A single-arm active assistive exoskeleton system for cardiopulmonary resuscitation

Abstract

This invention relates to the field of medical upper limb assistive exoskeletons, and discloses a single-arm active assistive exoskeleton system for cardiopulmonary resuscitation (CPR), comprising: a wearable exoskeleton structure including a back exoskeleton and an upper limb exoskeleton; the back exoskeleton is connected to the upper limb exoskeleton, the upper limb exoskeleton including a shoulder exoskeleton structure, an arm exoskeleton structure, a wrist exoskeleton structure, and a palm exoskeleton structure connected in sequence; a motor drive mechanism including joint motors disposed at the shoulder and elbow for providing active assistance for the upper limb compression action; a sensing unit disposed at the joints of the back exoskeleton, the upper limb exoskeleton, and the palm exoskeleton structure for acquiring real-time motion state information of the rescuer's upper limb and compression state information of the palm; and a control module connected to the motor drive mechanism and the sensing unit for adjusting the motor output in real-time according to the motion state information and compression state information to assist the rescuer in maintaining standard CPR compression operations.

Description

Technical Field

[0001] This invention relates to the field of medical upper limb assistive exoskeleton technology, and in particular to a single-arm active assistive exoskeleton system for cardiopulmonary resuscitation. Background Technology

[0002] Cardiopulmonary resuscitation (CPR) is an emergency procedure that demands a high level of physical and technical skill from the rescuer. Especially when performing high-quality chest compressions, rescuers are required to maintain a steady rhythm, appropriate depth, and sufficient frequency of continuous compressions, typically at a rate of 100 to 120 compressions per minute and a depth of 5 to 6 centimeters – a high-intensity, repetitive motion. This procedure often leads to fatigue in the rescuer's upper limb muscles and postural instability, which in turn affects the quality of compressions and reduces the success rate. Therefore, there is an urgent need for assistive devices to reduce the burden on the rescuer's upper limb muscles and improve the quality and efficiency of compressions.

[0003] Currently, the most widely used CPR assistive devices are automatic chest compression devices, such as the LUCAS (Lund University Cardiac Assist System) and Autopulse Resuscitation System from abroad; the E7 / E8 series of electric and electronically controlled cardiopulmonary resuscitation machines from Anbao Medical, and the WFS-02B manual chest compression resuscitation device from Anbei Medical; as well as the mechanical chest compression device with published patent number CN223183777U, the wearable defibrillator with published patent number CN220967891U, and the handheld defibrillator with published patent number CN221512780U. These devices are all fixed to the torso of the person being rescued, thus ensuring good consistency in compressions. However, this fixed form has the following obvious limitations: 1. The equipment is bulky and heavy, which is not conducive to mobile deployment and is limited in use in ambulances, narrow spaces, or complex post-disaster environments; 2. The patient needs to be fixed before operation, which takes a long time to prepare and is not suitable for dynamic judgment and rapid response scenarios in emergency care; 3. The patient's body shape is poorly adapted, and there are problems such as failure to fit properly or excessive error, which affects the compression effect; 4. The equipment has low adjustability during compression and lacks a feedback mechanism for clinical judgment.

[0004] In addition to the aforementioned automated compression devices specifically designed for CPR, there are many exoskeleton-type assistive devices on the market, such as the shoulderX series upper limb assistive exoskeleton and Ottobock. Paexo series industrial exoskeletons; domestically, there are exoskeletons such as the UGO lower limb rehabilitation exoskeleton from Chengtian Technology; and upper limb assistive exoskeleton devices with published patent numbers CN114714325A and CN108839006A, and lower limb assistive exoskeleton devices with published patent numbers CN106109186A and CN114260879A. However, these devices are mostly used in medical rehabilitation or industrial scenarios and have the following limitations: 1. Exoskeletons in the field of medical rehabilitation are mostly used to assist patients with limb dysfunction in training. Their wearing process is complex and time-consuming, making it difficult to meet the rapid wearing requirements of CPR rescue; 2. Exoskeletons used in industrial applications are highly action-oriented. Their structure is often designed for a specific action (such as bending over, raising the upper body, raising the arm, etc.). The assistance based on this is usually passive and lacks the ability to adjust in real time according to the work posture and intensity; 3. Exoskeletons in the above two scenarios are intended to assist large-amplitude, low-speed movements, which cannot meet the high-frequency, low-amplitude operation requirements of CPR rescue. Summary of the Invention

[0005] To address the aforementioned problems, the purpose of this invention is to provide a single-arm active assistive exoskeleton system for cardiopulmonary resuscitation (CPR) to solve the technical challenges of decreased compression frequency and effectiveness caused by the high physical exertion of the rescuer and the difficulty in maintaining the quality of movements over a long period during existing CPR procedures.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a single-arm active assistive exoskeleton system for cardiopulmonary resuscitation (CPR), employing a single-arm assistive structure, with the other arm remaining idle for auxiliary positioning or system fine-tuning; the single-arm active assistive exoskeleton system for CPR includes: a wearable exoskeleton structure comprising a back exoskeleton and an upper limb exoskeleton; the back exoskeleton is connected to the upper limb exoskeleton, the upper limb exoskeleton comprising a shoulder exoskeleton structure, an arm exoskeleton structure, a wrist exoskeleton structure, and a palm exoskeleton structure connected in sequence; a motor drive mechanism including joint motors located at the shoulder and elbow, for providing active assistance for the upper limb compression movements; a sensing unit located at the joints of the back exoskeleton, the upper limb exoskeleton, and the palm exoskeleton structure, for acquiring real-time information on the rescuer's upper limb movement and palm compression; and a control module connected to the motor drive mechanism and the sensing unit, for adjusting the motor output in real-time based on the movement and compression information to assist the rescuer in maintaining standard CPR compression techniques.

[0007] Furthermore, the back exoskeleton includes a back plate, a rotatable locking structure, a hardware installation cavity, and an upper limb structure support arm. The back plate has an embedded Y-shaped internal skeleton support strip, and multiple layers of cushioning pads on its outer side. A hardware installation cavity is located at the center of the inner side of the back plate for modular installation of control and power modules. The rotatable locking structure is located on the outer side of the back plate. One end of the upper limb structure support arm is connected to the back plate via the rotatable locking structure, forming a rotatable support arm. The other end of the upper limb structure support arm is connected to the upper limb exoskeleton. An adjustable limiter is provided on the rotatable support arm to adjust the fixed angle of the support arm according to the rescuer's left or right hand habits. A shoulder strap installation structure is encircled on the back plate, with both ends connected to the sides of the back plate for securing and tightening the shoulder straps. Annular chest tightening mechanisms are located at both ends of the back plate for securing and tightening the chest straps.

[0008] Furthermore, the shoulder exoskeleton structure includes a spherical universal joint, which consists of two mutually perpendicular rotation axes to simulate the degrees of freedom of the human shoulder joint. A first motor base is located at the end of the spherical universal joint for mounting the shoulder joint motor. The arm exoskeleton structure includes an upper arm body and a forearm body. The upper arm body extends from the first motor base of the spherical universal joint to the elbow region, serving as a first retractable drive rod. An arm body fixing structure is located on the inner side of the upper arm body, and a second motor base is located at its end. The forearm body is located between the elbow and wrist joint, connected to the upper arm body via the second motor base at the end of the upper arm body, and serves as a second retractable drive rod. The retractable drive rod incorporates a magnetic linear encoder to detect changes in the rescuer's arm length for adaptive adjustment. The wrist exoskeleton structure includes a forearm connection structure, a semi-circular locking structure, and a wrist telescopic rod. It consists of a forearm end connected to a semi-circular locking structure. The forearm end has a mounting hole for installing the semi-circular locking structure, which is suspended below the forearm connection structure through the mounting hole. The semi-circular locking structure is a semi-circular bearing, with its shaft inserted into the mounting hole at the forearm end. One end of the wrist telescopic rod connects to the semi-circular locking structure, and the other end connects to the palm exoskeleton structure to accommodate different users. The palm exoskeleton structure includes a dual-sided universal palm glove structure and a pressure module mounting position. The other end of the wrist telescopic rod connects to the dual-sided universal palm glove structure, which is a fingerless glove. A palm support component, in the shape of a fitted plate, is provided at the palm of the dual-sided universal palm glove structure, with a surface covered with a flexible, high-friction material for stable support of the pressing area. A pressure module mounting position, containing a pressure monitoring module, is located at the bottom of the dual-sided universal palm glove structure.

[0009] Furthermore, the control module includes: a kinematics module, used to establish an analytical mapping relationship between the shoulder and elbow joint angles based on the end-compression target point, and to provide training data for the deep learning model, while also serving as a real-time inverse kinematics model for compensation; a compression depth compensation module, used to determine the rescuer's trunk displacement based on the real-time pressure value collected by the palm pressure monitoring module and the preset chest elasticity model, calculate the depth error and correct the end-compression target point, triggering the deep learning module to regenerate the angle and torque sequences driving the shoulder and elbow joint motor movements; and a deep learning module, which uses data collected by sensors in conjunction with the kinematic model... The block and dynamic modeling solves for high-precision motor angle and torque sequences, and then trains a deep learning model. The deep learning module deploys a trained multilayer perceptron (MLP) network, which is used to instantly match and generate a target angle and target torque sequence that drives the shoulder and elbow joint motors within a standard compression cycle based on the initial state information of the rescuer collected by the sensing unit and the corrected end-press target point. The motor drive control unit generates motor drive signals based on the target angle and target torque sequences, driving the joint motors to output downward pressure torque during the pressing phase and rebound torque during the lifting phase.

[0010] Furthermore, the input parameters of the deep learning module include static feature parameters and dynamic posture parameters; among them, the static feature parameters are the length of the upper arm exoskeleton and the length of the forearm exoskeleton, which are collected by a magnetic linear encoder set in the telescopic drive rod; the dynamic posture parameters are quaternion posture data collected by IMU posture sensors set at three locations: the top and side of the universal joint, and the connection between the forearm body and the upper arm body.

[0011] Furthermore, the output parameters of the deep learning module are one-dimensional vectors, which correspond to the target angle of the shoulder joint, the target angle of the elbow joint, the target torque of the shoulder joint, and the target torque of the elbow joint in each of the multiple time steps discretized within a standard pressing cycle.

[0012] Furthermore, the compression depth compensation module also includes an online chest elasticity model identification unit, which is used to establish the current patient's chest elasticity benchmark based on real-time pressure value and compression depth during the first three compression cycles of cardiopulmonary resuscitation, so as to serve as a reference for judging the rescuer's trunk displacement during subsequent compressions.

[0013] Furthermore, during subsequent compressions, the compression depth compensation module uses Hooke's Law to deduce the error in the final compression depth caused by the rescuer's trunk displacement based on the deviation between the real-time pressure value collected by the palm pressure monitoring module and the chest elastic reference. It then corrects the final compression target point based on the depth error. After correcting the final compression target point, the deep learning module is triggered to regenerate the target angle sequence and target torque sequence to achieve adaptive compensation for the rescuer's trunk displacement.

[0014] Furthermore, the motor drive control unit adopts a hybrid control strategy that combines feedforward and feedback. The target torque sequence is used as the feedforward command, and a compensation torque is generated by the PID controller based on the error between the target angle sequence and the actual angle in real time feedback. The feedforward command and the compensation torque are superimposed to generate the final motor drive signal, so that the joint motors at the shoulder and elbow output the downward pressure torque during the downward phase and the rebound torque during the lifting phase.

[0015] Furthermore, the deep learning module is obtained by collecting initial state vectors from multiple groups of experimental subjects of different body types under various typical initial postures as input, and using the standard compression cycle driving vector generated by the kinematics module and the compression depth compensation module as output labels, and then conducting offline training.

[0016] Because of the above technical solutions, this invention has the following advantages compared with existing CPR-assisted compression devices and exoskeleton assistive devices for other scenarios:

[0017] 1. The skeleton of this invention has few wearing points, making it convenient and quick to put on, and has a size adaptation device to fit different rescuer body types; the skeleton has sufficient ergonomic degrees of freedom, that is, it can be highly adapted to the human body after being worn, and the rescuer can flexibly adjust the body posture at the rescue site according to their own comfort and fatigue points; the compression degree of freedom of the skeleton is in the form of external assistance, which can actively adjust the compression frequency and force according to different rescuer postures; the ergonomic degrees of freedom of the skeleton and the compression degree of freedom are completely decoupled, and its compression stability will not be affected by changes in the rescuer's posture.

[0018] 2. The ergonomic degrees of freedom and compression degrees of freedom of the skeleton in this invention can be switched to a low-damping state to support the elbow movements required for artificial respiration. It is also equipped with a posture monitoring system, which can monitor the rescuer's movements in real time and provide feedback on compression frequency and depth information to ensure the standardization of compressions; it is also equipped with deep learning function to achieve human-machine collaboration in compression frequency and force, that is, while maintaining the advantage of human perception, it effectively provides assistance to reduce the fatigue of the rescuer.

[0019] 3. The shoulder ergonomic freedom of this invention has excellent flexibility and stability: the shoulder universal joint allows the rescuer to freely rotate the shoulder according to the characteristics of the site and their own comfort, thereby flexibly adjusting the posture of the entire arm; the two rotating shafts of the shoulder universal joint are orthogonal to the motor shafts of the shoulder and elbow, avoiding the motor base posture from being offset by the reaction of the motor torque, ensuring the stability of the motor base, and thus ensuring the stability of the motor output torque.

[0020] 4. The shoulder and elbow pressing degrees of freedom in this invention possess responsiveness and stability: the output shafts of the shoulder and elbow motors always remain parallel, ensuring the system's force transmission efficiency and stability; the elbow motor base is located at the output end of the shoulder motor, so when adjusting the shoulder posture, the two motors of the shoulder and elbow can adjust synchronously. The shoulder universal joint structure can conform to the rescuer's shoulder movement trajectory, achieving adaptive matching to different arm posture changes.

[0021] 5. The wrist freedom of this invention is flexible and stable: the wrist universal joint structure allows the rescuer to adjust the angle of their hand according to their own comfort; the telescopic rod connected to the wrist universal joint, after being locked in length, can provide stable support for the wrist during the pressing action.

[0022] 6. The electronic control system of this invention adopts a dual-processor collaborative architecture. An STM32 microcontroller completes high-frequency data acquisition from various sensors, ensuring the real-time performance of core parameters such as posture, frequency, and compression depth. Simultaneously, a Jetson Nano is used to perform deep learning and generate rescue strategies, enabling personalized human-machine collaborative control. Through the fusion of data from multiple sensors, including the IMU posture sensor and the hand pressure sensor, the system can acquire real-time dynamic rescue data and compare it with standard CPR action requirements. When the rescuer deviates from the expected action, the system can immediately correct it through voice prompts and automatically adjust the motor output frequency, achieving dynamic compensation and precise control during compressions, thereby ensuring the standardization and consistency of chest compressions. When artificial respiration is required, the system can control the motors and joint structures to enter a decoupled state, cooperating with the wrist's quick-release device to provide ample operating space for the rescuer, avoiding the impact of exoskeleton constraints on the continuity of the rescue process. This mechanism significantly improves the safety and applicability of the exoskeleton in actual emergency scenarios. It provides local storage and timestamp management functions for rescue data, enabling complete tracking of the rescue process and providing objective evidence for quality assessment and medical research. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall structure of the upper limb assistive exoskeleton in an embodiment of the present invention; Figure 2 This is a schematic diagram of the shoulder structure of the upper limb assistive exoskeleton in an embodiment of the present invention; Figure 3 This is a schematic diagram of the back structure of the upper limb assistive exoskeleton in an embodiment of the present invention; Figure 4 This is a schematic diagram of the rotatable locking structure of the back exoskeleton in an embodiment of the present invention; Figure 5 This is a schematic diagram of the upper limb assistive exoskeleton worn in an embodiment of the present invention; Figure 6 This is a schematic diagram of the electrical control system cavity in the back exoskeleton of this invention. Figure 7 This is a schematic diagram of the shoulder universal joint connection component of the upper limb assistive exoskeleton in an embodiment of the present invention; Figure 8 This is a schematic diagram of the wrist joint and hand support components of the upper limb assistive exoskeleton in an embodiment of the present invention; Figure 9 This is a schematic diagram illustrating how the upper limb assistive exoskeleton can be adapted to different rescuers' preferred wearing habits in an embodiment of the present invention; Figure 10 This is a schematic diagram illustrating the comfort effect of the shoulder universal joint of the upper limb assistive exoskeleton in an embodiment of the present invention; Figure 11 This is a schematic diagram illustrating the comfort effect of the shoulder universal joint of the upper limb assistive exoskeleton in an embodiment of the present invention; Figure 12 This is a schematic diagram illustrating the assistive effect of the upper limb assistive exoskeleton on the right hand in an embodiment of the present invention; Figure 13 This is a schematic diagram illustrating the left-handed assistance effect of the upper limb assistive exoskeleton in an embodiment of the present invention; Figure 14 This is a schematic diagram of the upper limb assistive exoskeleton system in an embodiment of the present invention; Figure 15 This is a schematic diagram illustrating the establishment of the kinematic coordinate system of the upper limb-assisted exoskeleton MDH in an embodiment of the present invention; Figure 16 A partial frontal view is established for the kinematic coordinate system of the upper limb-assisted exoskeleton MDH in this embodiment of the invention; Figure 17 A partial side view is established for the kinematic coordinate system of the upper limb-assisted exoskeleton MDH in this embodiment of the invention; Figure 18 This is a schematic diagram of the maximum pressing position of the upper limb assistive exoskeleton arm in four initial postures in an embodiment of the present invention; Figure 19 This is a schematic diagram of the deep learning model (MLP) network architecture of the upper limb assistive exoskeleton in an embodiment of the present invention; Figure 20 This is a schematic diagram of the sample acquisition and network training process of the upper limb assistive exoskeleton deep learning model (MLP) in an embodiment of the present invention; Figure 21 This is a schematic diagram of the motor control system of the upper limb assistive exoskeleton based on a hybrid strategy in an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0025] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0026] In one embodiment of the present invention, a single-arm active assistive exoskeleton system for cardiopulmonary resuscitation (CPR) is provided. This active assistive exoskeleton can naturally coordinate with the rescuer's upper limb during rescue operations. Through an integrated motor drive unit and sensor module, it provides active assistance support for chest compressions and senses the pressure and frequency of compressions in real time. This helps the rescuer maintain high-quality, consistent chest compressions, reduces operational deviations caused by muscle fatigue, and improves the overall effectiveness and reliability of CPR. In this embodiment, a single-arm assistive structure is used, allowing for interchangeable use of the left and right hands, with the other arm remaining idle for positioning assistance or system fine-tuning. Simultaneously, the single-arm assistive structure significantly reduces the overall weight of the system, and the idle arm can assist in arm stability and allow for free access to objects. Figures 1 to 21 As shown, the single-arm active assistive exoskeleton system for cardiopulmonary resuscitation includes: a wearable exoskeleton structure, a motor drive mechanism, a sensing unit, and a control module.

[0027] A wearable exoskeleton structure is used to fit and connect with the rescuer's key upper limb joints (including shoulder, elbow and wrist joints). It includes a back exoskeleton and an upper limb exoskeleton. The back exoskeleton is connected to the upper limb exoskeleton, which includes a shoulder exoskeleton structure, an arm exoskeleton structure, a wrist exoskeleton structure and a palm exoskeleton structure connected in sequence.

[0028] The motor drive mechanism includes joint motors located at the shoulder and elbow, which provide active assistance for the pressing motion of the upper limbs. The joint motors execute assistance commands issued by the control module, providing power assistance for the lifting and pressing processes.

[0029] The sensing unit is installed on the joints of the back exoskeleton, the upper limb exoskeleton, and the palm exoskeleton structure to acquire real-time information on the movement of the rescuer's upper limbs and the pressure of the palm.

[0030] The control module, connected to the motor drive mechanism and sensing unit, adjusts the motor output in real time based on motion and compression information to assist rescuers in maintaining standard CPR compression procedures.

[0031] When in use, the wearing effect of the wearable upper limb assistive exoskeleton system is as follows: Figure 5 As shown. Addressing the high demands for efficiency, ergonomics, and operational flexibility during emergency rescue, the design incorporates targeted structural optimizations: When wearing this active assistive exoskeleton, the standard rescue posture for the rescuer is a semi-forward lean, performing chest compressions with one arm. This posture is suitable for kneeling or standing positions, requiring only that the rescuer's upper limb exoskeleton compressions are perpendicular to the patient's chest compression point, thus ensuring that the compression trajectory and direction comply with the rescue requirements of the "Guidelines for Cardiopulmonary Resuscitation." The system employs a modular design, with all functional modules capable of independent installation and disassembly. The main connection points of the exoskeleton are equipped with Velcro straps and snap-on closure mechanisms. The user simply needs to carry the back module on their back and sequentially wear and adjust the shoulder straps, chest strap, and upper limb components to complete the overall donning process. When wearing the exoskeleton, starting with the back module, the rescuer places the exoskeleton in a fixed position and secures it using the chest tightening mechanism located in front of the chest.

[0032] In a preferred embodiment, the back exoskeleton balances load-bearing capacity, counterweight stability, and comfort. (e.g., back exoskeleton) Figure 3 (As shown) includes a back plate 4, a rotatable locking structure 7, a hardware installation cavity (and an upper limb structure support arm 5, which is mainly used to support the control unit, energy supply unit and upper limb exoskeleton structure of the system).

[0033] The back panel 4 has an embedded Y-shaped internal skeleton support strip, and the outer side of the back panel 4 is provided with multiple layers of cushioning pads. The inner center of the back panel 4 has a hardware installation cavity for modular installation of control modules and power modules. In this embodiment, the back panel 4 is a rigid load-bearing back panel, positioned close to the human body, with an overall curved, form-fitting design based on the natural curvature of the human spine. The outer shell of the back panel 4 is made of lightweight carbon fiber, with a flexible inner section to improve wearing comfort. The outer multi-layer cushioning pads include a honeycomb EVA foam layer, a breathable mesh layer, and an antibacterial lining material. From the outside in, the back panel 4 consists of multiple layers of cushioning pads, an antibacterial lining material, and a Y-shaped skeleton.

[0034] like Figure 3As shown, the rotatable locking structure 7 on the back is located on the outside of the back plate 4.

[0035] like Figure 3 As shown, one end of the upper limb structural support arm 5 is connected to the back plate 4 via a rotatable locking structure 7 on the back, so that the upper limb structural support arm 5 forms a rotatable support arm; the other end of the upper limb structural support arm 5 is connected to the upper limb exoskeleton. Figure 4 As shown, the rotatable support arm is equipped with an angle-adjustable limiter 72, which is used to adjust the fixed angle of the support arm according to the rescuer's left and right hand habits.

[0036] like Figure 1 As shown, the shoulder strap mounting structure 3 is encircled on the back plate 4, with both ends of the shoulder strap mounting structure 3 connected to the sides of the back plate 4 for securing and tightening the shoulder straps. Annular chest tightening mechanisms 4 are located at both ends of the back plate 4 for securing and tightening the chest straps. In this embodiment, both the shoulder strap mounting structure 3 and the chest tightening mechanism 4 employ a button-type buckle structure, allowing the wearer to tighten and lock the shoulder straps and chest straps with a single hand, significantly improving wearing efficiency and stability.

[0037] In the above embodiments, in this embodiment, as Figure 4 As shown, the rotatable locking structure 7 on the back is composed of a limited toothed rail 71 and an adjusting limiter 72; the adjusting limiter is set on a toothed groove on the outer side of the back plate 4.

[0038] The adjustable limiter 72 structure includes a toothed U-shaped block and a top eccentric pressure rod mounted on the U-shaped block. The U-shaped block engages and disengages with the teeth in the groove, adjusting the angle range from 0° to 180°. Locking and loosening are achieved via the top eccentric pressure rod, thus adapting to the angle adjustment of the upper limb support arm 5. This structure primarily caters to the rescuer's habits, specifically accommodating individual differences in whether CPR is performed using the left or right arm. The user can rotate the structure with their other hand to adjust it to a left or right rescue position. Figure 9 The diagram illustrates the left-handed and right-handed rescuers wearing the present invention. The rescuers follow... Figure 9 Simply adjust the icons in the middle right corner to switch arms. After rotation and adjustment, the support arm can be instantly locked by the self-locking mechanism mentioned above to prevent displacement caused by loosening.

[0039] In the above embodiments, such as Figure 6As shown, a voice interaction module connected to the control module is installed in the hardware installation cavity of the back panel 4. This module provides real-time prompts when rescue actions are not performed correctly or when physical strength is waning. The power module uses a small-sized solid-state battery. During use, the rescuer wears the complete exoskeleton and manually powers it on. The system first enters the sensor self-check and posture calibration phase. If the rescuer's posture does not conform to the CPR operation range at this time, the system will correct it through voice prompts. Once the rescue posture meets the requirements, the monitoring module continuously collects data on compression frequency, depth, posture, and motor status, and transmits the information to the top-level processing unit (NANO).

[0040] In the above embodiment, the upper limb structure support arm 5 has a hollow structure inside to facilitate the wiring of the electrical control part of the upper arm exoskeleton.

[0041] In a preferred embodiment, such as Figure 2 The shoulder exoskeleton structure shown is a shoulder joint connection assembly, which includes a spherical universal joint (such as...). Figure 7 As shown), this type of spherical universal joint consists of two mutually perpendicular rotation axes, used to simulate the degree of freedom of movement of the human shoulder joint. Specifically, it can realize movements such as forward and backward swinging, lifting and pressing, and internal and external rotation. Functionally, it simulates the spherical degree of freedom of movement of the human shoulder joint, ensuring that the upper limb exoskeleton can complete the compression trajectory and angle adjustment required for standard cardiopulmonary resuscitation (CPR) movements, and realize the arm's pitching and swinging functions and adapt to the displacement error generated in the shoulder during arm movement. This allows the rescuer to still have good wearing comfort without deliberately adapting to the resistance torque generated by arm assistance (e.g., Figure 10 , 11 (As shown). A first motor base 12 is provided at the end of the spherical universal joint for mounting the joint motor of the shoulder.

[0042] like Figure 1 The illustrated arm exoskeleton structure includes an upper arm body 1 and a forearm body 2. The upper arm body 1 extends from a first motor base 12 (quasi-spherical universal joint) to the elbow region. The upper arm body 1 is a first retractable drive rod 11. An arm body fixing structure 10 is located on the inner side of the upper arm body 1, and a second motor base 22 is located at its end. The arm body fixing structure 10 is an elastic wrapping structure consisting of fixing straps and cushioning pads. During use, it can be secured with Velcro to ensure the exoskeleton is firmly attached to the user's upper arm surface. The forearm body 2 is located between the elbow and wrist joint, and its structure is the same as the upper arm body 1. It is connected to the upper arm body 1 via the second motor base 22 at the end of the upper arm body 1. The forearm body 2 is a second retractable drive rod 21. The retractable drive rod has a built-in magnetic linear encoder used to detect changes in the length of the rescuer's upper arm and forearm for adjustment.

[0043] like Figure 1As shown, the wrist exoskeleton structure 8 includes a forearm connection structure, a semi-circular locking structure 81, and a wrist telescopic rod 82. The forearm connection structure consists of the end of the forearm body 2 connected to the semi-circular locking structure 81. The end of the forearm body 2 has a mounting hole for installing the semi-circular locking structure 81, which is suspended below the forearm connection structure through the mounting hole. The semi-circular locking structure 81 is a semi-circular bearing, with a rotating shaft inserted into the mounting hole at the end of the forearm body 2 to achieve rotational connection between the end of the forearm body 2 and the semi-circular locking structure 81. The ring-shaped locking structure 81 enables quick donning and locking. One end of the wrist telescopic rod 82 is connected to the semi-circular locking structure 81, and the other end is connected to the palm exoskeleton structure to accommodate different users.

[0044] like Figure 8 As shown, the palm exoskeleton structure includes a dual-sided universal palm glove structure 83 and a pressure module mounting position 84. The other end of the wrist telescopic rod 82 is connected to the dual-sided universal palm glove structure 83, which is a fingerless glove. A palm support assembly is provided at the palm of the dual-sided universal palm glove structure 83, which is in the form of a fitted support plate and is covered with a flexible high-friction material to stably support the pressing area. The pressure module mounting position 84 is located at the bottom of the dual-sided universal palm glove structure 83. The pressure module mounting position 84 is equipped with a pressure monitoring module to collect the palm pressing depth and force to determine whether it meets the CPR standard.

[0045] When in use, the semi-circular locking structure 81 is initially in the open state (e.g., Figure 8 (As shown in the left image), allowing the user to quickly insert their wrist; subsequently, the other free arm rotates the inner half-ring inside the locking structure, causing it to slide relative to the outer structure and close, forming a complete ring (i.e., entering the closed state, as shown in the left image). Figure 8 (As shown in the right figure), thus achieving rapid locking of the wrist; subsequently, the dual universal palm sleeve structure 83 moves to the palm position for wearing. The palm support component at the palm of the dual universal palm sleeve structure 83 is used to support the hand posture during rescue and ensure the stability of the compression path.

[0046] In the above embodiments, such as Figure 1 As shown, IMU attitude sensors 91, 92, and 93 are respectively installed at the top of the universal joint, the first motor base 12, and the second motor base 22 to monitor the wearer's posture and acceleration changes in real time, so as to reflect his movement intention.

[0047] In the above embodiments, the telescopic drive rods 11 and 21 are telescopic rods with adjustment functions, allowing the user to make quick fine adjustments according to their own arm length.

[0048] In summary, considering the operational flexibility and coordination during rescue operations, the wearable exoskeleton structure of this invention adopts a single-arm assist design, keeping the other arm always idle. This allows for auxiliary positioning of the patient's chest and confirmation of the compression area, and also facilitates real-time fine-tuning of the exoskeleton system during rescue operations, such as arm length adjustment or position calibration. This design maximizes the rescuer's operational freedom without sacrificing the assistive effect, enhancing the system's applicability and coordination in real-world rescue scenarios.

[0049] In a preferred embodiment, the control module includes a deep learning module, a kinematics module, and a pressure depth compensation module. Wherein: The kinematics module is used to establish an analytical mapping relationship between the shoulder and elbow joint angles based on the end-press target point, and to provide training data for the deep learning model. It also serves as a real-time inverse kinematics model for compensation. In this embodiment, the kinematics module provides data support for the output of the deep learning training. However, due to the complex kinematics inverse model requiring continuous real-time solving, high-frequency computation can lead to significant control delays, causing the motor response to lag behind the actual pressing rhythm. To address this issue, this embodiment abandons continuous real-time model solving and adopts a control strategy based on neural network gap prediction and periodic generation.

[0050] The compression depth compensation module is used to determine the rescuer's torso displacement based on the real-time pressure value collected by the palm pressure monitoring module and the preset chest elasticity model, calculate the depth error and correct the end compression target point, and trigger the deep learning module to regenerate the angle sequence and torque sequence that drive the shoulder and elbow joint motors.

[0051] The deep learning module uses data collected by sensors, along with kinematics and dynamics modeling, to calculate high-precision motor angle and torque sequences, and then trains a deep learning model. The deep learning module is equipped with a trained multilayer perceptron (MLP) network, which is used to instantly match and generate a target angle and torque sequence that drives the shoulder and elbow joint motors within a standard compression cycle based on the rescuer's initial state information collected by the sensing unit and the corrected end-press target point.

[0052] The motor drive control unit generates motor drive signals based on the target angle sequence and the target torque sequence, driving the joint motor to output downward pressure torque during the downward pressing phase and rebound torque during the lifting phase.

[0053] In the above embodiments, to achieve precise tracking and high-frequency assistance of the rescuer's chest compression movements, the upper limb assistive exoskeleton system of this invention is modeled and analyzed kinematically and dynamically using a kinematic module. This establishes a mapping relationship between the exoskeleton's mechanical structure and the end-compression trajectory, providing a theoretical basis for subsequent deep learning models and control algorithms. In this embodiment, as... Figures 15 to 17 As shown, kinematic and dynamic modeling and analysis includes the following steps: S1. Parameter Acquisition: Obtain the link length information of the exoskeleton. In this embodiment, the link is the telescopic rod in the above embodiments, hereinafter collectively referred to as the link.

[0054] S2. Update the existing MDH parameter table and transformation matrix based on the measured link length and attitude information.

[0055] S3. Model Construction: Generate an inverse kinematic model that matches the current user information based on the above transformation matrix.

[0056] S4. Labeling the pressing task planning and kinematic inverse derivation: Using the above kinematic inverse model, the intention trajectory of the end effector is converted into the joint angle sequence of each joint of the exoskeleton.

[0057] S5. Construct and perform inverse dynamic solution and torque calculation through dynamic equations: Based on the above joint angle sequence and its derivatives combined with dynamic parameters, generate inverse dynamic equations to obtain the torques of each joint.

[0058] like Figure 16 As shown, in step S1 above, the link length information of the exoskeleton is obtained. This is acquired through the corresponding sensor module, specifically for the link length information (i.e., the length of the upper arm of the exoskeleton). Forearm length This is measured by a magnetic linear encoder placed in the connecting rod of the upper arm and forearm sections. For example... Figure 17 As shown, IMU posture sensors (such as the MPU9250) placed at the shoulder and elbow are used to acquire the current posture of the human body. Their purpose is to measure the quaternions of each link in the exoskeleton in three-dimensional space, and to use these quaternions to determine the angles of each joint in the initial posture, defined as... Used to build MDH models The yaw angle (YAW) represents the rotation of the shoulder IMU sensor around the Z-axis. This represents the yaw angle (YAW) of the elbow IMU sensor rotating around the Z-axis. The wrist joint provides passive assistance, allowing the rescuer's hand free movement to find the optimal compression position.

[0059] This section will introduce the method of measuring joint angles using an IMU: The core of this method is not to directly measure angles. Taking the shoulder as an example, it measures the absolute attitude of the IMUs set up on the torso and upper arm, then calculates their relative attitude, and finally converts it into joint angles through calibration. Specifically, it includes the following steps: 1. Obtain data on attitude relative to the world.

[0060] A 6-DoF fusion algorithm was run separately on IMU_T (torso) and IMU_A (upper arm), using only gyroscope angular velocity data and accelerometer gravity vector data. The attitude was corrected and output. The fusion algorithm used gravity vector correction to adjust the gyroscope's roll and pitch integrals, and output the attitude of each relative to the world coordinate system (W). (recorded as) )and (recorded as) ), where each pose The components contained in a vector are denoted as .

[0061] In this step, a magnetometer is not used to calibrate the yaw angle. This is because: (1) the gyroscope's integration of angular velocity will inevitably produce cumulative error (drift); (2) the accelerometer can only measure the direction of gravity, and therefore can only correct the drift of the Roll and Pitch axes; (3) although the magnetometer can correct the yaw angle, the exoskeleton's motors and metal frame will severely interfere with it. Therefore, the output and The absolute yaw component will inevitably drift over time.

[0062] The solution to this yaw angle drift is based on the fact that while the absolute yaw angles of both IMUs drift over time, their drift trends are highly consistent because they use the same fusion algorithm and inertial reference (gravity direction). In the relative attitude calculation step, the quaternion of the upper arm relative to the torso is calculated. These common drift components are mathematically canceled out. Therefore, from The yaw angle extracted reflects the rotation of the upper arm relative to the torso around the Z-axis (i.e., the relative yaw angle of the joint), rather than the absolute yaw angle in the world coordinate system. This relative yaw angle is stable over time and can accurately characterize the amplitude of upper limb flexion, extension, or elevation.

[0063] 2. Calculate the relative posture between IMUs (taking the shoulder joint angle as an example).

[0064] The posture of the upper arm relative to the torso is calculated using quaternion multiplication. :

[0065] in The value is equal to conjugate, This represents quaternion multiplication operations.

[0066] 3. Calibrate the zero-point posture.

[0067] The exoskeleton is held in a pre-defined "zero-point pose" (arms pointing vertically downwards). The relative pose of the upper arm IMU_A with respect to the torso IMU_T is measured and recorded using the method described above, and this is defined as the "initial zero-point quaternion". .

[0068]

[0069] 4. Real-time calculation of joint rotation.

[0070] During motion, instantaneous calculations are performed in real time. The rotation quaternion of the joint is obtained through calculation. :

[0071] Should The actual meaning it represents is: the change in posture of the upper arm link relative to the torso from the "zero point posture".

[0072] 5. Convert to Euler angles.

[0073] The joint rotation quaternion obtained above To maximize the computational efficiency of the STM32, the conversion to Euler angles is omitted, eliminating the need for Roll and Pitch value conversions. Only the required Yaw value calculations are performed.

[0074] To align with the specifications in subsequent MDH modeling (i.e.) Defined as the link about its axis of rotation In this invention, through the aforementioned zero-point calibration and coordinate transformation process, a measurement coordinate system is mathematically established. This coordinate system is aligned with the zero-point attitude of the MDH model. Therefore, the calculated yaw angle in this coordinate system physically represents the joint variable defined by the MDH model. The above changes over time Value directly assigned to This can be used as a real-time shoulder joint angle input in kinematic and dynamic models.

[0075] For elbow joint angle To obtain the forearm relative to the upper arm, simply replace the posture data of IMU_T (torso) and IMU_A (upper arm) in the above calculation process with the posture data of IMU_A (upper arm) and IMU_F (forearm).

[0076] Before proceeding with kinematic and dynamic modeling, a key simplification is first established: as described in step S1, the wrist joint is a "passively assisted" joint. Structurally, it is designed as a universal joint that adaptively conforms to the patient's chest without requiring motor assistance; therefore, only the shoulder joint needs to be incorporated into the modeling. ) and elbow joint ( These two active joints can be calculated. The wrist joint is considered as a passive degree of freedom, and its mathematical value is set to 0 in the calculation. This simplification reduces the model to a two-degree-of-freedom system. This is also the core basis for the inverse kinematics and dynamics solutions.

[0077] In step S2 above, the pre-generated MDH parameter table (as shown in Table 1) and transformation matrix are obtained by using an improved Denavit algorithm. Hartenberg notation (i.e., MD) H) As shown above, some of the data in Table 1 has already been introduced.

[0078] Table 1. Parameters of Exoskeleton MDH

[0079] In exoskeleton systems, link torsion angle Indicates Looking in that direction, arrive The included angle between them. In this model, the shoulder motor and elbow motor are mounted in two parallel planes, so their default value is 0 (this will not be explained separately later). Link length. Indicates along direction, arrive The distance between links is a parameter describing the geometric dimensions of each link segment (such as the upper arm segment and forearm segment) on the exoskeleton. Due to individual differences among different populations, it is measured by a magnetic linear encoder. Link offset Indicates along direction arrive The distance between joints reflects the misalignment of each joint in the vertical or normal direction; in this model, it is the motor thickness. Joint angle. In the Cartesian coordinate system, around The rotation angle of the axis is a variable. The initial value is measured by the IMU after the rescuer has put on the equipment. Among them: shoulder joint rotation axis θ1: responsible for characterizing the pitch of the upper arm to achieve the overall downward pressing action; elbow joint rotation axis θ2: used to characterize the linear displacement of the upper arm to complete the compression depth; The length of the human upper arm exoskeleton. This refers to the length of the human forearm exoskeleton, set by default. =0.3m, =0.26m. The system will read the current position of the magnetic encoder during self-test and update it again. d1 and d2 are the linkage offsets, where d1 is used to characterize the axial offset of the shoulder connector in the base direction; d2 is used to characterize the axial offset of the elbow connector in the base direction. The default values ​​are d1=0.05m and d2=0.03m.

[0080] Thus, a kinematic model was established using the improved Denavit-Hartenberg (MDH) notation. Its parameter table (Table 1) defines the coordinate transformation from the base ({F0}, i.e., the torso) to the distal end ({F3}, i.e., the pressure point), as follows: Figure 10 As shown.

[0081] Line 1 (Shoulder Joint): This line defines the transformation from the base coordinate system {F0} (torso) to the coordinate system {F1} (located at the shoulder joint).

[0082] Line 2 (Elbow Joint): This line defines the transformation from coordinate system {F1} (shoulder joint) to coordinate system {F2} (located at the elbow joint).

[0083] Line 3 (End): This line defines the transformation from coordinate system {F2} (elbow joint) to coordinate system {F3} (i.e., "end effector", palm pressure point).

[0084] Based on the MDH parameter table above, the transformation matrix of the exoskeleton is calculated, as shown in the following formula: The homogeneous transformation matrix of each link is represented as follows:

[0085] Based on the above parameter definitions, the overall terminal transformation matrix of the system can be expressed as:

[0086] That is, transformation matrix form one:

[0087] This matrix is ​​used to describe the position and orientation changes of the palm tip in three-dimensional space. The first three columns of the matrix are orientation rotation matrices, and the fourth column is a displacement vector.

[0088] Therefore, the equation for the end position can be expressed as follows:

[0089] When the pressing action is mainly performed in the vertical plane, the x component can be regarded as a displacement constraint in the direction of pressing depth.

[0090] Here we define a standard “end-point pressing task”, which is the complete trajectory sequence of the end-point tool (palm).

[0091] In standard CPR, the rescuer's palm should press vertically downwards (i.e., along the X-axis of the base coordinate system {F0}), and should not create a horizontal rubbing motion on the patient's chest (i.e., Y-axis displacement). Therefore, the initial horizontal position can be calculated using the aforementioned end-position equation and the initial angle measured by the IMU. Throughout the entire cycle, The value remains unchanged.

[0092] To ensure a smooth 6cm press depth, the optimal function that simultaneously satisfies all the aforementioned position and velocity constraints is the cosine function. (Define the press depth.) =0.06m, a complete compression-lift cycle is T=0.5s (this cycle can ensure that the exoskeleton-assisted cardiopulmonary resuscitation compression rate reaches 120 times / minute).

[0093] Therefore, the expression for the palm displacement along the pressing direction can be obtained as follows: ,

[0094] Therefore, the equation for the complete pressing trajectory can be written as:

[0095] This complete pressing trajectory equation will be used in subsequent deep learning strategies.

[0096] In step S3 above, the process of generating the inverse kinematic model using the aforementioned end-position equation is as follows: During chest compressions, the target point for the final compression is known. Then, the joint angles can be solved using a geometric method. Here, the shoulder joint angles obtained after the inverse solution are defined as follows: The elbow joint angle is (Here) , (To obtain the accurate value of the joint angle corresponding to the end position during the motion through inverse kinematics during the assist process) Squaring the equations of the final positions and then adding them together eliminates... ,get:

[0097] According to the Law of Cosines, we have:

[0098] We can obtain:

[0099] correspond At this time, the corresponding elbow is pointing upwards. The value is This represents the natural flexion of the elbow joint, that is, the elbow naturally bends inward (towards the torso), which is a solution that conforms to the biomechanical meaning.

[0100] or:

[0101] correspond At this time, the elbow should be pointing downwards. The value is This means the elbow joint bends in the opposite direction (i.e., away from the torso). Therefore, while this solution is mathematically valid, it is invalid and dangerous in biomechanical and exoskeleton system applications and must be discarded.

[0102] Find the working conditions that are consistent with reality. Then it can be calculated again based on the positive equation. The forward equation can be rewritten as:

[0103] in, ,

[0104] Ultimately, we can obtain:

[0105] Using the above formulas, a complete inverse kinematics model can be constructed. This model can be based on the given end effector target point. The target rotation angles required for the shoulder and elbow joints of the exoskeleton at this location are determined. , This provides the core algorithm for converting the target intent into joint angles in step S4.

[0106] To fully demonstrate the broad applicability of the above-disclosed examples to the exoskeleton-assisted control, this embodiment provides four simulation examples for different initial rescuer postures. The core objective of these examples is to demonstrate that regardless of changes in the rescuer's height, arm length, or compression posture, the control method of this invention can accurately adapt and achieve standard, efficient chest compression assistance. All the examples below are based on the same task objective: the end effector (i.e., the hand) performs a standard vertical chest compression to a depth of 6 cm while maintaining a stable horizontal position, such as... Figure 18 As shown.

[0107] Example 1: Arm-driven pressing posture (Initial posture 1: ) Practical significance: This corresponds to a chest compression operation mode that does not require the rescuer to use their torso weight to stabilize their overall rescue posture. In this mode, the rescuer can complete a standard chest compression operation by using the exoskeleton to provide rhythmic flexion and extension assistance to their forearm, supplemented by slight coordinated movements of the upper arm.

[0108] System performance: This demonstrates that the invention can precisely coordinate multi-joint movements to support this driving mode. Furthermore, this mode reduces the rescuer's reliance on their core trunk muscles, proving the invention's compatibility with diverse operational strategies.

[0109] Example 2: Standard core pose (Initial pose 2: ) Practical significance: It simulates the standard ergonomic chest compression posture and simulates the most efficient range of assistance.

[0110] System performance: Optimal dynamic performance, lowest peak joint angular velocity and angular acceleration, achieving the smoothest and lowest energy consumption assist control.

[0111] Example 3: Close-quarters compact stance (Initial stance 3: ) Practical significance: Simulates scenarios involving close-range pressing by individuals who are petite or in confined spaces.

[0112] System performance: It exhibits extremely high control linearity, with joint angle changes being directly proportional to pressing depth, resulting in uniform and highly predictable power assist output and ensuring stability during close-range operation.

[0113] Example 4: Fatigue Compensation Posture (Initial Posture 4: ) Practical significance: This simulates the compensatory posture of a rescuer who has become fatigued due to prolonged work and whose center of gravity has shifted backward, testing the robustness of the system under boundary conditions.

[0114] System performance: The system was still able to complete the intended task even under conditions close to the kinematic boundaries. This demonstrates that the invention has a wide operating range and tolerance for attitude deviations, ensuring the continuity and effectiveness of assistance under non-standard attitudes.

[0115] After completing the above-mentioned kinematic modeling of the exoskeleton, in order to further achieve precise control of the output torque of each drive motor, the embodiments disclosed in this invention continue to establish a dynamic equation model of the exoskeleton system in step S5. Through this model, the force law (such as gravity, inertial force, etc.) of the exoskeleton when performing tasks is revealed, providing a theoretical basis for calculating the required drive torque for motor control and deep learning models.

[0116] In step S5 above, the upper limb part of the exoskeleton system of the present invention can be simplified as a two-dimensional planar two-bar linkage mechanism, which consists of a shoulder joint and an elbow joint. The linkage parameter definitions are consistent with the aforementioned MDH modeling, and the dynamic parameter definitions are as follows (the kinematic parameter values ​​introduced here are the current data acquired in real time during the exoskeleton assistance process), as shown in Table 2: Table 2 Definitions of Dynamic Parameters

[0117] in , , , Same as the definition in the above kinematics, These are the corresponding angular velocities and angular accelerations, used to characterize the joint motion state. , The distance between the center of mass of each link is the distance from its proximal joint to the center of mass of the link. The mass of the connecting rod (including structural components and actuators). The moment of inertia is the moment of inertia of each link about its own center of mass. This refers to the pressure applied at the end of the arm. Let gravitational acceleration be taken here. (The direction of gravity is the same as the x-axis of the base coordinate system). To drive the joint torque.

[0118] In cardiopulmonary resuscitation, the motor must overcome an external pressure, which is the reaction force when pressing on the patient's chest. Therefore, the total torque that the motor needs to output = internal torque + external torque.

[0119] Combining the Lagrange dynamics modeling method, the following dynamic equations for the lower limb exoskeleton can be obtained:

[0120] in This represents the external force applied to the end effector (hand). Vertical component. It is a dynamic variable that increases linearly with compression depth (derived from the thoracic elasticity model described later). (where k is the thoracic elasticity coefficient), its peak value at the maximum compression depth is set to 450N, and the Jacobian matrix transpose is used. ,Will This is mapped to the equivalent driving torque required by the shoulder and elbow joint motors. .

[0121] In the formula, The driving torque matrix of the exoskeleton. ; For joint angle vectors, ; The joint angular velocity vector; Joint angular acceleration vector; It is a real symmetric inertial matrix; This is the matrix of Coriolis force and centrifugal force related terms; Here is the gravity correlation term matrix; Jacobian matrix. It is achieved by solving the terminal position equations in the inverse kinematics stage. Regarding its The partial derivative is obtained; Inertia matrix The expression is:

[0122] in

[0123]

[0124]

[0125] Gravity matrix The expression is:

[0126] Where (gravity g is along the X-axis of the base coordinate system)

[0127]

[0128] Correlation matrix of Correlation terms for Correlation force and centrifugal force The expression is:

[0129] First, define the intermediate terms.

[0130] , , ,

[0131] End pressure vector The expression is:

[0132] Jacobian matrix Defined as:

[0133]

[0134]

[0135]

[0136] Finally, we can deduce that: Shoulder torque ( ):

[0137] Elbow torque ( ):

[0138] Because CPR compressions are a high-frequency activity, relying on inverse kinematics models for real-time calculations would result in significant control delays, causing the motor response to lag behind the standard compression cycle. To address this delay, the deep learning module of this invention is designed as a fast matching model for motor drive parameters within a standard compression cycle. Based on the different rescuer postures and initial poses of the rescue actions obtained from sensors, it instantly matches and provides a complete set of optimal parameter sequences that cover the entire kinematic and dynamic inverse kinematic solution for a standard compression cycle, while simultaneously using a hybrid control strategy combining feedforward and feedback to obtain the corresponding motor drive parameters.

[0139] In the above embodiments, the deep learning module employs a feedforward neural network (i.e., MLP), such as... Figure 19 As shown, this network is a classic and efficient model for implementing nonlinear function mapping. The network is built with fully connected layers, which ensures a simplified model structure and provides extremely high inference speed and low latency on edge computing devices such as Jetson Nano. The hidden layers are modified with the ReLU activation function to accelerate network convergence and improve computational efficiency. The output layer uses a linear activation function to ensure that the model can directly fit the specific regression values ​​of the output motor drive parameters (such as torque, angle, etc.).

[0140] The number of hidden layers and nodes directly affects the network's prediction accuracy. Therefore, the optimal range for the number of hidden layer nodes L is typically determined using the following empirical formula: (Where: m and n are the number of nodes in the input and output layers, respectively; constants) =1~3. Where the number of input layer nodes is... The number of output layer nodes is (Corresponding to the sampling points of the pressing trajectory). Substituting into the calculation yields... To meet the system's requirements for nonlinear fitting accuracy, and to balance real-time performance and low latency on the Jetson Nano platform, the empirical coefficients were adjusted. Cross-validation and testing were performed within the range of 1.5 to 2.5, and the results showed that when This approach achieves a good balance between accuracy and speed. The final number of hidden layers was determined to be two, with the first hidden layer containing 128 nodes. The number of nodes in the second hidden layer... A layer-by-layer compression strategy is adopted. Experiments have shown that when the number of nodes in the second hidden layer is 64, the network can significantly reduce the number of parameters and latency while maintaining stable error convergence.

[0141] In this embodiment, the input parameters of the deep learning module include static feature parameters and dynamic attitude parameters. The static feature parameters are the lengths of the upper arm exoskeleton and the forearm exoskeleton, which are collected by a magnetic linear encoder installed in the telescopic drive rod. The dynamic attitude parameters are quaternion attitude data collected by IMU attitude sensors located at three locations: the top and side of the universal joint, and the connection between the forearm body 2 and the upper arm body 1.

[0142] Specifically, the input parameters are divided into two main categories: static features and dynamic posture, totaling 14 independent input parameters. Among them, there are two static feature parameters, the purpose of which is to describe the inherent arm length characteristics of the rescuer. These are collected once during system initialization (wearable device) and used as data for individualized model adaptation. Definition: The length of the human upper arm exoskeleton. The length of the human forearm exoskeleton (consistent with the definition of intrinsic parameters in the kinematic model).

[0143] Twelve dynamic attitude parameters are used to describe the rescuer's instantaneous initial posture as they prepare to begin a new set of chest compressions. These parameters are collected in real-time as the rescuer changes posture, serving as the core basis for the model's "instant matching." Selection criteria: Quaternions are used to describe the current posture of each body segment. Compared to Euler angles, quaternions provide stable and continuous input data, fundamentally avoiding gimbal lock-up and angle jump problems, making them the most robust attitude input for MLP models. Each IMU (e.g., MPU9250) will output a set of values ​​for four components. This set of dynamic parameters consists of 12 items, uniformly named: Torso posture acquired by IMU-1 (fixed to the torso). Upper arm posture data acquired by IMU-2 (fixed to the upper arm) and forearm posture data acquired by IMU-3 (fixed to the forearm). (using shoulder posture parameters) For example, it is a collection of [ , , , (A set of four components).

[0144] At any moment when "instant matching" is needed, the complete vector input into the MLP model ( () is a one-dimensional vector composed of the above 14 parameters: =[ , , , , , , , , , , , , , ].

[0145] In this embodiment, the output parameters of the deep learning module are one-dimensional vectors, which correspond to the target angle of the shoulder joint, the target angle of the elbow joint, the target torque of the shoulder joint, and the target torque of the elbow joint in each of the multiple time steps discretized within a standard pressing cycle.

[0146] Specifically, the output of the MLP model is a one-dimensional vector that logically represents the sequence of parameters from all kinematic and dynamic inverse solutions for a standardized complete pressing cycle, comprising 200 parameters (i.e., one "press-release" process). This set of output parameters is defined as... =[ , , , (Parameters for time step t=1)... , , , (Parameters for time step t=50).

[0147] This collection is defined by the following two aspects: Time dimension: Determine the standard press frequency ( According to CPR standards, the compression rate should be 100-120 compressions per minute. To balance rescue efficiency with the need for smooth movement of the "machine-guided rescuer," the following was selected: =120 times / minute is used as the "optimal pressing frequency" in this embodiment. Based on this frequency, the exact time required to perform one complete "press-rebound" action (i.e., one standard cycle) can be calculated. Second.

[0148] Determine the time step ( In the MLP model, it is defined as one of the above... Within a standard period of seconds, the total number of time steps for discretization (sampling) is represented by the symbol... Let t = 1, 2, ..., k. To ensure the smoothness of the motion trajectory and the control accuracy of the Jetson Nano (achieving a control rate of 100Hz), k = 50 is set. That is, a complete pressing-lifting cycle is discretized into 50 sampling points. After deep learning matching by the high-level controller (Jetson Nano), the controller (STM32) will sequentially read the parameters of these 50 steps and use a hybrid control strategy to accurately control the motor to perform the pressing assist action.

[0149] Parameter Dimensions: To achieve "machine-guided human" behavior, two parameters must be provided for each joint at each time step t: joint angle and torque. These parameters can be obtained before model training through the inverse kinematic and dynamic model derivation described above. Shoulder joint angle ( ):

[0150] elbow joint angle ( ):

[0151] Here As can be seen from the definition of the complete trajectory equation of the press in the aforementioned kinematics:

[0152]

[0153] in, The expression for the vertical displacement as a function of time. The vertical pressing depth is set to 6cm in this system.

[0154] Shoulder torque ( ):

[0155] Elbow torque ( ):

[0156] In summary, the output of an MLP model is a one-dimensional vector. This vector contains the information in... Within seconds (i.e., k=50 time steps), determine all the inverse kinematic and dynamic parameters required to drive the shoulder and elbow joints.

[0157] The total dimension of the output vector = k (time step) × j (number of joints) × p (parameters) = 50 × 2 × 2 = 200 parameters.

[0158] In the above embodiments, the compression depth compensation module also includes an online chest elasticity model identification unit, which is used to establish the current patient's chest elasticity benchmark based on the real-time pressure value and compression depth during the first three compression cycles of cardiopulmonary resuscitation, so as to serve as a reference for judging the rescuer's trunk displacement during subsequent compressions.

[0159] In this embodiment, during subsequent compressions, the compression depth compensation module uses Hooke's Law to deduce the end-compression depth error caused by the rescuer's trunk displacement based on the deviation between the real-time pressure value collected by the palm pressure monitoring module and the chest elastic reference. It then corrects the end-compression target point based on the depth error. After correcting the end-compression target point, the deep learning module is triggered to regenerate the target angle sequence and target torque sequence to achieve adaptive compensation for the rescuer's trunk displacement.

[0160] In the above embodiments, the motor drive control unit adopts a hybrid control strategy that combines feedforward and feedback. The target torque sequence is used as the feedforward command, and a compensation torque is generated by the PID controller based on the error between the target angle sequence and the actual angle fed back in real time. The feedforward command and the compensation torque are superimposed to generate the final motor drive signal, so that the joint motors at the shoulder and elbow output the downward pressure torque during the downward phase and the rebound torque during the lifting phase.

[0161] In the above embodiments, the deep learning module is obtained by collecting the initial state vectors of multiple groups of experimental personnel of different body types under various typical initial postures as input, and using the standard compression cycle driving vector generated by the kinematics module and the compression depth compensation module as the output label, and performing offline training.

[0162] Specific examples Figure 20As shown, during the training process of the deep learning module, training samples are generated by organizing N experimental personnel of different body types to wear this exoskeleton device. The system acquires the static parameters of each experimental personnel through a magnetic linear encoder, and collects their different body posture data and initial dynamic parameters under different preparatory postures (e.g., changing posture due to fatigue, repositioning after artificial respiration) through an IMU posture sensor.

[0163] To construct a training dataset that covers real-world working conditions, a multi-dimensional sampling strategy is adopted: I. Body type sampling Group): First, organize Experimenters of different body types wore this exoskeleton device. The system acquired the static parameters of each experimenter through a magnetic linear encoder. , This step is used to cover the variable of "different body types".

[0164] II. Typical pose sampling (Group), requiring each experimenter to simulate in turn. Group( =3) Typical initial preparation positions during CPR. These positions include: Position 1 (Standard): The standard initial position before compressions. Position 2 (Fatigue): Simulating postural changes due to fatigue after prolonged compressions (e.g., the torso leans further forward, the shoulders drop). Position 3 (Repositioning): Simulating the repositioning position after artificial respiration.

[0165] III. Attitude perturbation sampling Group 1): In each of the typical postures described above (e.g., in the "standard posture"), the experimenter is required to maintain that posture for several seconds. During this period, the system continuously acquires data at a high frequency (e.g., 10 Hz). Group dynamic attitude parameters ( , , This step is used to collect “minor differences” in attitude, namely natural swaying and fine-tuning around a typical attitude, to ensure the robustness of the data cluster.

[0166] Using this method, a total of N= The "initial state vector" obtained from the group of actual measurements (14-dimensional). Subsequently, a standard pressing task is defined (i.e., a standard end-point trajectory of k=50 steps). For N sets of actual measurements For each set of data, the inverse kinematics (IK) model is used to calculate the corresponding k-step target angle change sequence. (Right now , Next, a comprehensive analysis... Based on its derivatives and dynamic parameters, the corresponding k-step target torque variation sequence is calculated using the inverse dynamics model (ID). (Right now , ).Will and A "standard driving vector" of size 200 is formed by combining time series data. .

[0167] Will (14-dimensional) as input features for neural networks (200 dimensions) are used as output labels. The neural network (MLP model) trained from this is thus generated as a... and The implicit physical mapping between them (i.e., approximate functions of inverse kinematics and dynamics) can be determined based on the initial state actually collected by the rescuer. It can instantly predict the complete parameter sequence of the inverse kinematics and dynamics of the exoskeleton under a corresponding standard compression cycle. The controller (STM32) reads the angle and torque commands of these 50 steps in sequence and generates corresponding control signals through a hybrid control strategy to precisely drive the motor.

[0168] like Figure 21 As shown, this demonstrates online matching and generalization of a deep learning model (handling new rescuers). The model's core advantage lies in its generalization ability, rather than simple "data lookup." When a new rescuer (i.e., whose body shape and initial posture have changed, i.e., the collected data)... When using this device (if the vector does not belong to the original N training sets), the system operates according to the following standard procedure: 1. Data Acquisition: The system acquires the new rescuer's 14-dimensional "initial state vector" through a hardware module. .

[0169] 2. Matching criteria: When training an MLP model offline, it learns... and The underlying physical mapping between them (i.e., approximate functions of inverse kinematics and dynamics), rather than memorizing N sets of specific data.

[0170] 3. Model reasoning: When this new When the vector is fed into the MLP, the model uses its learned physical laws to perform interpolation generalization on the new input, thereby instantly predicting a "standard driving vector" that conforms to the current "standard press task". (200 dimensions).

[0171] 4. The vector is then loaded into the controller's memory for subsequent motor control execution.

[0172] To facilitate future improvements, the system will include these new rescuers. Data (i.e., inputs not seen by the model) is logged and sent back to the offline development environment. In the offline environment, these newly collected data... The inputs (i.e., those not seen by the MLP) are then re-introduced into the high-precision inverse kinematics (IK) and inverse dynamics (ID) models. In this way, the corresponding, accurate "standard driving vectors" can be derived for these new inputs. . These newly generated [ , Once the data is added back to the original N training sets, the MLP model can be retrained offline. By continuously expanding the training set, the model's generalization ability will be continuously enhanced, making it more accurate in matching new body types and poses encountered in the future.

[0173] Specifically, motor control execution: based on the output of the deep learning model (i.e. This method can obtain all the inverse kinematic and dynamic parameters of a complete downward and upward motion during the rescue process, avoiding the problem of motor control lag caused by complex kinematic calculations.

[0174] During execution, the controller (STM32) employs a hybrid control strategy combining feedforward and feedback. Specifically, at each step t in k=50 time steps, Target torque in vector (Right now , ) is used as a feedforward command to actively compensate for system inertia, gravity, and elbow drag; simultaneously, the target angle (Right now , This is used as the setpoint for the feedback position loop. Its ultimate goal is to calculate an accurate final output torque. :

[0175] For feedback instructions The definition is as follows: This is a high-precision compensation command, calculated in real time by a PID feedback position loop (i.e., a PID controller). The complete implementation of this closed-loop circuit is as follows: A. Define error variables :

[0176] in (Set point): For MLP output In the vector, the target angle at the current k steps (i.e. , ). (Actual position): The actual angle of the motor measured in real time by the controller (STM32) through the joint motor encoder.

[0177] B. Calculate the incremental compensation torque :

[0178] in, and These are the errors from the previous moment and the two moments before, respectively.

[0179] The proportional term P is This term calculates the rate of change of the error (first-order difference). It works by quickly responding to the changing trend of the error, providing the main instantaneous compensation torque. The integral term I is... This term calculates the cumulative current error. It is achieved by continuously applying a torque to eliminate small steady-state deviations (e.g., caused by friction or incompletely compensated gravity). The differential term D is... This calculation measures the rate of change of the error (second-order difference). It works by predicting the "acceleration" of the error and suppressing overshoot and oscillation during high-frequency pressing of the arm (T=0.5s).

[0180] C. Calculate the total compensation torque This refers to the final output "total compensation torque" of the incremental PID controller. It is the sum of the torque at the previous moment and the increment at the current moment.

[0181]

[0182] At time k, the controller (STM32) will combine the above two components ( and The components are superimposed to obtain the final high-level torque command sent to the lower-level driver:

[0183]

[0184] This concludes the control strategy for motor torque compensation. When the shoulder data is input (i.e., input...), , When the data from the shoulder motor is input (i.e., inputting the elbow motor data), the required compensation torque is obtained. , The torque obtained at that time is the compensation torque required by the elbow motor.

[0185] Finally, the STM32's low-level control (FOC algorithm) will... The torque command is converted into a target current I, and then converted into a low-level PWM electrical signal through a high-speed PID current loop and SVM modulation to drive the motor to actually generate the torque.

[0186] Specifically, the adaptive compensation strategy based on depth compensation and state reconstruction is as follows: In actual rescue operations, to address the issue of rescuer fatigue leading to possible forward or backward tilting of the torso, where the original motor-driven data cannot achieve the predetermined compression depth due to the rescuer's forward or backward tilting or shifting center of gravity during assisted compressions. This system further introduces an adaptive compensation mechanism based on neural network reprogramming. This mechanism utilizes the coupling characteristics of the aforementioned inverse kinematics model and deep learning model to achieve real-time correction of the compression trajectory. The specific steps are as follows: (1) Error Perception: A depth error calculation method based on Hooke's Law is proposed. First, a thoracic elastic baseline model of the patient needs to be established at the initial stage of rescue. Considering the individual differences in the thoracic cavity of different patients, and to prevent sternal damage due to excessive compression, an online stiffness identification process including a safety threshold determination is set up. This process is executed during the first three compression cycles of cardiopulmonary resuscitation, assuming that the rescuer is in an initial stable state (no fatigue or trunk displacement). The specific implementation steps are as follows: Safety threshold and initial parameter settings: The system presets a safety threshold for the pressure applied. This value is set based on the typical force required to compress the chest cavity 6cm in a normal healthy adult, and it is also the ultimate protective force to prevent sternal collapse or fracture. At the start of the rescue, the system executes a standard initialization procedure: first, the IMU is used to determine the rescuer's initial pose; then, the state parameters are input into the deep learning module (MLP), which matches and outputs a standard motor motion control sequence.

[0187] Elasticity coefficient calculation: During the first three pressing actions, the system reads the encoder values ​​of the joint motor in real time, and combines them with the aforementioned forward kinematics model to calculate the actual vertical displacement of the palm end in real time. The system is set to ensure that the target stroke for a single press must reach the standard depth. (This is the effective limit of chest compression depth in humans). During compression, the system simultaneously monitors the pressure sensor readings on the palm. .

[0188] The logic for handling branches with different degrees of thoracic cavity elasticity: Based on the pressure feedback during the pressing process, the system identifies parameters under two different operating conditions: Operating Condition A: Conventional elastic thoracic cavity (F<450N). If, at the instant the compression displacement x(t) reaches 6cm, the sensor reads the real-time pressure value... A pressure less than the safe threshold of 450N indicates that the patient's chest cavity elasticity is within the normal range. At this point, the system records the pressure value F1 at a compression depth of 6cm, and records the patient's chest cavity elasticity coefficient as F1. According to Hooke's Law Calculate the patient's thoracic elasticity.

[0189]

[0190] Condition B: Hard, elastic chest cavity (F>450N). If the sensor reading reaches or exceeds 450N before the compression displacement reaches 6cm, it indicates that the patient's chest cavity is hard. In this case: Immediate fuse protection: To prevent chest cavity damage, the system immediately stops the compression command and performs a rebound action. Progressive torque compensation: In the subsequent second and third compression strokes, the system calls the preset "torque compensation curve" (obtained from simulation experiments), gradually increasing the output torque of the joint motor to improve end-effector drive capability within a safe range, bringing the compression depth close to or reaching 6cm. Parameter correction: Record the actual pressure value when the maximum feasible depth (6cm or close to 6cm) is finally reached. and corresponding displacement Calculate the elastic modulus under high impedance conditions. :

[0191] Finally, the system calculates the elasticity coefficient for the first three pressing cycles. The average value was calculated and confirmed as the final thoracic elasticity benchmark for this patient. This benchmark value This will serve as the core basis for monitoring the rescuer's torso displacement and calculating depth compensation during subsequent compression cycles.

[0192] Specifically, depth error detection and compensation are as follows: As the system continues to assist chest compressions (i.e., after the initial three cycles), based on the previously calculated elasticity model corresponding to the patient and according to CPR guidelines, a compression depth of 6cm is generally considered effective. Here, for this patient, the maximum tolerance force at a chest compression depth of 6cm is recorded as follows: This can be calculated using Hooke's Law:

[0193] During each downward press of the assisted hand, the system obtains real-time pressure values ​​through the palm pressure monitoring module. Utilizing the properties of Hooke's Law, if at this point... and If a deviation occurs, it can be determined that the rescuer's torso posture has changed. The system will automatically initiate a compensation program, calculating the end-effector depth error caused by torso displacement based on the force deviation. .

[0194] (when When the value is >0, the torso leans forward. <0, i.e., the torso leans back) Vertical trunk displacement "and "changes in end-press depth" "They are completely equivalent, and through geometric derivation, it can be seen that the horizontal movement of the torso has a negligible impact on the pressure depth of the palm (calculated with a total arm length of 56cm, a horizontal sway of 6cm has only a 0.25cm impact on the distal end depth). Therefore, the change in the horizontal direction does not need to be considered in the distal end position; only the change in the vertical direction needs to be introduced. Then, the expression for the actual position of the distal end pressure target point after the torso change can be derived." (Here, x0 and y0 represent the final positions of the arm extension during the assist process.) Perform attitude state reconstruction: reconstruct the corrected end effector pressure target point. The inverse kinematics algorithm established in stage S3 is then substituted into the terminal position equation. Let... , To determine the new initial joint angle required for the palm to just touch the patient's chest wall at the revised pressure point, we can use the law of cosines:

[0195] By the Pythagorean theorem, we can obtain:

[0196] We can solve for:

[0197]

[0198]

[0199] That is, after the motor completes the pressing stroke with an error, it rotates to the latest position (i.e., , At this point, the IMU rereads the torso's current state data and generates new neural network input parameters. .

[0200] Specifically, motor drive sequence regeneration: The reconstructed sequence... The vector is then fed back into the second control unit (Jetson Nano). Based on its generalization capabilities, the MLP model performs real-time inference for this "new initial physical configuration," outputting a matching full-cycle drive vector. This process is completed during the interval of pressing and rebounding. Updated The vector not only contains the corrected pressing position trajectory, but its feedforward torque sequence is also dynamically optimized for the new arm length posture, thereby ensuring that the system can still output compliant assistance in accordance with biomechanics while compensating for trunk displacement.

[0201] In summary, the hybrid control strategy employed in this invention is executed completely once. The parameter period is 0.5 seconds. To achieve the optimal compression frequency of 120 compressions per minute as specified in the standard compression protocol, the aforementioned hybrid control flow containing 50 steps only needs to be repeated 120 times in the controller (STM32). This process has virtually no delay. When the system detects a deviation in force, it autonomously corrects the rescuer's arm posture and re-matches the latest motor drive sequence. When the accumulated time reaches the artificial respiration stage, the motor will automatically stop running. After artificial respiration, the posture is read again, and the motor is re-matched with the drive sequence. This achieves high-frequency, accurate, and "load-free" guided motor control in CPR.

[0202] Operational example: providing bidirectional assistance via a drive motor (assistance effect as follows) Figure 12 , Figure 13 As shown in the diagram: During the compression phase, the motor outputs assist torque to ensure that the chest compression depth meets medical requirements; during the release phase, the motor provides rebound assistance to ensure that the compression and release rhythms conform to standards. The motor drive strategy is jointly determined by the real-time status and the output strategy of the deep learning module, constructing a closed-loop control.

[0203] This invention utilizes two sets of joint drive motors located at key joints of the upper limb exoskeleton, acting on the shoulder and elbow joints respectively, to achieve precise assist control during the two phases of pressing: upward force accumulation and downward force release. Lifting Assist Phase: In the lifting phase, in preparation for the next press, the control system drives two sets of motors at the shoulder and elbow joints to work together to perform the following movements: the shoulder joint drives the upper arm to abduct (i.e., the upper arm is raised away from the body in the coronal plane); at the same time, the elbow joint drives the forearm to adduct (i.e., the forearm bends towards the chest). The combination of the two actions creates a bent-arm, power-gathering state in the upper limb, which not only increases the structural tension before pressing down but also effectively reduces the muscle load on the rescuer, providing mechanical potential energy reserves for the next compression action.

[0204] Assisted Compression Phase: During the actual chest compression, the control system sends commands to two sets of motors to coordinate the compression and release process: the shoulder joint drives the upper arm to adduct (i.e., the upper arm moves towards the center of the body); simultaneously, the elbow joint drives the forearm to abduct (i.e., the forearm extends outward and downward). This phase of the movement completes the overall process of the arm moving from flexion to extension, that is, from a folded state to a straight arm state, constituting a complete and powerful downward vertical compression. Because the system has completed energy accumulation during the lifting phase, this release action can achieve standard depth compression of the patient's sternum, while also possessing good frequency control capabilities, significantly improving the consistency of compression quality and operational efficiency.

[0205] Through the above specific implementation methods, the present invention achieves high adaptability, rapid wearability, and flexible rescue operation support without adding extra complex structures. It is particularly suitable for single-person or group rescue missions in emergency medical rescue scenarios such as CPR, and has broad application prospects and promotional value.

[0206] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A single-arm active assistive exoskeleton system for cardiopulmonary resuscitation, characterized in that, It adopts a single-arm assist structure, while the other arm remains idle for auxiliary positioning or system fine-tuning; This single-arm active assistive exoskeleton system for cardiopulmonary resuscitation includes: The wearable exoskeleton structure includes a back exoskeleton and an upper limb exoskeleton; the back exoskeleton is connected to the upper limb exoskeleton, and the upper limb exoskeleton includes a shoulder exoskeleton structure, an arm exoskeleton structure, a wrist exoskeleton structure and a palm exoskeleton structure connected in sequence. The motor drive mechanism includes joint motors located at the shoulder and elbow for providing active assistance for the pressing action of the upper limb; The sensing unit is installed on the joints of the back exoskeleton, the upper limb exoskeleton, and the palm exoskeleton structure to acquire real-time information on the movement of the rescuer's upper limbs and the pressure of the palm. The control module, connected to the motor drive mechanism and sensing unit, adjusts the motor output in real time based on motion and compression information to assist rescuers in maintaining standard CPR compression procedures.

2. The single-arm active assistive exoskeleton system for cardiopulmonary resuscitation as described in claim 1, characterized in that, The back exoskeleton includes a back plate (4), a rotatable locking structure (7), a hardware installation cavity, and an upper limb structure support arm (5). The back plate (4) is embedded with a Y-shaped inner skeleton support strip, and the outer side of the back plate (4) is provided with a multi-layer buffer pad; the inner center of the back plate (4) is provided with a hardware device installation cavity for modular installation of control module and power module; A rotatable locking structure (7) is provided on the outside of the back plate (4); One end of the upper limb structure support arm (5) is connected to the back plate (4) through the back rotatable locking structure (7) so that the upper limb structure support arm (5) forms a rotatable support arm; the other end of the upper limb structure support arm (5) is connected to the upper limb exoskeleton; The rotatable support arm is equipped with an angle-adjustable limiter, which is used to adjust the fixed angle of the support arm according to the rescuer's left and right hand habits; The shoulder strap mounting structure (3) is encircled on the back plate (4). The two ends of the shoulder strap mounting structure (3) are connected to the two sides of the back plate (4) respectively to fix and tighten the shoulder strap. The two ends of the back plate (4) are provided with a ring-shaped chest tightening mechanism (4) to fix and tighten the chest strap.

3. The single-arm active assistive exoskeleton system for cardiopulmonary resuscitation as described in claim 1, characterized in that, The shoulder exoskeleton structure includes a spherical universal joint, which consists of two mutually perpendicular rotation axes (6) to simulate the degree of freedom of movement of the human shoulder joint; a first motor base (12) is provided at the end of the spherical universal joint for mounting the joint motor of the shoulder. The arm exoskeleton structure includes an upper arm body (1) and a forearm body (2); the upper arm body (1) is an arm body that extends from the first motor base (12) of the spherical universal joint to the elbow area. The upper arm body (1) is a first telescopic drive rod (11). An arm body fixing structure (10) is provided on the inner side of the upper arm body (1), and a second motor base (22) is provided at the end. The forearm body (2) is located between the elbow and the wrist joint and is connected to the upper arm body (1) through the second motor base (22) at the end of the upper arm body (1). The forearm body (2) is a second telescopic drive rod (21). The telescopic drive rod has a built-in magnetic linear encoder for detecting changes in the rescuer's arm length for adaptive adjustment. The wrist exoskeleton structure includes a forearm connection structure, a semi-ring locking structure (81), and a wrist telescopic rod (82). The forearm connection structure is formed by connecting the end of the forearm body (2) to the semi-ring locking structure (81). The end of the forearm body (2) is provided with a mounting hole for installing the semi-ring locking structure (81). The ring locking structure (81) is suspended below the forearm connection structure through the mounting hole. The semi-ring locking structure (81) is a semi-circular bearing. The shaft on the semi-circular bearing is inserted into the mounting hole at the end of the forearm body (2). One end of the wrist telescopic rod (82) is connected to the semi-ring locking structure (81), and the other end is connected to the palm exoskeleton structure to adapt to different people wearing it. The palm exoskeleton structure includes a dual universal palm glove structure (83) and a pressure module mounting position (84); the other end of the wrist telescopic rod (82) is connected to the dual universal palm glove structure (83), which is a fingerless glove. A palm support component is provided at the palm of the dual universal palm glove structure (83), which is in the form of a fitted support plate. The surface is covered with a flexible high-friction material to stabilize and support the pressing part; a pressure module mounting position (84) is provided at the bottom of the dual universal palm glove structure (83), and a pressure monitoring module is provided in the pressure module mounting position (84).

4. The single-arm active assistive exoskeleton system for cardiopulmonary resuscitation as described in claim 1, characterized in that, The control module includes: The kinematics module is used to establish an analytical mapping relationship between the shoulder and elbow joint angles based on the end-press target point, and to provide training data for the deep learning model, while also serving as a real-time inverse solution model during compensation. The compression depth compensation module is used to determine the rescuer's trunk displacement based on the real-time pressure value collected by the palm pressure monitoring module and the preset chest elasticity model, calculate the depth error and correct the end compression target point, and trigger the deep learning module to regenerate the angle sequence and torque sequence that drive the shoulder and elbow joint motors. The deep learning module uses data collected by sensors, along with kinematics and dynamics modeling, to calculate high-precision motor angle and torque sequences, and then trains a deep learning model. The deep learning module is equipped with a trained multilayer perceptron (MLP) network, which is used to instantly match and generate a target angle and torque sequence that drives the shoulder and elbow joint motors within a standard compression cycle based on the rescuer's initial state information collected by the sensing unit and the corrected end-press target point. The motor drive control unit generates motor drive signals based on the target angle sequence and the target torque sequence, driving the joint motor to output downward pressure torque during the downward pressing phase and rebound torque during the lifting phase.

5. The single-arm active assistive exoskeleton system for cardiopulmonary resuscitation as described in claim 4, characterized in that, The input parameters of a deep learning module include static feature parameters and dynamic pose parameters; Among them, the static feature parameters are the length of the upper arm exoskeleton and the length of the forearm exoskeleton, which are collected by a magnetic linear encoder set in the telescopic drive rod; The dynamic attitude parameters are quaternion attitude data collected by IMU attitude sensors located at three locations: the top and side of the universal joint, and the connection between the forearm body (2) and the upper arm body (1).

6. The single-arm active assistive exoskeleton system for cardiopulmonary resuscitation as described in claim 4, characterized in that, The output parameters of the deep learning module are a one-dimensional vector, which corresponds to the target angle of the shoulder joint, the target angle of the elbow joint, the target torque of the shoulder joint, and the target torque of the elbow joint in each of the discretized time steps within a standard pressing cycle.

7. The single-arm active assistive exoskeleton system for cardiopulmonary resuscitation as described in claim 4, characterized in that, The compression depth compensation module also includes an online chest elasticity model identification unit, which is used to establish the current patient's chest elasticity benchmark based on real-time pressure value and compression depth during the first three compression cycles of cardiopulmonary resuscitation, so as to serve as a reference for judging the rescuer's trunk displacement during subsequent compressions.

8. The single-arm active assistive exoskeleton system for cardiopulmonary resuscitation as described in claim 7, characterized in that, During subsequent compressions, the compression depth compensation module uses Hooke's Law to deduce the error in the final compression depth caused by the rescuer's torso displacement based on the deviation between the real-time pressure value collected by the palm pressure monitoring module and the chest elastic reference, and corrects the final compression target point based on the depth error. After correcting the target point of the end-press, the deep learning module is triggered to regenerate the target angle sequence and the target torque sequence to achieve adaptive compensation for the rescuer's torso displacement.

9. The single-arm active assistive exoskeleton system for cardiopulmonary resuscitation as described in claim 4, characterized in that, The motor drive control unit adopts a hybrid control strategy that combines feedforward and feedback. The target torque sequence is used as the feedforward command, and a compensation torque is generated by the PID controller based on the error between the target angle sequence and the actual angle in real time feedback. The feedforward command and the compensation torque are superimposed to generate the final motor drive signal, so that the joint motors at the shoulder and elbow output the downward pressure torque during the downward phase and the rebound torque during the lifting phase.

10. The single-arm active assistive exoskeleton system for cardiopulmonary resuscitation as described in claim 4, characterized in that, The deep learning module is obtained by collecting initial state vectors of multiple groups of experimental personnel of different body types under various typical initial postures as input, and using the standard compression cycle driving vector generated by the kinematics module and the compression depth compensation module as output labels, and then training it offline.

Images