Multi-modal sensor adaptive adjustment method and device, electronic device, and storage medium

Abstract

The application discloses a multi-modal sensing adaptive adjustment method and device, electronic equipment and storage medium, comprising: acquiring multi-modal sensing data including pressure distribution data and attitude angle data collected by sensing components distributed on a bearing main body; determining a current sitting posture category of a target object for the bearing main body through a preset sitting posture classification model according to the data; fusing historical preference data of the target object, current attitude change trend characteristics and pressure distribution change direction characteristics to infer an intended sitting posture category of the target object; querying a preset attitude mapping table to obtain target attitude parameters according to the intended sitting posture category; generating an adjustment track based on the target attitude parameters and a preset speed curve model, and instructing a driving mechanism to perform corresponding attitude adjustment. The application introduces a multi-dimensional intention inference mechanism, effectively filters the interference caused by the transitional attitude, improves the accuracy of the system attitude adjustment, and realizes the closed-loop adaptive optimization of the control parameters.

Description

Technical Field

[0001] This invention relates to the fields of data processing and automatic control technology, and more specifically, to a multimodal sensing adaptive adjustment method, device, electronic device, and storage medium. Background Technology

[0002] With the development of intelligent control technology, load-bearing entities (such as adjustable support equipment) equipped with electric multi-dimensional adjustment mechanisms are becoming increasingly common. Existing attitude adaptive adjustment methods typically collect pressure distribution data of a specified area through sensors, and then directly issue control commands based on the identified current force posture to drive the actuator to adjust the angle and position of the components.

[0003] However, existing technologies have the following technical shortcomings in practical applications: First, existing solutions employ a linear response logic of "current posture is desired posture," lacking an active inference mechanism for the intention of posture changes. When the target object is in a transitional posture or an unconscious temporary movement, existing systems are prone to misidentifying non-target postures as desired postures and executing adjustments, leading to frequent erroneous adjustments and reducing the accuracy and robustness of system control.

[0004] Second, existing solutions lack quantitative assessment mechanisms and closed-loop feedback mechanisms that are linked to human physiological health indicators. The system cannot assess potential physical strain risks based on local stress time and posture change frequency, nor can it extract feedback signals from the target object's secondary manual correction behavior in response to the adjustment results. This leads to the long-term stagnation of the underlying control strategy and the inability to achieve continuous optimization of personalized parameters. Summary of the Invention

[0005] The purpose of this invention is to provide a multimodal sensing adaptive adjustment method, device, electronic device and storage medium to solve the technical problems in the prior art that the lack of intent inference leads to low adjustment accuracy and the lack of closed-loop learning mechanism leads to rigid adjustment strategy.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The first aspect of this invention provides a multimodal sensing adaptive adjustment method applied to a data processing device. The method includes: acquiring multimodal sensing data collected by sensing components distributed on a support body; wherein the multimodal sensing data includes pressure distribution data collected by a pressure sensing matrix and attitude angle data collected by an inertial measurement unit; determining the current sitting posture category of a target object relative to the support body based on the pressure distribution data and the attitude angle data, using a preset sitting posture classification model; inferring the target object's intended sitting posture category based on the current sitting posture category by fusing the target object's historical preference data, current posture change trend characteristics determined based on the attitude angle data, and pressure distribution change direction characteristics determined based on the pressure distribution data; querying a preset posture mapping table according to the intended sitting posture category to obtain target posture parameters; and generating an adjustment trajectory based on the target posture parameters according to a preset velocity curve model, and instructing the drive mechanism of the support body to perform corresponding posture adjustment according to the adjustment trajectory.

[0007] Preferably, determining the current sitting posture category of the target object relative to the supporting body based on the pressure distribution data and the posture angle data using a preset sitting posture classification model includes: Calculate the principal tilt angle of the target object based on the attitude angle data; Based on the main tilt angle, determine the first-level sitting posture category of the target object in the preset sitting posture classification system; Within the first-level sitting posture category, feature vectors of the pressure distribution data are extracted; the feature vectors include regional pressure peak value, pressure center of gravity position, pressure distribution symmetry coefficient, and pressure distribution entropy value. The feature vector is identified by the sitting posture classification model to determine the current sitting posture category of the target object relative to the supporting body, and the classification confidence score is obtained.

[0008] Preferably, the step of inferring the target object's intended sitting posture category based on the current sitting posture category, by fusing the target object's historical preference data, the current posture change trend characteristics determined based on the posture angle data, and the pressure distribution change direction characteristics determined based on the pressure distribution data, includes: Based on the historical preference data, calculate the first conditional probability of the current sitting posture category shifting to each candidate intended sitting posture; Based on posture angle data from multiple consecutive sampling periods, a posture change trend vector is calculated; and based on the posture change trend vector, a second conditional probability is calculated of the target object tending to each candidate intentional sitting posture. Based on pressure distribution data from multiple consecutive sampling periods, a pressure distribution change vector is calculated; and based on the pressure distribution change vector, a third conditional probability is calculated for the target object's tendency toward each candidate intended sitting posture. The first conditional probability, the second conditional probability, and the third conditional probability are weighted and summed according to preset weighting coefficients to obtain the comprehensive probability of each candidate intended sitting posture. The candidate intended sitting posture corresponding to the highest comprehensive probability is determined as the intended sitting posture category of the target object.

[0009] Preferably, after the drive mechanism that instructs the supporting body to perform corresponding attitude adjustment according to the adjustment trajectory, the method further includes: Within a preset time window, acquire the manual correction signal triggered by the target object against the carrier; the manual correction signal includes the corrected final attitude parameters; The historical preference data of the target object is updated based on the difference between the target attitude parameters and the final attitude parameters; Based on the gradient direction corresponding to the difference, update the weight coefficients used when inferring the intended sitting posture category.

[0010] Preferably, the method further includes: Based on the current sitting posture category, query the preset sitting posture load mapping table to obtain the corresponding first load coefficient; The second load coefficient is calculated based on the duration parameter of the current sitting posture category and the posture change frequency parameter. The first load factor and the second load factor are weighted and calculated to obtain the strain risk value of the target object; If the risk value of strain exceeds a preset risk threshold, a health intervention signal is generated, and the drive mechanism is instructed to perform adjustments according to the safe posture parameters corresponding to the health intervention signal.

[0011] Preferably, the step of weighting the first load factor and the second load factor to obtain the strain risk value of the target object includes: Obtain the biometric attribute information of the target object; the biometric attribute information includes age range and body mass index; The first weight corresponding to the first load coefficient and the second weight corresponding to the second load coefficient are determined based on the biometric attribute information. Based on the first weight and the second weight, the first load coefficient and the second load coefficient are weighted and summed to obtain the strain risk value.

[0012] Preferably, after acquiring the multimodal sensing data collected by the sensing components distributed on the supporting body, the method further includes: Perform connected component analysis on the pressure distribution data to extract multiple independent pressure regions that exceed a preset activation threshold; Each independent pressure region is assigned a unique object identifier, and the attitude angle data is mapped and allocated according to the positional relationship of the independent pressure regions. For each individual object identifier, the processing steps from determining the current sitting posture category to inferring the intended sitting posture category are executed in parallel. Based on a preset control priority strategy, a multi-target adjustment trajectory is collaboratively generated for the carrier.

[0013] A second aspect of the present invention provides a multimodal sensing adaptive adjustment device, comprising: The data acquisition module is used to acquire multimodal sensing data collected by sensing components distributed on the supporting body; wherein, the multimodal sensing data includes pressure distribution data collected by the pressure sensing matrix and attitude angle data collected by the inertial measurement unit; The sitting posture recognition module is used to determine the current sitting posture category of the target object relative to the supporting body based on the pressure distribution data and the posture angle data and through a preset sitting posture classification model. The intention inference module is used to infer the target object's intention sitting posture category based on the current sitting posture category, by integrating the target object's historical preference data, the current posture change trend characteristics determined based on the posture angle data, and the pressure distribution change direction characteristics determined based on the pressure distribution data. The parameter determination module is used to query a preset posture mapping table based on the intended sitting posture category to obtain the target posture parameters; The trajectory adjustment module is used to generate an adjustment trajectory based on the target attitude parameters and a preset velocity curve model, and to instruct the drive mechanism of the supporting body to perform corresponding attitude adjustment according to the adjustment trajectory.

[0014] A third aspect of the present invention provides an electronic device, comprising: a processor and a memory; The memory stores computer programs that can be executed by the processor; When the processor executes the computer program, it implements the multimodal sensing adaptive adjustment method as described above.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the multimodal sensing adaptive adjustment method as described in any one of the preceding descriptions.

[0016] Beneficial effects Improved accuracy of posture adjustment: This invention introduces an intentional posture category inference mechanism between sensing the current sitting posture and executing posture adjustment. By fusing historical preference data, current posture change trend characteristics, and pressure distribution change direction characteristics for multi-dimensional probabilistic inference, it effectively filters out interference signals caused by transitional postures and temporary movements, avoiding the defect of using non-target postures as the basis for adjustment, and improving the system's accuracy in recognizing the true adjustment intention.

[0017] This invention achieves closed-loop adaptive optimization of the adjustment strategy: It utilizes the manual correction signal triggered by the target object as feedback input, extracts the difference between the target attitude parameters and the final attitude parameters, and updates the weight coefficients in the intention inference model based on the gradient direction corresponding to the difference. This closed-loop learning mechanism enables the data processing device to dynamically adjust control parameters according to usage behavior, overcoming the technical bottleneck that fixed models cannot adapt to individual differences.

[0018] This invention provides a proactive health intervention mechanism: by integrating a static first load coefficient and a dynamic second load coefficient, and combining them with biometric attribute information for weighted calculation, it achieves quantitative monitoring of strain risk values. When the risk exceeds the limit, it proactively generates a health intervention signal and executes safe posture parameter adjustments, reducing the risk of physiological strain caused by prolonged improper stress posture in the target object.

[0019] Improved smoothness of drive adjustment: Before issuing control commands, this invention generates a continuous adjustment trajectory for the target attitude parameters through a preset speed curve model, and implements strict kinematic constraints on acceleration and jerk during operation, reducing mechanical impact when multiple motors work together and ensuring the smoothness of attitude transition. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a main flowchart of a multimodal sensing adaptive adjustment method provided in an embodiment of the present invention; Figure 2 This is a first sub-flowchart of a multimodal sensing adaptive adjustment method provided in an embodiment of the present invention; Figure 3 This is a second sub-flowchart of a multimodal sensing adaptive adjustment method provided in an embodiment of the present invention; Figure 4This is a third sub-flowchart of a multimodal sensing adaptive adjustment method provided in an embodiment of the present invention; Figure 5 This is a fourth sub-flowchart of a multimodal sensing adaptive adjustment method provided in an embodiment of the present invention; Figure 6 This is the fifth sub-flowchart of a multimodal sensing adaptive adjustment method provided in an embodiment of the present invention; Figure 7 This is a block diagram of a multimodal sensing adaptive adjustment device provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0023] It should be noted that in the description of this invention, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0024] like Figures 1 to 7 As shown, embodiments of the present invention provide a multimodal sensing adaptive adjustment method, apparatus, electronic device, and storage medium. The technical solution of the present invention will be described in detail below.

[0025] like Figure 1 As shown, Figure 1 This is a main flowchart of a multimodal sensing adaptive adjustment method provided in an embodiment of the present invention. The method is applied to a data processing device, which can be a microcontroller unit, an application-specific integrated circuit, or a cloud server node built into a host device. The method includes the following steps: S110 acquires multimodal sensing data collected by sensing components distributed on the supporting body; wherein, the multimodal sensing data includes pressure distribution data collected by the pressure sensing matrix and attitude angle data collected by the inertial measurement unit; in this embodiment, the data processing device establishes a low-level data transmission link with the sensing components distributed on the supporting body (such as a smart sofa or smart seat equipped with an electric adjustment mechanism) through a preset hardware communication interface (such as SPI bus, I2C bus or CAN bus). The physical architecture of the sensing components includes a two-layer network: the first layer is a pressure sensing matrix, which consists of a large number of piezoresistive or capacitive sensing units distributed in a grid pattern, deployed on the seat surface, backrest, armrests and leg rest of the supporting body, etc., which are in physical contact with the target object, and configured to sense the two-dimensional pressure distribution state of the contact surface. The second layer is an inertial measurement unit, which includes a multi-axis accelerometer and a gyroscope, fixed to each adjustable mechanical component of the supporting body, and configured to sense the spatial attitude change and angular velocity information of each component. The data processing device executes a periodic hardware interrupt service routine to synchronously acquire the pressure distribution data output by the pressure sensing matrix at a preset sampling frequency (optimized to a sampling period of 100 milliseconds to 500 milliseconds in this embodiment). and the attitude angle data output by the inertial measurement unit ,in This indicates the current sampling time. After being acquired into the memory buffer, the data processing device performs preprocessing steps on the raw data: First, a digital low-pass filter is invoked to filter out high-frequency mechanical vibrations and electromagnetic noise; second, baseline calibration processing for zero-point offset is performed based on the reference value of the load-bearing body under no-load conditions; finally, the ADC sampling values ​​of each pressure sensing unit are normalized to a unified dimensionless numerical range through linear mapping, thereby constructing a high signal-to-noise ratio multimodal sensing data queue.

[0026] Based on the pressure distribution data and posture angle data, S120 determines the current sitting posture category of the target object relative to the supporting body using a preset sitting posture classification model. After acquiring preprocessed multimodal sensing data, the data processing device performs multimodal data fusion and recognition calculations using the preset sitting posture classification model loaded in memory. The preset sitting posture classification model is instantiated during system initialization, and its internal logic is mapped to a multi-level classification system containing 6 major categories and 19 subcategories. The 6 major categories include: upright sitting posture (tilt angle 0°-10°), backrest sitting posture (10°-45°), deep reclining posture (45°-180°), cross-legged / curled sitting posture (limb folding features), side sitting / slanted sitting posture (body asymmetry features), and special postures (multiple people / unconventional features). The data processing device concatenates the pressure distribution data and posture angle data into a high-dimensional input tensor, inputs it into the feature extraction network of the sitting posture classification model, and calculates a probability distribution vector of length 19 through nonlinear activation functions and fully connected layers. The category index corresponding to the element with the largest value in this vector is determined as the category of the target object in relation to the current sitting posture of the carrier. .

[0027] S130, based on the current sitting posture category, integrates the target object's historical preference data, the current posture change trend characteristics determined based on the posture angle data, and the pressure distribution change direction characteristics determined based on the pressure distribution data to infer the target object's intended sitting posture category. After determining the current sitting posture category, the data processing device does not directly use it as the control target, but instead initiates the intention inference module to perform probabilistic graphical reasoning to filter transitional postures or unconscious temporary movements. This process is implemented in software logic by constructing three parallel computing branches: The first branch extracts historical dimension features: the data processing device accesses a preset local storage medium or cloud database, reads the target object's historical preference data, and calculates the prior probability and conditional probability of the current sitting posture category transitioning to other categories in the Markov chain. The second branch extracts kinematic dimension features: the data processing device retrieves historical posture angle data within a sliding time window, and calculates the target object's motion vector in three-dimensional space, i.e., the current posture change trend characteristics, through finite difference method or derivative operation, to indicate the expected termination position of the target object's limb movement. The third branch extracts force-related features: The data processing device performs matrix subtraction on the pressure distribution data within the sliding time window to obtain the spatial displacement vector of the pressure center of gravity, i.e., the direction feature of pressure distribution change. The data processing device then calls a multilayer perceptron (MLP) or a preset weighted fusion algorithm to jointly calculate the features of the above three dimensions. By fusing the data of these three dimensions, the system can effectively distinguish whether the target object is in a true posture adjustment intention or is merely performing a brief physical stretch, thereby outputting the final inference result and determining it as the target object's intended sitting posture category.

[0028] S140 queries a preset posture mapping table based on the intended sitting posture category to obtain the target posture parameters. After obtaining the intended sitting posture category, the data processing device performs a hash table or relational database query operation in a preset storage medium. The preset posture mapping table stores the nonlinear mapping relationship between 19 intended sitting posture categories and the control parameters of the underlying mechanical components. The data processing device uses the unique identifier (ID) of the intended sitting posture category as the primary key to retrieve the corresponding target posture parameters in the posture mapping table. The target posture parameter is a multi-dimensional floating-point vector, and its data structure contains the target values ​​of all mechanical degrees of freedom required to achieve the intended sitting posture category, such as the target backrest angle, the target seat cushion tilt angle, the target leg rest lifting angle, and the target inflation pressure or mechanical protrusion of the lumbar support airbag. The data processing device loads the obtained target posture parameters into the instruction cache of the motion control thread.

[0029] Based on the target attitude parameters, S150 generates an adjustment trajectory according to a preset velocity curve model and instructs the drive mechanism of the supporting body to perform corresponding attitude adjustments according to the adjustment trajectory. Before executing the underlying hardware drive, to avoid physical shocks caused by sudden mechanical changes, the data processing device calls a preset velocity curve model (such as an S-shaped velocity curve model or a trapezoidal velocity curve model) for path planning. Taking the current attitude parameters of each component of the supporting body as the starting point and the target attitude parameters as the ending point, the data processing device introduces maximum velocity, maximum acceleration, and maximum jerk as kinematic constraints to generate a discrete, continuous, and smooth adjustment trajectory in the time domain. After the adjustment trajectory is generated, the data processing device converts it into a series of underlying pulse width modulation signals or stepping control commands, and sends them to the drive mechanism of the supporting body (including multiple sets of servo motors or stepper motors and transmission links) via the fieldbus. The control logic instructs each set of motors to perform coordinated operation according to the assigned adjustment trajectory, thereby achieving smooth and stepless corresponding attitude adjustments, enabling the supporting body to accurately reach the physical position corresponding to the target attitude parameters.

[0030] like Figure 2 As shown, Figure 2 The first sub-flowchart of a multimodal sensing adaptive adjustment method provided in this embodiment of the invention details the implementation process of the above step S120, including: S210 calculates the principal tilt angle of the target object based on the attitude angle data; in the first-level identification stage, the data processing device extracts the gyroscope and acceleration measurement values ​​output by each inertial measurement unit from the attitude angle data. To eliminate calculation errors caused by local deformation of the supporting body, the data processing device uses an algorithm combining weighted averaging and inverse trigonometric functions to calculate the overall torso tilt. Specifically, the principal tilt angle is calculated using the following formula: ,in For the first The gyroscope measurements of each inertial measurement unit. This corresponds to the installation angle offset of the sensor. This main tilt angle objectively reflects the angle between the target object's spine and the direction of gravity.

[0031] S220 determines the first-level sitting posture category of the target object in a preset sitting posture classification system based on the main tilt angle. After obtaining the main tilt angle, the data processing device compares it with multiple threshold ranges in the preset sitting posture classification system. As mentioned above, the classification system is divided into 6 major categories. The data processing device executes a series of logical condition judgment branches: if the main tilt angle is in the range of 0°-10°, the first-level sitting posture category is determined to be "upright sitting posture"; if it is in the range of 10°-45°, it is determined to be "backrest sitting posture"; if it is in the range of 45°-180°, it is determined to be "deep reclining posture". Through this step, the search space for subsequent fine feature matching is greatly reduced, and the execution efficiency of the algorithm is improved.

[0032] S230 extracts feature vectors from the pressure distribution data within the first-level sitting posture category. These feature vectors include regional pressure peak values, pressure center of gravity positions, pressure distribution symmetry coefficients, and pressure distribution entropy values. Moving to the second-level fine recognition stage, the data processing device performs matrix feature decomposition on the two-dimensional pressure distribution data within the defined first-level sitting posture category range. The data processing device extracts and calculates a multi-dimensional feature vector. This feature vector contains the following key physical quantities: the regional pressure peak determined by finding local maxima through matrix traversal; the pressure centroid location calculated using a weighted average formula of the coordinates of each sensing unit and its pressure value; the pressure distribution symmetry coefficient calculated by comparing the pressure matrix differences on both sides of the central axis of the load-bearing body; and the pressure distribution entropy value, a statistical indicator used to measure the degree of pressure dispersion. These features can accurately characterize the microscopic force patterns exerted by the muscles and bones of the target object on the load-bearing body.

[0033] S240 identifies the feature vector using the posture classification model to determine the current posture category of the target object relative to the supporting body and obtains the classification confidence score. The data processing device calls a pre-trained posture classification model (e.g., a Support Vector Machine (SVM) or a Lightweight Convolutional Neural Network (CNN) classifier) ​​and uses the extracted feature vector as model input for forward propagation calculation. Based on a multi-dimensional decision boundary or hyperplane, the model outputs the probability distribution of each specific subclass (e.g., "standard backrest sitting," "semi-reclining backrest sitting," or "cross-leg backrest sitting") under the first-level category. The data processing device determines the subclass corresponding to the maximum probability as the current posture category. And use the maximum probability value as the classification confidence score. If the classification confidence level is lower than the preset confidence level threshold (e.g., 0.85), the data processing device will maintain the classification result of the previous sampling period unchanged to prevent classification jumps caused by sensor transient noise.

[0034] like Figure 3 As shown, Figure 3 This is a second sub-flowchart of a multimodal sensing adaptive adjustment method provided in an embodiment of the present invention, which details the mathematical model for multi-dimensional inference in step S130: S310 calculates the first conditional probability of transitioning from the current sitting posture category to each candidate intended sitting posture based on the historical preference data; the data processing device accesses the name... The user preference database is used to retrieve historical behavioral statistics for the target individual. The data processing device extracts the current sitting posture category. The corresponding state transition matrix is ​​used to calculate the first conditional probability using Bayesian inference. The specific calculation logic is as follows: for any candidate intended sitting posture... Get the category of the current sitting posture Subsequent transfer to transition probability and combined Prior probability The first conditional probability inferred from historical habits is calculated. This probability reflects the tendency of the target object to shift its posture purely based on personal usage habits in the absence of external physical interference.

[0035] S320 calculates a posture change trend vector based on posture angle data from multiple consecutive sampling periods; and calculates a second conditional probability that the target object tends to each candidate intended sitting posture based on the posture change trend vector; in the time series processing dimension, the data processing device reads posture angle data from multiple consecutive sampling periods (e.g., window length N is 5 to 10 periods) stored in a circular buffer. to Through the formula Calculate the attitude change trend vector ,in The sampling interval is defined as follows. Subsequently, the data processing device correlates the posture change trend vector with each candidate intended sitting posture. Spatial geometric distance is calculated between feature centers in a multidimensional pose space. The distance is mapped to a probability using a Gaussian kernel function, and the calculation formula is as follows: ,in For spatial distance, This is the kernel bandwidth parameter. The result of this formula is the second conditional probability. This indicates the probability that the current physical kinematic trend points to each target location.

[0036] S330 calculates a pressure distribution change vector based on pressure distribution data from multiple consecutive sampling periods; and calculates the third conditional probability of the target object's tendency toward each candidate intended sitting posture based on the pressure distribution change vector; similar to the processing of posture angles, the data processing device extracts pressure distribution data with a window length of M from memory for the pressure field data, and uses the matrix difference formula... Calculate the pressure distribution change vector This vector, in matrix form, represents the dynamic shift of the center of force on the seat or backrest (e.g., increased pressure at the front indicates forward leaning). Subsequently, the data processing device compares the currently calculated pressure distribution change vector with the pre-set standard pressure change feature centers for each sitting posture category in the database. Calculate the Euclidean distance. Similarly, use the Gaussian probability density function formula: After transformation, the third conditional probability is obtained. This parameter delves into the physical essence of force changes in muscle groups.

[0037] S340 performs a weighted summation of the first conditional probability, the second conditional probability, and the third conditional probability according to preset weight coefficients to obtain the comprehensive probability of each candidate intended sitting posture; after obtaining the conditional probabilities of the above three independent dimensions, the fusion engine of the data processing device performs vector dot product operation. The engine extracts the three preset weight coefficients. , , (and satisfy) ), The above is a weighted summation formula, calculated for each candidate intended sitting posture category, and finally outputs a comprehensive probability array for all candidate categories. By properly configuring the weights, the system can achieve the optimal mathematical expectation between historical prior knowledge and real-time sensor feedback.

[0038] S350 determines the candidate intended sitting posture corresponding to the highest comprehensive probability as the intended sitting posture category of the target object. After weighted fusion calculation, the data processing device traverses and sorts the array storing comprehensive probabilities, and executes the argmax mathematical operator to extract the desired posture. Maximize the category index ,Right now This category index The corresponding data structure node is uniquely and precisely identified as the intended sitting posture category of the target object, serving as an unchangeable source of instruction input for the subsequent drive control phase.

[0039] like Figure 4 As shown, Figure 4The third sub-flowchart of a multimodal sensing adaptive adjustment method provided in this embodiment of the invention illustrates the process of continuous optimization of the system through a feedback mechanism: S410 acquires the manual correction signal triggered by the target object against the carrier body within a preset time window; the manual correction signal contains the corrected final attitude parameters; after instructing the drive mechanism to complete the adjustment in step S150 above, the data processing device automatically starts a system timer to enable a listening service for a preset time window (preferably set to 3 to 5 minutes). During this window period, the data processing device continuously monitors the control panel, touch screen interface, or voice receiving module of the carrier body. If the data processing device receives a hardware interrupt through the I / O port during this period, it indicates that the target object has intervened, and the data processing device will capture the event and parse it as a manual correction signal. When the manual adjustment action stops, the data processing device reads the absolute position encoder feedback values ​​of all mechanical sensors of the carrier body at this time and encapsulates them into a corrected final attitude parameter matrix.

[0040] S420 updates the historical preference data of the target object based on the difference between the target attitude parameters and the final attitude parameters; upon receiving feedback, the data processing device performs matrix subtraction to calculate the target attitude parameters (denoted as a vector) originally generated by the system. The final attitude parameters (denoted as vectors) after user manual intervention are compared with the final attitude parameters after user manual intervention. The algebraic differences across various dimensions constitute an error vector. The data processing device uses this error vector as the backpropagation signal (i.e., the learning signal) in a reinforcement learning environment to access the user preference database in the storage medium. The system adjusts the specific values ​​of the corresponding conditional probability and prior probability in this scenario using a cover or smooth update algorithm, causing a shift in the model features of the intended sitting posture in the system's memory, thus strictly aligning with the latest preferences of the target object.

[0041] S430 updates the weight coefficients used to infer the intended sitting posture category based on the gradient direction corresponding to the difference. In addition to updating the static database, the data processing device further employs a gradient ascent strategy to dynamically update the weights at the bottom layer of the algorithm model. The data processing device calculates the difference between each component of the target posture parameter vector (e.g., backrest angle, leg rest angle) and obtains the partial derivative of the intended sitting posture probability with respect to each weight coefficient (i.e., the sensitivity gradient). Using the formula... .in The preset learning rate, The process iterates through all mechanical degrees of freedom parameters. By calculating the gradient direction corresponding to the difference, the data processing device precisely adjusts the weighting coefficients in the aforementioned fusion formula. , , This allows the calculated output parameters to more closely approximate the coordinates of the final attitude parameters in multidimensional space under the same multimodal sensing data input.

[0042] like Figure 5 As shown, Figure 5 The fourth sub-flowchart of a multimodal sensing adaptive adjustment method provided in this embodiment of the invention illustrates the quantitative assessment of physiological health dimensions: S510 queries a preset sitting posture load mapping table based on the current sitting posture category to obtain the corresponding first load coefficient. To establish a quantitative assessment related to the target object's physiological health indicators, the data processing device, within each assessment cycle, uses the current sitting posture category to query the preset sitting posture load mapping table stored in flash memory via the data bus. This mapping table is based on an ergonomic biomechanical model and records the ratio of the physical pressure borne by the intervertebral discs under a specific skeletal posture to that under a standard upright coordinate system. Upon a successful query, the data processing device extracts and temporarily stores the static biomechanical multiplier associated with the current sitting posture, i.e., the first load coefficient. .

[0043] S520 calculates the second load coefficient based on the duration parameter of the current sitting posture category and the posture change frequency parameter; based on the obtained static indicators, the data processing device reads the system global timer to obtain the absolute duration parameter of the target object maintaining the current sitting posture category. Simultaneously, by calculating the number of times the sitting posture state changes over a period of time, the posture change frequency parameter is extracted. The data processing equipment substitutes these dynamic time variables into the ergonomic fatigue cumulative integral formula, performs definite integral calculations over the time domain, and then uses the formula... (in These are weighting coefficients. (Using the reference transformation frequency), the output is a second load coefficient that characterizes the degree of nonlinear accumulation of muscle strain over time.

[0044] S530 acquires the biometric attribute information of the target object; the biometric attribute information includes age range and body mass index; to achieve individualized adaptation of the evaluation model, the data processing device acquires the biometric attribute information of the target object pre-entered in the cloud server or local configuration through a network interface or local APP synchronization interface. This data structure contains specific integer or floating-point fields, explicitly indicating the target object's physiological age range (e.g., 20-40 years, 40-60 years, over 60 years) and the body mass index (BMI) calculated from height and weight.

[0045] S540 determines the first weight corresponding to the first load coefficient and the second weight corresponding to the second load coefficient based on the biometric attribute information; after extracting the biometric attribute information, the data processing device allocates weights through conditional branch judgment logic. Taking an age range as an example: if it is determined to be in the 20-40 age range, then the first weight is configured. Second weight If the age range is 40-60, then allocate the first weight. Second weight If the age is greater than 60, both weights are set to 0.5. The system further fine-tunes the first and second weights based on the deviation of the body mass index from the standard value using a linear compensation function to generate a final weight matrix specific to the target object.

[0046] S550, based on the first weight and the second weight, performs a weighted summation of the first load coefficient and the second load coefficient to obtain the strain risk value. The data processing equipment extracts the calculated static and dynamic parameters and applies a linear weighted formula. Floating-point arithmetic is performed. The scalar value output by the arithmetic is defined as the strain risk value of the target object. This value is normalized to form a quantitative indicator that objectively represents the current fatigue level of the spine and muscle groups, and is stored in a memory variable for periodic polling by the health monitoring thread.

[0047] S560 generates a health intervention signal when the strain risk value exceeds a preset risk threshold, and instructs the drive mechanism to perform adjustments according to the safe posture parameters corresponding to the health intervention signal. The health monitoring thread runs with high priority. Once the comparator determines that the strain risk value read from memory is algebraically greater than the pre-programmed safety boundary—the risk threshold... The data processing device immediately triggers the highest-level software interrupt. This interrupt service routine packages and sends a health intervention signal containing a specific identifier. Instead of performing the usual intentional sitting posture inference, the system forcibly reads a set of preset safe posture parameters aimed at relieving spinal pressure and instructs the drive mechanism to take over the motor control of the load-bearing body, slowly and smoothly performing interventional displacement adjustment, thereby actively preventing further accumulation of fatigue at the physical level.

[0048] like Figure 6 As shown, Figure 6 The fifth sub-flowchart of a multimodal sensing adaptive adjustment method provided in this embodiment of the invention is specifically designed to handle complex spatial concurrency requirements: S610 performs connected component analysis on the pressure distribution data to extract multiple independent pressure regions exceeding a preset activation threshold. In complex spatial scenarios such as shared homes, when the data processing device acquires multimodal sensor data containing multiple physical entity features such as "two people sitting close together," the system calls a connected component labeling algorithm from the field of image processing (such as a connected component analysis algorithm based on depth-first or breadth-first search) to perform a topological scan of the two-dimensional pressure distribution matrix. By comparing with the preset activation threshold, the data processing device segments out the cluster of sensor units that are physically isolated and whose pressure intensity exceeds the limit, thereby generating and extracting multiple independent pressure region structures that do not intersect in the physical coordinate system in the memory.

[0049] S620 assigns a unique object identifier to each independent pressure region and maps the attitude angle data according to the positional relationship of the independent pressure regions. After successfully segmenting the matrix, the operating system kernel of the data processing device instantiates a process or thread for each independent pressure region and assigns it a unique object identifier (UserID, such as User1, User2) in memory. Next, the data processing device queries the hardware assembly topology table of the carrier to calculate the Euclidean distance between the three-dimensional spatial coordinates of each inertial measurement unit and the centroid of each independent pressure region. Based on the principle of minimum distance and the mechanical transmission relationship, the overall acquired attitude angle data is decoupled and accurately mapped to the corresponding object identifier, ensuring that each thread has independent context data.

[0050] For each independent object identifier, S630 executes the processing steps from determining the current sitting posture category to inferring the intended sitting posture category in parallel. The data processing device utilizes the concurrent computing capabilities of the underlying multi-core processor, employing a multi-threaded concurrency mechanism to ensure that the logical units corresponding to each object identifier run the recognition algorithm synchronously and independently. Each thread independently calls the classification model and probabilistic inference network, executing the aforementioned steps S120 to S130 in parallel. Ultimately, each thread will output the current sitting posture category and the inferred intended sitting posture category specific to its corresponding target object within its own memory space, without interfering with each other.

[0051] S640 collaboratively generates a multi-target adjustment trajectory for the load-bearing entity based on a preset control priority strategy. When the inference results of all concurrent threads are aggregated to the main control scheduler, the data processing device invokes the preset control priority strategy to resolve conflicts due to the physical rigidity constraints of the load-bearing entity's mechanical structure. For example, the system calculates the distance matrix between the intended sitting posture action request point and the physical positions of each mechanical actuator, assigning higher priority flags to adjustment requests that are closer or involve the main load-bearing beam. Based on the arbitration-based decision matrix, the control algorithm collaboratively generates a set of multi-target adjustment trajectories that ensure no collisions occur in the mechanical structure while taking into account the needs of multiple parties. The data processing device then converts this into a synchronous control frame and sends it through the communication bus, realizing unified and coordinated driving of the complex physical structure of the load-bearing entity.

[0052] like Figure 7 As shown, Figure 7 This is a block diagram of a multimodal sensing adaptive adjustment device provided in an embodiment of the present invention. The device, burned into the memory of a data processing device in the form of software code or firmware, includes: a data acquisition module, configuring an interface driver to establish a hardware handshake, for acquiring multimodal sensing data collected by sensing components distributed on a support body; wherein the multimodal sensing data includes pressure distribution data collected by a pressure sensing matrix and attitude angle data collected by an inertial measurement unit; a posture recognition module, encapsulating a neural network or SVM inference engine, for performing tensor calculations based on the pressure distribution data and the attitude angle data, using a preset posture classification model to determine the current posture category of the target object relative to the support body; and an intention inference module, including a Markov chain and a Gaussian probability density calculation unit, for inferring the current... The sitting posture category is determined by integrating the target object's historical preference data, the current posture change trend characteristics determined based on the posture angle data, and the pressure distribution change direction characteristics determined based on the pressure distribution data, and inferring the target object's intended sitting posture category through vector weighting. The parameter determination module is configured with an efficient hash lookup algorithm to query a preset posture mapping table based on the intended sitting posture category, converting logical identifiers into target posture parameters for physical hardware. The trajectory adjustment module embeds a kinematic curve generation algorithm and PID closed-loop control logic to generate a smooth adjustment trajectory based on the target posture parameters and a preset velocity curve model, and instructs the drive mechanism of the supporting body to perform corresponding posture adjustments according to the adjustment trajectory.

[0053] This invention also provides an electronic device, including: a central processing unit (CPU) or a microcontroller (MCU) and a memory (such as DDR SDRAM and non-volatile Flash); the memory stores computer program instructions that can be executed by the processor; when the processor reads and executes the computer program through a data bus, it implements all the steps of the multimodal sensing adaptive adjustment method as described in the above method embodiments.

[0054] This invention also provides a computer-readable storage medium, such as EEPROM, flash memory, or any non-volatile solid-state memory, on which a compiled binary computer program is stored. When the program is read and executed by a processor, it implements the steps of the multimodal sensing adaptive adjustment method as described in any of the above embodiments.

[0055] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multimodal sensing adaptive adjustment method, applied to data processing equipment, characterized in that, The method includes: Acquire multimodal sensing data collected by sensing components distributed on the supporting body; wherein, the multimodal sensing data includes pressure distribution data collected by the pressure sensing matrix and attitude angle data collected by the inertial measurement unit; Based on the pressure distribution data and the posture angle data, the current sitting posture category of the target object relative to the supporting body is determined by a preset sitting posture classification model. Based on the current sitting posture category, the target object's intended sitting posture category is inferred by integrating the target object's historical preference data, the current posture change trend characteristics determined based on the posture angle data, and the pressure distribution change direction characteristics determined based on the pressure distribution data. The target posture parameters are obtained by querying a preset posture mapping table based on the intended sitting posture category. Based on the target attitude parameters, an adjustment trajectory is generated according to a preset speed curve model, and the drive mechanism of the supporting body is instructed to perform corresponding attitude adjustments according to the adjustment trajectory.

2. The multimodal sensing adaptive adjustment method according to claim 1, characterized in that, The step of determining the current sitting posture category of the target object relative to the supporting body based on the pressure distribution data and the posture angle data, using a preset sitting posture classification model, includes: Calculate the principal tilt angle of the target object based on the attitude angle data; Based on the main tilt angle, determine the first-level sitting posture category of the target object in the preset sitting posture classification system; Within the first-level sitting posture category, feature vectors of the pressure distribution data are extracted; the feature vectors include regional pressure peak value, pressure center of gravity position, pressure distribution symmetry coefficient, and pressure distribution entropy value. The feature vector is identified by the sitting posture classification model to determine the current sitting posture category of the target object relative to the supporting body, and the classification confidence score is obtained.

3. The multimodal sensing adaptive adjustment method according to claim 1, characterized in that, The step of inferring the target object's intended sitting posture category based on the current sitting posture category, by fusing the target object's historical preference data, the current posture change trend characteristics determined based on the posture angle data, and the pressure distribution change direction characteristics determined based on the pressure distribution data, includes: Based on the historical preference data, calculate the first conditional probability of the current sitting posture category shifting to each candidate intended sitting posture; Based on posture angle data from multiple consecutive sampling periods, a posture change trend vector is calculated; and based on the posture change trend vector, a second conditional probability is calculated of the target object tending to each candidate intentional sitting posture. Based on pressure distribution data from multiple consecutive sampling periods, a pressure distribution change vector is calculated; and based on the pressure distribution change vector, a third conditional probability is calculated of the target object's tendency toward each candidate intended sitting posture. The first conditional probability, the second conditional probability, and the third conditional probability are weighted and summed according to preset weighting coefficients to obtain the comprehensive probability of each candidate intended sitting posture. The candidate intended sitting posture corresponding to the highest comprehensive probability is determined as the intended sitting posture category of the target object.

4. The multimodal sensing adaptive adjustment method according to claim 1, characterized in that, After the drive mechanism that instructs the supporting body to perform the corresponding attitude adjustment according to the adjustment trajectory, the method further includes: Within a preset time window, acquire the manual correction signal triggered by the target object against the carrier; the manual correction signal includes the corrected final attitude parameters; The historical preference data of the target object is updated based on the difference between the target attitude parameters and the final attitude parameters; Based on the gradient direction corresponding to the difference, update the weight coefficients used when inferring the intended sitting posture category.

5. The multimodal sensing adaptive adjustment method according to claim 1, characterized in that, The method further includes: Based on the current sitting posture category, query the preset sitting posture load mapping table to obtain the corresponding first load coefficient; The second load coefficient is calculated based on the duration parameter of the current sitting posture category and the posture change frequency parameter. The first load factor and the second load factor are weighted and calculated to obtain the strain risk value of the target object; If the risk value of strain exceeds a preset risk threshold, a health intervention signal is generated, and the drive mechanism is instructed to perform adjustments according to the safe posture parameters corresponding to the health intervention signal.

6. The multimodal sensing adaptive adjustment method according to claim 5, characterized in that, The step of weighting the first load factor and the second load factor to obtain the strain risk value of the target object includes: Obtain the biometric attribute information of the target object; the biometric attribute information includes age range and body mass index; The first weight corresponding to the first load coefficient and the second weight corresponding to the second load coefficient are determined based on the biometric attribute information. Based on the first weight and the second weight, the first load coefficient and the second load coefficient are weighted and summed to obtain the strain risk value.

7. The multimodal sensing adaptive adjustment method according to claim 1, characterized in that, After acquiring the multimodal sensing data collected by the sensing components distributed on the supporting body, the method further includes: Perform connected component analysis on the pressure distribution data to extract multiple independent pressure regions that exceed a preset activation threshold; Each independent pressure region is assigned a unique object identifier, and the attitude angle data is mapped and allocated according to the positional relationship of the independent pressure regions. For each individual object identifier, the processing steps from determining the current sitting posture category to inferring the intended sitting posture category are executed in parallel. Based on a preset control priority strategy, a multi-target adjustment trajectory is collaboratively generated for the carrier.

8. A multimodal sensing adaptive adjustment device, characterized in that, include: The data acquisition module is used to acquire multimodal sensing data collected by sensing components distributed on the supporting body; wherein, the multimodal sensing data includes pressure distribution data collected by the pressure sensing matrix and attitude angle data collected by the inertial measurement unit; The sitting posture recognition module is used to determine the current sitting posture category of the target object relative to the supporting body based on the pressure distribution data and the posture angle data and through a preset sitting posture classification model. The intention inference module is used to infer the target object's intention sitting posture category based on the current sitting posture category, by integrating the target object's historical preference data, the current posture change trend characteristics determined based on the posture angle data, and the pressure distribution change direction characteristics determined based on the pressure distribution data. The parameter determination module is used to query a preset posture mapping table based on the intended sitting posture category to obtain the target posture parameters; The trajectory adjustment module is used to generate an adjustment trajectory based on the target attitude parameters and a preset velocity curve model, and to instruct the drive mechanism of the supporting body to perform corresponding attitude adjustment according to the adjustment trajectory.

9. An electronic device, characterized in that, include: Processor and memory; The memory stores computer programs that can be executed by the processor; When the processor executes the computer program, it implements the multimodal sensing adaptive adjustment method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the steps of the multimodal sensing adaptive adjustment method as described in any one of claims 1 to 7.

Images