Animal experiment acupoint massage parameter self-correction method combined with deep learning

Abstract

The invention discloses an animal experiment acupuncture point massage parameter self-correction method combined with deep learning, and relates to the technical field of animal intervention experiments. The specific operation of the animal experiment acupoint massage parameter self-correction method combined with deep learning comprises the steps of simulating human hand tactile characteristic device design, deep learning model construction and training and massage parameter self-correction operation flow, and the human hand tactile characteristic device design comprises a tactile sensing module and a force feedback execution module. The step of deep learning model construction and training comprises data acquisition and labeling, model structure design and model training and optimization, and the step of massage parameter self-correction operation comprises an experiment preparation stage, an experiment proceeding stage and an experiment ending stage. Spatial feature data, including pressure distribution and strain images, collected by a multi-mode sensor can be effectively processed.

Description

Technical Field

[0001] The present invention relates to the technical field of animal intervention experiments, and specifically to a method for self-correcting acupoint massage parameters in animal experiments combined with deep learning. Background Art

[0002] In animal experiments, acupoint massage is a common intervention method. The accuracy and stability of its parameters have a key impact on the reliability of experimental results. Traditional acupoint massage operations mainly rely on manual labor, which has problems such as inconsistent techniques and difficulty in accurately controlling strength and frequency. Some existing mechanical massage devices cannot accurately simulate the tactile characteristics of human hands, resulting in deviations between the physiological responses of animals and their natural state when receiving massage stimulation, which in turn affects the accuracy of experimental data and the repeatability of experimental results. Existing equipment such as the XYZ-2000 massage instrument has a pressure control error range of ±15%, which cannot simulate the dynamic feedback of human hands, resulting in a deviation rate of animal stress response as high as 30%. To this end, we propose a self-correction method for acupoint massage parameters in animal experiments combined with deep learning. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for self-correction of acupoint massage parameters in animal experiments combined with deep learning.

[0004] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a method for self-correction of acupoint massage parameters in animal experiments combined with deep learning, including a method for self-correction of acupoint massage parameters in animal experiments combined with deep learning. The specific operations of the method for self-correction of acupoint massage parameters in animal experiments combined with deep learning include the design of a device for simulating the tactile characteristics of human hands, the construction and training of a deep learning model, and a self-correction operation process of massage parameters. The design of the device for simulating the tactile characteristics of human hands includes a tactile perception module and a force feedback execution module. The construction and training of the deep learning model include data acquisition and labeling, model structure design, and model training and optimization. The labeling is completed by three independent experimenters, and the labeling consistency is verified by the Kappa coefficient (K≥0.85). The labeling tool adopts LabelStudio 3.0, which supports multi-modal data synchronous labeling. The self-correction operation process of massage parameters includes an experimental preparation stage, an experimental stage, and an experimental end stage.

[0005] As a further solution of the present invention: the tactile perception module is to install a multimodal sensor integrated device on the massage head of the massage device. The multimodal sensor integrated device includes a pressure sensor, a strain sensor, a vibration sensor and a temperature sensor. The pressure sensor is used to accurately measure the pressure applied to the animal's acupoints during massage. The strain sensor is used to sense the deformation of the tissue at the acupoints during massage, and feedback the elastic characteristics of the tissue based on the deformation. The vibration sensor is used to capture the subtle vibration signals generated by human hand massage, simulating the rhythm and force changes of human hand massage. The temperature sensor is used to monitor the temperature changes of the skin at the acupoints in real time. The force feedback execution module is designed based on the data collected by the tactile perception module. The force feedback execution device adopts an electromagnetic drive method and quickly adjusts the massage force and direction of the massage head according to the received control signal. When the tactile perception module detects an increase in the hardness of the tissue at the acupoint, the force feedback actuator automatically increases the massage force to maintain a stimulation effect similar to that of human hand massage. When the animal is detected to have an avoidance reaction, the massage direction and force are adjusted to simulate the dynamic adjustment of the human hand based on the animal feedback.

[0006] As a further solution of the present invention: the data set collection and annotation operation specifically involves conducting acupoint massage experiments on experimental animals of different species and different health statuses, and collecting massage data, including the strength, frequency, and duration of the massage. At the same time, multimodal sensor data simulating the tactile characteristics of human hands and physiological feedback data of animals, including heart rate variability, blood fluctuations, and neural electrical activity in specific brain areas, are collected. The collected data are annotated in detail, including the species of the animal, health status, experimental purpose, corresponding massage parameters, and expected physiological response information.

[0007] As a further solution of the present invention: the model structure is designed to construct a deep learning model that combines a convolutional neural network and a recurrent neural network. The convolutional neural network part is used to process the spatial feature data collected by the multimodal sensor, including pressure distribution and strain images, and extract local features therein. The recurrent neural network part is used to process the time series data during the massage process, including the changes in massage force over time and the dynamic changes in animal physiological indicators, and capture the time dependence and trend of the data.

[0008] As a further solution of the present invention, the convolutional neural network part constructs 3-5 convolution layers, with the input pressure distribution data as the main body. The first convolution layer adopts a 3×3 convolution kernel, a step size of 1, and a padding of 1. The number of input channels is determined according to the array structure of the pressure sensor, and the output channels are set to 16. The convolution operation formula is as follows: in, It is The convolutional layer is at position ( ), It is The convolution kernel weights of the layer, It is Layer at position ( ), It is The bias of the layer, and is the size of the convolution kernel; The pooling layer after the convolutional layer uses maximum pooling with a pooling kernel size of 2×2 and a stride of 2. The maximum pooling operation can be expressed as: Flatten the feature map and input it into the fully connected layer; The recurrent neural network uses the long short-term memory network as the specific implementation of the recurrent neural network. The kneading intensity changes according to time and the time series data are input into the LSTM unit according to the time step. LSTM is the long short-term memory network. In the LSTM unit, the input gate , Forget Gate , output gate and memory cells The calculation formula is as follows: in, is the sigmoid activation function, is the hyperbolic tangent activation function, represents the weight matrix, represents the bias vector, represents element-wise multiplication, is the input of the current time step, is the hidden state at the previous time step, It is the memory cell state of the previous time step. The flattened features are concatenated with the LSTM hidden state and input into the fully connected layer for feature fusion and prediction.

[0009] As a further solution of the present invention: the model training and optimization specifically includes data set division, loss function definition, training process, regularization and prevention of overfitting. The data set division is to divide the labeled large-scale data set into 70% training set, 20% validation set and 10% test set. The training set is used to update the model parameters, the validation set is used to evaluate the performance of the model during the training process, adjust the hyperparameters, including the learning rate and the regularization coefficient, and the test set is used to evaluate the generalization ability of the model. The loss function definition specifically selects the mean square error as the loss function to measure the difference between the animal physiological response predicted by the model and the actual observed physiological response. The mean square error loss function is: in, is the physiological response predicted by the model, is the actual observed value, is the sample size; The training process includes initializing the model parameters, inputting the training set data into the model in batches of 32 samples, and in each training batch, forward propagating the prediction structure of the model. The loss between the prediction result and the actual observation value is calculated according to the loss function. The model parameters are updated according to the loss using the backpropagation algorithm, mainly using the Adagrad adaptive learning rate adjustment algorithm. The parameter update formula is: in, and are the updated and current model parameters, respectively, is the initial learning rate, is the time step The sum of squared gradients up to , is a constant to prevent division by zero, In the current parameter Loss function During training, the performance of the model is evaluated on the validation set each time, including calculating the loss value and accuracy on the validation set. Training is stopped when the performance on the validation set stops improving. The regularization and prevention of overfitting introduce L1 and L2 regularization techniques, adding regularization terms to the loss function. The L1 regularization term is , the L2 regularization term is ,in and is the regularization coefficient, is the model parameter, and the modified loss function is: Adjust the regularization coefficient to balance the model's fitting ability and generalization ability.

[0010] As a further solution of the present invention, the specific operations of the experimental preparation stage in the massage parameter self-correction operation process are as follows: The massage device that simulates the tactile sensation of human hands is connected and calibrated with the animal experimental platform with high precision. The massage head of the massage device is accurately brought into contact with the animal acupoints. The initial massage parameter range is extracted from historical experimental data and relevant literature. Combined with the characteristics of the animal model, a deep learning model is used for preliminary evaluation to determine the initial massage parameter values ​​and set the expected animal physiological response targets, including the range of changes in specific physiological indicators and the degree of improvement in behavioral learning performance. The deep learning model is preheated, the pre-trained model parameters are loaded, and it is integrated with the real-time data acquisition system and control module of the experimental equipment. The model is used to receive and process experiments in real time.

[0011] As a further solution of the present invention, the specific operations in the experiment phase of the massage parameter self-correction operation process are as follows: Start the massage device and start massaging the animal's acupoints according to the set initial parameters to simulate the human hand's touch. During the massage process, multimodal sensors are used to collect real-time massage force, frequency, mechanical and temperature characteristics data of the tissue at the acupoints, and physiological signal data of the animal, and transmit the data to the deep learning model. After receiving the real-time data, the deep learning model preprocesses the data, including data cleaning, normalization, and feature extraction operations. The processed data is input into the model for prediction to obtain the expected physiological response of the animal under the current massage parameters. The model compares and analyzes the predicted results with the set expected targets and calculates the difference between the two. During the massage process, the deep learning model continuously monitors the experimental data and massage effects, dynamically adjusts the massage parameters, and records the process and results of each parameter adjustment.

[0012] As a further solution of the present invention, the specific operations at the end of the experiment in the massage parameter self-correction operation process are as follows: After the experiment, all data generated in the experiment will be comprehensively organized and archived, including detailed parameter changes during the massage process, full data collected by multimodal sensors, complete physiological response data of animals, and operation records of the deep learning model. The newly generated experimental data and historical data will be merged, and the deep learning model will be incrementally trained and optimized. The model will be retrained to learn the characteristics and patterns of the experimental data. At the same time, the massage equipment will be inspected and maintained every week, and the hardware and software of the equipment will be adjusted and upgraded according to the operation status during the experiment.

[0013] By adopting the above technical solution, compared with the prior art, the beneficial effects of the present invention are: 1. By using convolutional neural networks in a deep learning model, the present invention can effectively process spatial feature data collected by multimodal sensors, including pressure distribution and strain images. Through multi-layer convolution and pooling operations, the convolutional neural network can extract local features in the data and identify the patterns of spatial features such as pressure and strain during massage of different acupoints. At the same time, it reduces the data dimension and the amount of calculation. In addition, when analyzing the pressure distribution images at massage points, the convolutional neural network can accurately extract key features such as pressure concentration areas and pressure change trends, providing an important basis for subsequent judgment of the massage effect, helping to accurately simulate the stimulation of acupoints by human tactile sensation in the spatial dimension, and improving the accuracy of experimental data. 2. By using a recurrent neural network in a deep learning model, the present invention excels at processing time series data during the massage process, including changes in massage intensity over time and dynamic changes in animal physiological indicators. The long-short-term memory network can effectively capture the time dependency and trend of the data. During the massage process, the recurrent neural network can analyze the changing patterns of massage intensity over a period of time and the dynamic changes in animal physiological indicators over time, thereby better understanding the temporal relationship between massage parameters and animal physiological responses, providing support for simulating dynamic changes during continuous massage by human hands, and improving the degree to which experimental data reflects the actual massage process. 3. The present invention uses the pressure, strain, vibration and temperature sensors in the tactile perception module in the deep learning model to work together to simulate the full range of tactile sensations of acupoints when a person presses and kneads the hands. The pressure sensor accurately measures the pressure of the massage, the vibration sensor restores the rhythm and force changes of the massage, and the temperature sensor monitors the temperature changes caused by friction. The data collected by these sensors provides rich information for simulating the real human hand touch, making the massage stimulation received by animals more in line with the natural state, thereby improving the accuracy of the experimental data. In the mouse experiment, it can accurately simulate the various tactile sensations when the human hand presses and kneads the mouse acupoints, obtain more realistic experimental data, and the deep learning model During the training process, the model learns the relationship between massage parameters and animal physiological responses from a large number of animal experimental data of different species and health conditions. The trained model is universal. When new experimental data is input, the massage parameters can be evaluated and adjusted based on the learned characteristic patterns and time association models. When different experimenters operate, the model can ensure that the massage parameters meet the standards, reduce the deviation of experimental results caused by personnel differences, and improve the repeatability of experimental results. In the rat nervous system regulation experiment, no matter which experimenter operates, the model can give consistent massage parameter guidance based on the training results, ensuring the stability and repeatability of the experimental results. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a flow chart of the method for self-correction of acupoint massage parameters in animal experiments combined with deep learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The specific embodiments of the present invention will be further described below in conjunction with the accompanying drawings. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention.

[0016] In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0017] Please see the attached Figure 1 The present invention provides a method for self-correction of acupoint massage parameters in animal experiments combined with deep learning, including the method for self-correction of acupoint massage parameters in animal experiments. The specific operations of the method for self-correction of acupoint massage parameters in animal experiments combined with deep learning include designing a device for simulating tactile characteristics of human hands, building and training a deep learning model, and self-correction operation process of massage parameters. The design of the device for simulating tactile characteristics of human hands includes a tactile perception module and a force feedback execution module. The deep learning model building and training include data collection and labeling, model structure design, and model training and optimization. The labeling is completed by three independent experimenters, and the labeling consistency is verified by the Kappa coefficient (K≥0.85). The labeling tool adopts LabelStudio3.0, which supports multi-modal data synchronous labeling. The self-correction operation process of massage parameters includes an experimental preparation stage, an experimental stage, and an experimental end stage.

[0018] In one embodiment of the present invention, the tactile perception module is a multimodal sensor integrated device installed on the massage head of the massage device. The multimodal sensor integrated device includes a pressure sensor, a strain sensor, a vibration sensor, and a temperature sensor. The pressure sensor is used to accurately measure the pressure applied to the animal's acupuncture points during massage. The strain sensor is used to sense the deformation of the tissue at the acupuncture points during massage and provide feedback on the elastic characteristics of the tissue based on the deformation. The vibration sensor is used to capture the subtle vibration signals generated by human hand massage to simulate the rhythm and force changes of human hand massage. The temperature sensor is used to monitor the temperature changes of the skin at the acupuncture points in real time. The force feedback execution module is a force feedback execution device designed based on the data collected by the tactile perception module. The electromagnetic drive method is used to quickly adjust the massage force and direction of the massage head according to the received control signal. When the tactile perception module detects an increase in the hardness of the tissue at the acupuncture point, the force feedback actuator automatically increases the massage force to maintain a stimulation effect similar to that of human hand massage. When the animal is detected to have an avoidance reaction, the massage direction and force are adjusted to simulate the dynamic adjustment of the human hand based on the animal's feedback. The pressure sensor is Honeywell FMA series, with a range of 0-10N and an accuracy of ±0.05N; the vibration sensor is PCB Piezotronics 352C03, with a frequency response range of 0.5-5000Hz; The strain sensor used is OMEGA SGD-3 / 350-LY13, with a sensitivity of 2.0mV / V; The vibration sensor uses PCB 352C03, with a frequency response of 0.5-5000Hz; The temperature sensor is TI TMP117 with an accuracy of ±0.1°C.

[0019] The force feedback actuator module uses an electromagnetic driver (Faulhaber 3242G024CR) with a response time of <10ms.

[0020] In one embodiment of the present invention: the data set collection and annotation operation specifically involves conducting acupoint massage experiments on experimental animals of different species and different health statuses, and collecting massage data, including the strength, frequency, and duration of the massage. At the same time, multimodal sensor data simulating the tactile characteristics of human hands and physiological feedback data of animals, including heart rate variability, blood fluctuations, and neural electrical activity in specific brain areas, are collected. The collected data are annotated in detail, including the species of animals, health status, experimental purpose, corresponding massage parameters, and expected physiological response information.

[0021] In one embodiment of the present invention: the model structure is designed to construct a deep learning model that combines a convolutional neural network and a recurrent neural network. The convolutional neural network part is used to process the spatial feature data collected by the multimodal sensor, including pressure distribution and strain images, and extract local features therein. The recurrent neural network part is used to process the time series data during the massage process, including the change of massage force over time and the dynamic changes of animal physiological indicators, and capture the time dependence and trend of the data.

[0022] In one embodiment of the present invention, the convolutional neural network part constructs 3-5 convolution layers, with the input pressure distribution data as the main body. The first convolution layer uses a 3×3 convolution kernel, a step size of 1, and a padding of 1. The number of input channels is determined according to the array structure of the pressure sensor, and the output channels are set to 16. The convolution operation formula is as follows: in, It is The convolutional layer is at position ( ), It is The convolution kernel weights of the layer, It is Layer at position ( ), It is The bias of the layer, and is the size of the convolution kernel; The pooling layer after the convolutional layer uses maximum pooling with a pooling kernel size of 2×2 and a stride of 2. The maximum pooling operation can be expressed as: Flatten the feature map and input it into the fully connected layer; The recurrent neural network uses the long short-term memory network as the specific implementation of the recurrent neural network. The kneading intensity changes according to time and the time series data are input into the LSTM unit according to the time step. LSTM is the long short-term memory network. In the LSTM unit, the input gate , Forget Gate , output gate and memory cells The calculation formula is as follows: in, is the sigmoid activation function, is the hyperbolic tangent activation function, represents the weight matrix, represents the bias vector, represents element-wise multiplication, is the input of the current time step, is the hidden state at the previous time step, It is the memory cell state of the previous time step. The flattened features are concatenated with the LSTM hidden state and input into the fully connected layer for feature fusion and prediction.

[0023] In one embodiment of the present invention, model training and optimization specifically include data set division, loss function definition, training process, regularization, and prevention of overfitting. Data set division is to divide the labeled large-scale data set into 70% training set, 20% validation set, and 10% test set. The training set is used to update the model parameters. The validation set is used to evaluate the performance of the model during the training process and adjust the hyperparameters, including the learning rate and the regularization coefficient. The test set is used to evaluate the generalization ability of the model. The loss function definition specifically selects the mean square error as the loss function to measure the difference between the animal physiological response predicted by the model and the actual observed physiological response. The mean square error loss function is: in, is the physiological response predicted by the model, is the actual observed value, is the sample size; The training process includes initializing the model parameters, inputting the training set data into the model in batches of 32 samples, and in each training batch, forward propagating the model's prediction structure. The loss between the prediction result and the actual observation value is calculated according to the loss function. The model parameters are updated according to the loss using the backpropagation algorithm, mainly using the Adagrad adaptive learning rate adjustment algorithm. The parameter update formula is: in, and are the updated and current model parameters, respectively, is the initial learning rate, is the time step The sum of squared gradients up to , is a constant to prevent division by zero, In the current parameter Loss function During training, the performance of the model is evaluated on the validation set each time, including calculating the loss value and accuracy on the validation set. Training is stopped when the performance on the validation set stops improving. Regularization and prevention of overfitting introduce L1 and L2 regularization techniques, add regularization terms to the loss function, and the L1 regularization term is , the L2 regularization term is ,in and is the regularization coefficient, is the model parameter, and the modified loss function is: Adjust the regularization coefficient to balance the model's fitting ability and generalization ability; During the training process, the initial learning rate was selected in {0.001, 0.01, 0.1} through grid search and was finally set to 0.01. The regularization coefficient was determined by 5-fold cross validation as =0.01, =0.001, and the early stopping condition for training is that the validation set loss does not decrease for 10 consecutive epochs.

[0024] In one embodiment of the present invention, the specific operations of the experimental preparation stage in the massage parameter self-calibration operation process are as follows: The massage device that simulates the tactile sensation of human hands is connected and calibrated with the animal experimental platform with high precision. The massage head of the massage device is accurately brought into contact with the animal acupoints. The initial massage parameter range is extracted from historical experimental data and relevant literature. Combined with the characteristics of the animal model, a deep learning model is used for preliminary evaluation to determine the initial massage parameter values ​​and set the expected animal physiological response targets, including the range of changes in specific physiological indicators and the degree of improvement in behavioral learning performance. The deep learning model is preheated, the pre-trained model parameters are loaded, and it is integrated with the real-time data acquisition system and control module of the experimental equipment. The model is used to receive and process experiments in real time.

[0025] As a further solution of the present invention, the specific operations in the experiment phase of the massage parameter self-correction operation process are as follows: Start the massage device and start massaging the animal's acupoints to simulate the human hand's touch according to the set initial parameters. During the massage process, use multimodal sensors to collect massage force, frequency, mechanical and temperature characteristics of tissue at the acupoints, and physiological signal data of the animal in real time, and transmit the data to the deep learning model. After receiving the real-time data, the deep learning model preprocesses the data, including data cleaning, normalization, and feature extraction operations. The processed data is input into the model for prediction to obtain the expected physiological response of the animal under the current massage parameters. The model compares and analyzes the predicted results with the set expected targets and calculates the difference between the two. If the mean square error (MSE) between the prediction and the actual exceeds the threshold of 0.1, the model calls the Bayesian optimization algorithm to calculate the massage force adjustment amount. , the constraints are | |≤0.2N / s, ensuring a smooth transition of stimulation. During the massage process, the deep learning model continuously monitors experimental data and massage effects, dynamically adjusts the massage parameters, and the model also records the process and results of each parameter adjustment.

[0026] In one embodiment of the present invention, the specific operations at the end of the experiment in the massage parameter self-calibration process are as follows: After the experiment, all data generated in the experiment will be comprehensively organized and archived, including detailed parameter changes during the massage process, full data collected by multimodal sensors, complete physiological response data of animals, and operation records of the deep learning model. The newly generated experimental data and historical data will be merged, and the deep learning model will be incrementally trained and optimized. The model will be retrained to learn the characteristics and patterns of the experimental data. At the same time, the massage equipment will be inspected and maintained every week, and the hardware and software of the equipment will be adjusted and upgraded according to the operation status during the experiment.

[0027] In one embodiment of the present invention: in regularization and prevention of overfitting, the fitting ability and generalization ability of the model are balanced by adjusting the regularization coefficient to prevent the model from overfitting. At the same time, random inactivation technology can also be used to randomly set the output of some neurons to 0 during the training process to further reduce the co-adaptability between neurons and reduce the risk of overfitting.

[0028] In one embodiment of the present invention: during the experiment stage in the massage parameter self-correction operation process, if the difference exceeds a preset threshold, the deep learning model immediately starts the parameter self-correction mechanism. The model calculates the adjusted massage parameters through an optimization algorithm based on the relationship between the massage parameters and the animal's physiological responses learned in advance, adjusts the increment or decrement of the massage intensity, changes the step size of the massage frequency, adjusts the massage rhythm pattern, etc. The model sends the adjustment instruction to the control module of the massage device, and the control module quickly adjusts the operating parameters of the massage device according to the instruction, thereby realizing real-time self-correction of the massage parameters.

[0029] Example Data collection and annotation operations: Experimental Design and Implementation: Acupoint massage experiments were designed for a variety of different species of experimental animals, including mice, rats, and rabbits, as well as those in normal condition and with various health conditions, such as those suffering from diabetes, arthritis, and neurological diseases. A stable and adaptable experimental platform was constructed to secure the animals to prevent them from moving during massage and reduce interference. Using a massage device that simulates the tactile properties of the human hand, massage force was set between 0.1N and 5N, frequency between 0.5 and 5 times per second, and duration between 5 and 30 minutes, depending on the animal species and experimental requirements. Massage was performed on specific acupoints on the animals. Simultaneously, a multimodal sensor integration device was activated. Based on the data variation characteristics and experimental accuracy requirements, the sampling rates of the pressure sensor were set to 100Hz, the strain sensor to 50Hz, the vibration sensor to 200Hz, and the temperature sensor to 20Hz. Pressure, strain, vibration, and temperature data were collected during the massage process. Heart rate variability data was collected during the massage process using implantable or surface electrodes. Blood pressure fluctuations were measured using a non-invasive blood pressure monitor. Neural electrical activity in specific brain regions was recorded using an electroencephalogram (EEG). Data annotation: Comprehensively annotate each set of collected data, clearly labeling the animal species. Mice should be labeled with "Musmusculusn." The health status should include information on whether the patient is healthy or not, the disease model, and the modeling method and time. The purpose of the experiment is to explore the effects of acupoint massage on pain relief or nervous system regulation. The massage intensity should be accurately labeled to 0.01N, the frequency to 0.01 times / second, and the duration to the second. Expected physiological responses should be clearly defined, including the range of changes in heart rate variability, expected blood pressure changes, and the target changes in the frequency and amplitude of neural electrical activity in specific brain regions. Model structure construction operation, Convolutional neural network (CNN) part: Construct 3-5 convolution layers, mainly based on the input pressure distribution data. The first convolution layer uses a 3x3 convolution kernel, a stride of 1, and a padding of 1. The number of input channels is determined by the pressure sensor array structure, and the number of input channels is set to 1 and the number of output channels is set to 16. Through convolution operations, local features in the data are extracted. Then, the pooling layer is connected, using a 2x2 pooling kernel and a maximum pooling method with a stride of 2 to reduce the data dimension and the amount of calculation. Finally, the feature maps output by the convolution layer and the pooling layer are flattened for input into the subsequent fully connected layer. Recurrent Neural Network (RNN): A long short-term memory (LSTM) network is used as the specific implementation of the RNN. Time series data, such as changes in massage intensity over time, is input into the LSTM unit at a time step per second. The LSTM unit captures the temporal dependencies and trends of the data through a series of operations, including input gates, forget gates, output gates, and memory cell operations. The flattened features output by the CNN are concatenated with the hidden state output by the LSTM and then input into the fully connected layer for feature fusion and prediction. The 128-dimensional feature vector extracted by CNN is concatenated with the 64-dimensional hidden state of LSTM and input into the fully connected layer.

[0030] Model training and optimization operations, Dataset division: The labeled large-scale dataset is divided into 70% training set, 20% validation set, and 10% test set. The training set is used to update model parameters, the validation set is used to evaluate model performance during training and adjust hyperparameters such as learning rate and regularization coefficient, and the test set is used to evaluate the generalization ability of the model. Loss function definition: The mean squared error (MSE) is selected as the loss function to measure the difference between the animal's physiological response predicted by the model and the actual observed physiological response. During the calculation process, the model prediction value and the actual observed value are substituted into the corresponding calculation formula to obtain the loss value; Training process: Initialize the model parameters and input the training set data into the model in batches of 32 samples. In each training batch, first perform forward propagation to calculate the prediction results. Then, use the loss function to calculate the loss between the prediction results and the actual observations. Use the backpropagation algorithm to update the model parameters based on the loss. The Adagrad adaptive learning rate adjustment algorithm is mainly used to adjust the model parameters based on factors such as the pre-set initial learning rate and the sum of squared gradients to the current time step. During the training process, the model performance is evaluated on the validation set each time, and indicators such as loss value and accuracy are calculated. When the performance on the validation set does not improve for 5-10 consecutive epochs, stop training; Regularization and prevention of overfitting: L1 and L2 regularization techniques are introduced, and corresponding regularization terms are added to the loss function. By adjusting the regularization coefficient, the model's fitting ability and generalization ability are balanced to prevent overfitting. At the same time, random dropout technology can be used to randomly set the output of some neurons to 0 during training, reducing the co-adaptability between neurons and reducing the risk of overfitting. Operation in the self-calibration process of acupoint massage parameters in animal experiments, Experimental preparation phase: The massage device that simulates human hand tactile sensation is connected and calibrated with the animal experimental platform with high precision to ensure that the massage head accurately contacts the animal's acupuncture points. The initial massage parameter range is extracted from historical experimental data and relevant literature. Combined with the purpose of this experiment and the characteristics of the animal model, it is input into the deep learning model for pre-evaluation to determine more reasonable initial massage parameter values. At the same time, the expected animal physiological response targets are set, including the range of changes in specific physiological indicators and the degree of improvement in behavioral performance. The deep learning model is preheated, pre-trained parameters are loaded, and integrated with the real-time data acquisition system and control module of the experimental equipment to enable the model to receive and process experimental data in real time. During the experiment: the experimenter starts the massage device, and the device simulates the human hand's touch to massage the animal's acupoints according to the initial parameters. The multimodal sensor collects the massage force, frequency, tissue mechanics and temperature characteristics data at the acupoints, and animal physiological signal data in real time, and transmits them to the deep learning model at high speed. After receiving the data, the model first cleans the data, removes outliers, and then normalizes the data to the range of 0-1. It then extracts key features such as the pressure distribution peak and the strain change slope. The processed data is input into the model to predict the expected physiological response of the animal under the current massage parameters, and the difference is compared with the set target. If the difference exceeds the preset threshold, the model Based on the pre-learned relationship between massage parameters and animal physiological responses, the model calculates the adjusted massage parameters through optimization algorithms such as gradient descent, adjusts the massage intensity increment or decrement, changes the massage frequency step size, and adjusts the massage rhythm pattern. The model sends the adjustment instructions to the massage device control module, which quickly adjusts the device operating parameters to achieve real-time self-calibration of the massage parameters. During the massage process, the model continuously monitors experimental data and massage effects, dynamically adjusts the massage parameters, and ensures that the animal's physiological response develops in the expected direction. At the same time, the model records the process and results of each parameter adjustment to provide data support for subsequent experimental analysis and model optimization. End of experiment stage: After the experiment is over, all data generated by this experiment will be comprehensively organized and archived, including detailed parameter changes during the massage process, full data from multimodal sensors, complete physiological response data of animals, and deep learning model operation records. The new data will be merged with historical data, and the deep learning model will be incrementally trained and optimized. The model will be retrained to enable it to learn more experimental data characteristics and patterns, further improving the prediction accuracy and parameter self-correction capabilities. The massage equipment will be inspected and maintained, and the equipment hardware and software will be adjusted and upgraded as necessary according to the experimental operation status to ensure stable and accurate operation of the equipment in the next experiment.

[0031] By using deep learning models to achieve relevant operations, in the animal experiment process, from determining the initial parameters and preheating the model in the experiment preparation stage, to real-time data collection, model prediction and self-correction of parameters in the experiment stage, and then to data collation, incremental training model and equipment maintenance at the end of the experiment, a complete and efficient self-correction system for animal experiment acupoint massage parameters has been built.

[0032] Although the present invention is disclosed above with reference to preferred embodiments, this is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent variations, and modifications made to the above embodiments in accordance with the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the scope of protection defined by the claims of the present invention.

Claims

1. A method for self-calibration of acupoint massage parameters in animal experiments combined with deep learning, including a method for self-calibration of acupoint massage parameters in animal experiments, characterized by: The specific operations of the animal experiment acupoint massage parameter self-correction method combined with deep learning include the design of a device simulating the tactile characteristics of human hands, the construction and training of a deep learning model, and the self-correction operation process of massage parameters. The design of the device simulating the tactile characteristics of human hands includes a tactile perception module and a force feedback execution module. The deep learning model construction and training include data collection and labeling, model structure design, and model training and optimization. The labeling is completed by three independent experimenters, and the labeling consistency is verified by the Kappa coefficient. The labeling tool uses LabelStudio3.0, which supports multi-modal data synchronous labeling. The self-correction operation process of massage parameters includes the experimental preparation stage, the experimental stage, and the experimental end stage.

2. The method for self-calibration of acupoint massage parameters in animal experiments combined with deep learning according to claim 1, characterized in that: The tactile perception module is a multimodal sensor integrated device installed on the massage head of the massage device. The multimodal sensor integrated device includes a pressure sensor, a strain sensor, a vibration sensor and a temperature sensor. The pressure sensor is used to accurately measure the pressure applied to the animal's acupoints during massage. The strain sensor is used to sense the deformation of the tissue at the acupoint during massage and feedback the elastic characteristics of the tissue based on the deformation. The vibration sensor is used to capture the subtle vibration signals generated by human hand massage to simulate the rhythm and force changes of human hand massage. The temperature sensor is used to monitor the temperature changes of the skin at the acupoints in real time. The force feedback execution module is designed based on the data collected by the tactile perception module. The force feedback execution device adopts an electromagnetic drive method and quickly adjusts the massage force and direction of the massage head according to the received control signal. When the tactile perception module detects an increase in the hardness of the tissue at the acupoint, the force feedback actuator automatically increases the massage force to maintain a stimulation effect similar to that of human hand massage. When the animal is detected to have an avoidance reaction, the massage direction and force are adjusted to simulate the dynamic adjustment of the human hand based on the animal feedback.

3. The method for self-calibration of acupoint massage parameters in animal experiments combined with deep learning according to claim 2, characterized in that: The data set collection and annotation operation specifically involves conducting acupoint massage experiments on experimental animals of different species and different health statuses, and collecting massage data, including the strength, frequency, and duration of the massage. At the same time, multimodal sensor data that simulates the tactile characteristics of human hands and physiological feedback data of animals, including heart rate variability, blood fluctuations, and neural electrical activity in specific brain areas, are collected. The collected data are detailedly annotated, including the animal's species, health status, experimental purpose, corresponding massage parameters, and expected physiological response information.

4. The method for self-calibration of acupoint massage parameters in animal experiments combined with deep learning according to claim 3, characterized in that: The model structure is designed to construct a deep learning model that combines a convolutional neural network and a recurrent neural network. The convolutional neural network part is used to process the spatial feature data collected by multimodal sensors, including pressure distribution and strain images, and extract local features therein. The recurrent neural network part is used to process the time series data during the massage process, including the changes in massage intensity over time and the dynamic changes in animal physiological indicators, and capture the time dependence and trend of the data.

5. The method for self-calibration of acupoint massage parameters in animal experiments combined with deep learning according to claim 4, characterized in that: The convolutional neural network part constructs 3-5 convolution layers, with the input pressure distribution data as the main body. The first convolution layer uses a 3×3 convolution kernel, a step size of 1, and a padding of 1. The number of input channels is determined according to the array structure of the pressure sensor, and the output channel is set to 16. The convolution operation formula is as follows: in, It is The convolutional layer is at position ( ), It is The convolution kernel weights of the layer, It is Layer at position ( ), It is The bias of the layer, and is the size of the convolution kernel; The pooling layer after the convolutional layer uses maximum pooling with a pooling kernel size of 2×2 and a stride of 2. The maximum pooling operation can be expressed as: Flatten the feature map and input it into the fully connected layer; The recurrent neural network uses the long short-term memory network as the specific implementation of the recurrent neural network. The kneading intensity changes according to time and the time series data are input into the LSTM unit according to the time step. LSTM is the long short-term memory network. In the LSTM unit, the input gate , Forget Gate , output gate and memory cells The calculation formula is as follows: in, is the sigmoid activation function, is the hyperbolic tangent activation function, represents the weight matrix, represents the bias vector, represents element-wise multiplication, is the input of the current time step, is the hidden state at the previous time step, It is the memory cell state of the previous time step. The flattened features are concatenated with the LSTM hidden state and input into the fully connected layer for feature fusion and prediction.

6. The method for self-calibration of acupoint massage parameters in animal experiments combined with deep learning according to claim 5, characterized in that: The model training and optimization specifically includes data set division, loss function definition, training process, regularization and prevention of overfitting. The data set division is to divide the labeled large-scale data set into 70% training set, 20% validation set and 10% test set. The training set is used to update the model parameters, the validation set is used to evaluate the performance of the model during the training process, adjust the hyperparameters, including the learning rate and the regularization coefficient, and the test set is used to evaluate the generalization ability of the model. The loss function definition specifically selects the mean square error as the loss function to measure the difference between the animal physiological response predicted by the model and the actual observed physiological response. The mean square error loss function is: in, is the physiological response predicted by the model, is the actual observed value, is the sample size; The training process includes initializing the model parameters, inputting the training set data into the model in batches of 32 samples, and in each training batch, forward propagating the prediction structure of the model. The loss between the prediction result and the actual observation value is calculated according to the loss function. The model parameters are updated according to the loss using the backpropagation algorithm, mainly using the Adagrad adaptive learning rate adjustment algorithm. The parameter update formula is: in, and are the updated and current model parameters, respectively, is the initial learning rate, is the time step The sum of squared gradients up to , is a constant to prevent division by zero, In the current parameter Loss function During training, the performance of the model is evaluated on the validation set each time, including calculating the loss value and accuracy on the validation set. Training is stopped when the performance on the validation set stops improving. The regularization and prevention of overfitting introduce L1 and L2 regularization techniques, adding regularization terms to the loss function. The L1 regularization term is , the L2 regularization term is ,in and is the regularization coefficient, is the model parameter, and the modified loss function is: Adjust the regularization coefficient to balance the model's fitting ability and generalization ability.

7. The method for self-calibration of acupoint massage parameters in animal experiments combined with deep learning according to claim 6, characterized in that: The specific operations in the experimental preparation stage of the massage parameter self-calibration operation process are as follows: The massage device that simulates the tactile sensation of human hands is connected and calibrated with the animal experimental platform with high precision. The massage head of the massage device is accurately brought into contact with the animal acupoints. The initial massage parameter range is extracted from historical experimental data and relevant literature. Combined with the characteristics of the animal model, a deep learning model is used for preliminary evaluation to determine the initial massage parameter values ​​and set the expected animal physiological response targets, including the range of changes in specific physiological indicators and the degree of improvement in behavioral learning performance. The deep learning model is preheated, the pre-trained model parameters are loaded, and it is integrated with the real-time data acquisition system and control module of the experimental equipment. The model is used to receive and process experiments in real time.

8. The method for self-calibration of acupoint massage parameters in animal experiments combined with deep learning according to claim 7, characterized in that: The specific operations in the experiment phase of the massage parameter self-calibration operation process are as follows: Start the massage device and start massaging the animal's acupoints according to the set initial parameters to simulate the human hand's touch. During the massage process, multimodal sensors are used to collect real-time massage force, frequency, mechanical and temperature characteristics data of the tissue at the acupoints, and physiological signal data of the animal, and transmit the data to the deep learning model. After receiving the real-time data, the deep learning model preprocesses the data, including data cleaning, normalization, and feature extraction operations. The processed data is input into the model for prediction to obtain the expected physiological response of the animal under the current massage parameters. The model compares and analyzes the predicted results with the set expected targets and calculates the difference between the two. During the massage process, the deep learning model continuously monitors the experimental data and massage effects, dynamically adjusts the massage parameters, and records the process and results of each parameter adjustment.

9. The method for self-calibration of acupoint massage parameters in animal experiments combined with deep learning according to claim 8, characterized in that: The specific operations at the end of the experiment in the massage parameter self-calibration operation process are as follows: After the experiment, all data generated in the experiment will be comprehensively organized and archived, including detailed parameter changes during the massage process, full data collected by multimodal sensors, complete physiological response data of animals, and operation records of the deep learning model. The newly generated experimental data and historical data will be merged, and the deep learning model will be incrementally trained and optimized. The model will be retrained to learn the characteristics and patterns of the experimental data. At the same time, the massage equipment will be inspected and maintained every week, and the hardware and software of the equipment will be adjusted and upgraded according to the operation status during the experiment.

Images