Intelligent shaking table system for pediatric nursing based on feature data feedback

Abstract

The invention discloses a pediatric nursing intelligent shaking table system based on feature data feedback, and the system comprises the steps: carrying out the preprocessing of multi-dimensional pediatric data, extracting long-term dependence features in the multi-dimensional pediatric data through a Transform encoder based on a CLP multi-mode base model, extracting spatial-temporal features through 3D-CNN, generating multi-scale acoustic features according to WaveNet, and carrying out the feedback of the multi-scale acoustic features. And constructing a causal graph of infant states and nursing actions according to the global state vector by using a PC algorithm, deducing a causal relationship between variables from historical data, simulating potential results under different nursing interventions through a BSTS Bayesian structure time sequence, and dynamically adjusting the action amplitude, temperature control parameters and sound pacifying parameters of the shaking table according to a shaking table control strategy. The nursing efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical equipment technology, and in particular to an intelligent pediatric nursing rocking bed system based on feature data feedback. Background Technology

[0002] Traditional pediatric nursing methods rely heavily on the experience and manual operation of nurses, resulting in low efficiency and insufficient precision. When determining the cause of an infant's crying, nurses often have to rely on limited observation and experience to guess, making it difficult to quickly and accurately find the root cause and take effective nursing measures, thus leading to low efficiency in pediatric nursing. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing an intelligent pediatric nursing rocking bed system based on feature data feedback.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following modules in the above-mentioned intelligent pediatric care rocking bed system based on feature data feedback:

[0005] A multi-dimensional data acquisition module is used to collect multi-dimensional pediatric data and preprocess the multi-dimensional pediatric data to obtain initial multi-dimensional pediatric data.

[0006] The multimodal feature fusion module is used to extract long-term dependency features from the multi-dimensional pediatric data based on the CLP multimodal base model, using the Transformer encoder, extracting spatiotemporal features through 3D-CNN, and generating multi-scale acoustic features based on WaveNet to obtain multimodal feature data.

[0007] The state vector unification module is used to align the feature representations of different modalities in the multimodal feature data through multimodal cross-attention, and generate a unified global state vector.

[0008] The rocking bed strategy generation module is used to construct a causal graph of infant state and care actions based on the global state vector using PC algorithm, infer causal relationships between variables from historical data, simulate potential outcomes under different care interventions through BSTS Bayesian structure time series simulation, and generate rocking bed control strategies.

[0009] The shaking table dynamic adjustment module is used to dynamically adjust the shaking table's motion amplitude, temperature control parameters, and sound soothing parameters according to the shaking table control strategy.

[0010] Furthermore, in the aforementioned intelligent pediatric nursing rocker system based on feature data feedback, the multi-dimensional data acquisition module includes the following units:

[0011] The data acquisition unit is used to collect infant physiological data, environmental parameters, nursing record data, and infant behavioral data to obtain multi-dimensional pediatric data.

[0012] The data cleaning unit is used to set reasonable threshold ranges for physiological data and environmental parameters, and to delete abnormal data points that exceed the threshold range.

[0013] The interpolation and deduplication unit is used to check for missing and duplicate values ​​in nursing record data. Missing values ​​are interpolated using the average of adjacent time points, and duplicate values ​​are deduplicated.

[0014] The normalization synchronization unit is used to convert data of different dimensions and scales into values ​​within the range of [0,1]. Based on the operation time in the nursing record data, the physiological data, environmental parameters and infant behavior data are time-aligned to obtain initial multidimensional pediatric data.

[0015] Furthermore, in the aforementioned intelligent pediatric nursing rocker system based on feature data feedback, the multimodal feature fusion module includes the following units:

[0016] The input processing unit is used to arrange the preprocessed initial multidimensional pediatric data in a time series. The input vector at each time point is composed of normalized vectors of physiological data, environmental parameters, nursing record data and infant behavior data.

[0017] The weight calculation unit is used in the Transformer encoder to generate the query vector, key vector, and value vector from the input vector through a linear transformation at each time point, calculate the dot product similarity between the query vector and the key vector, and obtain the attention weights.

[0018] The attention output unit is used to perform softmax normalization on the attention weights, multiply them by the value vector to obtain a weighted sum, and then obtain the self-attention output.

[0019] Furthermore, in the aforementioned intelligent pediatric nursing rocker system based on feature data feedback, the multimodal feature fusion module further includes the following units:

[0020] The 3D convolution operation unit is used to perform convolution operations simultaneously in the spatial and temporal dimensions using the convolution kernels of 3D-CNN to extract the spatiotemporal features of video data in the initial multidimensional pediatric data.

[0021] The feature output unit is used to flatten the output of the last layer into a one-dimensional vector after multiple 3D convolution and pooling operations, and output the spatiotemporal feature representation of infant behavioral data.

[0022] Furthermore, in the aforementioned intelligent pediatric care rocker system based on feature data feedback, the state vector unification module includes the following units:

[0023] The multimodal feature representation unit is used to unify the dimensions of the long-term dependency features extracted by the Transformer encoder, the spatiotemporal features extracted by 3D-CNN, and the multi-scale acoustic features generated by WaveNet, and transform them into feature vectors of the same dimension through linear transformation to obtain multimodal feature data.

[0024] The intermodal attention calculation unit is used to calculate the cross attention between every two modalities of the multimodal feature data. Through a multi-layer cross attention mechanism, the features of each modality can pay attention to the feature representation, and the cross attention output is obtained.

[0025] The global state vector generation unit is used to fuse the cross-attention outputs using a weighted summation method to generate a unified global state vector.

[0026] Furthermore, in the aforementioned intelligent pediatric care rocking bed system based on feature data feedback, the rocking bed strategy generation module includes the following units:

[0027] The independence test unit is used to perform conditional independence tests through the chi-square test, delete edges that do not meet the conditional independence, and construct a causal graph structure between infant state and care actions;

[0028] The causal relationship inference unit is used to infer the strength of causal relationships between variables from historical data based on the constructed causal graph, and to estimate the weights of each edge in the causal graph using regression analysis to obtain the target causal graph structure.

[0029] Furthermore, in the aforementioned intelligent pediatric care rocking bed system based on feature data feedback, the rocking bed strategy generation module further includes the following units:

[0030] The potential outcome simulation unit is used to simulate the trend of changes in infant state under different nursing interventions based on the target causal graph structure using the BSTS model.

[0031] The control strategy generation unit is used to select the optimal nursing intervention plan from the potential results of the simulation with the goal of optimizing the infant's state, and generate the rocking crib control strategy.

[0032] Furthermore, in the method for implementing a pediatric nursing intelligent rocking bed system based on feature data feedback, the method includes the following steps:

[0033] Collect multi-dimensional pediatric data, and preprocess the multi-dimensional pediatric data to obtain initial multi-dimensional pediatric data;

[0034] Based on the CLP multimodal base model, the Transformer encoder is used to extract long-term dependency features from the multidimensional pediatric data, spatiotemporal features are extracted by 3D-CNN, and multi-scale acoustic features are generated according to WaveNet to obtain multimodal feature data.

[0035] By aligning the feature representations of different modalities in the multimodal feature data through multimodal cross-attention, a unified global state vector is generated;

[0036] Using the PC algorithm, a causal graph of infant state and nursing actions is constructed based on the global state vector. Causal relationships between variables are inferred from historical data. The potential outcomes under different nursing interventions are simulated through BSTS Bayesian time series simulation to generate a rocking crib control strategy.

[0037] The shaking table's motion amplitude, temperature control parameters, and sound soothing parameters are dynamically adjusted according to the shaking table control strategy.

[0038] Furthermore, in the method for implementing a pediatric nursing intelligent rocking bed system based on feature data feedback, the method includes the following steps:

[0039] Conditional independence was tested using the chi-square test, edges that did not satisfy conditional independence were removed, and a causal graph structure between infant state and care actions was constructed.

[0040] Based on the constructed causal graph, the strength of the causal relationship between variables is inferred from historical data. Regression analysis is used to estimate the weights of each edge in the causal graph, thus obtaining the target causal graph structure.

[0041] Furthermore, in the method for implementing a pediatric nursing intelligent rocking bed system based on feature data feedback, the method includes the following steps:

[0042] Based on the target cause-effect diagram structure, the BSTS model was used to simulate the trend of infant state changes under different nursing interventions.

[0043] With the goal of optimizing the infant's condition, the optimal nursing intervention plan is selected from the potential results of the simulation using the particle swarm optimization algorithm, and a rocking bed control strategy is generated.

[0044] Its beneficial effects are as follows: 1. It avoids misjudgments caused by analysis of a single data dimension, providing a reliable basis for subsequent nursing strategy formulation. It can more accurately determine whether an infant's crying is due to hunger, physical discomfort, or environmental factors, improving the pertinence and effectiveness of nursing care. 2. It can dynamically adjust the amplitude of the rocking motion, temperature control parameters, and soothing sound parameters according to the infant's real-time status, providing personalized nursing services, improving nursing efficiency and infant comfort. 3. It improves nursing efficiency and quality. Nursing staff can focus more on complex nursing tasks, while reducing potential risks to the infant caused by improper manual operation, ensuring the infant's health and safety. Attached Figure Description

[0045] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0046] Figure 1 This is a schematic diagram of the first embodiment of the intelligent pediatric nursing rocking bed system based on feature data feedback in this invention.

[0047] Figure 2 This is a schematic diagram of the second embodiment of the intelligent pediatric nursing rocking bed system based on feature data feedback in this invention.

[0048] Figure 3 This is a schematic diagram of the third embodiment of the intelligent pediatric nursing rocking bed system based on feature data feedback in this invention. Detailed Implementation

[0049] 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.

[0050] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0051] The present invention will now be described in detail with reference to the accompanying drawings, such as... Figure 1 As shown, the intelligent pediatric nursing shaker system based on feature data feedback includes the following modules:

[0052] The multi-dimensional data acquisition module is used to collect multi-dimensional pediatric data, preprocess the multi-dimensional pediatric data, and obtain initial multi-dimensional pediatric data.

[0053] Specifically, this embodiment also includes a data acquisition unit for collecting infant physiological data, environmental parameters, nursing record data, and infant behavioral data to obtain multi-dimensional pediatric data;

[0054] The data cleaning unit is used to set reasonable threshold ranges for physiological data and environmental parameters, and to delete abnormal data points that exceed the threshold range.

[0055] The interpolation and deduplication unit is used to check for missing and duplicate values ​​in nursing record data. Missing values ​​are interpolated using the average of adjacent time points, and duplicate values ​​are deduplicated.

[0056] The normalization synchronization unit is used to convert data of different dimensions and scales into values ​​within the range of [0,1]. Based on the operation time in the nursing record data, the physiological data, environmental parameters and infant behavior data are time-aligned to obtain initial multidimensional pediatric data.

[0057] Specifically,

[0058] (I) Multidimensional Pediatric Data Collection

[0059] Infant physiological data acquisition: Wearable physiological monitoring devices, such as smart baby clothes, integrate high-precision sensors (a PPG photoplethysmography (PPG) heart rate sensor with an accuracy of ±1 beat / minute; a DS18B20 digital temperature sensor with an accuracy of ±0.5℃; and a piezoelectric film sensor for respiratory rate monitoring, which can capture chest rise and fall signals in real time), to collect real-time physiological data such as the infant's heart rate, body temperature, respiratory rate, and blood oxygen saturation. The acquisition frequency is set to once per second to ensure the real-time nature and accuracy of the data.

[0060] Environmental parameter acquisition: An environmental monitoring module is deployed in the infant's care environment, including a temperature and humidity sensor (SHT30, measurement range: temperature -40℃~125℃, humidity 0%RH~100%RH, accuracy: temperature ±0.2℃, humidity ±2%RH), a light intensity sensor (TSL2561, measurement range: 0.1~400000 lux), and a noise sensor (LM393, capable of detecting sound intensity from 30dB to 130dB), to collect environmental parameters such as temperature, humidity, light intensity, and noise decibels in real time, with a sampling frequency of once per minute.

[0061] Nursing record data collection: A dedicated mobile application was developed for nursing staff. During nursing procedures such as feeding, diaper changing, bathing, and soothing, staff can use this application to record data in real time, including the type, time, duration, and method of the procedure (e.g., amount of milk, formula brand; diaper wetness / dryness). The system automatically records the entry time and synchronizes it with other data.

[0062] Infant Behavioral Data Acquisition: High-definition cameras (1920×1080 resolution, 30fps) are installed around the crib. Combined with computer vision technology, these cameras capture real-time data on the infant's physical movements (such as waving arms, kicking legs, rolling over, etc.) and facial expressions (such as smiling, frowning, and facial features when crying). Simultaneously, microphones capture the infant's cries, laughter, and other sound signals for subsequent acoustic feature analysis. The camera and microphone capture data at the same frequency as the video frame rate, 30fps.

[0063] (ii) Data preprocessing

[0064] Data cleaning: A data cleaning strategy combining statistical methods and machine learning algorithms was employed. First, for physiological data and environmental parameters, reasonable threshold ranges were set (e.g., normal heart rate range of 60-160 beats / minute, normal body temperature range of 36-37.5℃) to remove outlier data points exceeding these ranges. For nursing record data, missing and duplicate values ​​were checked; missing values ​​were imputed using the mean of adjacent time points, and duplicate values ​​were deduplicated. For infant behavioral data, video denoising algorithms (e.g., median filtering, mean filtering) were used to remove noise interference from the video, and audio denoising techniques (e.g., wavelet denoising) were used to remove background noise from the audio signal.

[0065] Data normalization: Converting data of different dimensions and units into values ​​within the range of [0,1]. For continuous data (such as physiological data and environmental parameters), the min-max normalization method is used. For discrete data (such as nursing operation types), one-hot encoding is used to convert each category into a binary vector.

[0066] Time synchronization: Based on the operation time in the nursing record data, physiological data, environmental parameters, and infant behavior data are time-aligned. For data with different collection frequencies, linear interpolation is used to unify them to the same time interval (set to 1 minute) to ensure that the data of each dimension corresponds one-to-one in time during subsequent analysis.

[0067] The multimodal feature fusion module is used to extract long-term dependency features from multi-dimensional pediatric data based on the CLP multimodal base model, extract spatiotemporal features through 3D-CNN, and generate multi-scale acoustic features based on WaveNet to obtain multimodal feature data.

[0068] Specifically, this embodiment also includes an input processing unit, which is used to arrange the preprocessed initial multi-dimensional pediatric data in a time series. The input vector at each time point is composed of normalized vectors of physiological data, environmental parameters, nursing record data and infant behavior data.

[0069] The weight calculation unit is used in the Transformer encoder to generate the query vector, key vector, and value vector from the input vector through a linear transformation at each time point, calculate the dot product similarity between the query vector and the key vector, and obtain the attention weights.

[0070] The attention output unit is used to perform softmax normalization on the attention weights, multiply them by the value vector to obtain a weighted sum, and then obtain the self-attention output.

[0071] The 3D convolution operation unit is used to perform convolution operations simultaneously in the spatial and temporal dimensions using the convolution kernels of 3D-CNN to extract the spatiotemporal features of video data in the initial multidimensional pediatric data.

[0072] The feature output unit is used to flatten the output of the last layer into a one-dimensional vector after multiple 3D convolution and pooling operations, and output the spatiotemporal feature representation of infant behavioral data.

[0073] Specifically,

[0074] I. Construction of CLP Multimodal Base Model

[0075] The CLP multimodal foundation model is the basic framework for the entire feature extraction process, consisting of three main layers. The input layer is responsible for receiving preprocessed initial multidimensional pediatric data, which covers different modalities such as infant physiological data, environmental parameters, nursing record data, and infant behavioral data.

[0076] The feature extraction layer is a crucial part of the model, employing Transformer encoders, 3D-CNN, and WaveNet to extract features tailored to the characteristics of different modalities. Transformer encoders excel at processing data with time-series properties, used to uncover long-term dependencies; 3D-CNN focuses on processing video data, extracting spatiotemporal features; and WaveNet is primarily used for audio data, generating multi-scale acoustic features.

[0077] The role of the fusion layer is to initially fuse the multimodal features extracted by the three models mentioned above, integrating features from different sources that reflect different states of the infant, and providing rich and valuable input for the subsequent multimodal cross-attention mechanism.

[0078] II. Long-term Dependency Feature Extraction (Transformer Encoder)

[0079] First, the preprocessed multidimensional pediatric data was organized and arranged into a time series according to chronological order. The input vector at each time point was composed of normalized vectors of physiological data, environmental parameters, nursing record data, and infant behavioral data, forming a multidimensional time series data containing various aspects of information.

[0080] The Transformer encoder uses a self-attention mechanism to process this data. It generates a query vector, a key vector, and a value vector for the data at each time point. By calculating the similarity between the query vector and the key vector, it determines the degree of attention each time point pays to the data from other time points, i.e., the attention weight.

[0081] To extract long-term dependency features from the data more deeply, a multi-layer Transformer encoder, specifically six layers, is employed. Each encoder layer contains two sub-layers: a self-attention mechanism and a feedforward neural network. Residual connections and layer normalization techniques are used between the sub-layers. Residual connections prevent gradient vanishing during training, while layer normalization helps accelerate training and improve generalization ability. Through this multi-layer encoding, the Transformer encoder can capture complex relationships between different time points and uncover long-term dependency features in the data.

[0082] III. Spatiotemporal Feature Extraction (3D-CNN)

[0083] For video sequences in infant behavior data, they are first split into a sequence of frames, each frame being a common RGB three-channel image with a size of 224×224 pixels. Sixteen consecutive frames are selected to form a video clip, which serves as input to a 3D-CNN, allowing for the complete capture of changes in an infant's behavior over a period of time.

[0084] 3D-CNN convolutional kernels perform convolution operations simultaneously along the width, height, and time dimensions of the image. The first 3D convolutional layer uses a 3×3×3 kernel with a stride of 1 and padding of 1, outputting a feature map with 64 channels. Subsequent 3D convolutional layers, from the second to the fourth, progressively increase the number of channels, setting them to 128, 256, and 512 channels respectively, and then adding 2×2×2 pooling layers after these layers for downsampling.

[0085] After multiple layers of 3D convolution and pooling operations, the output of the last layer is flattened into a one-dimensional vector. This one-dimensional vector integrates the spatial action features and temporal variation features of infant behavior, serving as a spatiotemporal feature representation of infant behavior data for subsequent analysis and processing.

[0086] IV. Multi-scale acoustic feature generation (WaveNet)

[0087] The acquired infant sound signals are first preprocessed, including framing and windowing. During framing, each frame is set to 512 sample points, with a frame shift of 256 sample points, dividing the continuous sound signal into smaller segments for processing. Hamming windows are used for windowing to reduce spectral leakage during framing, making the signal more suitable for subsequent processing. After these operations, the sound signal is converted into a spectrogram and input into WaveNet.

[0088] WaveNet employs a dilated causal convolutional structure with convolutional layers using different dilation rates: 1, 2, 4, 8, 16, and 32. This structure enables the model to capture the temporal dependencies of audio signals at different scales, effectively capturing both short-term changes and long-term trends in sound. Each convolutional layer is configured with 256 filters, using the ReLU function as the activation function, and a gating mechanism is introduced, employing the sigmoid function to control the flow of information, determining which information can pass through and which needs to be suppressed.

[0089] After undergoing multiple layers of dilated causal convolution, WaveNet outputs multi-scale acoustic feature vectors. These feature vectors contain acoustic features such as pitch, loudness, and tempo of the baby's cry, providing important information for analyzing the baby's emotional state and needs.

[0090] The state vector unification module is used to align the feature representations of different modalities in multimodal feature data through multimodal cross-attention, and generate a unified global state vector.

[0091] Specifically, this embodiment also includes a multimodal feature representation unit, which is used to unify the dimensions of the long-term dependency features extracted by the Transformer encoder, the spatiotemporal features extracted by the 3D-CNN, and the multi-scale acoustic features generated by WaveNet, and convert them into feature vectors of the same dimension through linear transformation to obtain multimodal feature data.

[0092] The intermodal attention computation unit is used to calculate the cross-attention between every two modalities of multimodal feature data. Through a multi-layer cross-attention mechanism, the features of each modality can pay attention to the feature representation, and the cross-attention output is obtained.

[0093] The global state vector generation unit is used to fuse the cross-attention outputs using a weighted summation method to generate a unified global state vector.

[0094] Specifically,

[0095] I. Multimodal Feature Representation Processing

[0096] After completing the multimodal feature extraction, we obtained the long-term dependency features extracted by the Transformer encoder, the spatiotemporal features extracted by the 3D-CNN, and the multi-scale acoustic features generated by WaveNet. First, we performed dimensionality unification on these features.

[0097] Through specific linear transformation operations, long-term dependent features, spatiotemporal features, and multi-scale acoustic features are all converted into feature vectors of the same dimension, which is uniformly set to 512 here. After processing, the feature data of different modalities are consistent in form, forming a multimodal feature data set.

[0098] II. Intermodal Cross-Attention Calculation

[0099] After obtaining multimodal feature data of a unified dimension, the next step is to use a cross-attention mechanism to calculate the attention relationships between different modalities. There are potential correlations between the multimodal data.

[0100] The calculated similarity is normalized so that the sum of all attention weights is 1, thus obtaining the relative importance of different behavioral information in the current physiological state. The normalized attention weights are multiplied by the value vector of the behavioral modality features and summed to obtain the cross-attention output between the physiological modality and the behavioral modality.

[0101] III. Global State Vector Generation

[0102] Through multi-layer cross-attention mechanism computation, the features of each modality fully interact and align with those of other modalities. Each layer of cross-attention computation makes the correlation between different modalities closer and the feature representation more coordinated. After multi-layer processing, multiple cross-attention outputs are obtained, which reflect the fusion information between multimodal data from different perspectives.

[0103] To integrate these scattered cross-attention outputs into a unified vector that comprehensively describes the infant's current state, a weighted summation method is used for fusion. After the cross-attention outputs of all modalities are weighted and summed according to the optimal weights, a unified global state vector is generated.

[0104] The rocking bed strategy generation module is used to construct a causal graph of infant state and care actions based on the global state vector using PC algorithm, infer the causal relationship between variables from historical data, simulate the potential results under different care interventions through BSTS Bayesian structure time series simulation, and generate rocking bed control strategies.

[0105] Specifically, this embodiment also includes an independence testing unit, which is used to perform conditional independence testing through chi-square test, delete edges that do not meet conditional independence, and construct a causal graph structure between infant state and care actions;

[0106] The causal relationship inference unit is used to infer the strength of causal relationships between variables from historical data based on the constructed causal graph, and to estimate the weights of each edge in the causal graph using regression analysis to obtain the target causal graph structure.

[0107] The potential outcome simulation unit is used to simulate the trend of changes in infant state under different nursing interventions based on the target causal graph structure using the BSTS model.

[0108] The control strategy generation unit is used to select the optimal nursing intervention plan from the potential results of the simulation with the goal of optimizing the infant's state, and generate the rocking crib control strategy.

[0109] Specifically,

[0110] (a) PC algorithm for constructing cause-effect graphs

[0111] Variable definition: Infant state variables (such as heart rate, body temperature, crying state, etc.) and nursing action variables (such as the amplitude of rocking motion, temperature control parameters, sound soothing parameters, etc.) are used as nodes in the cause-and-effect graph.

[0112] Conditional Independence Test: A stepwise screening process using the PC algorithm is employed. First, the correlation coefficients between all variable pairs are calculated, and edges with correlation coefficients below a threshold (set to 0.3) are removed. Then, given other variables, conditional independence is tested using the chi-square test or Fisher's exact test, progressively removing edges that do not meet the conditional independence requirement, ultimately constructing a causal graph structure between infant state and care actions.

[0113] Causal relationship inference: Based on the constructed causal graph, the strength of the causal relationship between variables is inferred from historical data. Regression analysis or structural equation modeling are used to estimate the weights of each side in the causal graph and clarify the degree and direction of the influence of nursing actions on the infant's condition.

[0114] (II) Potential Results of Bayesian Structure-Time Series Simulation of BSTS

[0115] Model Construction: The BSTS model decomposes time series data into trend, seasonal, and irregular components, while considering the structural relationships between variables. For each nursing action variable (such as the amplitude of rocking motion), a dynamic model is established between it and the infant's state variable. The model parameters are estimated using Bayesian methods, and the posterior distribution of the parameters is obtained by sampling through the Markov Chain Monte Carlo (MCMC) algorithm.

[0116] Potential Outcome Simulation: Based on the known current infant state and cause-effect graph structure, the BSTS model is used to simulate the trend of changes in infant state under different nursing interventions (such as different rocking motion amplitudes and different combinations of temperature control parameters) and predict possible potential outcomes (such as whether the infant's heart rate returns to normal or whether the crying stops).

[0117] Control strategy generation: With the goal of optimizing the infant's condition (such as stabilizing the infant's heart rate within the normal range and reducing crying time), optimization algorithms (such as particle swarm optimization algorithm and genetic algorithm) are used to screen the optimal nursing intervention plan from the potential results of the simulation and generate the rocking crib control strategy, including the adjustment range and frequency of the rocking crib movement, the set value of the temperature control parameter, and the sound soothing parameters (such as the frequency, intensity and duration of the sound).

[0118] The shaking table dynamic adjustment module is used to dynamically adjust the shaking table's motion amplitude, temperature control parameters, and sound soothing parameters according to the shaking table control strategy.

[0119] Specifically, in this embodiment,

[0120] (a) Adjustment of movement range

[0121] Based on the generated rocking crib control strategy, the rocking crib's movement amplitude is adjusted in real time. The rocking crib's movement amplitude is divided into horizontal rocking amplitude and vertical undulation amplitude. The horizontal rocking amplitude range is set to 0-10 cm, with an adjustment accuracy of 1 cm; the vertical undulation amplitude range is set to 0-5 cm, with an adjustment accuracy of 0.5 cm. When the baby shows signs of restlessness such as crying or frequent limb movements, the rocking crib's movement amplitude is gradually increased to simulate the rocking sensation of a mother holding the baby and soothe the baby. When the baby's state becomes stable, the movement amplitude is gradually decreased to avoid excessive shaking that may cause discomfort to the baby.

[0122] (ii) Temperature control parameter adjustment

[0123] The temperature control system dynamically adjusts the temperature inside the rocker based on ambient temperature and the baby's body temperature. The ideal temperature range inside the rocker is set to 24-26℃. When the ambient temperature is below 24℃, the heating device is activated, and the heating power automatically adjusts according to the temperature difference (10W increase for every 1℃ increase in temperature difference). When the ambient temperature is above 26℃, the cooling device is activated, and the cooling power also automatically adjusts according to the temperature difference (15W increase for every 1℃ increase in temperature difference). Simultaneously, the system fine-tunes the temperature control parameters in real time based on changes in the baby's body temperature to ensure the baby is in a comfortable temperature environment.

[0124] (III) Adjustment of sound soothing parameters

[0125] The sound soothing module generates appropriate soothing sounds based on the characteristics of the baby's cries and the decibel level of ambient noise. The frequency range of the sound is set to 200-1000Hz (simulating the sound of a mother's heartbeat and soft music), and the intensity range is 40-60dB (to avoid excessively high or low sounds that may stimulate the baby). When the baby's cries are high-frequency and high-intensity, a lower-frequency, moderately intense soothing sound is played; when there is high ambient noise, the intensity of the soothing sound is appropriately increased to mask the noise interference. The duration of the sound is adjusted according to the baby's state. After the baby stops crying and falls asleep, the sound continues to play for 5-10 minutes, then the sound intensity is gradually reduced until it is turned off.

[0126] Its beneficial effects are as follows: 1. It avoids misjudgments caused by analysis of a single data dimension, providing a reliable basis for subsequent nursing strategy formulation. It can more accurately determine whether an infant's crying is due to hunger, physical discomfort, or environmental factors, improving the pertinence and effectiveness of nursing care. 2. It can dynamically adjust the amplitude of the rocking motion, temperature control parameters, and soothing sound parameters according to the infant's real-time status, providing personalized nursing services, improving nursing efficiency and infant comfort. 3. It improves nursing efficiency and quality. Nursing staff can focus more on complex nursing tasks, while reducing potential risks to the infant caused by improper manual operation, ensuring the infant's health and safety.

[0127] Please see Figure 2 In the pediatric nursing intelligent rocking bed system based on feature data feedback, the multi-dimensional data acquisition module includes the following units:

[0128] The data acquisition unit is used to collect infant physiological data, environmental parameters, nursing record data, and infant behavioral data to obtain multi-dimensional pediatric data.

[0129] The data cleaning unit is used to set reasonable threshold ranges for physiological data and environmental parameters, and to delete abnormal data points that exceed the threshold range.

[0130] The interpolation and deduplication unit is used to check for missing and duplicate values ​​in nursing record data. Missing values ​​are interpolated using the average of adjacent time points, and duplicate values ​​are deduplicated.

[0131] The normalization synchronization unit is used to convert data of different dimensions and scales into values ​​within the range of [0,1]. Based on the operation time in the nursing record data, the physiological data, environmental parameters and infant behavior data are time-aligned to obtain initial multidimensional pediatric data.

[0132] Please see Figure 3 In the pediatric nursing intelligent rocker system based on feature data feedback, the state vector unification module includes the following units:

[0133] The multimodal feature representation unit is used to unify the dimensions of the long-term dependency features extracted by the Transformer encoder, the spatiotemporal features extracted by 3D-CNN, and the multi-scale acoustic features generated by WaveNet, and transform them into feature vectors of the same dimension through linear transformation to obtain multimodal feature data.

[0134] The intermodal attention computation unit is used to calculate the cross-attention between every two modalities of multimodal feature data. Through a multi-layer cross-attention mechanism, the features of each modality can pay attention to the feature representation, and the cross-attention output is obtained.

[0135] The global state vector generation unit is used to fuse the cross-attention outputs using a weighted summation method to generate a unified global state vector.

[0136] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A pediatric nursing intelligent rocking bed system based on feature data feedback, characterized in that, The intelligent pediatric nursing rocking bed system includes the following modules: A multi-dimensional data acquisition module is used to collect multi-dimensional pediatric data and preprocess the multi-dimensional pediatric data to obtain initial multi-dimensional pediatric data. The multimodal feature fusion module is used to extract long-term dependency features from the multi-dimensional pediatric data based on the CLP multimodal base model, using the Transformer encoder, extracting spatiotemporal features through 3D-CNN, and generating multi-scale acoustic features based on WaveNet to obtain multimodal feature data. The state vector unification module is used to align the feature representations of different modalities in the multimodal feature data through multimodal cross-attention, and generate a unified global state vector. The rocking bed strategy generation module is used to construct a causal graph of infant state and care actions based on the global state vector using PC algorithm, infer causal relationships between variables from historical data, simulate potential outcomes under different care interventions through BSTS Bayesian structure time series simulation, and generate rocking bed control strategies. The shaking table dynamic adjustment module is used to dynamically adjust the shaking table's motion amplitude, temperature control parameters, and sound soothing parameters according to the shaking table control strategy.

2. The intelligent pediatric nursing rocking bed system based on feature data feedback as described in claim 1, characterized in that, The multi-dimensional data acquisition module includes the following units: The data acquisition unit is used to collect infant physiological data, environmental parameters, nursing record data, and infant behavioral data to obtain multi-dimensional pediatric data. The data cleaning unit is used to set reasonable threshold ranges for physiological data and environmental parameters, and to delete abnormal data points that exceed the threshold range. The interpolation and deduplication unit is used to check for missing and duplicate values ​​in nursing record data. Missing values ​​are interpolated using the average of adjacent time points, and duplicate values ​​are deduplicated. The normalization synchronization unit is used to convert data of different dimensions and scales into values ​​within the range of [0,1]. Based on the operation time in the nursing record data, the physiological data, environmental parameters and infant behavior data are time-aligned to obtain initial multidimensional pediatric data.

3. The intelligent pediatric nursing rocking bed system based on feature data feedback as described in claim 1, characterized in that, The multimodal feature fusion module includes the following units: The input processing unit is used to arrange the preprocessed initial multidimensional pediatric data in a time series. The input vector at each time point is composed of normalized vectors of physiological data, environmental parameters, nursing record data and infant behavior data. The weight calculation unit is used in the Transformer encoder to generate the query vector, key vector, and value vector from the input vector through a linear transformation at each time point, calculate the dot product similarity between the query vector and the key vector, and obtain the attention weights. The attention output unit is used to perform softmax normalization on the attention weights, multiply them by the value vector to obtain a weighted sum, and then obtain the self-attention output.

4. The intelligent pediatric nursing rocking bed system based on feature data feedback as described in claim 3, characterized in that, The multimodal feature fusion module also includes the following units: The 3D convolution operation unit is used to perform convolution operations simultaneously in the spatial and temporal dimensions using the convolution kernels of 3D-CNN to extract the spatiotemporal features of video data in the initial multidimensional pediatric data. The feature output unit is used to flatten the output of the last layer into a one-dimensional vector after multiple 3D convolution and pooling operations, and output the spatiotemporal feature representation of infant behavioral data.

5. The intelligent pediatric nursing rocking bed system based on feature data feedback as described in claim 1, characterized in that, The state vector unification module includes the following units: The multimodal feature representation unit is used to unify the dimensions of the long-term dependency features extracted by the Transformer encoder, the spatiotemporal features extracted by 3D-CNN, and the multi-scale acoustic features generated by WaveNet, and transform them into feature vectors of the same dimension through linear transformation to obtain multimodal feature data. The intermodal attention calculation unit is used to calculate the cross attention between every two modalities of the multimodal feature data. Through a multi-layer cross attention mechanism, the features of each modality can pay attention to the feature representation, and the cross attention output is obtained. The global state vector generation unit is used to fuse the cross-attention outputs using a weighted summation method to generate a unified global state vector.

6. The pediatric nursing intelligent rocking bed system based on feature data feedback as described in claim 1, characterized in that, The shaking table strategy generation module includes the following units: The independence test unit is used to perform conditional independence tests through the chi-square test, delete edges that do not meet the conditional independence, and construct a causal graph structure between infant state and care actions; The causal relationship inference unit is used to infer the strength of causal relationships between variables from historical data based on the constructed causal graph, and to estimate the weights of each edge in the causal graph using regression analysis to obtain the target causal graph structure.

7. The pediatric nursing intelligent rocking bed system based on feature data feedback as described in claim 1, characterized in that, The shaking table strategy generation module also includes the following units: The potential outcome simulation unit is used to simulate the trend of changes in infant state under different nursing interventions based on the target causal graph structure using the BSTS model. The control strategy generation unit is used to select the optimal nursing intervention plan from the potential results of the simulation with the goal of optimizing the infant's state, and generate the rocking crib control strategy.

8. A method for implementing the intelligent pediatric nursing rocking bed system based on feature data feedback as described in claim 1, characterized in that, The method includes the following steps: Collect multi-dimensional pediatric data, and preprocess the multi-dimensional pediatric data to obtain initial multi-dimensional pediatric data; Based on the CLP multimodal base model, the Transformer encoder is used to extract long-term dependency features from the multidimensional pediatric data, spatiotemporal features are extracted by 3D-CNN, and multi-scale acoustic features are generated according to WaveNet to obtain multimodal feature data. By aligning the feature representations of different modalities in the multimodal feature data through multimodal cross-attention, a unified global state vector is generated; Using the PC algorithm, a causal graph of infant state and nursing actions is constructed based on the global state vector. Causal relationships between variables are inferred from historical data. The potential outcomes under different nursing interventions are simulated through BSTS Bayesian time series simulation to generate a rocking crib control strategy. The shaking table's motion amplitude, temperature control parameters, and sound soothing parameters are dynamically adjusted according to the shaking table control strategy.

9. A method for implementing the intelligent pediatric nursing rocking bed system based on feature data feedback as described in claim 1, characterized in that, The method includes the following steps: Conditional independence was tested using the chi-square test, edges that did not satisfy conditional independence were removed, and a causal graph structure between infant state and care actions was constructed. Based on the constructed causal graph, the strength of the causal relationship between variables is inferred from historical data. Regression analysis is used to estimate the weights of each edge in the causal graph, thus obtaining the target causal graph structure.

10. A method for implementing the intelligent pediatric nursing rocking bed system based on feature data feedback as described in claim 1, characterized in that, The method includes the following steps: Based on the target cause-effect diagram structure, the BSTS model was used to simulate the trend of infant state changes under different nursing interventions. With the goal of optimizing the infant's condition, the optimal nursing intervention plan is selected from the potential results of the simulation using the particle swarm optimization algorithm, and a rocking bed control strategy is generated.

Images