Digital twin driven medical service robot control safety monitoring system

Abstract

This application provides a digital twin-driven safety monitoring system for medical service robot control, relating to the field of medical service robot control safety monitoring technology. It organizes data such as joint angles, end effector torque, and motor current into spatiotemporal telemetry features, and performs simulations under two assumptions: mechanical degradation and physiological disturbance. The system then decouples the disturbance source by comparing the residuals between the real and simulated states, determining the source and intensity of the anomaly. Subsequently, the decoupling results are input into a safety incident knowledge graph to form a dynamic risk assessment, which is used to match candidate intervention strategies. Before being issued, candidate actions are reinjected into the digital twin for secondary rehearsal, eliminating action sequences that may cause secondary damage. Finally, safe and reliable intervention control commands are generated, thereby improving the accuracy of anomaly attribution, risk prediction, and control intervention.

Description

Technical Field

[0001] This application relates to the field of medical service robot control safety monitoring technology, and more specifically, to a digital twin-driven medical service robot control safety monitoring system. Background Technology

[0002] As the application of surgical robots, rehabilitation robots, and nursing robots in clinical settings continues to expand, medical service robots have gradually transformed from auxiliary information interaction devices into intelligent equipment that directly participates in patient physical contact and the execution of treatment actions. In contact-based tasks such as rehabilitation traction and intraoperative soft tissue traction, a closed mechanical transmission chain is formed between the robot's end effector and the patient's tissue. Any control deviation, mechanical deterioration, or sudden physiological reaction from the patient can be rapidly amplified through this mechanical chain, potentially leading to risks such as over-traction, tissue damage, or treatment interruption.

[0003] Existing safety monitoring technologies for medical service robots primarily rely on threshold alarms, single-type fault detection, or post-incident anomaly handling. Even with the introduction of digital twins, they are typically used only as passive mirrors of the robot's operational state, making it difficult to accurately attribute the sources of abnormal disturbances. Especially in contact-based medical tasks, the robot's internal mechanical degradation and the patient's external physiological disturbances are not always independent: for example, high-frequency vibrations from early wear of joint reducers may induce muscle spasms in patients, and the resulting reverse impact load from these spasms can further exacerbate wear on the transmission chain. If existing solutions still employ a hard-score approach, choosing between mechanical faults and physiological disturbances, it's easy to forcibly attribute a problem to a single source in a dual-source coupled state, leading to the omission of the other disturbance source and causing subsequent intervention strategies to deviate from the actual risk. Furthermore, if emergency actions are not validated through secondary rehearsals using a digital twin, secondary injuries may occur during emergency stops, locking, or compliant following, failing to meet the requirements of clinical scenarios for safety monitoring accuracy and intervention reliability.

[0004] Therefore, an optimized safety monitoring system for the control of medical service robots is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a digital twin-driven medical service robot control safety monitoring system, comprising:

[0006] The spatiotemporal telemetry feature construction module is used to perform spatiotemporal phase alignment and Kalman filtering denoising on the joint angle data stream, end torque data stream and motor current data stream collected by the medical service robot in contact tasks to obtain the spatiotemporal telemetry feature matrix.

[0007] The dual-branch digital twin simulation module is used to inject the spatiotemporal telemetry feature matrix into the dual positive simulation branches with mutually exclusive boundary conditions in the digital twin engine to perform hypothesis testing-based dynamic deduction in order to obtain the mechanical degradation simulation tensor and the physiological disturbance simulation tensor.

[0008] The disturbance source decoupling module is used to perform spatiotemporal dual-domain residual analysis and precise decoupling of the disturbance source from the spatiotemporal telemetry feature matrix, the mechanical degradation simulation tensor, and the physiological disturbance simulation tensor to obtain the decoupled disturbance source vector.

[0009] The dynamic safety assessment module is used to map the abnormal source type identifier and deviation intensity value in the decoupled disturbance source vector into a pre-trained safety incident knowledge graph based on graph neural network to obtain a dynamic safety assessment report.

[0010] The intervention strategy verification and control module is used to adaptively match candidate intervention strategies based on the risk level and anomaly source type in the dynamic safety assessment report. The candidate intervention action set is injected back into the digital twin for secondary pre-run verification to eliminate action sequences with secondary damage risk. The selected optimal strategy is compiled into intervention control instructions and sent to the robot actuator.

[0011] Compared with existing technologies, this application provides a digital twin-driven safety monitoring system for medical service robots. It collects operational data such as joint angles, end effector torque, and motor current during contact tasks to form a spatiotemporal telemetry feature matrix characterizing the robot's real-time force and motion state. This feature matrix is ​​then fed into the digital twin simulation branches corresponding to the mechanical degradation and physiological disturbance assumptions for parallel simulation. Residual analysis is performed between the actual telemetry state and the two types of simulation results to identify whether abnormal disturbances are closer to mechanical-side changes or patient-side disturbances. Furthermore, coupling degree estimation and soft contribution decomposition are used to express the combined effect of two sources, avoiding the simplistic attribution of complex contact risks to a single source. Subsequently, the decoupled abnormal source type, deviation intensity, and coupling state are mapped to a safety incident knowledge graph to generate dynamic safety assessment results, which are then used to match candidate intervention strategies. Candidate actions are pre-validated in the digital twin before execution, thereby achieving early risk identification and accurate attribution while reducing the risk of misinterpretation and secondary injuries. Attached Figure Description

[0012] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0013] Figure 1 This is a system block diagram of a digital twin-driven medical service robot control safety monitoring system according to an embodiment of this application;

[0014] Figure 2 This is a schematic diagram of data flow in a digital twin-driven medical service robot control safety monitoring system according to an embodiment of this application;

[0015] Figure 3 This is a block diagram of a dual-branch digital twin simulation module in a digital twin-driven medical service robot control safety monitoring system according to an embodiment of this application;

[0016] Figure 4 This is a block diagram of a disturbance source decoupling module in a digital twin-driven medical service robot control safety monitoring system according to an embodiment of this application;

[0017] Figure 5 This is a block diagram of a disturbance source screening and decoupling unit in a digital twin-driven medical service robot control safety monitoring system according to an embodiment of this application;

[0018] Figure 6 This is a block diagram of the intervention strategy verification and control module in a digital twin-driven medical service robot control safety monitoring system according to an embodiment of this application. Detailed Implementation

[0019] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0020] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0021] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0022] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0023] Currently, medical service robots, in contact-based tasks such as rehabilitation traction and nursing assistance, are easily affected by both internal mechanical degradation of the robot and external physiological disturbances from patients. Existing safety monitoring methods often only detect abnormalities in torque, current, or posture, but struggle to accurately pinpoint the source of these abnormalities. Especially in coupled scenarios where mechanical vibrations induce muscle spasms in patients, and the patient's reverse impact exacerbates wear on transmission components, traditional mutually exclusive discrimination methods easily misclassify complex risks as single faults, causing intervention strategies to deviate from the true risk. Therefore, this application proposes a digital twin-driven control safety monitoring system for medical service robots. The system first performs spatiotemporal alignment, filtering and noise reduction, and feature matrix construction on data such as joint angles, end torque, and motor current. Then, this feature matrix is ​​input in parallel into the mechanical degradation hypothesis branch and the physiological disturbance hypothesis branch. Two hypothetical states are obtained through forward simulation using a digital twin, and the residuals are compared with the actual telemetry states. Based on this, instead of a simple either-or hard decision approach, the system calculates the independent activation energies of the mechanical and physiological channels separately, further estimating the cross-source coupling degree between them. A decoupled disturbance source vector containing mechanical contribution, physiological contribution, and coupling degree is formed through soft contribution decomposition. Subsequently, the system inputs this vector into a safety accident knowledge graph and a dynamic risk assessment module to identify single-source faults or dual-source coupled co-existence, and matches corresponding candidate intervention strategies. Finally, the candidate strategies are injected back into the digital twin for secondary pre-playing. Action sequences that may cause collisions, traction, or secondary injuries are eliminated before the optimal intervention command is issued to the robot actuator, thus balancing fault tracing accuracy, intervention safety, and clinical application reliability.

[0024] Figure 1 This is a system block diagram of a digital twin-driven medical service robot control safety monitoring system according to an embodiment of this application. Figure 2 This is a schematic diagram of data flow in a digital twin-driven medical service robot control safety monitoring system according to an embodiment of this application. Figure 1 and Figure 2As shown, a digital twin-driven medical service robot control safety monitoring system 100 according to an embodiment of this application includes: a spatiotemporal telemetry feature construction module 110, used to perform spatiotemporal phase alignment and Kalman filtering denoising on the joint angle data stream, end torque data stream, and motor current data stream collected by the medical service robot in contact tasks to obtain a spatiotemporal telemetry feature matrix; a dual-branch digital twin simulation module 120, used to inject the spatiotemporal telemetry feature matrix in parallel into the dual positive simulation branches with mutually exclusive boundary conditions in the digital twin engine to perform hypothesis testing-based dynamic deduction to obtain a mechanical degradation simulation tensor and a physiological disturbance simulation tensor; and a disturbance source decoupling module 130, used to decouple the spatiotemporal telemetry feature matrix and the mechanical degradation... The simulation tensor and physiological disturbance simulation tensor are subjected to spatiotemporal dual-domain residual analysis and precise decoupling of disturbance sources to obtain decoupled disturbance source vectors; the dynamic safety assessment module 140 is used to map the abnormal source type identifier and deviation intensity value in the decoupled disturbance source vector into a pre-trained safety accident knowledge graph based on graph neural network to obtain a dynamic safety assessment report; the intervention strategy verification and control module 150 is used to adaptively match candidate intervention strategies based on the risk level and abnormal source type in the dynamic safety assessment report. In this module, the candidate intervention action set is back-injected into the digital twin for secondary pre-run verification to eliminate action sequences with secondary damage risk. The selected optimal strategy is compiled into intervention control instructions and sent to the robot actuator.

[0025] In the aforementioned digital twin-driven medical service robot control and safety monitoring system 100, the spatiotemporal telemetry feature construction module 110 is used to perform spatiotemporal phase alignment and Kalman filtering denoising on the joint angle data stream, end-effector torque data stream, and motor current data stream collected by the medical service robot during contact tasks to obtain a spatiotemporal telemetry feature matrix. It should be noted that when the medical service robot performs contact tasks such as rehabilitation traction, the sampling frequencies and spatial coordinate references of the three heterogeneous sensors—joint encoders, end-effector six-dimensional torque sensors, and motor current sensors—are different, and the original signals are mixed with transient noise introduced by electromagnetic interference and electromechanical coupling. Unprocessed heterogeneous data will not be accurately consumed by the subsequent twin engine due to temporal phase misalignment and noise superposition. Based on this, the technical solution of this application first performs spatiotemporal phase alignment and Kalman filtering denoising on the joint angle data stream, end-effector torque data stream, and motor current data stream collected by the medical service robot during contact tasks to obtain a spatiotemporal telemetry feature matrix. Through the above processing, a structured feature input with unified spatiotemporal reference and controlled noise can be provided for subsequent dual-branch hypothesis testing simulation.

[0026] More specifically, in a concrete example of this application, a time phase alignment operation is first performed on the joint encoder angle data stream, the end effector six-dimensional torque data stream, and the motor current data stream, using a global microsecond-level master clock as a unified time reference. Due to the difference in hardware sampling frequencies among the three types of sensors—the joint encoder outputs angle values ​​at 1kHz while the end effector torque sensor outputs six-dimensional force components at 500Hz—the sampling points of each channel are not naturally aligned under the same microsecond-level timestamp. Therefore, using the master clock time axis as a reference, a cubic spline interpolation algorithm is used to compensate for the channels with lower sampling frequencies, strictly mapping the data points of all channels to a unified time scale. After completing the time axis alignment, a homogeneous spatial coordinate transformation is further performed on the end effector torque data, transforming it from the end effector's local coordinate system to the robot's absolute base coordinate system, ensuring that the three data streams maintain spatial reference consistency, ultimately yielding aligned sensor data.

[0027] Furthermore, based on the statistical characteristics of the aligned sensor data within the local historical time window, adaptive Kalman filtering is performed for noise reduction. During rehabilitation traction, the mechanical response generated when the robot's end effector contacts the patient's limb exhibits time-varying characteristics. The rate of change of the torque signal amplitude at the moment of contact differs from that in the non-contact phase. If a standard Kalman filter with a fixed noise covariance parameter is used, it is prone to over-smoothing the true high-frequency mechanical response during the contact transition phase. Therefore, before each frame filtering iteration, the sliding variance of the aligned sensor data within a local historical time window of approximately 10 milliseconds is calculated, and the diagonal elements in the observation noise covariance matrix are dynamically updated using this variance value. This tightens the noise margin in stable signal regions to fully suppress interference, and relaxes the noise margin in rapidly changing signal regions to preserve the true physical abrupt changes. Based on this, Kalman filtering state prediction and observation updates are performed frame by frame. The update process for the posterior state estimation is as follows:

[0028]

[0029] in, Let be the posterior optimal state estimation vector of the k-th frame after filtering and correction. These are prior predictions derived from the state transition equation. Let Kalman gain be the Kalman gain matrix for the k-th frame. Let be the actual observation vector obtained from the aligned sensor data in the k-th frame. This is the observation model matrix that maps the state space to the observation space. After frame-by-frame iterative filtering, the denoised sensor data is output.

[0030] After obtaining the denoised sensor data, the feature matrix is ​​concatenated and reconstructed within a preset time window. A fixed-length time window is set, and within the time interval covered by the window, data slices of three physical channels—joint angle subarray, motor current subarray, and end torque subarray—are extracted from the denoised sensor data. These three subarrays are then horizontally concatenated along the column direction, i.e., the feature dimension direction, to form a high-dimensional matrix with a row dimension equal to the number of time sampling points and a column dimension equal to the sum of the number of all physical channels. Subsequently, principal component analysis is used to perform dimensionality reduction on this high-dimensional matrix to remove redundant feature components, retaining the top principal components whose cumulative variance contribution rate reaches a preset threshold (e.g., 95%). Finally, the spatiotemporal telemetry feature matrix is ​​obtained, which serves as the unified input for subsequent dual-branch digital twin simulations.

[0031] In the aforementioned digital twin-driven medical service robot control and safety monitoring system 100, the dual-branch digital twin simulation module 120 is used to inject the spatiotemporal telemetry feature matrix in parallel into the dual positive simulation branches with mutually exclusive boundary conditions in the digital twin engine for hypothesis testing-based dynamic deduction to obtain the mechanical degradation simulation tensor and the physiological disturbance simulation tensor. It should be noted that, given that the torque fluctuation information carried by the spatiotemporal telemetry feature matrix is ​​a composite result of the robot's internal mechanical state and the external patient's physiological response, during rehabilitation traction, the early tooth surface wear of the joint harmonic reducer causing micro-amplitude vibrations in the transmission chain and the protective muscle spasms caused by the patient's discomfort will both cause similar abnormal surges in the end torque channel. However, the current integration depth of digital twin technology with control and safety monitoring is insufficient; relying solely on the passive state mirror of a single twin cannot determine the excitation source of this deviation residual from a dynamic analytical perspective. Based on this, the technical solution of this application further injects the spatiotemporal telemetry feature matrix into the dual positive simulation branches with mutually exclusive boundary conditions in the digital twin engine for hypothesis testing-based dynamic deduction to obtain the mechanical degradation simulation tensor and the physiological disturbance simulation tensor. Through the above processing, independent expected dynamic response benchmarks under two mutually exclusive assumptions of internal mechanical degradation and external physiological disturbance can be constructed in virtual space, providing a calculable reference for accurately locating the source of disturbance through residual comparison.

[0032] Figure 3 This is a block diagram of a dual-branch digital twin simulation module in a digital twin-driven medical service robot control and safety monitoring system according to an embodiment of this application. Figure 3As shown, the dual-branch digital twin simulation module 120 includes: a state parameter extraction unit 121, used to decompose the spatiotemporal telemetry feature matrix to extract the kinematic state vector and dynamic state vector at the current moment, thereby obtaining the observed state parameter vector; a mechanical degradation simulation unit 122, used to inject the observed state parameter vector into the first simulation branch for forward dynamic deduction, thereby obtaining the mechanical degradation simulation tensor; and a physiological disturbance simulation unit 123, used to inject the observed state parameter vector into the second simulation branch for forward dynamic deduction, thereby obtaining the physiological disturbance simulation tensor.

[0033] In the aforementioned digital twin-driven medical service robot control and safety monitoring system 100, the state parameter extraction unit 121 is used to decompose the spatiotemporal telemetry feature matrix to extract the kinematic state vector and dynamic state vector at the current moment, thereby obtaining the observed state parameter vector. It should be noted that, since the spatiotemporal telemetry feature matrix is ​​a high-dimensional composite structure formed by horizontally cascading joint angle subarrays, current subarrays, and torque subarrays along the column directions, the physical dimensions and mechanical meanings corresponding to each column are different, making it unsuitable as direct input for solving the Newton-Euler dynamics equations. Based on this, the technical solution of this application further decomposes the spatiotemporal telemetry feature matrix to extract the kinematic state vector and dynamic state vector at the current moment, thereby obtaining the observed state parameter vector. Through the above processing, the high-dimensional cascaded matrix can be restored to a standardized parameter input with clear physical dimensions that can be directly called by the forward dynamics equations.

[0034] More specifically, in a concrete example of this application, the spatiotemporal telemetry feature matrix is ​​decomposed dimensionally according to column index intervals. Based on the preset channel arrangement order during cascading, column intervals corresponding to the joint angle subarray and the end effector torque subarray are extracted respectively. After obtaining the joint angle subarray, the joint angle values ​​corresponding to the current moment (i.e., the last frame of the sliding window) are extracted to form an N-dimensional joint angle column vector, where N is the number of mechanical degrees of freedom of the medical service robot. In the rehabilitation traction task, this angle column vector reflects the spatial geometric configuration of each joint of the robot in the current traction posture. On this basis, a first-order numerical derivative operation is performed on the historical sequence of joint angles within the sliding window to obtain the angular velocity values ​​of each joint at the current moment, forming an N-dimensional joint angular velocity column vector. This vector represents the instantaneous motion rate of each joint during the robot's traction motion. Simultaneously, the six-dimensional torque observation value at the current moment is extracted from the end effector torque subarray. This torque value reflects the actual interactive load borne by the robot's end effector at the interface between the robot's end effector and the patient's limb. The joint angle column vector, joint angular velocity column vector, and six-dimensional end moment observation values ​​are vertically concatenated to construct the observation state parameter vector:

[0035]

[0036] in, The vector of observed state parameters generated at time t. This is the N-dimensional joint angle column vector extracted from the joint angle subarray. This is the N-dimensional column vector of joint angular velocities obtained through numerical differentiation. This is the 6-dimensional end moment observation vector extracted from the moment subarray, with dimensions of [missing information]. It covers all the initial kinematic and dynamic conditions required for the subsequent forward dynamic derivation of the two branches.

[0037] In the aforementioned digital twin-driven medical service robot control safety monitoring system 100, the mechanical degradation simulation unit 122 is used to inject the observed state parameter vector into the first simulation branch for forward dynamic deduction to obtain the mechanical degradation simulation tensor. It should be noted that since the torque anomaly contained in the observed state parameter vector may originate from the mechanical degradation of the robot's internal joint transmission chain, it is necessary to construct an independent simulation branch in the virtual space that assumes a completely normal external environment while only allowing the internal mechanical parameters to deteriorate, in order to generate the expected dynamic response under this assumption. Based on this, the technical solution of this application further injects the observed state parameter vector into the first simulation branch for forward dynamic deduction to obtain the mechanical degradation simulation tensor. Through the above processing, a theoretical torque response benchmark completely attributed to the evolution of internal mechanical faults can be obtained, providing a first hypothetical reference for subsequent residual comparison with physical observation values.

[0038] More specifically, in a concrete example of this application, the joint angle column vector and joint angular velocity column vector in the observed state parameter vector are injected into the digital twin engine as the initial kinematic conditions of the first simulation branch. This branch constrains the external interaction environment between the robot's end effector and the patient's limb to a state of rated constant stiffness, assuming that the patient's end does not exhibit any abnormal muscle spasms or resistance during the entire rehabilitation traction process, and the external contact force remains at the rated ideal interaction torque. Simultaneously, the friction coefficient and gear backlash of the robot's internal joint transmission chain are set as nonlinear time-varying variables that deteriorate over time, simulating the mechanical degradation evolution process of gradually increasing harmonic reducer tooth surface wear and continuously increasing Coulomb friction and viscous friction. Under the above boundary constraints, the current kinematic state is solved and deduced based on the Newton-Euler forward dynamic equations, calculating the torque response sequence that each joint should exhibit under this mechanical degradation assumption:

[0039]

[0040] in, This is the expected joint torque vector calculated under the assumption of mechanical degradation. Let be the symmetric positive definite inertia matrix of the robot under the current joint configuration. The joint angular acceleration vector. The matrix represents the centrifugal force and the Coriolis force. The joint angular velocity vector. This is the gravity compensation vector. The constructed time-varying mechanical degradation friction compensation function includes Coulomb friction terms and viscous friction terms that continuously increase due to gear wear. Let be the transpose of the Jacobian matrix in robot space. Let be the rated ideal interaction torque of the external contact environment. The torque response sequences obtained above are stacked along the time sliding window dimension to finally obtain the mechanical degradation simulation tensor.

[0041] In the aforementioned digital twin-driven medical service robot control and safety monitoring system 100, the physiological disturbance simulation unit 123 is used to inject the observed state parameter vector into the second simulation branch for forward dynamic deduction to obtain the physiological disturbance simulation tensor. It should be noted that since the torque anomaly contained in the observed state parameter vector may also originate from sudden muscle spasms and protective resistance caused by pain or discomfort during rehabilitation traction, it is necessary to construct an independent simulation branch in virtual space that assumes the robot's internal mechanical state is completely ideal and only allows external physiological disturbances to exist, in order to generate the expected dynamic response under this assumption. Based on this, the technical solution of this application further injects the observed state parameter vector into the second simulation branch for forward dynamic deduction to obtain the physiological disturbance simulation tensor. Through the above processing, a theoretical torque response benchmark completely attributed to external patient physiological disturbances can be obtained, forming a dual-assumption reference system with mutually exclusive boundary conditions with the mechanical degradation simulation tensor of the first simulation branch.

[0042] More specifically, in a concrete example of this application, the joint angle column vector and joint angular velocity column vector in the observed state parameter vector are injected into the digital twin engine as the kinematic initial conditions of the second simulation branch. This branch is strictly mutually exclusive with the first simulation branch in its boundary condition settings, that is, it forces all mechanical parameters of the robot's internal joint transmission chain to be in an absolutely ideal, wear-free state, with the friction coefficient and gear backlash remaining at their factory calibration values ​​without any degradation. Simultaneously, a time-varying nonlinear spring-damped model is attached at the contact interface between the end effector and the virtual patient's limb to simulate the protective muscle spasm behavior that suddenly occurs in the patient during passive traction due to joint pain or muscle fatigue. The stiffness and damping parameters in this model dynamically change over time to reflect the physiological evolution of the spasm from triggering to peak and then to decay. Under the above boundary condition constraints, the current kinematic state is forward-deduced based on the rigid body dynamics equations to calculate the torque response sequence that each joint should exhibit under this physiological perturbation assumption:

[0043]

[0044] in, This is the expected joint torque vector calculated under the physiological perturbation assumption. , , These are the inertia matrix, centrifugal force and Coriolis force matrices, and gravity compensation vector, respectively, assuming the internal mechanical transmission is in an absolutely ideal, wear-free state. Let be the transpose of the Jacobian matrix in robot space. This is a time-varying nonlinear muscle stiffness diagonal matrix used to simulate sudden spasms in patients. This is a time-varying viscous damping diagonal matrix used to simulate the spastic resistance of patients. and These are the displacement and velocity deviation vectors between the end effector and the contact area of ​​the patient's limb, respectively. The torque response sequences obtained from the above solutions are stacked along the time sliding window dimension to finally obtain the physiological perturbation simulation tensor.

[0045] In the aforementioned digital twin-driven medical service robot control safety monitoring system 100, the disturbance source decoupling module 130 is used to perform spatiotemporal dual-domain residual analysis and precise decoupling of the disturbance source from the spatiotemporal telemetry feature matrix to the mechanical degradation simulation tensor and the physiological disturbance simulation tensor, thereby obtaining the decoupled disturbance source vector. It should be noted that, given that the mechanical degradation simulation tensor and the physiological disturbance simulation tensor represent theoretical torque responses under two mutually exclusive assumptions, and the physical reality observed value carried by the spatiotemporal telemetry feature matrix is ​​the result of the combined action of one or both of the aforementioned potential disturbance sources, in rehabilitation traction scenarios, the physically closed mechanical transmission chain formed between the robot's end effector and the patient's body makes the representations of the two types of disturbances highly similar in the end-torque channel. A quantitative residual comparison mechanism is needed to determine which assumption branch the physical reality matches more closely, thereby pinpointing the actual origin of abnormal torque fluctuations. Based on this, the technical solution of this application further performs spatiotemporal dual-domain residual analysis and precise decoupling of the disturbance source from the spatiotemporal telemetry feature matrix to the mechanical degradation simulation tensor and the physiological disturbance simulation tensor to obtain the decoupled disturbance source vector. Through the above processing, the previously indistinguishable composite torque deviation can be attributed to specific disturbance source categories and its deviation intensity can be quantified, providing a clear source determination basis for the accurate formulation of downstream risk assessment and intervention strategies.

[0046] Figure 4 This is a block diagram of a disturbance source decoupling module in a digital twin-driven medical service robot control safety monitoring system according to an embodiment of this application. Figure 4As shown, the disturbance source decoupling module 130 includes: a mechanical residual analysis unit 131, used to calculate the mechanical residual distance between the spatiotemporal telemetry feature matrix and the mechanical degradation simulation tensor under the optimal regularized path, using the spatiotemporal telemetry feature matrix as a reference sequence; a physiological residual analysis unit 132, used to calculate the physiological residual distance between the spatiotemporal telemetry feature matrix and the physiological disturbance simulation tensor under the optimal regularized path, using the spatiotemporal telemetry feature matrix as a reference sequence; and a disturbance source screening and decoupling unit 133, used to perform competitive screening of the mechanical residual distance and the physiological residual distance to obtain the decoupled disturbance source vector.

[0047] In the aforementioned digital twin-driven medical service robot control and safety monitoring system 100, the mechanical residual analysis unit 131 is used to calculate the mechanical residual distance between the spatiotemporal telemetry feature matrix and the mechanical degradation simulation tensor under the optimal regularization path, using the spatiotemporal telemetry feature matrix as a reference sequence. It should be noted that since the spatiotemporal telemetry feature matrix and the mechanical degradation simulation tensor originate from two independent paths—physical sensing acquisition and virtual dynamics deduction—there may be local phase drift on the time axis caused by the inconsistency between the simulation step size and the physical sampling period. Therefore, a distance metric method that can tolerate temporal scaling is needed to calculate the degree of deviation between them. Based on this, the technical solution of this application further uses the spatiotemporal telemetry feature matrix as a reference sequence to calculate the mechanical residual distance between the spatiotemporal telemetry feature matrix and the mechanical degradation simulation tensor under the optimal regularization path. Through the above processing, a quantitative deviation index between the physical reality and the mechanical degradation hypothesis in the spatiotemporal dual domains can be obtained, providing a first reference value for subsequent comparison with the physiological residual distance.

[0048] More specifically, in a concrete example of this application, the spatiotemporal telemetry feature matrix is ​​used as the reference sequence, and the mechanical degradation simulation tensor is used as the sequence to be compared. Both are input into a dynamic time warping algorithm for residual analysis. First, a two-dimensional cumulative distance matrix is ​​constructed. The row indices of this matrix correspond to the time sampling frames of the spatiotemporal telemetry feature matrix, and the column indices correspond to the simulation inference frames of the mechanical degradation simulation tensor. Each element in the matrix stores the Euclidean distance between two corresponding frames. Based on the cumulative distance matrix, starting from the initial node, a warping path that minimizes the sum of cumulative distances is searched step-by-step according to local path continuity constraints. This path allows the two sequences to be nonlinearly scaled and aligned on the time axis, thereby eliminating local phase shifts caused by differences between the simulation step size and the physical sampling period. Along this optimal warping path, the Euclidean distances at all aligned nodes on the path are weighted and summed to obtain the mechanical residual distance:

[0049]

[0050] in, To calculate the mechanical residual distance, The set of optimal paths obtained by dynamic time warping search consists of a series of aligned node pairs. composition, This represents the total length of the optimal path, which is the total number of nodes after alignment. Let be the weight coefficient of the k-th node on the path. For the spatiotemporal telemetry feature matrix at the time index The feature vector at that location, For mechanical degradation simulation tensor in time index The feature vector at that location, This is the L2 norm. The smaller this mechanical residual distance value, the higher the degree of agreement between the actual physical state and the mechanical degradation assumption, meaning that the current torque anomaly is more likely to be caused by internal mechanical failure.

[0051] In the aforementioned digital twin-driven medical service robot control safety monitoring system 100, the physiological residual analysis unit 132 is used to calculate the physiological residual distance between the spatiotemporal telemetry feature matrix and the physiological disturbance simulation tensor under the optimal regularization path, using the spatiotemporal telemetry feature matrix as a reference sequence. It should be noted that since the mechanical residual distance only reflects the degree of deviation between the actual physical state and the mechanical degradation hypothesis, it cannot independently determine the cause of the disturbance source. It is necessary to simultaneously obtain the degree of deviation between the actual physical state and the physiological disturbance hypothesis in order to determine which hypothesis branch is more consistent with the current observation in subsequent competitive comparisons. Based on this, the technical solution of this application further uses the spatiotemporal telemetry feature matrix as a reference sequence to calculate the physiological residual distance between the spatiotemporal telemetry feature matrix and the physiological disturbance simulation tensor under the optimal regularization path. Through the above processing, a quantitative deviation index between the actual physical state and the physiological disturbance hypothesis in the spatiotemporal dual domain can be obtained, which, together with the mechanical residual distance, constitutes a dual-path residual input, providing a complete comparison basis for subsequent competitive screening of disturbance sources.

[0052] More specifically, in a concrete example of this application, the spatiotemporal telemetry feature matrix is ​​used as the reference sequence, and the physiological perturbation simulation tensor is used as the sequence to be compared. The same dynamic time warping algorithm as used for calculating the mechanical residual distance is employed for residual analysis. A two-dimensional cumulative distance matrix is ​​constructed between the spatiotemporal telemetry feature matrix and the physiological perturbation simulation tensor. Each element in the matrix stores the Euclidean distance between the corresponding frame of the reference sequence and the corresponding frame of the simulation sequence. On this cumulative distance matrix, an optimal warping path is searched according to the local path continuity constraint to minimize the sum of the cumulative distances. This path also allows for nonlinear scaling and alignment of the two sequences on the time axis to eliminate the phase shift between physical sampling and simulation. Along this optimal warping path, the Euclidean distances at all aligned nodes are weighted and summed to obtain the physiological residual distance.

[0053]

[0054] in, The calculated physiological residual distance, This represents the set of optimal regularized paths that can be solved independently for the physiological perturbation branch. This is the total length of the path. Let be the weight coefficient of the k-th node on the path. For the spatiotemporal telemetry feature matrix at the time index The feature vector at that location, Simulation tensor for physiological perturbation in time index The feature vector at that location, The value represents the L2 norm. The smaller this physiological residual distance value, the higher the degree of agreement between the actual physical state and the physiological perturbation hypothesis, meaning that the current torque abnormality is more likely to be caused by sudden muscle spasms or protective resistance generated by the patient during rehabilitation traction.

[0055] In the aforementioned digital twin-driven medical service robot control safety monitoring system 100, the disturbance source screening and decoupling unit 133 is used to perform competitive screening of mechanical residual distance and physiological residual distance to obtain a decoupled disturbance source vector. It should be noted that, given that the mechanical residual distance and physiological residual distance quantify the degree of deviation between the physical reality and two mutually exclusive hypotheses, respectively, these two independent scalar values ​​cannot directly indicate the source of the current torque anomaly. A competitive probabilistic discrimination mechanism is needed to compare and distinguish the two residuals, locking the hypothesis branch with the smaller residual (i.e., higher consistency) as the actual origin of the disturbance, and integrating the judgment result with the corresponding confidence strength into a structured vector expression. Based on this, the technical solution of this application further performs competitive screening of mechanical residual distance and physiological residual distance to obtain a decoupled disturbance source vector. Through the above processing, the previously indistinguishable composite torque deviation can be clearly attributed to either internal mechanical deterioration or external patient physiological disturbance, and simultaneously carry the disturbance source type identifier and deviation intensity information in vector form, providing structured decision input for downstream risk mapping and intervention strategy matching.

[0056] Figure 5 This is a block diagram of a disturbance source screening and decoupling unit in a digital twin-driven medical service robot control safety monitoring system according to an embodiment of this application. Figure 5As shown, the disturbance source screening and decoupling unit 133 includes: an independent channel activation energy calculation subunit 1331, used to perform residual-driven independent channel activation energy calculation on the mechanical residual distance and the physiological residual distance to obtain the mechanical activation energy and the physiological activation energy; a cross-source coupling degree estimation subunit 1332, used to estimate the cross-source coupling degree of the mechanical activation energy and the physiological activation energy to obtain the cross-source coupling degree; and a soft contribution decomposition subunit 1333, used to perform soft contribution decomposition of the mechanical activation energy and the physiological activation energy based on the cross-source coupling degree to obtain the decoupled disturbance source vector.

[0057] In the aforementioned digital twin-driven medical service robot control safety monitoring system 100, the independent channel activation energy calculation subunit 1331 is used to perform residual-driven independent channel activation energy calculation on the mechanical residual distance and physiological residual distance to obtain the mechanical activation energy and physiological activation energy. It should be noted that, given that existing methods, based on the negative exponential probability discriminant function of the Softmax structure, force the sum of the two probabilities to be always equal to 1 during normalization, this complementary constraint implicitly assumes that the two disturbance sources are completely mutually exclusive at the mathematical level. However, in the real clinical scenario of contact rehabilitation traction, mechanical faults and patient physiological disturbances are not naturally mutually exclusive, but rather have a causal coupling and symbiotic relationship. When the values ​​of the two residual distances are close, the winning probability output by Softmax is only slightly greater than 0.5, resulting in extremely low decision confidence. However, the Boolean indicator function still forcibly attributes it to a single source, masking the actual uncertainty of the decision. Based on this, the technical solution of this application further performs residual-driven independent channel activation energy calculation on the mechanical residual distance and physiological residual distance to obtain the mechanical activation energy and physiological activation energy. Through the above processing, the complementary constraints imposed by Softmax normalization can be broken, making the two activation energies mathematically independent and both able to obtain high values ​​simultaneously, laying the prerequisite for accurate estimation of cross-source coupling degree in the future.

[0058] More specifically, in a particular example of this application, negative exponential transformations are applied independently to the mechanical residual distance and the physiological residual distance, mapping the unbounded distance metric to a bounded activation energy value. The mechanical activation energy is obtained after applying a negative exponential transformation to the mechanical residual distance, and its calculation process is as follows:

[0059]

[0060] in, For mechanical activation energy, This is a sensitivity adjustment hyperparameter used to control the amplification effect of residual differences. The mechanical residual distance. The physiological activation energy is obtained by applying a negative exponential transformation to the physiological residual distance; the calculation process is as follows:

[0061]

[0062] in, As physiological activation energy, The physiological residual distance is represented by the negative exponential transformation. The physical meaning of this transformation is that a smaller residual corresponds to a higher activation energy, indicating a stronger fit between the hypothetical branch and the actual physical state. Unlike Softmax normalization, the two activation energies are mathematically independent and unconstrained, and both can simultaneously achieve high values. This precisely corresponds to the objective fact that both hypotheses under coupled symbiosis highly match physical reality. In rehabilitation traction scenarios, when early tooth surface wear occurs in the robot's joint harmonic reducer, the micro-amplitude high-frequency vibrations generated by the transmission chain are transmitted to the skin and muscle tissue at the patient's contact point via the end effector, inducing protective muscle spasms due to discomfort. This means that internal mechanical deterioration directly triggers external physiological disturbances. At this time, both the mechanical and physiological residual distances are at low levels. After the negative exponential transformation, both the mechanical and physiological activation energies show high values, accurately reflecting the physical fact that both sources exist simultaneously, rather than being forcibly attributed to one of them.

[0063] In the aforementioned digital twin-driven medical service robot control safety monitoring system 100, the cross-source coupling degree estimation subunit 1332 is used to estimate the cross-source coupling degree of mechanical activation energy and physiological activation energy to obtain the cross-source coupling degree. It should be noted that, given that two independent activation energies can only reflect the degree of agreement between their respective channels and the actual physical state, they cannot characterize whether two disturbance sources exist simultaneously and mutually excite each other. Relying solely on symmetry measurement carries a potential risk of misjudgment. Specifically, under normal robot operation, the residual distances of both paths are large, corresponding to extremely low levels of both activation energies. In this case, the symmetry ratio of the activation energies may accidentally approach 1 due to numerical noise, leading to the normal operation state being misjudged as coupled co-existing. Based on this, the technical solution of this application further estimates the cross-source coupling degree of mechanical and physiological activation energy to obtain the cross-source coupling degree. Through the above processing, both dual-source symmetry and signal effectiveness can be simultaneously considered, accurately capturing coupled co-existing states and suppressing noise coupling misjudgments under low signal-to-noise ratio conditions.

[0064] More specifically, in a concrete example of this application, the ratio of the geometric mean to the arithmetic mean of the mechanical activation energy and the physiological activation energy is first calculated as a symmetry measure. According to the mathematical properties of the mean inequality, this ratio reaches its maximum value of 1 if and only if the two activation energies are completely equal; any deviation from either will cause the ratio to decrease. Therefore, it is naturally suitable as a measure of the degree of coexistence of two sources. In a contact rehabilitation traction scenario, when mechanical vibration induces spasm in the patient, and the spasm impact in turn exacerbates gear wear, the two activation energies exhibit a characteristic pattern of both being high and having similar values. At this time, the geometric-arithmetic mean ratio approaches 1. However, in the case of a single-source failure, only one activation energy is high, while the other is low, and the ratio decreases accordingly.

[0065] Furthermore, to eliminate noise interference in the low activation energy region during normal operation, a hyperbolic tangent saturation gating function based on the total activation energy of both sources is introduced. When the total activation energy is extremely low, corresponding to the background noise region during normal operation, the output of the gating function approaches zero, forcibly suppressing the coupling degree to near zero to inhibit noise coupling misjudgment; when the total activation energy exceeds the effective signal threshold, the gating function quickly saturates to 1, no longer suppressing the coupling degree. Multiplying the symmetry metric factor by the saturation gating factor yields the cross-source coupling coefficient that simultaneously considers dual-source symmetry and signal effectiveness. The calculation process is as follows:

[0066]

[0067] in, This is the cross-source coupling coefficient, with a value ranging from 0 to 1. 0 represents a purely independent single-source fault, while 1 represents a fully coupled and symbiotic fault. For mechanical activation energy, As physiological activation energy, The geometric-arithmetic mean ratio measures the symmetry of the dual-path activation energy if and only if The maximum value is 1. The saturation-gated slope hyperparameter controls the sensitivity of the total activation energy transitioning from the noise region to the effective signal region. It is a hyperbolic tangent saturation gate function that maps the total activation energy to the interval [0,1). When the total activation energy is low, the output approaches zero to suppress noise coupling misjudgment.

[0068] In a specific rehabilitation traction scenario, assuming the robot's joint harmonic reducer is in the early tooth surface wear stage, the micro-vibrations generated by the transmission chain are transmitted to the patient's forearm muscles via the end effector, inducing protective muscle spasms. At this time, both the mechanical residual distance and the physiological residual distance are at a low level. After negative exponential transformation, the mechanical activation energy and physiological activation energy are 0.82 and 0.78, respectively. Both values ​​are close and at a high level. The geometric-arithmetic mean ratio is 0.9997, approaching 1. Meanwhile, the total activation energy of the two paths is 1.60. After hyperbolic tangent gating, the saturated output is close to 1. Finally, the cross-source coupling coefficient... The calculated result is 0.9997, close to 1, accurately reflecting the current physical fact that the current situation is a dual-source coupling and symbiosis of mechanical degradation and physiological disturbance. However, under normal robot operation, the residual distances of both paths are relatively large, with corresponding mechanical activation energies of 0.005 and physiological activation energies of 0.006. Although the geometric-arithmetic mean ratio of the two is 0.9985, the total activation energy of the two paths is only 0.011. After hyperbolic tangent gating, the output approaches zero, and the cross-source coupling coefficient is ultimately suppressed to 0.055, effectively preventing the normal operation state from being misjudged as a coupled symbiotic state.

[0069] In the aforementioned digital twin-driven medical service robot control and safety monitoring system 100, the soft contribution decomposition subunit 1333 is used to perform soft contribution decomposition based on cross-source coupling degree to couple and modulate mechanical activation energy and physiological activation energy to obtain a decoupled disturbance source vector. It should be noted that, given that the cross-source coupling coefficient has quantified the causal coupling degree between mechanical degradation and physiological disturbance, but the contribution weights of the two disturbance sources have not been redistributed accordingly, and the Softmax normalization in existing methods, in the coupled state, forcibly attributes a composite fault that is essentially a superposition of two sources to a single source, the root cause being that its output structure only allows binary representations that are either one or the other. Based on this, the technical solution of this application further performs soft contribution decomposition based on cross-source coupling degree to couple and modulate mechanical activation energy and physiological activation energy to obtain a decoupled disturbance source vector. Through the above processing, a linear mixture between independent normalized proportion and equal distribution can be performed using the cross-source coupling coefficient as the interpolation factor. This maintains backward compatibility with the original solution in single-source failure scenarios and acknowledges the simultaneous existence and equal contribution of two sources in coupled co-ecosystem scenarios.

[0070] More specifically, in a concrete example of this application, a linear mixture is performed between independent normalized proportions and equal allocation using the cross-source coupling coefficient as an interpolation factor to achieve soft contribution decomposition. When the coupling coefficient approaches 0, the mixing result degenerates into a standard normalized proportion, which is strictly equivalent to the original Softmax behavior, ensuring backward compatibility in single-source fault scenarios; when the coupling coefficient approaches 1, the contribution weights of both paths are pulled towards 0.5, acknowledging the simultaneous existence and equal contribution of the two sources, thereby avoiding forcibly attributing composite faults to a single source. The calculation process of the mechanical contribution weight is as follows:

[0071]

[0072] in, The soft contribution weight of the mechanically degraded channel after coupling modulation. For cross-source coupling coefficients, For mechanical activation energy, As physiological activation energy, These are the weighting coefficients for the independent decision components; the stronger the coupling, the weaker the influence of the independent decisions. The equal contribution terms of the coupled co-generated components are considered; the stronger the coupling, the more equal the weights of the two sources. The calculation process for the physiological contribution weights is as follows:

[0073]

[0074] in, The soft contribution weights of the physiological perturbation channels are defined by coupling modulation, and the other symbols are defined as above. After obtaining the dual-channel soft contribution weights, they are vertically concatenated with the cross-source coupling coefficients to construct a three-dimensional structured column vector. The output is an enhanced decoupled perturbation source vector, and its construction process is as follows:

[0075]

[0076] in, To decouple the disturbance source vector, the first dimension The soft contribution weights for the mechanical degradation channel, the second dimension The soft contribution weights for physiological perturbation channels, the third dimension It serves as a cross-source coupling coefficient, used for downstream risk mapping to identify coupled ecosystems and trigger dual-source joint intervention strategies.

[0077] For example, continuing the coupling state where early tooth surface wear of the aforementioned harmonic reducer induces protective muscle spasms in patients, the cross-source coupling coefficient at this time... Approaching 1, both mechanical and physiological activation energies are high and similar in value. After soft contribution decomposition, the weights of both pathways approach 0.5, accurately reflecting the co-existing coupling of the two sources with comparable contributions. The high coupling coefficient carried in the third dimension of the final decoupled perturbation source vector will trigger downstream risk assessment to match the dual-source joint intervention strategy, avoiding the clinical safety risks caused by neglecting the treatment of the other source while only developing intervention measures for a single source. In the single mechanical failure scenario, only mechanical activation energy is high while physiological activation energy is low. The cross-source coupling coefficient is suppressed to near zero by saturation gating. After soft contribution decomposition, the mechanical channel weight is absolutely dominant while the physiological channel weight approaches zero. The output behavior is strictly equivalent to the original normalized result, verifying complete backward compatibility in the single-source failure scenario without introducing any performance regression risk.

[0078] In the aforementioned digital twin-driven medical service robot control safety monitoring system 100, the dynamic safety assessment module 140 is used to map and inject the abnormal source type identifier and deviation intensity value in the decoupled disturbance source vector into a pre-trained safety accident knowledge graph based on a graph neural network to obtain a dynamic safety assessment report. It should be noted that since the decoupled disturbance source vector carries the disturbance source contribution weight and coupling information in the form of a numerical column vector, it has not yet been correlated with the evolutionary patterns of historical safety accidents, making it impossible to directly determine the probability of the current situation triggering a clinical safety accident within a future time window and its corresponding risk management level. Based on this, the technical solution of this application further uses a graph neural network to map and inject the abnormal source type identifier and deviation intensity value in the decoupled disturbance source vector into a pre-trained safety accident knowledge graph to obtain a dynamic safety assessment report. Through the above processing, the current decoupled situation and historical accident evolution paths can be topologically correlated and retrieved, outputting a structured assessment report containing risk probability scores and discretized management levels, providing a decision-making basis for the accurate matching of downstream intervention strategies.

[0079] More specifically, in a concrete example of this application, the decoupled perturbation source vector is first extracted by index slicing. The decoupled perturbation source vector is a three-dimensional structured column vector, where the first and second dimensions are the soft contribution weights of the mechanical degradation channel and the physiological perturbation channel, respectively, and the third dimension is the cross-source coupling coefficient. The vector is sliced ​​by row index, and the channel identifier corresponding to the larger of the soft contribution weights is extracted as the anomaly source type identifier. Simultaneously, the soft contribution weight value of this channel, together with the cross-source coupling coefficient, constitutes the deviation strength value, preparing the query entity for subsequent map retrieval.

[0080] Furthermore, the anomaly source type identifier and deviation intensity value are concatenated to form an initial node feature vector, which is then injected into a pre-trained and deployed medical service robot concurrent safety incident knowledge graph for risk probability density retrieval. This knowledge graph is constructed based on a graph attention network. Nodes in the graph represent snapshots of historical safety incidents. The initial features of each node include the disturbance source type, deviation intensity, coupling coefficient, and incident severity level at the time of the incident. Edges represent the causal evolution relationship between incidents, and edge weights represent the strength of the causal association. This graph attention network contains L message passing layers (e.g., L=3), each employing a multi-head attention aggregation mechanism. The node feature update process in the l-th layer is as follows: the features of the query node and its neighboring nodes are concatenated, and then linearly transformed and activated with LeakyReLU to calculate the attention coefficient. The attention coefficient is then used as the weight to perform weighted aggregation of the neighboring node features. The concatenated initial node feature vector is embedded as the 0th layer of the query node, triggering the message passing mechanism of the graph attention network. L-layer feature aggregation and weight updates are performed between the query node and adjacent historical incident nodes in the graph, evaluating the position of the current disturbance potential in the graph topology space and its trend towards evolving into a catastrophic incident node. After L-layer aggregation, the node embedding features of the final layer are fed into a fully connected layer and a sigmoid activation function to calculate the probability density of the current perturbation potential evolving into a clinical safety accident within the next 500 milliseconds, outputting a future risk prediction score. During the training phase, training samples are constructed based on abnormal events and safety accident cases recorded in the historical operation logs of the medical service robot. Whether an accident actually occurred is used as the binary classification label, and the parameters of the graph attention network are optimized end-to-end using a binary cross-entropy loss function. In the rehabilitation traction scenario, when the third-dimensional cross-source coupling coefficient of the decoupled perturbation source vector is high, the query node will have a strong association with historical accident nodes of the dual-source coupling type in the graph. The retrieved risk probability density is higher than that of the single-source fault situation, reflecting the superposition effect of safety risks under coupled co-ecosystem conditions.

[0081] After obtaining the future risk prediction score, it is discretized and graded based on a step-threshold piecewise function. Two dividing points are set: a high-risk circuit breaker probability threshold and an early warning probability threshold. When the future risk prediction score reaches or exceeds the high-risk circuit breaker probability threshold, it is classified as emergency blocking level (i.e., level 3); when the score is between the early warning probability threshold and the high-risk circuit breaker probability threshold, it is classified as compliant intervention level (i.e., level 2); when the score is below the early warning probability threshold, it is classified as routine monitoring level (i.e., level 1). The determined risk level, future risk prediction score, and determination timestamp are structured and encapsulated to obtain a dynamic security assessment report.

[0082] In the aforementioned digital twin-driven medical service robot control safety monitoring system 100, the intervention strategy verification and control module 150 is used to adaptively match candidate intervention strategies based on the risk level and anomaly source type in the dynamic safety assessment report. Specifically, the candidate intervention action set is reinjected into the digital twin for secondary pre-test verification to eliminate action sequences with secondary injury risks. The optimal strategy after screening is compiled into intervention control instructions and issued to the robot actuator. It should be noted that while the dynamic safety assessment report clearly identifies the risk level and anomaly source type of the current situation, directly issuing emergency control instructions based on the risk level poses a risk of secondary injury. In rehabilitation traction scenarios, if a rigid emergency stop is directly executed due to patient spasm, the robot's high rigidity will cause tearing force on the patient's ligaments or soft tissues at the moment of emergency stop. Existing technologies, when issuing emergency control instructions directly after risk assessment, do not evaluate the safety of the intervention action itself in the digital space, lacking a forward-looking verification step for candidate strategies. Based on this, the technical solution of this application further adaptively matches candidate intervention strategies based on the risk level and anomaly source type in the dynamic safety assessment report. The candidate intervention action set is then reinjected into the digital twin for secondary pre-performance verification to eliminate action sequences with secondary injury risks. The selected optimal strategy is compiled into intervention control instructions and sent to the robot actuator. Through the above processing, collision and stress pre-performance assessments of candidate actions can be completed in virtual space before the intervention instructions actually act on the physical robot, selecting the optimal compliant strategy with global safety and smoothness. This avoids secondary physical damage to the patient due to inappropriate intervention measures, thus closing the entire safety monitoring loop of perception, decoupling, assessment, verification, and control.

[0083] Figure 6 This is a block diagram of the intervention strategy verification and control module in a digital twin-driven medical service robot control safety monitoring system according to an embodiment of this application. Figure 6 As shown, the intervention strategy verification and control module 150 includes: a candidate strategy matching unit 151, used to decode the dynamic safety assessment report and perform two-dimensional lookup table retrieval and parameter matching in the compliant impedance strategy matrix library according to the risk level and anomaly source type to obtain a set of candidate intervention actions; a pre-playing screening unit 152, used to inject the set of candidate intervention actions one by one into the digital twin engine for forward collision and stress distribution pre-playing and optimization screening to obtain the optimal strategy instruction; and a control instruction issuing unit 153, used to compile and encapsulate the optimal strategy instruction into an intervention control instruction compatible with the underlying servo protocol and issue it to the robot actuator.

[0084] In the aforementioned digital twin-driven medical service robot control safety monitoring system 100, the candidate strategy matching unit 151 is used to decode the dynamic safety assessment report and perform a two-dimensional lookup table search and parameter matching in the compliant impedance strategy matrix library based on the risk level and anomaly source type to obtain a set of candidate intervention actions. It should be noted that since the dynamic safety assessment report encapsulates the judgment results such as risk level and anomaly source type in structured data form, the physical intervention requirements corresponding to different combinations of risk levels and anomaly source types are fundamentally different. Therefore, it is necessary to retrieve impedance control parameters matching the current situation from a preset strategy space based on the characteristics of this combination. Based on this, the technical solution of this application further decodes the dynamic safety assessment report and performs a two-dimensional lookup table search and parameter matching in the compliant impedance strategy matrix library based on the risk level and anomaly source type to obtain a set of candidate intervention actions. Through the above processing, the abstract risk assessment results can be transformed into a combination of impedance control parameters with clear physical dimensions, providing a candidate action pool for subsequent twin pre-simulation verification.

[0085] More specifically, in a concrete example of this application, the dynamic security assessment report is first decoded to extract the current discrete security threat level and abnormal disturbance source type. The security threat level includes three discrete values: emergency blocking level, compliant intervention level, and routine monitoring level. The abnormal source type includes three categories: internal mechanical degradation, external physiological disturbance, and dual-source coupling symbiosis. Using the above two dimensions as row and column indices, a two-dimensional lookup table is performed in a pre-built compliant impedance strategy matrix library. Each cell of this strategy matrix library stores one or more pairs of impedance control parameters consisting of a Cartesian space virtual target stiffness diagonal matrix and a virtual target damping diagonal matrix. In rehabilitation traction scenarios, when the risk level is emergency closure and the anomaly source is internal mechanical gear fracture, the retrieved parameter pair is a high-stiffness, high-damping position-locking impedance configuration to quickly fix the joint in its current position to prevent mechanical loss of control. When the risk level is compliant intervention and the anomaly source is external patient spasticity, the retrieved parameter pair is a low-stiffness, zero-force-field compliant following impedance configuration, allowing the robot end effector to passively retract in accordance with the direction of patient spasticity to release contact stress. When the anomaly source is a dual-source coupled co-existence scenario, the retrieved parameter pair simultaneously includes a joint limiting protection configuration for mechanical degradation and an end effector compliant following configuration for patient spasticity, achieving dual-source joint intervention. Since there may be multiple similar feasible solutions in the fuzzy boundary region of the risk level, all impedance control parameter pairs retrieved from the lookup table are packaged and encapsulated as a candidate intervention action set output, providing a complete action pool for subsequent prospective screening in the digital twin.

[0086] In the aforementioned digital twin-driven medical service robot control safety monitoring system 100, the pre-screening unit 152 is used to inject candidate intervention action sets one by one into the digital twin engine for prospective collision and stress distribution pre-screening and optimization to obtain the optimal strategy instruction. It should be noted that since the candidate intervention action set contains multiple pairs of impedance control parameters, different parameter configurations may trigger differentiated dynamic responses after being applied to the robot. Some candidate actions may cause the joint motion trajectory to collide with the patient's limb or generate excessive stress at the contact interface under specific postures. Therefore, safety verification needs to be completed in virtual space before the intervention instruction is actually applied to the physical robot. Based on this, the technical solution of this application further injects the candidate intervention action set one by one into the digital twin engine for prospective collision and stress distribution pre-screening and optimization to obtain the optimal strategy instruction. Through the above processing, the execution consequences of each set of candidate actions can be pre-evaluated and action sequences with secondary injury risks can be eliminated without any actual impact on the physical robot and the patient, thus selecting the optimal compliant strategy with global safety and smoothness.

[0087] More specifically, in a concrete example of this application, each set of impedance control parameters in the candidate intervention action set is injected one by one into the isolation sandbox of the digital twin engine. Using the robot's current kinematic and dynamic states as initial conditions, the Runge-Kutta numerical integration method is used to extrapolate the robot's kinematic and dynamic evolution trajectory within a future look-ahead time window for each set of candidate impedance parameters. During the extrapolation process, the instantaneous angular velocities of each joint of the robot and the peak contact force applied by the end effector to the virtual patient's limb contact interface are calculated in real time for each evolution trajectory. In a rehabilitation traction scenario, when a set of candidate actions is configured with high stiffness and position lock, the pre-simulation results may show that the peak contact force generated by the end effector on the patient's forearm muscle tissue at the moment of sudden stop exceeds the soft tissue safety tolerance threshold. This action sequence is then marked as having a risk of secondary tearing injury and is eliminated. For each evolution trajectory that is not eliminated, its cost penalty function value within the look-ahead time window is calculated. This cost function comprehensively evaluates two indicators: the surge in joint kinetic energy and the peak end effector contact force.

[0088]

[0089] in, The selected optimal strategy instruction. For a single candidate control parameter pair during the traversal iteration. For the candidate intervention action set, For the current moment, The integration time window for forward-looking rehearsals and These are positive definite weighting coefficients used to balance the two indices of joint kinetic energy and contact force. To apply candidate actions Later derived by the twin engine The joint angular velocity vector at any moment, To apply candidate actions Later derived by the twin engine The contact torque vector acting on the patient at the end of time. Among all candidate actions that were not eliminated, a pair of impedance control parameters that minimizes the above cost integral is selected as the optimal strategy command output.

[0090] In the aforementioned digital twin-driven medical service robot control safety monitoring system 100, the control command issuing unit 153 is used to compile and encapsulate the optimal strategy command into an intervention control command compatible with the underlying servo protocol and issue it to the robot actuator. It should be noted that, since the optimal strategy command expresses physical dimensional parameters such as the virtual target stiffness and virtual target damping in Cartesian space in floating-point form, while the robot's underlying servo driver can only receive binary message frames conforming to a specific communication bus protocol specification, the two are incompatible in data format and encoding method, making it impossible to directly inject floating-point parameters into the actuator. Based on this, the technical solution of this application further compiles and encapsulates the optimal strategy command into an intervention control command compatible with the underlying servo protocol and issues it to the robot actuator. Through the above processing, the optimal compliant strategy verified through twin pre-performance can be transformed into a control frame that can be directly parsed and executed by the physical actuator, closing the entire safety monitoring loop from sensor acquisition, disturbance source decoupling, risk assessment, pre-performance verification to physical intervention.

[0091] More specifically, in a concrete example of this application, firstly, based on the hardware quantization resolution of the robot's underlying servo driver, the floating-point parameters in the optimal strategy instruction are subjected to fixed-point quantization scaling and rounding processing. The floating-point format virtual target stiffness and virtual target damping values ​​are divided by the physical quantity increment corresponding to each pulse of the servo controller, and then rounded to the nearest integer to convert them into hexadecimal dimensionless machine code. After the quantization conversion is completed, the message frames are assembled and encapsulated according to the protocol specifications of the robot's underlying real-time communication bus. The fixed synchronization frame header byte of the communication bus and the safety control command word identifier are assembled at the beginning of the quantized data. This command word identifier is used to specify the functional nature of this data packet, corresponding to specific intervention types such as compliance mode switching or joint limit protection in rehabilitation traction scenarios. A cyclic redundancy check code is calculated and appended to the end of the quantized data to prevent bit errors caused by electromagnetic interference during bus transmission. The frame header, command word, quantized payload data, and check code are concatenated and linked together to form a complete binary message. This message serves as an intervention control command, which is then transmitted from the master control unit to the servo drives of each joint of the robot via the real-time communication bus. The actuators are driven to perform safety intervention actions according to the optimal impedance parameters verified by twin simulation.

[0092] In summary, a digital twin-driven safety monitoring system for medical service robots, according to an embodiment of this application, is described. It unifies multi-source operational data of the medical service robot during contact tasks, such as joint angles, end effector torque, and motor current, into spatiotemporal telemetry features. Simultaneously, two types of simulation results are generated in the digital twin environment, one under the assumption of mechanical degradation and the other under the assumption of physiological disturbance. The actual telemetry state is then compared with the two types of simulation states using residuals, and the disturbance source is decoupled to identify whether the anomaly is closer to a mechanical fault, a patient-side physiological disturbance, or a coupled state of the two. Subsequently, the decoupling results are mapped to a safety incident knowledge graph to obtain the current risk level and future risk prediction results, and candidate intervention strategies are matched accordingly. Before actual deployment, the candidate actions are re-enacted using the digital twin to eliminate action sequences that may cause collisions, overloads, or tissue traction damage, ultimately forming a verified intervention control command. Therefore, the solution can achieve integrated linkage of anomaly source identification, risk assessment, and safety intervention, improving the monitoring accuracy and handling reliability of contact-based medical service robots.

Claims

1. A digital twin-driven control and safety monitoring system for medical service robots, characterized in that, include: The spatiotemporal telemetry feature construction module is used to perform spatiotemporal phase alignment and Kalman filtering denoising on the joint angle data stream, end torque data stream and motor current data stream collected by the medical service robot in contact tasks to obtain the spatiotemporal telemetry feature matrix. The dual-branch digital twin simulation module is used to inject the spatiotemporal telemetry feature matrix into the dual positive simulation branches with mutually exclusive boundary conditions in the digital twin engine to perform hypothesis testing-based dynamic deduction in order to obtain the mechanical degradation simulation tensor and the physiological disturbance simulation tensor. The disturbance source decoupling module is used to perform spatiotemporal dual-domain residual analysis and precise decoupling of the disturbance source from the spatiotemporal telemetry feature matrix, the mechanical degradation simulation tensor, and the physiological disturbance simulation tensor to obtain the decoupled disturbance source vector. The dynamic safety assessment module is used to map the abnormal source type identifier and deviation intensity value in the decoupled disturbance source vector into a pre-trained safety incident knowledge graph based on graph neural network to obtain a dynamic safety assessment report. The intervention strategy verification and control module is used to adaptively match candidate intervention strategies based on the risk level and anomaly source type in the dynamic safety assessment report. The candidate intervention action set is injected back into the digital twin for secondary pre-run verification to eliminate action sequences with secondary damage risk. The selected optimal strategy is compiled into intervention control instructions and sent to the robot actuator.

2. The digital twin-driven medical service robot control safety monitoring system according to claim 1, characterized in that, The spatiotemporal telemetry feature construction module includes: The phase alignment unit is used to perform interpolation compensation and coordinate system homogeneous transformation on the joint angle data stream, end torque data stream and motor current data stream with the global microsecond-level master clock as a unified reference to obtain aligned sensing data. An adaptive filtering unit is used to dynamically adjust the observation noise covariance matrix and perform adaptive Kalman filtering frame by frame based on the sliding variance of the aligned sensor data within the local historical time window to obtain the denoised sensor data. The feature matrix generation unit is used to perform horizontal cascading and dimensionality reduction of the joint angle subarray, current subarray and torque subarray along the column direction of the denoised sensing data within a preset time sliding window to obtain the spatiotemporal telemetry feature matrix.

3. The digital twin-driven medical service robot control safety monitoring system according to claim 1, characterized in that, The dual-branch digital twin simulation module includes: The state parameter extraction unit is used to decompose the spatiotemporal telemetry feature matrix in order to extract the kinematic state vector and dynamic state vector at the current moment, and obtain the observed state parameter vector. The mechanical degradation simulation unit is used to inject the observed state parameter vector into the first simulation branch to perform forward dynamic deduction and obtain the mechanical degradation simulation tensor. The physiological disturbance simulation unit is used to inject the observed state parameter vector into the second simulation branch for forward dynamic deduction to obtain the physiological disturbance simulation tensor.

4. The digital twin-driven medical service robot control safety monitoring system according to claim 1, characterized in that, The disturbance source decoupling module includes: The mechanical residual analysis unit is used to calculate the mechanical residual distance between the spatiotemporal telemetry feature matrix and the mechanical degradation simulation tensor under the optimal regular path, with the spatiotemporal telemetry feature matrix as the reference sequence. The physiological residual analytical unit is used to calculate the physiological residual distance between the spatiotemporal telemetry feature matrix and the physiological perturbation simulation tensor under the optimal regularization path, with the spatiotemporal telemetry feature matrix as the reference sequence. The disturbance source screening and decoupling unit is used to perform competitive screening of mechanical residual distance and physiological residual distance to obtain the decoupled disturbance source vector.

5. The digital twin-driven medical service robot control safety monitoring system according to claim 1, characterized in that, The dynamic security assessment module includes: The vector slicing parsing unit is used to extract index slices from the decoupled disturbance source vectors to obtain the anomaly source type identifier and deviation intensity value; The risk scoring retrieval unit is used to concatenate the anomaly source type identifier and the deviation intensity value into an initial node feature vector and inject it into the safety accident knowledge graph to perform knowledge graph-driven risk probability density retrieval to obtain a future risk prediction score. The risk level encapsulation unit is used to discretize the future risk prediction score based on the step threshold piecewise function, and encapsulate the determined risk level and the future risk prediction score to obtain a dynamic security assessment report.

6. The digital twin-driven medical service robot control safety monitoring system according to claim 1, characterized in that, The intervention strategy verification and control module includes: The candidate strategy matching unit is used to decode the dynamic security assessment report and perform two-dimensional table lookup and parameter matching in the compliant impedance strategy matrix library according to the risk level and anomaly source type to obtain a set of candidate intervention actions. The pre-screening unit is used to inject candidate intervention action sets one by one into the digital twin engine for forward collision and stress distribution pre-screening and optimization to obtain the optimal strategy instruction. The control command issuing unit is used to compile and encapsulate the optimal strategy command into intervention control commands compatible with the underlying servo protocol and issue them to the robot actuator.

7. The digital twin-driven medical service robot control safety monitoring system according to claim 4, characterized in that, The disturbance source screening and decoupling unit includes: The independent channel activation energy calculation subunit is used to perform residual-driven independent channel activation energy calculation on the mechanical residual distance and physiological residual distance to obtain the mechanical activation energy and physiological activation energy. The cross-source coupling degree estimation subunit is used to estimate the cross-source coupling degree of mechanical activation energy and physiological activation energy to obtain the cross-source coupling degree. The soft contribution decomposition subunit is used to obtain the decoupled perturbation source vector by performing soft contribution decomposition on the coupling modulation of mechanical activation energy and physiological activation energy based on cross-source coupling degree.

Images