triggering mechanism, the expression of which is:

[0153]

[0154] In the formula: z(t) represents the trigger counter and z(t) = 1, which is specifically incremented by 1 after each trigger moment of the first event triggering mechanism; This indicates the time when the (l+1)th event is triggered; The l-th event is triggered at time t; u represents the abbreviation of u(t); q represents a sufficiently small scalar of the design.

[0155] S52: To reduce data security issues caused by exposure to the network, watermarking technology is introduced into the controller-actuator channel to encrypt the control signals transmitted to the actuator via a pre-set communication network, thereby obtaining encrypted control signals.

[0156] The formula for obtaining the encryption control signal is as follows:

[0157]

[0158] In the formula: Indicates encryption control signal; This represents an invertible matrix used for multiplication encryption; express Time-based control input;

[0159] S53: The control signal transmitted to the actuator is decrypted using watermarking technology. Modified to

[0160]

[0161] S54: Considering the transmission delay factor in the network between the unmanned vessel load switching model (5) and the safety controller (26), that is... Trigger data generated at any time Will The moment arrives at the actuator;

[0162] in, Indicates in The state signal x at time t; Indicates in The control signal u at time; This represents the state x or the control input u; Indicates the trigger time;

[0163] Under the second event triggering mechanism, by setting

[0164] Then it can divide the interval The form of partitioning into subsets and their union is:

[0165]

[0166]

[0167] In the formula: v represents the number of times the trigger is performed; Indicates in The vth sampling time thereafter; Indicates in The (v+1)th sampling time thereafter; Indicates in The next The sampling time; ∈ l Indicates the delay of the l-th transmission; Indicates the (l+1)th trigger time; ∈ l+1 This indicates the transmission delay of the (l+1)th transmission.

[0168] Let the parameter And τ(t) satisfies τ =∈ m , Where ∈ m Indicates the lower bound of the transmission delay; ∈ M Indicates the upper bound of the transmission delay; h represents the sampling step size;

[0169] Let the parameter in express Real-time status signals of the unmanned vessel; express Real-time status signals of the unmanned vessel;

[0170] S55: Combine the unmanned vessel load switching model with the state observer, i.e., the unmanned vessel (5) and the observer (15) to obtain the error system, the expression of which is:

[0171]

[0172] In the formula: e(t) represents the state error; Denote the first derivative of e(t);

[0173] S56: Modify the control signal in S53 Replace u(t) in the unmanned surface vessel (USV) load switching model and combine it with the error system to obtain the closed-loop system of the USV load switching system, whose expression is:

[0174]

[0175] In the formula: ξ(t) represents the state vector and ξ(t) = [x T (t) e T [(t)], initial condition ξ(t)=ξ0,

[0176] Indicates load mode and Denotes the intermediate parameter matrix, and

[0177]

[0178] Indicates load mode The system matrix below; Indicates load mode The observation matrix below; Indicates load mode The input matrix below; Indicates load mode The controller gain matrix under;

[0179] By acquiring the closed-loop system of the load-switching unmanned surface vessel, safe dynamic positioning control of the load-switching unmanned surface vessel based on deception attack detection can be achieved.

[0180] In this embodiment, the dynamic positioning problem of the load-switching unmanned vessel (5) is transformed into two control objectives of the closed-loop switching system (31) through the closed-loop system of the load-switching unmanned vessel:

[0181] (i) When w(t) = 0, the closed-loop switching system (31) is asymptotically stable.

[0182] (ii) When w(t)≠0, (31) has H∞ property, and for b>0 it has

[0183]

[0184] To achieve the above control objectives, it is necessary to analyze the subsystem modes σ(t) and controller modes in the closed-loop system (31). The switching behavior. Without loss of generality, for any two adjacent switching intervals [t] k-1 ,t k ) and [t k ,t k+1 Within t, assume the subsystem mode σ(t) is in t k The switch occurs from i to j, while the controller mode... Different modal signals will be adopted based on the detection results of the A2C algorithm. For [t] k ,t k+1 ), The controller switching time is calculated by (24); and These represent the controller switching times caused by the detection of the ιth spoofing start and end, respectively. The mode selected by the controller for the ιth attack, when detected hour, otherwise The switching behavior between the system and the controller is analyzed in detail below, and the results are given in Table 1. Details of the changes.

[0185] Case A: The secondary attack range did not cover the controller switching time. When alarm value At that time, the deception attack did not tamper with the modal signal σ(t), and At any time, based on what is sent from the network Select controller mode. When When a deception attack is detected that alters the modal signal σ(t), the controller mode is calculated according to Algorithm 2 and Formula (24). Timing controller mode Switch from i to When the attack is detected to have ended exist From time to time Switch to i. Therefore, in the interval superior, In the interval superior, In the interval superior In the interval and superior

[0186] Table 1. Three scenarios of closed-loop system switching behavior

[0187]

[0188] Case B: The moment when the coverage controller switches during the i-th attack interval. When alarm value When the deception attack was detected, the modal signal σ(t) was not tampered with. and At any time, based on what is sent from the network Select controller mode. When When a deception attack is detected that alters the modal signal σ(t), the controller mode is calculated according to Algorithm 2 and Formula (24). The moment will cause controller mode Switch from i to When the attack is detected to have ended exist From time to time Switch to i. Therefore, in the interval superior, In the interval superior, In the interval and superior

[0189] Case C: The attack interval of the i-th attack covers the entire interval [t] k,t k+1 When alarm value When the deception attack was detected, the modal signal σ(t) was not tampered with. The time controller mode switches from i to j. When... When a deception attack is detected that alters the modal signal σ(t), the controller mode is calculated according to Algorithm 2 and Formula (24). exist The time is switched from i to Therefore, in the interval superior, In the interval superior,

[0190] Theorem 1: Given a positive constant δ i <1, τ , λ and b>0, κ1>0, κ2<0, β1>1, 0<β2<1 where And i≠j, The control objective of the unmanned vessel closed-loop system (31) is achievable if:

[0191] (1) There exists a matrix Make

[0192]

[0193] Δ>0, (33)

[0194]

[0195]

[0196] and attack frequency λ f The maximum duration of the attack τ max ,satisfy

[0197] 0≤λ f ≤κ2 / (κ2-κ1), (36)

[0198] h≤τ max ≤-h ln(β2) / ln(β1), (37)

[0199] Among them, the 12×12 matrix in (32) It consists of the following matrices:

[0200]

[0201] The remaining matrix blocks are zero matrices of appropriate dimensions.

[0202] (2) There exists an indeterminate matrix of appropriate dimension. Make

[0203]

[0204] Among them, matrix It consists of the following matrices:

[0205]

[0206]

[0207] Proof: First, we prove that the system (30) is asymptotically stable when the control objective (i) is w(t) = 0. We construct the following Lyapunov function.

[0208]

[0209] in

[0210]

[0211] Differentiating (37) yields

[0212]

[0213] right Applying lemma and condition (33) to the integral term, we can obtain

[0214]

[0215] Where X T (t)=[ξ T (t- τ )-ξ T (t-τ(t)))

[0216] Then, and Substitution able to obtain Substitute it with (42) and (43) Then, by using (32) and (34), we can obtain

[0217]

[0218] Where Γ(t)=b 2 w T (t)w(t)-y T (t)y(t),

[0219]

[0220] Next, based on the situation, the system mode σ(t) and controller mode are analyzed using AC. The changes in system (31) are discussed in any switching interval [t]. k ,t k+1 Changes within )

[0221] Scenario A: In Integrating over (44) within the interval yields...

[0222]

[0223] The controller switching moment when the end of the spoofing attack is detected. Depend on And (35) obtained

[0224]

[0225] For detecting the first Secondary attack zone Integrating both sides of (44) yields

[0226]

[0227] The controller switching moment when the spoofing attack is detected to have started. Depend on And (34) obtained

[0228]

[0229] For attacking the dormant zone Integrating both sides of (44) yields

[0230]

[0231] For the moment of controller switching Depend on get

[0232]

[0233] In the interval Integrating both sides of (44) yields

[0234]

[0235] Combining (45) and (51), we can deduce that in case A, (41) in [t k ,t k+1 The evolution of ) is as follows:

[0236]

[0237] in,

[0238] Reuse (39) to obtain Derivation:

[0239]

[0240] Scenario B: In Within the interval, the integral of (44) yields...

[0241]

[0242] The switching moment when the end of the spoofing attack is detected. Depend on And (35) obtained

[0243]

[0244] The controller switching moment when the spoofing attack is detected to have started. Depend on And (34) obtained

[0245]

[0246] For attacking the dormant zone Integrating both sides of (44) yields

[0247]

[0248] Using (32) in the interval Integral, to obtain

[0249]

[0250] The controller switching moment when the spoofing attack is detected to have started. Depend on And (35) can be obtained

[0251]

[0252] By combining (54) and (59), we can deduce that in case B, (41) in [t k ,t k+1 The evolution of ) is as follows:

[0253]

[0254] in, From (40), we can obtain It can be deduced that:

[0255]

[0256] Case C: In Within the interval, according to (32), we can obtain

[0257]

[0258] Controller switching moment when a spoofing attack is detected use And (35) can be obtained

[0259]

[0260] For interval Integrating (40) yields

[0261]

[0262] By combining (62) and (64), we can deduce that in case C, (41) in [t k ,t k+1 The evolution of ) is as follows:

[0263]

[0264] Therefore, for case AC, (41) in any interval [t] k ,t k+1 The changes can be summarized as follows:

[0265]

[0266] in,

[0267] Subsequently, the subsystem switching point t was discussed. k The preceding and following (33) can be obtained

[0268]

[0269] From (65)-(67), we can obtain

[0270]

[0271] Next, we will discuss the derivation of the relationship between V(t) and V(t0) over the entire operating interval [0,t). Using (39) and (68), we can obtain the result recursively.

[0272]

[0273] Where, N p (0,t) represents the number of times an attack was detected, which satisfies λ. f t / τ max ≤N p (0,t)≤λf t / h; From (36)-(37), we can obtain κ1λ f +κ2(1-λ f )<0 and ln(β1) / h+ln(β2) / τ max <0 and summarize (69) as

[0274]

[0275] Thus, when w(t) = 0, due to the properties of (32) and positive definite matrices, it satisfies and It is possible to draw

[0276]

[0277] in,

[0278] Therefore, the system is asymptotically stable, and the control objective (i) is achieved.

[0279] Next, we will demonstrate the control objective (ii):

[0280] When w(t)≠0, the evolution relationship from V(t) to V(t0) can be derived from (71) as follows:

[0281]

[0282] Therefore, under zero initial conditions

[0283]

[0284] This means that when w(t)≠0, control objective (ii) is achieved, Q.E.D.

[0285] Below, we will discuss Theorem 1 (32). and (38) The nonlinear terms are linearized.

[0286] Theorem 2: Given positive constants τ, λ and And i≠j, If the conditions in Theorem 1 hold, and there exists a matrix as well as satisfy

[0287]

[0288] Therefore, the closed-loop system is asymptotically stable and has H∞ performance, where the 14×14 matrix in (72) is... It consists of the following matrices

[0289]

[0290] The other matrix blocks are zero matrices of appropriate dimension.

[0291] The 7×7 matrix in formula (76) It consists of the following matrices:

[0292]

[0293] and The remaining matrix blocks are zero matrices of appropriate dimension.

[0294] Proof: Multiply both sides of (72) by the matrix respectively. And its transpose, and then using Theorem 1, we can obtain (32) holds, where Similarly, multiply the left and right sides of (76) by the matrix respectively. Its transpose, can yield (38) which holds true, where Therefore, according to Theorem 1, it can be proved that the control objectives (i) and (ii) of the closed-loop system (31) are valid.

[0295] Application Example: This embodiment uses a load-switching unmanned surface vessel (USV) example to verify the effectiveness of the proposed A2C algorithm's optimal switching rule and the A2C algorithm-based spoofing attack detection mechanism. Using the USV parameters, matrix M can be calculated. σ(t) A σ(t) B σ(t) C σ(t) Assume the unmanned vessel disturbance w(t) = 0.0025 × col{-0.01cos(ψ(t)-π / 6),0.1sin(ψ(t)-π / 6),sin(2ψ(t)-π / 6)}.

[0296] The input parameters for Algorithm 1 are: discount factor ε = 0.99; training iterations j = 500; action set In the reward function (6), ρ1 = -0.05, ρ2 = -0.0002, and ρ3 = -10. -6 Maximum number of samples to store The Actor network has 64 hidden layer neurons (n ​​= 64); the Actor network has a learning rate of α = 0.00001; the Critic network has [missing information - likely a number of neurons in its hidden layer]. The learning rate β of the Critic network is 0.0001.

[0297] The input parameter for Algorithm 2 is: discount factor Action Set In the reward function (6) ρ3=-10 -6 c1 = 0.5, c2 = -0.5; store the maximum number of samples. n =20; Number of neurons in the Actor network The learning rate of the Actor network Number of neurons in the hidden layer of the Critic network Learning rate of Critic network The estimated state residual norm is divided into: Modal estimation residual norms are divided into: Given constants δ1 = 0.2, δ2 = 0.1, τ =0.001, h = 0.02s, λ = 0.002, g 12 =-0.3, g 13 =-0.3, κ1=4, κ2=-1, β1=1.0618, β2=0.001. Solving (36) and (37) in Theorem 1, we can obtain that the attack frequency satisfies 0≤λ. f ≤0.2, the maximum duration of the attack satisfies τ max = 2.3s. Solving Theorem 2 yields the controller matrix and observation matrix as follows:

[0298]

[0299]

[0300] Table 2 shows the performance comparison results of the A2C switching rule and the traditional state-dependent switching rule in this embodiment, and Table 3 shows the attack detection results of the A2C algorithm and the switching Q learning algorithm.

[0301] Table 2. Performance Comparison of A2C Switching Rules and State-Dependent Switching Rules (T=80s)

[0302]

[0303] Table 3. Attack detection results of A2C algorithm and switching Q-learning algorithm (T=60s)

[0304]

[0305] like Figure 2The diagram illustrates the structure of the unmanned surface vessel (USV) undergoing load switching. On one hand, an optimal switching rule based on the A2C algorithm is designed, defining the USV's modes as the action space and incorporating a reward function that includes state convergence factors. Through iterative learning, the A2C algorithm selects the optimal load mode based on the current USV state information, achieving efficient dynamic positioning for load switching. On the other hand, a defense strategy based on deception attack detection is designed. First, a mode estimation and detection mechanism based on the A2C algorithm is proposed. This mechanism incorporates reward functions that include observer state convergence and alarm information accuracy. During iterative learning, it accurately estimates the USV's load modes and detects whether mode and state information have been tampered with, effectively reducing errors in load mode estimation. A security controller is designed based on the attack detection results, and watermark encryption technology is applied to the control signals transmitted back to the USV to resist tampering.

[0306] The simulation portion of the method described in this embodiment considers the safe dynamic positioning of an unmanned surface vessel undergoing load switching. For example... Figure 5 Displayed position error rr*, heading angle Surge speed Swing speed The heading angular velocity γ curve approaches 0 in the initial state x(0)=col{2,3,0.3,0.3,0.2,0.3}. Figure 6 The optimal unmanned surface vessel mode σ(t) generated based on A2C and the controller-side mode are shown under a deception attack. from Figure 6 It can be seen that, due to the introduction of the attack detection results of Algorithm 2, even if the unmanned vessel mode is attacked within 16s-18s, 21s-23.5s, and 31s-33.5s, the correct controller mode can still be selected using formula (16), and the green shading in the figure represents (α). x ,α σ (1,0) indicates that the status and control signals sent by the unmanned vessel have been tampered with by a deception attack. The pink shading represents (α). x ,α σ ) = (1,1) indicates that the state, mode, and control signals transmitted by the unmanned vessel have all been tampered with by a deception attack. The blue shading represents (α). x ,α σ The expression () = (0,1) indicates that the modalities and control signals transmitted by the unmanned surface vessel have been tampered with by a deception attack. Further comparison is made of the dynamic response performance of state-dependent switching rules and switching rules generated by the A2C algorithm. For example... Figure 7 The image shows the unmanned surface vessel modes under two switching rules. Figure 8The results show that, under the same parameters, the state-dependent switching rule requires approximately 75 seconds for the entire unmanned vessel to reach a steady state, while the A2C switching rule only requires 55 seconds, a reduction of 26.7%. This indicates that the A2C algorithm optimizes the switching timing decision through learning, significantly improving the speed of dynamic positioning. Secondly, comparative analysis of the data in Table 2 shows that there are significant differences in the settling time and peak performance between the A2C switching rule and the state-dependent switching rule in various unmanned vessel states. Taking position error as an example, the adjustment time for the A2C switching rule is 14.87s, with a peak value of 2.03m, while the adjustment time for the state-dependent switching rule is 15.71s, with a peak value of 2.04m; in rr * Under positional error, the A2C handover rule adjustment time is 49.87s, while the state-dependent handover rule takes as long as 69.86s. Although both have a peak value of 3.01m, the A2C handover rule exhibits a faster response speed. Further observation of the heading angle ψ and surge velocity... Swing speed With state data such as heading angular velocity γ, the adjustment time for A2C switching rules is shorter, and in some states ψ, Mid-peak control is superior. Considering all indicators, the dynamic performance of the state-dependent switching rule is generally inferior to that of the A2C switching rule in terms of settling time and peak control, thus indicating that the A2C switching rule constructed in this embodiment has a greater advantage in the dynamic response of unmanned vessel dynamic positioning.

[0307] Next, the attack detection results of the A2C algorithm and the switching Q-learning algorithm are compared. Specifically, this invention divides the observer residual state in Algorithm 2, which detects spoofing attacks, into... Modal residuals are divided into Similarly, to ensure the accuracy of the comparison, the Q-learning algorithm is switched to divide the observer residual state into... Modal residuals are divided into The Q-learning algorithm was trained for 500 rounds. The comparison results are shown in Table 3. It can be seen that when the unmanned vessel suffers a (1,1) type attack, that is, when both the modality and state are attacked and tampered with, the detection accuracy of the A2C algorithm is significantly higher than that of the Q-learning algorithm. For other types of attacks, the detection accuracy and false alarm rate of the A2C algorithm proposed in this invention are also better than those of the Q-learning algorithm. Figure 9 Compare the detection results α of Algorithm 3.2 and the Switching Q-Learning Algorithm respectively. x (t) and αx(t). In Figure 10 In this embodiment, the watermarking technology effectively eliminates the fluctuations in control input u(t) caused by deception attacks during the operation of the unmanned vessel.

[0308] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A secure dynamic positioning control method for unmanned surface vessels with load switching based on deception attack detection, characterized in that, Specifically, the following steps are included: S1: Construct a uniformly segmented unmanned vessel load switching model; S2: Construct switching rules based on the A2C algorithm according to the unmanned vessel load switching model; The method for constructing switching rules based on the A2C algorithm includes: a training phase for switching rules and a generation phase for switching rules; The training phase of the switching rules: that is, constructing a reward function based on the unmanned vessel load switching model, and training the network parameters of the Actor network based on the reward function to obtain the optimized Actor network; The generation phase of switching rules: that is, based on the optimized Actor network, the optimal action for switching the load mode of the unmanned vessel is selected according to the environmental variables of the current unmanned vessel. S3: Based on the unmanned vessel load mode corresponding to the optimal action, and according to the constructed first event triggering mechanism, the unmanned vessel load mode and status signal are transmitted to the remote land control station through a pre-set communication network. S4: By constructing a state observer and using the A2C algorithm modal estimation / detection mechanism, it confirms whether the data signal transmitted to the remote land control station has been tampered with by a deception attack, and then outputs the observation status of the unmanned vessel; and based on the constructed safety controller, it obtains the control signal of the unmanned vessel according to the observation status. S5: Based on the constructed second event triggering mechanism, the control signal is transmitted to the actuator through the preset communication network, thereby realizing safe dynamic positioning control of the unmanned vessel for load switching based on deception attack detection.

2. The safe dynamic positioning control method for a load-switching unmanned surface vessel based on deception attack detection as described in claim 1, characterized in that, The method for constructing the uniformly segmented unmanned surface vessel load switching model in S1 specifically includes: S11: Obtain the unmanned vessel model, its expression is: In the formula: η(t) represents the position vector and r(t) represents the lateral and longitudinal positions; ψ(t) represents the heading angle; v(t) represents the velocity vector and Indicates oscillation velocity, Let γ(t) represent the sway velocity, γ(t) represent the heading angular velocity, and u(t) represent the control input vector, where u(t) = [u1(t) u2(t) u3(t)]. T u1(t), u2(t), u3(t) represent the sway force, roll force, and bow moment, respectively; w(t) represents the disturbance vector describing the continuous impact of wind, waves, and ocean currents on the unmanned vessel in the marine environment; M(m(t)), N, and R(ψ) represent the inertia matrix, damping matrix, and coordinate transformation matrix, respectively. Indicates time-varying load, Indicates the maximum load weight. m Indicates the unloaded weight; Indicates the additional hydrodynamic load, x g I represents the longitudinal component of the difference vector from the origin to the centroid in the preset coordinate system. Z Represents the inertia matrix. Y γ N v N γ Indicates the linear damping coefficient; S12: Set the dynamic positioning control target for the unmanned vessel: η(t) → η * v(t)→0, where η * Describe the desired position vector and Represents the desired position coordinates for dynamic positioning; S13: Transfer time-varying load Dividing it into L equal parts, then any load interval segment Can be represented as In the formula: θ q Indicates intermediate parameters and q represents the interval index; and when When m(t) is considered as m +θ q-1 And denote the interval index q as the subsystem mode, i.e., σ(t) = q. At this time, the inertia matrix M(m(t)) in the unmanned ship model can be written as: And according to M σ(t) Able to define intermediate variable A σ(t) B σ(t) C σ(t) for: Let the position vector x(t) = [η T (t)-η *T v T (t)] T Therefore, the uniformly segmented unmanned vessel load switching model is constructed as follows: In the formula: T represents the transpose; σ(t) represents the load mode of the unmanned vessel; D σ(t) y(t) represents the output matrix corresponding to the load mode; y(t) represents the measurement output of the unmanned vessel's dynamic positioning. Let x(t) represent the first derivative of x(t); x(t) represents the state signal of the unmanned vessel.

3. The safe dynamic positioning control method for unmanned surface vessels with load switching based on deception attack detection according to claim 2, characterized in that, The training phase of the switching rules in S2 specifically includes the following steps: S21: Define the quintuple used for training the A2C algorithm. in, Represents the set of environment variables and Indicates the first The environment state x(kh) sampled in the kth round of training. Represents an action set and 'a' represents the set of load modes in the unmanned vessel load switching model. The value obtained is the action; For strategy The abbreviation of indicates the first Environment variables in step k of round training Using the parameters of the first Actor network The probability of performing action 'a'; ε represents the discount factor used to balance long-term and short-term returns, and ε∈[0,1]. Let be the reward function, representing the th Environment variables in the k-th step of the training round Execute action The reward received later; S22: Construct a reward function for model system stabilization based on the unmanned vessel load switching model; its expression is as follows: In the formula: ρ1, ρ2, ρ3 represent design parameters and ρ1 < 0, ρ2 < 0, ρ3 < 0; Represents an asynchronous matrix; Describes the symmetric matrix under load mode σ(kh) and This represents a symmetric matrix under load mode σ(kh-h); Represents a symmetric matrix; Indicates the first The environment state sampled in the (k-1)th round of training; σ(kh) represents the load mode at time kh; σ(kh-h) represents the load mode at time kh-h; h represents the sampling step size; S23: Initialize the network parameters of the A2C algorithm k = 1, And make the experience buffer in, Indicates the first The initial values ​​for the actions performed in each round of training; This represents the initial values ​​of the parameters of the first Actor network; Indicates the first The final step in the wheel The corresponding first Actor network parameters; This represents the initial values ​​of the parameters of the first Critic network; No. The final step in the wheel The corresponding first Critic network parameters; S24: Using the first Actor network parameters in the A2C algorithm Calculate and obtain the first During the training round, the environment variables at step k. The action score for action a is expressed as follows: In the formula: Indicates action score; Represents the weight matrix The row vector corresponding to action 'a' in the data; Represents the weight matrix; The bias vector is represented by h1 and h2, which represent intermediate variables. S25: Scoring based on action Calculate the probability distribution of action a And obtain the environment variable at step k based on the probability distribution. The cumulative probability of the corresponding action is expressed as follows: In the formula: L represents the number of neurons in the output layer of the first Actor network; q represents the set of load modes in the unmanned vessel load switching model. Elements in; P q,k Represents the environment variable at step k. The cumulative probability of the actions of the first q elements; S26: Set a random number μ and μ∈[0,1], and the cumulative probability of the action satisfies P. q-1,k <μ<P q,k When selecting an action And obtain the environment variable at step k based on the reward function. By performing actions Rewards received Will be determined by action Environment variables and rewards Combinations Store to experience buffer S27: Determine current environment variables Does the set condition meet? and in Represents the experience buffer The maximum number of samples that can be stored; Indicates the first termination environment variable set and k represents the number of iterations; % represents the modulo operator; If the condition is met, proceed to step S29; if the condition is not met, proceed to step S28. S28: Based on the current experience buffer Combined with the first Critic network parameters Construct the reward function, whose expression is: In the formula: Indicates the first During the first round of training The return value of step environmental variables; Indicate design parameters and h4 and h5 represent the value function of the first Critic network; h4 and h5 represent intermediate variables. This represents the weight matrix of the first Critic network; represents the bias vector of the first Critic network; l represents the number of neurons in the output layer of the first Critic network. Indicates the first discount factor; Based on the return function and the current experience buffer Update the first Actor network parameters With the first Critic network parameters Its expression is In the formula: Represents the parameters of the first Actor network. The update gradient; α j This represents the learning rate of the first Actor network and α represents the learning constant of the first Actor network; Represents the parameters of the first Actor network. The vector formed by taking the partial derivatives of each parameter in the vector; Represents the parameters of the first Actor network. The update gradient; This represents the learning rate of the first Critic network and β represents the learning constant of the first Critic network; express The vector formed by taking the partial derivatives of each parameter in the vector; These represent the updated parameters of the first Actor network and the first Critic network, respectively. Let the current experience buffer If the set is empty, check the current environment variables. Is it a termination environment variable? If not, set k = k + 1 and continue to step S29; if yes, continue to step S30. S29: Determine if k satisfies in, This indicates the maximum number of steps the system can run; if this is met, then repeat step S24; otherwise, continue with step S30. S30: Update training rounds And determine whether the current training round meets the termination condition. If the conditions are met, the optimized parameters of the first Actor network are output after training. With the first Critic network parameters Otherwise, repeat step S23.

4. The safe dynamic positioning control method for unmanned surface vessels with load switching based on deception attack detection according to claim 3, characterized in that, In S2, select the optimal action a corresponding to the switching of the unmanned vessel's load mode σ(t). k The expression is In the formula: Indicates the parameters of the corresponding first Actor network. The cumulative probability of the action obtained; kh represents the kth sampling time; kh+h represents the (k+1)th sampling time.

5. A safe dynamic positioning control method for unmanned surface vessels with load switching based on deception attack detection, as described in claim 4, is characterized in that... The first event triggering mechanism in S3 The construction formula is In the formula: Indicates the time of the l-th trigger; Indicates in The vth sampling time thereafter; Indicates intermediate variables; Indicates state error and express Real-time status signals of the unmanned vessel; express Real-time status signals of the unmanned vessel; Indicates the parameters for state error detection; Indicates the set trigger threshold and Describes a weighted matrix and These represent the load mode matching condition parameters for the unmanned surface vessel. Represents a set of integers; express The load mode corresponding to the time; express The load mode corresponding to the time; h represents the sampling step size.

6. A secure dynamic positioning control method for a load-switching unmanned surface vessel based on deception attack detection, as described in claim 4, is characterized in that... S4 specifically includes the following steps: S41: Construct a state observer, and the expression of the state observer is: In the formula: The observer state represents the unmanned vessel's state signal; p(t) represents the observation residual. Represents the observation matrix; This represents the payload modes that the remote land control station expects to receive from the unmanned surface vessel; Γ l Indicates a time parameter; This indicates the actual status signal of the unmanned vessel; S42: Construct an A2C algorithm mode estimation / detection mechanism to confirm whether the data signal transmitted to the remote land control station through the state observer has been tampered with by a deception attack, and then output the observation state of the unmanned vessel. The construction method of the A2C algorithm mode estimation / detection mechanism includes: mechanism training phase and attack detection / mode selection phase. The training phase of the mechanism specifically includes the following steps: S100: Set the quintuple used during the mechanism training phase. in, Indicates the number of iterations; Represents the set of modes and alarm signals of the state observer and This indicates that the mode is estimated according to the A2C algorithm. No status signal was detected. With load mode The parameter under attack; This indicates that the mode is estimated according to the A2C algorithm. And only status signals were detected. Parameters that have been manipulated by deception attacks; This indicates that the mode is estimated according to the A2C algorithm. And load mode was detected With status signals All parameters were manipulated by deception attacks; This indicates that the mode is estimated according to the A2C algorithm. And only load mode was detected Parameters that have been manipulated by deception attacks; L represents the total number of action types; L represents the total number of modalities. Represents the set of environment variables and Indicates the first The environment variables sampled in the kth round of training; col represents a column vector; Indicates the first The observer state at step k in the training round; Indicates the first Modal residual norm at step k in round training Indicates the first The modal residual at step k in the training round; This indicates the load mode received by the controller; Indicates the estimated load mode; The observation residual norm of the state observer This represents the observation residuals of the state observer; Representation Strategy The abbreviation form is for the first time. In the k-th round of training, the environment variables Using the parameters of the second Actor network Execute action The probability of; Discount factor; Indicates the first On the k-th step of training environment variables Execute action The reward received later; S101: Constructing a system for detecting load modes With status signals The reward function for determining whether an attack has occurred is expressed as follows: In the formula: c1 and c2 represent given constants and c1>0, c2<0; Indicates the first The observer state residual in the k-th step of the training round; Indicates the first The observer state at step k in the training round; Indicates the first The observer state at step k-1 in the training round; This represents the estimation of the modal asynchronous matrix; Indicates load mode The positive definite matrix under and Indicates load mode The positive definite matrix under the given conditions; Represents a symmetric matrix; This represents the known load modes of the unmanned vessel during training; This indicates the probability of a known attack occurring during training; This represents the estimated load mode at the k-th sampling time. This represents the estimated load mode at the (k-1)th sampling time. S102: Initialize the second network parameters for the A2C algorithm And make the experience buffer in, Indicates (the first) (Initial values ​​for the actions performed in each round of training); This represents the initial values ​​of the parameters of the second Actor network; Indicates the first The final step in the wheel The corresponding second Actor network parameters; This represents the initial values ​​of the parameters of the second Critic network; No. The final step in the wheel The corresponding second Critic network parameters; S103: Through the second Actor network parameters in the A2C algorithm Calculate and obtain the first During the training round, the environment variables at step k. Corresponding action The action score is expressed as follows: In the formula: Represents the weight matrix; Represents the bias vector; Represents the weight matrix Corresponding actions in The row vector; S104: Scoring based on action Calculate actions probability distribution And according to the probability distribution Get the environment variables in step k The cumulative probability of the corresponding action is expressed as follows: In the formula: 4L represents the number of neurons in the output layer of the second Actor network; Indicating the index of the action type and Represents the environment variable at step k. Corresponding to the front The cumulative probability of each element's action; S105: Set random number and And the cumulative probability of the action satisfies When selecting an action And obtain the environment variable at step k based on the reward function. By performing actions Rewards received Will be determined by action Environment variables and rewards Combinations Store to experience buffer S106: Determine current environment variables Does the set condition meet: k%? n / 3)≠0 and in n Represents the experience buffer The maximum number of samples that can be stored; s This indicates the second termination environment variable set and k represents the number of iterations; % represents the modulo operator; If the condition is met, proceed to step S108; if the condition is not met, proceed to step S107. S107: Based on the current experience buffer Combined with the second Critic network parameters Construct the reward function, whose expression is: In the formula: Indicates the first During the first round of training The return value of step environmental variables; Indicate design parameters and This represents the value function of the second Critic network; Indicates intermediate variables; This represents the weight matrix of the second Critic network; represents the bias vector of the second Critic network; l represents the number of neurons in the output layer of the second Critic network; Indicates the second discount factor; According to the return function With the current experience buffer Update the second Actor network parameters With the second Critic network parameters Its expression is In the formula: Indicates the parameters of the second Actor network. The update gradient; Indicates the learning rate of the second Actor network and This represents the learning constant of the second Actor network; express Each variable pair The vector formed by taking partial derivatives; Indicates the parameters of the second Actor network. The update gradient; Indicates the learning rate of the second Critic network and This represents the learning constant of the second Critic network; express The vector formed by taking the partial derivatives of each parameter in the vector; These represent the updated parameters of the second Actor network and the second Critic network, respectively. Let the current experience buffer If the set is empty, check the current environment variables. Is it a termination environment variable? If not, set k = k + 1 and continue to step S108; if yes, continue to step S109. S108: Determine if k satisfies the condition. in, This indicates the maximum number of steps the system can run; if this is met, then repeat step S103; otherwise, continue with step S109. S109: Update training rounds And determine whether the current training round meets the termination condition. If the conditions are met, the optimized parameters of the second Actor network are output after training. With the second Critic network parameters Otherwise, repeat step S102; S110: Based on optimizing the parameters of the second Actor network Based on the environmental variables observed so far for unmanned surface vessels Obtain the action probability distribution to select the optimal action for the corresponding unmanned surface vessel mode estimation / detection. Its expression is In the formula: Indicates the parameters of the corresponding second Actor network. The cumulative probability of the acquired action; S111: Based on the optimal action Select modal signals and state information; Define it as in Represents the estimated load mode signal; υ x (t) represents the attack detection result of the state; υ σ (t) represents the result of modality attack detection; And according to Build and obtain a detection model for detecting deception attacks, and a state output model for determining the observed state; And the expression for the detection model is: In the formula: This represents the output of the detection model; This indicates the load mode signal transmitted through the communication network; This represents the estimated load mode signal; Indicates the time of the l-th trigger; Indicates the delay of the l-th transmission; This indicates the (l+1)th trigger time; This indicates the transmission delay of the (l+1)th transmission. The expression for the state output model is: In the formula: This represents the output of the state output model; This represents the estimated state signal; Status signals transmitted through communication networks; S112: Construct a safety controller based on the detection model and the state output model, and obtain the control signal of the unmanned vessel based on the observed state according to the constructed safety controller; The expression for the security controller is: In the formula: u(t) represents the output of the safety controller; This represents the gain matrix of the safety controller.

7. A secure dynamic positioning control method for a load-switching unmanned surface vessel based on deception attack detection, as described in claim 6, is characterized in that... S5 specifically includes the following steps: S51: Construct a second event triggering mechanism, the expression of which is: In the formula: z(t) represents the trigger counter and z(t) = 1, This indicates the (l+1)th event trigger time for the control input u; The l-th event trigger time is indicated by the control input u; u represents the abbreviation of u(t); q represents the scalar of the design. S52: By introducing watermarking technology into the controller-actuator channel, the control signal transmitted to the actuator through the preset communication network is encrypted with watermark to obtain the encrypted control signal; The formula for obtaining the encryption control signal is as follows: In the formula: Indicates encryption control signal; This represents an invertible matrix used for multiplication encryption; express Time-based control input; S53: The control signal transmitted to the actuator is decrypted using watermarking technology. Modified to S54: Considering the transmission delay factor in the network between the unmanned vessel load switching model and the safety controller, that is... Trigger data generated at any time Will The moment arrives at the actuator; in, Indicates in The state signal x at time t; Indicates in The control signal u at time; The general term representing state x or control input u; Indicates the trigger time; And by setting Then the interval The form of partitioning into subsets and their union is: In the formula: v represents the number of times the trigger is performed; This indicates the time of the vth trigger. This indicates the (v+1)th trigger time; Indicates the first Next trigger time; ∈ l Indicates the delay of the l-th transmission; Indicates the (l+1)th trigger time; ∈ l+1 This indicates the transmission delay of the (l+1)th transmission. Let the parameter And τ(t) satisfies τ =∈ m , Where ∈ m Indicates the lower bound of the transmission delay; ∈ M Indicates the upper bound of the transmission delay; h represents the sampling step size; Let the parameter in express Real-time status signals of the unmanned vessel; express Real-time status signals of the unmanned vessel; S55: By combining the unmanned surface vessel load switching model with the state observer, the error system is obtained, and its expression is: In the formula: e(t) represents the state error; Denote the first derivative of e(t); S56: Modify the control signal in S53 Replace u(t) in the unmanned surface vessel (USV) load switching model and combine it with the error system to obtain the closed-loop system of the USV load switching system, whose expression is: In the formula: ξ(t) represents the state vector and ξ(t) = [x T (t) e T [(t)], initial condition ξ(t)=ξ0, Represents the load modal index and Denotes the intermediate parameter matrix, and Indicates load mode The system matrix below; Indicates load mode The observation matrix below; Indicates load mode The input matrix below; Indicates load mode The controller gain matrix under; By acquiring the closed-loop system of the load-switching unmanned surface vessel, safe dynamic positioning control of the load-switching unmanned surface vessel based on deception attack detection can be achieved.

Images