Bath chair self-adaptive adjusting system based on multi-mode sensing

Abstract

The invention relates to the technical field of health care, and discloses a bath chair self-adaptive adjusting system based on multi-modal sensing. According to the system, a pressure sensor array, an infrared thermal imager and an inertial measurement unit are used for collecting multi-dimensional real-time data, and the multi-modal sensing fusion module is used for processing the multi-dimensional real-time data to generate a structured physiological feature vector. The user intention recognition module predicts user behavior intention and gives a comfort level score on the basis, and the posture self-adaptive adjustment module optimizes seat supporting curved surface parameters according to the comfort level score to achieve real-time adjustment. In addition, the safety monitoring module evaluates the slip risk, the closed-loop control module guarantees safety, the online learning module optimizes the adjustment strategy, and the abnormal response module deals with the emergency situation. According to the system, the comfort and safety of the bath chair are improved, personalized adjustment is achieved, and the system has high practical value.

Description

Technical Field

[0001] This invention relates to the field of smart home and health care technology, specifically to an adaptive adjustment system for a bath chair based on multimodal sensing. Background Technology

[0002] In modern life, bathing is an important daily activity for cleaning and relaxation, and it is crucial for people's physical and mental health. However, traditional bath chairs have many design shortcomings and fail to meet the diverse needs of users, especially in terms of comfort and safety, where they face serious challenges.

[0003] From a comfort perspective, traditional shower chairs typically have a fixed shape and cannot be adjusted to suit different users' body characteristics and sitting habits. Individuals vary significantly in height, weight, and body type, making it difficult to obtain ideal support when using a fixed-shape shower chair. For example, taller users may find the seat depth insufficient, leaving their legs without adequate support; while heavier users may feel cramped due to insufficient seat width. Furthermore, users' postures during showering are not static; they may adjust their sitting position due to fatigue over time. However, traditional shower chairs cannot detect these changes and cannot adjust the seat posture in a timely manner, leading to a continuous decline in user comfort and potentially causing muscle fatigue and physical discomfort.

[0004] In terms of safety, the unique characteristics of the bathing environment, such as slippery floors and confined spaces, increase the risk of falls. Traditional bath chairs lack effective safety monitoring mechanisms and cannot detect the user's movement or changes in the surrounding environment in real time. If a user slips or loses balance while bathing, the bath chair cannot respond in time, easily leading to serious injuries. This safety hazard is particularly prominent for the elderly, those with mobility issues, and those who are physically weak. According to relevant statistics, falls during bathing account for a considerable proportion of household accidents, placing a heavy burden on individuals and families.

[0005] Furthermore, with the continuous advancement of technology and people's increasing demands for quality of life, smart home products have gradually become a market hotspot. In the field of bathing equipment, the demand for intelligent and personalized products is also growing stronger. However, the level of intelligence in current bath chairs on the market is generally low, with most only possessing simple functions such as heating and massage, lacking adaptive adjustment functions based on the user's real-time status. They cannot integrate with modern smart home systems and fail to meet users' pursuit of a convenient and intelligent lifestyle. Summary of the Invention

[0006] The purpose of this invention is to provide an adaptive adjustment system for a bath chair based on multimodal sensing, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an adaptive adjustment system for a bath chair based on multimodal sensing, the system comprising:

[0008] The multimodal sensing fusion module is used to collect multi-dimensional real-time data on user sitting posture, body surface temperature distribution and motion state through pressure sensor array, infrared thermal imager and inertial measurement unit, and generate structured physiological feature vectors.

[0009] The user intent recognition module is used to input the structured physiological feature vector into a deep neural network model based on a multi-head attention mechanism, and combine it with a temporal sliding window segmentation and feature cross-fusion algorithm to output user comfort score and behavioral intent prediction results.

[0010] The posture adaptive adjustment module is used to construct a joint angle optimization model based on a dynamic weighted graph convolutional network according to the predicted behavioral intention, solve for the ergonomic optimal support surface parameters through a multi-objective differential evolution algorithm, and generate real-time adjustment commands.

[0011] Preferably, the deep neural network model based on the multi-head attention mechanism includes:

[0012] The structured physiological feature vector is divided into time windows, and the pressure distribution entropy value and temperature gradient change within each window are extracted.

[0013] A self-attention mechanism is applied to model cross-channel feature correlations in sensor data from different modalities;

[0014] Temporal dependencies are captured by a gated recurrent unit, and spatial and temporal information is fused by a feature cross-layer to generate a discrete probability distribution of user behavioral intent.

[0015] Preferably, the step of solving for the optimal ergonomic support surface parameters using a multi-objective differential evolution algorithm includes:

[0016] Define a dual-objective cost function for the curvature of the support surface and the joint torque;

[0017] A population initialization strategy based on Pareto dominance is constructed, and an adaptive mutation operator is introduced to balance global exploration and local exploitation;

[0018] The optimal parameter combination that meets the muscle fatigue constraint is selected from the non-dominated solution set using the entropy weight method.

[0019] Preferably, the system further includes:

[0020] The safety monitoring module is used to extract point cloud features of the surrounding environment of the shower chair using a lightweight 3D point cloud processing network, and calculate the probability of slippage risk by combining it with a historical fall event database to generate a dynamic safety assessment matrix.

[0021] The closed-loop control module is used to design an adaptive fuzzy PID controller based on the dynamic safety assessment matrix to dynamically adjust the motor drive parameters and output a multi-joint cooperative motion control signal.

[0022] Preferably, the calculation of the slip risk probability includes:

[0023] Voxel grid downsampling was performed on the point cloud data around the shower chair to extract the ground normal vector distribution features;

[0024] The normal vector distribution features are dynamically time-warped and matched with the motion trajectories of historical fall events to calculate the slip risk similarity.

[0025] By fusing similarity and user weight distribution data through a Bayesian network, a dynamic risk probability value is output.

[0026] Preferably, the adaptive fuzzy PID controller dynamically adjusts the motor drive parameters, including:

[0027] The input variables are defined as the slip risk level and the attitude deviation, and the output variable is the PID gain coefficient adjustment.

[0028] Construct a fuzzy rule base for describing input variables using triangular membership functions;

[0029] The area averaging method is used to defuzzify the motor and generate precise motor speed and torque control commands.

[0030] Preferred methods for generating structured physiological feature vectors include:

[0031] Spatiotemporal alignment of multimodal sensing data, including timestamp synchronization of pressure sensor signals and spatial registration of thermal imaging data;

[0032] A robust principal component analysis algorithm is used to separate sparse noise and low-rank background components from sensor data.

[0033] The maximum information coefficient is used to select features from the denoised data, and a multidimensional feature tensor is constructed and subjected to hierarchical normalization.

[0034] Preferably, the system further includes:

[0035] The online learning module is used to collect user feedback data and employs a meta-reinforcement learning algorithm to incrementally optimize the posture adjustment strategy.

[0036] Preferably, the incremental optimization of the meta-reinforcement learning algorithm includes:

[0037] Construct a dual-channel architecture of policy network and value network, and design a reward function based on inverse dynamics model;

[0038] The network parameters are updated by optimizing the algorithm through a near-end policy, and the exploration and utilization are balanced by using an experience replay buffer pool.

[0039] Preferably, the system further includes:

[0040] Anomaly response module, used to identify sudden attitude loss events based on a hierarchical threshold detection algorithm;

[0041] The emergency braking protocol is triggered through an event-driven architecture, generating a seat locking command and activating an emergency alarm signal.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] In enhancing user comfort, the system's multimodal sensor fusion module plays a crucial role. The pressure sensor array, like countless "tactile neurons," accurately senses the pressure distribution between different parts of the body and the seat, keenly detecting subtle movements and postural changes. The infrared thermal imager acts like a "temperature detector," monitoring surface temperature distribution and determining whether prolonged localized pressure has led to poor blood circulation or abnormal temperature. The inertial measurement unit functions as a "motion tracker," capturing the user's movement in real time. The structured physiological feature vectors generated from this multidimensional data provide a reliable basis for subsequent precise adjustments.

[0044] Based on this, the user intent recognition module uses advanced algorithms for in-depth analysis, outputting user comfort scores and behavioral intent prediction results. For example, when a user gradually feels fatigued during a bath and unconsciously adjusts their posture, the module can promptly capture these signals. After receiving the prediction results, the posture adaptive adjustment module quickly constructs a joint angle optimization model, solves for the optimal ergonomic support surface parameters, and allows the seat back, seat, and other parts to adjust in real time according to the user's needs, providing just the right amount of support to the body, effectively relieving muscle fatigue, and greatly improving comfort during the bathing process.

[0045] In terms of safety performance, the safety monitoring module utilizes a lightweight 3D point cloud processing network to quickly extract point cloud features of the surrounding environment of the shower chair. Through voxel grid downsampling and ground normal vector distribution feature extraction, it accurately analyzes the ground conditions to determine whether there are dangerous situations such as slipperiness or unevenness. Combining a historical fall event database, it uses dynamic time warping matching and Bayesian networks to accurately calculate the probability of slippage. Once a risk is detected, the adaptive fuzzy PID controller of the closed-loop control module responds quickly, dynamically adjusts the motor drive parameters, and outputs multi-joint coordinated motion control signals to stabilize the seat in a timely manner, preventing the user from slipping and effectively ensuring the user's personal safety.

[0046] In terms of system self-optimization and personalized customization, the online learning module collects user feedback data and uses meta-reinforcement learning algorithms to incrementally optimize the posture adjustment strategy. As the number of times a user uses the system increases, it continuously learns the user's preferences and habits, making the adjustments more tailored to individual needs and achieving truly personalized service. The anomaly response module, based on a hierarchical threshold detection algorithm, can quickly identify sudden posture loss of control events. Through an event-driven architecture, it triggers an emergency braking protocol, locks the seat, and activates emergency alarm signals, adding a solid layer of protection for user safety.

[0047] From a smart home ecosystem perspective, this system can interconnect with other smart home devices and integrate into the overall smart home system. Users can control the shower chair and other devices through smart terminals, enabling intelligent linkage of living scenarios, improving the overall intelligence and convenience of the home environment, and creating a more comfortable, safe, and intelligent living experience for users. Attached Figure Description

[0048] Figure 1 This is a schematic diagram illustrating the working principle of the multimodal sensing-based adaptive adjustment system for a bath chair as described in this invention.

[0049] Figure 2 This is a diagram illustrating the working principle of a deep neural network model based on multi-head attention mechanism.

[0050] Figure 3 A step-by-step diagram for solving the optimal ergonomic support surface parameters using a multi-objective differential evolution algorithm;

[0051] Figure 4 A diagram illustrating the steps involved in generating structured physiological feature vectors. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Please see Figures 1-4 This invention provides an adaptive adjustment system for a shower chair based on multimodal sensing, aiming to improve user comfort and safety during the showering process. The specific implementation of this system is described in detail below.

[0054] The system mainly consists of a multimodal sensing fusion module, a user intent recognition module, and a posture adaptive adjustment module. First, the multimodal sensing fusion module collects multi-dimensional real-time data on the user's sitting posture, body surface temperature distribution, and movement state through a pressure sensor array, an infrared thermal imager, and an inertial measurement unit, generating a structured physiological feature vector. Specifically, the pressure sensor array is positioned on key parts of the shower chair that come into contact with the human body, such as the seat and backrest, to acquire pressure information from various parts of the user's body against the seat; the infrared thermal imager is installed in a suitable location to capture the user's body surface temperature distribution in real time; and the inertial measurement unit monitors the user's movement state, such as whether there are any body movements like swaying or twisting.

[0055] After acquiring multimodal data, the multimodal sensor fusion module processes the data. First, it performs spatiotemporal alignment on the multimodal sensor data, including timestamp synchronization of pressure sensor signals and spatial registration of thermal imaging data, ensuring temporal and spatial consistency of data acquired by different sensors. Next, an improved robust principal component analysis algorithm is used to separate sparse noise and low-rank background components from the sensor data, removing noise interference. Then, the denoised data is used for feature selection with the maximum information coefficient, constructing a multidimensional feature tensor and performing hierarchical normalization to finally generate a structured physiological feature vector.

[0056] The user intent recognition module inputs structured physiological feature vectors into a deep neural network model based on a multi-head attention mechanism. Combining temporal sliding window segmentation and feature cross-fusion algorithms, it outputs user comfort scores and behavioral intent predictions. In this process, the model first divides the structured physiological feature vectors into time windows, extracting the pressure distribution entropy and temperature gradient changes within each window. Then, a self-attention mechanism is applied to model cross-channel feature correlations of sensor data from different modalities, uncovering potential connections between different sensor data. Next, a gated recurrent unit captures temporal dependencies, and a feature cross-fusion layer fuses spatial and temporal information to generate a discrete probability distribution of user behavioral intent, yielding user comfort scores and behavioral intent predictions.

[0057] The posture adaptive adjustment module constructs a joint angle optimization model based on a dynamically weighted graph convolutional network based on the behavioral intention prediction results. It then uses a multi-objective differential evolution algorithm to solve for the ergonomically optimal support surface parameters and generates real-time adjustment commands. The module first defines a bi-objective cost function for the support surface curvature and joint torque to measure the rationality of the seat's support for the human body under different parameter combinations. Next, a population initialization strategy based on Pareto dominance is constructed, introducing an adaptive mutation operator to balance global exploration and local exploitation, searching for a better solution among numerous possible parameter combinations. Finally, the entropy weight method is used to select the optimal parameter combination that meets the muscle fatigue constraint from the non-dominated solution set, obtaining the ergonomically optimal support surface parameters, and then generating real-time adjustment commands to drive the corresponding components of the shower chair to adjust, providing the user with more comfortable support.

[0058] The technical solution of the present invention will be further illustrated below through specific embodiments.

[0059] Example 1:

[0060] In this embodiment, the specific implementation process of the deep neural network model based on the multi-head attention mechanism and the detailed method for generating structured physiological feature vectors are further described.

[0061] To generate structured physiological feature vectors, consider data collected by a pressure sensor array, an infrared thermal imager, and an inertial measurement unit. During the spatiotemporal alignment stage, each pressure sensor signal carries its own timestamp. Hardware clock synchronization technology or software time calibration algorithms are used to precisely synchronize the timestamps of all pressure sensor signals, ensuring the temporal consistency of the acquired pressure data. For spatial registration of thermal imaging data, specific markers pre-set on the shower chair are used, combined with feature matching algorithms in image processing, such as SIFT (Scale Invariant Feature Transform), to correspond the human body parts in the thermal imaging data to their actual positions on the chair, achieving spatial registration.

[0062] When separating data components using an improved robust principal component analysis (PCA) algorithm, the collected sensor data is treated as a matrix. The algorithm decomposes this matrix into a low-rank matrix and a sparse matrix. The low-rank matrix represents the main features of the data, i.e., the low-rank background component, while the sparse matrix contains outliers such as noise, i.e., sparse noise. In this way, noise in the data is effectively removed, improving data quality.

[0063] When using the maximum information coefficient for feature selection, the maximum information coefficient between each feature and other features is calculated, and features with larger coefficients are selected, as these features contain more effective information. When constructing a multidimensional feature tensor, the selected features are arranged in a specific dimensional order to form a tensor structure. Then, hierarchical normalization is performed, normalizing features at different levels separately to ensure the data is processed on the same scale.

[0064] In deep neural network models based on multi-head attention mechanisms, when dividing structured physiological feature vectors into time windows, the window size can be set according to the actual situation, for example, set to 1 second, meaning that data from every second constitutes a window. Within each window, the pressure distribution entropy value is calculated. The entropy value measures the degree of disorder in the pressure distribution; the more uniform the pressure distribution, the larger the entropy value; conversely, the smaller the entropy value. The temperature gradient change is obtained by calculating the difference between temperature values ​​at adjacent time points, reflecting the trend of temperature change.

[0065] When applying the self-attention mechanism, for sensor data of different modalities, such as pressure and temperature data, the importance of different data in feature association is determined by calculating their attention weights. The gated recurrent unit selectively retains or updates information in the time series through a gating mechanism, effectively capturing temporal dependencies. The feature cross-layer fuses features in the spatial dimension (such as pressure and temperature features at different locations) and features in the temporal dimension (such as changes in features at different times), ultimately generating a discrete probability distribution of user behavioral intent, providing an accurate basis for subsequent user intent recognition.

[0066] Example 2:

[0067] This embodiment details the specific steps for solving the optimal ergonomic support surface parameters using a multi-objective differential evolution algorithm.

[0068] When defining the bi-objective cost function of support surface curvature and joint torque, let the support surface curvature be... The joint torque is Support surface curvature This reflects the curvature of the seat support surface, which significantly impacts the support provided to the human body. Excessive or insufficient curvature can lead to discomfort. Joint torque. This represents the torque exerted on a human joint while maintaining a certain sitting posture. Excessive torque can easily lead to muscle fatigue. The bi-objective cost function can be expressed as: ,in and This is a weighting coefficient used to balance the importance of the two objectives, and can be adjusted according to actual needs and experimental results.

[0069] When constructing a population initialization strategy based on Pareto dominance relations, a set of initial support surface parameter combinations is randomly generated as the population. For any two individuals in the population... and If an individual It is not inferior to the individual in all targets (support surface curvature and joint torque). And superior to individuals in at least one objective. Then it is called an individual Dominant Individual Through this dominance relationship, non-dominant individuals are selected to form the initial population.

[0070] An adaptive mutation operator is introduced to balance global exploration and local exploitation. During evolution, the mutation operator is used to generate new individuals. The adaptive mutation operator adjusts the mutation magnitude based on the population's evolutionary state. When the population gets stuck in a local optimum during global search, the mutation magnitude is increased to encourage individuals to explore new solution spaces over a larger area; when the population gradually approaches the optimal solution, the mutation magnitude is decreased to perform a more refined local search and improve solution accuracy.

[0071] The entropy weight method is used to screen the optimal parameter combination that meets the muscle fatigue constraint from the non-dominated solution set. The entropy weight method is a method that determines weights based on the degree of variation of indicator data. For each solution in the non-dominated solution set, the entropy value of each indicator (support surface curvature and joint torque) is calculated. The smaller the entropy value, the greater the degree of variation of that indicator in the solution set, the greater its influence on the decision, and the higher its corresponding weight. This is combined with the muscle fatigue constraint, such as setting an upper limit value for joint torque. Screening for joints with torque less than Furthermore, the optimal combination of parameters with the best overall weight is used as the optimal support surface parameter for ergonomics.

[0072] Example 3:

[0073] The safety monitoring module utilizes a lightweight 3D point cloud processing network to extract point cloud features of the environment surrounding the shower chair. Combined with a historical fall event database, it calculates the probability of slippage risk and generates a dynamic safety assessment matrix. In practical applications, depth cameras installed around the shower chair acquire point cloud data of the surrounding environment. The lightweight 3D point cloud processing network processes this point cloud data, extracting geometric and density features. For example, by calculating the normal vector distribution characteristics of the point cloud, the flatness of the ground can be determined; the more uniform the normal vector distribution, the flatter the ground; conversely, there may be slopes or uneven areas.

[0074] The extracted normal vector distribution features are dynamically time-warped and matched with the motion trajectories of historical fall events. The motion trajectories of historical fall events are stored in a historical fall event database. The dynamic time warping algorithm can find the optimal matching path between two time series (the normal vector distribution feature sequence and the historical fall event motion trajectory sequence) and calculate their similarity.

[0075] By fusing similarity and user weight distribution data using a Bayesian network, a dynamic risk probability value is output. A Bayesian network is a graphical model based on probabilistic reasoning. It calculates the probability of slippage risk based on known conditional probability relationships, combined with similarity and user weight distribution data. For example, if the similarity is high and the user weight distribution is uneven, the Bayesian network will output a higher slippage risk probability value.

[0076] The closed-loop control module, based on a dynamic safety assessment matrix, designs an adaptive fuzzy PID controller to dynamically adjust motor drive parameters and output multi-joint coordinated motion control signals. The input variables of the adaptive fuzzy PID controller are the slip risk level and the attitude deviation. The slip risk level is classified according to the risk probability value in the dynamic safety assessment matrix; for example, a risk probability value of 0-0.3 indicates a low risk level, 0.3-0.7 indicates a medium risk level, and 0.7-1 indicates a high risk level. The attitude deviation is obtained by calculating the difference between the current seat posture and the ideal posture.

[0077] The output variable is the PID gain coefficient adjustment. A fuzzy rule base describing the input variables is constructed using triangular membership functions. For example, for slip risk levels, when the risk level is low, the triangular membership function has a high membership degree in the low-risk region; when the risk level is medium, it has a high membership degree in the medium-risk region. The area averaging method is used to defuzzify the data. Based on the results of fuzzy inference, precise motor speed and torque control commands are calculated, thereby achieving precise control of the shower chair motor and ensuring user safety during the showering process.

[0078] Example 4:

[0079] This embodiment details the specific process of calculating the slip risk probability.

[0080] Voxel downsampling is performed on the point cloud data around the shower chair to reduce data volume and highlight key features. The point cloud data is divided into small voxel grids, assuming the side length of the voxel grid is set to... ,For example Meters. Within each voxel grid, the orientation information of the point cloud is obtained by calculating its normal vector. The normal vector is a vector perpendicular to the surface of the point cloud, reflecting the surface orientation of that local area. For each voxel grid, the mean, standard deviation, and other statistics of the point cloud normal vectors are calculated to serve as the characteristics of the ground normal vector distribution. The mean indicates the approximate direction of the normal vector in that area, while the standard deviation reflects the dispersion of the normal vector direction; the greater the dispersion, the more uneven the ground.

[0081] The extracted normal vector distribution features are dynamically time-warped (DTW) matched with the motion trajectories of historical fall events. These historical fall trajectories are obtained by monitoring and recording fall events in real-world scenarios, containing data on the changes in position and posture of the human body over time during the fall. The core idea of ​​the DTW algorithm is to find the optimal matching path between two time series (here, the normal vector distribution feature sequence and the historical fall event trajectory sequence), making them as similar as possible in time and shape. During the matching process, the similarity between corresponding elements of the two sequences is measured by calculating the distance, such as Euclidean distance. The calculated similarity reflects the degree of similarity between the current environment surrounding the bath chair and the historical fall scenario.

[0082] By fusing similarity and user weight distribution data using a Bayesian network, a dynamic risk probability value is output. A Bayesian network is a graphical model based on probabilistic reasoning, consisting of nodes and directed edges. Nodes represent random variables, and edges represent dependencies between variables. In this embodiment, the nodes of the Bayesian network include nodes representing slip risk probability, nodes representing similarity, and nodes representing user weight distribution data. Conditional probability relationships between nodes are learned through extensive historical data; that is, the probability distribution of other node values ​​given the values ​​of certain nodes. When the currently calculated similarity and user weight distribution data (such as the user's center of gravity position on the seat) are input, the Bayesian formula is applied... (in Indicates a slippage risk event. (Including evidence such as similarity and user weight distribution data), the probability value of slip risk is calculated. This value is the dynamic risk probability value, which is used to assess the likelihood of a user slipping and falling in the current bathing scenario.

[0083] Example 5:

[0084] The adaptive fuzzy PID controller defines the input variables as the slip risk level and the attitude deviation, and the output variable as the PID gain coefficient adjustment. The slip risk level is determined based on the dynamic risk probability value calculated in Example 4. For example, the dynamic risk probability value is divided into three intervals: when the probability value is within... When defined as a low-risk level; At that time, it was classified as a medium-risk level; At this time, it is considered a high-risk level. The posture deviation is obtained by using sensors to measure in real time the difference between the current posture of the bath chair (such as the seat back angle, seat tilt, etc.) and the preset ideal posture.

[0085] A fuzzy rule base describing the input variables is constructed using triangular membership functions. Taking the low-risk level as an example, a triangular membership function is constructed for the slip risk level. Assuming This indicates the numerical value corresponding to the slip risk level, when The lower limit of the low-risk level range (like )hour, ;when Reaching the midpoint of the interval (like )hour, ;when Reaching the upper limit of the interval (like )hour, The function is triangular in shape. Thus, based on the input slip risk level value, its membership degree in the low-risk fuzzy set can be calculated. Similarly, for the attitude deviation, a similar triangular membership function is constructed to represent fuzzy sets such as small deviation, medium deviation, and large deviation, and its membership degree in the corresponding fuzzy set is determined based on the magnitude of the attitude deviation.

[0086] The area averaging method is used for defuzzification to generate precise motor speed and torque control commands. The area averaging method is a commonly used defuzzification method. Its principle is to average the areas of the membership function graphs corresponding to the fuzzy output set obtained from fuzzy inference to obtain a precise value, which is the adjustment amount of the PID gain coefficient. Assume the fuzzy output set is... Its membership function is The output value after defuzzification Through formula The calculations are as follows: Based on the adjusted PID gain coefficient and combined with the control model of the shower chair motor, the motor speed and torque control commands are calculated, thereby achieving precise control of the shower chair motor and ensuring that the shower chair can adjust its posture in a timely manner according to the current risk situation and posture deviation, thus ensuring the safety and comfort of the user.

[0087] Example 6:

[0088] The online learning module collects user feedback data and uses a meta-reinforcement learning algorithm to incrementally optimize the posture adjustment strategy. In practical applications, by setting up a feedback interface on the shower chair, users can directly evaluate the comfort of the seat, such as "comfortable," "average," or "uncomfortable," and this evaluation information serves as user feedback data. Simultaneously, it can also collect data on users' manual adjustments to the seat during use, such as the number and extent of adjustments made to the backrest angle and seat height.

[0089] When employing meta-reinforcement learning algorithms, a dual-channel architecture of policy network and value network is constructed. The policy network is used to generate the posture adjustment strategy for the bath chair. Its input is the current state information of the bath chair (such as data collected by sensors on the user's sitting posture, body surface temperature distribution, and motion state), and its output is different actions (such as adjusting the seat back angle to increase...). Degree, reduce seat height The probability distribution (e.g., centimeters, etc.) is used. A value network is used to evaluate the effectiveness of a strategy; the input is the current state information of the bath chair, and the output is a value estimate of executing a certain strategy in the current state. A reward function based on an inverse dynamics model is designed. The inverse dynamics model calculates the required force and torque based on the current state and desired state of the bath chair. The reward function design is based on several factors, such as the closeness of the seat posture to the user's comfortable posture. It is assumed that the seat back angle corresponding to the user's comfortable posture is... The current seat back angle is Therefore, the reciprocal of the absolute value of the difference between the two can be used as part of the reward, that is... (when hour, We take a large positive value, such as 10); we also consider the smoothness of the movement, and give a negative reward when the movement changes too much. The total reward value is obtained by combining these factors.

[0090] Network parameters are updated using a proximal policy optimization algorithm, which limits the magnitude of policy network updates to prevent performance degradation due to excessively large updates. An experience replay buffer is used to balance exploration and utilization. The experience replay buffer stores past experience data, including states, actions, rewards, and the next state. During training, data is randomly drawn from the experience replay buffer for learning, thus avoiding continuous learning of similar data and improving the stability and efficiency of learning.

[0091] The anomaly response module identifies sudden attitude loss events based on a hierarchical threshold detection algorithm. This algorithm sets multiple levels of thresholds, for example, a primary threshold. and secondary threshold ,and When the posture data detected by the sensors (such as the rate of change of the seat's tilt angle, the shift of the user's center of gravity, etc.) exceeds the primary threshold... Initially, it was determined that there might be an abnormal posture; then, further checks were conducted to see if the data exceeded a secondary threshold. If the threshold is exceeded, it is determined that a sudden loss of posture has occurred.

[0092] An event-driven architecture triggers an emergency braking protocol. Upon detecting an abnormal event, the event-driven architecture rapidly activates the emergency braking protocol. This protocol generates a seat locking command, stopping the shower chair's motor and fixing the seat's position and posture to prevent the user from falling. Simultaneously, emergency alarm signals are activated, such as a loud alarm sounding through a speaker and flashing lights, to notify the user and surrounding personnel of the abnormal situation and allow for timely measures to ensure the user's safety.

[0093] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0094] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A shower chair adaptive adjustment system based on multimodal sensing, characterized in that, The system includes: The multimodal sensing fusion module is used to collect multi-dimensional real-time data on user sitting posture, body surface temperature distribution and motion state through pressure sensor array, infrared thermal imager and inertial measurement unit, and generate structured physiological feature vectors. The user intent recognition module is used to input the structured physiological feature vector into a deep neural network model based on a multi-head attention mechanism, and combine it with a temporal sliding window segmentation and feature cross-fusion algorithm to output user comfort score and behavioral intent prediction results. The posture adaptive adjustment module is used to construct a joint angle optimization model based on a dynamic weighted graph convolutional network according to the predicted behavioral intention, solve for the ergonomic optimal support surface parameters through a multi-objective differential evolution algorithm, and generate real-time adjustment commands.

2. The adaptive adjustment system for a shower chair based on multimodal sensing as described in claim 1, characterized in that, The deep neural network model based on the multi-head attention mechanism includes: The structured physiological feature vector is divided into time windows, and the pressure distribution entropy value and temperature gradient change within each window are extracted. A self-attention mechanism is applied to model cross-channel feature correlations in sensor data from different modalities; Temporal dependencies are captured by a gated recurrent unit, and spatial and temporal information is fused by a feature cross-layer to generate a discrete probability distribution of user behavioral intent.

3. The adaptive adjustment system for a shower chair based on multimodal sensing as described in claim 2, characterized in that, The process of solving for the optimal ergonomic support surface parameters using a multi-objective differential evolution algorithm includes: Define a dual-objective cost function for the curvature of the support surface and the joint torque; A population initialization strategy based on Pareto dominance is constructed, and an adaptive mutation operator is introduced to balance global exploration and local exploitation; The optimal parameter combination that meets the muscle fatigue constraint is selected from the non-dominated solution set using the entropy weight method.

4. The adaptive adjustment system for a shower chair based on multimodal sensing as described in claim 1, characterized in that, The system also includes: The safety monitoring module is used to extract point cloud features of the surrounding environment of the shower chair using a lightweight 3D point cloud processing network, and to calculate the probability of slippage risk by combining it with a historical fall event database, thereby generating a dynamic safety assessment matrix. The closed-loop control module is used to design an adaptive fuzzy PID controller based on the dynamic safety assessment matrix to dynamically adjust the motor drive parameters and output a multi-joint cooperative motion control signal.

5. The adaptive adjustment system for a shower chair based on multimodal sensing as described in claim 4, characterized in that, The calculation of the slip risk probability includes: Voxel grid downsampling was performed on the point cloud data around the shower chair to extract the ground normal vector distribution features; The normal vector distribution features are dynamically time-warped and matched with the motion trajectories of historical fall events to calculate the slip risk similarity. By fusing similarity and user weight distribution data through a Bayesian network, a dynamic risk probability value is output.

6. The adaptive adjustment system for a shower chair based on multimodal sensing as described in claim 5, characterized in that, The adaptive fuzzy PID controller dynamically adjusts the motor drive parameters, including: The input variables are defined as the slip risk level and the attitude deviation, and the output variable is the PID gain coefficient adjustment. Construct a fuzzy rule base for describing input variables using triangular membership functions; The area averaging method is used to defuzzify the motor and generate precise motor speed and torque control commands.

7. The adaptive adjustment system for a shower chair based on multimodal sensing as described in claim 1, characterized in that, Methods for generating structured physiological feature vectors include: Spatiotemporal alignment of multimodal sensing data, including timestamp synchronization of pressure sensor signals and spatial registration of thermal imaging data; A robust principal component analysis algorithm is used to separate sparse noise and low-rank background components from sensor data. The maximum information coefficient is used to select features from the denoised data, and a multidimensional feature tensor is constructed and subjected to hierarchical normalization.

8. The adaptive adjustment system for a shower chair based on multimodal sensing as described in claim 1, characterized in that, Also includes: The online learning module is used to collect user feedback data and employs a meta-reinforcement learning algorithm to incrementally optimize the posture adjustment strategy.

9. The adaptive adjustment system for a shower chair based on multimodal sensing as described in claim 8, characterized in that, The incremental optimization of the meta-reinforcement learning algorithm includes: Construct a dual-channel architecture of policy network and value network, and design a reward function based on inverse dynamics model; The network parameters are updated by optimizing the algorithm through a near-end policy, and the exploration and utilization are balanced by using an experience replay buffer pool.

10. The adaptive adjustment system for a shower chair based on multimodal sensing as described in any one of claims 1 to 9, characterized in that, The system also includes: Anomaly response module, used to identify sudden attitude loss events based on a hierarchical threshold detection algorithm; The emergency braking protocol is triggered through an event-driven architecture, generating a seat locking command and activating an emergency alarm signal.

Images