Adaptive moxibustion robot system based on multi-modal physiological feedback and control method

Abstract

This invention discloses an adaptive moxibustion robot system and control method based on multimodal physiological feedback, belonging to the field of intelligent medical equipment for traditional Chinese medicine. The system includes a main control computer, a power supply unit, a six-degree-of-freedom robotic arm, and an end effector module. The end effector integrates a carbon fiber infrared heating module, a high-resolution infrared thermal imager, and a laser Doppler blood flow probe. The method uses a multimodal sensing module to non-contactly acquire video streams of skin heat distribution and time-series data of microcirculation blood perfusion in the moxibustion area. After preprocessing, this data is input into a moxibustion sensation quantification mapping model of a dual-stream spatiotemporal neural network, achieving a quantification mapping between objective physiological indicators and the three phases of moxibustion sensation. Then, through a deep reinforcement learning adaptive control algorithm, using the quantified moxibustion sensation value as the state observation, the robotic arm posture and heating power are dynamically adjusted to form a closed-loop control. This invention breaks through the subjective barrier of moxibustion, achieving personalized adaptive moxibustion and improving the safety and clinical efficacy of moxibustion.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of TCM intelligent medical equipment, multimodal sensing technology and artificial intelligence control, and specifically relates to an adaptive moxibustion robot system and control method based on multimodal physiological feedback. Background Technology

[0002] Moxibustion is an important component of traditional Chinese medicine's external therapies. In clinical practice, the efficacy of moxibustion highly depends on the precise control of the "moxibustion dosage," which is directly reflected in the patient's "moxibustion sensation" (such as local heat, penetrating heat, and sensations along the meridians). However, existing moxibustion robots or intelligent moxibustion devices have the following significant drawbacks:

[0003] 1. The feedback mechanism is simplistic and lacks objective evaluation: Existing devices mostly rely on single-point temperature sensors for simple threshold control (e.g., rigidly maintaining the skin surface temperature at 42°C), ignoring individual differences among patients in terms of skin thickness, microcirculation, and heat tolerance. This "constant-temperature electric heating" control cannot induce a deep "moxibustion sensation."

[0004] 2. Lack of digital and quantitative representation of "moxibustion sensation": Traditional moxibustion sensation mainly relies on the patient's subjective description and the doctor's experience judgment, lacking objective physiological indicators, which makes it difficult for "moxibustion dosage" to be quantified and learned by computer systems.

[0005] 3. Rigid control strategy: The movement trajectory and heating power of the existing robotic arms are mostly preset open-loop control, which cannot be dynamically and adaptively adjusted according to the real-time physiological reactions of the patient during moxibustion (such as vasodilation, blood flow acceleration, and local heat accumulation), making it difficult to achieve personalized and precise moxibustion. Summary of the Invention

[0006] This invention aims to solve the technical problems of low precision in traditional moxibustion processes, difficulty in quantifying "moxibustion dosage", and lack of adaptive adjustment capabilities in existing moxibustion robots. It provides an adaptive moxibustion robot system and control method based on multimodal physiological feedback, which realizes the acquisition of objective physiological indicators, quantitative mapping of moxibustion sensation, and closed-loop adaptive control of the moxibustion process, thereby achieving personalized, precise, and safe moxibustion treatment.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] An adaptive moxibustion robot system and control method based on multimodal physiological feedback includes a main control computer, a power supply unit, a six-degree-of-freedom robotic arm, and an end effector module installed at the end of the six-degree-of-freedom robotic arm.

[0009] The power supply unit supplies power to all components of the system, and the main control computer is connected to the six-degree-of-freedom robotic arm via an industrial bus to send motion control commands and receive status feedback.

[0010] The end effector module integrates a carbon fiber infrared heating module, a high-resolution infrared thermal imager, and a laser Doppler blood flow probe, wherein:

[0011] The carbon fiber infrared heating module is used to emit far-infrared thermal radiation toward the treatment target area;

[0012] The high-resolution infrared thermal imager is used to collect real-time video streams of skin thermal distribution in the moxibustion area.

[0013] The laser Doppler blood flow probe is used for non-contact monitoring of blood perfusion in the subcutaneous microcirculation;

[0014] The system also has a hardware-level data synchronization mechanism. The main control computer outputs a pulse signal through the data acquisition card. The rising edge of the pulse signal is simultaneously connected to the external trigger pin of the infrared thermal imager and the acquisition gate of the laser Doppler blood flow probe, so that the time deviation between the i-th frame image of the infrared thermal imager and the i-th data packet of the blood flow probe is less than 1ms.

[0015] As a further technical solution of the present invention: the six-degree-of-freedom robotic arm is a UR5e series collaborative robotic arm with an end-effector repeatability of ±0.03mm and a working radius of 850mm.

[0016] The control method for an adaptive moxibustion robot based on multimodal physiological feedback, applied to the above-mentioned system, is characterized by including the following steps:

[0017] S1. Multimodal physiological data acquisition: Through the high-resolution infrared thermal imager and laser Doppler blood flow probe of the end effector module, non-contact real-time acquisition of skin heat distribution video stream and microcirculation blood perfusion volume time series data of the moxibustion area is carried out, and the time synchronization of the two types of data is achieved by using a hardware-level data synchronization mechanism.

[0018] S2. Data preprocessing: The collected infrared thermal image data and blood perfusion data are denoised and effective region extracted to obtain preprocessed feature data.

[0019] S3. Moxibustion Sensation Quantification Mapping: The preprocessed feature data is input into the moxibustion sensation quantification mapping model of the dual-stream spatiotemporal neural network. The spatial clustering features of the thermal imaging data and the temporal fluctuation features of the blood flow data are extracted. After feature fusion, the matching probability distribution of the moxibustion state and the three phases of moxibustion sensation is output, realizing the quantitative mapping from objective physiological indicators to the theory of moxibustion sensation in traditional Chinese medicine.

[0020] S4. Adaptive Closed-Loop Control: The moxibustion sensation quantification value output by the moxibustion sensation quantification mapping model is used to observe the environmental state. Through an adaptive control algorithm based on deep reinforcement learning, a reward function is constructed with the goal of maintaining the best moxibustion sensation and avoiding thermal damage. Control commands are output in real time to dynamically adjust the spatial pose of the six-degree-of-freedom robotic arm end and the output power of the carbon fiber infrared heating module. At the same time, a safety forced intervention mechanism is set to achieve closed-loop precise moxibustion.

[0021] As a further technical solution of the present invention: the preprocessing of infrared thermal image data in step S2 specifically includes:

[0022] A Gaussian smoothing filter with a standard deviation of σ=1.5 was used to denoise the infrared thermal image matrix I(x,y) to eliminate high-frequency noise; then, the Otsu method was used to automatically calculate the adaptive threshold, generate a binary mask, and extract the thermal image data of the moxibustion center target area.

[0023] As a further technical solution of the present invention: the preprocessing of blood perfusion data in step S2 specifically includes:

[0024] Daubechies4 was selected as the wavelet basis function to perform a 5-level wavelet transform decomposition on the original blood flow signal. The high-frequency detail coefficients of the first to third levels were set to zero, and the signal was reconstructed using the remaining low-frequency approximation coefficients to extract the effective feature signal reflecting the microcirculation perfusion.

[0025] As a further technical solution of the present invention: the moxibustion sensation quantification mapping model in step 3 includes a spatial feature extraction stream, a temporal feature extraction stream, and a cross-attention feature fusion module, and the specific processing procedure is as follows:

[0026] The spatial feature extraction stream takes a 16-frame sequence of thermal images of the moxibustion area ROI as input, with tensor dimensions of (B,C,K,H,W), and extracts spatiotemporal features through a 3D convolutional neural network, outputting a flattened spatial feature vector V_space.

[0027] The temporal feature extraction stream takes the blood perfusion time sequence X={x1,x2,...,x_T} as input, extracts temporal features through a Long Short-Term Memory (LSTM) network, and outputs a temporal feature vector V_time.

[0028] The cross-attention feature fusion module uses V_space as the query vector Q, V_time as the key vector K and value vector V, calculates the fusion feature V_fusion through the attention mechanism, and then outputs the moxibustion sensation probability distribution P_sensation through a fully connected layer and a Softmax function, which includes four states: no sensation, accumulation at the point of application, sensation along the meridian, and heat pain warning.

[0029] As a further technical solution of the present invention: the formula for calculating the feature map of the lth layer of the 3D convolutional neural network is:

[0030]

[0031] Where w is the weight of the 3D convolution kernel, and b is the bias term;

[0032] The core update formula for the LSTM unit is as follows:

[0033]

[0034] The final output is a time-series feature vector. .

[0035] The formula for calculating the attention mechanism is:

[0036]

[0037] Where d_k is the dimension of the key vector.

[0038] As a further technical solution of the present invention: the deep reinforcement learning algorithm in step S4 adopts the Deep Deterministic Policy Gradient (DDPG) algorithm, which realizes the control of the continuous action space based on the actor-commentator architecture, specifically as follows:

[0039] The actor network is based on the current state Directly output the defined action value This enables smooth and continuous adjustment of the robotic arm's position and heating power;

[0040] The critic network evaluates the merits of actions by calculating Q-values, which guides the parameter updates of the actor network.

[0041] As a further technical solution of the present invention: the state space Defined as a vector Where Temp_max is the highest skin temperature in the moxibustion area, P_sensation represents the change in blood perfusion and the probability distribution of moxibustion sensation.

[0042] Action space Defined as a vector ,in This refers to the height adjustment amount at the end of the robotic arm. This refers to the power adjustment amount of the carbon fiber infrared heating module.

[0043] As a further technical solution of the present invention: the reward function The calculation formula is

[0044]

[0045] in, The probability of meridian sensing output by the model. This is the therapeutic effect reward coefficient; The highest skin temperature is 43, and the comfortable temperature threshold is 43. This is the temperature penalty coefficient. For indicator functions, For the probability of a heat pain warning, This is a safety penalty coefficient; This is a stability penalty coefficient. =0.1;

[0046] The mandatory safety intervention mechanism is as follows: when it is detected that... or At that time, ignoring the actor network output, the robotic arm was forcibly raised by 50mm and the power supply to the carbon fiber infrared heating module was cut off.

[0047] This technology proposes an adaptive moxibustion robot system and control method based on multimodal physiological feedback, which has the following advantages and beneficial effects:

[0048] 1. Breaking down the subjective barrier of traditional Chinese medicine moxibustion: For the first time, multimodal physical characteristics such as infrared thermal imaging and laser Doppler blood perfusion are mathematically quantified and mapped with the "three phases of moxibustion sensation" theory of traditional Chinese medicine. The output includes the probability distribution of moxibustion sensation, including no sensation, accumulation at the point of action, sensation along the meridian, and heat and pain warning. A new objective and calculable standard for evaluating moxibustion dosage has been established, transforming the "moxibustion sensation" of traditional Chinese medicine moxibustion from subjective experience into objective data.

[0049] 2. Achieved truly personalized adaptive treatment: Abandoning the traditional "fixed distance, constant temperature" open-loop control mode of moxibustion equipment, the system monitors the patient's skin temperature distribution and microcirculation status in real time through multimodal sensors, "reading the room" like an experienced traditional Chinese medicine doctor, and dynamically adjusts the robotic arm posture and heating power through deep reinforcement learning algorithms, so that the moxibustion process is "different for each person and according to the time", adapting to the individual differences of different patients and the real-time physiological changes during the moxibustion process;

[0050] 3. Significantly improves the safety and clinical efficacy of moxibustion: Through high-frequency multimodal data monitoring (30fps thermal imaging acquisition, 100Hz blood flow sampling) and the ultra-fast response mechanism of deep reinforcement learning, it can effectively stimulate the patient's deep-level "penetrating heat" and "meridian sensation" moxibustion feeling, thereby improving the clinical treatment effect; at the same time, a multi-layer safety protection mechanism is set up, including temperature and heat pain penalty of reward function and hardware-level safety mandatory intervention, which completely eliminates the common risk of moxibustion burns in clinical practice and ensures the safety of the moxibustion process;

[0051] 4. The system has strong compatibility and practicality: It adopts a standardized industrial bus and a universal collaborative robotic arm. The end effector module has a high degree of integration, and the data acquisition has hardware-level synchronization. The preprocessing algorithm and deep learning model are implemented on the main control computer. The overall system structure is simple and the control is precise. It can be adapted to the moxibustion treatment needs of different parts and diseases in traditional Chinese medicine clinical practice, which is convenient for clinical promotion and application. Attached Figure Description

[0052] Figure 1 This is a diagram showing the overall hardware architecture of the adaptive moxibustion robot system of the present invention;

[0053] Figure 2 This is a schematic diagram of the multimodal physiological signal processing and moxibustion sensation quantification mapping model architecture of the present invention;

[0054] Figure 3 This is a flowchart of the robot adaptive closed-loop control logic based on reinforcement learning according to the present invention. Detailed Implementation

[0055] The present invention will be further described below with reference to the embodiments. It should be noted that these are merely examples and descriptions of the inventive concept. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the inventive concept or exceed the scope defined in the claims, they should all be considered to fall within the protection scope of the present invention.

[0056] Example 1: System Hardware Architecture and Multimodal Data Acquisition

[0057] like Figure 1 As shown, this invention provides an adaptive moxibustion robot system based on multimodal physiological feedback. The system mainly includes: a main control computer, a power supply unit, a six-degree-of-freedom robotic arm, and an end effector module installed at the end of the robotic arm. Specifically, the six-degree-of-freedom robotic arm preferably adopts the UR5e series collaborative robotic arm, with an end-effector repeatability of 0.03 mm and a working radius of 850 mm. The main control computer is connected to the robotic arm via an industrial bus and is used to send motion control commands and receive status feedback.

[0058] More specifically, the end effector module integrates a carbon fiber infrared heating module, a high-resolution infrared thermal imager, and a laser Doppler blood flow probe.

[0059] The carbon fiber infrared heating module is used to emit wavelengths of [wavelength value missing] towards the treatment target area. The far-infrared thermal radiation has an adjustable rated power range of [range missing]. The response time is less than 2 seconds.

[0060] The high-resolution infrared thermal imager is used to acquire real-time video streams of skin thermal distribution in the moxibustion area. Preferably, its resolution is not lower than... Pixel, thermal sensitivity (NETD) is better than The frame rate was set to 30fps.

[0061] The laser Doppler blood flow probe is used for non-contact monitoring of blood perfusion in the subcutaneous microcirculation. Preferably, its laser wavelength is 780nm and the sampling frequency is set to 100Hz.

[0062] Furthermore, the system also includes a hardware-level data synchronization mechanism. The main control computer outputs a 30Hz TTL level pulse signal through a data acquisition card. The rising edge of this pulse signal is simultaneously connected to the external trigger pin of the infrared thermal imager and the acquisition gating terminal of the laser Doppler blood flow probe. This mechanism ensures that the time deviation between the $i$-th frame image acquired by the infrared thermal imager and the i-th data packet acquired by the blood flow probe is less than 1ms.

[0063] Example 2: Data Preprocessing Algorithm

[0064] Before inputting the collected raw data into the deep learning model, the main control computer first performs a preprocessing step.

[0065] For infrared thermal imaging matrix The system first employs a Gaussian filter for noise reduction. Specifically, this Gaussian filter is a linear smoothing filter that eliminates high-frequency noise by applying a weighted average across the entire image. Its convolution kernel calculation formula is as follows:

[0066]

[0067] in The standard deviation is set to 1.5 in this embodiment. This step effectively suppresses random thermal noise generated by the sensor. Subsequently, the system automatically calculates an adaptive threshold using the Otsu method (maximum inter-class variance method) to generate a binary mask to extract the target area at the moxibustion center.

[0068] More specifically, Otsu's Method is a global adaptive image thresholding segmentation algorithm. Its core idea is to iterate through all possible gray levels in the image to find an optimal threshold T such that the inter-class variance between the two classes of pixels segmented according to this threshold (i.e., the foreground "high-temperature moxibustion area" and the background "low-temperature skin area") is... Maximum. Its variance is calculated using the following formula:

[0069]

[0070] in, , These represent the ratio of foreground and background pixels, , These represent the average gray levels of the foreground and background, respectively. By maximizing the inter-class variance, the system can stably and automatically separate the moxibustion heat field from the surrounding skin background under different patients and ambient temperatures.

[0071] For blood flow signals, the system employs wavelet transform for denoising. More specifically, wavelet transform is a time-scale analysis method for signals, featuring multi-resolution analysis capabilities. In this embodiment, Daubechies 4 (db4) is selected as the wavelet basis function to perform a 5-level decomposition of the original signal. The system forcibly sets the high-frequency detail coefficients of levels 1 to 3 to zero, as these high-frequency components typically correspond to physiological disturbances caused by respiration and heartbeat; subsequently, the remaining low-frequency approximation coefficients are used for reconstruction, thereby accurately extracting the effective feature signals reflecting microcirculation perfusion.

[0072] Example 3: A Moxibustion Sensation Quantification Mapping Model Based on a Dual-Stream Spatiotemporal Network

[0073] The moxibustion sensation quantification mapping model of the dual-stream spatiotemporal network is as follows: Figure 2 As shown, this invention provides a deep learning-based method for quantifying and mapping the sensation of moxibustion. The model includes a spatial feature extraction stream, a temporal feature extraction stream, and a cross-attention feature fusion module.

[0074] Specifically, the input data for the spatial feature extraction stream (3D-CNN) is a sequence of K consecutive frames (K=16) of ROI thermal images, with tensor dimensions of (B, C, K, H, W). A 3D convolutional neural network is used to extract spatiotemporal features, and the feature map calculation formula for the l-th layer is:

[0075]

[0076] Where w represents the weights of the 3D convolution kernel. The output is a flattened spatial feature vector. Temporal Feature Extraction Stream (LSTM): Input data is a blood flow sequence. The Long Short-Term Memory (LSTM) network is used for processing. The core update formula for the LSTM unit is as follows:

[0077]

[0078] The final output is a time-series feature vector. .

[0079] The cross-attention feature fusion is defined as follows: For query vector , Key vector Sum value vector Utilizing attention mechanisms to compute fused features :

[0080]

[0081] Finally, the probability distribution of the moxibustion sensation is output through a fully connected layer. It includes four states: no sensation, accumulation at the point of action, sensation along the meridian, and heat and pain warning.

[0082] Example 4: Adaptive Closed-Loop Control Based on Deep Deterministic Policy Gradient

[0083] The adaptive closed-loop control of the gradient of the deep deterministic policy is as follows: Figure 3 As shown, this invention provides a robot adaptive control method based on deep reinforcement learning. The method employs a deep deterministic policy gradient algorithm.

[0084] Specifically, the DDPG (Deep Deterministic Policy Gradient) algorithm is a deep reinforcement learning algorithm based on an "actor-critic" architecture, specifically designed to solve problems with continuous action spaces. The actor network is responsible for adjusting the current state... Directly output the defined action value ;

[0085] The "critic" network is responsible for evaluating the quality of the action, guiding its parameter updates by calculating the Q-value (action value function). Compared to discrete control algorithms, the DDPG algorithm can achieve smooth and continuous adjustment of the robotic arm's position and heating power, avoiding motion jitter.

[0086] Specifically, state space Defined as a vector .

[0087] Action space Defined as a vector To balance efficacy and safety, this invention designs a specific composite reward function. :

[0088]

[0089] in:

[0090] The probability of "meridian sensation" output by the model. This is the therapeutic effect reward coefficient;

[0091] The highest skin temperature is 43, and the comfortable temperature threshold is 43. This is the temperature penalty coefficient;

[0092] Let be the indicator function, representing the predicted probability of a "heat pain warning". If the value is greater than 0.8, take 1; otherwise, take 0. This is a safety penalty coefficient; =0.1 is the stability penalty coefficient, used to prevent the robotic arm from making frequent large movements.

[0093] In addition, the system has a security-mandated intervention mechanism: when it detects... or At that time, the Actor network output is ignored, the robotic arm is forced to rise 50mm and the heating power is cut off.

[0094] The above is an exemplary description of the invention. Obviously, the specific implementation of the invention is not limited to the above-described manner. Any non-substantial improvement made using the inventive concept and technical solution of the invention, or the direct application of the inventive concept and technical solution to other situations without modification, is within the protection scope of the invention.

Claims

1. An adaptive moxibustion robot system based on multimodal physiological feedback, characterized in that, It includes a main control computer, a power supply unit, a six-degree-of-freedom robotic arm, and an end effector module installed at the end of the six-degree-of-freedom robotic arm; The power supply unit supplies power to all components of the system, and the main control computer is connected to the six-degree-of-freedom robotic arm via an industrial bus to send motion control commands and receive status feedback. The end effector module integrates a carbon fiber infrared heating module, a high-resolution infrared thermal imager, and a laser Doppler blood flow probe, wherein: The carbon fiber infrared heating module is used to emit far-infrared thermal radiation toward the treatment target area; The high-resolution infrared thermal imager is used to collect real-time video streams of skin thermal distribution in the moxibustion area. The laser Doppler blood flow probe is used for non-contact monitoring of blood perfusion in the subcutaneous microcirculation; The system also features a hardware-level data synchronization mechanism. The main control computer outputs pulse signals through the data acquisition card. The rising edge of the pulse signal is simultaneously connected to the external trigger pin of the infrared thermal imager and the acquisition gating terminal of the laser Doppler blood flow probe.

2. The adaptive moxibustion robot system based on multimodal physiological feedback according to claim 1, characterized in that, The six-degree-of-freedom robotic arm is a UR5e series collaborative robotic arm with an end-effector repeatability of ±0.03mm and a working radius of 850mm.

3. A control method for an adaptive moxibustion robot based on multimodal physiological feedback, applied to the system described in any one of claims 1-2, characterized in that, Includes the following steps: S1. Multimodal physiological data acquisition: Through the high-resolution infrared thermal imager and laser Doppler blood flow probe of the end effector module, non-contact real-time acquisition of skin heat distribution video stream and microcirculation blood perfusion volume time series data of the moxibustion area is carried out, and the time synchronization of the two types of data is achieved by using a hardware-level data synchronization mechanism. S2. Data preprocessing: The collected infrared thermal image data and blood perfusion data are denoised and effective region extracted to obtain preprocessed feature data. S3. Moxibustion Sensation Quantification Mapping: The preprocessed feature data is input into the moxibustion sensation quantification mapping model of the dual-stream spatiotemporal neural network. The spatial clustering features of the thermal imaging data and the temporal fluctuation features of the blood flow data are extracted. After feature fusion, the matching probability distribution of the moxibustion state and the three phases of moxibustion sensation is output, realizing the quantitative mapping from objective physiological indicators to the theory of moxibustion sensation in traditional Chinese medicine. S4. Adaptive Closed-Loop Control: The moxibustion sensation quantification value output by the moxibustion sensation quantification mapping model is used to observe the environmental state. Through an adaptive control algorithm based on deep reinforcement learning, a reward function is constructed with the goal of maintaining the best moxibustion sensation and avoiding thermal damage. Control commands are output in real time to dynamically adjust the spatial pose of the six-degree-of-freedom robotic arm end and the output power of the carbon fiber infrared heating module. At the same time, a safety forced intervention mechanism is set to achieve closed-loop precise moxibustion.

4. The control method for the adaptive moxibustion robot based on multimodal physiological feedback according to claim 3, characterized in that, The preprocessing of infrared thermal image data in step S2 is as follows: A Gaussian smoothing filter with a standard deviation of σ=1.5 was used to denoise the infrared thermal image matrix I(x,y) to eliminate high-frequency noise; then, the Otsu method was used to automatically calculate the adaptive threshold, generate a binary mask, and extract the thermal image data of the moxibustion center target area.

5. The control method for the adaptive moxibustion robot based on multimodal physiological feedback according to claim 3, characterized in that, The preprocessing of blood perfusion data in step S2 is as follows: Daubechies4 was selected as the wavelet basis function to perform a 5-level wavelet transform decomposition on the original blood flow signal. The high-frequency detail coefficients of the first to third levels were set to zero, and the signal was reconstructed using the remaining low-frequency approximation coefficients to extract the effective feature signal reflecting the microcirculation perfusion.

6. The control method for the adaptive moxibustion robot based on multimodal physiological feedback according to claim 3, characterized in that, The moxibustion sensation quantification mapping model in step 3 includes a spatial feature extraction stream, a temporal feature extraction stream, and a cross-attention feature fusion module. The specific processing procedure is as follows: The spatial feature extraction stream takes a 16-frame continuous thermal image sequence of the moxibustion area ROI as input, with tensor dimensions of B, C, K, H, W, and extracts spatiotemporal features through a 3D convolutional neural network, outputting a flattened spatial feature vector V_space. The temporal feature extraction stream takes the blood perfusion time sequence X={x1,x2,...,x_T} as input, extracts temporal features through a long short-term memory network LSTM, and outputs a temporal feature vector V_time; The cross-attention feature fusion module uses V_space as the query vector Q, V_time as the key vector K and value vector V, calculates the fusion feature V_fusion through the attention mechanism, and then outputs the moxibustion sensation probability distribution P_sensation through a fully connected layer and a Softmax function, which includes four states: no sensation, accumulation at the point of application, sensation along the meridian, and heat pain warning.

7. The control method for the adaptive moxibustion robot based on multimodal physiological feedback according to claim 3, characterized in that, The formula for calculating the feature map of the l-th layer of a 3D convolutional neural network is: Where w is the weight of the 3D convolution kernel, and b is the bias term; The core update formula for the LSTM unit is as follows: The final output is a time-series feature vector. ; The formula for calculating the attention mechanism is: Where d_k is the dimension of the key vector.

8. The control method for the adaptive moxibustion robot based on multimodal physiological feedback according to claim 3, characterized in that, The deep reinforcement learning algorithm in step S4 employs the Deep Deterministic Policy Gradient (DDPG) algorithm, which uses an actor-critic architecture to control the continuous action space. Specifically: The actor network is based on the current state Directly output the defined action value This enables smooth and continuous adjustment of the robotic arm's position and heating power; The critic network evaluates the merits of actions by calculating Q-values, which guides the parameter updates of the actor network.

9. The control method for an adaptive moxibustion robot based on multimodal physiological feedback according to claim 8, characterized in that, state Space is defined as a vector Where Temp_max is the highest skin temperature in the moxibustion area. P_sensation represents the change in blood perfusion volume, and P_sensation represents the probability distribution of moxibustion sensation. Action value Space is defined as a vector ,in This refers to the height adjustment amount at the end of the robotic arm. This refers to the power adjustment amount of the carbon fiber infrared heating module.

10. The control method for the adaptive moxibustion robot based on multimodal physiological feedback according to claim 9, characterized in that, The reward function The calculation formula is in, The probability of meridian sensing output by the model. This is the therapeutic effect reward coefficient; The highest skin temperature is 43, and the comfortable temperature threshold is 43. This is the temperature penalty coefficient. For indicator functions, For the probability of a heat pain warning, This is a safety penalty coefficient; This is a stability penalty coefficient. =0.1; The mandatory safety intervention mechanism is as follows: when it is detected that... or At that time, ignoring the actor network output, the robotic arm was forcibly raised by 50mm and the power supply to the carbon fiber infrared heating module was cut off.

Images