Control method and system of intelligent massage robot

Abstract

The invention discloses a control method and system of an intelligent massage robot, and relates to the field of intelligent robots. The method comprises the following steps: performing multi-modal dynamic perception fusion, and processing sensor data by using an attention mechanism deep learning model to obtain a body state; modeling and judging a muscle state based on the bio-electricity signal; adjusting a massage reference in combination with environmental parameters; a personalized massage strategy is generated through reinforcement learning; and real-time feedback correction ensures the effect and safety. The system comprises a multi-mode sensing module and the like. The method solves the problems that in the prior art, accuracy is insufficient, and the method is difficult to adapt to individual and environment differences, and has creativity and practicability.

Description

Technical Field

[0001] This invention relates to the field of intelligent robots, specifically to a control method and system for an intelligent massage robot. Background Technology

[0002] As people's living standards improve and their health awareness increases, intelligent massage robots are gradually entering homes and offices, providing convenient massage services. However, most existing intelligent massage robot control methods are based on preset programs or simple pressure feedback, which have many shortcomings.

[0003] In existing technologies, massage robots often operate according to fixed patterns, making it difficult to personalize the massage based on individual user differences (such as physical condition and muscle state). This results in low massage precision and an inability to effectively relieve muscle fatigue and tension. Furthermore, current technologies rarely consider the impact of environmental factors on massage effectiveness; the same massage parameters may produce different experiences and results under different environmental conditions. In addition, existing feedback mechanisms are relatively simple and cannot perceive changes in the user's physical state during the massage in real time and comprehensively. This makes it difficult to make timely and accurate adjustments to the massage strategy, potentially leading to problems such as excessive or insufficient massage intensity or inappropriate frequency, affecting the massage effect or even causing harm to the user.

[0004] Therefore, a more innovative intelligent massage robot control method and system is needed to solve the problems existing in the above-mentioned prior art. Summary of the Invention

[0005] The purpose of this invention is to provide a control method and system for an intelligent massage robot to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a control method for an intelligent massage robot, comprising the following steps:

[0007] Step S1: Multimodal dynamic perception fusion. The user's bioelectrical signals are collected in real time through a flexible tactile sensor array, an infrared thermal imaging sensor and wearable devices set on the massage end of the massage robot. The collected data is then fused using a deep learning model based on an attention mechanism to obtain comprehensive information about the user's current physical state.

[0008] Step S2: Muscle state modeling based on bioelectric signals. Feature extraction is performed on the user's bioelectric signals collected in step S1. A mapping relationship model between muscle tension and bioelectric signal features is constructed by combining human anatomical data to judge the user's muscle tension and fatigue level in real time.

[0009] Step S3: Combined with adaptive adjustment of environmental parameters, the temperature, humidity and noise parameters of the current environment are collected by the environmental sensors set on the robot, and the preset massage intensity, frequency and time reference parameters are initially adjusted according to the environmental parameters.

[0010] Step S4: Massage strategy generation based on body state and environmental parameters. Based on the comprehensive body state information obtained in step S1, the muscle state obtained in step S2, and the baseline parameters adjusted in step S3, a reinforcement learning algorithm is used to generate a personalized massage path, intensity, frequency, and time control strategy.

[0011] Step S5: Real-time feedback correction. During the massage, the system continuously collects data on changes in the user's physical state and feedback signals through multimodal sensors, compares them with the expected results, and uses a PID control algorithm to correct the massage control strategy in real time to ensure the massage effect and safety.

[0012] Preferably, the specific implementation logic of using a deep learning model based on an attention mechanism to fuse the collected data in step S1 to obtain comprehensive information about the user's current physical state is as follows:

[0013] Step S11, Data Preprocessing: For the pressure distribution and contact area data collected by the flexible tactile sensor array, due to the noise and drift inherent in the sensor array, a mean filtering method is first used to remove high-frequency noise. Then, a linear interpolation method is used to supplement the missing sampling points, unifying the data format into a 20×20 matrix, where each element represents the pressure value at the corresponding location. For the temperature distribution image collected by the infrared thermal imaging sensor, bad pixel repair is first performed. For abnormally high or low temperature points in the image, the average value of the surrounding pixels is used for replacement. Then, the image is processed... Normalization is performed to map temperature values ​​to a range of 0-1 to reduce the impact of different temperature ranges on model training. Finally, the image size is adjusted to the standard size of 64×64. The user's bioelectrical signals (including electromyography signals, electrocardiogram signals, etc.) collected by the wearable device are first filtered by a bandpass filter to remove power frequency interference (50Hz) and low frequency drift. Then, wavelet transform is used to denoise the signal to remove noise such as motion artifacts. The processed signal is then segmented, with each second as a data segment. The feature values ​​of each data segment are extracted to form a feature vector of length 1000.

[0014] Step S12, Feature Extraction: For the processed flexible tactile sensor array data, CNN is used for feature extraction. Through two convolutional layers and pooling layers, the first convolutional layer uses 32 3×3 convolutional kernels with a stride of 1 to perform preliminary feature extraction, and then a 2×2 pooling layer is used for downsampling; the second convolutional layer uses 64 3×3 convolutional kernels with a stride of 1 to further extract deep features, and then a 2×2 pooling layer is used to process, finally obtaining a 16×16×64 feature matrix; the image data after infrared thermal imaging processing is also extracted using CNN. Using convolution and pooling operations similar to those used in flexible tactile sensor data processing, the first convolutional layer uses 32 3×3 convolutional kernels with a stride of 1, followed by a 2×2 pooling layer; the second convolutional layer uses 64 3×3 convolutional kernels with a stride of 1, followed by a 2×2 pooling layer, resulting in an 8×8×64 feature matrix. The feature vector of the bioelectrical signal is extracted using LSTM, which is input into a two-layer LSTM network. The first LSTM layer has 128 hidden units, and the second LSTM layer has 64 hidden units. The LSTM network captures the temporal features of the bioelectrical signal, and finally outputs a feature vector of length 64.

[0015] Step S13, Attention weight calculation: The flexible tactile sensor feature matrix, infrared thermal imaging feature matrix and bioelectric signal feature vector extracted in step 12 are flattened and converted into one-dimensional vectors with lengths of 16×16×64=16384, 8×8×64=4096 and 64 respectively.

[0016] Introduce a learnable weight matrix W and a bias vector b, and perform linear transformations on the three one-dimensional vectors respectively to obtain three vectors with the same dimensions, denoted as V1, V2, and V3, each with a dimension of 256.

[0017] Calculate the attention score for each vector using the softmax function, with the formula: Attention Score(V) i ) = exp(Vi·u) / (exp(V1·u)+exp(V2·u)+exp(V3·u)), where u is a learnable query vector; the weight of each data feature is determined based on the attention score;

[0018] Step S14, Feature Fusion: Multiply the three feature vectors extracted in Step 12 by their corresponding attention weights to obtain weighted feature vectors W1×F1, W2×F2, and W3×F3, where W1, W2, and W3 are the attention weights for the flexible tactile sensor feature, infrared thermal imaging feature, and bioelectric signal feature, respectively, and F1, F2, and F3 are the corresponding feature vectors. Concatenate the three weighted feature vectors to obtain a fused feature vector with a length of 16384 + 4096 + 64 = 20544.

[0019] Step S15, Output of comprehensive body state information: The fused feature vector is input into a fully connected neural network. This network contains two hidden layers. The first hidden layer has 1024 neurons and uses the ReLU activation function; the second hidden layer has 256 neurons and also uses the ReLU activation function.

[0020] The output layer of a fully connected network contains multiple neurons, each corresponding to a different bodily state indicator, such as stress tolerance, muscle activity intensity, and blood circulation. The output value of each neuron is mapped to a range of 0-1 using a sigmoid function, representing the degree of the corresponding bodily state indicator.

[0021] The output values ​​of each neuron in the output layer are then integrated to form comprehensive information on the user's current physical state. This information fully reflects the user's physical condition and provides a basis for generating subsequent massage strategies.

[0022] Step S16, Model Training and Optimization: Train and optimize the model from step S15;

[0023] Specifically, in order for the model to accurately perform data fusion and assess body condition, it needs to be trained and optimized.

[0024] A large amount of labeled data is prepared, including data collected by three types of sensors and corresponding comprehensive information on the user's physical condition. Professional massage therapists then label the data according to the user's actual situation.

[0025] The labeled data is divided into training set, validation set and test set in a ratio of 7:2:1.

[0026] The mean squared error was used as the loss function, and the Adam optimizer was used to train the model. The initial learning rate was set to 0.001, and the learning rate was gradually reduced as the number of training rounds increased.

[0027] During training, the model's performance is evaluated using a validation set. Training is stopped when the loss on the validation set no longer decreases to prevent overfitting.

[0028] The trained model is tested using a test set to evaluate metrics such as accuracy and recall. If the metrics are not up to standard, the model parameters are adjusted, such as the weight matrix in the attention mechanism and the number of neurons in the fully connected network, and the model is retrained until its performance meets the requirements.

[0029] Preferably, the specific implementation steps of step S2 are as follows:

[0030] Step S21, Bioelectric signal feature extraction: Extract feature parameters that can effectively reflect muscle state from the preprocessed bioelectric signal in step S1, including time domain features, frequency domain features and time-frequency domain features;

[0031] Specifically, time-domain feature extraction involves calculating the peak value (the maximum value in the signal), mean (the arithmetic mean of all signal values ​​within the data segment), variance (reflecting the dispersion of the signal), and root mean square (RMS) value (quantification of signal energy) for each data segment. For example, a larger RMS value for an electromyographic signal indicates higher muscle activity intensity.

[0032] Frequency domain feature extraction: The time-domain signal is converted into a frequency-domain signal using Fast Fourier Transform (FFT) to obtain the power spectral density of the signal. The center frequency (the frequency corresponding to the centroid of the power spectrum) and the median frequency (the frequency at which the cumulative power in the power spectrum reaches 50%) are extracted. When muscles are fatigued, the median frequency of the electromyographic signal shifts towards lower frequencies, which is an important indicator for judging muscle fatigue.

[0033] Time-frequency domain feature extraction: Wavelet transform is used to perform time-frequency analysis on the signal. The db4 wavelet is selected as the basis function, and the signal is decomposed into five levels to obtain wavelet coefficients for different frequency bands. The energy values ​​of the wavelet coefficients in each frequency band are calculated as time-frequency domain features. Significant differences in energy distribution across frequency bands exist under different muscle activity states, which can be used to distinguish the degree of muscle tension.

[0034] Step S22, Human Anatomy Data Matching: Accurately match the bioelectric signal acquisition location with human anatomical data to determine the target muscle corresponding to the signal;

[0035] Specifically, an anatomical database will be established, storing detailed information on the major muscles of the human body, including muscle name, anatomical location (e.g., the trapezius muscle is located subcutaneously in the neck and back, forming a triangle on one side, and the two sides together form a rhomboid shape), origin and insertion points (the trapezius muscle originates from the superior nuchal line, external occipital protuberance, nuchal ligament and all spinous processes of the thoracic vertebrae, and inserts into the lateral 1 / 3 of the clavicle, acromion and scapular spine), and muscle fiber direction (the upper muscle fibers of the trapezius muscle run obliquely downward and outward, the middle muscle fibers run horizontally outward, and the lower muscle fibers run obliquely upward and outward).

[0036] Location tracking: Wearable devices have a built-in positioning module (such as a UWB-based positioning component) to acquire the spatial coordinates of the bioelectrical signal acquisition electrodes in real time. Using a coordinate transformation algorithm, the electrode coordinates are mapped onto a standard 3D human body model to determine the muscle area covered by the electrodes.

[0037] Muscle matching verification: Combining the body surface temperature distribution collected by infrared thermal imaging sensors, when a muscle is in a state of tension, the temperature of the corresponding body surface area may rise slightly, thereby helping to verify the accuracy of the matching between the electrode and the target muscle.

[0038] Step S23: Construction of the mapping relationship model: Based on the extracted bioelectrical signal features and the matched human anatomical data, a mapping relationship model between muscle tension and bioelectrical signal features is constructed.

[0039] Specifically, the model training data preparation involves recruiting volunteers of different ages and body types to collect data. Volunteers are asked to maintain specific postures under different muscle tension states (such as relaxation, mild tension, moderate tension, and severe tension), and the bioelectrical signal characteristics and muscle tension levels (1-4, with level 1 being the most relaxed and level 4 being the most tense) are recorded simultaneously to form a training dataset.

[0040] Model Selection and Training: A mapping model was constructed using the Support Vector Regression (SVR) algorithm, with extracted bioelectrical signal features (time domain, frequency domain, and time-frequency domain combination features) as input and muscle tension level as output. The penalty coefficient C and kernel function parameter g of the SVR model were optimized using a grid search method, and 5-fold cross-validation was used to evaluate the model's performance to ensure its generalization ability.

[0041] Model optimization: During model training, an attention mechanism is introduced to assign different weights to different features. For example, for judging muscle tension, the median frequency and root mean square value of the electromyographic signal have higher weights; for assessing fatigue, the proportion of low-frequency energy in the frequency domain features has higher weights, enabling the model to more accurately capture the correlation between key features and muscle state.

[0042] Step S24: Real-time assessment of muscle state: Using the constructed mapping model, the bioelectrical signal characteristics are analyzed in real time to output the user's muscle tension and fatigue level.

[0043] Specifically, real-time feature input: The features of the real-time acquired and preprocessed bioelectric signals are extracted according to the method in step S22 to form feature vectors, which are then input into the mapping relationship model in real time.

[0044] Model inference calculation: The model performs inference calculations based on the input feature vector, and outputs the muscle tension level (1-4) and fatigue index (0-1, where 0 indicates no fatigue and 1 indicates extreme fatigue). For example, when the model outputs a trapezius muscle tension level of 3 and a fatigue index of 0.6, it indicates that the muscle is in a state of moderate tension and moderate fatigue.

[0045] Results Validation and Update: The muscle state results output by the model are compared with the temperature changes of the corresponding muscle areas collected by the infrared thermal imaging sensor. If the temperature increase trend is consistent with the judgment of muscle tension, the credibility of the results is enhanced. If there is a deviation, the parameters of the mapping relationship model are fine-tuned through an online learning mechanism to improve the accuracy of long-term judgment.

[0046] Preferably, the specific implementation logic of step S4 is as follows:

[0047] Step S41, State Space Construction:

[0048] Body area division: Based on the human anatomical structure, the user's body is divided into 12 key massage areas. Each area is located using three-dimensional coordinates (x, y, z). The origin of the coordinates is set at the midpoint of the suprasternal notch. The x-axis is along the coronal axis to the right, the y-axis is along the sagittal axis to the front, and the z-axis is along the vertical axis to the top.

[0049] State parameter quantification: The state parameters for each region include: muscle tension level (levels 1-4, corresponding to the output of step S2), muscle fatigue index (0-1, corresponding to the output of step S2), skin temperature (°C, taken from the infrared thermal imaging data in step S1), and pressure tolerance threshold (N, calculated based on the flexible tactile sensor data in step S1, which is the maximum comfortable pressure that the user can withstand in this region); environmental parameters include: temperature (°C), humidity (%RH), and noise (dB), all of which are taken from the original data before adjustment in step S3.

[0050] State vector construction: The state parameters of 12 regions are combined with 3 environmental parameters to form a state vector S, with a vector dimension of 12×4+3=51. For example, the state parameters of the left side of the neck region are [3,0.6,36.2,45], which means that the muscle tension level of this region is 3, the fatigue index is 0.6, the skin temperature is 36.2℃, and the pressure tolerance threshold is 45N.

[0051] Step S42, Action Space Definition: Define the actions and parameter range that the massage robot can perform, and form the action space for reinforcement learning;

[0052] Specifically, the basic movement types are defined as follows: five core massage movements are defined, including: pressing (force applied perpendicular to the body surface), kneading (force applied in a relative rotational manner), massage (force applied parallel to the body surface), tapping (high-frequency impact force applied), and vibration (low-frequency reciprocating force applied). Each movement corresponds to a specific motion trajectory of the robot's end effector; for example, pressing corresponds to linear extension and contraction, and kneading corresponds to relative rotation of two fingers.

[0053] Action parameter range: The parameters for each action include: force (N), ranging from 5-50N with a step size of 1N; frequency (Hz), ranging from 0.5-3Hz with a step size of 0.1Hz; and duration (s), ranging from 2-10s with a step size of 1s. The parameter range is constrained by the reference parameters adjusted in step S3. For example, when the reference force value is adjusted to 35N in step S3, the upper limit of the force parameter in this area is set to 40N (reference value + 5N), and the lower limit is set to 30N (reference value - 5N).

[0054] Action vector encoding: The action type and parameters are combined into an action vector A. One-hot encoding is used to represent the action type (5-dimensional). The parameter part is the normalized value (0-1) of force, frequency, and duration. The total dimension of the vector is 5+3=8. For example, [1,0,0,0,0,0.6,0.3,0.5] represents a pressing action with a force of 30N ((30-5) / (50-5)=0.6), a frequency of 1.2Hz ((1.2-0.5) / (3-0.5)=0.3), and a duration of 6s ((6-2) / (10-2)=0.5).

[0055] Step S43, Reward Function Design: Construct a reward function centered on massage effect and user comfort to guide the reinforcement learning algorithm to learn the optimal strategy;

[0056] Specifically, the instant reward calculation is as follows:

[0057] Muscle condition improvement reward: When the muscle tension level of a certain area decreases by ΔT after the exercise (e.g., from level 3 to level 2), the reward value R1 = ΔT × 20; the fatigue index decreases by ΔF (e.g., from 0.6 to 0.4), the reward value R2 = ΔF × 50. If the condition does not improve or worsens, R1 and R2 are negative.

[0058] Comfort reward: Based on the flexible tactile sensor data in step S1, calculate the deviation rate D between the actual massage force and the pressure tolerance threshold: D = |actual force - 0.8 × tolerance threshold| / tolerance threshold. When D < 0.1, the reward R3 = 20; when 0.1 ≤ D < 0.2, R3 = 10; when D ≥ 0.2, R3 = -10 (0.8 × tolerance threshold is the optimal comfort force).

[0059] Environmental adaptation bonus: According to the adjustment rules in step S3, a bonus is given when the motion parameters meet the environmental adaptation requirements. For example, if the force is ≥ 1.1 times the baseline value in a low-temperature environment, the bonus is R4 = 5; otherwise, R4 = -3.

[0060] Total reward function: Total reward R = R1 + R2 + R3 + R4, with a single-step reward range of -50 to 150. Additionally, a terminal reward is set: when muscle tension in all areas is ≤2 and fatigue index is ≤0.3, an extra reward of 500 is awarded, triggering the termination of the massage process.

[0061] Step S44: Reinforcement learning model training:

[0062] The model is trained using a deep deterministic gradient algorithm: Actor network: input state vector S (51-dimensional), output action vector A (8-dimensional) through a 3-layer fully connected neural network (hidden layer dimensions are 128, 64, and 32 respectively), and the output layer uses the tanh activation function to constrain the parameter range;

[0063] Critic Network: Input state vector S and action vector A, output Q-value (action value estimate) through a 3-layer fully connected neural network (hidden layer dimensions are 128, 64 and 32 respectively), which is used to evaluate the quality of the current action;

[0064] Specific training process: Initialization of experience replay pool: Collect 10,000 sets of operation data from manual massage experts, including state S, action A, reward R, and next state S′, and store them in the experience replay pool.

[0065] Iterative training: In each iteration, 256 sets of data are randomly sampled from the replay pool. Actions are generated using the Actor network, and the Q-value is calculated using the Critic network. Gradient descent is used to update the network parameters. The Actor network aims to maximize the Q-value, while the Critic network aims to minimize the Q-value prediction error. The training cycle consists of 50,000 iterations, with a learning rate of 0.001 and a discount factor γ = 0.95.

[0066] Exploration Strategy: In the early stages of training, an ε-greedy strategy is adopted (ε decays linearly from 0.9 to 0.1) to balance the model's exploration of new actions with the utilization of known effective actions. In the later stages of training, a Gaussian noise addition strategy is adopted, which adds Gaussian noise with a mean of 0 and a standard deviation of 0.1 to the output actions of the Actor network to enhance the robustness of the strategy.

[0067] Step S45: Personalized strategy generation: Using the trained reinforcement learning model, combined with real-time input state data, a massage strategy is generated.

[0068] Specifically, the initial state input is as follows: the initial state vector S0 constructed in step S41 is input into the Actor network, and the first action A0 is output (e.g., [0,1,0,0,0,0.7,0.4,0.6], corresponding to the kneading action, with a force of 32N, a frequency of 1.5Hz, and a duration of 7s).

[0069] Massage path planning: A greedy algorithm is used to determine the order of area visits, prioritizing areas with muscle tension ≥3 or fatigue index ≥0.5. For example, in the initial state, the deltoid muscle area of ​​the shoulder (tension level 4) and the trapezius muscle area of ​​the back (fatigue index 0.7) are listed as priority areas, forming the following path: deltoid muscle area of ​​the shoulder → trapezius muscle area of ​​the back → both sides of the neck → ...

[0070] Policy Iteration Update: After each action is executed, new state data S1 is collected through a multimodal sensor, the reward R0 is calculated, and (S0, A0, R0, S1) is stored in the experience replay pool for online learning. The Actor network generates the next action A1 based on S1, and this process is repeated until the terminal reward condition is triggered.

[0071] Dynamic parameter adjustment: During strategy execution, if the pressure feedback in a certain area exceeds 1.2 times the tolerance threshold (taken from real-time data in step S1), parameter adjustment is immediately triggered, reducing the intensity by 20% while keeping the frequency unchanged to ensure massage safety.

[0072] Step S46, Policy Output and Transformation: Convert the abstract action vectors generated by reinforcement learning into control instructions that the robot can execute.

[0073] Specifically, motion parameter decoding: convert the normalized parameters in motion vector A into actual physical quantities, such as force = 5 + (force component in motion vector) × 45, frequency = 0.5 + (frequency component in motion vector) × 2.5, and duration = 2 + (time component in motion vector) × 8.

[0074] Trajectory planning: Generate the motion trajectory of the end effector according to the action type. For example, the trajectory of the pressing action is: move from the initial position (5cm away from the body surface) to the target depth (the displacement corresponding to the force, calculated through the force-displacement mapping relationship) at a speed of 50mm / s, hold for the duration, and then return to the initial position at a speed of 30mm / s.

[0075] Control command generation: The trajectory parameters are converted into control commands in the robot joint space. The inverse kinematics algorithm is used to calculate the angle changes of the six joints and send them to the servo motors through pulse width modulation (PWM) signals. The control signal frequency is 1kHz to ensure the accuracy of motion execution.

[0076] Preferably, a control system for an intelligent massage robot includes: a multimodal perception module, which consists of a flexible tactile sensor array, an infrared thermal imaging sensor, a bioelectric signal acquisition unit, and an environmental sensor, used to collect user body state data and environmental parameters;

[0077] The data fusion processing module uses a deep learning model based on an attention mechanism to fuse the data collected by the multimodal perception module to obtain comprehensive information about the user's physical state.

[0078] The muscle state modeling module extracts features from bioelectrical signals and combines them with human anatomical data to construct a muscle state mapping model to determine muscle tension and fatigue levels.

[0079] The environmental parameter adjustment module makes preliminary adjustments to the preset massage reference parameters based on the parameters collected by the environmental sensors.

[0080] The massage strategy generation module generates personalized massage control strategies based on comprehensive body condition information, muscle condition, and adjusted baseline parameters using reinforcement learning algorithms.

[0081] The real-time feedback correction module collects feedback data during the massage process and corrects the massage control strategy in real time through a PID control algorithm.

[0082] The execution module receives control signals from the massage strategy generation module and the real-time feedback correction module, and controls the massage mechanism of the massage robot to perform corresponding massage actions.

[0083] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention, through multimodal dynamic perception fusion, can comprehensively and accurately acquire user's physical state information; based on bioelectrical signal-based muscle state modeling, it achieves precise judgment of muscle state; combined with adaptive adjustment of environmental parameters, it makes massage more in line with human sensation under environmental conditions; and the personalized massage strategy generated by reinforcement learning algorithms and the real-time feedback correction mechanism ensure the effectiveness and safety of the massage. Compared with existing technologies, this invention has outstanding inventiveness and significant progress, and can provide users with higher-quality intelligent massage services. Attached Figure Description

[0084] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0085] Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0086] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0087] Example 1

[0088] Please see Figure 1 The present invention provides a technical solution: a control method for an intelligent massage robot, comprising the following steps:

[0089] Step S1: Multimodal dynamic perception fusion. This involves real-time acquisition of multi-dimensional data using a flexible tactile sensor array, an infrared thermal imaging sensor, and wearable devices mounted on the massage end of the massage robot. A deep learning model based on an attention mechanism is then used to fuse the acquired data, resulting in comprehensive information about the user's current physical state. The specific implementation logic is as follows:

[0090] Step S11, Data Preprocessing: For the pressure distribution and contact area data collected by the flexible tactile sensor array, due to the noise and drift inherent in the sensor array, a mean filtering method is first used to remove high-frequency noise. Then, a linear interpolation method is used to supplement the missing sampling points, unifying the data format into a 20×20 matrix, where each element represents the pressure value at the corresponding location. For the temperature distribution image collected by the infrared thermal imaging sensor, bad pixel repair is first performed. For abnormally high or low temperature points in the image, the average value of the surrounding pixels is used for replacement. Then, the image is processed... Normalization is performed to map temperature values ​​to a range of 0-1 to reduce the impact of different temperature ranges on model training. Finally, the image size is adjusted to the standard size of 64×64. The user's bioelectrical signals (including electromyography signals, electrocardiogram signals, etc.) collected by the wearable device are first filtered by a bandpass filter to remove power frequency interference (50Hz) and low frequency drift. Then, wavelet transform is used to denoise the signal to remove noise such as motion artifacts. The processed signal is then segmented, with each second as a data segment. The feature values ​​of each data segment are extracted to form a feature vector of length 1000.

[0091] Step S12, Feature Extraction: For the processed flexible tactile sensor array data, CNN is used for feature extraction. Through two convolutional layers and pooling layers, the first convolutional layer uses 32 3×3 convolutional kernels with a stride of 1 to perform preliminary feature extraction, and then a 2×2 pooling layer is used for downsampling; the second convolutional layer uses 64 3×3 convolutional kernels with a stride of 1 to further extract deep features, and then a 2×2 pooling layer is used to process, finally obtaining a 16×16×64 feature matrix; the image data after infrared thermal imaging processing is also extracted using CNN. Using convolution and pooling operations similar to those used in flexible tactile sensor data processing, the first convolutional layer uses 32 3×3 convolutional kernels with a stride of 1, followed by a 2×2 pooling layer; the second convolutional layer uses 64 3×3 convolutional kernels with a stride of 1, followed by a 2×2 pooling layer, resulting in an 8×8×64 feature matrix. The feature vector of the bioelectrical signal is extracted using LSTM, which is input into a two-layer LSTM network. The first LSTM layer has 128 hidden units, and the second LSTM layer has 64 hidden units. The LSTM network captures the temporal features of the bioelectrical signal, and finally outputs a feature vector of length 64.

[0092] Step S13, Attention weight calculation: The flexible tactile sensor feature matrix, infrared thermal imaging feature matrix and bioelectric signal feature vector extracted in step 12 are flattened and converted into one-dimensional vectors with lengths of 16×16×64=16384, 8×8×64=4096 and 64 respectively.

[0093] Introduce a learnable weight matrix W and a bias vector b, and perform linear transformations on the three one-dimensional vectors respectively to obtain three vectors with the same dimensions, denoted as V1, V2, and V3, each with a dimension of 256.

[0094] Calculate the attention score for each vector using the softmax function, with the formula: Attention Score(Vi)=exp(Vi·u) / (exp(V1·u)+exp(V2·u)+exp(V3·u)), where u is a learnable query vector; determine the weight of each data feature based on the attention score;

[0095] Step S14, Feature Fusion: Multiply the three feature vectors extracted in Step 12 by their corresponding attention weights to obtain weighted feature vectors W1×F1, W2×F2, and W3×F3, where W1, W2, and W3 are the attention weights for the flexible tactile sensor feature, infrared thermal imaging feature, and bioelectric signal feature, respectively, and F1, F2, and F3 are the corresponding feature vectors. Concatenate the three weighted feature vectors to obtain a fused feature vector with a length of 16384 + 4096 + 64 = 20544.

[0096] Step S15, Output of comprehensive body state information: The fused feature vector is input into a fully connected neural network. This network contains two hidden layers. The first hidden layer has 1024 neurons and uses the ReLU activation function; the second hidden layer has 256 neurons and also uses the ReLU activation function.

[0097] The output layer of a fully connected network contains multiple neurons, each corresponding to a different bodily state indicator, such as stress tolerance, muscle activity intensity, and blood circulation. The output value of each neuron is mapped to a range of 0-1 using a sigmoid function, representing the degree of the corresponding bodily state indicator.

[0098] The output values ​​of each neuron in the output layer are then integrated to form comprehensive information on the user's current physical state. This information fully reflects the user's physical condition and provides a basis for generating subsequent massage strategies.

[0099] Step S16, Model Training and Optimization: Train and optimize the model from step S15.

[0100] Specifically, in order for the model to accurately perform data fusion and assess body condition, it needs to be trained and optimized.

[0101] A large amount of labeled data is prepared, including data collected by three types of sensors and corresponding comprehensive information on the user's physical condition. Professional massage therapists then label the data according to the user's actual situation.

[0102] The labeled data is divided into training set, validation set and test set in a ratio of 7:2:1.

[0103] The mean squared error was used as the loss function, and the Adam optimizer was used to train the model. The initial learning rate was set to 0.001, and the learning rate was gradually reduced as the number of training rounds increased.

[0104] During training, the model's performance is evaluated using a validation set. Training is stopped when the loss on the validation set no longer decreases to prevent overfitting.

[0105] The trained model is tested using a test set to evaluate metrics such as accuracy and recall. If the metrics are not up to standard, the model parameters are adjusted, such as the weight matrix in the attention mechanism and the number of neurons in the fully connected network, and the model is retrained until its performance meets the requirements.

[0106] Step S2: Muscle state modeling based on bioelectrical signals. Features are extracted from the user's bioelectrical signals collected in Step S1. A mapping model between muscle tension and bioelectrical signal features is constructed using human anatomical data to determine the user's muscle tension and fatigue level in real time. Specific implementation steps are as follows:

[0107] Step S21, Bioelectric signal feature extraction: Extract feature parameters that can effectively reflect muscle state from the preprocessed bioelectric signal in step S1, including time domain features, frequency domain features and time-frequency domain features;

[0108] Specifically, time-domain feature extraction involves calculating the peak value (the maximum value in the signal), mean (the arithmetic mean of all signal values ​​within the data segment), variance (reflecting the dispersion of the signal), and root mean square (RMS) value (quantification of signal energy) for each data segment. For example, a larger RMS value for an electromyographic signal indicates higher muscle activity intensity.

[0109] Frequency domain feature extraction: The time-domain signal is converted into a frequency-domain signal using Fast Fourier Transform (FFT) to obtain the power spectral density of the signal. The center frequency (the frequency corresponding to the centroid of the power spectrum) and the median frequency (the frequency at which the cumulative power in the power spectrum reaches 50%) are extracted. When muscles are fatigued, the median frequency of the electromyographic signal shifts towards lower frequencies, which is an important indicator for judging muscle fatigue.

[0110] Time-frequency domain feature extraction: Wavelet transform is used to perform time-frequency analysis on the signal. The db4 wavelet is selected as the basis function, and the signal is decomposed into five levels to obtain wavelet coefficients for different frequency bands. The energy values ​​of the wavelet coefficients in each frequency band are calculated as time-frequency domain features. Significant differences in energy distribution across frequency bands exist under different muscle activity states, which can be used to distinguish the degree of muscle tension.

[0111] Step S22, Human Anatomy Data Matching: Accurately match the bioelectric signal acquisition location with human anatomical data to determine the target muscle corresponding to the signal;

[0112] Specifically, an anatomical database will be established, storing detailed information on the major muscles of the human body, including muscle name, anatomical location (e.g., the trapezius muscle is located subcutaneously in the neck and back, forming a triangle on one side, and the two sides together form a rhomboid shape), origin and insertion points (the trapezius muscle originates from the superior nuchal line, external occipital protuberance, nuchal ligament and all spinous processes of the thoracic vertebrae, and inserts into the lateral 1 / 3 of the clavicle, acromion and scapular spine), and muscle fiber direction (the upper muscle fibers of the trapezius muscle run obliquely downward and outward, the middle muscle fibers run horizontally outward, and the lower muscle fibers run obliquely upward and outward).

[0113] Location tracking: Wearable devices have a built-in positioning module (such as a UWB-based positioning component) to acquire the spatial coordinates of the bioelectrical signal acquisition electrodes in real time. Using a coordinate transformation algorithm, the electrode coordinates are mapped onto a standard 3D human body model to determine the muscle area covered by the electrodes.

[0114] Muscle matching verification: Combining the body surface temperature distribution collected by infrared thermal imaging sensors, when a muscle is in a state of tension, the temperature of the corresponding body surface area may rise slightly, thereby helping to verify the accuracy of the matching between the electrode and the target muscle.

[0115] Step S23: Construction of the mapping relationship model: Based on the extracted bioelectrical signal features and the matched human anatomical data, a mapping relationship model between muscle tension and bioelectrical signal features is constructed.

[0116] Specifically, the model training data preparation involves recruiting volunteers of different ages and body types to collect data. Volunteers are asked to maintain specific postures under different muscle tension states (such as relaxation, mild tension, moderate tension, and severe tension), and the bioelectrical signal characteristics and muscle tension levels (1-4, with level 1 being the most relaxed and level 4 being the most tense) are recorded simultaneously to form a training dataset.

[0117] Model Selection and Training: A mapping model was constructed using the Support Vector Regression (SVR) algorithm, with extracted bioelectrical signal features (time domain, frequency domain, and time-frequency domain combination features) as input and muscle tension level as output. The penalty coefficient C and kernel function parameter g of the SVR model were optimized using a grid search method, and 5-fold cross-validation was used to evaluate the model's performance to ensure its generalization ability.

[0118] Model optimization: During model training, an attention mechanism is introduced to assign different weights to different features. For example, for judging muscle tension, the median frequency and root mean square value of the electromyographic signal have higher weights; for assessing fatigue, the proportion of low-frequency energy in the frequency domain features has higher weights, enabling the model to more accurately capture the correlation between key features and muscle state.

[0119] Step S24: Real-time assessment of muscle state: Using the constructed mapping model, the bioelectrical signal characteristics are analyzed in real time to output the user's muscle tension and fatigue level.

[0120] Specifically, real-time feature input: The features of the real-time acquired and preprocessed bioelectric signals are extracted according to the method in step S22 to form feature vectors, which are then input into the mapping relationship model in real time.

[0121] Model inference calculation: The model performs inference calculations based on the input feature vector, and outputs the muscle tension level (1-4) and fatigue index (0-1, where 0 indicates no fatigue and 1 indicates extreme fatigue). For example, when the model outputs a trapezius muscle tension level of 3 and a fatigue index of 0.6, it indicates that the muscle is in a state of moderate tension and moderate fatigue.

[0122] Results Validation and Update: The muscle state results output by the model are compared with the temperature changes of the corresponding muscle areas collected by the infrared thermal imaging sensor. If the temperature increase trend is consistent with the judgment of muscle tension, the credibility of the results is enhanced. If there is a deviation, the parameters of the mapping relationship model are fine-tuned through an online learning mechanism to improve the accuracy of long-term judgment.

[0123] Step S3: Combined with adaptive adjustment of environmental parameters, the temperature, humidity and noise parameters of the current environment are collected by the environmental sensors set on the robot, and the preset massage intensity, frequency and time reference parameters are initially adjusted according to the environmental parameters.

[0124] Step S4: Massage strategy generation based on body state and environmental parameters. Based on the comprehensive body state information obtained in Step S1, the muscle state obtained in Step S2, and the baseline parameters adjusted in Step S3, a reinforcement learning algorithm is used to generate personalized control strategies for massage path, intensity, frequency, and duration. The specific implementation logic is as follows:

[0125] Step S41, State Space Construction:

[0126] Body area division: Based on the human anatomical structure, the user's body is divided into 12 key massage areas. Each area is located using three-dimensional coordinates (x, y, z). The origin of the coordinates is set at the midpoint of the suprasternal notch. The x-axis is along the coronal axis to the right, the y-axis is along the sagittal axis to the front, and the z-axis is along the vertical axis to the top.

[0127] State parameter quantification: The state parameters for each region include: muscle tension level (levels 1-4, corresponding to the output of step S2), muscle fatigue index (0-1, corresponding to the output of step S2), skin temperature (°C, taken from the infrared thermal imaging data in step S1), and pressure tolerance threshold (N, calculated based on the flexible tactile sensor data in step S1, which is the maximum comfortable pressure that the user can withstand in this region); environmental parameters include: temperature (°C), humidity (%RH), and noise (dB), all of which are taken from the original data before adjustment in step S3.

[0128] State vector construction: The state parameters of 12 regions are combined with 3 environmental parameters to form a state vector S, with a vector dimension of 12×4+3=51. For example, the state parameters of the left side of the neck region are [3,0.6,36.2,45], which means that the muscle tension level of this region is 3, the fatigue index is 0.6, the skin temperature is 36.2℃, and the pressure tolerance threshold is 45N.

[0129] Step S42, Action Space Definition: Define the actions and parameter range that the massage robot can perform, and form the action space for reinforcement learning;

[0130] Specifically, the basic movement types are defined as follows: five core massage movements are defined, including: pressing (force applied perpendicular to the body surface), kneading (force applied in a relative rotational manner), massage (force applied parallel to the body surface), tapping (high-frequency impact force applied), and vibration (low-frequency reciprocating force applied). Each movement corresponds to a specific motion trajectory of the robot's end effector; for example, pressing corresponds to linear extension and contraction, and kneading corresponds to relative rotation of two fingers.

[0131] Action parameter range: The parameters for each action include: force (N), ranging from 5-50N with a step size of 1N; frequency (Hz), ranging from 0.5-3Hz with a step size of 0.1Hz; and duration (s), ranging from 2-10s with a step size of 1s. The parameter range is constrained by the reference parameters adjusted in step S3. For example, when the reference force value is adjusted to 35N in step S3, the upper limit of the force parameter in this area is set to 40N (reference value + 5N), and the lower limit is set to 30N (reference value - 5N).

[0132] Action vector encoding: The action type and parameters are combined into an action vector A. One-hot encoding is used to represent the action type (5-dimensional). The parameter part is the normalized value (0-1) of force, frequency, and duration. The total dimension of the vector is 5+3=8. For example, [1,0,0,0,0,0.6,0.3,0.5] represents a pressing action with a force of 30N ((30-5) / (50-5)=0.6), a frequency of 1.2Hz ((1.2-0.5) / (3-0.5)=0.3), and a duration of 6s ((6-2) / (10-2)=0.5).

[0133] Step S43, Reward Function Design: Construct a reward function centered on massage effect and user comfort to guide the reinforcement learning algorithm to learn the optimal strategy;

[0134] Specifically, the instant reward calculation is as follows:

[0135] Muscle condition improvement reward: When the muscle tension level of a certain area decreases by ΔT after the exercise (e.g., from level 3 to level 2), the reward value R1 = ΔT × 20; the fatigue index decreases by ΔF (e.g., from 0.6 to 0.4), the reward value R2 = ΔF × 50. If the condition does not improve or worsens, R1 and R2 are negative.

[0136] Comfort reward: Based on the flexible tactile sensor data in step S1, calculate the deviation rate D between the actual massage force and the pressure tolerance threshold: D = |actual force - 0.8 × tolerance threshold| / tolerance threshold. When D < 0.1, the reward R3 = 20; when 0.1 ≤ D < 0.2, R3 = 10; when D ≥ 0.2, R3 = -10 (0.8 × tolerance threshold is the optimal comfort force).

[0137] Environmental adaptation bonus: According to the adjustment rules in step S3, a bonus is given when the motion parameters meet the environmental adaptation requirements. For example, if the force is ≥ 1.1 times the baseline value in a low-temperature environment, the bonus is R4 = 5; otherwise, R4 = -3.

[0138] Total reward function: Total reward R = R1 + R2 + R3 + R4, with a single-step reward range of -50 to 150. Additionally, a terminal reward is set: when muscle tension in all areas is ≤2 and fatigue index is ≤0.3, an extra reward of 500 is awarded, triggering the termination of the massage process.

[0139] Step S44: Reinforcement learning model training:

[0140] The model is trained using a deep deterministic gradient algorithm: Actor network: input state vector S (51-dimensional), output action vector A (8-dimensional) through a 3-layer fully connected neural network (hidden layer dimensions are 128, 64, and 32 respectively), and the output layer uses the tanh activation function to constrain the parameter range;

[0141] Critic Network: Input state vector S and action vector A, output Q-value (action value estimate) through a 3-layer fully connected neural network (hidden layer dimensions are 128, 64 and 32 respectively), which is used to evaluate the quality of the current action;

[0142] Specific training process: Initialization of experience replay pool: Collect 10,000 sets of operation data from manual massage experts, including state S, action A, reward R, and next state S′, and store them in the experience replay pool.

[0143] Iterative training: In each iteration, 256 sets of data are randomly sampled from the replay pool. Actions are generated using the Actor network, and the Q-value is calculated using the Critic network. Gradient descent is used to update the network parameters. The Actor network aims to maximize the Q-value, while the Critic network aims to minimize the Q-value prediction error. The training cycle consists of 50,000 iterations, with a learning rate of 0.001 and a discount factor γ = 0.95.

[0144] Exploration Strategy: In the early stages of training, an ε-greedy strategy is adopted (ε decays linearly from 0.9 to 0.1) to balance the model's exploration of new actions with the utilization of known effective actions. In the later stages of training, a Gaussian noise addition strategy is adopted, which adds Gaussian noise with a mean of 0 and a standard deviation of 0.1 to the output actions of the Actor network to enhance the robustness of the strategy.

[0145] Step S45: Personalized strategy generation: Using the trained reinforcement learning model, combined with real-time input state data, a massage strategy is generated.

[0146] Specifically, the initial state input is as follows: the initial state vector S0 constructed in step S41 is input into the Actor network, and the first action A0 is output (e.g., [0,1,0,0,0,0.7,0.4,0.6], corresponding to the kneading action, with a force of 32N, a frequency of 1.5Hz, and a duration of 7s).

[0147] Massage path planning: A greedy algorithm is used to determine the order of area visits, prioritizing areas with muscle tension ≥3 or fatigue index ≥0.5. For example, in the initial state, the deltoid muscle area of ​​the shoulder (tension level 4) and the trapezius muscle area of ​​the back (fatigue index 0.7) are listed as priority areas, forming the following path: deltoid muscle area of ​​the shoulder → trapezius muscle area of ​​the back → both sides of the neck → ...

[0148] Policy Iteration Update: After each action is executed, new state data S1 is collected through a multimodal sensor, the reward R0 is calculated, and (S0, A0, R0, S1) is stored in the experience replay pool for online learning. The Actor network generates the next action A1 based on S1, and this process is repeated until the terminal reward condition is triggered.

[0149] Dynamic parameter adjustment: During strategy execution, if the pressure feedback in a certain area exceeds 1.2 times the tolerance threshold (taken from real-time data in step S1), parameter adjustment is immediately triggered, reducing the intensity by 20% while keeping the frequency unchanged to ensure massage safety.

[0150] Step S46, Policy Output and Transformation: Convert the abstract action vectors generated by reinforcement learning into control instructions that the robot can execute.

[0151] Specifically, motion parameter decoding: convert the normalized parameters in motion vector A into actual physical quantities, such as force = 5 + (force component in motion vector) × 45, frequency = 0.5 + (frequency component in motion vector) × 2.5, and duration = 2 + (time component in motion vector) × 8.

[0152] Trajectory planning: Generate the motion trajectory of the end effector according to the action type. For example, the trajectory of the pressing action is: move from the initial position (5cm away from the body surface) to the target depth (the displacement corresponding to the force, calculated through the force-displacement mapping relationship) at a speed of 50mm / s, hold for the duration, and then return to the initial position at a speed of 30mm / s.

[0153] Control command generation: The trajectory parameters are converted into control commands in the robot joint space. The inverse kinematics algorithm is used to calculate the angle changes of the six joints and send them to the servo motors through pulse width modulation (PWM) signals. The control signal frequency is 1kHz to ensure the accuracy of motion execution.

[0154] Step S5: Real-time feedback correction. During the massage, the system continuously collects data on changes in the user's physical state and feedback signals through multimodal sensors, compares them with the expected results, and uses a PID control algorithm to correct the massage control strategy in real time to ensure the massage effect and safety.

[0155] Example 2

[0156] Please see Figure 2 A control system for an intelligent massage robot includes: a multimodal perception module, consisting of a flexible tactile sensor array, an infrared thermal imaging sensor, a bioelectric signal acquisition unit, and an environmental sensor, used to collect user body state data and environmental parameters;

[0157] The data fusion processing module uses a deep learning model based on an attention mechanism to fuse the data collected by the multimodal perception module to obtain comprehensive information about the user's physical state.

[0158] The muscle state modeling module extracts features from bioelectrical signals and combines them with human anatomical data to construct a muscle state mapping model to determine muscle tension and fatigue levels.

[0159] The environmental parameter adjustment module makes preliminary adjustments to the preset massage reference parameters based on the parameters collected by the environmental sensors.

[0160] The massage strategy generation module generates personalized massage control strategies based on comprehensive body condition information, muscle condition, and adjusted baseline parameters using reinforcement learning algorithms.

[0161] The real-time feedback correction module collects feedback data during the massage process and corrects the massage control strategy in real time through a PID control algorithm.

[0162] The execution module receives control signals from the massage strategy generation module and the real-time feedback correction module, and controls the massage mechanism of the massage robot to perform corresponding massage actions.

[0163] This invention discloses a control method and system for an intelligent massage robot, relating to the field of intelligent robots. The method includes: multimodal dynamic perception fusion; processing sensor data to obtain body state using an attention mechanism deep learning model; judging muscle state based on bioelectrical signal modeling; adjusting massage benchmarks in conjunction with environmental parameters; generating personalized massage strategies using reinforcement learning; and real-time feedback correction to ensure effectiveness and safety. The system includes modules such as multimodal perception. This invention addresses the problems of insufficient accuracy and difficulty in adapting to individual and environmental differences in existing technologies, possessing both inventiveness and practicality.

[0164] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A control method for an intelligent massage robot, characterized in that, Includes the following steps: Step S1: Multimodal dynamic perception fusion. The user's bioelectrical signals are collected in real time through a flexible tactile sensor array, an infrared thermal imaging sensor and wearable devices set on the massage end of the massage robot. The collected data is then fused using a deep learning model based on an attention mechanism to obtain comprehensive information about the user's current physical state. Step S2: Muscle state modeling based on bioelectric signals. Feature extraction is performed on the user's bioelectric signals collected in step S1. A mapping relationship model between muscle tension and bioelectric signal features is constructed by combining human anatomical data to judge the user's muscle tension and fatigue level in real time. Step S3: Combined with adaptive adjustment of environmental parameters, the temperature, humidity and noise parameters of the current environment are collected by the environmental sensors set on the robot, and the preset massage intensity, frequency and time reference parameters are initially adjusted according to the environmental parameters. Step S4: Massage strategy generation based on body state and environmental parameters. Based on the comprehensive body state information obtained in step S1, the muscle state obtained in step S2, and the baseline parameters adjusted in step S3, a reinforcement learning algorithm is used to generate a personalized massage path, intensity, frequency, and time control strategy. Step S5: Real-time feedback correction. During the massage, the system continuously collects data on changes in the user's physical state and feedback signals through multimodal sensors, compares them with the expected results, and uses a PID control algorithm to correct the massage control strategy in real time to ensure the massage effect and safety.

2. The control method for an intelligent massage robot according to claim 1, characterized in that: The specific implementation logic of step S1, which uses a deep learning model based on an attention mechanism to fuse the collected data and obtain comprehensive information about the user's current physical state, is as follows: Step S11, Data Preprocessing: For the pressure distribution and contact area data collected by the flexible tactile sensor array, the mean filtering method is first used to remove high-frequency noise, and then the missing sampling points are supplemented by linear interpolation. For the temperature distribution image collected by the infrared thermal imaging sensor, bad pixel repair is first performed. For abnormal high or low temperature points in the image, the average value of the surrounding pixels is used to replace them, and then image normalization is performed. For the user's bioelectrical signals collected by the wearable device, the power frequency interference and low frequency drift are first removed by the bandpass filter, and then the signal is denoised by wavelet transform to remove noise such as motion artifacts. Then the processed signal is segmented, with each second as a data segment. The feature value of each data segment is extracted to form a feature vector with a length of 1000. Step S12, Feature Extraction: For the processed flexible tactile sensor array data, CNN is used for feature extraction, ultimately obtaining a 16×16×64 feature matrix; the infrared thermal imaging image data is also processed using CNN for feature extraction. Convolution and pooling operations similar to those used in the flexible tactile sensor data processing are employed to obtain an 8×8×64 feature matrix; the feature vector of the bioelectrical signal is extracted using LSTM, capturing the temporal features of the bioelectrical signal through the LSTM network, ultimately outputting a feature vector of length 64. Step S13, Attention Weight Calculation: The flexible tactile sensor feature matrix, infrared thermal imaging feature matrix and bioelectric signal feature vector extracted in step 12 are flattened and converted into one-dimensional vectors. Introduce a learnable weight matrix W and a bias vector b, and perform linear transformations on the three one-dimensional vectors respectively to obtain three vectors with the same dimensions, denoted as V1, V2, and V3, each with a dimension of 256. Calculate the attention score for each vector using the softmax function, with the formula: AttentionScore(Vi)=exp(Vi·u) / (exp(V1·u)+exp(V2·u)+exp(V3·u)), where u is a learnable query vector; determine the weight of each data feature based on the attention score; Step S14, Feature Fusion: Multiply the three feature vectors extracted in step 12 by their corresponding attention weights to obtain weighted feature vectors W1×F1, W2×F2, and W3×F3, where W1, W2, and W3 are the attention weights of the flexible tactile sensor feature, infrared thermal imaging feature, and bioelectric signal feature, respectively, and F1, F2, and F3 are the corresponding feature vectors. Concatenate the three weighted feature vectors to obtain a fused feature vector. Step S15, Output of comprehensive body state information: The fused feature vector is input into a fully connected neural network; then the output values ​​of each neuron in the output layer are integrated to form comprehensive information on the user's current body state, providing a basis for the generation of subsequent massage strategies; Step S16, Model Training and Optimization: Train and optimize the model from step S15.

3. The control method for an intelligent massage robot according to claim 1, characterized in that: The specific implementation steps of step S2 are as follows: Step S21, Bioelectric signal feature extraction: Extract feature parameters that can effectively reflect muscle state from the preprocessed bioelectric signal in step S1, including time domain features, frequency domain features and time-frequency domain features; Step S22, Human Anatomy Data Matching: Accurately match the bioelectric signal acquisition location with human anatomical data to determine the target muscle corresponding to the signal; Step S23: Construction of the mapping relationship model: Based on the extracted bioelectrical signal features and the matched human anatomical data, a mapping relationship model between muscle tension and bioelectrical signal features is constructed. Step S24: Real-time assessment of muscle state: Using the constructed mapping model, the bioelectrical signal characteristics are analyzed in real time to output the user's muscle tension and fatigue level.

4. The control method for an intelligent massage robot according to claim 1, characterized in that: The specific implementation logic of step S4 is as follows: Step S41, State Space Construction: Body area division: Based on the human anatomical structure, the user's body is divided into 12 key massage areas. Each area is located using three-dimensional coordinates (x, y, z). The origin of the coordinates is set at the midpoint of the suprasternal notch. The x-axis is along the coronal axis to the right, the y-axis is along the sagittal axis to the front, and the z-axis is along the vertical axis to the top. State parameter quantification: The state parameters for each region include: muscle tension level, muscle fatigue index, skin temperature, and pressure tolerance threshold; environmental parameters include: temperature, humidity, and noise, all of which are taken from the original data before the adjustment in step S3. State vector construction: Combine the state parameters of 12 regions with 3 environmental parameters to form a state vector S; Step S42, Action Space Definition: Define the actions and parameter range that the massage robot can perform, and form the action space for reinforcement learning; Step S43, Reward Function Design: Construct a reward function centered on massage effect and user comfort to guide the reinforcement learning algorithm to learn the optimal strategy; Step S44: Reinforcement learning model training: The model is trained using a deep deterministic gradient algorithm: Actor network: input state vector S, through a 3-layer fully connected neural network, output action vector A, and the output layer uses the tanh activation function to constrain the parameter range; Critic network: It takes a state vector S and an action vector A as input, and outputs a Q-value through a 3-layer fully connected neural network to evaluate the quality of the current action; Step S45: Personalized strategy generation: Using the trained reinforcement learning model, combined with real-time input state data, a massage strategy is generated. Step S46, Policy Output and Transformation: Convert the abstract action vectors generated by reinforcement learning into control instructions that the robot can execute.

5. The control system for an intelligent massage robot according to claim 1, characterized in that, include: The multimodal sensing module consists of a flexible tactile sensor array, an infrared thermal imaging sensor, a bioelectric signal acquisition unit, and an environmental sensor, and is used to collect user physical status data and environmental parameters. The data fusion processing module uses a deep learning model based on an attention mechanism to fuse the data collected by the multimodal perception module to obtain comprehensive information about the user's physical state. The muscle state modeling module extracts features from bioelectrical signals and combines them with human anatomical data to construct a muscle state mapping model to determine muscle tension and fatigue levels. The environmental parameter adjustment module makes preliminary adjustments to the preset massage reference parameters based on the parameters collected by the environmental sensors. The massage strategy generation module generates personalized massage control strategies based on comprehensive body condition information, muscle condition, and adjusted baseline parameters using reinforcement learning algorithms. The real-time feedback correction module collects feedback data during the massage process and corrects the massage control strategy in real time through a PID control algorithm. The execution module receives control signals from the massage strategy generation module and the real-time feedback correction module, and controls the massage mechanism of the massage robot to perform corresponding massage actions.

Images