Image-based analysis of neonatal intestinal distension and feeding position correlation and correction system

Abstract

The application discloses an image-based correlation analysis and correction system for neonatal intestinal distension and feeding posture, and belongs to the technical field of infant care, and comprises the following modules: a multi-modal image acquisition module, which acquires infant facial images, depth images and temperature distribution images; a differential geometry posture analysis module, which constructs an infant head and a feeding bottle position into a parameterized curved surface manifold, calculates posture curvature features, predicts posture evolution trends, and plans an optimal correction path; an intestinal distension state monitoring module, which evaluates an intestinal distension risk level; a posture-intestinal distension correlation analysis module, which establishes a personalized correlation model; a correction strategy generation module, which generates an optimal correction path and a control parameter sequence; and an execution control module, which drives a bed body and a feeding bottle support system to perform posture correction actions. The application introduces differential geometry theory, upgrades feeding posture analysis from traditional plane coordinate analysis to curved surface manifold space analysis, and improves posture recognition accuracy from 75% to 85% to above 95%.

Description

Technical Field

[0001] This invention relates to the field of infant and toddler care technology, and in particular to an image-based system for analyzing and correcting the correlation between neonatal gas and feeding posture. This system uses image acquisition and analysis technology to monitor, analyze, and automatically correct the feeding posture of newborns, thereby reducing the incidence of gas and improving the feeding comfort of newborns. Background Technology

[0002] Neonatal gas refers to abdominal distension and pain caused by the accumulation of gas in the intestines. Statistics show that approximately 60% to 70% of newborns experience varying degrees of gas symptoms in the first three months after birth. Gas can lead to infant fussiness, poor sleep quality, and seriously affect the newborn's physical and mental health as well as the quality of family life.

[0003] Currently, it is widely accepted in clinical practice that improper feeding posture is one of the main causes of neonatal gas. Traditional prevention and treatment methods mainly rely on the experience and judgment of caregivers, manually adjusting the infant's feeding posture. This method suffers from problems such as strong subjectivity, inconsistent standards, and difficulty in quantification. Furthermore, for inexperienced new parents, correctly mastering the infant feeding posture is a significant challenge.

[0004] Existing technologies attempt to monitor infant feeding postures using image acquisition devices, but they primarily focus on two-dimensional planar image analysis, failing to accurately capture the complex postural relationships in three-dimensional space. Furthermore, most systems only provide simple posture assessments, lacking the ability to predict trends in posture changes and automatic correction functions. In addition, current technologies generally lack in-depth analysis of the correlation between intestinal gas and feeding posture, making it impossible to develop personalized prevention plans.

[0005] Therefore, there is an urgent need for a system that can accurately analyze infant feeding postures, predict trends in posture changes, automatically implement corrections, and establish personalized gas prevention plans. Summary of the Invention

[0006] The purpose of this invention is to provide an image-based system for analyzing and correcting the correlation between neonatal gas and feeding posture. By introducing differential geometry theory to construct a mathematical model of feeding posture, the system can accurately represent, dynamically predict, and intelligently correct feeding posture, thereby reducing the incidence of neonatal gas and improving feeding comfort.

[0007] This invention proposes an image-based system for analyzing and correcting the correlation between neonatal gas and feeding posture, comprising:

[0008] The multimodal image acquisition module is used to acquire facial images, depth images, and temperature distribution images during infant feeding.

[0009] An image preprocessing module, connected to the multimodal image acquisition module, is used to perform noise filtering and key point extraction on the acquired images;

[0010] The differential geometry posture analysis module, connected to the image preprocessing module, is used to construct a parameterized curved manifold between the infant's head position and the bottle position, calculate posture curvature characteristics, predict posture evolution trends based on curvature flow theory, and plan the optimal correction path on the posture manifold.

[0011] The gas monitoring module, connected to the multimodal image acquisition module, is used to monitor the infant's respiratory rate, abdominal activity, and body temperature distribution, and to assess the risk level of gas.

[0012] The posture-bloating correlation analysis module is connected to the differential geometric posture analysis module and the bloating state monitoring module. It is used to analyze the statistical correlation between different postures and bloating state and establish a personalized correlation model.

[0013] The correction strategy generation module is connected to the differential geometric posture analysis module and the posture-intestinal gas correlation analysis module, and is used to generate the optimal correction path and control parameter sequence based on the current posture risk assessment results.

[0014] The execution control module, connected to the correction strategy generation module, is used to receive the control parameter sequence, drive the bed adjustment system and the bottle holder system to perform posture correction actions, and feed back the execution status to the correction strategy generation module.

[0015] Preferably, the multimodal image acquisition module includes:

[0016] Area scan camera, used to capture high-resolution images of the baby's head, face, and upper body;

[0017] Depth cameras are used to acquire three-dimensional spatial information and construct depth maps;

[0018] Infrared thermal imagers are used to monitor the distribution of an infant's body temperature.

[0019] The time synchronization unit is connected to the area array camera, the depth camera, and the infrared thermal imager to ensure the time consistency of multi-source image data.

[0020] Preferably, the differential geometric attitude analysis module includes:

[0021] Surface manifold building units are used to construct parametric surface manifolds for baby head pose and bottle position based on key points;

[0022] The curvature feature calculation unit, connected to the surface manifold construction unit, is used to calculate the Gaussian curvature and average curvature on the manifold and extract the attitude feature vector;

[0023] The curvature flow prediction unit, connected to the curvature feature calculation unit, is used to construct discrete curvature flow equations, predict posture evolution trajectories, and identify equilibrium points and unstable regions in the posture space.

[0024] The Riemann path planning unit, connected to the curvature flow prediction unit, is used to calculate the geodesic path from the current pose to the ideal pose on the risk-weighted attitude manifold and decompose the path into a sequence of control parameters.

[0025] Preferably, the curvature feature calculation unit extracts attitude features through the following steps:

[0026] Divide the key points of the infant's head and face into multiple local areas;

[0027] Calculate the principal curvature and principal direction for each local region;

[0028] Construct a curvature histogram descriptor to represent local shape features;

[0029] The features of each region are weighted and fused to form a global pose feature vector.

[0030] Preferably, the curvature flow prediction unit employs a multi-timescale prediction strategy:

[0031] Short-timescale forecasting uses a locally linearized model for real-time feedback control;

[0032] Mid-timescale prediction combined with local nonlinear dynamics is used for attitude stability assessment;

[0033] Long-term scale prediction integrates statistical and dynamic models for individualized pattern analysis.

[0034] Preferably, the intestinal gas monitoring module includes:

[0035] The physiological signal analysis unit is used to analyze the infant's respiratory rate and abdominal movements;

[0036] A facial expression recognition unit is used to identify infant discomfort expressions and crying patterns;

[0037] Thermal imaging analysis unit, used to detect abnormal temperature distribution in the abdomen;

[0038] The risk assessment unit, connected to the physiological signal analysis unit, the facial expression recognition unit, and the thermal imaging analysis unit, is used to comprehensively assess the risk level of intestinal bloating based on multidimensional indicators.

[0039] Preferably, the posture-bloating correlation analysis module includes:

[0040] A time-series data storage unit is used to record posture sequences and corresponding intestinal bloating states;

[0041] The association pattern mining unit is connected to the time-series data storage unit and is used to analyze the temporal correlation between posture and flatulence using a sliding time window.

[0042] Individual difference learning units, connected to the association pattern mining unit, are used to adapt to the physiological characteristics of specific infants;

[0043] The model update unit, connected to the individual difference learning unit, is used to continuously optimize the posture-bloating association model.

[0044] Preferably, the correction strategy generation module includes:

[0045] Risk assessment unit, used to assess risk based on current posture and predicted trajectory;

[0046] A target pose selection unit, connected to the risk assessment unit, is used to select the most suitable target pose from a low-risk pose library;

[0047] The path planning unit, connected to the target pose selection unit, is used to calculate the optimal path from the current pose to the target pose;

[0048] A control sequence generation unit, connected to the path planning unit, is used to convert the path into a sequence of control parameters.

[0049] Preferably, the execution control module includes:

[0050] The bed control unit is used to adjust the bed surface height, tilt angle, and curvature;

[0051] The bottle holder control unit is used to precisely adjust the position and angle of the bottle;

[0052] A status monitoring unit, connected to the bed control unit and the bottle holder control unit, is used to monitor the execution effect in real time.

[0053] The human-computer interaction unit, connected to the status monitoring unit, is used to provide status feedback and suggestions to the caregiver.

[0054] Preferably, the bed control unit includes:

[0055] The lifting mechanism is used to adjust the height of the bed surface;

[0056] The X-direction rotation mechanism is used to drive the bed surface to rotate in the X direction on the horizontal plane of the bed surface.

[0057] The Y-direction curvature adjustment mechanism is used to drive the bed surface to rotate in the Y direction according to the curvature of the front end of the bed surface;

[0058] The movement direction of the Y-direction curvature adjustment mechanism is related to the curvature of the front end of the bed, so that the baby's head is kept in a preset direction.

[0059] The beneficial effects of this invention include:

[0060] 1. By introducing differential geometry theory to construct a surface manifold representation of feeding posture, the feeding posture is upgraded from traditional planar coordinate analysis to surface manifold space analysis, which greatly improves the accuracy of posture recognition from 75% to 85% of the traditional method to more than 95%;

[0061] 2. Based on curvature flow theory, dynamic prediction of posture change trends can be achieved, which can predict potential high-risk posture changes 5 to 10 seconds in advance, providing a time window for preventive intervention;

[0062] 3. By constructing an optimal control system under Riemannian geometry, a smooth transition from the current pose to the ideal pose was achieved, and the control accuracy was improved from ±5° to ±1° compared to the traditional method;

[0063] 4. A correlation analysis model between intestinal gas status and feeding posture was established, enabling personalized risk prediction and prevention programs;

[0064] 5. Preclinical trials have shown that this system can reduce the incidence of neonatal colic by 65% ​​and significantly improve infant comfort and sleep quality;

[0065] 6. The system adopts an adaptive learning mechanism, which can continuously optimize the association model according to the individual characteristics of different infants, thereby improving the system's adaptability and accuracy. Attached Figure Description

[0066] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention;

[0067] Figure 2 This is a schematic diagram of the structure of the multimodal image acquisition module of the present invention;

[0068] Figure 3 This is a flowchart of the differential geometry attitude analysis module of the present invention;

[0069] Figure 4 This is a schematic diagram of the intestinal gas monitoring module of the present invention;

[0070] Figure 5 This is a flowchart of the posture-intestinal bloating correlation analysis module of the present invention;

[0071] Figure 6 This is a schematic diagram of the structure of the correction strategy generation module of the present invention;

[0072] Figure 7This is a schematic diagram of the working principle of the execution control module of the present invention. Detailed Implementation

[0073] Please refer to Figures 1-7 The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Those skilled in the art should understand that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0074] Reference Figure 1 The present invention provides an image-based neonatal gas and feeding posture correlation analysis and correction system, comprising a multimodal image acquisition module 1, an image preprocessing module 2, a differential geometric posture analysis module 3, a gas state monitoring module 4, a posture-gas correlation analysis module 5, a correction strategy generation module 6, and an execution control module 7. The modules exchange data and transmit signals via a data bus, collectively forming a closed-loop control system.

[0075] Reference Figure 2 The multimodal image acquisition module 1 includes an area array camera 11, a depth camera 12, an infrared thermal imager 13, and a time synchronization unit 14.

[0076] The area scan camera 11 is used to acquire high-resolution images of the infant's head, face, and upper body. In a preferred embodiment of the invention, the area scan camera 11 is a 24-megapixel color camera with a field of view of 90° and a sampling frequency of 30 frames per second. These parameters ensure that subtle facial expression changes of the infant are captured, providing high-quality image data for subsequent analysis.

[0077] The depth camera 12 is used to acquire three-dimensional spatial information and construct a depth map. Preferably, the depth camera 12 is a structured light or time-of-flight (TOF) camera with a resolution of 640×480, 16-bit depth accuracy, and a detection range of 0.3–3 meters. This setup can accurately capture the relative positional relationship between the baby and the bottle in three-dimensional space, providing the necessary spatial data for attitude manifold construction.

[0078] Infrared thermal imager 13 is used to monitor the infant's body temperature distribution. In this embodiment, thermal imager 13 is a non-contact thermal imaging device with a resolution of 384×288 and a temperature accuracy of ±0.5℃, which acquires temperature distribution maps of the infant's abdominal area at a frequency of 5 frames / second to help determine the state of intestinal gas.

[0079] The time synchronization unit 14 is connected to the area array camera 11, the depth camera 12, and the infrared thermal imager 13 to ensure the temporal consistency of multi-source image data. This unit uses high-precision timestamp technology to mark data collected by different devices according to a unified time base, with a time synchronization accuracy better than 5 milliseconds, ensuring accurate and reliable temporal correlation analysis of multimodal data.

[0080] Image preprocessing module 2 is used to perform noise filtering and key point extraction on the images acquired by multimodal image acquisition module 1. This module receives image data from multimodal image acquisition module 1, first applying Gaussian filtering and median filtering algorithms to remove ambient light interference and random noise, then locating 68 key feature points of the infant's face using an Active Appearance Model (AAM), and simultaneously extracting the bottle contour points using edge detection and ellipse fitting algorithms. To reduce the impact of jitter, this module also applies a Kalman filter to smooth the key point trajectory, with the smoothing window size set to 5 frames. This parameter setting effectively suppresses jitter interference while ensuring real-time performance.

[0081] Reference Figure 3 The differential geometric attitude analysis module 3 is the core innovative module of this system, which includes a surface manifold construction unit 31, a curvature feature calculation unit 32, a curvature flow prediction unit 33, and a Riemann path planning unit 34.

[0082] The surface manifold construction unit 31 is used to construct parametric surface manifolds for the baby's head pose and bottle position based on key points. This unit first uses the set of facial key points extracted by the image preprocessing module 2. and the set of outline points of the baby bottle As input, a parameterized surface manifold for the infant's head position is then constructed using a quadratic spline interpolation method. Parametric manifold for bottle position ,in For parameterized coordinates. Specifically, for the infant head manifold , can be represented as:

[0083] ,

[0084] in: A parameterized surface manifold for the position of the baby's head; These are surface parameters, and their values ​​range from [value range missing]. ; These are quadratic B-spline basis functions used for interpolation to fit surfaces; For the first The three-dimensional coordinates of facial key points are: One key point; This represents the total number of facial key points, which is set to 68 in this embodiment. Indicates all The summation operation is performed on the key points.

[0085] Similarly, the bottle position manifold It can be represented as:

[0086] ,

[0087] in: The parameterized surface manifold for the position of the baby bottle; These are surface parameters, and their values ​​range from [value range missing]. ; These are quadratic B-spline basis functions used for interpolation to fit surfaces; For the first The three-dimensional coordinates of the bottle outline points are: One outline point; The total number of points on the outline of the baby bottle is 32 in this embodiment; Indicates all Summation is performed on each contour point. After the two surface manifolds are constructed, the feeding pose manifold space is defined. As a complete representation of the infant's head posture relative to the bottle. Among them: For feeding posture manifold space; This represents the Cartesian product operation, used to construct a combined space of two surface manifolds.

[0088] The curvature feature calculation unit 32 is connected to the surface manifold construction unit 31 and is used to calculate the Gaussian curvature and mean curvature on the manifold to extract the pose feature vector. This unit first divides the key point region of the infant's head and face into multiple local regions such as the forehead, cheeks, and chin, and calculates the principal curvature for each local region. , and principal direction. Gaussian curvature and mean curvature The calculation is as follows:

[0089] ,

[0090] ,

[0091] in: Gaussian curvature characterizes the intrinsic geometric properties of a surface; The mean curvature characterizes the extrinsic geometric properties of the surface; , Principal curvature, that is, the extreme value of the normal curvature at a point on the surface; Indicates multiplication operation; It represents the arithmetic mean of the two principal curvatures.

[0092] For local areas Construct curvature histogram descriptor Characterizing local shape features:

[0093] ,

[0094] in: Let be the curvature histogram descriptor for the i-th local region, which is a vector. This is a Gaussian curvature histogram, describing the distribution of Gaussian curvature in this region; This is a histogram of mean curvature, describing the distribution of mean curvature in this region; This represents a vector concatenation operation, merging two histograms into a single feature vector. In this embodiment, each histogram contains 10 bins, therefore... It is a 20-dimensional vector.

[0095] The features from each region are weighted and fused to form a global pose feature vector. :

[0096] ,

[0097] in: This is the global pose feature vector, with a dimension of 20; Let be the weight coefficient of the i-th region, representing the contribution of that region to the global features; The total number of regions is 5 in this embodiment, corresponding to the forehead, cheeks, chin, and other regions respectively. This represents a weighted summation over all r regions; This represents the multiplication operation between a scalar and a vector. In this embodiment, the weight of the forehead region is set to 0.3, the weight of each cheek region is 0.2, the weight of the chin region is 0.15, and the weight of the remaining regions is 0.15. This weight allocation reflects the difference in the contribution of different regions to the feeding posture.

[0098] The curvature flow prediction unit 33 is connected to the curvature feature calculation unit 32, and is used to construct discrete curvature flow equations, predict attitude evolution trajectories, and identify equilibrium points and unstable regions in attitude space. This unit constructs discrete curvature flow equations based on curvature feature time series:

[0099] ,

[0100] in: The attitude manifold, which varies with time, is a function of time t; This represents the rate of change of the manifold over time, i.e., the time derivative of the manifold; For curvature, the average curvature H is used in this embodiment; It is a unit normal vector, perpendicular to the manifold surface, with a length of 1; The negative value indicates that the manifold evolves along the direction of decreasing curvature.

[0101] To improve computational efficiency, the forward difference method is used to discretize the equation:

[0102] ,

[0103] in: For time t+ The manifold state; Let be the manifold state at time t; The time step is 0.1 seconds in this embodiment; Let be the curvature at time t; Let be the unit normal vector at time t; This indicates scalar multiplication.

[0104] In practical applications, this unit adopts a multi-timescale prediction strategy: short-timescale (3-5 seconds) prediction uses a locally linearized model for real-time feedback control; medium-timescale (30 seconds-2 minutes) prediction combines local nonlinear dynamics for posture stability assessment; and long-timescale (the entire feeding cycle) prediction integrates statistical and dynamic models for individualized pattern analysis.

[0105] This unit also searches for the equilibrium point of the curvature flow using an iterative method. The equilibrium point satisfies the following condition:

[0106] ,

[0107] Right now or .in: represents the time derivative of the manifold; 0 represents the zero vector, indicating that the manifold no longer changes; Indicates a curvature of zero; This indicates that the normal vector is zero, which will not occur in practice because N is a unit vector.

[0108] For each equilibrium point Calculate the Jacobian matrix of its vicinity. Through analysis The stability of an equilibrium point is determined by its eigenvalues. An equilibrium point where all eigenvalues ​​are negative real parts is a stable equilibrium point (ideal posture), while an equilibrium point with positive real parts is an unstable equilibrium point (risk posture). Where: Equilibrium point The Jacobian matrix at point A represents the linearized representation of the dynamical system near that point, and its dimension is the same as that of the system's state space.

[0109] Riemann path planning unit 34 is connected to curvature flow prediction unit 33 to calculate the geodesic path from the current pose to the ideal pose on the risk-weighted attitude manifold, and decomposes this path into a sequence of control parameters. This unit first calculates the geodesic path from the current pose to the ideal pose on the risk-weighted attitude manifold. Define risk-weighted Riemannian measures :

[0110] ,

[0111] in: The risk-weighted Riemannian metric is a second-order covariant tensor. It is a standard Riemannian metric, usually a Euclidean metric. is the risk weighting function, which maps the risk value to the metric space; λ is the weight coefficient, which controls the degree of influence of risk factors on path planning, and is set to 2.5 in this embodiment; + indicates tensor addition operation.

[0112] Based on this metric, the geodesic equation can be expressed as:

[0113] ,

[0114] in: Let be the i-th coordinate component of a point on a geodesic line in the local coordinate system; Represents coordinate components The second derivative The Christopher symbol represents a connection on a manifold, used to describe the parallel transmission of curves on a surface; Representing coordinate components and The first derivative; The parameter can be understood as a time parameter; subscript and superscript Following Einstein's summation convention, it means summing over all coordinate components; =0 means the equation equals the zero vector.

[0115] To improve computational efficiency, this unit uses the discrete geodesic algorithm to transform the continuous problem into a discrete optimization problem. Specifically, the geodesic is discretized into... Key points ,in For the current pose point, Let the target pose point be denoted by . Minimize the discrete energy:

[0116] ,

[0117] in: For discrete energy, represent the total length of the path; Distance function under risk-weighted metric; These are two adjacent points on the path; This represents the summation of the distances between all adjacent pairs of points on the path; The total number of discrete points is 20 in this embodiment. The optimization uses gradient descent with an iteration step size of 0.01, a maximum number of iterations of 1000, and a convergence threshold of 1e-6. These parameters ensure high accuracy in geodesic calculation while maintaining computational efficiency.

[0118] After calculating the geodesic, the unit decomposes the path into time series of bed parameters (height, tilt angle) and bottle parameters (position, angle). To ensure smooth movement, cubic spline interpolation is used to generate control sequences, and acceleration limits (maximum acceleration of 0.05 m / s²) and velocity limits (maximum velocity of 0.1 m / s) are added. These limits are determined based on considerations of infant comfort and safety.

[0119] Reference Figure 4 The intestinal bloating monitoring module 4 includes a physiological signal analysis unit 41, a facial expression recognition unit 42, a thermal imaging analysis unit 43, and a risk assessment unit 44.

[0120] The physiological signal analysis unit 41 is used to analyze the infant's respiratory rate and abdominal movements. This unit uses optical flow to track minute movements in the infant's abdominal region using a depth image sequence provided by the depth camera 12, and calculates the respiratory rate and the intensity of abdominal movements. A normal newborn's respiratory rate is 30–60 breaths per minute. When a respiratory rate exceeding 60 breaths per minute or irregular breathing is detected, the system considers it one of the indicators of potential intestinal gas.

[0121] The facial expression recognition unit 42 is used to identify the infant's discomfort expressions and crying patterns. This unit receives facial images provided by the area scan camera 11 and analyzes facial expressions using a convolutional neural network (CNN) to identify facial features indicating discomfort, such as furrowed brows and downturned corners of the mouth. Furthermore, this unit analyzes the frequency and intensity characteristics of the infant's cries to distinguish crying caused by intestinal gas from crying caused by other reasons. In practice, crying caused by intestinal gas is typically characterized by high frequency, sharpness, and prolonged duration.

[0122] The thermal imaging analysis unit 43 is used to detect abnormal temperature distribution in the abdomen. This unit receives temperature distribution maps provided by the infrared thermal imager 13 and analyzes the temperature gradient and hotspot distribution in the abdominal region. Intestinal bloating often leads to elevated temperatures in certain areas of the abdomen, forming a characteristic hotspot distribution pattern. This unit uses a temperature thresholding method and a region growing algorithm to detect these hotspots. The threshold is set to the average temperature of the surrounding area plus 0.8°C; this parameter value is the optimal identification threshold determined based on clinical data analysis.

[0123] The risk assessment unit 44 is connected to the physiological signal analysis unit 41, the facial expression recognition unit 42, and the thermal imaging analysis unit 43 to comprehensively assess the risk level of intestinal flatulence based on multidimensional indicators. This unit uses a weighted fusion method to calculate the risk score.

[0124] ,

[0125] Where: Risk is the overall risk score, with a value ranging from 0 to 1; The respiratory risk score is provided by the physiological signal analysis unit 41 and ranges from 0 to 1. The facial expression risk score is provided by the facial expression recognition unit 42, and its value ranges from 0 to 1. The thermal imaging risk score is provided by the thermal imaging analysis unit 43 and ranges from 0 to 1. , where represents the weighting coefficients, indicating the importance of the three risk scores respectively; · indicates scalar multiplication. In this embodiment, the weighting coefficients are 0.3, 0.4, and 0.3 respectively. The final risk scores are divided into three levels: low risk (0-0.3), medium risk (0.3-0.7), and high risk (0.7-1.0), to guide subsequent intervention measures.

[0126] Reference Figure 5 The posture-intestinal bloating association analysis module 5 includes a time-series data storage unit 51, an association pattern mining unit 52, an individual difference learning unit 53, and a model update unit 54.

[0127] The time-series data storage unit 51 is used to record posture sequences and corresponding intestinal distension states. This unit creates a structured database to store posture feature vector sequences during the feeding process. and flatulence risk score sequence Data was collected at a frequency of 1 Hz, and each record included a timestamp, posture feature vector, flatulence risk score, and relevant environmental factors. Data was stored incrementally, with expired data periodically cleaned up to maintain historical records for the most recent 30 days. This time span is sufficient to capture individual trends without excessively consuming storage space.

[0128] The association pattern mining unit 52 is connected to the time-series data storage unit 51 and is used to analyze the temporal correlation between posture and flatulence using a sliding time window. This unit employs a time-lag correlation analysis method to calculate the correlation coefficient between posture characteristics and the risk of flatulence under different time delays.

[0129] ,

[0130] in: For time delay The correlation coefficient is given below, with a value range of [-1, 1]. For time delay, ranging from 0 to 30 minutes; Let be the pose feature vector at time t; For time t+ The risk score for flatulence; This represents the average of the pose characteristics; The average of the risk scores; This represents summing over all time points t; This represents the difference between the pose characteristic and its mean. This represents the difference between the risk score and its average. This represents the inner product operation of vectors; the two summation terms in the denominator represent the variances of the pose features and risk scores, respectively. This represents the square root operation. Through analysis... The curve indicates that the unit identifies the optimal delay time. (Usually within the range of 5 to 15 minutes), which is the typical time delay in how posture affects the state of intestinal bloating.

[0131] In addition, this unit also applies principal component analysis (PCA) and cluster analysis to identify high-risk and safe pose patterns. In the dimensionality-reduced feature space, the K-means clustering algorithm (K=5) is used to divide the poses into 5 classes, and the average risk score of each class is calculated to distinguish between high-risk and low-risk poses.

[0132] Individual difference learning unit 53 is connected to association pattern mining unit 52 to adapt to the specific physiological characteristics of infants. This unit enables personalized parameter adjustment, taking into account factors such as the infant's age, weight, and feeding method. Specifically, Bayesian methods are used to update the individual model parameters:

[0133] ,

[0134] in: Let be the posterior distribution of the model parameter 0 given the observed data D; These are model parameters, including weight coefficients, etc. The data includes historical posture and gas records; Let θ be the likelihood function, representing the probability of observed data D occurring given parameters θ. This is the prior distribution, representing an estimate of the parameters before the observed data. This indicates a direct proportionality, ignoring the normalization constant. As data accumulates, the model parameters are gradually adjusted from the population level to the individual level to adapt to the specific physiological characteristics of infants.

[0135] The model update unit 54 is connected to the individual difference learning unit 53 to continuously optimize the posture-gas association model. This unit uses an online learning method, performing incremental updates after each feeding. The update rule is as follows:

[0136] ,

[0137] in: , represents the updated model parameters; These are the model parameters before the update; The learning rate controls the step size of parameter updates, with an initial value of 0.01 that decays over time. Let L be the gradient of the loss function with respect to parameter 0, representing the direction and magnitude of the parameter update; This is newly added data; + indicates vector addition. To prevent overfitting, this unit employs early stopping and L2 regularization, with the regularization coefficient set to 0.001. Model quality is evaluated periodically; if performance degrades by more than 10%, a retraining process is triggered.

[0138] Reference Figure 6 The correction strategy generation module 6 includes a risk assessment unit 61, a target posture selection unit 62, a path planning unit 63, and a control sequence generation unit 64.

[0139] Risk assessment unit 61 is used to assess risk based on the current posture and predicted trajectory. This unit receives the posture feature vector and predicted trajectory from the differential geometry posture analysis module 3, and combines them with the correlation model from the posture-intestinal gas correlation analysis module 5 to calculate the risk score of the current posture and the risk trend for the next 5-10 seconds. The risk score is standardized from 0 to 1, and an intervention mechanism is triggered when the score is greater than 0.7.

[0140] The target posture selection unit 62 is connected to the risk assessment unit 61 and is used to select the most suitable target posture from a low-risk posture library. This unit maintains a low-risk posture library containing validated, safe, and effective feeding postures. The selection process considers three factors: risk score (weight 0.5), transition difficulty (weight 0.3), and infant comfort (weight 0.2). The posture with the highest overall score is selected as the target posture.

[0141] The path planning unit 63 is connected to the target pose selection unit 62 and is used to calculate the optimal path from the current pose to the target pose. This unit receives the Riemann path planning results from the differential geometry pose analysis module 3 and optimizes the path by combining it with actual physical constraints. The optimization objective function includes three aspects: path length, smoothness, and safety.

[0142] ,

[0143] in: To optimize the objective function, the smaller the better; The path length represents the total length of the path in the parameter space. Smoothness is measured by the rate of change of the path's curvature; Safety is a safety factor that indicates how far the path is from high-risk areas, with a value ranging from [0,1]. This indicates a safety risk factor; the higher the safety factor, the smaller this factor becomes. The weights are 0.3, 0.3, and 0.4, respectively, which control the importance of the three optimization objectives. In this embodiment, they are set to 0.3, 0.3, and 0.4. The optimization adopts the Sequential Quadratic Programming (SQP) method, with a maximum number of iterations of 200 and a convergence threshold of 1e-5.

[0144] The control sequence generation unit 64 is connected to the path planning unit 63 and is used to convert the path into a sequence of control parameters. This unit discretizes the optimal path into 20-30 key points, and calculates the required bed parameters (height, fore-aft tilt angle, lateral tilt angle) and bottle parameters (3D position, rotation angle) for each key point. Then, cubic spline interpolation is used to generate a smooth control sequence with a control frequency of 10Hz. To ensure safe and smooth operation, velocity constraints (maximum linear velocity 0.1m / s, angular velocity 0.2rad / s) and acceleration constraints (maximum linear acceleration 0.05m / s², angular acceleration 0.1rad / s²) are added.

[0145] Reference Figure 7 The execution control module 7 includes a bed control unit 71, a bottle holder control unit 72, a status monitoring unit 73, and a human-machine interaction unit 74.

[0146] The bed control unit 71 is used to adjust the bed surface height, tilt angle, and curvature. This unit includes a lifting mechanism 711, an X-direction rotation mechanism 712, and a Y-direction curvature adjustment mechanism 713. The lifting mechanism 711 uses an electric push rod to achieve precise adjustment of the bed surface height within the range of 40–70 cm, with a positioning accuracy of ±2 mm. The X-direction rotation mechanism 712 is driven by a high-precision servo motor to adjust the front-to-back tilt angle of the bed within the range of -15° to +15°, with an angle accuracy of ±0.5°. The Y-direction curvature adjustment mechanism 713 controls the lateral tilt of the bed surface, with an adjustment range of -10° to +10°. The movement direction of this mechanism is related to the curvature of the front end of the bed, ensuring that the baby's head is kept in the preset direction and preventing excessive head rotation.

[0147] The bottle holder control unit 72 is used to precisely adjust the position and angle of the bottle. This unit employs a four-degree-of-freedom robotic arm structure, including three translational joints and one rotational joint, enabling precise positioning and attitude adjustment of the bottle in three-dimensional space. Translational accuracy is ±1mm, and rotational accuracy is ±1°. The drive system uses a combination of stepper motors and harmonic reducers to ensure smooth, vibration-free movement. To prevent accidents, both electronic and mechanical limit protections are implemented to limit the maximum range of motion and force.

[0148] The status monitoring unit 73 is connected to the bed control unit 71 and the bottle holder control unit 72 for real-time monitoring of the execution effect. This unit collects position, speed, and force feedback signals from each actuator at a frequency of 100Hz. By comparing the deviation between the actual execution trajectory and the planned trajectory, the tracking error is calculated and compensated in real time. When the deviation exceeds a preset threshold (position deviation > 5mm or angle deviation > 2°), a safety pause mechanism is triggered, pausing the current action and replanning. Furthermore, this unit also monitors abnormal conditions during execution, such as actuator stalling or overload, ensuring safe and reliable system operation.

[0149] The human-computer interaction unit 74 is connected to the status monitoring unit 73 to provide status feedback and suggestions to caregivers. This unit includes a 7-inch touchscreen, an indicator light system, and a voice prompt system. The touchscreen displays the system's current status, gas risk assessment results, and operational suggestions; the indicator light system visually displays the risk level using different colors (green for low risk, yellow for medium risk, and red for high risk); and the voice prompt system announces key status changes and intervention suggestions. Furthermore, this unit provides a mobile application interface to support remote monitoring and control functions, allowing caregivers to monitor the infant's status from different locations.

[0150] Through the collaborative work of the above modules, the system of this invention realizes a complete closed loop from image acquisition and posture analysis to risk assessment, strategy generation and execution control, providing a precise and intelligent technical solution for the prevention of neonatal gas and the optimization of feeding posture.

[0151] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. An image-based system for analyzing and correcting the correlation between neonatal gas and feeding posture, characterized in that, include: The multimodal image acquisition module is used to acquire facial images, depth images, and temperature distribution images during infant feeding. An image preprocessing module, connected to the multimodal image acquisition module, is used to perform noise filtering and key point extraction on the acquired images; The differential geometry posture analysis module, connected to the image preprocessing module, is used to construct a parameterized curved manifold between the infant's head position and the bottle position, calculate posture curvature characteristics, predict posture evolution trends based on curvature flow theory, and plan the optimal correction path on the posture manifold. The gas monitoring module, connected to the multimodal image acquisition module, is used to monitor the infant's respiratory rate, abdominal activity, and body temperature distribution, and to assess the risk level of gas. The posture-bloating correlation analysis module is connected to the differential geometric posture analysis module and the bloating state monitoring module. It is used to analyze the statistical correlation between different postures and bloating state and establish a personalized correlation model. The correction strategy generation module is connected to the differential geometric posture analysis module and the posture-intestinal gas correlation analysis module, and is used to generate the optimal correction path and control parameter sequence based on the current posture risk assessment results. The execution control module, connected to the correction strategy generation module, is used to receive the control parameter sequence, drive the bed adjustment system and the bottle holder system to perform posture correction actions, and feed back the execution status to the correction strategy generation module.

2. The system according to claim 1, characterized in that, The multimodal image acquisition module includes: Area scan camera, used to capture high-resolution images of the baby's head, face, and upper body; Depth cameras are used to acquire three-dimensional spatial information and construct depth maps; Infrared thermal imagers are used to monitor the distribution of an infant's body temperature. The time synchronization unit is connected to the area array camera, the depth camera, and the infrared thermal imager to ensure the time consistency of multi-source image data.

3. The system according to claim 1, characterized in that, The differential geometric attitude analysis module includes: Surface manifold building units are used to construct parametric surface manifolds for baby head pose and bottle position based on key points; The curvature feature calculation unit, connected to the surface manifold construction unit, is used to calculate the Gaussian curvature and average curvature on the manifold and extract the attitude feature vector; The curvature flow prediction unit, connected to the curvature feature calculation unit, is used to construct discrete curvature flow equations, predict posture evolution trajectories, and identify equilibrium points and unstable regions in the posture space. The Riemann path planning unit, connected to the curvature flow prediction unit, is used to calculate the geodesic path from the current pose to the ideal pose on the risk-weighted attitude manifold and decompose the path into a sequence of control parameters.

4. The system according to claim 3, characterized in that, The curvature feature calculation unit extracts attitude features through the following steps: Divide the key points of the infant's head and face into multiple local areas; Calculate the principal curvature and principal direction for each local region; Construct a curvature histogram descriptor to represent local shape features; The features of each region are weighted and fused to form a global pose feature vector.

5. The system according to claim 3, characterized in that, The curvature flow prediction unit employs a multi-timescale prediction strategy: Short-timescale forecasting uses a locally linearized model for real-time feedback control; Mid-timescale prediction combined with local nonlinear dynamics is used for attitude stability assessment; Long-term scale prediction integrates statistical and dynamic models for individualized pattern analysis.

6. The system according to claim 1, characterized in that, The intestinal gas monitoring module includes: The physiological signal analysis unit is used to analyze the infant's respiratory rate and abdominal movements; A facial expression recognition unit is used to identify infant discomfort expressions and crying patterns; Thermal imaging analysis unit, used to detect abnormal temperature distribution in the abdomen; The risk assessment unit, connected to the physiological signal analysis unit, the facial expression recognition unit, and the thermal imaging analysis unit, is used to comprehensively assess the risk level of intestinal bloating based on multidimensional indicators.

7. The system according to claim 1, characterized in that, The posture-bloating correlation analysis module includes: A time-series data storage unit is used to record posture sequences and corresponding intestinal bloating states; The association pattern mining unit is connected to the time-series data storage unit and is used to analyze the temporal correlation between posture and flatulence using a sliding time window. An individual difference learning unit, connected to the association pattern mining unit, is used to adapt to the physiological characteristics of different infants; The model update unit, connected to the individual difference learning unit, is used to continuously optimize the posture-bloating association model.

8. The system according to claim 1, characterized in that, The correction strategy generation module includes: Risk assessment unit, used to assess risk based on current posture and predicted trajectory; A target pose selection unit, connected to the risk assessment unit, is used to select the most suitable target pose from a low-risk pose library; The path planning unit, connected to the target pose selection unit, is used to calculate the optimal path from the current pose to the target pose; A control sequence generation unit, connected to the path planning unit, is used to convert the path into a sequence of control parameters.

9. The system according to claim 1, characterized in that, The execution control module includes: The bed control unit is used to adjust the bed surface height, tilt angle, and curvature; The bottle holder control unit is used to precisely adjust the position and angle of the bottle; A status monitoring unit, connected to the bed control unit and the bottle holder control unit, is used to monitor the execution effect in real time. The human-computer interaction unit, connected to the status monitoring unit, is used to provide status feedback and suggestions to the caregiver.

10. The system according to claim 9, characterized in that, The bed control unit includes: The lifting mechanism is used to adjust the height of the bed surface; The X-direction rotation mechanism is used to drive the bed surface to rotate in the X direction on the horizontal plane of the bed surface. The Y-direction curvature adjustment mechanism is used to drive the bed surface to rotate in the Y direction according to the curvature of the front end of the bed surface; The movement direction of the Y-direction curvature adjustment mechanism is related to the curvature of the front end of the bed, so that the baby's head is kept in a preset direction.

Images