Control method and system of massage device for assisting sleep

Abstract

The invention discloses a control method of a massage device for assisting sleep, which comprises the following steps: S1, collecting a multi-modal physiological signal of a user in a sleep process in real time through a non-contact or micro-contact sensor, and preprocessing the multi-modal physiological signal; s2, inputting the collected multi-modal physiological signals into a pre-trained sleep analysis model, and calculating and outputting sleep stage probability distribution and sleep stability index of the current user in real time; s3, dynamically generating an optimal massage control instruction according to the sleep stage probability distribution and the sleep stability index at the current moment and a preset intervention target library; and S4, recording physiological signal change and sleep stage evolution data before and after massage intervention, and continuously optimizing the sleep analysis model and the massage decision strategy in combination with the morning subjective sleep quality score of the user.

Description

Technical Field

[0001] This invention relates to the field of physiotherapy equipment control technology, and in particular to a control method and system for a massage device used to assist sleep. Background Technology

[0002] Sleep disorders have become a widespread public health problem globally, with traditional solutions including pharmacological interventions, cognitive behavioral therapy, and various sleep aids. Among these, physical sleep aids with massage functions have gained widespread attention due to their non-invasiveness and soothing properties. Currently available sleep-aid massage devices, such as massage pillows and mattresses, mostly rely on simple timers or preset, fixed program cycles for control. These programs are typically designed based on general relaxation logic, such as performing a massage of a fixed duration and using fixed techniques before the user goes to sleep.

[0003] However, this "one-size-fits-all" control model has significant technical limitations: First, it lacks perception. The device cannot sense and identify the user's actual sleep state (such as whether they have fallen asleep or what sleep stage they are in) in real time. A user trying to fall asleep and a user already in deep sleep receive the same mechanical stimulation. Second, it suffers from rigid decision-making. Its intervention is open-loop and blind, unable to dynamically adjust based on the user's real-time physiological feedback. This not only significantly reduces the sleep aid effect but may also disrupt sleep continuity and lead to fragmented sleep by applying inappropriate intensity or frequency of stimulation at inappropriate times (such as during deep sleep). Third, it lacks personalization and evolutionary capabilities. The device cannot learn and optimize its strategies from long-term interactions with specific users; its functions are fixed at the factory and cannot adapt to individual differences and changing sleep needs over time.

[0004] To enhance intelligence, some high-end products have begun integrating single biosensors (such as heart rate sensors) in an attempt to acquire user status. However, relying solely on single-modal signals like heart rate makes it difficult to accurately and robustly distinguish complex sleep stages (such as differentiating between N1 light sleep and REM sleep), let alone quantify sleep "stability." Their decision-making logic often remains based on finite state machines with simple threshold judgments, failing to achieve true understanding and decision-making based on multi-dimensional physiological context. Therefore, existing technological solutions suffer from a gap in the core chain of "precise perception – intelligent decision-making – personalized adaptation." Summary of the Invention

[0005] In order to solve the above-mentioned technical problems, the present invention provides a control method and system for a massage device for assisting sleep.

[0006] The technical solution of this invention is implemented as follows: A method for controlling a massage device for assisting sleep, comprising: S1. Collect multimodal physiological signals of the user during sleep in real time using non-contact or micro-contact sensors, and preprocess the multimodal physiological signals. S2. Input the collected multimodal physiological signals into the pre-trained sleep analysis model, and calculate and output the probability distribution of the current user's sleep stages and the sleep stability index in real time. S3. Based on the current sleep stage probability distribution and sleep stability index, as well as the preset intervention target library, dynamically generate the optimal massage control instructions. S4. Record changes in physiological signals and sleep stage evolution data before and after massage intervention, and combine them with the user's subjective sleep quality score the following morning to continuously optimize the sleep analysis model and massage decision-making strategy.

[0007] A control system for a massage device used to assist sleep includes: The signal acquisition and preprocessing module includes a millimeter-wave radar sensor and a piezoelectric thin film sensor, which are used to acquire the user's raw physiological signals and preprocess them to remove body motion interference. The sleep state analysis module has a built-in deep learning model, which is used to receive pre-processed physiological signals, calculate and output the user's sleep stage probability distribution and sleep stability index in real time. The adaptive decision-making module integrates an intervention strategy knowledge base and an effect prediction model, which is used to generate and output optimized massage control instructions based on the probability distribution of the sleep stages and the sleep stability index. The massage execution module includes multiple independently controllable miniature massage actuators for executing the massage control commands; The user interaction and learning module is used to provide a user interface and collect subjective sleep quality scores. Through reinforcement learning algorithms, it continuously optimizes the sleep state analysis module and the adaptive decision-making module by utilizing changes in physiological signals before and after intervention, sleep stage evolution data, and the subjective scores. The central controller is used to coordinate and control the operation of the above modules.

[0008] Compared with the prior art, the present invention has the following advantages: This invention collects and analyzes multimodal physiological signals of users during sleep in real time, and can objectively identify different real-time sleep stages and stable states such as difficulty falling asleep, light sleep with easy awakening, or deep sleep. This allows massage intervention to be triggered based on an understanding of the user's current physiological state, changing the blind working mode of existing technologies that ignores the user's state and only performs fixed operations according to preset time or sequence. Based on the analysis of real-time sleep states, this invention dynamically generates and executes massage control commands that match the state. Therefore, this invention can provide soothing guidance when users need help falling asleep, provide stabilizing assistance when sleep is unstable, and actively reduce interference during deep sleep. This provides differentiated intervention strategies for different states such as difficulty falling asleep and unstable sleep, thereby improving the pertinence and effectiveness of the intervention. This invention records physiological data changes before and after each intervention and integrates user subjective feedback, enabling continuous optimization of the analysis model and decision-making strategy. This allows the device to adapt to individual user sleep patterns and response preferences as it is used, thus solving the problem of existing technologies lacking learning and evolution capabilities. Through the synergistic effect of real-time perception, dynamic decision-making, and continuous learning capabilities, this invention provides a solution that can more effectively assist users in improving sleep onset efficiency, maintaining sleep continuity, and reducing sleep interruptions. Its intervention process is coordinated with the user's natural sleep circadian rhythm, and compared with existing technologies based on fixed cycles, it has significant advantages in improving sleep quality and user satisfaction. Attached Figure Description

[0009] Figure 1 This is a flowchart of a control method for a massage device for assisting sleep according to the present invention; Figure 2 This is a system framework diagram of a massage device control system for assisting sleep according to the present invention. Detailed Implementation

[0010] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0011] Example 1 like Figure 1 As shown, this embodiment provides a control method for a massage device used to assist sleep, including: S1. Collect multimodal physiological signals of the user during sleep in real time using non-contact or micro-contact sensors, and preprocess the multimodal physiological signals. S2. Input the collected multimodal physiological signals into the pre-trained sleep analysis model, and calculate and output the probability distribution of the current user's sleep stages and the sleep stability index in real time. S3. Based on the current sleep stage probability distribution and sleep stability index, as well as the preset intervention target library, dynamically generate the optimal massage control instructions. S4. Record changes in physiological signals and sleep stage evolution data before and after massage intervention, and combine them with the user's subjective sleep quality score the following morning to continuously optimize the sleep analysis model and massage decision-making strategy.

[0012] The multimodal physiological signals in step S1 include at least body movement signals, heart rate variability signals, and respiratory rhythm signals.

[0013] In step S2, the sleep stability index integrates the coordination of body movement frequency, heart rate and breathing, as well as the smoothness of the transition between each stage.

[0014] The preprocessing of multimodal physiological signals in step S1 includes: By comparing and analyzing the phase change of the millimeter-wave echo signal with the charge change amplitude of the piezoelectric signal, interference data segments caused by large-amplitude body motion are identified and eliminated, resulting in anti-interference millimeter-wave phase sequence and piezoelectric charge sequence. The above anti-interference sequence was filtered and blind source separation was performed to initially separate the signal components of physiological information; Specifically, it can be illustrated with the following example: In this embodiment, the core of the process for extracting anti-interference physiological signals from raw sensor data lies in utilizing the complementarity of multi-sensor information and following the processing flow of "interference identification - signal separation" to ultimately provide high-quality breathing and heartbeat signals for sleep analysis. The key lies in distinguishing between micro-movements caused by breathing and heartbeat and body movement interference caused by turning over, as detailed below: Simultaneously acquire the complex baseband signal (I / Q signal) of the millimeter-wave radar and the original voltage signal of the piezoelectric thin film; For millimeter-wave signals, calculate their phase sequence. This phase is extremely sensitive to changes in distance; both chest rise and fall and limb movement will cause changes in it. For piezoelectric signals, calculate their energy amplitude sequence within a short time window (e.g., 1 second). This value directly reflects the intensity of the pressure change acting on the sensor; Since large-amplitude body movements simultaneously cause dramatic phase jumps in millimeter waves and a sudden increase in piezoelectric signal energy, while breathing and heartbeats only cause regular periodic phase changes and slight fluctuations in piezoelectric energy, a combined body movement index can be constructed. To identify interference:

[0015] in, This is the absolute value of the phase change (differential calculation). and These are normalized weighting coefficients used to balance the dimensions and importance of the two terms; Set a dynamic threshold ,when At that moment, it is determined that the situation is under "significant physical disturbance". Mark all consecutive time points of interference states as interference data segments to be removed; The millimeter-wave phase value and piezoelectric voltage value marked as interference data segment are either set to zero or linearly interpolated to obtain an anti-interference millimeter-wave phase sequence. and anti-interference piezoelectric charge sequence In these sequences, outliers caused by large-scale body movements have been effectively suppressed; Respiratory and heartbeat signal components suitable for analysis were isolated from the anti-interference sequence: right and Bandpass filtering is performed separately to remove DC offset and high-frequency noise. The passband for the respiratory signal is typically set to 0.1Hz to 0.8Hz, and the passband for the heartbeat signal is set to 0.8Hz to 2.5Hz. The filtered signals are then the band-limited signals from both sensors. and ,Will and These are considered as observed signals, which are a mixture of independent respiratory and cardiac sources. Independent component analysis (ICA) is used for blind source separation; specifically, the two observed signals are combined into a vector. The ICA algorithm finds a demixing matrix W through iterative optimization, such that the two components of the output Y(t) = WX(t) are equal. and Statistically as independent as possible; Based on the separated signal and Based on the spectral characteristics (main peak frequency) and waveform periodicity, it automatically identifies which channel corresponds to the respiratory signal. Which path corresponds to the heartbeat signal? For example, a signal with a dominant frequency of 0.2-0.3 Hz is a respiratory signal, and a signal with a dominant frequency of 1.0-1.5 Hz is a heartbeat signal; The final output of the preprocessing process is the initially separated respiratory signal component. and heartbeat signal components These signals have eliminated body movement interference to the greatest extent possible and completed the initial source separation, becoming the input for subsequent sleep stage and stability analysis steps.

[0016] Step S2 specifically includes: S21. Input the preprocessed multimodal physiological signal sequence into a pre-trained deep neural network to generate context-aware feature vectors.

[0017] S22. Input the feature vector into the classification layer and output the real-time probability distribution of the user's sleep stage.

[0018] S23. Based on the feature vector and stage probability, calculate the dynamic time regularization distance between the current feature and recent historical features and the transition entropy of the stage probability sequence, and integrate the above indicators and stage prior values ​​to normalize them into a sleep stability index.

[0019] The deep neural network adopts a hybrid architecture of bidirectional recurrent neural network and attention mechanism; Specifically, it can be illustrated with the following example: In this embodiment, the preprocessed physiological signals are transformed into quantified sleep stage probabilities and stability indices. The core technology lies in using deep time series models to extract contextual features and combining information theory indicators for multi-dimensional evaluation. Context-aware feature extraction based on Bi-LSTM and attention mechanism: The continuous respiratory and heartbeat signal streams are divided into analysis segments (epochs) of fixed duration (e.g., 30 seconds). Each segment corresponds to a time point t to be analyzed. The dual-channel signals within each segment are standardized, and a series of primary features in the time and frequency domains (such as mean, variance, peak frequency, spectral entropy, etc.) are calculated to form the initial feature vector of that segment. ; The initial feature sequence of the current segment and several segments before and after it (e.g., 5 segments before and 5 segments after, for a total of 11 segments). As input to the model, to provide contextual information; Bidirectional LSTM layer (Bi-LSTM): This layer receives the input sequence X. The forward LSTM processes the sequence in chronological order, and the backward LSTM processes the sequence in reverse order. For a given time point t, the Bi-LSTM hides its forward state. and backward hidden state By concatenating the components, we obtain the hidden state containing complete contextual information. This state captures the dependencies of the sequence before and after time t; Attention mechanism layer: Not all historical contexts are equally important in determining the current state. The attention layer represents the hidden state at each time step in the sequence. Calculate a weight The weights reflect the importance of the information at that moment in determining the state of the current central segment t. The calculation process is as follows: Each is calculated using a small neural network. Score , where W, b, and v are learnable parameters; The attention weights are obtained by normalizing all scores using the softmax function. ; Calculate the context vector: This vector is a weighted sum of all hidden states, highlighting information from key time periods; Finally, the original hidden state of the central segment t is... Attention-weighted context vector The feature vectors are fused (e.g., concatenated) and then reduced and refined through a fully connected layer to output the final context-aware feature vector. The vector It forms the basis for subsequent classification and stability calculations; eigenvectors The input consists of a classification layer (usually two layers) made up of a fully connected network, and finally passes through a Softmax activation function; The Softmax function transforms the network output into a probability distribution. , where each component This represents the probability that, in the current analysis segment t, the user is awake (W, N1, N2, N3, REM), and the sum of all components is 1. Dynamic time warping (DTW) is used to measure the similarity between two time series that may be of different lengths, and it can effectively handle fluctuations and scaling. The feature vector at the current time t Compared with all feature vectors over a past period (e.g., the first 10 minutes, approximately 20 segments). Calculate the DTW distance separately. During the calculation, each feature vector is treated as a "point" in a multidimensional sequence. The core of DTW is to find a regular path that minimizes the cumulative distance after aligning two sequences. Here, we use this minimum cumulative distance as the feature drift. , The smaller the value, the more stable and continuous the recent physiological pattern is; Transition entropy is a metric in information theory that measures the dynamic complexity of a process. Here, it is used to quantify the unpredictability or disorder of sleep stage transitions. First, based on probability distributions… The stage with the highest probability is selected as the prediction stage for the current segment. This yields a recent phase sequence (such as the first 20 segments). The transition entropy TE of this stage sequence can be calculated using a simplified approximation based on a first-order Markov chain, as shown in the formula:

[0020] in, It is a joint probability. It is the transition probability. These are marginal probabilities, all derived from historical sequences. The higher the TE value, the more random and unstable the stage transition. Assign a baseline stability value to different sleep stages. (For example, N3 is highest during deep sleep and W is lowest during wakefulness). By combining the above indicators with the prior values ​​of the current dominant stage (the stage with the highest probability), an initial stability score is calculated. :

[0021] in, This means normalizing the index to [0, 1]. , , These are the weighting coefficients, and their sum is 1. Finally, The sleep stability index is output by mapping to the (0,1) interval using the Sigmoid function. The closer the index is to 1, the more stable and continuous the sleep is.

[0022] The process of generating the optimal massage control command in step S3 specifically includes: S31. Based on the current dominant sleep stage and stability index, generate a multi-dimensional intervention demand vector through a demand calculation function; S32. Query the intervention strategy knowledge base using the current state tuple, and dynamically synthesize a preliminary massage strategy; S33. Input the preliminary strategy into the effect prediction model to predict the changes in physiological state after its execution. Under safety constraints, with the goal of maximizing prediction stability and the probability of deep sleep, fine-tune the strategy parameters to obtain optimized massage instructions and execute them.

[0023] In step S32, the dynamic synthesis of the preliminary massage strategy adopts nearest neighbor matching or a neural network strategy generator. Specifically, it can be illustrated with the following example: Dominant sleep stages and stability indices are quantified into specific intervention needs: From the probability distribution of the current sleep stages, select the stage with the highest probability as the dominant sleep stage (e.g., =0.7, then the dominant stage is N3). Get the current sleep stability index (0-1); Define a three-dimensional intervention demand vector These represent the intensity of the need to induce sleep, maintain stability, and remain quiet, respectively. The demand calculation function F is based on the current state. Calculated using a set of heuristic rules or by querying a predefined demand mapping table: Promote the need to fall asleep The dominant phase is "awake (W)" or "N1 phase," and A value less than 0.3 is considered high; it can be set to: If ,but =1- Otherwise, it is 0; Maintain stable demand The dominant phase is either "N2 phase" or "REM phase," and When the value is in a moderate range (e.g., 0.3-0.7) or shows a downward trend, it can be set to: If ,but (This indicates high demand when deviating from the ideal stability value), otherwise it is 0; The need to maintain quiet The dominant phase is "N3" or A value >0.8 is considered high; it can be set to: If or ,but Otherwise, it is 0; For the calculated Normalize the vector so that the sum of all components is 1 to obtain the final demand vector. Based on the current status and needs, a preliminary massage plan is retrieved and combined from the strategy knowledge base: The dominant phase Stability Index and normalized demand vector Combine them into a current state tuple, which will serve as the lookup key; The intervention strategy knowledge base stores historical success records, and each record contains: State: Historical state tuple.

[0024] Action: Massage strategy parameters to be performed, including location (head, neck, back, etc.), technique (pressing, kneading, vibration, heat application), intensity (0-1), frequency (Hz), and duration (seconds); Effect: Short-term stability change after execution and the change in the probability of deep sleep ; Calculate the similarity between the current state tuple and each historical state tuple in the knowledge base, for numerical features. Use Euclidean distance; for categorical features A similarity score is 1, and a different score is 0. The total similarity is obtained by weighted summation. Retrieve the K records with the highest similarity (e.g., K=5). For each parameter of the massage action (e.g., intensity), calculate a weighted average of the K retrieved parameter values ​​according to their similarity weights to obtain the parameters of the initial strategy:

[0025] in, Let be the similarity of the i-th record. Its strength value; Similarly, parameters such as body part (which can be weighted using multi-thermal coding), technique (voting or weighted), and frequency are integrated to form a preliminary massage strategy. ; The initial strategy is validated for security and effectiveness, and then optimized before execution: Effect prediction model: This model is a neural network (such as a fully connected network), whose input is the current feature vector. Given the massage strategy parameter A to be evaluated, the output is a prediction of the change in the sleep stability index over a future period (as shown in the next analysis segment) after implementing the strategy. Changes in the probability of deep sleep ; Define the optimization objective: The objective is to maximize the positive effects of the intervention while minimizing the interference. Construct the objective function. :

[0026] Where α, β, and γ are weights, and Penalty(A) is a penalty term for unsafe or inappropriate parameters (such as excessive intensity or using active massage when there is a high need for quiet). Optimization must adhere to hard constraints, such as: During this time, all active massage techniques are prohibited; only gentle heat application is permitted. Under no circumstances should the intensity exceed 80% of the user's personal historical maximum tolerance intensity. With initial strategy Starting with the model as a starting point, and under the premise of satisfying safety constraints, a simple local search or gradient descent (if the model is differentiable) is used to make small-scale adjustments to the adjustable parameters (intensity, frequency) to find the optimal parameters. Maximize the optimization of massage commands ; Will The commands are interpreted as specific control commands (e.g., activate the back airbag group, apply wave-like pressure at an intensity of 0.3 and a frequency of 0.5Hz for 120 seconds) and sent to the massage actuator, which then precisely performs the massage operation on the designated body part.

[0027] The process of continuously optimizing the sleep analysis model and massage decision-making strategy in step S4 specifically includes: S41. Record the complete experience data package for each intervention, including: pre-intervention state, performed actions, post-intervention state transition, and immediate reward calculated from physiological changes; S42. Combining the user's subjective sleep quality score the following morning, the accumulated experience data package is periodically used to update the policy network and effect prediction model through reinforcement learning algorithm, and the sleep analysis model can be fine-tuned optionally. S43. Update the newly learned strategy knowledge to the intervention strategy knowledge base and adjust the user's personalized parameters; Specifically, it can be illustrated with the following example: After each massage instruction is executed, the system immediately encapsulates and stores a complete decision-feedback experience. Each intervention experience is encapsulated into a 5-tuple data packet. : Pre-intervention status : Feature vector at the moment before instruction execution Dominant sleep stage and sleep stability index ; Execute action : Optimized massage instructions executed Parametric descriptions (location, technique, intensity, frequency, duration); Post-intervention state transition The average feature vector, dominant sleep stage, and stability index within a time window (such as the next 5 minutes) after the instruction execution is completed; Instant rewards : The reward value calculated based on short-term physiological changes after intervention. Its calculation function aims to quantify the immediate effect of a single intervention, for example, defined as:

[0028] in, , , These are the weighting coefficients. The first two rewards, stability and the probability of deep sleep, are considered positive. The cost of the action (e.g., positively correlated with massage intensity and duration) is used to encourage efficient and gentle interventions; Subjective reward Initially empty, after the user submits the subjective sleep quality score the following morning (e.g., a VAS scale score of 0-10 via a mobile app), all intervention data packages for that night are associated with this score as a long-term, global reward signal. The system periodically (e.g., every morning) utilizes a pool of experience data accumulated over a period of time (e.g., a week) to perform batch offline training. Policy Network The parameter θ defines the mapping from state to action (i.e., the decision logic of S3), which is updated using the Actor-Critic reinforcement learning framework. First, train a Critic network. To evaluate the expected long-term reward of performing action A in state S, temporal difference (TD) learning is used to minimize the loss function:

[0029] Where R is the combined reward that combines immediate reward and final subjective reward, and γ is the discount factor; The policy network is updated using the policy gradient method, with the goal of maximizing the expected return. The gradient is approximated as:

[0030] By updating the parameter θ through gradient ascent, a higher expected return can be generated given the state S. The probability of action A being selected increases; The update of the result prediction model (corresponding to the model in step S3) is performed as a supervised learning task; Its training data consists of triples from the experience data set. The input is the state. and actions The output is the predicted next state. ; By minimizing the predicted state and the actual observed state The mean squared error (MSE) between the two sides is used to update the model parameters. ; Once a large amount of high-confidence user data has been accumulated (e.g., through multiple manual calibrations or by using high-stability period data as self-supervised labels), fine-tuning of the sleep analysis model in step S2 can be initiated. Using user-specific "physiological signals - sleep stages" paired data, a small number of iterative transfer learning processes are performed on the pre-trained model to improve the accuracy of identifying the individual physiological characteristics of the user. By accumulating the "knowledge" gained from offline learning into the system, its online decision-making becomes more intelligent: Reinforcement learning strategy network High-value actions generated under new conditions, or those discovered during the exploration process and proven to be highly effective (high-value actions). and New strategy parameters are added to the knowledge base as new "successful strategy fragments". The knowledge base is managed by priority queues or sliding windows, and old and ineffective records are regularly eliminated, while efficient and novel strategies are retained to ensure the timeliness and quality of the knowledge base. Analyze users' personal patterns from historical data and dynamically adjust internal parameters, for example: Sensitivity coefficient: Update the weights in the demand calculation function (S31) and the λ coefficient in the reward function (S41) to match the user's actual physiological and psychological response to various interventions; Based on historical tolerance records, the safety limits for massage intensity and duration are adjusted individually. These parameters are continuously updated as a user's "personalized profile," making the system's decisions increasingly tailored to the individual.

[0031] Example 2 like Figure 2 As shown, this embodiment provides a control system for a massage device used to assist sleep, aiming to implement a control method for a massage device used to assist sleep. The system includes: The signal acquisition and preprocessing module includes a millimeter-wave radar sensor and a piezoelectric thin film sensor, which are used to acquire the user's raw physiological signals and preprocess them to remove body motion interference. The sleep state analysis module has a built-in deep learning model, which is used to receive pre-processed physiological signals, calculate and output the user's sleep stage probability distribution and sleep stability index in real time. The adaptive decision-making module integrates an intervention strategy knowledge base and an effect prediction model, which is used to generate and output optimized massage control instructions based on the probability distribution of the sleep stages and the sleep stability index. The massage execution module includes multiple independently controllable miniature massage actuators for executing the massage control commands; The user interaction and learning module is used to provide a user interface and collect subjective sleep quality scores. Through reinforcement learning algorithms, it continuously optimizes the sleep state analysis module and the adaptive decision-making module by utilizing changes in physiological signals before and after intervention, sleep stage evolution data, and the subjective scores. The central controller is used to coordinate and control the operation of the above modules.

Claims

1. A method for controlling a massage device for assisting sleep, characterized in that, include: S1. Collect multimodal physiological signals of the user during sleep in real time using non-contact or micro-contact sensors, and preprocess the multimodal physiological signals. S2. Input the collected multimodal physiological signals into the pre-trained sleep analysis model, and calculate and output the probability distribution of the current user's sleep stages and the sleep stability index in real time. S3. Based on the current sleep stage probability distribution and sleep stability index, as well as the preset intervention target library, dynamically generate the optimal massage control instructions. S4. Record changes in physiological signals and sleep stage evolution data before and after massage intervention, and combine them with the user's subjective sleep quality score the following morning to continuously optimize the sleep analysis model and massage decision-making strategy.

2. The control method for a massage device for assisting sleep according to claim 1, characterized in that, The multimodal physiological signals in step S1 include at least body movement signals, heart rate variability signals, and respiratory rhythm signals.

3. The control method for a massage device for assisting sleep according to claim 1, characterized in that, In step S2, the sleep stability index integrates the coordination of body movement frequency, heart rate and breathing, as well as the smoothness of the transition between each stage.

4. The control method for a massage device for assisting sleep according to claim 1, characterized in that, The preprocessing of multimodal physiological signals in step S1 includes: By comparing and analyzing the phase change of the millimeter-wave echo signal with the charge change amplitude of the piezoelectric signal, interference data segments caused by large-amplitude body motion are identified and eliminated, resulting in anti-interference millimeter-wave phase sequence and piezoelectric charge sequence. The above anti-interference sequence was filtered and blind source separation was performed to initially separate the signal components of physiological information.

5. The control method for a massage device for assisting sleep according to claim 1, characterized in that, Step S2 specifically includes: S21. Input the preprocessed multimodal physiological signal sequence into a pre-trained deep neural network to generate context-aware feature vectors. S22. Input the feature vector into the classification layer and output the real-time probability distribution of the user in each sleep stage; S23. Based on the feature vector and stage probability, calculate the dynamic time regularization distance between the current feature and recent historical features and the transition entropy of the stage probability sequence, and integrate the above indicators and stage prior values ​​to normalize them into a sleep stability index.

6. The control method for a massage device for assisting sleep according to claim 5, characterized in that, The deep neural network adopts a hybrid architecture of bidirectional recurrent neural network and attention mechanism.

7. The control method for a massage device for assisting sleep according to claim 1, characterized in that, The process of generating the optimal massage control command in step S3 specifically includes: S31. Based on the current dominant sleep stage and stability index, generate a multi-dimensional intervention demand vector through a demand calculation function; S32. Query the intervention strategy knowledge base using the current state tuple, and dynamically synthesize a preliminary massage strategy; S33. Input the preliminary strategy into the effect prediction model to predict the changes in physiological state after its execution. Under safety constraints, with the goal of maximizing prediction stability and the probability of deep sleep, fine-tune the strategy parameters to obtain optimized massage instructions and execute them.

8. A method for controlling a massage device for assisting sleep according to claim 7, characterized in that, In step S32, the dynamic synthesis of the preliminary massage strategy employs nearest neighbor matching or a neural network strategy generator.

9. A control method for a massage device for assisting sleep according to claim 1, characterized in that, The process of continuously optimizing the sleep analysis model and massage decision-making strategy in step S4 specifically includes: S41. Record the complete experience data package for each intervention, including: pre-intervention state, performed actions, post-intervention state transition, and immediate reward calculated from physiological changes; S42. Combining the user's subjective sleep quality score the following morning, the accumulated experience data package is periodically used to update the policy network and effect prediction model through reinforcement learning algorithm, and the sleep analysis model can be fine-tuned optionally. S43. Update the learned new strategy knowledge to the intervention strategy knowledge base and adjust the user's personalized parameters.

10. A control system for a massage device for assisting sleep, used to implement the control method for a massage device for assisting sleep as described in any one of claims 1-9, characterized in that, include: The signal acquisition and preprocessing module includes a millimeter-wave radar sensor and a piezoelectric thin film sensor, which are used to acquire the user's raw physiological signals and preprocess them to remove body motion interference. The sleep state analysis module has a built-in deep learning model, which is used to receive pre-processed physiological signals, calculate and output the user's sleep stage probability distribution and sleep stability index in real time. The adaptive decision-making module integrates an intervention strategy knowledge base and an effect prediction model, which is used to generate and output optimized massage control instructions based on the probability distribution of the sleep stages and the sleep stability index. The massage execution module includes multiple independently controllable miniature massage actuators for executing the massage control commands; The user interaction and learning module is used to provide a user interface and collect subjective sleep quality scores. Through reinforcement learning algorithms, it continuously optimizes the sleep state analysis module and the adaptive decision-making module by utilizing changes in physiological signals before and after intervention, sleep stage evolution data, and the subjective scores. The central controller is used to coordinate and control the operation of the above modules.

Images