A bilateral teleoperation detection and drive elasticity control method, system and device

Abstract

This application discloses a bilateral teleoperation detection-driven elastic control method, system, and device, relating to the field of bilateral teleoperation control. The method includes: acquiring motion and interaction signals from the master and slave ends of the bilateral teleoperation system to construct detection features; calculating the attack probability through a detector and generating a reliable gating quantity corresponding to the communication direction; updating the reliable state register and Kalman filter under normal conditions; freezing the reliable state and performing non-recursive reconstruction under attack conditions to obtain the compensated position and compensated velocity; constructing the master-slave control error; and outputting master-slave control commands through a bounded bilateral controller. The non-recursive reconstruction under attack conditions uses the frozen nearest reliable position as a reference, combines the Kalman predicted velocity and the nearest reliable velocity to generate a position extrapolation auxiliary velocity for position extrapolation, and maintains the nearest reliable velocity as the compensation velocity. This avoids the destruction of the control closed loop by distorted data and provides elastic control with deterministic error bounds.

Description

Technical Field

[0001] This application relates to the field of bilateral teleoperation control, and in particular to a bilateral teleoperation detection-driven elastic control method, system and equipment. Background Technology

[0002] Bidirectional teleoperation systems (PTS) enable remote operation by transmitting status information bidirectionally between the master and slave ends via a communication network. However, in practical applications, they often face threats such as time-varying communication latency and additive spoofing attacks. In existing security control schemes, attack detection and resilient control are often separated. Detectors can only output alarm signals, lacking effective data eligibility assessment and blocking mechanisms. This allows corrupted data packets, though formatted correctly, to directly enter the controller, compromising the system's synchronization stability. Furthermore, in the data compensation phase after attack detection, existing technologies typically employ a traditional recursive compensation method, using the compensation output from the previous moment as the compensation benchmark for the current moment. This recursive mechanism leads to the continuous accumulation and amplification of reconstruction errors over time during the attack's duration, failing to provide bounded error guarantees and reliable resilient control for PTS. Summary of the Invention

[0003] The purpose of this application is to provide a bilateral teleoperation detection-driven elastic control method, system, and device to solve the problems of disrupting the synchronization stability of the system and the inability to provide bounded error guarantees and reliable elastic control for the bilateral teleoperation system.

[0004] To achieve the above objectives, this application provides the following solution.

[0005] In a first aspect, this application provides a bilateral teleoperation detection-driven elastic control method, comprising the following steps.

[0006] Based on the detector, the motion and interaction signals of the master and slave ends of the bilateral teleoperation system are used to construct detection features. The attack probability is calculated through the detector and a trusted gating quantity corresponding to the communication direction is generated.

[0007] Based on the detection-switching elastic compensator, the trusted state register and Kalman filter are updated under normal conditions according to the trusted gating quantity. Under attack conditions, the trusted state is frozen and non-recursive reconstruction is performed to obtain the compensation position and compensation speed.

[0008] Based on a bounded bilateral controller, a master-slave control error is constructed according to the compensation position and the compensation speed, and a master-slave control command is output through the bounded bilateral controller.

[0009] In the attack state, the non-recursive reconstruction is based on the frozen nearest trusted position, and the position extrapolation auxiliary speed is generated by combining the Kalman prediction speed and the nearest trusted speed, and the nearest trusted speed is kept as the compensation speed.

[0010] Secondly, this application provides a bilateral teleoperation detection-driven elastic control system, including the following modules.

[0011] The detector is used to acquire motion and interaction signals from the master and slave ends of the bilateral teleoperation system to construct detection features. The detector is used to calculate the attack probability and generate a trusted gating quantity for the corresponding communication direction.

[0012] The detection switching elastic compensator is used to update the trusted state register and Kalman filter under normal conditions according to the trusted gating quantity, and freeze the trusted state and perform non-recursive reconstruction under attack conditions to obtain the compensation position and compensation speed.

[0013] A bounded bilateral controller is used to construct a master-slave control error based on the compensation position and the compensation speed, and output master-slave control commands through the bounded bilateral controller.

[0014] In the attack state, the non-recursive reconstruction is based on the frozen nearest trusted position, and the position extrapolation auxiliary speed is generated by combining the Kalman prediction speed and the nearest trusted speed, and the nearest trusted speed is kept as the compensation speed.

[0015] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described bilateral teleoperation detection-driven elastic control method.

[0016] According to the specific embodiments provided in this application, this application has the following technical effects: This application constructs detection features by acquiring motion and interaction signals and calculates attack probabilities to generate a trusted gating quantity. Based on the trusted gating quantity, the trusted state register and Kalman filter are updated under normal conditions. Under attack conditions, the trusted state is frozen, effectively blocking the path of contaminated data into the controller and avoiding the destruction of the control closed loop by distorted data. At the same time, under attack conditions, the trusted state is frozen and non-recursively reconstructed to obtain the compensation position and compensation velocity. Based on the frozen nearest trusted position, the position extrapolation auxiliary velocity is generated by combining the Kalman predicted velocity and the nearest trusted velocity, and the nearest trusted velocity is kept as the compensation velocity. This avoids the problem of geometric accumulation and amplification of errors caused by traditional recursive compensation methods during the duration of the attack. Based on the compensation position and compensation velocity, the master-slave control error is constructed, and the master-slave control command is output through the bounded bilateral controller. Thus, in the environment where time-varying delay and additive false data injection attacks coexist, it provides elastic control with deterministic error bounds for the bilateral teleoperation system, improving the synchronization accuracy and operational stability of the system under complex network attacks. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a master-slave teleoperated robotic arm control method provided in one embodiment of this application.

[0019] Figure 2 This is a schematic diagram of the integrated bilateral teleoperation system structure for detection, compensation, and bounded control provided in an embodiment of this application.

[0020] Figure 3 A schematic diagram of the ResCNN-BiLSTM fake data injection attack detector architecture provided in an embodiment of this application. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] To make the objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] like Figure 1 As shown in the figure, this application provides a bilateral teleoperation detection-driven elastic control method, which includes the following steps.

[0024] S1: Based on the detector, obtain the motion and interaction signals of the master and slave ends of the bilateral teleoperation system to construct detection features, calculate the attack probability through the detector and generate the trusted gating quantity for the corresponding communication direction.

[0025] S2: Based on the detection switching elastic compensator, according to the trusted gating quantity, the trusted state register and Kalman filter are updated in the normal state, and the trusted state is frozen and non-recursive reconstruction is performed in the attack state to obtain the compensation position and compensation speed.

[0026] S3: Based on the bounded bilateral controller, construct the master-slave control error according to the compensation position and the compensation speed, and output the master-slave control command through the bounded bilateral controller.

[0027] In the attack state, the non-recursive reconstruction is based on the frozen nearest trusted position, and the position extrapolation auxiliary speed is generated by combining the Kalman prediction speed and the nearest trusted speed, and the nearest trusted speed is kept as the compensation speed.

[0028] This application targets bilateral remote operating systems that simultaneously suffer from bounded time-varying communication delays and additive spoofing injection attacks, such as... Figure 2 As shown, an integrated bilateral teleoperation system structure of detection-compensation-bounded control is constructed, consisting of a detector, a detection switching elastic compensator, and a bounded bilateral controller. The detector is a Residual Convolutional Neural Network (ResCNN)-Bidirectional Long Short-Term Memory (BiLSTM) detector, and the bounded bilateral controller is a like nonlinear proportion plus nonlinear damping (nP+nD-like) controller.

[0029] The operator applies human-machine interaction torque to the main robotic arm. The robot arm changes its pose, and the joint angles of the master robot arm are transmitted through the forward channel network. Angular velocity of the main robotic arm joint Among them, affected by time-varying delay and fake data injection attacks The impact of this, resulting in data packets. The signal is transmitted to the slave ResCNN-BiLSTM detector, which analyzes and outputs the gating signal. Switch the elastic compensator, and the output of the elastic compensator is the compensated joint angle. and joint velocity The controller receives the control torque from the slave robotic arm. The input is given to the slave robot arm, which then changes its pose and interacts with the environment. The environment provides the slave robot arm with an environmental torque. Finally, the changed joint angles With joint velocity Feedback is transmitted via the backward channel network, which is subject to time-varying delays. and fake data injection attacks The impact of this, resulting in data packets. The signal is transmitted to the ResCNN-BiLSTM detector at the master end, which analyzes and outputs the gating signal. Given a switching elastic compensator, the angle estimation at the slave end under the compensator output delay. Estimation of end-to-end angular velocity under time delay The controller receives the control torque of the master robotic arm. The feedback is sent to the main robotic arm, which then applies the torque. Feedback is given to the operator.

[0030] The speeds and equivalent interactive torques corresponding to the two and four joint positions of the master and slave ends are transmitted through a bidirectional communication channel; received packets in either direction first enter the detection and trust determination stage, and only data that passes the trust determination is eligible to enter the state update and control calculation.

[0031] This application employs a strict "detection-gating-compensation-control" data qualification chain. The detector outputs a normal / attack determination and a trusted gating value. In the normal state, trusted packets pass directly, refreshing the trusted state register and Kalman filter. In the attack state, the currently contaminated packet is isolated, and the detection switching elastic compensator non-recursively generates a compensation position based on the frozen nearest trusted position, continuous replacement count, sampling period, and position extrapolation auxiliary speed. The frozen nearest trusted speed is held at zero order to obtain the compensation speed. The bounded bilateral controller constructs the master-slave control error based on the compensation position and the compensation speed, and outputs the master-slave control command through the bounded bilateral controller. The forward and reverse channels use the same structure but maintain their respective trusted states, filtering states, and replacement counts.

[0032] In an exemplary embodiment, S1 involves acquiring motion and interaction signals from the master and slave ends of a bilateral teleoperation system to construct detection features, including the following steps.

[0033] S11.1: Collect motion and interaction signals from the master and slave ends of the bilateral teleoperation system to form a 10-dimensional basic signal vector; the motion and interaction signals include the positions of the two joints of the master end, the positions of the two joints of the slave end, the velocities corresponding to the four joint positions, and the equivalent interaction torque; the equivalent interaction torque includes the equivalent human-machine interaction torque and the equivalent environmental interaction torque.

[0034] S11.2: Calculate the first-order and second-order differences of the 10-dimensional fundamental signal vector.

[0035] S11.3: By concatenating the 10-dimensional basic signal vector, the first-order difference, and the second-order difference in that order, a 30-dimensional enhanced feature is obtained.

[0036] S11.4: Standardize the 30-dimensional enhanced features and update the scrolling window of a preset length to obtain the detection features.

[0037] In practical applications, both the master end (i.e., the master robotic arm) and the slave end (i.e., the slave robotic arm) are described using Eulerian-Lagrange dynamics, and the control input considers component saturation; the communication delay from the master end to the slave end and from the slave end to the master end are represented by the time-varying delay in the corresponding direction; the attack adopts a model that additively tampers with the numerical components of the arrived data packets.

[0038] The dynamics of the master and slave robotic arms and the saturation function of the actuator are shown in equation (1).

[0039] (1) in, The inertial matrix of the master robotic arm. The joint angle of the main robotic arm. The joint angular acceleration of the master robotic arm The Coriolis force matrix of the master robotic arm. The angular velocity of the main robotic arm joints The gravity matrix of the main robotic arm. Human-machine interaction torque, For the constrained master end control torque, The control torque of the main robotic arm For the inertia matrix of the slave robotic arm, From the angle of the end-effector joint, For the joint angular acceleration of the end-effector robot arm, For the Coriolis force matrix of the slave robotic arm, For the angular velocity of the robotic arm joints at the end, For the gravity matrix of the slave robotic arm, For the constrained control torque of the slave robotic arm, To control the torque of the slave robotic arm, This refers to the torque exerted by the environment.

[0040] The component saturation function is shown in equation (2).

[0041] (2) in, It is a saturation function. For coefficient functions, =1,2,...,n, For a conforming function, For the first One actuator input parameter, For the first The saturation limit of each actuator.

[0042] The time-varying delay constraint is shown in equation (3).

[0043] (3) in, For delay function, This is the upper limit of latency. Let be the derivative of the time delay function. It is a constant less than 1.

[0044] Additive fake data injection is shown in equation (4).

[0045] (4) in, The data packet after being injected with fake data, For the joint angle of the robotic arm, For the angular velocity of the robotic arm joints, This is the original data packet. For injected fake data functions, The lower bound of the function. For the first In the communication direction, the first Fake data functions injected into each executor This represents the upper limit of the function.

[0046] The rate of change condition in equation (3) makes the data packet source time monotonically increase with the reception time, thus providing a positive time interval for the prediction of the variable step size Kalman filter; equation (4) indicates that the target of the attack is a valid and arrived communication packet, rather than a communication interruption. Therefore, the controller must perform a content trust determination before using the data.

[0047] Detector according to Figure 3 The process, from left to right, involves teleoperation signal flow, feature construction, online processing, residual convolution-bidirectional long short-term memory neural network model, and decision generator, forming a continuous online processing chain.

[0048] The detector receives a 30-dimensional teleoperation signal stream. ), where the 10-dimensional fundamental signal vector is Including the positions of the first two joints at the main end From the two anterior joint positions The corresponding angular velocities of the four joints mentioned above Simultaneously, the L2 norms of the human-machine interaction torque vector and the environmental interaction torque vector are respectively used as the equivalent human-machine interaction torque. and equivalent environmental torque Then, a 30-dimensional enhanced feature signal with differential information is constructed through first-order and second-order differences. Then use the first The mean of each feature and the Standard deviation of each characteristic The samples are normalized, and then the most recent 40 samples are cached to form a detection window with a dimension of 30×40. The data is then fed into a residual convolutional-bidirectional long short-term memory neural network model. Local temporal patterns are first extracted using residual block 1 and residual block 2, and then fed into the bidirectional long short-term memory neural network for temporal modeling. The modeled data is then output to a classification head, and finally, an activation function is used to output the attack probability. This information is then fed into the decision-making layer; when the attack probability is greater than or equal to the detection threshold... If this occurs, it is considered an attack and the gating signal will be activated. Set to 0; otherwise, it is considered normal and the gate signal is set. Set to 1.

[0049] At each sampling moment, the detector synchronously receives motion and interaction measurements in the bilateral teleoperation loop. In a preferred embodiment, eight motion channels are selected: the positions of the first two joints at the master end, the positions of the first two joints at the slave end, and the corresponding velocities of the aforementioned four joints. Simultaneously, the L2 norms of the human-machine interaction torque vector and the environmental interaction torque vector are used as the equivalent human-machine interaction torque and the equivalent environmental interaction torque, respectively, forming two interaction channels. These ten channels are arranged in a fixed order to form a 10-dimensional basic signal vector. The original channel does not make an attack decision independently, but rather collectively constructs detection features in the next step.

[0050] For a 10-dimensional fundamental signal vector Calculate the first difference = - Then calculate the second difference. = - Then, following the sequence of "10-dimensional basic signal vector—first-order difference—second-order difference," the 30-dimensional enhanced features consistent with the network interface are obtained. This is the 10-dimensional fundamental signal vector from the previous time step. It is the first-order difference of the 10-dimensional fundamental signal vector at the previous moment.

[0051] In the initial sampling phase, if the historical data is insufficient, the initial value can be maintained or the window accumulation can be started after the differential history is complete.

[0052] 30-dimensional enhanced features As shown in equation (5).

[0053] (5) in, T It is the transpose symbol; It is a 30-dimensional entity set.

[0054] Explicit differential channels preserve position, velocity, and interaction changes over a short period of time; subsequently, ResCNN learns higher-level local patterns within the time window of these 30-dimensional enhanced features, without replacing the aforementioned basic feature construction.

[0055] Each augmentation feature is standardized using the mean and standard deviation pre-computed and fixed at deployment time from the corresponding training subset, without using validation or test set statistics. The standardized observations are sequentially written to a rolling buffer of length L=40, with the forty most recent samples in the buffer forming a 30×40 detection window. .

[0056] The standardization and time window are shown in Equation (6).

[0057] (6) in, For the standardized first 1 eigenvalue, For the first 1 eigenvalue, For the first The mean of each feature, For the first Standard deviation of each characteristic For scrolling windows The standardized eigenvalues ​​at each time step The standardized feature value at the current time in the scrolling window. d For characteristic number, for In this embodiment, the entity set is used. d =30.

[0058] Upon online startup, a complete window is initially accumulated. After the window is formed, each decision update deletes the oldest sample, adds the latest sample, and performs network forward propagation. Therefore, the reliability of the detection includes the cumulative effect of the initial or rolling window and the time required for the current decision update.

[0059] In an exemplary embodiment, S1 calculates the attack probability using a detector, including the following steps.

[0060] S12.1: Input the detection features into the detector, extract local temporal patterns sequentially through residual convolutional blocks, perform temporal modeling through a bidirectional long short-term memory network, and output the conditional probability of the current window belonging to the normal class and the attack class through the classification head.

[0061] S12.2: The conditional probability of the attack class is used as the attack probability.

[0062] In practical applications, ResCNN local feature extraction includes: Residual Block 1, which consists of one-dimensional convolution (Conv1D), batch normalization (BN), rectified linear unit (ReLU) activation, max pooling (MaxPool1D), and dropout. Both one-dimensional convolutional layers use 32 kernels (i.e., 32 filters) with a kernel size of 5. After the first convolution, BN and ReLU are performed sequentially; after the second convolution, BN is performed. The input branch is projected onto 32 channels via a 1×1 convolution and then added to the main branch, followed by ReLU. It is then downsampled using max pooling with a pooling size of 2 and a stride of 2, and finally, Dropout with a probability of 0.2 is applied.

[0063] Residual block 2 maintains the same operation order, adjusting the number of convolution kernels to 64 (i.e., 64 filters), keeping the kernel size at 5, and using 1×1 convolution to complete the 64-channel residual projection; after summing the residuals and ReLU, max pooling and Dropout with a probability of 0.2 are performed again. Residual block 1 and residual block 2 thus extract and compress the local temporal patterns in the window sequentially.

[0064] BiLSTM temporal modeling: The first layer, a bidirectional long short-term memory network, has 48 hidden units in each direction (i.e., BiLSTM units=48) and operates in sequence output mode to preserve the complete time series after convolutional downsampling; its output is first subjected to Dropout with a probability of 0.2. The second layer, a BiLSTM, has 24 hidden units in each direction (i.e., BiLSTM units=24) and operates in terminal output mode, combining the forward and backward terminal states into a 48-dimensional fixed-length temporal summary, which is then subjected to Dropout with a probability of 0.3 and fed into the classification head. This sequence allows for a continuous connection between local convolutional features and bidirectional dependencies over a longer time range.

[0065] Classification Head and Attack Probability Generation: The classification head sequentially executes a 32-unit fully connected layer (FC(32)), ReLU, Dropout with a probability of 0.3, a 2-unit fully connected layer (FC(2)), and a normalized exponential function (Softmax), outputting the conditional probabilities that the current window belongs to the Normal and Attack classes. Classification cross-entropy is used during training; when classes are imbalanced, the weights of each class are determined by dividing the total number of training sequences by twice the number of sequences of that class. The classification head only generates probabilities; the final on / off state is determined by the validation set threshold in the next step.

[0066] In an exemplary embodiment, generating a trusted gating quantity corresponding to the communication direction in S1 includes the following steps.

[0067] S13.1: Search the validation set for the candidate value that maximizes the F1 score and use it as the detection threshold.

[0068] S13.2: When the attack probability is not less than the detection threshold, output the attack label and set the trusted gating quantity to 0.

[0069] S13.3: When the attack probability is less than the detection threshold, output a normal label and set the trusted gating quantity to 1.

[0070] In practical applications, the detection decision and compensation switch interface does not use the default threshold of 0.5, nor does it utilize test set information. Instead, it searches for the candidate value with the highest F1 score among the candidate values ​​in the validation set as the detection threshold. When multiple candidates achieve the same maximum F1 score, the minimum threshold among them is taken as the detection threshold. This candidate value is the attack probability candidate value. When the attack probability of the current window is not less than... Output Attack if it is true, otherwise output Normal, and map the label to a binary trusted gating quantity for the corresponding communication direction.

[0071] The maximum F1 threshold and trusted gating of the validation set are shown in Equation (7).

[0072] (7) in, To set a threshold, The F1 threshold, The candidate threshold is formed by the attack probability of the validation set and the endpoints of the interval. For trusted gating quantities, This represents the conditional probability.

[0073] Trusted Gated Quantity =1 indicates Normal, which allows trusted packets to enter the normal path of the compensator; =0 indicates an Attack, initiating an alternative reconstruction in that direction. This trusted gating quantity is not only an alarm signal but also a common qualification control quantity for packet passage, trusted state register overlay, and Kalman filter measurement updates. The Industrial Internet of Things (IIoT) dataset architecture selects data that directly provides 30-dimensional features, while the Robotic Surgical Maneuver Dataset (ROSMA) and hardware domain are arranged according to... , and Construct interfaces with the same dimensions.

[0074] In one exemplary embodiment, S2 includes the following steps.

[0075] S21: When the trusted gating value is 1, under normal conditions, the current trusted packet is allowed to pass through, and the latest trusted location, latest trusted speed and source timestamp are refreshed using the current delayed packet.

[0076] S22: Execute the prediction step and measurement update step of the Kalman filter, reset the continuous replacement count, and update the confidence status register and Kalman filter.

[0077] S23: When the trusted gating quantity is 0, under attack conditions, the current trusted packet is isolated, the nearest trusted location, the nearest trusted velocity and the source timestamp are frozen, the measurement update step of the Kalman filter is skipped, and only the prediction step of the Kalman filter is executed.

[0078] S24: Generate the compensation position non-recursively based on the frozen nearest reliable position, the successive replacement count, the sampling period, and the position extrapolation auxiliary speed.

[0079] S25: The frozen most recent reliable velocity is preserved at zero order to obtain the compensated velocity.

[0080] In practical applications, a trusted status register is set up for each communication direction, storing at least the most recent trusted location, the most recent trusted speed, and the corresponding source timestamp. Only when... When =1, the current trusted packet overwrites the trusted status register; when When the value changes from 1 to 0, the most recent trusted reception time and its source timestamp are frozen and used as the common initial value for subsequent non-recursive reconstruction.

[0081] The conditions for updating the trusted status register are shown in equation (8).

[0082] (8) in, The joint position at the time of the frozen trusted source; For the moment when the trusted source is frozen; Joint positions under time-varying and time-delay conditions; Joint velocities under time-varying delays; To update the symbols.

[0083] Variable step size Kalman prediction and reliable measurement update: The filter state of the Kalman filter consists of position and velocity, employing a constant velocity motion model. The source timestamp interval carried by adjacent received packets... Used for online construction of state transition matrices and process noise covariance under discrete white noise acceleration model The measurement matrix only extracts the position component, and the measurement noise covariance is determined by... Sure.

[0084] The variable step size state equation is shown in equation (9).

[0085] (9) in, For process noise covariance, For variable step size parameters, For variable step size parameters, It is an n-order identity matrix. For the first The robotic arm is constantly inputting its status. For the first The robotic arm is constantly inputting its status. For the first The robotic arm output status is constantly monitored. It is a row-block matrix.

[0086] The prediction steps are shown in equation (10).

[0087] (10) in, Estimating the state from the previous moment, Forward covariance of the previous moment, The process noise covariance of the previous time step. The process noise covariance of the previous moment. For the state estimation of the previous moment, This is the forward covariance of the previous time step.

[0088] The measurement update performed only when trusted is shown in Equation (11).

[0089] (11) in, For profit value, To measure the noise covariance matrix, For the observation matrix, Estimate the state at the current moment. Forward covariance at the current moment.

[0090] State and covariance predictions based on equation (10) can be performed in each cycle; only The measurement correction in equation (11) is performed only when the value is 1. During the attack, the measurement update is skipped and the predicted covariance is carried into the next cycle, so the attack data cannot contaminate the subsequent estimates through the filter correction stage.

[0091] When the detection result is Normal, the received position and velocity remain unchanged and are sent to the controller. Simultaneously, the trusted position, trusted velocity, and source timestamp are refreshed using the current trusted packet, and the Kalman measurement update is completed. Continuous replacement of the count reset ensures that the initial reconstruction value for the next attack comes from the latest trusted state.

[0092] When the detection result is Attack, the current trusted packet must not enter the controller, must not overwrite the trusted state register, and must not participate in Kalman measurement updates. The compensation speed uses the zero-order hold of the most recent trusted speed; the compensation position uses the frozen trusted position as a fixed reference, superimposed with the continuous replacement count, sampling period, and position extrapolation auxiliary speed obtained by combining the Kalman predicted speed and the most recent trusted speed according to the position hold gain.

[0093] communication direction The unified switching compensation is shown in equation (12).

[0094] (12) in, The joint position after compensation; The compensated joint velocity; The current moment; For time-varying delay; For continuous replacement counting; The sampling period; To assist in position extrapolation speed; The joint velocity at the moment the trusted source is frozen.

[0095] The position extrapolation auxiliary speed is shown in equation (13).

[0096] (13) in, Maintain gain for position; Predicting speed for Kalman.

[0097] The upper bound of the position error during continuous replacement is shown in equation (14).

[0098] (14) in, For discrete Lyapunov functions, For speed tracking error, For the first Disturbance suppression boundaries in each channel direction.

[0099] Each attack step in equation (12) is based on Using the same benchmark, and through It reflects the cumulative time since the most recent trusted state; it does not simply re-input the previous compensation position into the next compensation. Therefore, the deterministic error bound of equation (14) can be obtained within the duration of the bounded attack.

[0100] In one exemplary embodiment, S3 includes the following steps.

[0101] S31: Construct position error and velocity error by using the real-time status of the master and slave ends and the compensated position and compensated velocity output from the opposite channel.

[0102] S32: The position error is obtained by applying a sign-power nonlinear mapping to obtain a position feedback term, and the velocity error is obtained by applying a sign-power nonlinear mapping to obtain a velocity feedback term and a local damping term.

[0103] S33: The position feedback term, the velocity feedback term, and the local damping term are superimposed, and gravity compensation is added to obtain the master-slave control command.

[0104] In practical applications, the master-to-slave and slave-to-master channels maintain trusted gating values, trusted state registers, Kalman states, and successive replacement counts, respectively, and perform compensation on their transmitted objects according to the same formula. When the attack returns to Normal, the current trusted packet immediately refreshes the trusted state register and triggers a Kalman measurement update, the replacement count is cleared to zero, and the controller reuses the real trusted packet; the perturbation terms related to attack reconstruction disappear accordingly, and the closed-loop residual set shrinks accordingly.

[0105] The controller receives the real-time status from the master and slave ends as well as the compensated delayed status output from the opposite channel. It forms a finite-time nominal control using symbolic power position feedback, symbolic power relative velocity feedback, and local damping feedback, and suppresses the command peak of the feedback channel through bounded continuous mapping.

[0106] The master end constructs the error by subtracting the slave end's delayed position and velocity after reverse channel compensation from the master and slave end's positions and velocities; the slave end constructs the master-slave control error by subtracting the master end's delayed state after forward channel compensation from the master and slave end's states. Thus, the detection and reconstruction results directly contribute to the master-slave control error.

[0107] The master-slave control error is shown in equation (15).

[0108] (15) in, To control the error of the main end angle. Main end angular velocity control error For time delay estimation from the end-user perspective, For estimation of end-to-end angular velocity under time delay, This is a conversion symbol.

[0109] In one exemplary embodiment, the sign power nonlinear mapping is shaped by a continuous bounded arctangent mapping.

[0110] The continuous bounded arctangent mapping remains linear in the neighborhood of the origin. When the absolute value of the input exceeds the joint threshold, the amplitude is compressed using an arctangent function that is continuously scaled according to the joint threshold.

[0111] In practical applications, position error is expressed exponentially. The sign power nonlinearity, relative velocity feedback, and local damping employ a finite-time exponent consistent with the position exponent. The sign exponentiation operation preserves the original sign of each joint component and raises the absolute value to the power, thus obtaining a non-Lipschitz finite-time feedback characteristic near the origin.

[0112] The exponential relationship is shown in equation (16).

[0113] (16) Continuous bounded arctangent mapping: Each finite-time exponential feedback channel is shaped using an odd function, strictly increasing, continuous, and bounded mapping. The preferred mapping maintains linearity in the origin neighborhood to preserve the local control structure; when the absolute value of the input exceeds the joint threshold... At that time, the amplitude is compressed using an arctangent function that scales continuously by a threshold.

[0114] The continuous bounded arctangent mapping function is shown in equation (17).

[0115] (17) in, For independent variable parameters, It is a continuous and bounded arctangent mapping function.

[0116] Bounded control law for master and slave ends: The master and slave ends respectively superimpose the bounded position term, bounded relative velocity term, and bounded local damping term to form the feedback control quantity, and then add the gravity compensation of the corresponding robotic arm. All three types of gains are positive definite diagonal matrices and can be independently calibrated for different joints and master-slave load differences.

[0117] The master-slave unified control law is shown in equation (18).

[0118] (18) in, For control laws, For the proportion term, For the angle error of the robotic arm, For proportional gain, For speed gain, For the speed term, For the speed error of the robotic arm, For damping gain, For damping term, To control the torque, For gravity compensation, This refers to the joint angle of the robotic arm.

[0119] The continuous bounded arctangent mapping limits the peak value of each nonlinear feedback channel, but cannot guarantee that the total torque after gravity compensation will never trigger actuator saturation. Therefore, the following finite time and actual elastic conclusions are limited to the allowable operating domain without active saturation.

[0120] Local finite time and actual elastic boundary: Under the nominal conditions of zero delay, no attacks, free motion, no active saturation, and using symmetric theoretical gain, the synchronization error system satisfies a negative-order homogeneous finite-time comparison relation near the synchronization manifold. When considering bounded time-varying delay, finite detection error, bounded reconstruction error, and continuous realization error, these factors are included as a total disturbance in the comparison equation.

[0121] The finite-time comparison formula for a zero-latency, attack-free nominal system is shown in equation (19).

[0122] (19) in, The derivative of the Lyapunov energy function. It is a local constant. For the index, For homogeneity, To maximize homogeneity, It represents the fractional power exponent.

[0123] The comparison formulas for delay, detection / reconstruction error and implementation error are shown in Equation (20).

[0124] (20) in, This is the upper bound of the total perturbation. This is the upper bound of the communication / computation latency. This is the upper bound for deception attacks suffered by the trusted interface. This represents the upper bound of the hardware / implementation error.

[0125] The final bounded set is shown in equation (21).

[0126] (twenty one) in, For the safety residual set, This represents the steady-state error boundary.

[0127] When there is zero latency, no attack, free movement, no active saturation, and symmetric theoretical gain, local finite-time synchronization can be obtained near the synchronous manifold. When there is persistent latency, finite detection errors, and bounded reconstruction errors, the system enters the local residual set of equation (21) in finite time. After the trusted packet is recovered and the trusted state register is refreshed, attack-related... The residual set disappears and shrinks.

[0128] The above conclusions do not extend to zero-error convergence under continuous non-zero time delay, nor do they claim that human-machine contact operation or active actuator saturation are covered by the finite-time theorem.

[0129] This application forms a closed-loop data qualification chain of detection, gating, reconfiguration and control, so that attack packets are excluded before they enter the controller, rather than just generating alarms.

[0130] Both the trusted state register and the Kalman measurement update are subject to the same trusted gating constraint, which prevents attack data from contaminating subsequent estimated states.

[0131] Each step of the location reconstruction starts from the same frozen reliable state and does not recursively use the previous compensation output, thus structurally avoiding the accumulation of geometric errors.

[0132] The velocity output uses the nearest reliable velocity zero-order hold, and the position extrapolation can utilize a combination of Kalman prediction and reliable velocity to take into account both short-term motion trends and robustness.

[0133] Symmetrical deployment of forward and reverse channels enables simultaneous handling of multi-channel position / velocity FDI from master to slave and from slave to master.

[0134] Bounded arctangent mapping limits the peak value of the sign power feedback channel, balancing finite-time nominal response and hardware command smoothing.

[0135] Stability analysis treats time-varying delay, finite detection error, and bounded reconstruction error as explicit perturbations, and gives a shrinkable local final bounded set.

[0136] The baseline experiment with the same controller isolated the closed-loop value of the detector and compensator combination. Under the test conditions, the double joint position error of the attack window was reduced by 93.4%, which is the root-mean-square error (RMSE).

[0137] This application provides a bilateral teleoperation detection-driven elastic control system, which includes the following modules.

[0138] The detector is used to acquire motion and interaction signals from the master and slave ends of the bilateral teleoperation system to construct detection features. The detector is used to calculate the attack probability and generate a trusted gating quantity for the corresponding communication direction.

[0139] The detection switching elastic compensator is used to update the trusted state register and Kalman filter under normal conditions according to the trusted gating quantity, and freeze the trusted state and perform non-recursive reconstruction under attack conditions to obtain the compensation position and compensation speed.

[0140] A bounded bilateral controller is used to construct a master-slave control error based on the compensation position and the compensation speed, and output master-slave control commands through the bounded bilateral controller.

[0141] In the attack state, the non-recursive reconstruction is based on the frozen nearest trusted position, and the position extrapolation auxiliary speed is generated by combining the Kalman prediction speed and the nearest trusted speed, and the nearest trusted speed is kept as the compensation speed.

[0142] The technical solution of this application is further illustrated below with specific embodiments.

[0143] In practical applications, the workflow of this application is as follows.

[0144] Step 1: Collect the joint positions, velocities, and equivalent interactive torques of the master and slave ends, and form a 10-dimensional basic signal vector in a fixed order.

[0145] Step 2: Calculate the first and second differences of the 10-dimensional basic signal vector and concatenate them to form a 30-dimensional enhanced feature; standardize using training set statistics and update the rolling window of length 40.

[0146] Step 3: After the window is complete, pass it through two ResCNN residual blocks, two BiLSTM layers, and a classification head in sequence to obtain the Normal probability and Attack probability.

[0147] Step 4: Compare the Attack probability with the maximum F1 threshold of the validation set, output the Normal / Attack label and generate the trusted gating quantities for the forward and reverse channels.

[0148] Step 5: If the trusted gating value is 1, the trusted packet passes directly, the most recent trusted location, velocity and source timestamp are refreshed, Kalman prediction and measurement update are performed, and the continuous replacement count is reset.

[0149] Step 6: If the trusted gating value is 0, isolate the current trusted packet, freeze the trusted state, and skip the Kalman measurement update; retain only the prediction step, start from the frozen baseline non-recursive reconstruction position, and maintain the trusted speed at order zero.

[0150] Step 7: Construct master and slave control errors using the compensation position and compensation speed of the opposing channel, sequentially form symbolic power feedback, perform bounded mapping and superimpose gravity compensation, and output master and slave control commands.

[0151] Step 8: Continuously monitor the detection results and continuous replacement count; immediately refresh the trust status register and filter after restoring trust and exit the replacement state, while recording the detection, switching, reconstruction and tracking quantities for running boundary checks.

[0152] Detector Training and Domain Transfer Implementation: First, BiLSTM, BiLSTM-Attention, and ResCNN-BiLSTM are compared on publicly available IIoT data using a uniform 30-dimensional augmented feature input, a non-overlapping window of length 40, and the same data partitioning to select the detection architecture. Then, starting from random initialization, the selected architecture is retrained on ROSMA surgical teleoperation data, instead of transferring the IIoT weights.

[0153] As shown in Table 1, the ROSMA data includes 206 trials involving 12 subjects, with training / validation / testing divided into 8 / 2 / 2 subjects. Each trial injected a synthetic attack interval of 100 samples, 2 seconds in length, and non-overlapping in three time periods, with eight position and velocity channels added. Noise. All test errors occurred in the label boundary window. This result only demonstrates separability under a specified synthetic attack protocol and does not constitute a general guarantee of detection for real, low-amplitude, slowly varying, correlated, or adaptive attacks.

[0154] Table 1. Training performance parameters of the detector on the two datasets

[0155] Wherein, TP: True Positive; TN: True Negative; FP: False Positive; FN: False Negative; F1: F1-Score; AUC: Area Under the ROC Curve.

[0156] Hardware Platform and Online Deployment Implementation Example: The hardware system consists of a master robotic arm haptic master device and a slave robotic arm industrial robot slave device. The master robotic arm is a SensAble Phantom Omni, and the slave robotic arm is a Denso VP-6242G. A ROSMA retrained model is deployed to this platform, enabling a cross-platform migration from the dVRK data domain to the PhantomOmni-Denso hardware domain.

[0157] The sampling period is controlled to be 1 ms. The forward and reverse target delays are sampled independently in each period, and a rate-of-change limiter is used to generate a sample that satisfies... s and Application latency. The attack intervals are 12-14 s, 24-26 s, and 36-38 s; each component of the position and velocity packets is added independently. Additive attacks, among which, is the derivative of the time delay function.

[0158] Comparative experiments and technical effects: Configuration C1 enables the detector, bidirectional switching compensator, and bounded nonlinear control; configuration C2 uses the same nP+nD-like controller but disables the detector and compensator. Both configurations replay the same full-duration delay curves and attack intervals to isolate the closed-loop value of the detection-driven compensation, as shown in Table 2.

[0159] The positional error of the attack window's two joints is shown in equation (22).

[0160] (twenty two) in, To address the positional error of the dual joints in the attack window; N Let A be the number of samples for each joint and configuration. Let be the rational numbers within the region. t For the current moment, This is the square of the joint 1 error. This is the square of the joint 2 error.

[0161] Table 2 Comparison and explanation of joint errors under two experimental configurations

[0162] The second joint still exhibits small bounded oscillations near the velocity reversal point, which may be related to friction, dead zone, backlash, gravity load, and velocity estimation / switching transients; the bounded mapping prevents these non-ideal factors from being amplified within the test attack range.

[0163] In practical applications, the detection input can be extended according to the specific robot degrees of freedom; the interactive torque can use direct measurements, equivalent norms, or task-related force / torque features.

[0164] While maintaining the principles of "only trusted data has the right to be updated" and "non-recursive reconstruction from frozen trusted states", the detector can be replaced by other temporal models that can output trusted probabilities, with ResCNN-BiLSTM being the preferred embodiment.

[0165] Bounded mappings can be other functions that satisfy singularity, strict monotonicity, continuity, local Lipschitz, linearity in the neighborhood of the origin, and global boundedness; continuous arctangent mappings are preferred embodiments.

[0166] The maximum consecutive replacement count should be determined by the allowed attack duration, movement speed range, and acceptable reconstruction error; if the upper limit is exceeded, the system can enter shutdown, degradation, or manual takeover mode.

[0167] Formal engineering deployment requires verification of safety boundaries under free motion, contact motion, active saturation, and different communication networks; the finite-time theorem in this paper does not automatically cover contact and active saturation conditions.

[0168] In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments. The computer device can be a server or a terminal. The computer device includes a processor, a memory, an input / output interface (I / O), and a communication interface. The processor, memory, and I / O interface are connected via a system bus, and the communication interface is connected to the system bus via the I / O interface. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device stores data to be processed. The I / O interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to connect and communicate with an external terminal via a network. When the computer program is executed by the processor, it implements the above-described methods.

[0169] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0170] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0171] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0172] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by computer program instructions and related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the above method embodiments. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0173] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0174] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0175] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A bilateral teleoperation detection-driven elastic control method, characterized in that, include: Based on the detector, the motion and interaction signals of the master and slave ends of the bilateral teleoperation system are used to construct detection features. The attack probability is calculated through the detector and a trusted gating quantity corresponding to the communication direction is generated. Based on the detection switching elastic compensator, according to the trusted gating quantity, the trusted state register and Kalman filter are updated in the normal state, and the trusted state is frozen and non-recursive reconstruction is performed in the attack state to obtain the compensation position and compensation speed. Based on a bounded bilateral controller, a master-slave control error is constructed according to the compensation position and the compensation speed, and a master-slave control command is output through the bounded bilateral controller; In the attack state, the non-recursive reconstruction is based on the frozen nearest trusted position, and the position extrapolation auxiliary speed is generated by combining the Kalman prediction speed and the nearest trusted speed, and the nearest trusted speed is kept as the compensation speed.

2. The bilateral teleoperation detection-driven elastic control method according to claim 1, characterized in that, The motion and interaction signals of the master and slave ends of the bilateral teleoperation system are used to construct detection features, including: The motion and interaction signals of the master and slave ends of the bilateral teleoperation system are collected to form a 10-dimensional basic signal vector. The motion and interaction signals include the positions of two joints of the master end, the positions of two joints of the slave end, the velocities corresponding to the four joint positions, and the equivalent interaction torque. The equivalent interaction torque includes the equivalent human-machine interaction torque and the equivalent environmental interaction torque. Calculate the first-order and second-order differences of the 10-dimensional fundamental signal vector; By concatenating the 10-dimensional basic signal vector, the first-order difference, and the second-order difference in that order, a 30-dimensional enhanced feature is obtained. The 30-dimensional enhanced features are standardized, and the scrolling window of a preset length is updated to obtain the detection features.

3. The bilateral teleoperation detection-driven elastic control method according to claim 2, characterized in that, The probability of an attack is calculated using a detector, including: The detection features are input into the detector, and local temporal patterns are extracted sequentially through residual convolutional blocks. Temporal modeling is performed through a bidirectional long short-term memory network, and the conditional probability of the current window belonging to the normal class and the attack class is output through the classification head. The conditional probability of the attack class is used as the attack probability.

4. The bilateral teleoperation detection-driven elastic control method according to claim 3, characterized in that, Generate trusted gating quantities for the corresponding communication direction, including: The candidate value that maximizes the F1 score in the validation set is used as the detection threshold. When the attack probability is not less than the detection threshold, an attack label is output and the trusted gating value is set to 0; When the attack probability is less than the detection threshold, a normal label is output and the trusted gating value is set to 1.

5. The bilateral teleoperation detection-driven elastic control method according to claim 4, characterized in that, Based on the trusted gating value, the trusted state register and Kalman filter are updated under normal conditions. Under attack conditions, the trusted state is frozen and non-recursive reconstruction is performed to obtain the compensation position and compensation speed, including: When the trusted gating value is 1, under normal conditions, the current trusted packet is allowed to pass through, and the latest trusted location, latest trusted speed and source timestamp are refreshed using the current delayed packet; Perform the prediction and measurement update steps of the Kalman filter, reset the continuous replacement count, and update the confidence state register and Kalman filter; When the trusted gating value is 0, under attack conditions, the current trusted packet is isolated, the nearest trusted location, the nearest trusted velocity and the source timestamp are frozen, the measurement update step of the Kalman filter is skipped, and only the prediction step of the Kalman filter is executed. The compensation position is generated non-recursively based on the nearest reliable frozen position, the successive replacement count, the sampling period, and the position extrapolation assist speed. The most recently reliable velocity that was frozen is preserved at zero order to obtain the compensated velocity.

6. The bilateral teleoperation detection-driven elastic control method according to claim 5, characterized in that, The compensation position is: ; in, The joint position after compensation; The compensated joint velocity; The current moment; For time-varying delay; For trusted gating quantities; Joint positions under time-varying and time-delay conditions; The joint velocity under time-varying delay; The joint position at the time of the frozen trusted source; For the moment when the trusted source is frozen; For continuous replacement counting; The sampling period; To assist in position extrapolation speed; The joint velocity at the moment the trusted source is frozen; For communication direction; The position extrapolation assist speed is: ; in, Maintain gain for position; Predicting speed for Kalman.

7. The bilateral teleoperation detection-driven elastic control method according to claim 1, characterized in that, Based on the compensation position and the compensation speed, a master-slave control error is constructed, and master-slave control commands are output through a bounded bilateral controller, including: Position error and velocity error are constructed by using the real-time status of the master and slave ends and the compensated position and compensated velocity output from the opposite channel; The position error is mapped using a sign-power nonlinear mapping to obtain a position feedback term, and the velocity error is mapped using a sign-power nonlinear mapping to obtain a velocity feedback term and a local damping term. The position feedback term, the velocity feedback term, and the local damping term are superimposed, and gravity compensation is added to obtain the master-slave control command.

8. The bilateral teleoperation detection-driven elastic control method according to claim 7, characterized in that, The symbolic power nonlinear mapping is shaped by a continuous bounded arctangent mapping. The continuous bounded arctangent mapping remains linear in the neighborhood of the origin. When the absolute value of the input exceeds the joint threshold, the amplitude is compressed using an arctangent function that is continuously scaled according to the joint threshold.

9. A bilateral teleoperation detection and drive elastic control system, characterized in that, The bilateral teleoperation detection-driven elastic control method according to any one of claims 1-8 includes: The detector is used to acquire motion and interaction signals from the master and slave ends of the bilateral teleoperation system to construct detection features. The detector is used to calculate the attack probability and generate a trusted gating quantity for the corresponding communication direction. The detection switching elastic compensator is used to update the trusted state register and Kalman filter under normal conditions according to the trusted gating quantity, and freeze the trusted state and perform non-recursive reconstruction under attack conditions to obtain the compensation position and compensation speed. A bounded bilateral controller is used to construct a master-slave control error based on the compensation position and the compensation speed, and output master-slave control commands through the bounded bilateral controller; In the attack state, the non-recursive reconstruction is based on the frozen nearest trusted position, and the position extrapolation auxiliary speed is generated by combining the Kalman prediction speed and the nearest trusted speed, and the nearest trusted speed is kept as the compensation speed.

10. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the bilateral teleoperation detection-driven elastic control method according to any one of claims 1-8.

Images