Intelligent closed-loop massage system based on brain-computer interface

Abstract

The invention discloses an intelligent closed-loop massage system based on a brain-computer interface. The intelligent closed-loop massage system comprises a diagnosis module, a physiotherapy module, a real-time optimization module and a curative effect evaluation module, the diagnosis module is used for automatically identifying discomfort of a user and generating an initial physiotherapy scheme; the physiotherapy module is used for executing accurate automatic massage; the real-time optimization module is an intelligent control core of the system and realizes self-adaptive adjustment in a physiotherapy process; and the curative effect evaluation module quantitatively evaluates the curative effect and drives the system to evolve for a long time. According to the invention, a massage system which is really intelligent and personalized and has continuously optimized curative effect is created through four technical supports of BCI neural signal dominated multi-modal fusion, acupoint-level precise targeting and expert manipulation reproduction, online reinforcement learning real-time optimization and long-term closed-loop learning evolution; the curative effect, the comfort level, the safety and the long-term health management value are all substantially improved.

Description

Technical Field

[0001] This invention relates to a massage therapy robot, and more particularly to an intelligent closed-loop massage system based on a brain-computer interface. Background Technology

[0002] With the rapid development of artificial intelligence, robotics, and biosignal sensing technologies, intelligent massage devices have gradually become a research hotspot in the field of health management. In existing technologies, massage robots mostly employ pre-programmed or simple force feedback control modes, capable of performing basic massage actions such as pressing and kneading. Meanwhile, adaptive massage systems based on physiological signals such as electromyography, heart rate, and skin conductance have been preliminarily explored, capable of adjusting massage intensity and rhythm by monitoring changes in the user's muscle tension or heart rate. Brain-computer interface (BCI) technology is also beginning to be applied in the fields of pain management and neurofeedback therapy, directly capturing neural signals in the brain related to pain and relaxation.

[0003] However, existing technologies mostly focus on a single technical path and have not yet achieved a complete closed loop of multimodal perception, real-time dynamic optimization, and long-term personalized learning. Furthermore, in the integration of traditional Chinese medicine meridian theory with precise acupoint therapy, existing technologies lack systematic and intelligent solutions.

[0004] For example, in existing technology 1, the academic paper "Intelligent acupuncture: data-driven revolution of traditional Chinese medicine" (Acupuncture and Herbal Medicine, vol.3, no.4, pp.271-284, 2023) proposes a data-driven intelligent acupuncture system. This system is a typical solution for the integration of traditional Chinese medicine acupuncture and artificial intelligence technology. The core technical content is as follows: 1. Intelligent diagnosis and localization of acupoints: Based on the acupoint sensitization theory, the electrophysiological characteristics of acupoints are detected by an impedance sensor, and the surface images are analyzed by a convolutional neural network (CNN) to achieve automatic acupoint recognition and localization, with the localization error controlled within ±1.2mm.

[0005] 2. Data-driven acupoint prescription generation: Using techniques such as association rule mining (Apriori algorithm) and complex network analysis, acupoint combination patterns are extracted from massive clinical data and ancient literature to generate acupoint combination schemes for specific diseases.

[0006] 3. Acupuncture efficacy evaluation: Functional near-infrared spectroscopy (fNIRS), electroencephalography (EEG) and other neuroimaging techniques are used to monitor the neural activity of the cerebral cortex during treatment. The efficacy is then quantitatively evaluated using the support vector machine (SVM) algorithm, and a stimulus-response correlation model is established.

[0007] 4. Digitalization of techniques: By collecting data on the force, frequency, and trajectory of acupuncture techniques by traditional Chinese medicine practitioners through PVDF thin-film tactile sensors, a mechanical-kinematic mathematical model is established to achieve standardized reproduction of acupuncture techniques.

[0008] The disadvantages of the existing technology 1 are: 1. Lack of BCI real-time neurofeedback closed loop: Although the system uses EEG and fNIRS to monitor neural activity, it is only used for postoperative efficacy assessment and does not feed the neural signals back to the treatment execution module in real time. It cannot dynamically adjust the manipulation parameters during the conditioning process and belongs to the open-loop mode of "diagnosis-treatment-assessment".

[0009] 2. The technology is limited to acupuncture and cannot meet the needs of massage: The system focuses on the digitization of acupuncture movements and does not cover core massage techniques such as kneading, rubbing, and rolling. Furthermore, it has not established a correlation model between the techniques and the depth of meridian stimulation, making it difficult to meet the "Qi sensation" requirement of massage.

[0010] 3. Insufficient personalization and long-term learning capabilities: Acupoint prescription generation relies on historical data mining, without combining the user's real-time physiological state and neural response characteristics for dynamic optimization, and lacks a user model update mechanism across treatment courses, thus failing to achieve personalized conditioning that becomes more accurate with use.

[0011] 4. Low degree of multimodal data fusion: The system does not integrate physical signals such as pressure sensing and human posture, and only relies on electrophysiological and neuroimaging data, which is easily affected by environmental interference and has insufficient robustness in diagnosis and efficacy evaluation.

[0012] For example, in existing technology 2, the academic paper "Multimodal User Comfort Monitoring for Massage Robots" (Mechanical Design and Manufacturing, DOI:10.19356 / j.cnki.1001-3997.20250604.004, 2025) proposes a method for monitoring the comfort of massage robots based on the fusion of contact force and facial expression. The core technical content is as follows: 1. Multimodal data acquisition: The normal and tangential force time-series data during the massage process are acquired through a six-dimensional force / torque sensor, and facial expression image data of the user are acquired through a camera.

[0013] 2. Construction of a single-modal monitoring model: A long short-term memory network (LSTM) is used to process the temporal features of the force signal and output the probability of "normal / uncomfortable" comfort level; a convolutional neural network (CNN) is used to process facial expression image features to complete the comfort level classification.

[0014] 3. Improved DSmT multimodal fusion algorithm: Based on the calculation of basic probability assignment using the model loss function, the temporal information of facial expressions within 5 seconds is introduced, and historical frames with the highest "discomfort" probability in the visual modality are selected and fused with the current force modality data to solve the problem of multimodal signal asynchrony.

[0015] The shortcomings of the existing technology 2 are: 1. Lack of neural-level deep feedback: It relies only on external physiological and behavioral signals of contact force and facial expression, which cannot capture the neural essence of the user's subjective comfort. It is easily interfered with by factors such as facial expression faking and muscle compensation, and the monitoring results have the limitation of "superficiality".

[0016] 2. Lack of integration with TCM meridian and acupoint theory: Comfort monitoring targets generalized muscle areas and does not achieve precise acupoint-level targeting. It cannot address issues such as blocked meridians, and its therapeutic effect is limited to superficial muscle relaxation.

[0017] 3. Lack of real-time dynamic optimization and long-term learning mechanism: This method is only a comfort monitoring module and does not form a closed loop with the massage execution mechanism. It cannot dynamically adjust parameters such as massage techniques, intensity, and trajectory based on monitoring results. At the same time, it lacks long-term user profiles and model evolution capabilities, and cannot achieve personalized improvement across sessions.

[0018] 4. Massage techniques are limited and not standardized: The experiment used a spherical end effector to perform only basic pressing actions, without covering the complex techniques of traditional Chinese massage such as kneading, pushing, and grasping, which cannot meet the needs of in-depth conditioning. Summary of the Invention

[0019] To address the aforementioned technical problems, this invention provides an intelligent closed-loop massage system based on a brain-computer interface, which features a complete closed loop of multimodal perception, real-time dynamic optimization, and long-term personalized learning.

[0020] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: an intelligent closed-loop massage system based on a brain-computer interface, comprising a diagnostic module, a physiotherapy module, a real-time optimization module, and a therapeutic effect evaluation module; the diagnostic module automatically identifies user discomfort and generates an initial physiotherapy plan; the diagnostic module receives three types of multimodal input data, namely, human force maps collected by a high-density pressure sensor array, EEG / EMG / GSR baseline signals collected by a brain-computer interface device, and human posture and surface images collected by a depth camera; the diagnostic module first completes high-precision automatic positioning of acupoints throughout the body through an acupoint positioning module, and then integrates the abnormal pressure data with the BCI... The system deeply integrates identified pain and relaxation nerve signals, ultimately outputting a discomfort heatmap centered on acupoints, a priority list of acupoints, and an initial physiotherapy plan including target acupoint sequences, recommended techniques, and intensity curves. The physiotherapy module executes precise automated massage. It loads the initial physiotherapy plan output by the diagnostic module and controls the massage components to perform massage actions. The module has a built-in "expert technique library" that precisely models the motion-mechanical parameters of various clinically validated massage techniques. Based on the initial plan, the module uses reinforcement learning algorithms to plan the optimal massage trajectory along the meridians, automatically generating a technique sequence matching the acupoint type, and calling the corresponding mathematical model to drive the actuator to reproduce the expert-level technique effect. The real-time optimization module achieves adaptive adjustment during the physiotherapy process. It collects three types of multimodal feedback signals in real time within the cycle, including BCI neural signals, end effector mechanical signals, EMG, and HRV. Physiological signals, after preprocessing, are input into a real-time optimization algorithm. This algorithm uses multimodal features as states, massage intensity, speed, frequency, contact area, acupoint dwell time, technique switching, and meridian direction parameters as actions, and pain signal reduction, comfortable force feedback, and physiological relaxation as reward objectives. It continuously iterates its strategy and outputs dynamic adjustment instructions to the physiotherapy module to achieve real-time optimization of massage parameters. The efficacy evaluation module quantifies the efficacy and drives the system's long-term evolution. After the treatment, the module compares multimodal data before and after the massage, using BCI signal changes as the core indicator to calculate the decrease in pain index, reduction rate of muscle tension, and improvement in physiological relaxation. A weighted total efficacy score and grade are derived. Based on the evaluation results, the module generates a personalized report including data comparison charts and natural language summaries. Simultaneously, it uses a Transformer model to update the user's long-term profile, predicting and generating future personalized physiotherapy plans. The updated model will be loaded into the diagnostic module the next time the user uses the system, achieving long-term closed-loop optimization across sessions.

[0021] Furthermore, the massage component is a multi-degree-of-freedom robotic arm or a massage mechanism.

[0022] Furthermore, the real-time optimization algorithm of this invention is the reinforcement learning algorithm PPO.

[0023] Furthermore, the real-time optimization algorithm of the present invention employs model-based predictive control (MPC) or adaptive fuzzy PID control.

[0024] Furthermore, the acupoint positioning module of the present invention is a depth camera and non-rigid registration; the human pose estimation is acquired by the depth camera and then non-rigidly registered with the standard acupoint database.

[0025] Furthermore, the diagnostic module of the present invention deeply integrates stress abnormality data with pain and relaxation nerve signals identified by BCI through a multimodal fusion neural network that combines CNN and Transformer.

[0026] The beneficial effects of adopting the above technical solution are: 1. This invention achieves precise and personalized treatment at the neural level, significantly improving therapeutic efficacy and comfort. Existing technologies mainly rely on indirect physiological signals such as pressure and electromyography, making it difficult to accurately capture the user's subjective pain and comfort levels. This invention uses brain-computer interface (BCI) neural signals (such as EEG pain signatures, alpha / theta relaxation waves, and bursts of sensation) as the core feedback source, and fuses them with multimodal mechanical and physiological signals to directly and objectively "read" the brain's true response to massage. Compared with existing technologies, this system improves the accuracy of pain perception detection from relying on indirect inference to objective quantification based on neural evidence. This allows the system to accurately identify discomfort target areas and determine in real time whether the massage is "effective" and "comfortable."

[0027] 2. Integrating Traditional Chinese Medicine (TCM) meridian theory with modern robotics technology, this invention achieves precise acupoint-level manipulation and deep conditioning. The system utilizes high-precision human posture estimation and acupoint database registration to automatically identify acupoints and perform massage using acupoints as the smallest operational unit. Combined with mathematical models of 15 expert techniques, the robotic arm can reproduce the effects of expert-level techniques. Existing massage robots mostly perform broad-area operations and cannot connect with TCM theory. This invention is the first to systematically integrate precise acupoint targeting with expert technique models, enabling massage to not only act on muscles but also stimulate meridians, inducing a "qi sensation." Clinically, in addition to muscle relaxation, it significantly improves meridian patency, achieving an upgrade in therapeutic effect from "muscle relaxation" to "meridian conditioning."

[0028] 3. Possesses real-time adaptive optimization capabilities within a single treatment session, achieving "increasing precision with each massage." Through online reinforcement learning (such as PPO) algorithms, this invention's system can dynamically adjust multi-dimensional parameters such as intensity, speed, frequency, technique combinations, and acupoint dwell time during the massage process based on multimodal feedback (centered on BCI). Compared to existing technologies, which rely on simple PID force control or threshold adjustment, this system achieves multi-dimensional, continuous, and intelligent parameter optimization. This allows the system to quickly converge to the user's current optimal state within a single treatment session, avoiding the problems of excessively light, excessively heavy, or ineffective stimulation caused by fixed procedures or simple feedback.

[0029] 4. Establish a closed-loop evolutionary mechanism based on long-term learning to achieve continuous personalized improvement. The efficacy assessment module of this invention uses pre- and post-BCI comparison data as its core, performs multi-dimensional quantitative scoring, and updates the user's long-term profile using a Transformer model. The results of each treatment session are used to optimize the next diagnosis and treatment plan. Existing systems lack effective efficacy quantification and long-term memory functions, making it impossible to achieve continuous user tracking and plan optimization. This invention forms a cross-session long-term closed loop of "assessment-learning-optimization," enabling the system to increasingly align with the user's personalized needs and changing trends as the number of uses increases, thereby achieving an intelligent transformation from "single-time relief" to "long-term cure," improving the sustainability and scientific rigor of health management.

[0030] 5. Improved system automation and safety. This invention's system requires no manual intervention. Multi-layered safety mechanisms (such as BCI-based acute pain signal recognition, force exceeding threshold braking, and priority of user emergency commands) ensure operational safety. While achieving a high degree of automation, direct monitoring of the pain center response via BCI allows for earlier and more sensitive detection and intervention of discomfort compared to systems relying on mechanical or physiological indicators. This significantly reduces the risk of injury due to improper force or technique, making it significantly safer than traditional pre-programmed or simple feedback systems.

[0031] In summary, this invention, through four key technological pillars—BCI neural signal-driven multimodal fusion, acupoint-level precise targeting and expert technique reproduction, online reinforcement learning for real-time optimization, and long-term closed-loop learning evolution—creates a truly intelligent, personalized massage system with continuously optimized therapeutic effects, resulting in substantial improvements in efficacy, comfort, safety, and long-term health management value. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the principle block of the present invention. Detailed Implementation

[0033] The present invention will now be described in further detail with reference to the accompanying drawings.

[0034] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0035] One embodiment of the present invention: as follows Figure 1 As shown, an intelligent closed-loop massage system based on brain-computer interface includes a diagnostic module, a physiotherapy module, a real-time optimization module, and a efficacy evaluation module. The core of these four modules is to achieve a complete closed loop through multimodal data flow and AI algorithm-driven operation.

[0036] The diagnostic module automatically identifies user discomfort and generates an initial physiotherapy plan. It receives three types of multimodal input data: human force maps acquired by a high-density pressure sensor array, EEG / EMG / GSR baseline signals acquired by a brain-computer interface device, and human posture and surface images acquired by a depth camera. The module first performs high-precision automatic localization of acupoints throughout the body using an acupoint positioning module. Then, it deeply integrates abnormal pressure data with neural signals such as pain and relaxation identified by the BCI, ultimately outputting a discomfort heatmap centered on acupoints, a priority list of acupoints, and an initial physiotherapy plan including target acupoint sequences, recommended techniques, and intensity curves.

[0037] Brain-computer interface (BCI) devices transmit EEG / EMG / GSR baseline signals via Bluetooth 5.0 or USB.

[0038] The depth camera transmits human image data streams via the USB 3.0 protocol and supports point cloud data output.

[0039] The specific working process of the diagnostic module is as follows: 1. The user wears a BCI device and lies on the physiotherapy bed. A depth camera collects full-body images and point cloud data, which are then resampled and smoothed. 2. Combining an expert manipulation library and a human stress map, target acupoint sequences, recommended manipulation techniques, and intensity curves are matched. 3. Based on point cloud registration, a non-rigid registration algorithm locates acupoints throughout the body, generating an acupoint coordinate mapping table. A multimodal fusion network integrates historical abnormal stress data and nerve pain signals, outputting a discomfort area determination result centered on acupoints. An initial physiotherapy plan is generated according to the acupoint generation trajectory. 4. Data is synchronously transmitted to the module server via the corresponding protocol, preprocessed with timestamp alignment, and records are generated.

[0040] The physiotherapy module performs precise automated massage; it loads the initial physiotherapy plan output by the diagnostic module and controls the massage components to complete the massage movements; the physiotherapy module has a built-in "expert technique library" that performs precise kinematic-mechanical parameterization modeling of various clinically validated massage techniques; based on the initial plan, the physiotherapy module plans the optimal massage trajectory along the meridians through reinforcement learning algorithms, automatically generates a sequence of techniques matching acupoint types, and calls the corresponding mathematical model to drive the actuator to reproduce the effects of expert-level techniques.

[0041] The sensing component employs a high-density pressure sensor array and transmits human body force point matrix data via SPI (Internet Protocol over Ethernet, TCP) protocol at a sampling rate of 100Hz. It receives the initial scheme from the diagnostic module, including parameters such as target acupoint coordinates, manipulation type, and force range. The interaction with the real-time optimization module is handled automatically by the industrial control computer.

[0042] The specific working process of the physiotherapy module is as follows: 1. Load the initial plan from the diagnostic module, parse the target acupoint list and manipulation parameters. Call the corresponding model from the expert manipulation library to generate the motion trajectory and pressure control curve of the robotic arm. 2. Drive the multi-degree-of-freedom robotic arm through the motion controller to perform massage actions according to the trajectory. 3. Receive adjustment instructions from the optimization module in real time and dynamically correct the parameters of the manipulation.

[0043] The real-time optimization module is the core of the system's intelligent control, enabling adaptive adjustments during the physiotherapy process. Within a cycle, the module continuously collects three types of multimodal feedback signals: BCI neural signals, end effector mechanical signals, and physiological signals such as EMG and HRV. After preprocessing, these signals are input into the real-time optimization algorithm, with a sampling period typically of 500ms. The algorithm uses multimodal features as states, massage intensity, speed, frequency, contact area, acupoint dwell time, technique switching, and meridian direction parameters as actions, and pain signal reduction, comfortable force feedback, and physiological relaxation as reward objectives. It continuously iterates its strategy and outputs dynamic adjustment commands to the physiotherapy module, achieving real-time optimization of massage parameters.

[0044] The multimodal feedback feature extraction includes: extracting pain signature (energy value of a specific frequency band) and relaxation wave (alpha wave proportion) from BCI signals; calculating the mean contact force and resistance change rate from force sensor data; and calculating muscle tension and heart rate variability indices from physiological signals.

[0045] The specific working process of the real-time optimization module is as follows: During the massage, BCI nerve signals, distal mechanical signals, and physiological signals are simultaneously acquired at 500ms intervals. The feature extraction module preprocesses the three types of data, removing noise and extracting key indicators. The real-time optimization algorithm calculates the optimal combination of motion parameters based on the current feature state and historical strategies. The industrial control computer sends adjustment commands to the physiotherapy module to correct the massage parameters in real time. The above steps are repeated until the end of a single treatment session, forming a real-time closed-loop optimization within a single treatment session.

[0046] The efficacy assessment module quantifies and evaluates treatment effectiveness and drives long-term system evolution. After the treatment, the module compares multimodal data before and after the massage, using changes in BCI signal as the core indicator to calculate quantitative indicators such as the decrease in pain index, reduction rate of muscle tension, and improvement in physiological relaxation. A weighted total efficacy score and grade are then derived. Based on the assessment results, the module generates a personalized report including data comparison charts and natural language summaries. Simultaneously, it uses a Transformer model to update the user's long-term profile, predicting and generating future personalized physiotherapy plans. The updated model will be loaded into the diagnostic module the next time the user uses the system, achieving long-term closed-loop optimization across sessions.

[0047] This includes receiving multimodal data before and after treatment: BCI, stress, and physiological data comparison sets before and after massage are obtained from the diagnostic and physiotherapy modules via TCP / IP and HTTP protocols. Long-term user profile updates: Updated model data is synchronized to a cloud database via HTTP protocol, supporting cross-device access.

[0048] The core algorithm of the efficacy assessment module includes efficacy quantification scoring: a weighted scoring model is constructed, using changes in BCI signals (the magnitude of pain index decrease and the rate of relaxation wave increase) as the core indicator, with a weighting of 60%; supplemented by indicators such as the rate of reduction in muscle tension and the rate of reduction in abnormal stress areas, the total efficacy score and grade are calculated using weighted averages. User model updates: a Transformer time-series model is used, inputting the user's efficacy data and multimodal characteristics from previous treatments, learning the user's physiological response patterns, generating personalized parameters such as acupoint sensitivity and manipulation preferences, and updating long-term records.

[0049] The specific workflow of the efficacy assessment module is as follows: 1. After the treatment course, multimodal data before and after the massage are automatically retrieved, and difference calculations and feature comparisons are performed. 2. The weighted scoring model outputs the total efficacy score, grade, and comparison charts of various indicators, generating a natural language report. 3. Based on the efficacy data, the Transformer model updates the user's long-term personalized model. The updated model is synchronized to the cloud via HTTP protocol. The next time it is used, the diagnostic module directly loads the model to generate an initial treatment plan that is more suitable for the user.

[0050] In summary, the four modules of this invention are seamlessly connected, forming a two-tiered closed loop. A single closed loop: diagnosis → physical therapy → (real-time optimization) → efficacy evaluation, completing a full personalized treatment cycle; a long-term closed loop: the efficacy evaluation results are fed back to the user model to optimize the next diagnosis and treatment plan, enabling the system to continuously learn and evolve.

[0051] This invention achieves highly personalized, adaptive, and continuously improving fully intelligent massage through multimodal fusion dominated by BCI neural signals, precise acupoint targeting, mathematical modeling of expert techniques, and online reinforcement learning optimization.

[0052] In other specific embodiments of the present invention, the remainder is the same as the embodiments described above, except that, as follows: Figure 1 As shown, the massage component is a multi-degree-of-freedom robotic arm or a massage mechanism.

[0053] In other specific embodiments of the present invention, the remainder is the same as the embodiments described above, except that, as follows: Figure 1 As shown, the real-time optimization algorithm is the reinforcement learning algorithm PPO. Online reinforcement learning optimization: The PPO (Proximal Policy Optimization) algorithm is used, with multimodal features as the state space, parameters such as massage intensity, speed, and frequency as the action space, and "pain signal decrease + force feedback comfort + physiological relaxation" as the composite reward function. The policy iteration is completed every 500ms, and the parameter adjustment instructions are output.

[0054] In other specific embodiments of the present invention, the remainder is the same as the embodiments described above, except that, as follows: Figure 1 As shown, the real-time optimization algorithm employs either Model-Based Predictive Control (MPC) or Adaptive Fuzzy PID Control. MPC, with its accurate user response model, can perform forward-looking optimization; fuzzy PID has clear rules and a fast response. However, both heavily rely on accurate mathematical models or expert-based rules. Given the complex, nonlinear, and highly individualized characteristics of BCI signals, establishing an accurate, universal model is extremely difficult. Online reinforcement learning, on the other hand, possesses the ability to learn and adapt to unknown models through interaction, better meeting the needs of highly personalized dynamic optimization—a capability that MPC or fuzzy PID cannot match.

[0055] In other specific embodiments of the present invention, the remainder is the same as the embodiments described above, except that, as follows: Figure 1 As shown, the acupoint positioning module is a depth camera and non-rigid registration; the depth camera collects human pose estimation and then non-rigidly registers it with a standard acupoint database.

[0056] In other specific embodiments of the present invention, the remainder is the same as the embodiments described above, except that, as follows: Figure 1 As shown, the diagnostic module deeply fuses stress anomaly data with neural signals such as pain and relaxation identified by BCI through a multimodal fusion neural network that integrates CNN and Transformer. Traditional feature-level fusion methods (such as weighted averaging and feature concatenation) can also replace the end-to-end fusion model of CNN+Transformer. However, while traditional feature fusion methods have low computational cost and low deployment threshold, they cannot effectively model the spatiotemporal correlation features between BCI signals, stress signals, and image signals, leading to a decrease in the accuracy and robustness of discomfort area identification.

[0057] Note that the above description is merely a preferred embodiment of the present invention. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A brain-computer interface-based intelligent closed-loop massage system, characterized in that, It includes a diagnostic module, a physiotherapy module, a real-time optimization module, and a efficacy evaluation module; The diagnostic module automatically identifies user discomfort and generates an initial physiotherapy plan. The diagnostic module receives three types of multimodal input data: human force maps acquired by a high-density pressure sensor array, EEG / EMG / GSR baseline signals acquired by a brain-computer interface device, and human posture and surface images acquired by a depth camera. The diagnostic module first completes high-precision automatic localization of acupoints throughout the body through the acupoint localization module, and then deeply integrates the abnormal pressure data with pain and relaxation nerve signals identified by BCI. Finally, it outputs a discomfort heat map centered on acupoints, a priority ranking list of acupoints, and an initial physiotherapy plan including target acupoint sequences, recommended techniques, and intensity curves. The physiotherapy module performs precise automated massage; it loads the initial physiotherapy plan output by the diagnostic module and controls the massage components to complete the massage movements; the physiotherapy module has a built-in "expert technique library" that performs precise kinematic-mechanical parameterization modeling of various clinically validated massage techniques; based on the initial plan, the physiotherapy module plans the optimal massage trajectory along the meridians through reinforcement learning algorithms, automatically generates a sequence of techniques matching acupoint types, and calls the corresponding mathematical model to drive the actuator to reproduce the effects of expert-level techniques. The real-time optimization module achieves adaptive adjustment during physiotherapy. It collects three types of multimodal feedback signals in real-time within a cycle: BCI neural signals, end effector mechanical signals, and EMG and HRV physiological signals. After preprocessing, these signals are input into the real-time optimization algorithm. The algorithm uses multimodal features as states, massage intensity, speed, frequency, contact area, acupoint dwell time, technique switching, and meridian direction parameters as actions, and pain signal reduction, comfortable force feedback, and physiological relaxation as reward objectives. It continuously iterates its strategy and outputs dynamic adjustment instructions to the physiotherapy module, achieving real-time optimization of massage parameters. The efficacy assessment module quantifies the efficacy and drives the long-term evolution of the system. After the treatment, the efficacy assessment module compares the multimodal data before and after the massage, uses the change in BCI signal as the core indicator, calculates the decrease in pain index, the reduction rate of muscle tension, and the quantitative indicators of the improvement in physiological relaxation, and calculates the total efficacy score and level by weighting. Based on the assessment results, the efficacy assessment module generates a personalized report that includes data comparison charts and natural language summaries. At the same time, it uses the Transformer model to update the user's long-term profile, predict and generate future personalized physiotherapy plans. The updated model will be loaded into the diagnostic module the next time the user uses it, achieving long-term closed-loop optimization across sessions.

2. The intelligent closed-loop massage system based on a brain-computer interface according to claim 1, characterized in that, The massage component is a multi-degree-of-freedom robotic arm or a massage mechanism.

3. A brain-computer interface-based intelligent closed-loop massage system according to claim 1 or 2, characterized in that, The real-time optimization algorithm is the reinforcement learning algorithm PPO.

4. A brain-computer interface-based intelligent closed-loop massage system according to claim 1 or 2, characterized in that, The real-time optimization algorithm employs model-based predictive control (MPC) or adaptive fuzzy PID control.

5. A brain-computer interface-based intelligent closed-loop massage system according to claim 1 or 2, characterized in that, The acupoint positioning module is a depth camera and non-rigid registration; the depth camera collects human pose estimation and then non-rigidly registers it with a standard acupoint database.

6. A brain-computer interface-based intelligent closed-loop massage system according to claim 1 or 2, characterized in that, The diagnostic module uses a multimodal fusion neural network that combines CNN and Transformer to deeply fuse abnormal stress data with pain and relaxation neural signals identified by BCI.

Images