Intelligent nursing robot system for assisting interventional operation based on deep learning

The deep learning-based intelligent nursing robot system solves the problems of long nurse training cycles, radiation hazards, and errors in instrument identification and delivery during interventional surgery, achieving efficient, safe, and standardized operation of interventional surgery.

CN121946574APending Publication Date: 2026-05-01BEIJING GENERAL AEROSPACE HOSPITAL
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610299412.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-03-12
Publication Date
2026-05-01

AI Technical Summary

Technical Problem

Existing interventional surgeries suffer from problems such as long nurse training cycles, high labor costs, serious radiation hazards, high risk of errors in instrument identification and delivery, and lagging vital sign monitoring. The lack of intelligent auxiliary means leads to low surgical efficiency and safety.

Method used

The system employs a deep learning-based intelligent nursing robot system, which includes a central control unit, a mobile robot platform, a multi-functional collaborative robotic arm, an intelligent instrument identification and delivery module, and a vital signs intelligent early warning module. It combines technologies such as multi-source perception navigation, flexible force-controlled grippers, and non-contact interactive terminals to achieve accurate instrument identification and delivery, early warning of vital signs, and obstacle avoidance navigation.

Benefits of technology

It reduces the risk of instrument delivery errors, reduces radiation exposure for medical staff, improves the standardization of surgical procedures, enables early warning of vital signs and efficient collaborative operation, and enhances the safety and efficiency of interventional surgery.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121946574A_ABST
    Figure CN121946574A_ABST
Patent Text Reader

Abstract

The invention provides an intelligent nursing robot system for assisting an interventional operation based on deep learning. The system comprises a central control unit, a mobile robot platform, a multifunctional cooperative mechanical arm, an intelligent instrument recognition and delivery module, a vital sign intelligent early warning module and a man-machine interaction module. The flexible damage-prevention clamping jaw at the tail end of the multifunctional cooperative mechanical arm is used for effectively preventing the thin-wall intervention equipment from being damaged; through a similar catheter vision distinguishing algorithm based on deep learning, fine-grained extraction is carried out on highly similar appearance features, and precise recognition and dual verification delivery of a complex interventional instrument are achieved; and meanwhile, a vital sign early warning mechanism based on a deep learning time sequence model is introduced, so that the trend of complications such as vagus nerve reflex and the like can be recognized in advance and early warning can be performed. The problems that in the high-radiation environment of the catheter room, nursing manpower is nervous, instrument delivery is prone to errors, complication intervention lags and the like are solved, and the safety and cooperation efficiency of an interventional operation are remarkably improved.
Need to check novelty before this filing date? Find Prior Art

Description

A Deep Learning-Based Intelligent Nursing Robot System for Assisting Interventional Surgery Technical Field

[0001] This invention relates to the field of medical device and robot control technology, and in particular to an intelligent nursing robot system for interventional surgery assistance based on deep learning. Background Technology

[0002] Interventional cardiology procedures, such as coronary angiography and stent implantation, are crucial for the clinical treatment of cardiovascular diseases. During routine procedures, scrub nurses must stand for extended periods, wearing heavy lead protective suits, and frequently, rapidly, and accurately deliver dozens of delicate and expensive interventional instruments, such as various catheters, guidewires, and stents, to the surgeon in a continuous environment of ionizing radiation. However, the existing clinical collaboration model has several significant drawbacks and limitations. First, training an experienced interventional scrub nurse is extremely time-consuming and costly, and the high-intensity radiation environment consistently poses a serious cumulative threat to the occupational health of healthcare workers.

[0003] Secondly, there is a high risk of human error in the instrument identification and delivery process. Consumables used in interventional procedures (such as JL and JR angiography catheters) often look extremely similar, making them difficult to distinguish quickly with the naked eye. In the high-intensity, high-stress environment of the operating room, nurses are highly susceptible to fatigue, leading to errors or delays in instrument delivery. Due to the current lack of intelligent auxiliary methods such as machine vision-based dual verification, this risk of human error severely restricts the overall efficiency and safety of the surgery. At the same time, individual differences in operating habits among nurses make it difficult to achieve complete standardization and normalization of the surgical procedure.

[0004] Furthermore, current equipment for monitoring patients' vital signs during surgery largely relies on passive alarm mechanisms triggered by physiological values ​​exceeding limits. For common and high-risk intraoperative complications, such as the vagal reflex that can easily cause a double drop in heart rate and blood pressure, existing systems lack the ability to predict the dynamic evolution of various vital signs in advance, resulting in a lack of early proactive warning mechanisms and often a lag in clinical intervention. Although some surgical robots are already being used in the surgical field, most are concentrated at the surgeon's operating end, and intelligent equipment specifically designed for assisting with nursing work in interventional catheterization labs is still lacking. Therefore, there is an urgent clinical need to develop an intelligent nursing robot system that can deeply understand the surgical procedure, accurately identify instruments based on deep learning technology, and efficiently collaborate with doctors. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent nursing robot system for interventional surgery based on deep learning, in order to solve the problems mentioned in the background art.

[0006] This invention provides an intelligent nursing robot system for interventional surgery assistance based on deep learning, comprising: a central control unit for processing data and sending control commands to schedule the collaborative operation of the entire system; a mobile robot platform for autonomous movement within the catheterization lab; a multi-functional collaborative robotic arm mounted on the mobile robot platform, the end effector of which is equipped with a flexible, damage-resistant gripper with integrated force feedback, used to non-destructively grasp and deliver interventional surgical instruments under the control of the central control unit; an intelligent instrument identification and delivery module, including a machine vision unit and a deep learning-based similar catheter visual differentiation model, used to perform fine-grained feature extraction and classification of multiple highly similar interventional instruments on the instrument table, and to perform at least two verifications before grasping and delivery to confirm the correctness of delivery; a vital signs intelligent early warning module, communicatively connected to the vital signs monitoring equipment within the catheterization lab, and incorporating a deep learning-based time-series data analysis model to acquire real-time vital signs data of the patient, analyze the trend of vital signs changes, and issue early warnings before reaching complication thresholds such as vagal reflexes; and a human-computer interaction module for receiving operator commands and providing status feedback through screen display and indicator lights.

[0007] Optionally, the mobile robot platform is equipped with an environmental perception and navigation unit, which includes a lidar and a depth camera. The central control unit is configured to use the depth camera to identify the geometric features of the C-arm robot's base and the features of the suspended rails on the ceiling in the duct room. It then uses the ICP algorithm to register the real-time perceived feature point cloud with a pre-stored reference template to calculate the absolute pose correction. This absolute pose correction is then introduced as a strong observation constraint into the particle filter of the adaptive Monte Carlo localization algorithm. The central control unit is also configured to execute a predictive obstacle avoidance strategy: real-time acquisition of the joint angles and angular velocities of the C-arm robot; prediction of the set of spatial regions swept by each rigid body of the C-arm robot within a future time window using a parametric kinematic model; mapping the scanned volume onto a local cost map and marking it as a dynamic no-entry layer; and dynamically adjusting the movement strategy based on the collision time when a planned path is detected crossing the scanned volume.

[0008] Optionally, the multifunctional collaborative robotic arm is a seven-degree-of-freedom redundant robotic arm with zero spatial degrees of freedom; the flexible anti-damage gripper integrates an array of piezoresistive sensors, which can sense the pressure distribution map on the contact surface; the central control unit is configured to monitor the peak pressure in the pressure distribution map, and if the peak pressure exceeds the preset yield limit threshold of the conduit material, the clamping current of the gripper is reduced in a closed loop; and the central control unit is also configured to perform frequency domain analysis on the pressure distribution map, and when a high-frequency micro-vibration signal along the direction of gravity is detected at the pressure center, it is determined to be a slip precursor, and the clamping force is dynamically increased in small increments until the micro-vibration signal disappears.

[0009] Optionally, the human-computer interaction module includes a non-contact intelligent interaction terminal, which has a built-in non-contact gesture recognition module for capturing the surgeon's hand movements in the sterile area; the system executes a two-level confirmation mechanism: entering the command listening state by recognizing a specific wake-up gesture; and then performing semantic parsing on the gesture flow of the spatial skeleton model within an effective time window and mapping it into corresponding robotic arm control commands.

[0010] Optionally, the mobile robot platform is specifically an omnidirectional mobile chassis equipped with Mecanum wheels; the system also includes a sterile consumable storage compartment, which is located inside the omnidirectional mobile chassis, has a controllable opening sealed door, and integrates a micro positive pressure air supply unit inside; the central control unit is configured to maintain the air pressure inside the compartment at a slightly positive pressure state by adjusting the fan speed of the micro positive pressure air supply unit, and to force the fan to run at full power before the sealed door is opened to form a high-speed clean airflow barrier at the compartment door.

[0011] Optionally, the sterile consumables storage compartment is equipped with a multi-polarized antenna array and an ultra-high frequency RFID reader / writer; the RFID reader / writer is configured to change the standing wave distribution mode of electromagnetic waves in the compartment by rapidly switching the carrier phase of each transmitting antenna; the central control unit is configured to calculate the presence confidence of the electronic tag by using the received signal strength indication and Kalman filtering smoothing processing, thereby automatically identifying and retrieving consumables and generating inventory change reports to be uploaded to the hospital material management system.

[0012] Optionally, the surface of the mobile robot platform is covered with a radiation-resistant coating of a flexible lead polymer plate made by calendering lead powder and an ethylene-vinyl acetate copolymer matrix; the system also includes a radiation dose monitor for real-time detection of the ambient X-ray dose rate; the central control unit is configured to associate the real-time dose rate with the robot's spatial position to construct a dynamic radiation heat map, and to include the radiation dose rate as a cost term in the cost function during path planning.

[0013] Optionally, the human-computer interaction module is equipped with a microphone ring array; the central control unit is configured to run an adaptive beamforming algorithm, which constructs a fixed beamformer, a blocking matrix, and an adaptive noise canceller submodule in the signal processing logic; the fixed beamformer is used to form a main lobe pointing towards the surgeon's position, the blocking matrix is ​​used to block the desired direction signal and output a noise reference, and the adaptive noise canceller uses an adaptive algorithm to adjust the weights so that the output noise approximates the noise component in the fixed beamformer branch, thereby forming a null in the direction of the interference noise source; the system also integrates a high-frequency spectrum prediction and reconstruction model based on generative adversarial networks to improve the accuracy of speech recognition caused by medical staff wearing masks.

[0014] Optionally, the central control unit analyzes the patient's heart rate, invasive arterial blood pressure, blood oxygen saturation, and electrocardiogram waveform data in real time through an isolation gateway. The central control unit is configured to calculate the vital signs instability index, and its calculation logic is to dynamically weight and fuse the heart rate deviation rate, mean arterial pressure deviation rate, blood oxygen saturation decrease, and the probability of arrhythmia inferred using a lightweight convolutional neural network. The system maintains a finite state machine. When the vital signs instability index exceeds the critical threshold for a continuous preset time or ventricular tachycardia / ventricular fibrillation characteristics are identified, the system automatically switches to emergency assistance mode. In the emergency assistance mode, the mobile robot platform is controlled to plan the shortest collision-free evacuation path to retreat to the non-clean area of ​​the catheterization lab, retract the multi-functional collaborative robotic arm, and automatically open the sterile consumable storage compartment containing emergency medications.

[0015] Optionally, the intelligent instrument identification and delivery module performs dual verification logic: before grasping, it obtains a global image of the instrument table through a global vision camera to extract candidate regions for interventional surgical instruments; after the robotic arm grasps the instrument and removes it from the table, it triggers a local vision camera at the end to collect a close-up image for secondary classification and comparison; the similar catheter visual differentiation model has a spatial attention module embedded in it, and its network weights are obtained through offline iterative training using a metric learning strategy that includes a triplet loss function. The training process is configured to increase the inter-class feature distance of different types of catheters in the feature mapping space and reduce the feature differences of the same type of catheter caused by the environment, thereby performing fine-grained classification of the catheter tip and tail connector.

[0016] This invention achieves the following beneficial effects: By configuring a deep learning-based visual differentiation model for similar catheters, this system can perform fine-grained feature extraction and multiple verifications of highly similar interventional devices, reducing the risk of device delivery errors caused by visual identification under fatigued conditions. Combined with a flexible, damage-resistant gripper and force control feedback mechanism integrated into the end effector of a multi-functional collaborative robotic arm, it avoids physical damage to the extremely thin tube walls and hydrophilic coatings caused by mechanical clamping, ensuring the structural integrity of medical consumables. In terms of vital sign monitoring, the system's built-in deep learning time-series data analysis model can jointly infer the dynamic decline gradient of heart rate and blood pressure, capturing the occurrence trend of complications such as vagal reflexes before physiological indicators fall below the alarm lower limit, transforming traditional passive over-limit alarms into early proactive warnings, providing a time window for medication preparation and intervention for the medical team. The system utilizes the collaborative operation of a mobile robot platform and a human-computer interaction module to replace nurses in performing high-frequency delivery tasks in the radiation area, reducing the cumulative radiation exposure of medical personnel. Furthermore, the unified automated control logic eliminates differences in human operating habits, improving the standardization and normalization of interventional catheterization lab procedures.

[0017] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 is a schematic diagram of the composition structure of an intelligent nursing robot system for interventional surgery assistance based on deep learning in an embodiment of the invention; Figure 2 is a schematic diagram of the workflow and control logic of the intelligent nursing robot system in an embodiment of the invention when performing dual-verification delivery of instruments and vital sign early warning. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] This invention provides an intelligent nursing robot system for interventional surgery assistance based on deep learning. This system addresses the special working conditions of interventional catheterization labs, including high radiation, strict aseptic requirements, and interference from complex dynamic equipment (such as large C-arm machines). It constructs an intelligent solution integrating omnidirectional mobility, multi-source perception fusion, dynamic obstacle avoidance, flexible force control, and multimodal contactless interaction.

[0022] As shown in Figure 1, the physical architecture of this system mainly includes: an omnidirectional mobile chassis, an environmental perception and navigation unit, a multi-degree-of-freedom collaborative robotic arm, an intelligent interactive terminal, a sterile consumable storage compartment, a radiation dose monitor, and a central control unit. Each functional component constructs an internal communication network through a highly interference-resistant industrial real-time bus (such as EtherCAT) and gigabit Ethernet, achieving control synchronization and throughput of large amounts of sensor information. The omnidirectional mobile chassis and the environmental perception and navigation unit mounted on it constitute the mobile robot platform of this system; the intelligent interactive terminal and the noise-reduced voice array constitute the human-machine interaction module of this system.

[0023] The omnidirectional mobile chassis serves as the robot's motion carrier, equipped with a high-power-density drive module for autonomous movement within the catheterization lab. Considering that the floor of the catheterization lab is typically covered with moving rails for imaging equipment, foot switches, and defibrillator cables, and that the passageway between the operating table and the C-arm machine frame is extremely narrow (usually less than 800mm), conventional differential drives or Ackerman steering mechanisms are insufficient to meet the mobility requirements for zero-radius attitude adjustment or lateral translation within a confined space. Therefore, the omnidirectional mobile chassis in this embodiment preferably adopts a four-wheel Mecanum wheel layout. Each Mecanum wheel consists of a hub and several spindle-shaped rollers distributed around the circumference of the hub, with the roller axis forming a 45-degree angle with the hub axis.

[0024] The drive module includes four independent servo drive units, which drive the Mecanum wheels in the front left, front right, rear left, and rear right positions, respectively. Each servo drive unit consists of a brushless DC servo motor, a precision planetary reducer, and a high-resolution absolute encoder (resolution not less than 23 bits).

[0025] In order to accurately convert the robot's overall desired motion trajectory into control commands for the four independent wheels and improve trajectory tracking accuracy, the central control unit integrates an omnidirectional kinematics calculation module, which performs calculations by constructing a kinematic coordinate system.

[0026] The specific calculation process is as follows: First, establish the robot chassis coordinate system, setting the origin O as the geometric center of the chassis, with the X-axis pointing towards the front of the vehicle and the Y-axis pointing towards the left side of the vehicle. Define the robot's desired motion state vector as... superscript Denotes the transpose of a matrix; where Let X be the longitudinal linear velocity (m / s) along the X-axis. Let be the transverse linear velocity (m / s) along the Y-axis. Let be the angular velocity of rotation about the geometric center (rad / s). Let the radius of the Mecanum wheel be... The geometric distribution of the centers of the four wheels forms a rectangle, with half the distance between the centers of the left and right wheels being... Half of the center distance between the front and rear wheels is The omnidirectional kinematics calculation module calculates the desired rotational speed of each wheel using the following inverse kinematics matrix equations. This enables a precise mapping from the overall chassis posture to the actions executed by the underlying motors. In the formula, Indicates the first The angular velocity of each wheel, measured in rad / s. , , as well as These represent the angular velocities of the 1st, 2nd, 3rd, and 4th wheels, respectively. All units are meters. The central control unit sends the calculated wheel speed commands to the servo driver. The driver uses a FOC (Field Oriented Control) vector control strategy combined with SVPWM (Space Vector Pulse Width Modulation) technology to finely adjust the motor stator current. Furthermore, to overcome wheel slippage caused by cable obstruction on the ductwork chamber floor or slippery coupling agent, a disturbance observer is introduced into the drive module. When the observed deviation between the actual motor current and the theoretical model current exceeds a preset threshold, the control algorithm automatically determines slippage or obstruction and dynamically adjusts the torque output of each wheel, correcting the vehicle's attitude through motion compensation of the remaining wheels.

[0027] The omnidirectional mobile chassis employs a multi-layer composite shielding design. The outer shell of the chassis is made of high-strength modified ABS engineering plastic, with the inner surface lined with a specially designed flexible lead polymer plate. This lead polymer plate is formed by mixing and calendering micron-sized lead powder with an ethylene-vinyl acetate copolymer (EVA) matrix, ensuring flexibility while maintaining a lead equivalent of 0.5 mmPb to 1.0 mmPb. This shielding layer completely covers the central control unit, battery management system (BMS), and motor drivers inside the chassis, effectively attenuating the primary radiation emitted by the DSA tube and scattered radiation reflected from the ground and walls, preventing high-energy photons from inducing single-event upsets or full-dose effects in semiconductor devices, thereby ensuring the long-term operational reliability of the robot in high-radiation fields.

[0028] The environmental perception and navigation unit is mounted on the top support of the omnidirectional mobile chassis and is used to construct a dynamic environmental map of the catheterization lab and perform real-time positioning. Given the presence of numerous highly reflective objects such as stainless steel cabinets and lead glass windows within the catheterization lab, and the variable lighting conditions during surgery (e.g., dimming lights during fluoroscopy), a single optical sensor cannot guarantee robust perception. Therefore, this embodiment employs a heterogeneous multi-sensor fusion scheme. Specific configurations include: a multi-line (e.g., 32-line) mechanical rotating or solid-state LiDAR located on the top of the chassis, used to acquire 3D environmental point clouds with a 360-degree horizontal direction and a certain field of view in the vertical direction; four sets of RGB-D depth cameras distributed at the front, sides, and rear of the chassis, used to detect low-lying obstacles (e.g., the surgeon's feet) and suspended obstacles (e.g., the edge of the operating table); and a 12-channel ultrasonic radar array surrounding the chassis skirt, used for near-field detection of optical blind spots such as transparent glass doors.

[0029] The central control unit runs a graph-optimized SLAM system. During the map building phase, the system calculates pose using a front-end odometry system (fusing wheel speed measurement and IMU data) and combines this with point cloud scanning and matching from LiDAR to construct a local sub-map. A loop closure detection mechanism identifies whether the robot has returned to a previous position, and nonlinear least squares methods (such as G2O or Ceres Solver) are used to optimize the global pose map, eliminating accumulated errors and generating a high-precision 3D occupancy grid map.

[0030] In the real-time localization stage, this system employs a strategy combining the Adaptive Monte Carlo Localization (AMCL) algorithm with C-arm robot feature matching. Due to the high repetition of features on the walls surrounding the catheterization chamber, conventional algorithms are prone to localization drift. Therefore, the system first uses a depth camera to identify the base geometry and ceiling rail features of the C-arm robot, whose position is absolutely fixed within the catheterization chamber. Subsequently, the central control unit uses the ICP (Iterative Closest Point) algorithm to register the real-time perceived feature point cloud with a pre-stored reference 3D template, accurately calculating the robot's absolute pose correction. Finally, this correction is directly introduced as a strong observation constraint into the particle filter of the AMCL algorithm, forcing convergence of the particle swarm distribution and thus eliminating accumulated localization errors.

[0031] Furthermore, C-arm robots project images from multiple angles during surgery. If the robot relies solely on sensors for passive obstacle avoidance, it is highly susceptible to getting stuck in blind spots. Therefore, this embodiment employs a predictive obstacle avoidance strategy, the specific execution logic of which is as follows: First, a parametric kinematic model of the C-arm robot is established, simplifying it into a series of rigid body systems.

[0032] The central control unit acquires real-time joint angle data vectors of the C-arm machine at a frequency of 20Hz via a communication interface. and the angular velocity (or linear velocity) data vectors of each joint. .

[0033] In the formula, The horizontal rotation angle of the C-arm machine base; The rotation angle of the L-shaped arm; The sliding angle of the C-arm; This refers to the vertical displacement of the operating table (considered as a translational joint). This refers to the longitudinal translational displacement of the operating table.

[0034] The second step, to quantify future collision risks, involves introducing the concept of scan volume based on forward kinematics. Let the current time be... The preset prediction time window is (e.g., 3 seconds), symbol This represents the union operation of sets. The central control unit uses numerical integration to predict the set of three-dimensional regions continuously swept by each rigid body of the C-arm in space within a future time window, i.e., the scan volume. : In the formula, The scan volume represents the set of absolute physical space regions continuously scanned and occupied by the rigid body components of the C-arm within a future prediction time window. In the navigation system, this region is mapped as a dynamic no-go layer in the local cost map. This represents the union operation of sets. Here, it means mathematically combining the spatial occupancy volumes at each tiny time step within a time window to form a continuous motion envelope. Indicates the start time of the current observation. This indicates the preset prediction time window length (e.g., 3 seconds). The selection of the time window directly determines the foresight of the obstacle avoidance strategy and the consumption of computational resources. Indicates the prediction time window A variable at any given time within the range. express This is a column vector of the absolute angle / displacement states of each joint of the C-arm boom at any given time. Specifically, it includes: horizontal rotation angle. L-arm rotation angle C-arm sliding angle Operating table lifting and displacement and longitudinal translational displacement . express The first-order kinematic derivatives of each joint at each instant, i.e., the column vectors of angular velocity or linear velocity. This term is used for first-order linear motion derivation over a short period of time. This represents the parametric forward kinematics solution function mapping for the C-arm machine. Its function is to transform the pose vector in joint space into a three-dimensional geometric volume model in the absolute Cartesian coordinate system of the catheterization lab, based on DH parameters or screw theory.

[0035] The third step, after completing the prediction, is for the navigation system to generate the scan volume. Dynamically map the data to a local cost map and perform obstacle avoidance path planning.

[0036] In a robot's navigation system, this scan volume The region is mapped onto a local cost map and marked as a dynamic forbidden layer, with its cost set to infinity. Simultaneously, to handle sensor errors and communication delays, the system... A safety expansion radius is superimposed on the geometric boundary. (e.g., 200mm).

[0037] The path planner employs either an improved A* algorithm or Dijkstra's algorithm to plan global paths. The local path planner uses the TEB algorithm, which introduces a time dimension when optimizing paths. This occurs when the planned path is detected to traverse the predicted scan volume. When the area is in use, the TEB algorithm dynamically adjusts the strategy based on the time to collision (TTC): if the TTC is long, it plans an alternative path; if the TTC is short and there is no alternative space, it controls the robot to immediately reverse or brake, and announces "Avoid C-arm robot" in voice, thereby effectively avoiding the risk of the robot colliding with expensive medical equipment.

[0038] The multi-DOF collaborative robotic arm is mounted on the front platform of the omnidirectional mobile chassis. This embodiment preferably employs a 7-DOF redundant robotic arm. Compared to a traditional 6-DOF robotic arm, the 7-DOF robotic arm has a redundant zero-space degree of freedom. This means that while maintaining the position of the end effector… and posture Even when locked, the robotic arm's elbow joint can still move in space. Utilizing this characteristic, the central control unit uses a zero-space projection control law to confine the elbow within a safe area that avoids the surgical light, suspended lead screen, and the surgeon's head, greatly improving the safety and psychological comfort of human-machine collaboration.

[0039] The end effector is a flexible adaptive gripper. Catheters, guidewires, and balloons commonly used in interventional surgery are mostly made of polymer materials with extremely thin walls (typically on the order of 0.1 mm) and hydrophilic coatings, making them highly susceptible to flattening or damage by rigid grippers. The flexible gripper in this embodiment employs an underactuated linkage mechanism, with the fingertip contact surface covered by a medical-grade silicone balloon, allowing it to conform to the shape of the object being gripped.

[0040] More importantly, the flexible adaptive gripper integrates a high-density array of force sensors within its fingertips. These sensors are composed of… (For example, 16x16) miniature piezoresistive sensing units arranged in a matrix can sense the pressure distribution pattern on the contact surface. The force control algorithm in the central control unit monitors this pressure distribution in real time and executes the following closed-loop control logic: the system calculates the peak pressure in the pressure distribution matrix in real time. .like Exceeding the preset yield strength threshold of the conduit material The control algorithm immediately reduces the current of the clamping motor through the PID controller until... It has fallen back to a safe range.

[0041] The system performs frequency domain analysis on the pressure distribution image. When a high-frequency micro-flutter signal along the direction of gravity is detected at the center of pressure (COP), it is identified as a slip precursor. At this point, the algorithm, while ensuring no overpressure, applies a small increment... Dynamically increase the clamping force until the micro-vibration signal disappears.

[0042] When delivering the catheter to the physician, the robotic arm enters admittance control mode. An end-effector dynamics model is established: ;in, These represent the virtual mass, damping, and stiffness matrices, respectively. and These represent the desired and actual positions of the robotic arm's end effector, respectively; each with a superscript. and These represent the desired velocity and actual velocity (i.e., the first derivative of the position) of the robotic arm's end effector, respectively; with two superscripts. and These represent the desired acceleration and the actual acceleration (i.e., the second derivative of the position) at the end of the robotic arm, respectively. The external interaction force is calculated using a joint torque sensor.

[0043] During the handover phase, and By lowering the parameters, the robotic arm exhibits the characteristics of a soft spring. The doctor only needs to apply a very small guiding force (such as 0.5N), and the robotic arm can move in the direction of the doctor's gestures, achieving a natural tactile sensation similar to that of a human nurse handing over a patient.

[0044] Furthermore, to achieve non-destructive gripping of interventional devices, the end effector of the multifunctional collaborative robotic arm is equipped with a flexible, damage-resistant gripper. This gripper integrates an array of piezoresistive sensors to acquire real-time pressure distribution data on the contact surface between the gripper and the device. The central control unit receives this pressure distribution data and executes closed-loop control: when the local pressure reaches a preset yield strength of the catheter material or a coating damage threshold, the gripper's drive mechanism is stopped from closing; during delivery, if a slippage tendency is detected by pressure center offset, the gripping force is finely adjusted and increased within a safe threshold range. Through the coordination of the above hardware configuration and control logic, the system can flexibly wrap around thin-walled polymer catheters with hydrophilic coatings, conforming to the device's shape. This maintains stability during gripping and transfer while avoiding plastic deformation of the tube wall or coating peeling caused by rigid mechanical compression, ensuring the structural integrity and safe use of high-value medical consumables.

[0045] During the grasping and delivery tasks performed by the multifunctional collaborative robotic arm, the system is guided and verified by a configured intelligent instrument recognition and delivery module. This module includes a global vision camera mounted on the mobile robot platform, a local vision camera at the end of the robotic arm, and a deep learning-based visual differentiation model for similar catheters running in the central control unit. Specifically, the system first acquires a global image of the instrument table using the global vision camera, and then uses a target detection network to extract candidate regions for each instrument on the table. Subsequently, the candidate region images are input into the visual differentiation model for similar catheters. This model uses a spatial attention mechanism to extract fine-grained features such as the bending angle of the catheter tip and the color of the connector at the tail, thereby classifying and identifying angiography catheters with highly similar overall appearances (such as JL and JR types). To ensure the accuracy of the delivery results, the system employs a dual verification logic: preliminary classification and coordinate positioning are performed based on the global image before grasping; after the robotic arm's flexible anti-damage gripper grasps the instrument and leaves the table, the end-effector local vision camera is triggered to acquire a close-up image for secondary classification and comparison. Only when the two classification results are consistent and match the current preset surgical plan can the central control unit control the robotic arm to perform the delivery action to the surgeon. This module overcomes the interference of local lighting changes and instrument stacking obstruction in the catheterization lab through fine-grained feature extraction and multiple verification mechanisms, avoiding the problems of misidentification and incorrect delivery of similar instruments that are very likely to occur when the surgeon is fatigued.

[0046] Furthermore, to achieve accurate classification of the aforementioned similar devices, the construction and offline training process of the visual differentiation model for similar catheters is as follows: First, the system acquires multi-angle sample images of various interventional devices under actual lighting conditions in the catheterization lab and in a sterile environment with stacked workbenches, constructing a training dataset. Professional personnel then meticulously annotate the bounding boxes and category labels of the morphology of the head end and the tail connector feature regions of various catheters in the images. Second, a deep learning architecture is constructed, with a convolutional neural network (CNN) as the backbone and a cascaded spatial attention module. During the model training phase, the training dataset is input into the network, and a metric learning strategy is employed, optimizing the network parameters through a triplet loss function. The working principle of this training logic is to force the network to continuously increase the feature distance between different types of catheters (negative samples) in the feature mapping space, while simultaneously reducing the feature differences of the same type of catheter (positive samples) caused by changes in lighting or posture. By setting a batch size and learning rate, the gradient descent algorithm is used for iterative training until the loss function converges and the model's recognition accuracy on the validation set reaches a preset threshold. Finally, the similar catheter visual differentiation model is generated and solidified, and deployed in the central control unit for real-time use during surgery.

[0047] To ensure the control system operates under aseptic conditions, the intelligent interactive terminal of this system adopts a dual-modal non-contact interaction scheme.

[0048] The non-contact gesture recognition module includes a binocular stereo camera or a ToF depth camera mounted on the robot's head, specifically designed to capture hand movements in the surgeon's chest operating area (sterile zone). At the algorithm level, a deep learning-based skeletal keypoint extraction network (such as ST-GCN) is employed. The system tracks 21 key points of the hand in real time (including the wrist joint, metacarpophalangeal joints, and interphalangeal joints) to construct a spatial skeletal model of the hand.

[0049] To prevent accidental triggering (such as a doctor's hand movements during normal surgery being mistaken for commands), the system implements a two-level confirmation mechanism: Level 1: Wake-up detection. The doctor needs to make a specific wake-up gesture (such as a "V" sign and hold it for 1 second) in the center of the depth camera's field of view. After the system recognizes the gesture, the indicator light turns green and emits a "beep" sound, indicating that the system has entered command listening mode.

[0050] Level 2: Command Recognition. Within a valid time window after wake-up (e.g., 5 seconds), the system performs semantic parsing of the gesture flow. For example, a leftward waving of the palm is mapped to a command to deliver the next specification of equipment, a circular motion of the index finger is mapped to a command to adjust the height of the robotic arm, and a clenched fist pressing down is mapped to a confirmation / release command. If no valid command is detected within the time limit, the system automatically returns to standby mode.

[0051] The noise-reducing voice array consists of six omnidirectional microphones arranged in a ring around the robot's neck. Employing adaptive beamforming technology, the system uses a sound source localization algorithm (such as GCC-PHAT) to lock onto the surgeon's azimuth in real time and form a pickup beam pointing in that direction, while simultaneously creating nulls in other directions to effectively suppress background noise from the C-arm cooling fan, ECG monitor alarms, and air conditioning vents. The enhanced voice signal is fed into an ASR (Automatic Speech Recognition) engine, which, combined with a cardiology-specific medical dictionary, performs semantic analysis to recognize professional terms such as "EBU3.5," "balloon inflation," and "angiography." The central control unit performs multimodal fusion verification of voice and gesture commands, executing critical actions only when both are semantically consistent or when the confidence level of a single modality is extremely high, ensuring system safety.

[0052] Radiation dose monitors are crucial components for ensuring the safety of medical personnel and the robots themselves. This monitor utilizes a high-sensitivity semiconductor detector (such as a cadmium zinc telluride (CdZnTe) detector), offering higher energy resolution and faster response speed compared to traditional Geiger tubes. The monitor outputs the ambient dose rate in real time. (unit: ) and cumulative dose .

[0053] The central control unit combines real-time dose rate data with the robot's spatial position. A correlation is established to construct a dynamic radiation heatmap internally. This heatmap is updated in real time according to the X-ray tube angle and exposure parameters of the C-arm X-ray machine. Based on this heatmap, the system implements the following active safety function: when planning the movement path, the radiation dose rate is considered as a cost item. Add it to the cost function of the A* algorithm: In the formula, Indicates the current path search node; For nodes The overall cost of the comprehensive assessment; To move from the starting point to the node The actual path cost (e.g., travel distance); For heuristic functions, it represents the function from the node. The estimated cost of moving to the target point (e.g., Euclidean distance); For nodes The local radiation dose cost at a given location is obtained by querying a dynamic radiation heat map in real time. The weighting coefficient for radiation cost can be dynamically adjusted according to the urgency of the surgical task. This algorithm design enables the robot to automatically select the path with the lowest radiation dose in non-emergency situations, thereby effectively reducing the cumulative radiation damage of high-energy rays to the chassis electronic components.

[0054] When an abnormally high instantaneous dose rate is detected in a certain area (such as next to a doctor's position) (e.g., the scattering shielding curtain is not fully closed), the robot immediately announces "Pay attention to radiation protection" via voice and displays a red warning icon on the interactive terminal screen to remind medical staff to adjust their position or check the protective facilities.

[0055] In this embodiment, an embedded high-cleanliness intelligent storage unit is set up in the hollow area of ​​the omnidirectional mobile chassis (i.e., the space above the chassis drive module and below the robotic arm base). The design of this unit is not only for physical storage, but also to solve the pain points of chaotic management, cumbersome retrieval records, and high risk of non-sterile transfer of high-value consumables (such as drug-eluting stents, balloon catheters, angiography catheters, etc.) in interventional surgery.

[0056] The shell of the sterile consumable storage compartment is made of 316L medical-grade stainless steel plate through laser seamless welding process, and the inner surface is electrolytically polished to achieve a surface roughness of [missing information]. This is to prevent the adhesion and growth of bacterial biofilms and to facilitate routine disinfection with vaporized hydrogen peroxide or ultraviolet light. The storage compartment is equipped with an electrically controlled, sealed door at its opening.

[0057] To maintain absolute sterility within the chamber under dynamic conditions, a miniature positive pressure air supply unit (FFU) is integrated inside the storage chamber. This unit includes a pre-filter and a high-efficiency particulate air (HEPA, H14 grade, filtration efficiency for 0.3μm particles). And DC brushless inverter fans.

[0058] The system's internal pressure sensors monitor the absolute pressure inside the cabin in real time at a frequency of 50Hz. absolute air pressure relative to the external environment The central control unit adjusts the fan speed using an incremental PID control algorithm to maintain the pressure difference between the inside and outside of the chamber. Under a slightly positive pressure state, the control objective formula is: When medical staff request consumables via command, the control unit forces the fan to run at full power within 2 seconds before the sealed door opens, forming a high-speed clean airflow barrier from the inside to the outside at the doorway, with a wind speed of over 0.5 m / s, effectively preventing unpurified external air from flowing back into the cabin at the moment of opening.

[0059] In terms of information technology for consumable management, the inner wall of the sterile consumable storage compartment integrates an ultra-high frequency (UHF) RFID reader and writer and a multi-polarized antenna array.

[0060] The RFID reader operates in the 920MHz-925MHz frequency band and supports the EPCClass1Gen2 standard.

[0061] Four sets of circularly polarized microstrip antennas are deployed on the top, bottom and two sides of the cabin to form an omnidirectional coverage field.

[0062] During the reading cycle, the reader rapidly switches the carrier phase of each transmitting antenna. This involves artificially altering the standing wave distribution pattern of electromagnetic waves inside the cabin, thereby activating electronic tags located at standing wave nodes (blind spots).

[0063] The system executes an anti-collision algorithm based on the dynamic Q-value time-slot ALOHA protocol. The central control unit uses Received Signal Strength Indication (RSSI) and phase information to locate and verify the presence of tags. Based on the modified Friis transmission equation, the first... The first label in the The received power at each antenna is : In the formula, This refers to the transmission power. Let be the gain of the j-th transmitting antenna; Antenna gain for electronic tags; The operating wavelength of the radio frequency signal; The radar cross section of the tag; Indicates the first The electronic tag and the first The spatial distance between each transmitting antenna; The polarization loss factor represents the signal energy loss caused by the mismatch between the antenna and tag polarization directions. The multipath fading factor represents the signal attenuation caused by complex electromagnetic wave reflections inside the metal cabin, and its fluctuations directly affect the stability of the reading.

[0064] The system smooths the RSSI values ​​read multiple times consecutively using a Kalman filter and calculates the tag presence confidence. Only when the confidence level exceeds a set threshold (such as 98%) is the inventory considered valid.

[0065] Once the sealed door is closed, the system automatically triggers a full inventory check, comparing the label lists before and after the door was opened, and generating an inventory change report (including consumable name, specifications, batch number, and expiration date). This report is uploaded in real time to the hospital's SPD (Supply, Production, and Distribution) material management system via a Wi-Fi 6 private network using the HL7FHIR standard protocol.

[0066] Considering the harsh acoustic environment of the catheterization lab, such as the noise of the cooling fan of the C-arm machine's high-voltage generator (mainly concentrated in the low frequency), the wind noise of the laminar flow purification air conditioner, the high-frequency alarm sound of the multi-parameter monitor, and the sound attenuation caused by doctors wearing masks, this system introduces a front-end processing (FEP) link.

[0067] The intelligent interactive terminal uses a 7-microphone circular array (6 microphones evenly distributed around the circumference, and 1 microphone located at the center). The central control unit runs an adaptive beamforming software algorithm based on a generalized sidelobe canceller (GSC) architecture.

[0068] Assume the mixed signal vector received by the microphone array is It consists of the target speech signal in the desired direction (the surgeon's position). With background noise vector In terms of signal processing logic, the GSC algorithm is divided into three main operational sub-modules: Fixed Beamformer (FBF), Blocking Matrix (BM), and Adaptive Noise Canceller (AIC).

[0069] The FBF module is used to enhance the desired direction signal: ;in, This is to enhance the output signal of the fixed beamformer branch; For a fixed beamformer, the fixed weight vector is indicated by the superscript. This represents the conjugate transpose of a matrix.

[0070] The BM module uses orthogonal matrix transformations to block the signal in the desired direction, outputting only a clean noisy reference signal vector: The AIC module adaptively adjusts the filter weight vector using the LMS (Least Mean Square) algorithm. Generate noise estimation signal: .in, This is the multi-channel noise reference signal output by the blocking matrix; For the blocking transformation matrix (satisfying) ); The adaptive filtering weight vector of the adaptive noise canceller, with superscript... Represents the conjugate transpose of a matrix; The output noise estimation signal (rather than a simple noise scalar value) is used to approximate the true noise component remaining in the FBF branch as closely as possible.

[0071] Finally, the system obtains the clean speech output signal by subtracting the noise estimation signal from the FBF branch signal. : The clean speech signal calculated using the adaptive beamforming algorithm based on the GSC architecture described above. It can form a high-gain main lobe (Beamwidth≈30°) in the direction of the surgeon and a null with a depth greater than 25dB in the direction of the interference noise source, thereby effectively suppressing background steady-state noise such as C-arm machine fan.

[0072] To address the issue of high-frequency speech component attenuation (>3kHz) caused by doctors wearing masks, the backend ASR (Automatic Speech Recognition) module introduced a bandwidth extension algorithm based on generative adversarial networks (GANs) to predict and reconstruct the lost high-frequency spectrum, significantly improving the recognition accuracy of high-frequency words specific to cardiology.

[0073] This system is physically connected to a multi-parameter monitor inside the conduit room via a built-in industrial-grade isolation gateway (with 4kV electrical isolation capability).

[0074] The central control unit analyzes the data stream output by the monitor at a frequency of 1Hz to obtain heart rate (HR) and invasive arterial blood pressure (ABP, including systolic blood pressure). diastolic blood pressure Average pressure ), blood oxygen saturation ( ) and real-time electrocardiogram (ECG) waveform.

[0075] Based on the real-time vital sign data acquired above, the intelligent vital sign early warning module configured within the system dynamically monitors the patient's condition and provides early warnings of complications by running a deep learning-based time-series data analysis model. Specifically, this module uses the average heart rate and blood pressure data from the patient's preoperative stable period as an individual reference baseline. During procedures that easily induce vagal reflexes, such as removing the arterial sheath or performing high-pressure balloon dilation, the time-series data analysis model continuously receives heart rate and blood pressure data within a preset time window and calculates their dynamic decline gradient and associated change characteristics. When the model determines that the dynamic rate of change of heart rate relative to the baseline decreases beyond a preset safety ratio, and is accompanied by a continuous downward trend in mean arterial pressure, the system determines that the patient meets the early evolutionary characteristics of a vagal reflex. At this point, before the aforementioned physiological values ​​fall below the fixed alarm lower limit set by a conventional multi-parameter monitor, the central control unit outputs a visual warning signal and issues a corresponding voice prompt through the screen of the human-computer interaction module. This early warning mechanism transforms passive over-limit alarms for a single parameter into early trend predictions based on the combined deep feature extraction of the temporal change trends of multiple physiological parameters. This provides medical staff with time to prepare and administer emergency drugs in advance, thereby reducing the risk of severe hypotension or shock in patients during interventional procedures.

[0076] To enable the intelligent vital signs early warning module to possess the aforementioned forward-looking early warning capabilities, the construction and offline training process of the time-series data analysis model is as follows: First, a large number of continuous vital signs data sequences from patients who have undergone previous interventional surgeries are extracted from the hospital's historical ECG monitoring database to establish a physiological time-series dataset. In the data preprocessing stage, continuous sampling points of heart rate and invasive arterial blood pressure within a specific time window (e.g., the first 3 to 5 minutes) before the diagnosis of complications such as vagal nerve reflexes are extracted as positive sample sequences, and isochronous sampling points within the normal stable period are extracted as negative sample sequences, and normalization is performed. In terms of model architecture, a Long Short-Term Memory (LSTM) network, capable of handling long-term dependencies in sequence data, is selected. During the model training stage, the positive and negative sample sequences are input into the LSTM network in batches according to a set sliding time window size. The network, through its internal forget gate, input gate, and output gate structure, extracts the coupling evolution characteristics of the heart rate decrease gradient and blood pressure decrease rate in the time dimension. During training, a binary cross-entropy loss function is used in conjunction with backpropagation to update network weights, enabling the model to fully learn and fit the nonlinear mapping relationship between the joint decline trend of multiple physiological parameters and the probability of complication. After the model training converges, the network weight parameters are exported and packaged into an early warning engine, which is then deployed within the system to perform real-time online inference.

[0077] To accurately assess critical situations and filter out artifacts caused by electrosurgical interference or wire pulling, the system constructs a vital sign instability index based on multi-source data fusion. Its calculation logic involves normalizing and dynamically weighting the deviations of various physiological parameters. In the formula, and These are real-time heart rate and mean arterial pressure, respectively. The patient's baseline heart rate (taken from the mean of preoperative stable period records). The patient's mean arterial pressure baseline value (taken from the mean value recorded during the preoperative stable period). The first two terms in this formula have the same structure, both calculating the absolute deviation rate of the real-time value relative to the baseline value, thereby unifying vital sign data of different dimensions to the same relative scale of change; This represents the real-time percentage of blood oxygen saturation. The probability of arrhythmia (inferred from ECG waveforms using a lightweight CNN network, such as identifying ventricular fibrillation (VF) and ventricular tachycardia (VT)); These are weighting coefficients dynamically configured according to different surgical types, used to assign different sensitivity priorities to different complication induction indicators. Specifically, the meanings of each weighting coefficient are as follows: Characterizes sensitivity to absolute deviation of heart rate. Characterizes the sensitivity to absolute deviation of mean arterial pressure. Characterizes sensitivity to the amount of decrease in blood oxygen saturation. This characterizes the sensitivity to the probability of cardiac arrhythmias. The aforementioned weighting coefficients typically satisfy the normalization condition in practical applications (i.e., ...). The following are examples of dynamically configured values ​​for different interventional surgical scenarios: Coronary intervention and catheter removal scenario: During procedures such as removing arterial sheaths or high-pressure balloon dilation that easily induce vagal reflexes, since the typical characteristic of the vagal reflex is a double decrease in heart rate and blood pressure, the system can dynamically configure the weights as follows: , , , This configuration significantly amplifies the proportion of heart rate and blood pressure fluctuations in the instability index calculation, thereby improving the early warning sensitivity to the early evolution trend of the vagal reflex.

[0078] In high-risk arrhythmia scenarios such as radiofrequency ablation: when treating surgical lesions that are highly likely to induce malignant arrhythmias, the system can adjust the weighting to... , , , This configuration assigns a very high response priority to arrhythmias such as ventricular fibrillation (VF) or ventricular tachycardia (VT) inferred from a lightweight CNN network, ensuring that the system can quickly reach the critical threshold and switch to emergency assistance mode.

[0079] Interventional scenarios involving deep sedation: When patients receive deep sedation or have underlying severe respiratory diseases, the risk of respiratory depression leading to hypoxemia during the procedure increases. In such cases, the system can adjust the weights to... , , , The focus is on strengthening the monitoring of indicators that indicate a continuous decline in blood oxygen saturation.

[0080] The system internally maintains a finite state machine (FSM). When it detects... Exceeding the critical threshold for 5 consecutive seconds When a clear ventricular fibrillation / ventricular tachycardia feature is identified, the system status immediately switches from routine support to emergency mode.

[0081] In this mode, the robot automatically executes the following high-priority strategy: the robot immediately interrupts its current navigation or grasping task. Utilizing the kinematic advantages of its omnidirectional mobile chassis, it quickly calculates its spatial topological relationship relative to the operating table head and defibrillator cart, planning a shortest collision-free evacuation path with the maximum permissible acceleration (…). Retreat to a corner of the non-clean area of ​​the catheterization lab and retract the robotic arm to a folded position. Announce via voice, "We are clearing the way, please provide first aid," ensuring that the space for medical personnel to perform endotracheal intubation or chest compressions is not obstructed.

[0082] If the system database shows that the robot's storage compartment contains adrenaline, atropine, or defibrillation electrode pads, the robot will automatically pop open the corresponding storage layer and turn on a red strobe light to indicate to the nurse that the medication can be retrieved at any time during evacuation.

[0083] The system automatically locks all panoramic videos, voice interaction recordings, and equipment operation logs within 5 minutes before and after the emergency response is triggered, timestamps them, and uploads them to the hospital server for archiving, providing objective evidence for postoperative medical review or legal proof.

[0084] To meet the power requirements of continuous multi-operation runs around the clock, this system is equipped with a high-safety lithium iron phosphate (LiFePO4) battery pack with a nominal voltage of 48V and a capacity of 60Ah. The BMS (Battery Management System) has an active balancing function. The system is designed with an opportunistic charging strategy: when the task queue is empty and the battery level is below 70%, the robot automatically navigates to a non-contact wireless charging station (based on magnetic coupling resonance technology, charging power of 3kW, efficiency >90%) in the corner of the catheterization lab for recharging; once a new wake-up command is received, the charging circuit is immediately cut off within milliseconds to resume operation.

[0085] Considering the presence of strong electromagnetic interference sources such as high-frequency electrosurgical units and radiofrequency ablation devices within the catheterization chamber, the system incorporates rigorous EMC design: all electronic components within the chassis are housed in conductive anodized aluminum alloy shielded boxes, with the boxes securely integrated with the chassis. All external cables utilize double-shielded twisted-pair cables, with ferrite cores installed at the ports.

[0086] A multi-stage LC low-pass filter and TVS (transient voltage suppressor) array are designed at the power input terminal and the signal terminals of key sensors (such as force sensors and IMUs) to withstand the load. Group pulse interference and The electrostatic discharge (ESD) meets the requirements of IEC 60601-1-2 standard, ensuring that the robot does not freeze or malfunction during the defibrillator discharge.

[0087] To further illustrate the collaborative mechanism between the modules of this invention, the specific workflow and control logic of the intelligent nursing robot system in actual interventional surgery in this embodiment are shown in Figure 2. It executes the following two core business processes in parallel: Firstly, in the instrument delivery workflow: when the central control unit receives the instrument retrieval instruction from medical personnel, it schedules the mobile robot platform to autonomously navigate to the instrument table; subsequently, the intelligent instrument identification and delivery module is activated, sequentially performing preliminary identification and positioning through global visual image acquisition and deep learning models, guiding the robotic arm to perform flexible grasping; after the instrument leaves the table, it triggers local visual image acquisition and deep learning models for secondary verification; if the verification passes, it controls the robotic arm to perform precise delivery to the surgeon and reports the task completion status; if the verification fails, it reports an error and re-executes the identification and grasping process.

[0088] On the other hand, in the vital signs early warning workflow: the intelligent vital signs early warning module continuously acquires data streams from vital signs monitoring devices and inputs them into a deep learning time-series data analysis model for real-time inference; when the model determines that dynamic changes in the patient's physiological characteristics trigger warning thresholds for complications such as vagal reflexes, it immediately triggers an active warning, providing an audible / visual / screen alarm through the human-computer interaction module, and coordinating with the central control unit to interrupt current routine tasks to respond to emergency assistance. The status of the above dual-line workflow is synchronized in real time to the human-computer interaction module for status display.

[0089] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A deep learning-based intelligent nursing robot system for assisting interventional surgery, characterized in that, include: The central control unit is used to process data and send control commands to schedule the coordinated operation of the entire system; A mobile robot platform for autonomous movement within the catheterization lab; A multifunctional collaborative robotic arm, mounted on the mobile robot platform, has an end effector equipped with a flexible, damage-resistant gripper with integrated force feedback, used for non-destructive grasping and delivery of interventional surgical instruments under the control of the central control unit. An intelligent instrument identification and delivery module, including a machine vision unit and a deep learning-based visual differentiation model for similar catheters, is used for fine-grained feature extraction and classification of multiple highly similar interventional instruments on the instrument table, and performs at least two verifications before grasping and delivery to confirm the correctness of the delivery. A vital signs intelligent early warning module communicates with the vital signs monitoring equipment in the catheterization lab, and incorporates a deep learning-based time-series data analysis model to acquire real-time vital signs data of patients, analyze trends in vital signs changes, and issue early warnings before complication thresholds such as vagal reflexes are reached. A human-machine interaction module receives operator commands and provides status feedback through screen display and indicator lights.

2. The intelligent nursing robot system for interventional surgery based on deep learning according to claim 1, characterized in that, The mobile robot platform is equipped with an environmental perception and navigation unit, which includes a lidar and a depth camera. The central control unit is configured to use the depth camera to identify the base geometry of the C-arm machine in the duct room and the features of the hanging rail on the ceiling. The unit then uses the ICP algorithm to register the real-time perceived feature point cloud with a pre-stored reference template to calculate the absolute pose correction. The absolute pose correction is then introduced as a strong observation constraint into the particle filter of the adaptive Monte Carlo localization algorithm. The central control unit is also configured to execute a predictive obstacle avoidance strategy: real-time acquisition of the joint angles and angular velocities of the C-arm, prediction of the set of spatial regions swept by each rigid body of the C-arm within a future time window using a parametric kinematic model, i.e., the scan volume, mapping the scan volume to a local cost map and marking it as a dynamic no-entry layer; when a planned path is detected to cross the scan volume, the movement strategy is dynamically adjusted according to the collision time.

3. The intelligent nursing robot system for interventional surgery based on deep learning according to claim 1, characterized in that, The multifunctional collaborative robotic arm is a redundant seven-degree-of-freedom robotic arm with zero spatial degrees of freedom; the flexible anti-damage gripper integrates an array of piezoresistive sensors, which can sense the pressure distribution map on the contact surface; the central control unit is configured to monitor the peak pressure in the pressure distribution map, and if the peak pressure exceeds the preset yield strength threshold of the conduit material, the clamping current of the gripper is reduced in a closed loop; the central control unit is also configured to perform frequency domain analysis on the pressure distribution map, and when a high-frequency micro-vibration signal along the direction of gravity is detected at the pressure center, it is determined to be a slip precursor, and the clamping force is dynamically increased in small increments until the micro-vibration signal disappears.

4. The intelligent nursing robot system for interventional surgery based on deep learning according to claim 1, characterized in that, The human-computer interaction module includes a non-contact intelligent interaction terminal, which has a built-in non-contact gesture recognition module for capturing the surgeon's hand movements in the sterile area. The system executes a two-level confirmation mechanism: it enters the command listening state by recognizing a specific wake-up gesture; then, within the effective time window, it performs semantic parsing on the gesture flow of the spatial skeleton model and maps it into the corresponding robotic arm control commands.

5. The intelligent nursing robot system for interventional surgery assistance based on deep learning according to claim 1, characterized in that, The mobile robot platform is specifically an omnidirectional mobile chassis equipped with Mecanum wheels; the system also includes a sterile consumable storage compartment, which is located inside the omnidirectional mobile chassis, has a controllable opening sealed door, and integrates a micro positive pressure air supply unit inside; the central control unit is configured to maintain the air pressure inside the compartment at a slightly positive pressure state by adjusting the fan speed of the micro positive pressure air supply unit, and to force the fan to run at full power before the sealed door is opened to form a high-speed clean airflow barrier at the compartment door.

6. The intelligent nursing robot system for interventional surgery based on deep learning according to claim 5, characterized in that, The sterile consumables storage compartment is equipped with a multi-polarized antenna array and an ultra-high frequency RFID reader / writer. The RFID reader / writer is configured to change the standing wave distribution mode of electromagnetic waves in the compartment by rapidly switching the carrier phase of each transmitting antenna. The central control unit is configured to calculate the presence confidence of electronic tags by using received signal strength indication and Kalman filtering smoothing processing, thereby automatically identifying and retrieving consumables and generating inventory change reports to be uploaded to the hospital's material management system.

7. The intelligent nursing robot system for interventional surgery assistance based on deep learning according to claim 1, characterized in that, The surface of the mobile robot platform is covered with a radiation-proof coating of a flexible lead polymer plate made by calendering lead powder and ethylene-vinyl acetate copolymer matrix; the system also includes a radiation dose monitor for real-time detection of the ambient X-ray dose rate; the central control unit is configured to associate the real-time dose rate with the robot's spatial position to construct a dynamic radiation thermogram, and to include the radiation dose rate as a cost term in the cost function during path planning.

8. The intelligent nursing robot system for interventional surgery assistance based on deep learning according to claim 1, characterized in that, The human-computer interaction module is equipped with a microphone ring array; the central control unit is configured to run an adaptive beamforming algorithm, which constructs a fixed beamformer, a blocking matrix, and an adaptive noise canceller submodule in the signal processing logic; the fixed beamformer is used to form a main lobe pointing towards the surgeon's position, the blocking matrix is ​​used to block the desired direction signal and output a noise reference, and the adaptive noise canceller uses an adaptive algorithm to adjust the weights so that the output noise approximates the noise component in the fixed beamformer branch, thereby forming a null in the direction of the interference noise source; the system also integrates a high-frequency spectrum prediction and reconstruction model based on generative adversarial networks to improve the accuracy of speech recognition caused by medical staff wearing masks.

9. The intelligent nursing robot system for interventional surgery assistance based on deep learning according to claim 1, characterized in that, The central control unit analyzes the patient's heart rate, invasive arterial blood pressure, blood oxygen saturation, and electrocardiogram waveform data in real time through an isolation gateway. The central control unit is configured to calculate the vital signs instability index, and its calculation logic is to dynamically weight and fuse the heart rate deviation rate, mean arterial pressure deviation rate, blood oxygen saturation decrease, and the probability of arrhythmia inferred using a lightweight convolutional neural network. The system maintains a finite state machine. When the vital signs instability index exceeds the critical threshold for a continuous preset time or ventricular tachycardia / ventricular fibrillation characteristics are identified, the system automatically switches to emergency assistance mode. In emergency assistance mode, the mobile robot platform is controlled to plan the shortest collision-free evacuation path to retreat to the non-clean area of ​​the catheterization lab, retract the multi-functional collaborative robotic arm, and automatically open the sterile consumable storage compartment containing emergency medications.

10. The intelligent nursing robot system for interventional surgery assistance based on deep learning according to claim 1, characterized in that, The intelligent instrument identification and delivery module executes a dual verification logic: before grasping, it uses a global vision camera to obtain a global image of the instrument table to extract candidate regions for interventional surgical instruments; after the robotic arm grasps the instrument and removes it from the table, it triggers a local vision camera at the end to capture a close-up image for secondary classification and comparison; the similar catheter visual differentiation model has an embedded spatial attention module, and its network weights are obtained through offline iterative training using a metric learning strategy that includes a triplet loss function. The training process is configured to increase the inter-class feature distance of different types of catheters in the feature mapping space and reduce the feature differences of the same type of catheter caused by the environment, thereby performing fine-grained classification of the catheter tip and tail connector.