Abdominal acupoint adaptive massage method based on neural network

Abstract

The invention discloses an abdominal acupoint adaptive massage method based on a neural network. The method comprises the following steps: applying machine disturbance to a set acupoint area position of an abdomen and collecting a tension rebound signal; performing two-dimensional compression Hilbert envelope analysis on the tension rebound signal to generate a tension echo spectrum; obtaining a cave area activity vector; constructing a multi-channel neurotransmitter prediction network, and performing reverse derivation to obtain a stimulation parameter set; matching the stimulation parameter set with the acupoint region activity vector, and performing sequence allocation optimization by adopting a probability uniformization allocation method based on diagonal guidance sparse constraint to generate an optimal massage action sequence; electromyographic signals of the target acupoint area are collected, and a prediction result is obtained; and when the deviation between the prediction result and the preset neurotransmitter target state exceeds a preset tolerance threshold, dynamically replacing and updating the optimal massage action sequence. Neural network reasoning and probability optimization distribution are fused, and intelligent acupoint massage action generation and closed-loop adjustment are achieved.

Description

Technical Field

[0001] This invention relates to the field of neural modulation and intelligent massage technology, and in particular to an adaptive abdominal acupoint massage method based on neural networks. Background Technology

[0002] With the development of neuromodulation therapy, personalized TCM intervention, and adaptive rehabilitation systems, intelligent modeling and real-time feedback control of the neurotransmitter regulatory effects of acupoint stimulation have become research frontiers. Existing abdominal massage systems mainly rely on fixed programmed movements or single electromyography (EMG) monitoring indicators for acupoint identification and intervention feedback. However, the following problems still exist in key aspects such as multi-source physiological data fusion, neural pathway modeling, and stimulation parameter optimization: The collected abdominal tissue tension signals and neural response indicators are significantly affected by subcutaneous tissue thickness, body position changes, and individual differences. This leads to key response features such as local stress release amplitude and rebound rate indicators being prone to drift and random distortion during dynamic massage, affecting the consistency and reliability of acupoint function assessment. Existing massage equipment mostly adopts a static mapping method of regularly arranged acupoints and fixed stimulation parameters, lacking precise modeling and back-inference of the key neurotransmitter states of serotonin, dopamine, and norepinephrine. This results in a disconnect between the massage application point and the target neural state, and the generation of movements lacks physiological directionality. Acupoint stimulation sequences are mostly generated by heuristic logic or simplified sorting, ignoring the complex mapping relationship between the stimulation triplets and acupoint activity. Existing scheduling strategies generally suffer from poor sparsity and high coupling in the allocation of stimulation resources, resulting in insufficient efficiency and targeting of massage movements, making it difficult to achieve continuous closed-loop massage strategy optimization.

[0003] Therefore, how to provide an adaptive abdominal acupoint massage method based on neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose an adaptive abdominal acupoint massage method based on neural networks. This invention integrates tension rebound analysis, multi-channel neurotransmitter prediction network and probability allocation optimization method to construct an adaptive abdominal acupoint massage process, realizing closed-loop control from physiological signal acquisition and stimulation parameter back-inference to action sequence optimization and neurotransmitter regulation. It has the advantages of accurate identification, targeted reasoning, adaptive adjustment and stable convergence, and improves the individual adaptability and neuromodulation effect of massage.

[0005] The adaptive abdominal acupoint massage method based on neural networks according to an embodiment of the present invention includes the following steps: Machine disturbances are applied to acupoints on the abdomen, and tension rebound signals of the corresponding acupoints are collected after each machine disturbance. Two-dimensional compressed Hilbert envelope analysis was performed on the tension rebound signal to extract the disturbance echo features and generate the tension echo spectrum of the corresponding acupoint area. Based on the tension echo spectrum, obtain the acupoint activity vector; Based on the preset neurotransmitter target state, a multi-channel neurotransmitter prediction network is constructed, and the set of stimulation parameters is derived in reverse. The set of stimulation parameters is matched with the acupoint activity vector, and the sequence allocation is optimized by a probability-consistent allocation method based on diagonal guided sparsity constraints to generate the optimal massage action sequence. Based on the optimal massage sequence, electromyographic signals of the target acupoint area are periodically collected to obtain the prediction results of the current massage on the change of neurotransmitter concentration. When the predicted result deviates from the preset neurotransmitter target state by more than a preset tolerance threshold, the optimal massage action sequence is dynamically replaced and updated to complete the adaptive adjustment of the massage.

[0006] Optionally, the machine disturbance is a periodic low-frequency mechanical vibration with a vibration frequency of 0.3Hz to 0.9Hz and a vibration duration of 0.5s. When applied to each set cavity location, it does not generate cross-cavity interference and maintains a constant external contact pressure before and after the disturbance.

[0007] Optionally, the step of performing two-dimensional compressed Hilbert envelope analysis on the tension rebound signal to extract disturbance echo features and generate a tension echo spectrum corresponding to the acupoint area specifically involves: Bandpass filtering was performed on the tension rebound signals collected from each acupoint area. The filtering frequency range was 0.1Hz to 10Hz to filter out high-frequency interference components and baseline drift terms. A Hilbert transform is applied to the filtered tension rebound signal to generate a complex analytic signal. The instantaneous amplitude signal is obtained by calculating the magnitude sequence of the complex analytic signal, forming the time domain envelope signal of the acupoint area. The time-domain envelope signals of each acupoint are stacked according to the order of acupoint arrangement to form a two-dimensional data matrix, which serves as a spatiotemporal joint response representation of tension rebound. In the spatial dimension, each column of the two-dimensional data matrix is ​​subjected to L1 norm normalization. Principal component analysis is performed on the two-dimensional data matrix in the time dimension. The first three principal components are extracted as the compression result. If the cumulative variance contribution rate of the first three principal components is less than 85%, the fourth principal component is extracted and added to the compression result. The envelope peak value, rising edge slope and full width at half height parameters of each cavity region are extracted from the compressed two-dimensional data matrix to construct the perturbation echo characteristics of each cavity region. The disturbance echo characteristics of each acupoint area are feature-coded, and a tension echo spectrum is constructed based on the acupoint area location index. The tension echo spectrum is used to represent the dynamic tissue response capability of each acupoint area to disturbance.

[0008] Optionally, obtaining the acupoint activity vector based on the tension echo spectrum specifically involves: The perturbation echo features contained in each cavity region of the tension echo spectrum are subjected to feature normalization processing. The Z-score normalization method is used to normalize the mean of the envelope peak value, the rising slope and the full width at half height to zero and the variance to unitize, respectively. The three normalized indicators were fused using a weighted superposition calculation method, ensuring that the sum of the weights was 1, and the functional activity value of each acupoint area was calculated. Based on the acupoint area arrangement index, the functional activity values ​​of each acupoint area are used to construct an acupoint area activity vector.

[0009] Optionally, the step of constructing a multi-channel neurotransmitter prediction network based on a preset neurotransmitter target state and then deriving the stimulus parameter set in reverse is as follows: The preset neurotransmitter target states include an increase of serotonin concentration of 10% to 30%, a dopamine concentration maintained within ±5% of the initial value, and a decrease of norepinephrine concentration of 5% to 20%. A multi-channel neurotransmitter prediction network was constructed, comprising a multi-input structure that receives inputs of stimulation frequency, force amplitude, and stimulation duration, as well as electromyographic feature inputs, for predicting neurotransmitter states at different stages; two hidden layers, the first hidden layer containing 32 nodes and the second hidden layer containing 16 nodes, with each hidden layer followed by a ReLU activation function for nonlinear transformation; and an output layer containing 3 nodes, corresponding to the predicted concentration changes of serotonin, dopamine, and norepinephrine, respectively. The loss function is defined as the weighted mean square error between the predicted concentration change of each neurotransmitter and the corresponding target concentration change. The gradient descent method is used to iteratively optimize the initial values ​​of the input stimulus frequency, force amplitude, and stimulus duration. The iteration stops when the set maximum number of iterations is reached. The stimulus frequency, force amplitude, and stimulus duration that cause the loss function to converge are used as the set of stimulus parameters for back-derived and output.

[0010] Optionally, the step of matching the set of stimulation parameters with the acupoint activity vector and using a probability-consistent allocation method based on diagonal-guided sparsity constraints to optimize sequence allocation and generate an optimal massage action sequence specifically involves: Each set of stimulus parameters in the stimulus parameter set is represented as a stimulus triplet containing stimulus frequency, force amplitude and stimulus duration, and numbered to form a stimulus parameter vector group. Initialize a matching probability matrix between a set of stimulus parameter vectors and acupoints, in which the initial allocation probability of each stimulus triplet to any acupoint is equal. By employing a diagonal guidance approach, an initial allocation trend dominated by diagonal elements is constructed; In each round of allocation update, a sparsity constraint rule is introduced, which allows the matching probability above the set threshold to be retained in only 2 acupoint areas for each set of stimulus parameters, and the matching probability of other positions is compressed according to the decreasing rule to sparsify the allocation results. The updated matching probability matrix is ​​subjected to probability uniformity processing, and the probabilities are normalized in the row direction and column direction respectively, so that the sum of the probabilities of each set of stimulation parameters being assigned to each acupoint area is 1, and the sum of the probabilities of all combinations of stimulation parameters received by each acupoint area remains consistent. When the magnitude of the change in the value of the corresponding position of the matching probability matrix is ​​less than the set convergence threshold in two consecutive iterations, and the total normalization error of the row vector and column vector is lower than the set error lower limit, the matching probability matrix is ​​considered to have reached a stable structure, the allocation optimization process is terminated, and the acupoint number with the highest probability corresponding to each set of stimulation parameters is extracted to form a one-to-one stimulation allocation relationship. The optimal massage action sequence is output according to the order of acupoints in the stimulation distribution relationship. The optimal massage action sequence consists of four components: acupoint number, stimulation frequency, force amplitude, and stimulation duration.

[0011] Optionally, the construction of a diagonally dominant initial allocation trend through diagonal guidance specifically involves: Cells in the matching probability matrix that are at the same or adjacent positions as the stimulation triplet number and the acupoint number are marked as the main diagonal region. The initial probability value in the main diagonal region is set as a high weight value, and the other off-diagonal regions are set as low weight values. The high weight value is between 0.8 and 1.0, and the low weight value is between 0.0 and 0.2. The matching position with an absolute value of 0 for the number difference is set as the highest initial probability value, the matching position with a number difference of 1 is set as the second highest initial probability value, and the matching position with a number difference greater than or equal to 2 is set as the basic initial probability value, thereby constructing an initial allocation trend centered on the diagonal and decreasing in the adjacent regions.

[0012] Optionally, based on the optimal massage sequence, periodically collecting electromyographic signals from the target acupoint area to obtain the predictive results of the current massage on changes in neurotransmitter concentration specifically involves: According to the execution cycle of each massage action in the optimal massage action sequence, electromyographic signals are collected from the target acupoint area. The sampling frequency is set to 500Hz to 1000Hz, and the sampling time window is within 1 second after the end of each massage action. The acquired electromyographic (EMG) signals were subjected to bandpass filtering with a frequency range of 20 Hz to 450 Hz. EMG features were extracted from the filtered EMG signals, including root mean square (RMS) value, mean absolute value, zero crossover rate, waveform factor, power spectral density center frequency, and median frequency. The extracted electromyographic features and corresponding stimulation parameters are input into the multichannel neurotransmitter prediction network to generate predicted values ​​of serotonin, dopamine and norepinephrine concentration changes under the current massage state, thereby obtaining the prediction results of the current massage on the neurotransmitter concentration changes.

[0013] Optionally, the dynamic replacement and updating of the optimal massage action sequence specifically involves: In the current optimal massage action sequence, identify acupoints where the predicted result deviates from the preset neurotransmitter target state by more than a preset tolerance threshold, and mark them as areas to be updated; For the region to be updated, re-search the stimulus parameter set for other stimulus triples that have not been assigned or have a low assignment probability, and construct a candidate massage action set; Based on the current functional activity value of the acupoint area and the direction of neurotransmitter prediction deviation, a set of stimulation parameters are selected from the candidate massage action set to generate a concentration change value that is corrected in the target direction in the multi-channel neurotransmitter prediction network. The new stimulation parameters are used to replace the corresponding positions in the current optimal massage action sequence, while keeping the original stimulation parameters of the remaining acupoints unchanged, thus completing the local update of the massage action sequence.

[0014] The beneficial effects of this invention are: This invention addresses the challenges of quantifying tension response, lack of neural targeting of stimulation parameters, and lag in massage strategy adjustment in abdominal acupoint massage by integrating two-dimensional compressed Hilbert envelope analysis of tension rebound signals with a probability-consistent allocation method based on diagonal guided sparsity constraints. It employs a unified machine perturbation standard and bandpass filtering mechanism, combined with L1 norm normalization and principal component analysis, to achieve compressed expression of perturbation echo features and construction of tension echo maps under high-noise conditions. During stimulus generation, a multi-channel neurotransmitter prediction network is constructed. Based on preset target states of serotonin, dopamine, and norepinephrine, a set of stimulation parameters consisting of stimulation frequency, force amplitude, and stimulation duration is derived using backpropagation and gradient descent. By matching the set of stimulation parameters with the acupoint activity vector, a system is constructed using stimulation triplet numbering and... The difference in acupoint numbers serves as the guiding diagonal dominant matching probability matrix. A sparsity constraint is introduced to compress the allocation range, and a probability uniformity normalization rule is applied, ultimately converging to generate the optimal massage action sequence. During execution, electromyographic signals from the target acupoints are periodically collected and input into a multi-channel neurotransmitter prediction network along with the current massage action to obtain the predicted neurotransmitter concentration changes caused by the current massage. When the predicted result deviates from the preset neurotransmitter target state by more than a preset tolerance threshold, unassigned or low-probability triples in the stimulus parameter set are re-searched based on the deviation direction, and the corresponding stimulus parameters in the optimal massage action sequence are replaced, completing dynamic replacement and updating. This ultimately achieves the overall integration from tension rebound acquisition, perturbation feature compression, neurotransmitter targeted derivation to closed-loop adaptive adjustment of massage actions, improving the physiological accuracy, parameter adaptability, and regulatory stability of abdominal acupoint massage. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the overall process of the adaptive abdominal acupoint massage method based on neural networks proposed in this invention. Figure 2 This is a flowchart of the acquisition of acupoint tension rebound signal and two-dimensional compressed Hilbert envelope analysis for the adaptive abdominal acupoint massage method based on neural networks proposed in this invention. Figure 3 This is a flowchart illustrating the probability-consistent allocation optimization of the stimulation parameter set and acupoint activity vector matching in the neural network-based adaptive abdominal acupoint massage method proposed in this invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figures 1-3 The adaptive abdominal acupoint massage method based on neural networks includes the following steps: Machine disturbances are applied to acupoints on the abdomen, and tension rebound signals of the corresponding acupoints are collected after each machine disturbance. Two-dimensional compressed Hilbert envelope analysis was performed on the tension rebound signal to extract the disturbance echo features and generate the tension echo spectrum of the corresponding acupoint area. Based on the tension echo spectrum, obtain the acupoint activity vector; Based on the preset neurotransmitter target state, a multi-channel neurotransmitter prediction network is constructed, and the set of stimulation parameters is derived in reverse. The set of stimulation parameters is matched with the acupoint activity vector, and the sequence allocation is optimized by a probability-consistent allocation method based on diagonal guided sparsity constraints to generate the optimal massage action sequence. Based on the optimal massage sequence, electromyographic signals of the target acupoint area are periodically collected to obtain the prediction results of the current massage on the change of neurotransmitter concentration. When the predicted result deviates from the preset neurotransmitter target state by more than a preset tolerance threshold, the optimal massage action sequence is dynamically replaced and updated to complete the adaptive adjustment of the massage.

[0018] This invention establishes a neural network-guided adaptive massage method, achieving for the first time a closed-loop optimized control of the entire process from acupoint tissue response signals to stimulation actions, significantly improving the individualization of massage stimulation and the consistency of physiological feedback. It employs perturbation acquisition and signal atlas construction to realistically reflect the dynamic tension of acupoint areas. By constructing acupoint activity representation and integrating neurotransmitter prediction and action optimization mechanisms, it solves the problems of coarse stimulation action settings and lagging feedback regulation in traditional massage methods, making it suitable for rehabilitation scenarios requiring continuous regulation of neurotransmitter states.

[0019] In this embodiment, the machine disturbance is a periodic low-frequency mechanical vibration with a vibration frequency of 0.3Hz to 0.9Hz and a vibration duration of 0.5s. When applied to each set cavity location, it does not cause cross-cavity interference and maintains a constant external contact pressure before and after the disturbance.

[0020] In this embodiment, the step of performing two-dimensional compressed Hilbert envelope analysis on the tension rebound signal to extract disturbance echo features and generate a tension echo spectrum corresponding to the acupoint area specifically involves: Bandpass filtering was performed on the tension rebound signals collected from each acupoint area. The filtering frequency range was 0.1Hz to 10Hz to filter out high-frequency interference components and baseline drift terms. A Hilbert transform is applied to the filtered tension rebound signal to generate a complex analytic signal. The instantaneous amplitude signal is obtained by calculating the magnitude sequence of the complex analytic signal, forming the time domain envelope signal of the acupoint area. The time-domain envelope signals of each acupoint are stacked according to the order of acupoint arrangement to form a two-dimensional data matrix, which serves as a spatiotemporal joint response representation of tension rebound. In the spatial dimension, each column of the two-dimensional data matrix is ​​subjected to L1 norm normalization. Principal component analysis is performed on the two-dimensional data matrix in the time dimension. The first three principal components are extracted as the compression result. If the cumulative variance contribution rate of the first three principal components is less than 85%, the fourth principal component is extracted and added to the compression result. The envelope peak value, rising edge slope and full width at half height parameters of each cavity region are extracted from the compressed two-dimensional data matrix to construct the perturbation echo characteristics of each cavity region. The disturbance echo characteristics of each acupoint area are feature-coded, and a tension echo spectrum is constructed based on the acupoint area location index. The tension echo spectrum is used to represent the dynamic tissue response capability of each acupoint area to disturbance.

[0021] By introducing two-dimensional compressed Hilbert envelope analysis into the tension rebound signal, not only is information fusion between the time and spatial domains achieved, but the redundancy of the original data is also effectively reduced. Combining filtering, normalization, principal component extraction, and feature combination, the constructed tension echo spectrum exhibits strong discriminative power and dynamic response representation capabilities, contributing to the accuracy of subsequent activity calculations and stimulus matching. Compared to traditional methods based on single-time-domain amplitude or frequency-domain spectral analysis, this method possesses greater advantages in time-space perception fusion, significantly enhancing the accuracy and stability of acupoint tissue dynamic response modeling.

[0022] In this embodiment, obtaining the acupoint activity vector based on the tension echo spectrum specifically involves: The perturbation echo features contained in each cavity region of the tension echo spectrum are subjected to feature normalization processing. The Z-score normalization method is used to normalize the mean of the envelope peak value, the rising slope and the full width at half height to zero and the variance to unitize, respectively. The three normalized indicators were fused using a weighted superposition calculation method, ensuring that the sum of the weights was 1, and the functional activity value of each acupoint area was calculated. Based on the acupoint area arrangement index, the functional activity values ​​of each acupoint area are used to construct an acupoint area activity vector.

[0023] By standardizing and weighting the perturbation echo features in the tension echo spectrum, a continuous acupoint activity vector representation is formed, which not only has physiological interpretability but also supports numerical alignment and spatial mapping with stimulation parameters.

[0024] In this embodiment, the step of constructing a multi-channel neurotransmitter prediction network based on a preset neurotransmitter target state and then deriving the stimulus parameter set in reverse is specifically as follows: The preset neurotransmitter target states include an increase of serotonin concentration of 10% to 30%, a dopamine concentration maintained within ±5% of the initial value, and a decrease of norepinephrine concentration of 5% to 20%. A multi-channel neurotransmitter prediction network was constructed, comprising a multi-input structure that receives inputs of stimulation frequency, force amplitude, and stimulation duration, as well as electromyographic feature inputs, for predicting neurotransmitter states at different stages; two hidden layers, the first hidden layer containing 32 nodes and the second hidden layer containing 16 nodes, with each hidden layer followed by a ReLU activation function for nonlinear transformation; and an output layer containing 3 nodes, corresponding to the predicted concentration changes of serotonin, dopamine, and norepinephrine, respectively. The loss function is defined as the weighted mean square error between the predicted concentration changes of each neurotransmitter and the corresponding target concentration changes: ; in, Represents the loss function. , and These represent the predicted values ​​of changes in the concentrations of serotonin, dopamine, and norepinephrine, respectively. , and These represent the corresponding target concentration changes. , and This represents the loss weight coefficient corresponding to each neurotransmitter, satisfying... ; The gradient descent method is used to iteratively optimize the initial values ​​of the input stimulus frequency, force amplitude, and stimulus duration. The iteration stops when the set maximum number of iterations is reached. The stimulus frequency, force amplitude, and stimulus duration that cause the loss function to converge are used as the set of stimulus parameters for back-derived and output.

[0025] This approach, by constructing a multi-channel neurotransmitter prediction network, establishes for the first time a direct mapping relationship between massage parameters and changes in the concentrations of three key neurotransmitters, enabling stimulation strategies to be derived in reverse based on the target neurotransmitter state. Through iterative iteration using a gradient optimization mechanism, a precise set of stimulation parameters is obtained, achieving synergistic regulation of serotonin, dopamine, and norepinephrine. It possesses advantages such as explicit target regulation, rapid parameter optimization convergence, and a compact network structure, significantly enhancing the initiative and precision of massage behavior in regulating the neurotransmitter system state.

[0026] In this embodiment, the process of matching the set of stimulation parameters with the acupoint activity vector and optimizing the sequence allocation using a probability-consistent allocation method based on diagonal guided sparsity constraints to generate the optimal massage action sequence specifically involves: Each set of stimulus parameters in the stimulus parameter set is represented as a stimulus triplet containing stimulus frequency, force amplitude and stimulus duration, and numbered to form a stimulus parameter vector group. Initialize a matching probability matrix between a set of stimulus parameter vectors and acupoints, in which the initial allocation probability of each stimulus triplet to any acupoint is equal. By employing a diagonal guidance approach, an initial allocation trend dominated by diagonal elements is constructed; In each round of allocation update, a sparsity constraint rule is introduced, which allows the matching probability above the set threshold to be retained in only 2 acupoint areas for each set of stimulus parameters, and the matching probability of other positions is compressed according to the decreasing rule to sparsify the allocation results. Let there be a total of A stimulus triplet constitutes a stimulus parameter vector set. : ; in, Indicates the first A stimulus triplet Indicates the first The stimulation frequency of a stimulus triplet Indicates the first The magnitude of the effect of the stimulus triplet Indicates the first The duration of stimulation in a stimulus triplet; Let the total number of acupoints in the abdomen be... Construct a matching probability matrix ,in Indicates the first The stimulus triplet is assigned to the first The probability of each acupoint area being allocated. Initialization: ; Construct a diagonally guided weight mask matrix ,in: ; in, The difference between the stimulus number and the acupoint number is... Initialization of guiding weights at that time This is the initial value representing the highest probability when the numbers are perfectly aligned. This is the second highest probability initial value when the numbers are adjacent. This is the initial value of the base probability when the numbers differ significantly, i.e., the low probability initial value; Each iteration Sparsification is performed on each... Keep the two largest ones in the corresponding row. The location, and the remaining elements multiplied by the compression factor. ,Right now: ; in, This represents the allocation probability after sparsification. Indicates the current number In the first iteration, the... The stimulus triplet is assigned to the first The probability of allocation to each acupoint area Indicates the first The column numbers corresponding to the two largest matching probabilities in the row; The updated matching probability matrix is ​​subjected to probability uniformity processing, and the probabilities are normalized in the row direction and column direction respectively, so that the sum of the probabilities of each set of stimulation parameters being assigned to each acupoint area is 1, and the sum of the probabilities of all combinations of stimulation parameters received by each acupoint area remains consistent. When the magnitude of the change in the value of the corresponding position of the matching probability matrix is ​​less than the set convergence threshold in two consecutive iterations, and the total normalization error of the row vector and column vector is lower than the set error lower limit, the matching probability matrix is ​​considered to have reached a stable structure, the allocation optimization process is terminated, and the acupoint number with the highest probability corresponding to each set of stimulation parameters is extracted to form a one-to-one stimulation allocation relationship. The optimal massage action sequence is output according to the order of acupoints in the stimulation distribution relationship. The optimal massage action sequence consists of four components: acupoint number, stimulation frequency, force amplitude, and stimulation duration.

[0027] A matching strategy between stimuli and acupoints was designed using diagonal guided sparsity constraints and a probability uniformity method. This not only significantly reduced allocation redundancy but also avoided allocation conflicts between acupoints. Through guided initialization, sparse compression, and bidirectional normalization iteration, stable convergence of the allocation probability matrix was achieved, improving the accuracy and controllability of massage action allocation. The resulting optimal massage action sequence exhibits high matching with the physiological state of acupoints, significantly outperforming traditional massage action selection based on heuristics or fixed templates.

[0028] In this embodiment, the construction of a diagonally dominant initial allocation trend through a diagonal guidance method specifically involves: Cells in the matching probability matrix that are at the same or adjacent positions as the stimulation triplet number and the acupoint number are marked as the main diagonal region. The initial probability value in the main diagonal region is set as a high weight value, and the other off-diagonal regions are set as low weight values. The high weight value is between 0.8 and 1.0, and the low weight value is between 0.0 and 0.2. The matching position with an absolute value of 0 for the number difference is set as the highest initial probability value, the matching position with a number difference of 1 is set as the second highest initial probability value, and the matching position with a number difference greater than or equal to 2 is set as the basic initial probability value, thereby constructing an initial allocation trend centered on the diagonal and decreasing in the adjacent regions.

[0029] By constructing an initialization probability weight structure centered on the main diagonal, the stimulation triplet and the acupoint area have a reasonable prior matching trend in the initial stage. Compared with the traditional uniform initialization method, it is more suitable for optimization modeling under small sample conditions and has the advantages of good initialization effect, fewer iterations and low risk of local optima.

[0030] In this embodiment, the step of periodically collecting electromyographic signals from the target acupoint area based on the optimal massage sequence to obtain the predictive result of the current massage on changes in neurotransmitter concentration specifically involves: According to the execution cycle of each massage action in the optimal massage action sequence, electromyographic signals are collected from the target acupoint area. The sampling frequency is set to 500Hz to 1000Hz, and the sampling time window is within 1 second after the end of each massage action. The acquired electromyographic (EMG) signals were subjected to bandpass filtering with a frequency range of 20 Hz to 450 Hz. EMG features were extracted from the filtered EMG signals, including root mean square (RMS) value, mean absolute value, zero crossover rate, waveform factor, power spectral density center frequency, and median frequency. The extracted electromyographic features and corresponding stimulation parameters are input into the multichannel neurotransmitter prediction network to generate predicted values ​​of serotonin, dopamine and norepinephrine concentration changes under the current massage state, thereby obtaining the prediction results of the current massage on the neurotransmitter concentration changes.

[0031] By collecting electromyographic signals and extracting multidimensional electromyographic features after each massage cycle, and combining them with the current stimulation parameters as input, a multi-channel neurotransmitter prediction network is used to dynamically estimate the neurotransmitter response. An effective perception-prediction closed-loop link is established, which breaks through the limitations of traditional methods that rely solely on external observation or indirect indicators to judge the regulatory effect. It can obtain the physiological response change trend in real time, and has the ability to evaluate the effect and make decision feedback in real time, effectively enhancing the physiological self-adaptation and feedback closed-loop characteristics of the entire system.

[0032] In this embodiment, the dynamic replacement and updating of the optimal massage action sequence specifically includes: In the current optimal massage action sequence, identify acupoints where the predicted result deviates from the preset neurotransmitter target state by more than a preset tolerance threshold, and mark them as areas to be updated; For the region to be updated, re-search the stimulus parameter set for other stimulus triples that have not been assigned or have a low assignment probability, and construct a candidate massage action set; Based on the current functional activity value of the acupoint area and the direction of neurotransmitter prediction deviation, a set of stimulation parameters are selected from the candidate massage action set to generate a concentration change value that is corrected in the target direction in the multi-channel neurotransmitter prediction network. The new stimulation parameters are used to replace the corresponding positions in the current optimal massage action sequence, while keeping the original stimulation parameters of the remaining acupoints unchanged, thus completing the local update of the massage action sequence.

[0033] When the current massage sequence does not meet the target neurotransmitter state, this method achieves timely optimization and fine-tuning of the massage strategy by identifying off-center acupoints, reselecting candidate stimuli, and updating the sequence locally. It achieves non-destructive replacement of the stimulation scheme through the feedback loop within the model, ensuring the continuity of massage movements and individual adaptability. In practical applications, it significantly improves the robustness, flexibility, and physiological consistency regulation ability of the system, which is an important foundation for realizing dynamic personalized massage intervention.

[0034] Example 1: To verify the feasibility of this invention in practice, it was applied to the intelligent massage clinical laboratory of the Rehabilitation Medicine Department of a tertiary-level traditional Chinese medicine hospital for clinical adjunctive treatment of patients with functional gastrointestinal disorders. The experimental subjects were 60 adult patients clinically diagnosed with functional gastrointestinal disorders, aged 25 to 60 years, with a male-to-female ratio close to 1:1. All patients possessed the ability to express themselves and cooperate. The main objective of the experiment was to verify the actual effectiveness of the abdominal acupoint adaptive massage method described in this invention in regulating patients' neurotransmitter levels, alleviating related symptoms, and improving individualized responses.

[0035] In this embodiment, the operator first applies a pre-defined perturbation program to the patient's abdomen, sequentially applying periodic low-frequency mechanical perturbations to the Zhongwan, Tianshu, and Daheng acupoints related to the digestive system. Each perturbation lasts 0.5 seconds, with a frequency set at 0.5 Hz. During the perturbation, a multi-point flexible contact probe is used to simultaneously acquire tension rebound signals. After bandpass filtering and Hilbert transform processing, the perturbation echo features, including envelope peak value, rising edge slope, and full width at half maximum (FWHM), are extracted to ultimately construct the individual patient's tension echo map.

[0036] The system further generates acupoint activity vectors using the PCA-compressed two-dimensional echo matrix and inputs them into a multi-channel neurotransmitter prediction network. Based on the set target states—serotonin concentration increasing by 10% to 30%, dopamine concentration maintained within ±5% of the initial value, and norepinephrine concentration decreasing by 5% to 20%—the system reverse-engineers the set of stimulation parameters that satisfy these targets. In the massage action allocation optimization stage, the system employs a probabilistic consistency algorithm with diagonally guided sparsity constraints. After constructing an initial allocation trend, it dynamically iterates to generate the individual optimal massage action sequence.

[0037] Each patient received 30 minutes of individualized massage therapy once daily for five consecutive days. Electromyography (EMG) data was collected before and after the massage, and changes in neurotransmitter concentrations were monitored. The concentration levels of relevant neurotransmitters in the blood were measured using ELISA. Neurotransmitter status was predicted based on the real-time EMG signals, and the movement sequence was dynamically adjusted as needed to ensure the continuous effectiveness of the regulatory loop.

[0038] A comparative analysis of data from 60 patients demonstrated that the method of this invention exhibits significant advantages in improving treatment response rate and the precision of neurotransmitter regulation. For clarity, Table 1 below lists the mean changes in serotonin, dopamine, and norepinephrine concentrations before and after the treatment in 10 randomly selected patients.

[0039] Table 1. Statistical table of neurotransmitter changes before and after adaptive massage. Patient number Serotonin levels before massage (ng / mL) Serotonin levels (ng / mL) after massage Changes in dopamine levels (%) Norepinephrine changes (%) P001 105.2 120.7 +3.1 -9.8 P002 98.6 112.4 +4.7 -10.5 P003 110.1 125.0 +2.5 -12.1 P004 102.3 116.9 +4.9 -11.7 P005 108.7 124.5 +3.8 -8.6 P006 100.5 115.1 +2.6 -10.2 P007 106.9 121.3 +4.1 -11.0 P008 99.8 113.2 +3.6 -9.4 P009 107.6 123.5 +2.9 -10.1 P010 101.2 115.7 +3.3 -8.9

[0040] Table 1 above systematically records and compares the changes in key neurotransmitter levels before and after the intervention in 10 randomly selected subjects who received adaptive massage intervention. Comparing serotonin levels before and after massage reveals that serotonin concentrations significantly increased in all subjects, with an average increase of approximately 14.2 ng / mL. The increase was generally between 13% and 16%, consistent with the pre-set target range (10% to 30%), indicating that the designed massage sequence has a good physiological effect in increasing serotonin levels. Regarding dopamine changes, the data fluctuations were relatively small, with all subjects' changes remaining within ±5%, mostly showing small positive increases. For example, P002 and P004 increased by 4.7% and 4.9%, respectively, demonstrating the system's precise and stable control over dopamine concentrations. Meanwhile, the results of norepinephrine changes also showed that the concentration was generally reduced, with a decrease ranging from 8.6% to 12.1%, with an average decrease of about 10.2%. This fully verifies that the set stimulation parameters can effectively inhibit the excessive secretion of stress-related neurotransmitters and help reduce sympathetic nerve tension.

[0041] This embodiment verifies the significant effect of the method of the present invention in regulating neurotransmitter levels by performing adaptive abdominal acupoint massage intervention on subjects in a real clinical setting. Experimental results show that the stimulation parameter derivation mechanism and optimal massage action sequence generation method proposed in this invention can significantly increase serotonin concentration, maintain stable dopamine levels, and effectively decrease norepinephrine, achieving precise regulation of the target state of neurotransmitters. Simultaneously, this invention possesses good individual adaptability and a closed-loop self-updating mechanism, completing multi-parameter dynamic optimization without increasing manual intervention, significantly improving the intelligence level and physiological response consistency of massage intervention, and possessing broad practical application value and promising prospects for promotion.

[0042] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A neural network-based adaptive abdominal acupoint massage method, characterized in that, The method comprises the following steps: applying machine disturbance to the set acupoint area position on the abdomen, and collecting tension rebound signals of the corresponding acupoint area after each machine disturbance; performing two-dimensional compressed Hilbert envelope analysis on the tension rebound signals to extract disturbance echo features and generate a tension echo atlas of the corresponding acupoint area; obtaining an acupoint activity vector according to the tension echo atlas; constructing a multi-channel neurotransmitter prediction network according to a preset neurotransmitter target state, and reversely deriving a stimulation parameter set; matching the stimulation parameter set with the acupoint activity vector, performing sequence allocation optimization by using a probability uniformization allocation method based on diagonal guide sparsification constraint, and generating an optimal massage action sequence; periodically collecting electromyographic signals of the target acupoint area based on the optimal massage action sequence to obtain a prediction result of the current massage on the change of the concentration of neurotransmitters; when the prediction result deviates from the preset neurotransmitter target state by more than a preset tolerance threshold, dynamically replacing and updating the optimal massage action sequence to complete adaptive adjustment of massage.

2. The neural network-based abdominal acupoint adaptive massage method according to claim 1, characterized in that, The machine disturbance is periodic low-frequency mechanical vibration, the vibration frequency is 0.3 Hz to 0.9 Hz, the vibration action time is 0.5 s, no cross-acupoint area interference is generated when the machine disturbance is applied to each set acupoint area position, and the external contact pressure is kept constant before and after the disturbance.

3. The neural network-based abdominal acupoint adaptive massage method according to claim 1, characterized in that, The two-dimensional compressed Hilbert envelope analysis on the tension rebound signals to extract disturbance echo features and generate a tension echo atlas of the corresponding acupoint area is specifically: performing a band-pass filtering operation on the tension rebound signals collected from each acupoint area, the filtering frequency range is 0.1 Hz to 10 Hz, and high-frequency interference components and baseline drift terms are filtered out; applying Hilbert transform to the filtered tension rebound signals to generate a complex analytic signal, obtaining an instantaneous amplitude signal by calculating the modulus sequence of the complex analytic signal, and forming a time-domain envelope signal of the acupoint area; stacking the time-domain envelope signals of each acupoint area according to the acupoint arrangement order to form a two-dimensional data matrix as a spatio-temporal joint response representation of tension rebound; performing L1 norm normalization processing on each column of the two-dimensional data matrix in the spatial dimension; performing principal component analysis on the two-dimensional data matrix in the time dimension, extracting the first three principal components as the compression result, and if the cumulative variance contribution rate of the first three principal components is less than 85%, the fourth principal component is continuously extracted and added to the compression result; extracting the envelope peak value, rising edge slope and full-width half-maximum parameters of each acupoint in the compressed two-dimensional data matrix to construct the disturbance echo features of each acupoint; performing feature encoding on the disturbance echo features of each acupoint, and constructing a tension echo atlas according to the acupoint position index, the tension echo atlas is used to represent the dynamic organizational response ability of each acupoint to disturbance.

4. The neural network-based abdominal acupoint adaptive massage method according to claim 1, characterized in that, The acupoint activity vector is obtained according to the tension echo atlas, specifically: performing feature normalization processing on the disturbance echo features contained in each acupoint in the tension echo atlas, and using the Z-score standardization method to perform mean zero and variance unitization on the envelope peak value, rising edge slope and full-width half-maximum three indicators, respectively. The three indicators after normalization are fused by a weighted superposition calculation method, and the sum of the weights is 1, and the functional activity value of each acupoint area is calculated; According to the acupoint area arrangement index, the functional activity value of each acupoint area is constructed into an acupoint area activity vector.

5. The neural network-based abdominal acupoint adaptive massage method according to claim 1, characterized in that, According to the preset neurotransmitter target state, a multi-channel neurotransmitter prediction network is constructed, and a set of stimulation parameters is obtained by reverse derivation, specifically: The preset neurotransmitter target state includes an increase of 10% to 30% in serotonin concentration, a maintenance of dopamine concentration within an initial value ± 5%, and a decrease of 5% to 20% in norepinephrine concentration; The construction of the multi-channel neurotransmitter prediction network includes a multi-input structure, which receives input of stimulation frequency, force amplitude and stimulation duration, and simultaneously receives input of electromyographic characteristics for neurotransmitter state prediction in different stages; two hidden layers, the first hidden layer containing 32 nodes and the second hidden layer containing 16 nodes, each hidden layer being subjected to nonlinear transformation by a ReLU activation function; and an output layer containing 3 nodes corresponding to prediction values of changes in concentrations of serotonin, dopamine and norepinephrine; The loss function is defined as the weighted mean square error between the prediction values of changes in concentrations of each neurotransmitter and the corresponding target concentration change values; The gradient descent method is used to iteratively optimize the initial values of input stimulation frequency, force amplitude and stimulation duration, and the iteration is stopped when the set maximum number of iterations is reached; The stimulation frequency, force amplitude and stimulation duration that make the loss function converge are taken as the set of stimulation parameters obtained by reverse derivation and are output.

6. The neural network-based abdominal acupoint adaptive massage method according to claim 1, characterized in that, The matching of the set of stimulation parameters and the acupoint area activity vector is performed by a probability uniformization allocation method based on diagonal guidance and sparsification constraint for sequence allocation optimization to generate an optimal massage action sequence, specifically: Each set of stimulation parameters in the set of stimulation parameters is represented as a stimulation triple containing stimulation frequency, force amplitude and stimulation duration, and is numbered to form a set of stimulation parameter vectors; A matching probability matrix between a set of stimulation parameter vectors and acupoint areas is initialized, and the initial allocation probability of each stimulation triple to any acupoint area in the matching probability matrix is an equal value; An initial allocation trend dominated by a diagonal is constructed by diagonal guidance; In each round of distribution update, a sparsification constraint rule is introduced, which allows each set of stimulation parameters to be retained in no more than 2 acupoint areas with a matching probability higher than a set threshold, and the matching probabilities at other positions are compressed according to a decreasing rule for sparsification of the distribution result; The updated matching probability matrix is subjected to probability uniformization processing, and the row direction and the column direction are subjected to probability normalization respectively, so that the probability sum of each set of stimulation parameters allocated to each acupoint area is 1, and the total probability sum of all stimulation parameter combinations accepted by each acupoint area remains consistent; When the value change amplitudes of the corresponding positions of the matching probability matrix in two consecutive iterations are both less than a set convergence threshold, and the total normalization error of the row vector and the column vector is lower than a set error lower limit, it is considered that the matching probability matrix reaches a stable structure, the distribution optimization process is terminated, and the acupoint area number corresponding to the maximum probability of each set of stimulation parameters is extracted to form a one-to-one stimulation allocation relationship. Output an optimal massage action sequence in the order of the acupoint area in the stimulation allocation relationship, the optimal massage action sequence consisting of four items of acupoint area number, stimulation frequency, force amplitude and stimulation duration.

7. The neural network-based abdominal acupoint adaptive massage method according to claim 6, characterized in that, The diagonal dominant initial allocation trend is constructed by diagonal guidance, specifically: Mark the cells in the matching probability matrix located at the same or adjacent positions of the stimulation triplet number and the acupoint area number as the main diagonal area, and set the initial probability value in the main diagonal area as a high weight value, and set the remaining non-diagonal area as a low weight value, the high weight value being between 0.8 and 1.0, and the low weight value being between 0.0 and 0.2; The matching position with an absolute value of 0 of the number difference is set as the highest probability initial value, the matching position with a number difference of 1 is set as the second highest probability initial value, and the matching position with a number difference greater than or equal to 2 is set as the basic probability initial value, thereby constructing an initial allocation trend with the diagonal line as the center and the adjacent area decreasing.

8. The neural network-based abdominal acupoint adaptive massage method according to claim 1, characterized in that, Based on the optimal massage action sequence, periodically collect the electromyographic signals of the target acupoint area to obtain the prediction result of the change of neurotransmitter concentration under the current massage, specifically: According to the execution period of each massage action in the optimal massage action sequence, collect the electromyographic signals of the target acupoint area, set the sampling frequency to 500Hz to 1000Hz, and set the sampling time window to 1 second after each massage action ends; Perform band-pass filtering on the collected electromyographic signals, set the filtering frequency range to 20Hz to 450Hz, extract electromyographic features from the filtered electromyographic signals, and the electromyographic features include root mean square value, average absolute value, zero-crossing rate, waveform factor, power spectral density center frequency and median frequency; Input the extracted electromyographic features and the corresponding stimulation parameters into the multi-channel neurotransmitter prediction network to generate the concentration change prediction values of serotonin, dopamine and norepinephrine under the corresponding current massage state, and obtain the prediction result of the change of neurotransmitter concentration under the current massage. 9.The neural network-based adaptive abdominal acupoint massage method according to claim 1, wherein, The optimal massage action sequence is dynamically replaced and updated, specifically: Identify the acupoint area in the current optimal massage action sequence whose prediction result deviates from the preset neurotransmitter target state beyond the preset tolerance threshold, and mark it as an area to be updated; Retrieving other stimulation triplets in the stimulation parameter set that are not allocated or have low allocation probability for the area to be updated, and constructing a candidate massage action set; According to the current acupoint area functional activity value and the neurotransmitter prediction deviation direction, select a group of stimulation parameters from the candidate massage action set, so that the concentration change value in the multi-channel neurotransmitter prediction network is corrected in the target direction; Replace the new stimulation parameters in the corresponding position of the current optimal massage action sequence, and keep the original stimulation parameters of the remaining acupoint areas unchanged to complete the local update of the massage action sequence.

Images