Intelligent medical diagnostic bed control system and method based on Internet of Things

Abstract

The present invention relates to the technical field of Internet of Things, and discloses an intelligent medical diagnostic bed control system and method based on the Internet of Things. The method comprises: performing multi-dimensional physiological data collection on a target patient based on a sensor group of the intelligent medical diagnostic bed to obtain a first physiological signal data set; performing third-order extended state observation analysis to obtain vital sign state estimation data; performing nonlinear tracking differential processing to obtain smoothed physiological signal data; performing intelligent diagnostic point planning modeling to obtain an optimized diagnostic node sequence; performing adaptive nonlinear control law calculation to obtain a first regulating control instruction; executing the first regulating control instruction and collecting a second physiological signal data set and performing deep learning analysis and dynamic adjustment of control instructions to obtain a second regulating control instruction. The present invention realizes dynamic adjustment of control instructions and can perform refined control according to real-time feedback from patients.

Description

Technical Field

[0001] The present invention relates to the technical field of Internet of Things, and in particular to an intelligent medical diagnostic bed control system and method based on the Internet of Things. Background Art

[0002] Current smart medical diagnostic beds often have problems such as single function, insufficient accuracy, and poor real-time performance, making it difficult to comprehensively and accurately collect and analyze patients' multi-dimensional physiological data.

[0003] In addition, existing medical diagnostic beds generally lack intelligent control and adaptive adjustment capabilities, and cannot be adjusted in time according to the patient's real-time status, which can easily cause patient discomfort or affect the diagnosis and treatment effect. Another prominent problem is that most medical devices have limited data processing capabilities and are unable to cope with complex nonlinear physiological signals, which affects the reliability and accuracy of diagnostic results. Summary of the invention

[0004] The present invention provides an intelligent medical diagnostic bed control system and method based on the Internet of Things. The present invention realizes dynamic adjustment of control instructions and can perform refined control according to real-time feedback from patients.

[0005] In a first aspect, the present invention provides a method for controlling an intelligent medical diagnostic bed based on the Internet of Things, the method comprising:

[0006] The sensor group based on the intelligent medical diagnostic bed collects multi-dimensional physiological data of the target patient to obtain a first physiological signal data set;

[0007] Performing a third-order extended state observation analysis on the first physiological signal data set to obtain vital sign state estimation data;

[0008] Performing nonlinear tracking differentiation processing on the vital sign state estimation data to obtain smoothed physiological signal data;

[0009] Performing intelligent diagnostic point planning modeling on the smoothed physiological signal data to obtain an optimized diagnostic node sequence;

[0010] Performing adaptive nonlinear control law calculation based on the optimized diagnostic node sequence to obtain a first adjustment control instruction for the intelligent medical diagnostic bed;

[0011] The first adjustment control instruction is executed and a second physiological signal data set of the target patient is collected, and deep learning analysis and dynamic adjustment of the control instruction are performed on the second physiological signal data set to obtain a second adjustment control instruction.

[0012] In a second aspect, the present invention provides an intelligent medical diagnostic bed control system based on the Internet of Things, and the intelligent medical diagnostic bed control system based on the Internet of Things includes:

[0013] An acquisition module, used for acquiring multi-dimensional physiological data of a target patient based on a sensor group of an intelligent medical diagnostic bed, to obtain a first physiological signal data set;

[0014] An analysis module, configured to perform a third-order extended state observation analysis on the first physiological signal data set to obtain vital sign state estimation data;

[0015] A processing module, used for performing nonlinear tracking differential processing on the vital sign state estimation data to obtain smoothed physiological signal data;

[0016] A modeling module, used for performing intelligent diagnostic point planning modeling on the smoothed physiological signal data to obtain an optimized diagnostic node sequence;

[0017] A calculation module, used for performing adaptive nonlinear control law calculation based on the optimized diagnosis node sequence to obtain a first adjustment control instruction of the intelligent medical diagnostic bed;

[0018] An adjustment module is used to execute the first adjustment control instruction and collect a second physiological signal data set of the target patient, and perform deep learning analysis and dynamic adjustment of the control instruction on the second physiological signal data set to obtain a second adjustment control instruction.

[0019] In the technical solution provided by the present invention, by integrating multiple sensors, the comprehensive collection of multi-dimensional physiological data such as patient body pressure distribution, bioelectric signals, surface temperature, micro-motion behavior and respiratory heart sounds is realized, providing a rich data basis for subsequent analysis. The third-order extended state observation and analysis technology is adopted, which can more accurately estimate the patient's vital sign state, including high-order derivative information that is difficult to measure directly, and improve the accuracy and reliability of state estimation. Through nonlinear tracking differential processing, noise and interference in physiological signals are effectively removed, and the key features of signals are retained. By using multi-objective optimization and timing constraint inspection, intelligent planning of diagnostic nodes is realized, while ensuring diagnostic efficiency and reducing risks, based on sliding mode control and adaptive law design, precise control of diagnostic beds is realized, which can quickly respond to changes in patient status and improve the robustness and adaptability of control. Combined with multi-channel convolutional neural networks, long short-term memory networks and attention mechanisms, physiological signals are deeply analyzed, complex spatiotemporal features are extracted, and the accuracy and interpretability of diagnosis are improved. Through Bayesian neural networks and fuzzy reasoning, dynamic adjustment of control instructions is realized, and refined control can be performed according to real-time feedback from patients, improving the comfort and safety of the diagnosis and treatment process. Comprehensively utilize data from multiple sensors to achieve information complementarity and coordination. Through parallel processing and efficient algorithm design, ensure the real-time response capability of the system to meet the timeliness requirements of medical diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0021] Figure 1 Schematic diagram of the steps of the control method of the intelligent medical diagnostic bed based on the Internet of Things in an embodiment of the present invention;

[0022] Figure 2 It is a structural schematic diagram of an intelligent medical diagnostic bed control system based on the Internet of Things in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] The embodiment of the present invention provides an intelligent medical diagnostic bed control system and method based on the Internet of Things. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0024] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 In one embodiment of the present invention, an embodiment of a smart medical diagnostic bed control method based on the Internet of Things includes:

[0025] Step S1, collecting multi-dimensional physiological data of a target patient based on a sensor group of an intelligent medical diagnostic bed to obtain a first physiological signal data set;

[0026] It is understandable that the execution subject of the present invention may be an intelligent medical diagnostic bed control system based on the Internet of Things, or may be a terminal or a server, which is not specifically limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0027] Specifically, the pressure sensor array of the intelligent medical diagnostic bed is read to obtain the body pressure distribution data of the target patient, and the posture feature of the target patient is extracted based on the body pressure distribution data to obtain the posture feature vector. The pressure sensor array is used to accurately capture the pressure distribution of various parts of the patient's body, and the posture features of the patient are analyzed through a specific algorithm. The bioelectric sensor of the intelligent medical diagnostic bed is used to collect signals to obtain the original bioelectric signal. The original bioelectric signal is subjected to wavelet transform denoising to obtain the filtered bioelectric signal. The signal quality and reliability are improved by eliminating the noise in the bioelectric signal. At the same time, the thermal imaging data collected by the infrared imaging device of the intelligent medical diagnostic bed is image segmented to obtain the body surface temperature distribution map of the target patient. Hot spot detection and temperature gradient analysis are performed on the body surface temperature distribution map to obtain abnormal body temperature area data. The patient's body surface temperature changes are monitored in real time through infrared imaging technology, and the abnormal body surface temperature area is accurately located through image processing technology. The image sequence collected by the high-definition camera of the intelligent medical diagnostic bed is motion estimated to obtain the micro-motion behavior data of the target patient, and the micro-motion behavior data is extracted in the time and frequency domain to obtain the activity state characteristics of the target patient. The high-definition camera captures the patient's subtle movements, and through motion estimation technology, these movements are quantitatively analyzed to understand the patient's activity status and provide auxiliary information for clinical diagnosis. At the same time, the audio signal collected by the sound sensor of the intelligent medical diagnostic bed is spectrally analyzed to obtain the characteristics of respiratory sounds and heart sounds. The sound sensor can collect the patient's breathing and heartbeat sounds, and extract the characteristic parameters of these sound signals through spectrum analysis technology. Multimodal data fusion is performed based on the posture feature vector, filtered bioelectric signals, abnormal body temperature area data, activity status characteristics, respiratory sounds and heart sound characteristics to obtain the first physiological signal data set. By fusing the data collected by multiple sensors and using data fusion technology, the characteristics of each physiological signal are comprehensively analyzed to form a complete physiological signal data set.

[0028] Step S2, performing a third-order extended state observation analysis on the first physiological signal data set to obtain vital sign state estimation data;

[0029] Specifically, the first physiological signal data set is subjected to nonlinear state space modeling to obtain a state equation and an observation equation, and a structure matrix of a third-order extended state observer is constructed based on the state equation and the observation equation. Nonlinear state space modeling can capture complex physiological dynamic changes by expressing the physiological signal data set as a functional relationship of state variables. The state equation describes the evolution of the system state, and the observation equation describes the relationship between the observed data and the system state. Through these equations, the structure matrix of the third-order extended state observer can be constructed. The structure matrix of the third-order extended state observer is initialized with parameters to obtain an initial state estimate and an observer gain matrix. The first physiological signal data set is linearly combined through the initial state estimate and the observer gain matrix to obtain an error feedback term of the state estimate. The error feedback term is the difference between the actual observed value and the estimated value. By analyzing the error feedback term, the output of the observer is continuously corrected to make the estimate more accurate. The error feedback term and the state equation are numerically integrated to obtain the target state estimate at the current moment, and the observation error is calculated based on the target state estimate to obtain an observation residual sequence. Numerical integration is to discretize the state equation, so as to calculate the change of the system state at each time point and obtain the target state estimate. The observation error is the difference between the target state estimate and the actual observation value. By calculating the observation error, the observation residual sequence is formed for subsequent analysis and adjustment. The observation residual sequence is adaptively weighted to obtain the adaptive adjustment factor, and the observer gain matrix is ​​dynamically updated according to the adaptive adjustment factor to obtain the optimized observer parameters. Adaptive weight calculation adaptively adjusts the observer gain matrix according to the change of the observation residual to improve the observer's ability to respond to dynamic changes. The observer parameters are adjusted in real time through the adaptive adjustment factor, so that the observer can maintain a high estimation accuracy under different states. Error covariance analysis is performed on the optimized observer parameters and the target state estimate to obtain the confidence interval of the state estimate, and data fusion and reasoning are performed based on the target state estimate and the confidence interval to obtain the vital sign state estimation data. Error covariance analysis calculates the confidence interval of the estimate through statistical analysis of the estimation error, provides confidence for the estimation result, and the reliability of the estimate can be judged by the confidence interval. Combining the target state estimation value and confidence interval, multi-source data is comprehensively analyzed through data fusion technology to obtain vital sign state estimation data.

[0030] Step S3, performing nonlinear tracking differentiation processing on the vital sign state estimation data to obtain smoothed physiological signal data;

[0031] Specifically, the time series of the vital sign state estimation data is decomposed to obtain the trend term, the periodic term and the random term, and the state equation of the nonlinear tracking differentiator is constructed according to the trend term to obtain the initial model of the differentiator. The trend term reflects the long-term change trend of the physiological signal, and these changes can be accurately captured by constructing the state equation of the nonlinear tracking differentiator. The initial model is discretized to obtain the difference equation of the discrete-time nonlinear tracking differentiator, and the fast convergence factor is designed according to the difference equation to obtain the fast tracking differentiation strategy. The discretization process is to realize the numerical calculation of the continuous model on the computer. By designing the fast convergence factor, it is ensured that the tracking differentiator can respond quickly to the change of the signal and improve the processing efficiency. The fast tracking differentiation strategy is subjected to bandwidth adaptive adjustment to obtain the dynamic bandwidth adjustment strategy, and the periodic term is filtered according to the dynamic bandwidth adjustment strategy to obtain a smooth periodic signal. The bandwidth adaptive adjustment dynamically adjusts the bandwidth of the differentiator so that it can adapt to signals with different frequency components, ensures that the filtered periodic term is smoother, and removes the high-frequency noise in the periodic signal. At the same time, the random term is subjected to wavelet threshold denoising to obtain a denoised random signal. Wavelet threshold denoising utilizes the good characteristics of wavelet transform in the time-frequency domain to effectively remove noise from random signals and improve the signal-to-noise ratio of signals. Signal reconstruction is performed based on the smoothed periodic signal and the denoised random signal to obtain a preliminary smoothed signal. The processed periodic signal and random signal are recombined to restore the main features of the original signal while eliminating noise and irregular fluctuations. The preliminary smoothed signal is edge-preserving filtered to obtain a smoothed signal with sharp change features. Edge-preserving filtering prevents the loss of important physiological information by retaining the sharp change features in the signal, so that the smoothed signal can more accurately reflect the patient's physiological state. Weighted fusion is performed based on the smoothed signal with sharp change features and the vital sign status estimation data to obtain smoothed physiological signal data. Weighted fusion comprehensively considers the characteristics of the two signals and assigns different weights to obtain more reliable and smooth physiological signal data.

[0032] The random item is decomposed by wavelet to obtain multi-scale wavelet coefficients, and the noise variance of each scale is calculated based on the multi-scale wavelet coefficients to obtain the noise level estimation value. Through wavelet decomposition, the random item signal is decomposed into wavelet coefficients of different scales, so that the noise and signal can be distinguished at different scales, thereby estimating the noise variance of each scale. The noise level estimation value is nonlinearly mapped to obtain an adaptive threshold function, and the multi-scale wavelet coefficients are soft-thresholded according to the adaptive threshold function to obtain the denoised wavelet coefficients. Nonlinear mapping can better adapt to the characteristics of the signal by converting the noise level estimation value into an adaptive threshold function, thereby effectively removing noise. In the soft threshold processing process, the wavelet coefficients are reduced when approaching the threshold instead of being directly set to zero, thereby retaining more signal details. The denoised wavelet coefficients are reconstructed by wavelet to obtain the denoised random signal. Wavelet reconstruction recombines the processed wavelet coefficients, restores the useful part of the original signal, removes the influence of noise, and obtains a smoother and more realistic random signal. A signal fusion model is constructed based on the smooth periodic signal and the denoised random signal, and the initial fusion weight is calculated. The signal fusion model combines the smooth periodic signal and the denoised random signal, and uses the initial fusion weight to perform a preliminary weighted combination of the two to combine the advantages of the two signals. The initial fusion weight is dynamically optimized to obtain an adaptive fusion weight coefficient, and the smooth periodic signal and the denoised random signal are weighted summed according to the adaptive fusion weight coefficient to obtain a fusion signal. The dynamic optimization process adjusts the initial fusion weight so that it can better adapt to the changes in the signal and improve the accuracy and reliability of the fusion signal. The fusion signal is subjected to the Hilbert-Huang transform to obtain the instantaneous frequency and amplitude of the signal, and the signal is reconstructed according to the instantaneous frequency and amplitude to obtain a preliminary smoothed signal. The Hilbert-Huang transform is an adaptive data analysis method that can capture the local characteristics of the signal by analyzing the instantaneous frequency and amplitude of the fusion signal, and improve the smoothness and accuracy of the signal. The process of reconstructing the signal restores the local characteristics of the fusion signal through the instantaneous frequency and amplitude information to obtain the final preliminary smoothed signal.

[0033] Step S4, performing intelligent diagnostic point planning modeling on the smoothed physiological signal data to obtain an optimized diagnostic node sequence;

[0034] Specifically, multi-scale feature extraction is performed on the smoothed physiological signal data to obtain a multi-dimensional feature vector. Through multi-scale feature extraction, the characteristics of the signal at different time scales are captured, including frequency components, amplitude changes and trend characteristics, forming a comprehensive multi-dimensional feature vector. According to the multi-dimensional feature vector, a dynamic time warping distance matrix is ​​constructed and a signal similarity metric is generated. The signal similarity metric is subjected to hierarchical clustering analysis to obtain an initial diagnostic point candidate set. The dynamic time warping distance matrix is ​​used to measure the similarity between different physiological signals. Through hierarchical clustering analysis, similar signals are clustered together to form an initial diagnostic point candidate set. According to the initial diagnostic point candidate set, a Markov decision process model is constructed and a state transition probability matrix is ​​generated. The state transition probability matrix is ​​solved by value iteration to obtain the optimal diagnostic point selection strategy. The Markov decision process model considers the state transition relationship between different diagnostic points and uses the state transition probability matrix for value iteration to find the optimal diagnostic point selection strategy and optimize the diagnostic process. The initial diagnostic point candidate set is screened according to the optimal diagnostic point selection strategy to obtain the key diagnostic point sequence, and the key diagnostic point sequence is tested for timing constraints to obtain a subset of diagnostic points that meets the timing logic. The screening of the key diagnostic point sequence is based on the optimal selection strategy, and the diagnostic points are optimized and selected. The timing constraint test ensures the timing logic relationship of the diagnostic points to form an effective subset of diagnostic points. According to the subset of diagnostic points that meets the timing logic, a multi-objective optimization model is constructed and a balance function of diagnostic efficiency and risk is generated. The multi-objective optimization model considers diagnostic efficiency and diagnostic risk, and weighs these two objectives through the balance function to ensure that the diagnostic efficiency is maximized while ensuring the diagnostic accuracy. The balance function is optimized and solved to obtain the Pareto optimal solution set, and the decision weights are allocated according to the Pareto optimal solution set to obtain the optimized diagnostic node sequence. By optimizing the balance function, a set of Pareto optimal solutions is obtained, which achieve the best balance between diagnostic efficiency and risk. Based on the optimal solution, the diagnostic node sequence is optimized through decision weight allocation to ensure that the final diagnostic node sequence is both efficient and reliable.

[0035] Step S5, performing adaptive nonlinear control law calculation based on the optimized diagnosis node sequence to obtain a first adjustment control instruction of the intelligent medical diagnostic bed;

[0036] Specifically, a time series analysis is performed on the optimized diagnostic node sequence to obtain the time distribution characteristics of the diagnostic nodes, and the state space equation of the diagnostic bed is constructed based on the time distribution characteristics of the diagnostic nodes. Time series analysis can capture the time variation law of the diagnostic nodes, and accurately describe the dynamic behavior of the intelligent medical diagnostic bed through the state space equation of the diagnostic bed. According to the state space equation of the diagnostic bed, the Lyapunov function is designed to obtain the system stability criterion, and the sliding surface equation is constructed according to the system stability criterion to obtain the sliding mode control law. The Lyapunov function can ensure the stability of the system. By constructing the sliding surface equation, the system state is constrained on the sliding surface to achieve stable control of the diagnostic bed. The sliding mode control law is designed with an exponential approach law to obtain a fast convergence control term, and the fast convergence control term is compensated to obtain a robust control term. The exponential approach law design improves the response speed of the system by accelerating the convergence speed of the system state to the sliding surface, and the compensated robust control term enhances the system's resistance to external interference and ensures the robustness of the system. According to the robust control term, an adaptive law is designed to obtain the parameter adaptive update rule, and the control gain is adjusted online according to the parameter adaptive update rule to obtain the adaptive control gain. The adaptive law updates the system parameters in real time, so that the control gain can be adjusted according to the change of the system state, thereby improving the control accuracy and adaptability of the system. The control input is calculated based on the adaptive control gain to obtain the initial control instruction, and the initial control instruction is processed by saturation constraint to obtain the first adjustment control instruction of the intelligent medical diagnostic bed. The calculation of the control input is based on the adaptive control gain, and the precise control of the intelligent medical diagnostic bed is achieved through the control input. The saturation constraint processing is to prevent the control instruction from exceeding the physical limitations of the system and ensure that the control instruction operates within a safe range.

[0037] Step S6, executing the first adjustment control instruction and collecting a second physiological signal data set of the target patient, and performing deep learning analysis and dynamic adjustment of the control instruction on the second physiological signal data set to obtain a second adjustment control instruction.

[0038] Specifically, the second physiological signal data set is preprocessed to obtain a standardized data set, and the noise and inconsistency in the data are eliminated. The standardized data set is input into the multi-channel convolutional neural network layer for feature extraction to obtain a multidimensional feature map, and the multidimensional feature map is input into the long short-term memory network and the gated recurrent unit network layer for time series analysis to obtain a dynamic feature sequence. The spatial features of the physiological signal are effectively extracted through the multi-channel convolutional neural network, while the long short-term memory network and the gated recurrent unit network can capture the time dependence of the physiological signal to obtain a dynamic feature sequence, which reflects the temporal change law of the signal. The dynamic feature sequence is input into the attention mechanism layer to calculate the importance weight of the feature to obtain a weighted feature vector, and the weighted feature vector is subjected to dimensionality reduction processing to obtain a compressed feature representation. The attention mechanism can highlight key features by calculating the importance weight of the feature, improve the interpretability and prediction accuracy of the model. The dimensionality reduction processing is to reduce the redundancy of the feature, improve the computational efficiency, and obtain a more concise compressed feature representation. The compressed feature representation is input into the multi-task learning layer to simultaneously predict multiple physiological indicators and body position parameters to obtain a preliminary prediction result. By learning multiple tasks simultaneously, the multi-task learning layer can make full use of the shared information in the data, improve the prediction performance of each task, and obtain preliminary prediction results of multiple physiological indicators and body position parameters. The Bayesian neural network layer calculates the confidence interval of the preliminary prediction results to obtain the probability distribution output, and the fuzzy reasoning layer optimizes the decision-making of the probability distribution output to obtain adjustment suggestions. The Bayesian neural network layer provides uncertainty information of the prediction by calculating the confidence interval of the prediction results, making the prediction results more reliable. The fuzzy reasoning layer optimizes the adjustment suggestions by processing the probability distribution output to improve the rationality and effectiveness of the decision. According to the adjustment suggestion, the first adjustment control instruction is corrected to obtain the corrected control instruction, and the corrected control instruction is constrained and smoothed to obtain the second adjustment control instruction. The corrected control instruction is an optimized adjustment of the original control instruction according to the adjustment suggestion to ensure that the control strategy is more in line with actual needs. Constraints and smoothing are to prevent the control instructions from exceeding the physical limitations of the system and reduce the fluctuation of the instructions, making the control instructions more stable and safe.

[0039] The standardized data set is channel-separated to obtain multiple single-channel data sequences. Through channel separation, each single-channel data sequence can be processed independently to capture the feature information of each channel. Multiple single-channel data sequences are processed by parallel convolution. The parallel convolution includes three convolution modules, each of which contains three convolution layers. Each convolution layer uses ReLU activation function and batch normalization. Through parallel convolution, features are extracted in parallel on different channels. ReLU activation function introduces nonlinear transformation, while batch normalization stabilizes the training process and accelerates convergence. The output of parallel convolution is fused by channel attention mechanism to obtain multi-dimensional feature map. Channel attention mechanism can effectively highlight key features and improve the expressiveness of feature map by giving higher weights to important channels. Spatial pyramid pooling is performed on multi-dimensional feature map to obtain multi-scale feature representation. Spatial pyramid pooling includes pooling operations of four different scales. Spatial pyramid pooling can capture feature information of different scales by performing pooling operations at different scales to form a richer multi-scale feature representation. The multi-scale feature representation is input into the fully connected layer for feature dimension reduction to obtain a compressed feature vector, and the compressed feature vector is reconstructed into a time series to obtain a time series feature sequence. The fully connected layer reduces the feature dimension to reduce the feature redundancy and improve the computational efficiency. The time series reconstruction reconstructs the compressed feature vector into a time series form for subsequent time series analysis. The time series feature sequence is input into the bidirectional long short-term memory network. The bidirectional long short-term memory network contains 3 layers of bidirectional LSTM units, and the hidden state dimension of each layer of LSTM units is 256. The bidirectional long short-term memory network can capture long-term dependencies in sequence data and improve the representation ability of feature sequences through forward and backward bidirectional LSTM units. The self-attention mechanism is applied to the output sequence of the bidirectional long short-term memory network to obtain context-enhanced features. The self-attention mechanism adopts a multi-head attention structure with the number of heads set to 8. The self-attention mechanism can highlight important features and enhance the context information of features by calculating the interdependence of features at different positions. The multi-head attention structure enriches feature representation by calculating different attention heads in parallel. The context-enhanced features are input into the gated recurrent unit network, which contains two layers of GRU units, and the hidden state dimension of each layer of GRU units is 128. The gated recurrent unit network can efficiently capture important information in the sequence through the gating mechanism of the GRU unit, reduce information loss, and improve the accuracy of feature extraction. The output sequence of the gated recurrent unit network is weighted by time step attention to obtain a dynamic feature sequence. Time step attention weighting can highlight the features of key time steps and improve the expressiveness of dynamic feature sequences by weighting the importance of each time step.

[0040] In the embodiment of the present invention, by integrating multiple sensors, the comprehensive collection of multi-dimensional physiological data such as patient body pressure distribution, bioelectric signals, surface temperature, micro-motion behavior, respiratory heart sounds, etc. is realized, providing a rich data basis for subsequent analysis. The third-order extended state observation and analysis technology is adopted, which can more accurately estimate the patient's vital sign state, including high-order derivative information that is difficult to measure directly, and improve the accuracy and reliability of state estimation. Through nonlinear tracking differential processing, noise and interference in physiological signals are effectively removed, and the key features of signals are retained. By using multi-objective optimization and timing constraint inspection, intelligent planning of diagnostic nodes is realized, while ensuring diagnostic efficiency and reducing risks. Based on sliding mode control and adaptive law design, precise control of diagnostic beds is realized, which can quickly respond to changes in patient status and improve the robustness and adaptability of control. Combined with multi-channel convolutional neural networks, long short-term memory networks and attention mechanisms, physiological signals are deeply analyzed, complex spatiotemporal features are extracted, and the accuracy and interpretability of diagnosis are improved. Through Bayesian neural networks and fuzzy reasoning, dynamic adjustment of control instructions is realized, and refined control can be performed according to real-time feedback from patients, improving the comfort and safety of the diagnosis and treatment process. Comprehensively utilize data from multiple sensors to achieve information complementarity and coordination. Through parallel processing and efficient algorithm design, ensure the real-time response capability of the system to meet the timeliness requirements of medical diagnosis.

[0041] In a specific embodiment, the process of executing step S1 may specifically include the following steps:

[0042] Reading data from the pressure sensor array of the intelligent medical diagnostic bed to obtain body pressure distribution data of the target patient, and extracting posture features of the target patient based on the body pressure distribution data to obtain a posture feature vector;

[0043] The bioelectric sensor of the intelligent medical diagnosis bed is used to collect signals to obtain the original bioelectric signal, and the original bioelectric signal is subjected to wavelet transform denoising to obtain the filtered bioelectric signal;

[0044] Perform image segmentation on the thermal imaging data collected by the infrared imaging device of the intelligent medical diagnostic bed to obtain the body surface temperature distribution map of the target patient, and perform hot spot detection and temperature gradient analysis on the body surface temperature distribution map to obtain abnormal body temperature area data;

[0045] Perform motion estimation on the image sequence captured by the high-definition camera of the intelligent medical diagnostic bed to obtain the micro-motion behavior data of the target patient, and perform time-frequency domain feature extraction on the micro-motion behavior data to obtain the activity state characteristics of the target patient;

[0046] Perform spectrum analysis on the audio signal collected by the sound sensor of the intelligent medical diagnostic bed to obtain the characteristics of respiratory sounds and heart sounds;

[0047] Multimodal data fusion is performed based on posture feature vectors, filtered bioelectric signals, abnormal body temperature area data, activity state features, respiratory sounds and heart sound features to obtain a first physiological signal data set.

[0048] Specifically, the pressure sensor array of the intelligent medical diagnostic bed is read to obtain the body pressure distribution data of the target patient, and the patient's posture is extracted through these data to obtain the posture feature vector. The pressure sensor array can capture the pressure distribution of the patient's body on the bed, and these data can be used to analyze the patient's posture. For example, if the pressure data is a matrix , where each element Indicates at location The pressure value at the pressure matrix Perform principal component analysis to extract posture feature vectors The specific formula is as follows:

[0049] ;

[0050] in, Indicates at location The pressure value of the posture feature vector obtained by principal component analysis It can summarize the patient's posture information. It collects signals from the bioelectric sensors of the intelligent medical diagnostic bed and obtains the original bioelectric signals. These signals may contain noise and need to be denoised by wavelet transform to obtain filtered bioelectric signals. Wavelet transform can decompose the signal and remove the noise component, making the signal clearer. Perform wavelet decomposition to obtain detail coefficients and approximate coefficients, and apply the threshold denoising method to remove noise. The formula is as follows:

[0051] ;

[0052] in, is the threshold function, is the wavelet basis function, is the wavelet coefficient. At the same time, the thermal imaging data collected by the infrared imaging device of the intelligent medical diagnostic bed is segmented to obtain the body surface temperature distribution map of the target patient. . Through image segmentation algorithms such as K-means clustering, Divide into different temperature areas. Identify temperature anomalies through hot spot detection and temperature gradient analysis . Temperature Matrix It is expressed as follows:

[0053] ;

[0054] The abnormal area is obtained through K-means clustering and temperature gradient analysis. The image sequence collected by the high-definition camera of the intelligent medical diagnostic bed is motion estimated to obtain the micro-motion behavior data of the target patient. The image sequence is analyzed by the optical flow method or deep learning method to obtain the motion vector field at each time point. The time-frequency domain features of the motion data are extracted, such as applying short-time Fourier transform to obtain the activity state characteristics of the patient. The audio signal collected by the sound sensor of the intelligent medical diagnostic bed is spectrally analyzed to obtain the respiratory sound and heart sound features. Then, the respiratory sound features and heart sound features are extracted by frequency band division. Multimodal data fusion is performed according to the posture feature vector, filtered bioelectric signal, abnormal body temperature area data, activity state characteristics, respiratory sound and heart sound characteristics to obtain the first physiological signal data set. Multimodal data fusion can be achieved by weighted average or deep learning methods.

[0055] In a specific embodiment, the process of executing step S2 may specifically include the following steps:

[0056] Performing nonlinear state space modeling on the first physiological signal data set to obtain a state equation and an observation equation, and constructing a structure matrix of a third-order extended state observer according to the state equation and the observation equation;

[0057] Initializing the parameters of the structure matrix of the third-order extended state observer to obtain an initial state estimation value and an observer gain matrix, and linearly combining the first physiological signal data set according to the initial state estimation value and the observer gain matrix to obtain an error feedback term of the state estimation;

[0058] Numerically integrate the error feedback term and the state equation to obtain the target state estimate at the current moment, and calculate the observation error based on the target state estimate to obtain the observation residual sequence;

[0059] Adaptive weight calculation is performed on the observation residual sequence to obtain an adaptive adjustment factor, and the observer gain matrix is ​​dynamically updated according to the adaptive adjustment factor to obtain the optimized observer parameters;

[0060] Error covariance analysis is performed on the optimized observer parameters and target state estimates to obtain the confidence interval of the state estimate. Data fusion and inference are performed based on the target state estimate and the confidence interval to obtain the vital sign state estimation data.

[0061] Specifically, a nonlinear state space model is performed on the first physiological signal data set to obtain a state equation describing the change of the system state over time and an observation equation describing the relationship between the observed data and the system state. Assume that the state equation can be expressed as:

[0062] ;

[0063] in, Indicates that the system is at time The state vector at time, is the control input vector, is the state transition function, is the process noise. The observation equation is expressed as:

[0064] ;

[0065] in, It's in time The observation vector at time, is the observation function, is the observation noise. The structure matrix of the third-order extended state observer is constructed based on the state equation and the observation equation. The extended state observer is used to estimate the state of the system and its unmodeled dynamics. Constructing the structure matrix of the extended state observer includes expanding the state variables and their high-order derivatives. Assume that the extended state vector is , then the structure matrix of the third-order extended state observer is , input matrix And the observation matrix C can be expressed as:

[0066] ;

[0067] in, is a system parameter, is the input gain. Initialize the parameters of the structure matrix of the third-order extended state observer and obtain the initial state estimate and the observer gain matrix The error feedback term of state estimation is obtained by linearly combining the first physiological signal data set according to the initial state estimation value and the observer gain matrix. Assume that the initial state estimation value is , the observer gain matrix is:

[0068] ;

[0069] in, is the observer gain. The state estimation error feedback term It is expressed as:

[0070] ;

[0071] Numerically integrate the error feedback term and the state equation to obtain the target state estimate at the current moment , and calculate the observation error based on the target state estimate to obtain the observation residual sequence. Numerical integration can be achieved through discretization methods, such as the Euler method:

[0072] ;

[0073] in, is the time step. The observation residual sequence is adaptively weighted to obtain the adaptive adjustment factor, and the observer gain matrix is ​​dynamically updated according to the adaptive adjustment factor to obtain the optimized observer parameters. The adaptive weight calculation can be implemented by the following formula:

[0074] ;

[0075] in, It is The weight of the moment, the adaptive adjustment factor Controls the rate at which weights are adjusted. Updates the observer gain matrix according to the adaptive adjustment factor:

[0076] ;

[0077] The error covariance analysis is performed on the optimized observer parameters and the target state estimation value to obtain the confidence interval of the state estimation, and data fusion and reasoning are performed based on the target state estimation value and the confidence interval to obtain the vital sign state estimation data. The error covariance analysis can be implemented by the following formula:

[0078] ;

[0079] in, It is The error covariance matrix at time , is the process noise covariance matrix, is the identity matrix.

[0080] In a specific embodiment, the process of executing step S3 may specifically include the following steps:

[0081] Decompose the vital sign state estimation data into time series to obtain trend terms, periodic terms and random terms, and construct the state equation of the nonlinear tracking differentiator based on the trend terms to obtain the initial model of the differentiator;

[0082] Discretize the initial model to obtain the difference equation of the discrete-time nonlinear tracking differentiator, and design a fast convergence factor based on the difference equation to obtain a fast tracking differentiation strategy;

[0083] Adaptively adjust the bandwidth of the fast tracking differential strategy to obtain a dynamic bandwidth adjustment strategy, and filter the periodic term according to the dynamic bandwidth adjustment strategy to obtain a smooth periodic signal;

[0084] Perform wavelet threshold denoising on the random items to obtain a denoised random signal, and perform signal reconstruction based on the smoothed periodic signal and the denoised random signal to obtain a preliminary smoothed signal;

[0085] The preliminary smoothed signal is subjected to edge-preserving filtering to obtain a smoothed signal retaining the sudden change feature, and the smoothed signal retaining the sudden change feature and the vital sign state estimation data are weightedly fused to obtain smoothed physiological signal data.

[0086] Specifically, the vital sign state estimation data is decomposed into trend terms, periodic terms and random terms. Assume that the vital sign state estimation data is , through time series decomposition, can be expressed as:

[0087] ;

[0088] in, Represents the trend item, represents the periodic term, Represents random terms. Through decomposition methods, long-term trends, periodic changes, and random fluctuations in the data are captured separately. Construct the state equation of the nonlinear tracking differentiator and obtain the initial model of the differentiator. Assume that the state equation of the nonlinear tracking differentiator can be expressed as:

[0089] ;

[0090] in, is the state vector, is a nonlinear function, is the control input, is the process noise. The initial model is discretized to obtain the difference equation of the discrete-time nonlinear tracking differentiator. The discretization can be achieved by the Euler method to obtain the difference equation:

[0091] ;

[0092] in, is the time step. In order to speed up the convergence, a fast convergence factor is designed. , we get the fast tracking differential strategy. The fast tracking differential strategy can be expressed as:

[0093] ;

[0094] in, Control the convergence speed. The fast tracking differentiation strategy is adaptively adjusted to obtain a dynamic bandwidth adjustment strategy. The bandwidth adaptive adjustment enables the tracking differentiator to adapt to signals with different frequency components by adjusting the bandwidth parameters in real time. The dynamic bandwidth adjustment strategy can be expressed as:

[0095] ;

[0096] in, It is Adaptive bandwidth parameters at the moment. Perform filtering to obtain a smooth periodic signal For random items Perform wavelet threshold denoising to obtain the denoised random signal Wavelet threshold denoising removes noise by decomposing the signal into different frequency components and applying threshold processing to these components. Let the wavelet coefficients be , the threshold function is , then the random signal after denoising is It can be expressed as:

[0097] ;

[0098] in, is the wavelet basis function. According to the smooth periodic signal And the random signal after noise reduction Reconstruct the signal to obtain a preliminary smoothed signal Signal reconstruction combines the smoothed periodic signal and the denoised random signal to form a composite signal. The formula is as follows:

[0099] ;

[0100] In order to improve the quality of the signal, the initial smoothing signal Edge-preserving filtering is performed to obtain a smooth signal that retains the sudden change features. Edge-preserving filtering prevents feature loss caused by over-smoothing during the smoothing process by retaining the sudden change features in the signal. For example, bilateral filtering is a commonly used edge-preserving filtering method, and the formula is as follows:

[0101] ;

[0102] in, is the edge-preserving filtered signal, is the spatial weight, is the intensity weight. According to the smooth signal that retains the sharp change characteristics and vital sign status estimates Perform weighted fusion to obtain the final smoothed physiological signal data . Weighted fusion can be achieved by the following formula:

[0103] ;

[0104] in, is the weighting coefficient.

[0105] In a specific embodiment, the execution step performs wavelet threshold denoising on the random item to obtain a denoised random signal, and performs signal reconstruction based on the smoothed periodic signal and the denoised random signal to obtain a preliminary smoothed signal. The process may specifically include the following steps:

[0106] Perform wavelet decomposition on random terms to obtain multi-scale wavelet coefficients, and calculate the noise variance of each scale based on the multi-scale wavelet coefficients to obtain the noise level estimation value;

[0107] Nonlinear mapping is performed on the noise level estimation value to obtain an adaptive threshold function, and soft threshold processing is performed on the multi-scale wavelet coefficients according to the adaptive threshold function to obtain the denoised wavelet coefficients;

[0108] Perform wavelet reconstruction on the denoised wavelet coefficients to obtain the denoised random signal;

[0109] A signal fusion model is constructed based on the smoothed periodic signal and the denoised random signal, and the initial fusion weight is calculated;

[0110] The initial fusion weight is dynamically optimized to obtain an adaptive fusion weight coefficient, and the smoothed periodic signal and the denoised random signal are weighted summed according to the adaptive fusion weight coefficient to obtain a fusion signal;

[0111] The fused signal is subjected to Hilbert-Huang transform to obtain the instantaneous frequency and amplitude of the signal, and the signal is reconstructed according to the instantaneous frequency and amplitude to obtain a preliminary smoothed signal.

[0112] Specifically, the random term is decomposed by wavelet to obtain multi-scale wavelet coefficients. Wavelet decomposition is a multi-resolution analysis method that decomposes the signal into details and approximate components at different scales. Assume that the random term signal is , can be expressed by wavelet transform as:

[0113] ;

[0114] in, are the wavelet coefficients, is the wavelet basis function, and Represent the scale and position index respectively. According to the multi-scale wavelet coefficients Calculate the noise variance at each scale to get the estimated value of the noise level. The calculation of the noise variance can be achieved by statistically analyzing the wavelet coefficients at each scale. For example, for scale The wavelet coefficients on It can be expressed as:

[0115] ;

[0116] in, It is a scale The number of wavelet coefficients on the scale. The noise level estimate is nonlinearly mapped to obtain an adaptive threshold function. The adaptive threshold function is used to determine the denoising threshold at each scale. One of the commonly used threshold functions is the soft threshold function, which is defined as follows:

[0117] ;

[0118] in, It is a scale The threshold on , is a sign function. According to the adaptive threshold function, the multi-scale wavelet coefficients are subjected to soft threshold processing to obtain the denoised wavelet coefficients. After soft threshold processing, the noise components in the wavelet coefficients are effectively removed and the main components of the signal are retained.

[0119] ;

[0120] The denoised wavelet coefficients are reconstructed by wavelet to obtain the denoised random signal. Wavelet reconstruction restores the signal by recombining the processed wavelet coefficients. The formula is as follows:

[0121] ;

[0122] According to the smooth periodic signal And the random signal after noise reduction Construct a signal fusion model and calculate the initial fusion weight. The signal fusion model forms a comprehensive signal by weighted combination of two signals. The initial fusion weight can be determined by simple weighted average:

[0123] ;

[0124] in, is the initial fusion weight. The initial fusion weight is dynamically optimized to obtain the adaptive fusion weight coefficient. Dynamic optimization adaptively adjusts the fusion weight so that the fusion signal can better reflect the actual situation. The adaptive fusion weight coefficient can be determined by minimizing the error function, such as the least squares method:

[0125] ;

[0126] in, is the original signal. According to the adaptive fusion weight coefficient The smoothed periodic signal and the denoised random signal are weighted and summed to obtain the final fused signal:

[0127] ;

[0128] The fused signal is subjected to Hilbert-Huang transform to obtain the instantaneous frequency and amplitude of the signal. The Hilbert-Huang transform calculates the instantaneous frequency and amplitude by decomposing the signal into a series of intrinsic mode functions (IMFs) and performing Hilbert transform on each IMF. Assume The first IMF is IM , then its instantaneous frequency and amplitude can be expressed as:

[0129] ;

[0130] in, is the instantaneous amplitude, is the instantaneous phase. By Hilbert transform, the instantaneous frequency It can be expressed as:

[0131] ;

[0132] Reconstruct the signal according to the instantaneous frequency and amplitude to obtain a preliminary smoothed signal The signal reconstruction is achieved by weighted combination of the instantaneous frequencies and amplitudes of all IMFs:

[0133] ;

[0134] In a specific embodiment, the process of executing step S4 may specifically include the following steps:

[0135] Perform multi-scale feature extraction on the smoothed physiological signal data to obtain a multi-dimensional feature vector;

[0136] A dynamic time warping distance matrix is ​​constructed based on the multi-dimensional feature vectors and a signal similarity metric is generated. A hierarchical clustering analysis is performed on the signal similarity metric to obtain an initial diagnostic point candidate set.

[0137] According to the initial diagnostic point candidate set, a Markov decision process model is constructed and a state transition probability matrix is ​​generated. The state transition probability matrix is ​​solved by value iteration to obtain the optimal diagnostic point selection strategy.

[0138] The initial diagnostic point candidate set is screened according to the optimal diagnostic point selection strategy to obtain a key diagnostic point sequence, and the key diagnostic point sequence is subjected to a timing constraint check to obtain a diagnostic point subset that satisfies the timing logic;

[0139] A multi-objective optimization model is constructed based on the subset of diagnostic points that satisfy temporal logic and a balance function between diagnostic efficiency and risk is generated;

[0140] The balance function is optimized and solved to obtain the Pareto optimal solution set, and the decision weights are allocated according to the Pareto optimal solution set to obtain the optimized diagnosis node sequence.

[0141] Specifically, multi-scale feature extraction is performed on the smoothed physiological signal data to obtain a multi-dimensional feature vector. Multi-scale feature extraction captures the details and overall trend of the signal by analyzing the characteristics of the signal at different time scales. Suppose the smoothed physiological signal data is , through multi-scale analysis, we can get the feature vectors on multiple scales These feature vectors are combined into a multi-dimensional feature vector , the formula is as follows:

[0142] ;

[0143] in, It is in The feature vectors extracted at each scale are is the number of scales. According to the multi-dimensional feature vector Construct a dynamic time warping distance matrix and generate a signal similarity measure. Dynamic time warping is an algorithm commonly used to measure the similarity of two time series. It measures their similarity by calculating the minimum alignment cost between the two sequences. Suppose the two multidimensional feature vectors are and , the DTW distance between them It can be expressed as:

[0144] ;

[0145] in, Represents the distance measure between two feature vectors, such as the Euclidean distance. Using the dynamic time warping distance matrix, the signal similarity measure is generated, and a hierarchical clustering analysis is performed on it to obtain the initial diagnostic point candidate set. Hierarchical clustering analysis builds a hierarchical structure by gradually merging or splitting data points to identify similarity patterns of signals. Initial diagnostic point candidate set is the result of cluster analysis, which contains all possible diagnostic points. Construct a Markov decision process model and generate a state transition probability matrix. Markov decision is a mathematical framework for modeling decision processes, which includes states, actions, and transition probabilities. Let the state set be , the action set is , the state transition probability matrix is ,in Indicates in status Select Action After transfer to state The probability of . The formula is as follows:

[0146] ;

[0147] The state transition probability matrix is ​​calculated by value iteration method. Solve and get the optimal diagnostic point selection strategy. The value iteration method repeatedly updates the state value and finally converges to the optimal strategy. Let the value function be , the update formula is as follows:

[0148] ;

[0149] in, is the reward function, is the discount factor. The initial diagnostic point candidate set is screened according to the optimal diagnostic point selection strategy to obtain the key diagnostic point sequence. The key diagnostic point sequence is subjected to a timing constraint check to ensure that these diagnostic points satisfy the timing logic, and a subset of diagnostic points that satisfy the timing logic is obtained. . Based on the subset of diagnostic points that satisfy the temporal logic A multi-objective optimization model is constructed and a balance function between diagnostic efficiency and risk is generated. The multi-objective optimization model considers multiple objectives at the same time to find an optimal balance point. Assume that the diagnostic efficiency is , the diagnosis risk is , then the balance function It can be expressed as:

[0150] ;

[0151] in, and is the weight coefficient, which is used to adjust the relative importance of efficiency and risk. Optimize and solve to obtain the Pareto optimal solution set. The Pareto optimal solution set contains all solutions that achieve the best balance between different objectives. According to the Pareto optimal solution set, the decision weight is allocated, and finally the optimized diagnosis node sequence is obtained. .

[0152] In a specific embodiment, the process of executing step S5 may specifically include the following steps:

[0153] Perform time series analysis on the optimized diagnostic node sequence to obtain the time distribution characteristics of the diagnostic nodes, and construct the diagnostic bed state space equation based on the time distribution characteristics of the diagnostic nodes;

[0154] According to the state space equation of the diagnostic bed, the Lyapunov function is designed to obtain the system stability criterion, and the sliding surface equation is constructed according to the system stability criterion to obtain the sliding mode control law;

[0155] The exponential approach law is designed for the sliding mode control law to obtain a fast convergence control term, and the fast convergence control term is compensated to obtain a robust control term;

[0156] An adaptive law is designed according to the robust control term to obtain a parameter adaptive update rule, and the control gain is adjusted online according to the parameter adaptive update rule to obtain an adaptive control gain;

[0157] The control input is calculated based on the adaptive control gain to obtain an initial control instruction, and the initial control instruction is subjected to saturation constraint processing to obtain a first adjustment control instruction of the intelligent medical diagnostic bed.

[0158] Specifically, the optimized diagnostic node sequence is analyzed in time series to obtain the time distribution characteristics of the diagnostic nodes. These characteristics include the frequency, periodicity and trend changes of the nodes. Suppose the optimized diagnostic node sequence is , through time series analysis, extract the time distribution characteristics of nodes , including the average interval, standard deviation and periodic components of nodes. According to the time distribution characteristics of the diagnosis nodes , construct the state space equation of the diagnostic bed. The state space equation is used to describe the dynamic changes of the system state and the impact of the input on the system. Let the state vector be , the control input is , the state space equation can be expressed as:

[0159] ;

[0160] in, is the state matrix, is the input matrix. According to the state space equation of the diagnostic bed, the Lyapunov function is designed to obtain the stability criterion of the system. The Lyapunov function is a scalar function used to judge the stability of the system. Let the Lyapunov function be , which is defined as follows:

[0161] ;

[0162] in, is a symmetric positive definite matrix. The criterion for system stability is the derivative of the Lyapunov function is negative throughout the state space, that is:

[0163] ;

[0164] The stability of the system under certain conditions is ensured by the Lyapunov stability criterion. The sliding surface equation is constructed according to the system stability criterion to obtain the sliding mode control law. Sliding mode control is a robust control method that achieves stable control of the system by restricting the system state to a sliding surface. Assume that the sliding surface equation is:

[0165] ;

[0166] in, is the sliding surface coefficient matrix. The sliding mode control law can be expressed as:

[0167] ;

[0168] in, is the control gain, is a sign function. The exponential reaching law is designed for the sliding mode control law to obtain a fast convergence control term. The exponential reaching law can accelerate the convergence speed of the system state to the sliding mode surface. The formula is as follows:

[0169] ;

[0170] in, and is the approach rate parameter. The fast convergence control term can be expressed as:

[0171] ;

[0172] The fast convergence control term is compensated to obtain the robust control term. The compensation control term can improve the robustness of the system to external interference. The formula is as follows:

[0173] ;

[0174] in, is a compensation term used to offset the uncertainty and disturbance in the system. Based on the robust control term, the adaptive law is designed to obtain the parameter adaptive update rule. The adaptive law enables the system to adapt to different environmental changes by adjusting the control parameters in real time. The formula is as follows:

[0175] ;

[0176] in, is an adaptive parameter, is the adaptive rate. According to the parameter adaptive update rule, the control gain is adjusted online to obtain the adaptive control gain:

[0177] ;

[0178] in, is the initial control gain. Based on the adaptive control gain, the control input is calculated to obtain the initial control command. The initial control command can be expressed as:

[0179] ;

[0180] The initial control command is processed by saturation constraint to obtain the first adjustment control command of the intelligent medical diagnostic bed. Saturation constraint processing can prevent the control input from exceeding the physical limit of the system. The formula is as follows:

[0181] ;

[0182] in, and are the upper and lower limits of the control input.

[0183] In a specific embodiment, the process of executing step S6 may specifically include the following steps:

[0184] performing data preprocessing on the second physiological signal data set to obtain a standardized data set;

[0185] The standardized data set is input into the multi-channel convolutional neural network layer for feature extraction to obtain a multi-dimensional feature map, and the multi-dimensional feature map is input into the long short-term memory network and the gated recurrent unit network layer for time series analysis to obtain a dynamic feature sequence;

[0186] The dynamic feature sequence is input into the attention mechanism layer to calculate the importance weight of the feature to obtain a weighted feature vector, and the weighted feature vector is reduced in dimension to obtain a compressed feature representation;

[0187] The compressed feature representation is input into the multi-task learning layer to simultaneously predict multiple physiological indicators and body position parameters to obtain preliminary prediction results;

[0188] The confidence interval of the initial prediction results is calculated through the Bayesian neural network layer to obtain the probability distribution output, and the fuzzy inference layer is used to optimize the decision-making of the probability distribution output to obtain adjustment suggestions;

[0189] The first regulating control instruction is corrected according to the adjustment suggestion to obtain a corrected control instruction, and the corrected control instruction is constrained and smoothed to obtain a second regulating control instruction.

[0190] Specifically, the second physiological signal data set is preprocessed to obtain a standardized data set. Data preprocessing includes steps such as removing noise, filling missing values, and standardization. The purpose of standardization is to convert data of different dimensions to the same scale for subsequent model training and prediction. Suppose the original data set is , the standardized dataset It can be calculated by the following formula:

[0191] ;

[0192] in, and Represent the mean and standard deviation of the data respectively. The standardized data set is input into the multi-channel convolutional neural network layer for feature extraction to obtain a multi-dimensional feature map. The multi-channel convolutional neural network extracts the local features of the data by applying convolution kernels on different channels. Suppose the standardized data set is , the convolution kernel is , the convolution operation can be expressed as:

[0193] ;

[0194] in, is a multi-dimensional feature map, * represents the convolution operation, is the bias term, and ReLU is the activation function. The multidimensional feature map is input into the long short-term memory network and the gated recurrent unit network layer for time series analysis to obtain a dynamic feature sequence. The long short-term memory network and the gated recurrent unit network layer are two commonly used recurrent neural network variants that can capture long-term dependencies in time series data. Suppose the multidimensional feature map is , the state transfer equation of the long short-term memory network can be expressed as:

[0195] ;

[0196] in, They represent the input gate, forget gate and output gate respectively. and Represent the cell state and hidden state respectively, is the sigmoid function, Denotes element-by-element product. The dynamic feature sequence is input into the attention mechanism layer to calculate the importance weight of the feature and obtain the weighted feature vector. The attention mechanism highlights the important features by calculating the weighted sum of the feature vector. Suppose the dynamic feature sequence is , the attention weight is , then the weighted eigenvector It can be expressed as:

[0197] ;

[0198] ;

[0199] in, is the attention weight matrix. The weighted feature vector is reduced in dimension to obtain a compressed feature representation. The dimension reduction can be achieved through a fully connected layer or other dimension reduction techniques. Suppose the weighted feature vector is , the compressed feature is expressed as , the dimension reduction operation can be expressed as:

[0200] ;

[0201] in, and Represent the weight and bias of dimensionality reduction respectively. The compressed feature representation is input into the multi-task learning layer to predict multiple physiological indicators and body position parameters at the same time to obtain preliminary prediction results. Multi-task learning improves the generalization ability of the model by sharing feature representation. Suppose the compressed feature representation is , the prediction results of physiological indicators are , the predicted results of body position parameters are , the output of multi-task learning can be expressed as:

[0202] ;

[0203] in, and Denote the weight and bias of multi-task learning respectively. The confidence interval of the initial prediction result is calculated through the Bayesian neural network layer to obtain the probability distribution output. The Bayesian neural network can provide uncertainty estimation of the prediction by introducing probability distribution on the parameters. Suppose the initial prediction result is , the probability distribution output of its confidence interval is , the output of the Bayesian neural network can be expressed as:

[0204] ;

[0205] in, are model parameters, is the training data. The probability distribution output is optimized through the fuzzy reasoning layer to obtain adjustment suggestions. Fuzzy reasoning can handle uncertainty through fuzzy logic and rules. Suppose the probability distribution output is , the adjustment suggestion is , the output of fuzzy reasoning can be expressed as:

[0206] ;

[0207] The first adjustment control instruction is modified according to the adjustment suggestion to obtain a modified control instruction, and the modified control instruction is constrained and smoothed to obtain a second adjustment control instruction. The modified control instruction can be expressed as:

[0208] ;

[0209] Constraining and smoothing the corrected control instructions can be achieved through the following formula:

[0210] ;

[0211] in, is a constraint operation, and They are the lower and upper limits of the control instruction respectively.

[0212] In a specific embodiment, the execution step inputs the standardized data set into the multi-channel convolutional neural network layer for feature extraction to obtain a multi-dimensional feature map, and the multi-dimensional feature map is input into the long short-term memory network and the gated recurrent unit network layer for time series analysis to obtain a dynamic feature sequence. The process can specifically include the following steps:

[0213] Perform channel separation on the standardized data set to obtain multiple single-channel data sequences;

[0214] Parallel convolution processing is performed on multiple single-channel data sequences. The parallel convolution processing includes 3 convolution modules, each of which contains 3 convolution layers. Each convolution layer uses the ReLU activation function and batch normalization.

[0215] The outputs of parallel convolution processing are fused through the channel attention mechanism to obtain a multi-dimensional feature map;

[0216] Perform spatial pyramid pooling on the multi-dimensional feature map to obtain multi-scale feature representation. Spatial pyramid pooling includes pooling operations at four different scales.

[0217] The multi-scale feature representation is input into the fully connected layer to reduce the dimension of the feature to obtain a compressed feature vector, and the compressed feature vector is reconstructed into a time series to obtain a time series feature sequence;

[0218] The time series feature sequence is input into the bidirectional long short-term memory network, which contains three layers of bidirectional LSTM units, and the hidden state dimension of each layer of LSTM units is 256;

[0219] The self-attention mechanism is applied to the output sequence of the bidirectional long short-term memory network to obtain context-enhanced features. The self-attention mechanism adopts a multi-head attention structure with the number of heads set to 8.

[0220] The context-enhanced features are input into the gated recurrent unit network, which consists of two layers of GRU units, and the hidden state dimension of each layer of GRU units is 128;

[0221] The output sequence of the gated recurrent unit network is weighted by time-step attention to obtain a dynamic feature sequence.

[0222] Specifically, the standardized data set is channel-separated to obtain multiple single-channel data sequences. Multiple single-channel data sequences are processed by parallel convolution, which includes three convolution modules, each of which contains three convolution layers. Each convolution layer uses the ReLU activation function and batch normalization to improve the nonlinear expression ability and stability of the model. Suppose the single-channel data sequence is , the convolution kernel is , No. The convolution operation of the layer can be expressed as:

[0223] ;

[0224] in, It is The feature map of the layer, * indicates the convolution operation, is the bias term, and BN is the batch normalization operation. The output of the parallel convolution processing is fused by the channel attention mechanism to obtain a multi-dimensional feature map. The channel attention mechanism improves the expression of important features by calculating the importance weight of each channel. Suppose the parallel convolution output is , the attention weight is , the output of the channel attention mechanism can be expressed as:

[0225] ;

[0226] ;

[0227] in, is the attention weight parameter, is a multi-dimensional feature map. Spatial pyramid pooling is performed on the multi-dimensional feature map to obtain a multi-scale feature representation. Spatial pyramid pooling captures multi-scale features by performing pooling operations at different scales. Spatial pyramid pooling includes four pooling operations at different scales, and the formula is as follows:

[0228] ;

[0229] in, It is Indicates The multi-scale feature representation is input into the fully connected layer for feature dimension reduction to obtain a compressed feature vector, and the compressed feature vector is reconstructed into a time series to obtain a time series feature sequence. Assume that the multi-scale feature representation is , the dimension reduction operation of the fully connected layer can be expressed as:

[0230] ;

[0231] in, is the compressed feature vector, and are the weights and biases of the fully connected layer respectively. The time series feature sequence is input into the bidirectional long short-term memory network (BiLSTM). The bidirectional long short-term memory network contains 3 layers of bidirectional LSTM units, and the hidden state dimension of each layer of LSTM units is 256. Bidirectional LSTM captures the long-term dependency in the sequence through forward and backward bidirectional processing. Suppose the time series feature sequence is , the state transition equation of the bidirectional LSTM can be expressed as:

[0232] ;

[0233] in, and Represent the hidden states of the forward and backward LSTMs, respectively. is the output of the bidirectional LSTM. The self-attention mechanism is applied to the output sequence of the bidirectional LSTM to obtain the context-enhanced features. The self-attention mechanism highlights important features by calculating the weighted sum of feature vectors. The self-attention mechanism adopts a multi-head attention structure with the number of heads set to 8. Suppose the output of the bidirectional LSTM is , the attention weight is , the output of the self-attention mechanism can be expressed as:

[0234] ;

[0235] ;

[0236] in, is the attention weight matrix. The context enhancement feature is input into the gated recurrent unit (GRU) network. The gated recurrent unit network contains 2 layers of GRU units, and the hidden state dimension of each layer of GRU units is 128. GRU effectively captures important information in the sequence through the gating mechanism and reduces information loss. Let the context enhancement feature be , the state transfer equation of GRU can be expressed as:

[0237] ;

[0238] in, and represent the update gate and reset gate respectively, is the hidden state, is a sigmoid function. The output sequence of the gated recurrent unit network is weighted by time-step attention to obtain a dynamic feature sequence. Time-step attention weighting highlights the features of key time steps by calculating the importance weight of each time step.

[0239] The above describes the control method of the intelligent medical diagnostic bed based on the Internet of Things in the embodiment of the present invention. The following describes the control system of the intelligent medical diagnostic bed based on the Internet of Things in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, an embodiment of a smart medical diagnostic bed control system based on the Internet of Things includes:

[0240] An acquisition module, used for acquiring multi-dimensional physiological data of a target patient based on a sensor group of an intelligent medical diagnostic bed, to obtain a first physiological signal data set;

[0241] An analysis module, used for performing a third-order extended state observation analysis on the first physiological signal data set to obtain vital sign state estimation data;

[0242] A processing module, used for performing nonlinear tracking differential processing on the vital sign state estimation data to obtain smoothed physiological signal data;

[0243] A modeling module is used to perform intelligent diagnostic point planning modeling on smoothed physiological signal data to obtain an optimized diagnostic node sequence;

[0244] A calculation module, used for performing adaptive nonlinear control law calculation based on an optimized diagnosis node sequence to obtain a first adjustment control instruction of the intelligent medical diagnostic bed;

[0245] The adjustment module is used to execute the first adjustment control instruction and collect the second physiological signal data set of the target patient, and perform deep learning analysis and dynamic adjustment of the control instruction on the second physiological signal data set to obtain the second adjustment control instruction.

[0246] Through the synergy of the above components and the integration of multiple sensors, the comprehensive collection of multi-dimensional physiological data such as patient body pressure distribution, bioelectric signals, surface temperature, micro-motion behavior, respiratory heart sounds, etc. is achieved, providing a rich data basis for subsequent analysis. The third-order extended state observation and analysis technology can more accurately estimate the patient's vital signs, including high-order derivative information that is difficult to measure directly, improving the accuracy and reliability of state estimation. Through nonlinear tracking differential processing, noise and interference in physiological signals can be effectively removed, and the key features of the signals can be retained. Multi-objective optimization and timing constraint inspection are used to realize intelligent planning of diagnostic nodes, while ensuring diagnostic efficiency and reducing risks. Based on sliding mode control and adaptive law design, precise control of diagnostic beds is achieved, which can quickly respond to changes in patient status and improve the robustness and adaptability of control. Combined with multi-channel convolutional neural networks, long short-term memory networks and attention mechanisms, in-depth analysis of physiological signals is carried out, complex spatiotemporal features are extracted, and the accuracy and interpretability of diagnosis are improved. Through Bayesian neural networks and fuzzy reasoning, dynamic adjustment of control instructions is realized, and refined control can be carried out according to real-time feedback from patients, improving the comfort and safety of the diagnosis and treatment process. Comprehensively utilize data from multiple sensors to achieve information complementarity and coordination. Through parallel processing and efficient algorithm design, ensure the real-time response capability of the system to meet the timeliness requirements of medical diagnosis.

[0247] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0248] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0249] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A control method for an intelligent medical diagnostic bed based on the Internet of Things, characterized in that: The method comprises: The sensor group based on the intelligent medical diagnostic bed collects multi-dimensional physiological data of the target patient to obtain a first physiological signal data set; The method comprises: performing a third-order extended state observation analysis on the first physiological signal data set to obtain vital sign state estimation data; specifically comprising: performing nonlinear state space modeling on the first physiological signal data set to obtain a state equation and an observation equation, and constructing a structure matrix of a third-order extended state observer according to the state equation and the observation equation; initializing parameters of the structure matrix of the third-order extended state observer to obtain an initial state estimation value and an observer gain matrix, and performing a linear combination of the first physiological signal data set according to the initial state estimation value and the observer gain matrix to obtain an error feedback term for state estimation; and performing a linear combination of the first physiological signal data set according to the initial state estimation value and the observer gain matrix to obtain an error feedback term for state estimation. The feed term and the state equation are numerically integrated to obtain the target state estimate at the current moment, and the observation error is calculated according to the target state estimate to obtain an observation residual sequence; adaptive weight calculation is performed on the observation residual sequence to obtain an adaptive adjustment factor, and the observer gain matrix is ​​dynamically updated according to the adaptive adjustment factor to obtain optimized observer parameters; error covariance analysis is performed on the optimized observer parameters and the target state estimate to obtain a confidence interval of the state estimate, and data fusion and reasoning are performed according to the target state estimate and the confidence interval to obtain vital sign state estimation data; Performing nonlinear tracking differentiation processing on the vital sign state estimation data to obtain smoothed physiological signal data; Intelligent diagnostic point planning modeling is performed on the smoothed physiological signal data to obtain an optimized diagnostic node sequence; specifically, the method includes: performing multi-scale feature extraction on the smoothed physiological signal data to obtain a multi-dimensional feature vector; constructing a dynamic time warping distance matrix based on the multi-dimensional feature vector and generating a signal similarity metric, and performing hierarchical clustering analysis on the signal similarity metric to obtain an initial diagnostic point candidate set; constructing a Markov decision process model based on the initial diagnostic point candidate set and generating a state transition probability matrix, and performing value iteration solution on the state transition probability matrix to obtain an optimal diagnostic point selection strategy; screening the initial diagnostic point candidate set according to the optimal diagnostic point selection strategy to obtain a key diagnostic point sequence, and performing a timing constraint check on the key diagnostic point sequence to obtain a diagnostic point subset that satisfies the timing logic; constructing a multi-objective optimization model based on the diagnostic point subset that satisfies the timing logic and generating a balance function of diagnostic efficiency and risk; optimizing and solving the balance function to obtain a Pareto optimal solution set, and performing decision weight allocation based on the Pareto optimal solution set to obtain an optimized diagnostic node sequence; Performing adaptive nonlinear control law calculation based on the optimized diagnostic node sequence to obtain a first adjustment control instruction for the intelligent medical diagnostic bed; The first adjustment control instruction is executed and a second physiological signal data set of the target patient is collected, and deep learning analysis and dynamic adjustment of the control instruction are performed on the second physiological signal data set to obtain a second adjustment control instruction.

2. The control method of the intelligent medical diagnostic bed based on the Internet of Things according to claim 1 is characterized in that: The sensor group based on the intelligent medical diagnostic bed collects multi-dimensional physiological data of the target patient to obtain a first physiological signal data set, including: Reading data from a pressure sensor array of an intelligent medical diagnostic bed to obtain body pressure distribution data of a target patient, and extracting posture features of the target patient according to the body pressure distribution data to obtain a posture feature vector; Collecting signals from the bioelectric sensor of the intelligent medical diagnostic bed to obtain original bioelectric signals, and performing wavelet transform denoising on the original bioelectric signals to obtain filtered bioelectric signals; Performing image segmentation on the thermal imaging data collected by the infrared imaging device of the intelligent medical diagnostic bed to obtain a body surface temperature distribution map of the target patient, and performing hot spot detection and temperature gradient analysis on the body surface temperature distribution map to obtain abnormal body temperature area data; Performing motion estimation on the image sequence captured by the high-definition camera of the intelligent medical diagnostic bed to obtain micro-motion behavior data of the target patient, and performing time-frequency domain feature extraction on the micro-motion behavior data to obtain activity state features of the target patient; Performing spectrum analysis on the audio signal collected by the sound sensor of the intelligent medical diagnostic bed to obtain respiratory sound and heart sound characteristics; Multimodal data fusion is performed based on the posture feature vector, the filtered bioelectric signal, the abnormal body temperature area data, the activity state characteristics, the respiratory sound and the heart sound characteristics to obtain a first physiological signal data set.

3. The control method of the intelligent medical diagnostic bed based on the Internet of Things according to claim 1 is characterized in that: The step of performing nonlinear tracking differentiation processing on the vital sign state estimation data to obtain smoothed physiological signal data includes: Performing time series decomposition on the vital sign state estimation data to obtain trend terms, periodic terms and random terms, and constructing a state equation of a nonlinear tracking differentiator based on the trend terms to obtain an initial model of the differentiator; Discretizing the initial model to obtain a differential equation of a discrete-time nonlinear tracking differentiator, and designing a fast convergence factor according to the differential equation to obtain a fast tracking differentiation strategy; Adaptively adjusting the bandwidth of the fast tracking differential strategy to obtain a dynamic bandwidth adjustment strategy, and filtering the periodic item according to the dynamic bandwidth adjustment strategy to obtain a smooth periodic signal; Performing wavelet threshold denoising on the random item to obtain a denoised random signal, and performing signal reconstruction based on the smoothed periodic signal and the denoised random signal to obtain a preliminary smoothed signal; The preliminary smoothed signal is subjected to edge-preserving filtering to obtain a smoothed signal retaining the sudden change feature, and weighted fusion is performed based on the smoothed signal retaining the sudden change feature and the vital sign state estimation data to obtain smoothed physiological signal data.

4. The control method of the intelligent medical diagnostic bed based on the Internet of Things according to claim 3 is characterized in that: The performing wavelet threshold denoising on the random item to obtain a denoised random signal, and performing signal reconstruction according to the smoothed periodic signal and the denoised random signal to obtain a preliminary smoothed signal, comprises: Performing wavelet decomposition on the random term to obtain multi-scale wavelet coefficients, and calculating the noise variance of each scale according to the multi-scale wavelet coefficients to obtain a noise level estimation value; Performing nonlinear mapping on the noise level estimation value to obtain an adaptive threshold function, and performing soft threshold processing on the multi-scale wavelet coefficients according to the adaptive threshold function to obtain denoised wavelet coefficients; Performing wavelet reconstruction on the denoised wavelet coefficients to obtain a denoised random signal; Building a signal fusion model according to the smoothed periodic signal and the denoised random signal, and calculating an initial fusion weight; Dynamically optimizing the initial fusion weight to obtain an adaptive fusion weight coefficient, and performing weighted summation on the smoothed periodic signal and the denoised random signal according to the adaptive fusion weight coefficient to obtain a fusion signal; The fused signal is subjected to Hilbert-Huang transform to obtain the instantaneous frequency and amplitude of the signal, and the signal is reconstructed according to the instantaneous frequency and amplitude to obtain a preliminary smoothed signal.

5. The control method of the intelligent medical diagnostic bed based on the Internet of Things according to claim 1 is characterized in that: The step of performing adaptive nonlinear control law calculation based on the optimized diagnostic node sequence to obtain a first adjustment control instruction for the intelligent medical diagnostic bed includes: Performing time series analysis on the optimized diagnostic node sequence to obtain time distribution characteristics of the diagnostic nodes, and constructing a diagnostic bed state space equation according to the time distribution characteristics of the diagnostic nodes; Designing a Lyapunov function according to the state space equation of the diagnostic bed to obtain a system stability criterion, and constructing a sliding surface equation according to the system stability criterion to obtain a sliding mode control law; Performing an exponential approach law design on the sliding mode control law to obtain a fast convergence control term, and compensating the fast convergence control term to obtain a robust control term; Designing an adaptive law according to the robust control item to obtain a parameter adaptive update rule, and adjusting the control gain online according to the parameter adaptive update rule to obtain an adaptive control gain; The control input is calculated based on the adaptive control gain to obtain an initial control instruction, and a saturation constraint process is performed on the initial control instruction to obtain a first adjustment control instruction of the intelligent medical diagnostic bed.

6. The control method of the intelligent medical diagnostic bed based on the Internet of Things according to claim 1 is characterized in that: The executing the first regulating control instruction and collecting the second physiological signal data set of the target patient, and performing deep learning analysis and dynamic adjustment of the control instruction on the second physiological signal data set to obtain the second regulating control instruction includes: performing data preprocessing on the second physiological signal data set to obtain a standardized data set; Inputting the standardized data set into a multi-channel convolutional neural network layer for feature extraction to obtain a multi-dimensional feature map, and inputting the multi-dimensional feature map into a long short-term memory network and a gated recurrent unit network layer for time series analysis to obtain a dynamic feature sequence; Inputting the dynamic feature sequence into the attention mechanism layer to calculate the importance weight of the feature to obtain a weighted feature vector, and performing dimensionality reduction processing on the weighted feature vector to obtain a compressed feature representation; Inputting the compressed feature representation into a multi-task learning layer to simultaneously predict multiple physiological indicators and body position parameters to obtain a preliminary prediction result; Calculating the confidence interval of the preliminary prediction result through the Bayesian neural network layer to obtain a probability distribution output, and optimizing the decision of the probability distribution output through the fuzzy reasoning layer to obtain adjustment suggestions; The first regulating control instruction is corrected according to the adjustment suggestion to obtain a corrected control instruction, and the corrected control instruction is constrained and smoothed to obtain a second regulating control instruction.

7. The control method of the intelligent medical diagnostic bed based on the Internet of Things according to claim 6 is characterized in that: The standardized data set is input into a multi-channel convolutional neural network layer for feature extraction to obtain a multi-dimensional feature map, and the multi-dimensional feature map is input into a long short-term memory network and a gated recurrent unit network layer for time series analysis to obtain a dynamic feature sequence, including: Performing channel separation on the standardized data set to obtain multiple single-channel data sequences; Performing parallel convolution processing on the multiple single-channel data sequences, the parallel convolution processing includes three convolution modules, each convolution module includes three convolution layers, and each convolution layer uses a ReLU activation function and batch normalization; The outputs of parallel convolution processing are fused through the channel attention mechanism to obtain a multi-dimensional feature map; Performing spatial pyramid pooling on the multidimensional feature map to obtain a multi-scale feature representation, wherein the spatial pyramid pooling includes pooling operations of four different scales; Input the multi-scale feature representation into a fully connected layer to perform feature dimension reduction to obtain a compressed feature vector, and perform time series reconstruction on the compressed feature vector to obtain a time series feature sequence; Inputting the temporal feature sequence into a bidirectional long short-term memory network, wherein the bidirectional long short-term memory network comprises three layers of bidirectional LSTM units, and the hidden state dimension of each layer of LSTM units is 256; Applying a self-attention mechanism to the output sequence of the bidirectional long short-term memory network to obtain context-enhanced features, wherein the self-attention mechanism adopts a multi-head attention structure, and the number of heads is set to 8; Inputting the context enhancement feature into a gated recurrent unit network, wherein the gated recurrent unit network comprises two layers of GRU units, and the hidden state dimension of each layer of GRU units is 128; The output sequence of the gated recurrent unit network is weighted by time-step attention to obtain a dynamic feature sequence.

8. An intelligent medical diagnostic bed control system based on the Internet of Things, characterized in that: Used to execute the control method of the intelligent medical diagnostic bed based on the Internet of Things as described in any one of claims 1 to 7, the system comprises: An acquisition module, used for acquiring multi-dimensional physiological data of a target patient based on a sensor group of an intelligent medical diagnostic bed, to obtain a first physiological signal data set; The analysis module is used to perform a third-order extended state observation analysis on the first physiological signal data set to obtain vital sign state estimation data; specifically comprising: performing nonlinear state space modeling on the first physiological signal data set to obtain a state equation and an observation equation, and constructing a structure matrix of a third-order extended state observer according to the state equation and the observation equation; initializing parameters of the structure matrix of the third-order extended state observer to obtain an initial state estimation value and an observer gain matrix, and performing a linear combination of the first physiological signal data set according to the initial state estimation value and the observer gain matrix to obtain an error feedback term for state estimation; The error feedback term and the state equation are numerically integrated to obtain the target state estimate at the current moment, and the observation error is calculated according to the target state estimate to obtain an observation residual sequence; adaptive weight calculation is performed on the observation residual sequence to obtain an adaptive adjustment factor, and the observer gain matrix is ​​dynamically updated according to the adaptive adjustment factor to obtain optimized observer parameters; error covariance analysis is performed on the optimized observer parameters and the target state estimate to obtain a confidence interval of the state estimate, and data fusion and reasoning are performed according to the target state estimate and the confidence interval to obtain vital sign state estimation data; A processing module, used for performing nonlinear tracking differential processing on the vital sign state estimation data to obtain smoothed physiological signal data; A modeling module is used to perform intelligent diagnostic point planning modeling on the smoothed physiological signal data to obtain an optimized diagnostic node sequence; specifically comprising: performing multi-scale feature extraction on the smoothed physiological signal data to obtain a multi-dimensional feature vector; constructing a dynamic time warping distance matrix based on the multi-dimensional feature vector and generating a signal similarity metric, and performing hierarchical clustering analysis on the signal similarity metric to obtain an initial diagnostic point candidate set; constructing a Markov decision process model based on the initial diagnostic point candidate set and generating a state transition probability matrix, and performing value iteration solution on the state transition probability matrix to obtain an optimal diagnostic point selection strategy; screening the initial diagnostic point candidate set according to the optimal diagnostic point selection strategy to obtain a key diagnostic point sequence, and performing a timing constraint check on the key diagnostic point sequence to obtain a diagnostic point subset that satisfies the timing logic; constructing a multi-objective optimization model based on the diagnostic point subset that satisfies the timing logic and generating a balance function of diagnostic efficiency and risk; optimizing and solving the balance function to obtain a Pareto optimal solution set, and performing decision weight allocation based on the Pareto optimal solution set to obtain an optimized diagnostic node sequence; A calculation module, used for performing adaptive nonlinear control law calculation based on the optimized diagnosis node sequence to obtain a first adjustment control instruction of the intelligent medical diagnostic bed; An adjustment module is used to execute the first adjustment control instruction and collect a second physiological signal data set of the target patient, and perform deep learning analysis and dynamic adjustment of the control instruction on the second physiological signal data set to obtain a second adjustment control instruction.

Images