Artificial intelligence assisted adaptive sickbed control system and method

Abstract

The invention relates to the technical field of artificial intelligence, and discloses an artificial intelligence-assisted adaptive sickbed control system and method, and the method comprises the steps: carrying out the noise reduction and normalization preprocessing of original physiological data collected by a multi-modal sensor, and obtaining standardized multi-modal data; performing time sequence fusion and dimension reduction to obtain low-dimensional fusion features; performing feature extraction through a multi-branch convolutional neural network to obtain generalization feature representation; the optimal posture demand of the patient is predicted through a long-short term memory network model with an attention mechanism, and an optimal posture demand prediction result is obtained; a layered sickbed adjustment strategy is generated through a double-time-scale reinforcement learning network comprising a priority experience playback mechanism; according to the intelligent sickbed control method and system, the multi-mode sensing data can be comprehensively utilized, the adaptive learning ability is achieved, the intelligent sickbed can be accurately controlled, and the performance and the application effect of the intelligent sickbed are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-assisted adaptive hospital bed control system and method. Background Art

[0002] With the aging of the population and the increase in the number of patients with chronic diseases, the demand for care for long-term bedridden patients has become increasingly prominent. Traditional manually adjustable beds have problems such as cumbersome operation, untimely adjustment, and difficulty in meeting personalized needs, which seriously affect the comfort and rehabilitation effect of patients. In recent years, the research on intelligent bed systems has made certain progress, but the existing technology still has shortcomings such as limited perception ability, single control strategy, and insufficient adaptability. Most smart beds can only be adjusted according to preset fixed modes, and it is difficult to make flexible adjustments according to the patient's real-time status and environmental changes. In addition, due to the lack of effective multimodal data fusion and analysis methods, it is difficult for the existing system to accurately grasp the potential needs of patients, resulting in unsatisfactory adjustment effects.

[0003] In order to solve these problems, researchers began to explore the introduction of artificial intelligence technology into bed control systems. However, how to effectively use multi-source heterogeneous data to achieve accurate perception and prediction of patient status remains a huge challenge. At the same time, considering the individual differences and changes in the condition of different patients, how to design a control algorithm with adaptive capabilities so that the bed can be intelligently adjusted according to the real-time needs of the patient is also a key issue that needs to be solved urgently. In addition, due to the particularity of the medical environment, the intelligent bed system also needs to have strong generalization and robustness to adapt to different application scenarios and patient groups. Summary of the invention

[0004] The present invention provides an artificial intelligence-assisted adaptive hospital bed control system and method. The present invention can comprehensively utilize multimodal sensor data, has adaptive learning capabilities, and can realize precise control of an intelligent hospital bed, thereby improving the performance and application effect of the intelligent hospital bed.

[0005] In a first aspect, the present invention provides an artificial intelligence-assisted adaptive hospital bed control method, the artificial intelligence-assisted adaptive hospital bed control method comprising: Perform noise reduction and normalization preprocessing on the original physiological data collected by the multimodal sensor to obtain standardized multimodal data; Performing time series fusion on the standardized multimodal data and performing dimensionality reduction through principal component analysis to obtain low-dimensional fusion features; Inputting the low-dimensional fusion features into a multi-branch convolutional neural network for feature extraction, and implementing domain adaptation through adversarial training to obtain a generalized feature representation; Taking the generalized feature representation as input, predicting the patient's optimal posture requirement through a long short-term memory network model with an attention mechanism, and obtaining an optimal posture requirement prediction result; Based on the optimal posture demand prediction results, a hierarchical bed adjustment strategy is generated through a dual-time scale reinforcement learning network including a priority experience playback mechanism; Based on the stratified bed adjustment strategy, combined with model predictive control and adaptive PID controller, real-time bed adjustment is performed and online learning and updating are performed to output optimized bed control instructions.

[0006] In a second aspect, the present invention provides an artificial intelligence-assisted adaptive hospital bed control system, the artificial intelligence-assisted adaptive hospital bed control system comprising: A preprocessing module is used to perform noise reduction and normalization preprocessing on the original physiological data collected by the multimodal sensor to obtain standardized multimodal data; A fusion module is used to perform time series fusion on the standardized multimodal data and perform dimensionality reduction through principal component analysis to obtain low-dimensional fusion features; An extraction module, used for inputting the low-dimensional fusion features into a multi-branch convolutional neural network for feature extraction, and realizing domain adaptation through adversarial training to obtain a generalized feature representation; A prediction module, used for taking the generalized feature representation as input, predicting the patient's optimal posture requirement through a long short-term memory network model with an attention mechanism, and obtaining an optimal posture requirement prediction result; A generation module, for generating a hierarchical bed adjustment strategy based on the optimal posture demand prediction result through a dual-time scale reinforcement learning network including a priority experience playback mechanism; The output module is used to perform real-time bed adjustment and online learning update based on the stratified bed adjustment strategy, combined with model predictive control and adaptive PID controller, and output optimized bed control instructions.

[0007] A third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned artificial intelligence-assisted adaptive bed control method.

[0008] A fourth aspect of the present invention provides a computer-readable storage medium, which stores instructions that, when executed on a computer, enable the computer to execute the above-mentioned artificial intelligence-assisted adaptive bed control method.

[0009] In the technical solution provided by the present invention, by performing noise reduction, normalization and time series fusion on the data collected by multiple sensors, and combining principal component analysis for dimensionality reduction, the key features of the patient's state are effectively extracted, data redundancy is reduced, and the efficiency of subsequent processing is improved. Multi-branch convolutional neural network and adversarial training are used to achieve domain adaptation, and feature representation with strong generalization ability is obtained, so that the system can adapt to different patient groups and application environments. The long short-term memory network model with attention mechanism can better capture the time dependence of the patient's state and improve the prediction accuracy of the optimal posture requirement. Through the dual-time scale reinforcement learning network, a hierarchical control strategy combining coarse adjustment and fine adjustment is realized, which not only ensures the accuracy of large-scale adjustment, but also realizes fine adjustment of small scale. By giving priority to the experience playback buffer, the utilization rate of important samples is improved, the convergence process of reinforcement learning is accelerated, and the learning efficiency is improved. Combining model predictive control with adaptive PID controller not only takes into account the prediction of future states, but also adjusts parameters according to real-time feedback, thereby improving the accuracy and robustness of control. The bed dynamics model is updated in real time through an incremental learning algorithm, allowing the system to continuously adapt to changes in patient status and environmental interference, maintaining long-term control effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] 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.

[0011] Figure 1 Schematic diagram of the steps of the artificial intelligence-assisted adaptive hospital bed control method in an embodiment of the present invention; Figure 2 Schematic diagram of the structure of an artificial intelligence-assisted adaptive bed control system in an embodiment of the present invention. DETAILED DESCRIPTION

[0012] Embodiments of the present invention provide an artificial intelligence-assisted adaptive bed control system and method. 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 interchanged 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.

[0013] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , an embodiment of the artificial intelligence-assisted adaptive hospital bed control method in the embodiment of the present invention includes: Step S1, performing noise reduction and normalization preprocessing on the original physiological data collected by the multimodal sensor to obtain standardized multimodal data; It is understandable that the execution subject of the present invention may be an artificial intelligence-assisted adaptive bed control system, or 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.

[0014] Specifically, the weight distribution data collected by the pressure sensor is subjected to wavelet transform, which can effectively remove the noise components in the data and obtain the denoised pressure data. The tilt angle data collected by the acceleration sensor is subjected to wavelet transform, and the denoised acceleration data is obtained by decomposing the signal and removing the high-frequency noise components. For the three-dimensional posture image collected by the depth camera, Gaussian filtering is used for processing. Gaussian filtering can effectively reduce the noise and irregularity in the image by performing convolution operations on the image to obtain smoothed image data. At the same time, for the environmental parameter data collected by the temperature and humidity sensor and the air quality sensor, moving average filtering is used for smoothing, and the window size is set to 5. By averaging the data points in the continuous data window, short-term fluctuations are eliminated to obtain smoothed environmental data. Based on the maximum-minimum normalization method, the denoised pressure data, the denoised acceleration data, the smoothed image data and the smoothed environmental data are normalized. The maximum-minimum normalization effectively eliminates the dimensional differences between different sensor data by mapping the data into a fixed range, ensures the uniformity and comparability of the data, and obtains standardized multimodal data.

[0015] Step S2, performing time series fusion on the standardized multimodal data, and performing dimensionality reduction through principal component analysis to obtain low-dimensional fusion features; Specifically, the standardized multimodal data is processed to construct a state vector containing information such as the patient's position, posture angle, and environmental parameters, which represents the initial state representation of the surroundings of the bed and the patient's current state. Based on the accuracy characteristics of each sensor, the covariance matrix of the observation noise is set to obtain the initial parameters of the Kalman filter. The Kalman filter algorithm is applied to the initial state representation, and the observation data from different sensors are effectively combined by executing the prediction step and the update step to obtain the fused state estimate. Based on the fused state estimate, the timestamp alignment processing is performed to synchronize all sensor data to a unified time base with a time resolution of 1 millisecond to obtain time-synchronized multimodal data and eliminate the time difference between the data. The covariance matrix of the time-synchronized multimodal data is calculated, and the eigenvalue decomposition of the covariance matrix is ​​performed to obtain the eigenvalue and the corresponding eigenvector. The eigenvectors are sorted according to the size of the eigenvalue, and the first K eigenvectors with a cumulative contribution rate of 95% are selected, and the projection matrix is ​​constructed using these eigenvectors. Through the projection matrix, the time-synchronized multimodal data is linearly transformed to achieve data dimensionality reduction processing and obtain the dimensionality-reduced feature data. In order to ensure the consistency of data and the effectiveness of processing, the feature data after dimensionality reduction is standardized again to finally obtain low-dimensional fusion features.

[0016] Step S3, inputting the low-dimensional fusion features into a multi-branch convolutional neural network for feature extraction, and implementing domain adaptation through adversarial training to obtain a generalized feature representation; Specifically, the low-dimensional fusion features are segmented, and the pressure data features, acceleration data features and image data features are extracted respectively. The pressure data features are input into the pressure branch network in the multi-branch convolutional neural network. The network structure contains three one-dimensional convolutional layers, each of which uses 32 convolution kernels of size 3×1, and the step size is set to 1. The LeakyReLU activation function is applied to the output in the convolution operation for nonlinear transformation, and batch normalization is performed to ensure the stability of the model and accelerate the training process. After the global maximum pooling operation, the pressure feature vector with global characteristics is extracted. The acceleration data features are input into the acceleration branch network in the multi-branch convolutional neural network. The network consists of three one-dimensional convolutional layers, each of which uses 64 convolution kernels of size 3×1 and a step size of 2. Each convolutional layer is followed by an ELU activation function to enhance the network's ability to handle negative inputs, and the Dropout regularization technology is used to prevent the model from overfitting. The acceleration feature vector is obtained through the global average pooling operation. At the same time, the image data features are input into the image branch network in the multi-branch convolutional neural network, which consists of 5 two-dimensional convolutional layers. The convolution kernel size of each convolutional layer is 7×7, 5×5, 3×3, 3×3, and 3×3, respectively, and the number of channels increases layer by layer, which are 64, 128, 256, 512, and 512, respectively. The ReLU activation function is applied after each convolutional layer, so that the network can effectively extract image features, and downsample through the maximum pooling layer to reduce the size of the feature map. The feature map is compressed into a fixed-size image feature vector using adaptive average pooling. The pressure feature vectors, acceleration feature vectors, and image feature vectors extracted from each branch network are concatenated and fused through a fully connected layer to obtain the initial fused features. In order to improve the generalization ability of the features, a domain adversarial network is constructed, which includes a feature extractor and a domain classifier. The feature extractor consists of a 3-layer fully connected network, each layer uses 256 neurons and adopts a ReLU activation function to ensure the nonlinear expression ability of the features. The domain classifier consists of a 2-layer fully connected network, with 128 neurons in each layer, and the robustness of the model is enhanced by the LeakyReLU activation function. The initial fused features are input into the feature extractor to obtain the intermediate feature representation, and the intermediate feature representation is input into the domain classifier. The parameters of the feature extractor are updated through the gradient reversal layer to perform adversarial training. This process aims to enable the feature extractor to learn feature representations that can be consistent in different data domains, thereby achieving domain adaptation. The attention mechanism is applied to the intermediate feature representation. By calculating the weights of the features, the key features are highlighted and the secondary features are suppressed. The intermediate feature representations are weighted and combined according to the feature weights to obtain the generalized feature representation. The generalized feature representation can adapt to changes in different environments and conditions.

[0017] The input image data features are zero-filled, and the border area of ​​the image is filled with zero values ​​of size 3 to keep the edge information from being lost during the convolution process, and the padded image data is obtained. The padded image data is input to the first convolution layer, which uses 64 convolution kernels of size 7×7 and sets the step size to 2 to perform convolution operations on the image data to obtain the first layer feature map. The ReLU activation function is then applied to the first layer feature map for nonlinear transformation, so that the network can capture more complex features. Then a 2×2 maximum pooling operation is performed to reduce the size of the feature map by pooling while retaining important feature information to obtain the first layer output feature map. The first layer output feature map is input to the second convolution layer, which uses 128 convolution kernels of size 5×5 and sets the step size to 1. The first layer output feature map is convolved to generate the second layer feature map. The ReLU activation function is also applied to the second layer feature map for activation processing, and the size of the feature map is further reduced by 2×2 maximum pooling to obtain the second layer output feature map. The output feature map of the second layer is input to the third convolution layer, which uses 256 convolution kernels of size 3×3 and a step size of 1 to perform convolution operations to generate the feature map of the third layer. After the feature map of the third layer is nonlinearly activated by the ReLU activation function, it continues to perform 2×2 maximum pooling to obtain the output feature map of the third layer. The output feature map of the third layer is input to the fourth convolution layer. The fourth layer uses 512 convolution kernels of size 3×3 and the step size is still set to 1. The fourth layer feature map is obtained by convolution operation. The fourth layer feature map is activated by the ReLU activation function and then subjected to 2×2 maximum pooling to obtain the output feature map of the fourth layer. The output feature map of the fourth layer is input to the fifth convolution layer. The fifth layer also uses 512 convolution kernels of size 3×3 and a step size of 1 to perform convolution operations on the output feature map of the fourth layer to generate the feature map of the fifth layer. The feature map of the fifth layer is ReLU activated again and subjected to 2×2 maximum pooling to finally obtain the output feature map of the fifth layer. An adaptive average pooling operation is performed on the output feature map of the fifth layer to adjust the size of the feature map to a fixed size H×W so as to maintain the consistency of features in subsequent processing steps. The adjusted fixed-size feature map is flattened to convert the two-dimensional feature map into a one-dimensional vector to obtain the image feature vector.

[0018] Step S4, taking the generalized feature representation as input, predicting the patient's optimal posture requirement through a long short-term memory network model with an attention mechanism, and obtaining an optimal posture requirement prediction result; Specifically, the generalized feature representation is segmented into time series, and the continuous feature data is divided into input sequences of fixed length T to ensure that the model can capture the dynamic changes of the patient's status. The input sequence is input into the LSTM layer of the long short-term memory network (LSTM) model with an attention mechanism. The LSTM layer consists of 128 hidden units. Through the internal gating mechanism (including forget gate, input gate and output gate) and the update of hidden states, the model can effectively memorize and extract the temporal information in the input sequence and generate the output sequence of the LSTM layer. The self-attention mechanism is applied to the sequence output by the LSTM layer. By calculating the query, key and value matrices and using the scaled dot product attention algorithm, the importance weight of the feature at each time step is obtained, and the attention weight is multiplied with the corresponding value matrix to generate the weighted feature representation. In order to enhance the stability and expressiveness of the feature representation, the weighted feature representation is residually connected with the output sequence of the original LSTM layer, retaining the original information while adding the enhanced features brought by the attention mechanism. The result after the residual connection is layer normalized to ensure that the feature maintains the numerical stability during the transmission process and obtain the fused feature sequence. The fused feature sequence is input into a feedforward neural network, which contains two fully connected layers. The first layer uses 256 neurons and introduces nonlinear characteristics through the ReLU activation function; the second layer uses 128 neurons and compresses the feature space through the tanh activation function to obtain the converted feature sequence. In order to improve the effect of feature representation, the converted feature sequence is residually connected with the previous fused feature sequence again, and the result is normalized for the second layer to obtain a more representative feature sequence. The temporal attention mechanism is applied to the final feature sequence, and the trend and law of feature changes over time are captured by calculating the importance weights of each time step. According to the importance weights, the final feature sequence is weighted averaged to obtain a context vector, which represents the patient's posture demand characteristics at different time points. The context vector is input into the last fully connected layer, and the output of the network is converted into a probability distribution of multiple posture categories through the softmax activation function. By selecting the category with the highest probability, the best posture demand prediction result of the patient is obtained.

[0019] The input fusion feature sequence is flattened, and the three-dimensional tensor is converted into a two-dimensional matrix to generate a flattened feature matrix, which contains linear expansion of information such as time steps and feature channels. The flattened feature matrix is ​​input into the first fully connected layer of the feedforward neural network. In this layer, the input data is linearly transformed using the weight matrix W1 and the bias vector b1 to obtain the intermediate feature vector. In order to introduce nonlinear features, the ReLU activation function is operated on the intermediate feature vector. By performing the max(0, x) operation on each element, the ReLU function suppresses all negative values ​​to zero while keeping the positive values ​​unchanged to obtain the activation output of the first layer. In order to prevent overfitting of the model, the Dropout regularization technique is applied to the activation output of the first layer. By randomly setting 50% of the neuron outputs to 0, Dropout randomly ignores some neurons during the training process, making the model more robust to different features of the input data, avoiding dependence on specific features or neurons, and generating the first layer of regularized output. The first layer of regularized output is input into the second fully connected layer of the feedforward neural network. The regularized output is linearly transformed using the weight matrix W2 and the bias vector b2 to generate the second layer of feature vectors. The tanh activation function is operated on the second-layer feature vector. The tanh function maps each element to the interval [-1, 1]. Through this activation function, the model can capture the symmetry and complex nonlinear relationship of the input features and obtain the second-layer activation output. In order to improve the stability and training efficiency of the model, batch normalization is performed on the second-layer activation output. The batch normalization process includes calculating the mean and variance of the activation output, and then scaling and translating it to obtain a normalized feature vector. Batch normalization can not only speed up the convergence of the model, but also alleviate the problem of gradient disappearance or explosion, so that the model maintains good performance at different training stages. The normalized feature vector is reshaped to the same time step as the input fusion feature sequence. Through this operation, the converted feature sequence can maintain consistency in the time dimension with the original input, and adapt to subsequent time series modeling or feature processing.

[0020] Step S5: for the optimal posture requirement prediction result, a hierarchical bed adjustment strategy is generated through a dual-time scale reinforcement learning network including a priority experience playback mechanism; Specifically, a state vector is constructed based on the prediction result of the optimal posture requirement. The state vector contains key information such as the current patient's posture, target posture, and environmental parameters, which together constitute the initial state of reinforcement learning. A large-scale policy network is applied to the initial state of reinforcement learning. The network uses the deep deterministic policy gradient (DDPG) algorithm to output coarse-grained adjustment actions and generate a large-scale adjustment policy. The DDPG algorithm determines the action through a deterministic policy network, and combines the optimization method of policy gradient, so that the large-scale policy network can effectively process the continuous action space and generate a coarse-grained adjustment scheme suitable for the current state. Based on the generated large-scale adjustment strategy, the bed performs the corresponding coarse adjustment action and collects the adjusted environmental feedback information in real time. The feedback information includes changes in the patient's posture, adjustments to environmental parameters, etc. These information and the executed actions together constitute large-scale experience tuples, and these tuples are stored in the priority experience replay buffer. In order to optimize the large-scale policy network and its corresponding value network, the experience batches that are most valuable to model learning are sampled from the priority experience replay buffer according to the priority of the temporal difference (TD) error. The parameters of the large-scale policy network and the value network are updated through these experiences, and the optimized large-scale network can better handle different coarse adjustment tasks. The large-scale adjustment strategy is input into the small-scale policy network as a constraint. The small-scale policy network uses the Soft Actor-Critic (SAC) algorithm to output fine adjustment actions and generate small-scale adjustment strategies. The SAC algorithm selects actions in a probabilistic manner by balancing exploration and utilization, and improves the stability and effectiveness of the strategy by minimizing the loss function. Based on the generated small-scale adjustment strategy, the bed performs more refined micro-adjustments to approach the target posture and collect environmental feedback information. These feedback information constitutes small-scale experience tuples and is stored in another priority experience replay buffer dedicated to small-scale adjustment. In order to optimize the small-scale policy network and its corresponding Q network, experience batches are sampled from the small-scale priority experience replay buffer. Through repeated training and updating of these experiences, the adaptability and accuracy of the network to fine-tuning tasks are enhanced. The updated large-scale network and small-scale network are combined to construct a hierarchical adjustment strategy tree. The strategy tree dynamically switches between large scale and small scale according to different adjustment requirements and outputs the final stratified bed adjustment strategy.

[0021] Step S6: Based on the stratified bed adjustment strategy, combined with model predictive control and adaptive PID controller, real-time bed adjustment is performed and online learning and updating are performed to output optimized bed control instructions.

[0022] Specifically, the hierarchical bed adjustment strategy is analyzed, and the large-scale adjustment target and small-scale fine-tuning parameters are extracted to obtain the hierarchical control target. The hierarchical control target specifically includes the adjustment requirements of the back, legs and bed height, as well as the corresponding fine-tuning parameters. The dynamic model of the bed is constructed based on the hierarchical control target. The model describes the motion characteristics of each adjustment mechanism of the bed, including the kinematic equations of the back, legs and height adjustment mechanisms, and obtains the state space representation of the bed. The state space model accurately reflects the physical parameters and constraints of the bed during the adjustment process. Using the state space representation of the bed, a model predictive controller is designed, in which the prediction time domain is defined as 10 steps and the control time domain is defined as 3 steps. The model predictive controller predicts the system behavior for several steps in the future, and solves the quadratic programming problem based on the deviation between the current state and the future target to obtain the optimal control sequence. The control sequence provides the specific operation steps that the bed should perform during the adjustment process to minimize the deviation and achieve the optimal control effect. An adaptive PID controller is applied to the first control quantity in the optimal control sequence. The adaptive PID controller automatically adjusts the PID parameters according to the current control error to improve the response speed and accuracy of the control and generate the initial execution instructions. The initial execution instructions are input into the actuator of the bed to prompt the bed to make corresponding adjustments. At the same time, the actual adjustment effect and environmental feedback are collected to construct an online learning data set. Based on the online learning data set, the dynamic model parameters of the bed are updated using the incremental learning algorithm to obtain the revised state space representation. The incremental learning algorithm can effectively integrate new and old data and dynamically update the physical parameters and model structure of the bed without completely retraining the model. Based on the revised state space representation, the parameters of the model predictive controller and the adaptive PID controller are recalculated to obtain the updated controller. The updated controller is applied in the next bed adjustment to ensure the accuracy and response speed of the adjustment. The updated controller generates optimized bed control instructions to guide the bed to perform the next adjustment operation.

[0023] In the embodiment of the present invention, by performing noise reduction, normalization and time series fusion on the data collected by multiple sensors, and combining principal component analysis for dimensionality reduction, the key features of the patient's state are effectively extracted, data redundancy is reduced, and the efficiency of subsequent processing is improved. Multi-branch convolutional neural network and adversarial training are used to achieve domain adaptation, and feature representation with strong generalization ability is obtained, so that the system can adapt to different patient groups and application environments. The long short-term memory network model with attention mechanism can better capture the time dependence of the patient's state and improve the prediction accuracy of the optimal posture demand. Through the dual-time scale reinforcement learning network, a hierarchical control strategy combining coarse adjustment and fine adjustment is realized, which not only ensures the accuracy of large-scale adjustment, but also realizes fine adjustment of small scale. By giving priority to the experience playback buffer, the utilization rate of important samples is improved, the convergence process of reinforcement learning is accelerated, and the learning efficiency is improved. Combining model predictive control with adaptive PID controller not only takes into account the prediction of future state, but also can adjust parameters according to real-time feedback, thereby improving the accuracy and robustness of control. The bed dynamics model is updated in real time through the incremental learning algorithm, so that the system can continuously adapt to the changes in patient status and environmental interference, and maintain long-term control effect.

[0024] In a specific embodiment, the process of executing step S1 may specifically include the following steps: Perform wavelet transform on the weight distribution data collected by the pressure sensor to obtain the pressure data after noise reduction; Performing wavelet transform on the tilt angle data collected by the acceleration sensor to obtain the acceleration data after noise reduction; Perform Gaussian filtering on the three-dimensional posture image collected by the depth camera to obtain smoothed image data; Perform moving average filtering on the environmental parameter data collected by the temperature and humidity sensor and the air quality sensor, with the window size set to 5, to obtain smoothed environmental data; Based on the maximum-minimum normalization method, the denoised pressure data, denoised acceleration data, smoothed image data and smoothed environmental data are normalized to obtain standardized multimodal data.

[0025] Specifically, the weight distribution data collected by the pressure sensor is subjected to wavelet transform. Wavelet transform is an effective noise reduction method that decomposes the signal into different scales and effectively removes noise while retaining important features. Suppose the original pressure data is , decompose it into components of different frequencies through wavelet transform, where high-frequency components usually contain noise and low-frequency components contain the main features of the signal. Assuming the wavelet basis function is used , then the wavelet transform can be expressed as: ; in, are the wavelet coefficients, and Represent the scale and position parameters respectively. By selecting the appropriate wavelet scale and threshold, the frequency component is retained and the high-frequency noise is removed to obtain the denoised pressure data. The tilt angle data collected by the acceleration sensor is subjected to wavelet transform. The tilt angle data of the acceleration sensor is used to monitor the tilt state of the bed. Through wavelet transform, the original tilt angle data is converted into Decompose into frequency components of different scales and remove noise components to obtain smoother and more accurate acceleration data. Gaussian filtering is used to smooth the 3D posture image collected by the depth camera. Gaussian filtering is a commonly used image processing technology. By convolving the image data with the Gaussian kernel, high-frequency noise in the image can be effectively removed while retaining the edge and contour features of the image. Assume that the depth image data is a matrix , the Gaussian kernel function is , then the image data after smoothing It can be expressed as: ; in, is the Gaussian kernel function, is the standard deviation, controlling the degree of smoothing, is the kernel size. After Gaussian filtering, the fine noise in the image is effectively suppressed, and the smoothed image data is obtained. At the same time, for the environmental parameter data collected by the temperature and humidity sensor and the air quality sensor, the moving average filter is used for smoothing, and the short-term fluctuations are eliminated by averaging the values ​​in the data window. Suppose the environmental parameter data is the sequence , the window size is 5, then the smoothed environmental data It can be expressed as: ; The average value of each time point is calculated by sliding the window, which effectively removes the random noise in the environmental data and retains the overall trend of the data. To normalize the data from different sources, the maximum-minimum normalization method is used to normalize the denoised pressure data, acceleration data, smoothed image data, and smoothed environmental data, and convert them to the same numerical range. Assume that the denoised data of a sensor is , whose maximum value is , the minimum value is , then the normalized data It is expressed as: ; In this way, data of different modalities are unified into the range of [0, 1] to obtain standardized multimodal data.

[0026] In a specific embodiment, the process of executing step S2 may specifically include the following steps: Constructing a state vector for the standardized multimodal data, including patient position, posture angle, and environmental parameters, to obtain an initial state representation; Based on the accuracy characteristics of each sensor, the observation noise covariance matrix is ​​set to obtain the initial parameters of the Kalman filter, the Kalman filter algorithm is applied to the initial state representation, and the prediction step and the update step are performed to obtain the fused state estimate; Based on the fused state estimation, timestamp alignment is performed to synchronize all sensor data to a unified time base with a time resolution of 1 millisecond to obtain time-synchronized multi-modal data. Calculate the covariance matrix for time-synchronized multimodal data and perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors; Sort the eigenvectors based on the eigenvalue size, select the first K eigenvectors with a cumulative contribution rate of 95%, and construct the projection matrix; The time-synchronized multimodal data is linearly transformed through the projection matrix to obtain the feature data after dimensionality reduction, and the feature data after dimensionality reduction is standardized to obtain low-dimensional fusion features.

[0027] Specifically, a state vector is constructed for the standardized multimodal data. The state vector usually contains key information such as the patient's position, posture angle, and environmental parameters. Assume that the patient's position is given by , the attitude angle is given by , the environmental parameters are given by Representation, then the initial state representation can be constructed as a vector containing these variables: ; According to the observation noise characteristics of each sensor, the observation noise covariance matrix is ​​set , which reflects the measurement accuracy of different sensors. For example, if the observation noise variances of patient position, posture angle, and environmental parameters are , and , then the observation noise covariance matrix can be expressed as: ; The initial parameters of the Kalman filter are obtained according to the covariance matrix. Kalman filtering is a recursive algorithm that optimizes the estimation of the state of the system by combining the prediction model and the observation data. Its core lies in two steps: prediction and update. The prediction step predicts the current state based on the prior state and the system model, while the update step corrects the prediction result according to the new observation value. Assume that the state transition equation of the state vector is: ; in, is the state transition matrix, is the control input matrix, is the control input, is process noise. The prediction step predicts the state of the next moment through this equation, while the update step is based on the observed value By continuously iterating these two steps, the Kalman filter finally outputs the fused state estimate: ; in, For the predicted state, is the observation matrix, is the Kalman gain, defined as: ; Through the Kalman filter algorithm, the observation data from multiple sensors can be fused to obtain a more accurate state estimate. Based on the fused state estimate, timestamp alignment is performed, and a time resolution of 1 millisecond is selected as a unified benchmark. Through methods such as linear interpolation, sensor data with different timestamps are aligned to a unified time axis to obtain time-synchronized multimodal data. The covariance matrix is ​​calculated for the time-synchronized multimodal data. The covariance matrix reflects the linear relationship between different data features. Assume that the time-synchronized multimodal data is , then the covariance matrix It can be expressed as: ; in, is the number of data samples, is the inner product matrix of the data. Perform eigenvalue decomposition on the covariance matrix to obtain a set of eigenvalues ​​and corresponding eigenvectors. The eigenvalue represents the importance of the corresponding eigenvector in the data. In order to achieve dimensionality reduction, the eigenvectors are sorted based on the size of the eigenvalue, and the top eigenvectors with a cumulative contribution rate of 95% are selected. The cumulative contribution rate refers to the previous The ratio of the eigenvalues ​​corresponding to the eigenvectors to the total eigenvalues. After selecting the eigenvectors, construct a projection matrix

[0028] ; in, is the selected eigenvector. Through the projection matrix, the time-synchronized multimodal data is linearly transformed to reduce the data to a low-dimensional space: ; The feature data after dimensionality reduction is standardized to scale the different features of the data to the same numerical range to eliminate the impact of dimensional differences on subsequent processing. Standardization usually uses the maximum-minimum normalization method to scale the data to the interval [0, 1] to obtain low-dimensional fusion features.

[0029] In a specific embodiment, the process of executing step S3 may specifically include the following steps: Segment the low-dimensional fusion features to obtain pressure data features, acceleration data features and image data features; The pressure data features are input into the pressure branch network in the multi-branch convolutional neural network. The pressure branch network contains three one-dimensional convolutional layers. Each layer uses 32 3×1 convolution kernels with a step size of 1. After LeakyReLU activation function and batch normalization processing, the pressure feature vector is finally obtained through global maximum pooling. The acceleration data features are input into the acceleration branch network in the multi-branch convolutional neural network. The acceleration branch network consists of three one-dimensional convolutional layers. Each layer uses 64 3×1 convolution kernels with a step size of 2. The ELU activation function and Dropout regularization are used, followed by global average pooling to obtain the acceleration feature vector. The image data features are input into the image branch network in the multi-branch convolutional neural network. The image branch network contains 5 two-dimensional convolutional layers. The convolution kernel sizes are 7×7, 5×5, 3×3, 3×3, and 3×3, respectively. The number of channels is 64, 128, 256, 512, and 512, respectively. Each layer is followed by a ReLU activation function and a maximum pooling layer. Finally, the image feature vector is obtained through adaptive average pooling. The pressure feature vector, acceleration feature vector and image feature vector are concatenated and fused through a fully connected layer to obtain the initial fusion feature; Construct a domain adversarial network, including a feature extractor and a domain classifier. The feature extractor uses a 3-layer fully connected network with 256 neurons and a ReLU activation function in each layer. The domain classifier uses a 2-layer fully connected network with 128 neurons and a LeakyReLU activation function in each layer. The initial fused features are input into the feature extractor to obtain the intermediate feature representation, and the intermediate feature representation is input into the domain classifier. The feature extractor parameters are updated through the gradient reversal layer to perform adversarial training. The attention mechanism is applied to the intermediate feature representation, the feature weights are calculated, and the intermediate feature representations are weighted combined according to the feature weights to obtain the generalized feature representation.

[0030] Specifically, the low-dimensional fusion features are segmented to obtain data features of different modalities, including pressure data features, acceleration data features, and image data features. The data features of each modality are input into the corresponding multi-branch convolutional neural network to extract its unique feature representation. The pressure data features are input into the pressure branch network in the multi-branch convolutional neural network. The pressure branch network contains 3 one-dimensional convolutional layers. Each convolutional layer uses 32 convolution kernels of size 3×1, and the step size is set to 1. The LeakyReLU activation function is applied to each convolutional layer. LeakyReLU is a commonly used nonlinear activation function. By allowing a small part of negative values ​​to pass, it alleviates the "neuron death" problem that may occur in ReLU when facing negative inputs. It can be expressed in formula as: ; in, is the input, is a very small positive number, usually set to 0.01. After activation by LeakyReLU, the output of the convolutional layer is batch normalized. Batch normalization helps to accelerate the training process of the network and reduce internal covariate shift. The output of the convolutional layer is compressed into a feature vector of fixed length through global maximum pooling to form a pressure feature vector. The acceleration data features are input into the acceleration branch network. The network consists of 3 one-dimensional convolutional layers, each layer uses 64 convolution kernels of size 3×1 and a step size of 2. The convolution layer of the acceleration branch network is followed by the ELU activation function. ELU is smoother than ReLU and gradually tends to a negative value when negative input is applied, making the network more robust. The formula of ELU is expressed as: ; in, is the input, Usually 1. To prevent overfitting, the acceleration branch network applies Dropout regularization after the output of the convolution layer. By randomly setting 50% of the neuron outputs to 0, Dropout effectively improves the generalization ability of the network. The output after convolution is compressed into an acceleration feature vector through global average pooling. For image data features, they are input into the image branch network for processing. The network contains 5 two-dimensional convolutional layers. The convolution kernel sizes of each convolution layer are 7×7, 5×5, 3×3, 3×3, and 3×3, respectively, and the number of channels are 64, 128, 256, 512, and 512, respectively. After each layer of convolution operation, the ReLU activation function is applied. ReLU is a widely used activation function. Its output is the non-negative part of the input, defined as: ; ReLU can introduce nonlinearity and has high computational efficiency. The maximum pooling layer in each convolutional layer reduces the size of the feature map, retains important features, and reduces the amount of computation. The feature map is compressed into a fixed-size image feature vector through adaptive average pooling, which can automatically adapt to the size of the input to ensure the consistency of the output size. The pressure feature vector, acceleration feature vector, and image feature vector extracted from the three branch networks are spliced ​​to obtain a complete feature representation. In order to extract high-level feature information, these spliced ​​feature vectors are input into a fully connected layer for fusion. The fully connected layer linearly combines all input feature information and implements nonlinear mapping through activation functions to obtain the initial fused features. Based on the initial fused features, a domain adversarial network is constructed to enhance the generalization ability of the features. The domain adversarial network consists of two main parts: a feature extractor and a domain classifier. The feature extractor consists of a 3-layer fully connected network, each layer contains 256 neurons, and uses the ReLU activation function. The domain classifier consists of a 2-layer fully connected network, each layer contains 128 neurons, and uses the LeakyReLU activation function. The initial fused features are input into the feature extractor to extract the intermediate feature representation. At the same time, the intermediate feature representation is input into the domain classifier, whose goal is to distinguish which domain these features come from (such as different patients or environmental conditions). In this way, the feature extractor is forced to extract more generalized feature representations, making it difficult for the domain classifier to accurately classify. In domain adversarial training, the parameters of the feature extractor are updated through the gradient reversal layer. The gradient reversal layer plays the role of transmitting signals during forward propagation, and during backpropagation, it multiplies the gradient by a negative value, thereby reversing the gradient direction, so that the learning direction of the feature extractor is opposite to the goal of the domain classifier. Through adversarial training, the feature extractor gradually learns feature representations that can be generalized in different domains. An attention mechanism is applied to the intermediate feature representation. The attention mechanism obtains feature weights by calculating the query, key, and value matrices. The attention mechanism can assign different weights according to the importance of the features. These weights are applied to each feature component of the intermediate feature representation, and a weighted combination is performed to obtain the final generalized feature representation.

[0031] In a specific embodiment, the execution step inputs the image data features into the image branch network in the multi-branch convolutional neural network, the image branch network includes 5 two-dimensional convolutional layers, the convolution kernel sizes are 7×7, 5×5, 3×3, 3×3, 3×3, the number of channels are 64, 128, 256, 512, 512, each layer is followed by a ReLU activation function and a maximum pooling layer, and finally, the image feature vector is obtained by adaptive average pooling. The process can specifically include the following steps: Zero padding is performed on the input image data features with a padding size of 3 to obtain the padded image data; Apply the first convolution layer to the padded image data, using 64 7×7 convolution kernels with a step size of 2 to obtain the first layer feature map, and perform ReLU activation and 2×2 maximum pooling on the first layer feature map to obtain the first layer output feature map; The output feature map of the first layer is input into the second convolution layer, and 128 5×5 convolution kernels with a step size of 1 are used to obtain the feature map of the second layer. The feature map of the second layer is then ReLU activated and 2×2 maximum pooled to obtain the output feature map of the second layer. The output feature map of the second layer is input into the third convolution layer, and 256 3×3 convolution kernels with a step size of 1 are used to obtain the feature map of the third layer. The feature map of the third layer is then activated by ReLU and 2×2 maximum pooling to obtain the output feature map of the third layer. The output feature map of the third layer is input into the fourth convolution layer, 512 3×3 convolution kernels are used with a step size of 1 to obtain the feature map of the fourth layer, and the feature map of the fourth layer is subjected to ReLU activation and 2×2 maximum pooling to obtain the output feature map of the fourth layer; The output feature map of the fourth layer is input into the fifth convolution layer, 512 3×3 convolution kernels are used with a step size of 1 to obtain the feature map of the fifth layer, and the feature map of the fifth layer is activated by ReLU and 2×2 maximum pooling to obtain the output feature map of the fifth layer; Adaptive average pooling is performed on the output feature map of the fifth layer, and the feature map size is adjusted to a fixed size H×W to obtain a fixed-size feature map. The fixed-size feature map is flattened to convert the two-dimensional feature map into a one-dimensional vector to obtain an image feature vector.

[0032] Specifically, the input image data features are zero-filled. By adding zero-valued pixels at the edge of the image, the convolution operation retains more edge information without reducing the size of the feature map. Assume that the original image data is a matrix , whose size is , perform zero padding of size 3, add 3 rows and 3 columns of zero values ​​on each boundary of the image, and obtain the padded image data matrix The image size after padding will become The padded image data is fed into the first convolutional layer. The convolutional layer uses 64 convolution kernels of size 7×7 and a step size of 2. The convolution operation extracts local spatial information by sliding the convolution kernel over the image and calculating the local weighted sum. Assume that the convolution kernel is a matrix , whose size is 7×7, the output feature map of the convolution operation can be expressed as: ; in, and Represents the position index of the output feature map, and is the index of the convolution kernel. Since the step size is 2, the size of the feature map after the convolution operation will be reduced to half of the original size. After the convolution operation, the ReLU activation function is applied to the first layer feature map. The formula of ReLU is: ; The ReLU function introduces nonlinearity and effectively suppresses the influence of negative values, improving the network's expressiveness. The output feature map of the first layer is further processed by the maximum pooling layer. The maximum pooling operation reduces the size of the feature map and retains the most significant features by selecting the maximum value in each 2×2 area. The output feature map of the first layer is input to the second convolution layer. This convolution layer uses 128 convolution kernels of size 5×5 and the step size is set to 1. More complex feature information is extracted by calculating the weighted sum of the local area. The activation function still uses ReLU, and then the pooling operation is performed through the 2×2 maximum pooling layer to reduce the size of the feature map and enhance the sparsity of the features. The feature map output by the second layer is input to the third convolution layer. This convolution layer uses 256 convolution kernels of size 3×3 and the step size is set to 1. The convolution kernel size is small and the step size is also small, which means that this convolution layer can capture more subtle local features. The feature map after convolution is nonlinearly mapped by the ReLU activation function and processed again by the 2×2 maximum pooling layer to obtain the output feature map of the third layer. The output feature map of the third layer is input to the fourth convolution layer. This convolution layer uses 512 convolution kernels of size 3×3, and the step size is still set to 1. Through the convolution operation, higher-level abstract features are extracted. The ReLU activation function is still used to introduce nonlinearity, and 2×2 maximum pooling is applied to the activated feature map, so that the feature map output by the fourth layer can further reduce the size while retaining the main information. The output feature map of the fourth layer is input into the fifth convolution layer. The fifth convolution layer uses 512 convolution kernels of size 3×3 and a step size of 1. Through this convolution layer, the network can extract the deepest features of the image. After these features are processed by the ReLU activation function, the fifth layer output feature map is obtained through a 2×2 maximum pooling layer. An adaptive average pooling operation is applied to the output feature map of the fifth layer to adjust the feature map to a fixed size. , adapting to input feature maps of different sizes, so that the network can maintain output consistency when processing images of different resolutions. Flatten the fixed-size feature map after adaptive average pooling, and convert the two-dimensional feature map into a one-dimensional vector so that the data can be input into the fully connected layer or other subsequent processing units. Assume that the flattened feature vector is , then its length is ,in is the number of channels.

[0033] In a specific embodiment, the process of executing step S4 may specifically include the following steps: The generalized feature representation is split into time series to obtain an input sequence of fixed length T; The input sequence is input into the LSTM layer in the long short-term memory network model with attention mechanism. The LSTM layer contains 128 hidden units. The output sequence of the LSTM layer is obtained through the gating mechanism and internal state update. Apply the self-attention mechanism to the LSTM layer output sequence, calculate the query, key and value matrices, use the scaled dot product attention algorithm to get the attention weights, and multiply the attention weights with the value matrix to get the weighted feature representation; The weighted feature representation is residually connected with the original LSTM layer output sequence, and layer normalization is performed to obtain a fused feature sequence; The fused feature sequence is input into a feedforward neural network, which contains two fully connected layers. The first layer uses 256 neurons and a ReLU activation function, and the second layer uses 128 neurons and a tanh activation function to obtain a converted feature sequence. The converted feature sequence is connected to the fused feature sequence for the second time with residual connection, and layer normalization is performed again to obtain the final feature sequence; Apply the temporal attention mechanism to the final feature sequence, calculate the importance weight of each time step, and perform weighted averaging on the final feature sequence according to the importance weight to obtain the context vector; The context vector is input into the last fully connected layer, and the softmax activation function is used to output the probability distribution of multiple posture categories, and the category with the highest probability is selected as the best posture requirement prediction result.

[0034] Specifically, the generalized feature representation is segmented into time series. Assuming there is a generalized feature sequence of length N, this sequence can be represented as , where each Represents a feature vector at a moment. The sequence is divided into multiple input sequences of fixed length T. The sequence form is ; Each input sequence Contains T consecutive feature vectors. Each input sequence is input into the LSTM layer in the long short-term memory network model with attention mechanism. The LSTM layer contains 128 hidden units. Through the gating mechanism (including forget gate, input gate and output gate) and internal state update, it can effectively process data with strong time series dependence. The LSTM layer can retain the information of past time steps when processing the data of each time step, solving the gradient vanishing problem of traditional RNN when processing long sequences. The hidden state update formula of LSTM is: ; in, is the activation value of the output gate, is the cell state, and © represents element-by-element multiplication. After processing by the LSTM layer, an output sequence is obtained: ; The output sequence retains the timing information in the input sequence. Apply the self-attention mechanism to the output sequence of the LSTM layer. The attention mechanism can adaptively assign different weights according to the importance of different parts of the input data. Calculate the query, key, and value matrices: ; in, and is a trainable weight matrix. We use the scaled dot product attention algorithm to calculate the attention weights and apply these weights to the value matrix to obtain the weighted feature representation: ; in, is the dimension of the key matrix, and the softmax function is used to normalize the dot product result into a probability distribution. In this way, the model can adaptively focus on the most important part of the input sequence and generate a weighted feature representation. The weighted feature representation is residually connected to the original LSTM layer output sequence. The residual connection can retain the original feature information while integrating the features enhanced by the attention mechanism to obtain a richer feature representation. The formula for the residual connection is ; The fused feature sequence Perform layer normalization, which can stabilize the training process and accelerate convergence, helping to avoid gradient vanishing or exploding. Input the fused feature sequence into the feedforward neural network. The feedforward neural network contains two fully connected layers. The first layer uses 256 neurons and performs nonlinear transformation through the ReLU activation function: ; The second layer uses 128 neurons and is processed by the tanh activation function: ; Through the processing of these two layers, the converted feature sequence is obtained. The converted feature sequence is residually connected with the fused feature sequence again, and the layer normalization is performed again to obtain the final feature sequence. The temporal attention mechanism is applied to the final feature sequence. The temporal attention mechanism can assign different weights according to the importance of different time steps, highlighting the most important time points for the task. By weighted averaging the importance weights of each time step, the context vector is obtained: ; in, is the time step The importance weight of is the context vector. The context vector is input to the last fully connected layer, and the softmax activation function is used to output the probability distribution of multiple posture categories. The formula of the softmax function is: ; in, It is The scores of the categories, is the total number of categories. The softmax activation function converts the scores of each category into a probability distribution, and finally selects the category with the highest probability as the prediction result of the best posture requirement.

[0035] In a specific embodiment, the execution step inputs the fused feature sequence into a feedforward neural network, the feedforward neural network includes two fully connected layers, the first layer uses 256 neurons and a ReLU activation function, and the second layer uses 128 neurons and a tanh activation function. The process of obtaining the converted feature sequence can specifically include the following steps: Flatten the fused feature sequence, convert the three-dimensional tensor into a two-dimensional matrix, and obtain a flattened feature matrix; The flattened feature matrix is ​​input into the first fully connected layer, and the weight matrix W1 and the bias vector b1 are used for linear transformation to obtain the intermediate feature vector, and the ReLU activation function operation is performed on the intermediate feature vector, and the max(0, x) operation is performed on each element to obtain the first layer activation output; Perform Dropout regularization on the activation output of the first layer, randomly set 50% of the neuron outputs to 0, and obtain the regularized output of the first layer; The first layer regularized output is input into the second fully connected layer, and the weight matrix W2 and the bias vector b2 are used for linear transformation to obtain the second layer feature vector, and the tanh activation function is performed on the second layer feature vector to map each element to the interval [-1, 1] to obtain the second layer activation output; A batch normalization operation is performed on the second layer activation output to calculate the mean and variance, and to perform scaling and translation to obtain a normalized feature vector, and the normalized feature vector is reshaped to the same time step as the input fused feature sequence to obtain a converted feature sequence.

[0036] Specifically, the fused feature sequence is flattened. Assume that the fused feature sequence is a three-dimensional tensor, represented by F, with a dimension of ,in represents the time step, represents the number of features, Indicates the number of channels. Convert the three-dimensional data to two-dimensional data so that it can be input into the fully connected layer for processing. The flattened two-dimensional matrix can be expressed as a flattened feature matrix: ; This operation converts the last two dimensions of the original 3D tensor and Flatten to a new dimension so that the dimension of the flattened feature matrix is The flattened matrix retains the information of the time dimension, but combines the feature and channel information to form a large feature vector. Input to the first fully connected layer. The fully connected layer passes through the weight matrix and the bias vector Perform a linear transformation on the input to obtain the intermediate feature vector: ; in, The dimension is , is the output dimension of the first fully connected layer, usually expressed as the number of neurons, and the bias vector The dimension is , used to translate the output of the linear transformation and enhance the expressiveness of the model. The intermediate feature vector after the linear transformation Nonlinear mapping is performed through the ReLU activation function. ReLU is a commonly used activation function, and its formula is: ; The ReLU activation function sets all negative values ​​to 0 while leaving positive values ​​unchanged. This operation can introduce nonlinearity and prevent the gradient vanishing problem without increasing computational complexity. After ReLU activation, the first layer activation output is obtained: ; In order to prevent the model from overfitting, the first layer activation output is regularized by Dropout. Dropout is a technique that randomly sets the output of some neurons to 0 during training to prevent the neural network from relying on specific neurons and enhance the generalization ability of the model. Assuming the dropout rate of Dropout is 50%, the output after Dropout processing is: ; After Dropout regularization, the regularized output of the first layer is input to the second fully connected layer. The second fully connected layer is similar to the first layer, through the weight matrix and the bias vector Perform a linear transformation to obtain the second-layer feature vector: ; in, The dimension is , is the output dimension of the second fully connected layer. The second layer feature vector is nonlinearly mapped through the tanh activation function. The formula of the tanh function is: ; The tanh function maps the input to the interval [-1, 1], which can centralize the data and improve the stability of the model. After tanh activation, the second layer activation output is obtained: ; In order to standardize the output and speed up the training process of the network, batch normalization is performed on the second layer activation output. Batch normalization scales and translates the output by calculating the mean and variance of small batches of data to make it more numerically stable. For each dimension of activation output, the batch normalization formula is: ; in, and are the mean and standard deviation of the small batch data, and are trainable scaling and translation parameters. The batch normalized output is more suitable for further processing of deep learning models, avoiding the problem of gradient vanishing or exploding. The normalized feature vector is reshaped to the same time step as the input fusion feature sequence to obtain the final transformed feature sequence. The reshaping operation is similar to the inverse operation of flattening, which rearranges the one-dimensional vector into a multi-dimensional matrix form, making the data consistent in the time dimension.

[0037] In a specific embodiment, the process of executing step S5 may specifically include the following steps: Based on the prediction result of the optimal posture requirement, a state vector is constructed, which includes the current posture, the target posture and the environmental parameters, and the initial state of reinforcement learning is obtained; Apply a large-time-scale policy network to the initial state of reinforcement learning, use a deep deterministic policy gradient algorithm, output coarse-grained regulation actions, and obtain a large-scale regulation policy; Based on the large-scale adjustment strategy, the beds are roughly adjusted, and environmental feedback is collected to construct large-scale experience tuples, which are stored in the priority experience playback buffer. Sample a batch of experiences from the priority experience playback buffer according to the TD error priority, update the parameters of the large-time-scale strategy network and value network, and obtain the updated large-scale network; The large-scale adjustment strategy is used as a constraint condition, input into the small-time-scale strategy network, and the Soft Actor-Critic algorithm is used to output fine-tuning actions to obtain the small-scale adjustment strategy. Based on the small-scale adjustment strategy, the bed is fine-tuned and environmental feedback is collected to construct a small-scale experience tuple, which is then stored in another priority experience playback buffer. Sample experience batches from the small-scale priority experience replay buffer, update the parameters of the small-time-scale strategy network and the Q network, and obtain the updated small-scale network; The updated large-scale network and small-scale network are combined to construct a hierarchical regulation strategy tree, and the hierarchical bed regulation strategy is output.

[0038] Specifically, based on the predicted optimal posture requirement, a state vector is constructed. Assume that the current patient posture is represented by the vector Represented by the target pose, the target pose is represented by the vector Indicates that environmental parameters such as temperature and humidity are represented by vector e, and this information is combined into a state vector: ; Initial state vector Covers the key information required for bed adjustment. Input the reinforcement learning initial state into the large time scale policy network. The large time scale policy network focuses on the coarse-grained adjustment of the bed and uses the deep deterministic policy gradient algorithm to generate adjustment actions. Deep deterministic policy gradient is a policy gradient-based reinforcement learning algorithm that can optimize in a continuous action space. The algorithm consists of a policy network and a value network, where the policy network generates actions and the value network estimates the value of actions in the current state. The output action of the policy network It can be expressed as: ; in, represents the output of the policy network, are the parameters of the policy network. By optimizing the policy network, coarse-grained adjustment actions are output, which are used for the initial adjustment of the bed. Based on the large-scale adjustment strategy, the bed performs coarse adjustments, such as adjusting the tilt angle or height of the bed. After the coarse adjustment action is performed, environmental feedback is collected, including changes in patient posture and changes in environmental parameters. This information is combined with the executed action and the initial state to construct a large-scale experience tuple, which is expressed as: ; in, represents the immediate reward after executing the action, Represents the new state after executing the action. This experience tuple is stored in the priority experience replay buffer for subsequent network updates. In order to optimize the large time scale policy network and value network, a batch of experiences are sampled from the priority experience replay buffer. The priority is calculated based on the temporal difference (TD) error. The experience with a larger TD error has a higher priority. The TD error is expressed as: ; in, and They are target value network and target strategy network. is the discount factor. By updating the parameters of the large-scale policy network and the value network, an updated large-scale network is obtained. These networks can more accurately generate coarse-grained actions suitable for bed adjustment. The updated large-scale adjustment strategy is used as a constraint and input into the small-scale policy network. The small-scale policy network uses the Soft Actor-Critic (SAC) algorithm to focus on finer adjustment actions. SAC is a reinforcement learning algorithm based on policy gradients. It combines the idea of ​​maximum entropy reinforcement learning and enhances exploration capabilities by maximizing the entropy of the strategy. The objective function of the SAC algorithm is: ; in, represents the entropy of the policy, is the weight coefficient, and by optimizing the objective function, a fine adjustment action suitable for the bed is generated. Based on the small-scale adjustment strategy, the bed performs micro-adjustments, such as fine-tuning the tilt angle or position of the bed to make the patient's posture closer to the target posture. After performing fine-tuning, environmental feedback is also collected, and a small-scale experience tuple is constructed, which is expressed as: ; in, is a fine-tuning action generated by a small time-scale strategy, It is the reward after execution. is the new state. This experience tuple is stored in another priority experience replay buffer. In order to optimize the small-time-scale policy network and the Q network, experience batches are sampled from the small-scale priority experience replay buffer. The network is updated with these experiences, and finally the optimized small-scale network is obtained. These networks can generate more precise adjustment actions, so that the bed adjustment is more in line with the needs of the patient. The updated large-scale network and the small-scale network are combined to construct a hierarchical adjustment strategy tree. The hierarchical adjustment strategy tree combines the advantages of coarse adjustment and fine adjustment, and can perform multi-level adjustments according to the real-time needs of the patient. Output the hierarchical bed adjustment strategy, so that the bed adjustment can better adapt to the patient's posture changes and environmental changes, and provide optimal comfort and support.

[0039] In a specific embodiment, the process of executing step S6 may specifically include the following steps: Analyze the stratified bed adjustment strategy, extract large-scale adjustment targets and small-scale fine-tuning parameters, and obtain stratified control targets; Based on the hierarchical control objectives, a dynamic model of the bed is constructed, including the kinematic equations of the back, legs, and height adjustment mechanisms, and the state space representation of the bed is obtained; The model predictive controller is designed using the state space representation of the bed, the prediction time domain is defined as 10 steps, the control time domain is defined as 3 steps, and the optimal control sequence is obtained by solving the quadratic programming problem; Apply an adaptive PID controller to the first control quantity of the optimal control sequence, automatically adjust the PID parameters according to the control error, and obtain the initial execution instruction; Input the initial execution instruction to the bed actuator, collect the actual adjustment effect and environmental feedback, build an online learning data set, and use the incremental learning algorithm to update the bed dynamics model parameters based on the online learning data set to obtain the corrected state space representation; The parameters of the model predictive controller and the adaptive PID controller are recalculated based on the revised state space representation to obtain an updated controller, which is then applied to the next bed adjustment to output an optimized bed control instruction.

[0040] Specifically, large-scale adjustment targets and small-scale fine-tuning parameters are extracted from the hierarchical bed adjustment strategy to determine the hierarchical control targets. Hierarchical control targets usually include the overall posture adjustment targets of the bed (such as the angle adjustment of the back and legs) and more detailed fine-tuning parameters (such as precise angle change values, speed control parameters, etc.). With this information, a complete bed dynamics model is constructed. Based on the extracted hierarchical control targets, a bed dynamics model is constructed. The model includes kinematic equations that describe the various adjustable parts of the bed (such as the back, legs, and height adjustment mechanisms). Assume that the back angle of the bed is , the leg angle is , the height is , the changing relationship of these variables can be described by the kinetic equation. The kinetic model can be expressed as: ; in, , and are the kinematic equations for back, leg, and height adjustment, respectively. , and is the control input. Through these equations, we get the state space representation of the bed, and define the state vector as: ; The state space equation can be expressed as: ; in, Describes the state changes of the system. is the control input vector. A model predictive controller is designed using the state space representation of the hospital bed. The model predictive controller is a model-based control method that optimizes control decisions by predicting system behavior over a period of time in the future. The model predictive controller generates a series of optimal control actions by solving an optimization problem at each moment, so that the system reaches the desired state in the future prediction domain. For the hospital bed control system, the prediction domain is defined as 10 steps and the control domain is defined as 3 steps. The prediction domain means that the system will predict the state changes in the next 10 time steps, while the control domain means that the system only performs the initial 3 control actions in these 10 steps. In the model predictive controller framework, the weighted sum of state deviation and control energy consumption is minimized. The optimization problem can be formalized as a quadratic programming problem: ; in, Indicates The predicted state at the moment, is the target state, is the weight matrix of state bias, is the weight matrix controlling the input, is the control input. By solving the optimization problem, a series of optimal control sequences are obtained. The first control quantity is extracted from the optimal control sequence and input into the adaptive PID controller for refinement. The adaptive PID controller responds to changes in the dynamic environment by adjusting the PID parameters in real time. Its control law is: ; in, is the control input, is the control error, , , They are proportional, integral and differential gain coefficients respectively. The adaptive PID controller automatically adjusts these parameters according to the real-time measured control error to generate more accurate control instructions. After the initial execution instruction is input into the bed actuator, the actual adjustment effect and environmental feedback are collected in real time to construct an online learning data set. The data set contains the current state, the executed control quantity, the adjusted state and the corresponding environmental parameters. Using the online learning data set, the incremental learning algorithm is used to update the dynamic model parameters of the bed to obtain the revised state space representation. The incremental learning algorithm can gradually absorb new environmental and system information without losing old knowledge, making the model more accurate. The revised state space representation can reflect the dynamic characteristics of the bed in actual operation. Based on the revised state space model, the parameters of the model predictive controller and the adaptive PID controller are recalculated to ensure that the control system can always adapt to the actual needs of the patient. The updated controller is applied to the next bed adjustment process. The new model predictive controller will generate a more accurate optimal control sequence based on the revised dynamic model, while the adaptive PID controller will further optimize these control instructions, making the bed adjustment more stable and accurate.

[0041] The above describes the artificial intelligence-assisted adaptive bed control method according to the embodiment of the present invention. The following describes the artificial intelligence-assisted adaptive bed control system according to the embodiment of the present invention. Figure 2 In one embodiment of the present invention, an artificial intelligence-assisted adaptive bed control system includes: A preprocessing module is used to perform noise reduction and normalization preprocessing on the original physiological data collected by the multimodal sensor to obtain standardized multimodal data; The fusion module is used to perform time series fusion on standardized multimodal data and reduce the dimension through principal component analysis to obtain low-dimensional fusion features; The extraction module is used to input the low-dimensional fusion features into the multi-branch convolutional neural network for feature extraction, and achieve domain adaptation through adversarial training to obtain generalized feature representation; The prediction module is used to take the generalized feature representation as input, predict the patient's optimal posture requirement through a long short-term memory network model with an attention mechanism, and obtain the optimal posture requirement prediction result; A generation module is used to generate a hierarchical bed adjustment strategy based on the optimal posture demand prediction results through a dual-time scale reinforcement learning network with a priority experience playback mechanism; The output module is used to perform real-time bed adjustment and online learning updates based on the stratified bed adjustment strategy, combined with model predictive control and adaptive PID controller, and output optimized bed control instructions.

[0042] Through the collaboration of the above components, the key features of the patient's state are effectively extracted by denoising, normalizing and time series fusion of data collected by multiple sensors, and combining principal component analysis for dimensionality reduction, reducing data redundancy and improving the efficiency of subsequent processing. Multi-branch convolutional neural network and adversarial training are used to achieve domain adaptation, and feature representation with strong generalization ability is obtained, so that the system can adapt to different patient groups and application environments. The long short-term memory network model with attention mechanism can better capture the time dependence of the patient's state and improve the accuracy of the prediction of the optimal posture requirement. Through the dual-time scale reinforcement learning network, a hierarchical control strategy combining coarse adjustment and fine adjustment is realized, which not only ensures the accuracy of large-scale adjustment, but also realizes fine adjustment of small scale. By giving priority to the experience replay buffer, the utilization rate of important samples is improved, the convergence process of reinforcement learning is accelerated, and the learning efficiency is improved. Combining model predictive control with adaptive PID controller not only considers the prediction of future state, but also adjusts parameters according to real-time feedback, improving the accuracy and robustness of control. The bed dynamics model is updated in real time through an incremental learning algorithm, allowing the system to continuously adapt to changes in patient status and environmental interference, maintaining long-term control effects.

[0043] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the artificial intelligence-assisted adaptive bed control method in the above-mentioned embodiments.

[0044] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the artificial intelligence-assisted adaptive bed control method.

[0045] 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.

[0046] 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.

[0047] 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. An artificial intelligence-assisted adaptive bed control method, characterized in that: The method comprises: Perform noise reduction and normalization preprocessing on the original physiological data collected by the multimodal sensor to obtain standardized multimodal data; Performing time series fusion on the standardized multimodal data and performing dimensionality reduction through principal component analysis to obtain low-dimensional fusion features; Inputting the low-dimensional fusion features into a multi-branch convolutional neural network for feature extraction, and implementing domain adaptation through adversarial training to obtain a generalized feature representation; Taking the generalized feature representation as input, predicting the patient's optimal posture requirement through a long short-term memory network model with an attention mechanism, and obtaining an optimal posture requirement prediction result; Based on the optimal posture demand prediction results, a hierarchical bed adjustment strategy is generated through a dual-time scale reinforcement learning network including a priority experience playback mechanism; Based on the stratified bed adjustment strategy, combined with model predictive control and adaptive PID controller, real-time bed adjustment is performed and online learning and updating are performed to output optimized bed control instructions.

2. The artificial intelligence-assisted adaptive bed control method according to claim 1, characterized in that: The raw physiological data collected by the multimodal sensor is subjected to denoising and normalization preprocessing to obtain standardized multimodal data, including: Perform wavelet transform on the weight distribution data collected by the pressure sensor to obtain the pressure data after noise reduction; Performing wavelet transform on the tilt angle data collected by the acceleration sensor to obtain the acceleration data after noise reduction; Perform Gaussian filtering on the three-dimensional posture image collected by the depth camera to obtain smoothed image data; Perform moving average filtering on the environmental parameter data collected by the temperature and humidity sensor and the air quality sensor, with the window size set to 5, to obtain smoothed environmental data; Based on the maximum-minimum normalization method, the denoised pressure data, the denoised acceleration data, the smoothed image data and the smoothed environmental data are normalized to obtain standardized multimodal data.

3. The artificial intelligence-assisted adaptive bed control method according to claim 1, characterized in that: The time series fusion of the standardized multimodal data and the dimensionality reduction through principal component analysis to obtain low-dimensional fusion features include: Constructing a state vector for the standardized multimodal data, including patient position, posture angle, and environmental parameters, to obtain an initial state representation; Based on the accuracy characteristics of each sensor, an observation noise covariance matrix is ​​set to obtain initial parameters of a Kalman filter, a Kalman filter algorithm is applied to the initial state representation, a prediction step and an update step are performed, and a fused state estimate is obtained; Based on the fused state estimation, a timestamp alignment process is performed to synchronize all sensor data to a unified time reference with a time resolution of 1 millisecond to obtain time-synchronized multimodal data; Calculating a covariance matrix for the time-synchronized multimodal data, and performing eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors; The eigenvectors are sorted based on the eigenvalues, the first K eigenvectors whose cumulative contribution rate reaches 95% are selected, and a projection matrix is ​​constructed; The time-synchronized multimodal data is linearly transformed through the projection matrix to obtain feature data after dimensionality reduction, and the feature data after dimensionality reduction is standardized to obtain low-dimensional fusion features.

4. The artificial intelligence-assisted adaptive bed control method according to claim 1, characterized in that: The step of inputting the low-dimensional fusion features into a multi-branch convolutional neural network for feature extraction, and implementing domain adaptation through adversarial training to obtain a generalized feature representation includes: Segmenting the low-dimensional fusion features to obtain pressure data features, acceleration data features, and image data features; The pressure data features are input into the pressure branch network in the multi-branch convolutional neural network, wherein the pressure branch network comprises three one-dimensional convolutional layers, each layer uses 32 3×1 convolution kernels, the step size is 1, and the pressure feature vector is obtained by LeakyReLU activation function and batch normalization, and finally by global maximum pooling; The acceleration data features are input into the acceleration branch network in the multi-branch convolutional neural network, wherein the acceleration branch network is composed of three one-dimensional convolutional layers, each layer uses 64 3×1 convolution kernels, the step size is 2, the ELU activation function and Dropout regularization are adopted, and then the acceleration feature vector is obtained by global average pooling; Input the image data features into the image branch network in the multi-branch convolutional neural network, the image branch network includes 5 two-dimensional convolutional layers, the convolution kernel sizes are 7×7, 5×5, 3×3, 3×3, 3×3, the number of channels are 64, 128, 256, 512, 512, each layer is followed by a ReLU activation function and a maximum pooling layer, and finally an image feature vector is obtained through adaptive average pooling; splicing the pressure feature vector, the acceleration feature vector and the image feature vector, and fusing them through a fully connected layer to obtain an initial fused feature; Construct a domain adversarial network, including a feature extractor and a domain classifier, wherein the feature extractor adopts a 3-layer fully connected network, each layer uses 256 neurons and a ReLU activation function, and the domain classifier adopts a 2-layer fully connected network, each layer uses 128 neurons and a LeakyReLU activation function; Inputting the initial fused features into the feature extractor to obtain an intermediate feature representation, and inputting the intermediate feature representation into the domain classifier, updating the feature extractor parameters through a gradient reversal layer, and performing adversarial training; An attention mechanism is applied to the intermediate feature representation, feature weights are calculated, and weighted combinations are performed on the intermediate feature representations according to the feature weights to obtain a generalized feature representation.

5. The artificial intelligence-assisted adaptive bed control method according to claim 4, characterized in that: The image data features are input into the image branch network in the multi-branch convolutional neural network, the image branch network comprises 5 two-dimensional convolutional layers, the convolution kernel sizes are 7×7, 5×5, 3×3, 3×3, 3×3, the number of channels are 64, 128, 256, 512, 512, each layer is followed by a ReLU activation function and a maximum pooling layer, and finally an image feature vector is obtained by adaptive average pooling, including: Zero padding is performed on the input image data features with a padding size of 3 to obtain the padded image data; Applying a first convolution layer to the padded image data, using 64 7×7 convolution kernels with a step size of 2 to obtain a first layer feature map, and performing ReLU activation and 2×2 maximum pooling on the first layer feature map to obtain a first layer output feature map; The output feature map of the first layer is input into the second convolution layer, 128 5×5 convolution kernels are used with a step size of 1 to obtain the second layer feature map, and the second layer feature map is subjected to ReLU activation and 2×2 maximum pooling to obtain the second layer output feature map; The output feature map of the second layer is input into the third convolution layer, 256 3×3 convolution kernels are used with a step size of 1 to obtain the feature map of the third layer, and ReLU activation and 2×2 maximum pooling are performed on the feature map of the third layer to obtain the output feature map of the third layer; Input the third layer output feature map into the fourth convolution layer, use 512 3×3 convolution kernels with a step size of 1 to obtain a fourth layer feature map, and perform ReLU activation and 2×2 maximum pooling on the fourth layer feature map to obtain a fourth layer output feature map; Inputting the fourth layer output feature map into the fifth convolution layer, using 512 3×3 convolution kernels with a step size of 1 to obtain a fifth layer feature map, and performing ReLU activation and 2×2 maximum pooling on the fifth layer feature map to obtain a fifth layer output feature map; Adaptive average pooling is performed on the fifth layer output feature map, the feature map size is adjusted to a fixed size H×W, a fixed-size feature map is obtained, and a flattening operation is performed on the fixed-size feature map to convert the two-dimensional feature map into a one-dimensional vector to obtain an image feature vector.

6. The artificial intelligence-assisted adaptive bed control method according to claim 1, characterized in that: The generalized feature representation is used as input, and the optimal posture requirement of the patient is predicted through a long short-term memory network model with an attention mechanism to obtain the optimal posture requirement prediction result, including: Performing time series segmentation on the generalized feature representation to obtain an input sequence with a fixed length of T; Input the input sequence into the LSTM layer in the long short-term memory network model with an attention mechanism, the LSTM layer includes 128 hidden units, and obtain the LSTM layer output sequence through the gating mechanism and internal state update; Applying a self-attention mechanism to the LSTM layer output sequence, calculating query, key, and value matrices, obtaining attention weights using a scaled dot product attention algorithm, and multiplying the attention weights by the value matrix to obtain a weighted feature representation; Performing residual connection between the weighted feature representation and the original LSTM layer output sequence, and performing layer normalization processing to obtain a fused feature sequence; Inputting the fused feature sequence into a feedforward neural network, the feedforward neural network comprising two fully connected layers, the first layer using 256 neurons and a ReLU activation function, and the second layer using 128 neurons and a tanh activation function, to obtain a converted feature sequence; Performing a second residual connection on the converted feature sequence and the fused feature sequence, and performing layer normalization again to obtain a final feature sequence; Applying a temporal attention mechanism to the final feature sequence, calculating the importance weight of each time step, and performing weighted averaging on the final feature sequence according to the importance weight to obtain a context vector; The context vector is input into the last fully connected layer, and the softmax activation function is used to output the probability distribution of multiple posture categories, and the category with the highest probability is selected as the best posture requirement prediction result.

7. The artificial intelligence-assisted adaptive bed control method according to claim 6, characterized in that: The fused feature sequence is input into a feedforward neural network, the feedforward neural network comprises two fully connected layers, the first layer uses 256 neurons and a ReLU activation function, and the second layer uses 128 neurons and a tanh activation function, to obtain a converted feature sequence, including: Performing a flattening operation on the fused feature sequence to convert the three-dimensional tensor into a two-dimensional matrix to obtain a flattened feature matrix; Input the flattened feature matrix into the first fully connected layer, use the weight matrix W1 and the bias vector b1 to perform linear transformation to obtain an intermediate feature vector, perform a ReLU activation function operation on the intermediate feature vector, perform a max(0, x) operation on each element, and obtain the first layer activation output; Perform Dropout regularization on the activation output of the first layer, randomly set 50% of the neuron outputs to 0, and obtain the first layer regularized output; The first layer regularized output is input into the second fully connected layer, and a linear transformation is performed using the weight matrix W2 and the bias vector b2 to obtain a second layer feature vector, and a tanh activation function operation is performed on the second layer feature vector to map each element to the interval [-1, 1] to obtain a second layer activation output; A batch normalization operation is performed on the second layer activation output to calculate the mean and variance, and to perform scaling and translation to obtain a normalized feature vector, and the normalized feature vector is reshaped to the same time step as the input fused feature sequence to obtain a converted feature sequence.

8. The artificial intelligence-assisted adaptive bed control method according to claim 1, characterized in that: The method generates a hierarchical bed adjustment strategy for the optimal posture requirement prediction result through a dual-time scale reinforcement learning network including a priority experience playback mechanism, including: Based on the optimal posture requirement prediction result, a state vector is constructed, including the current posture, the target posture and the environmental parameters, to obtain the initial state of reinforcement learning; Applying a large time scale policy network to the reinforcement learning initial state, using a deep deterministic policy gradient algorithm, outputting a coarse-grained adjustment action, and obtaining a large-scale adjustment policy; Based on the large-scale adjustment strategy, the bed is roughly adjusted, and environmental feedback is collected to construct a large-scale experience tuple, which is stored in a priority experience playback buffer; Sampling a batch of experiences from the priority experience playback buffer according to the TD error priority, updating the parameters of the large-time-scale strategy network and the value network, and obtaining an updated large-scale network; The large-scale adjustment strategy is used as a constraint condition, input into a small-time-scale strategy network, and a Soft Actor-Critic algorithm is used to output a fine adjustment action to obtain a small-scale adjustment strategy; Based on the small-scale adjustment strategy, micro-adjust the bed, collect environmental feedback, construct a small-scale experience tuple, and store it in another priority experience playback buffer; Sampling experience batches from the small-scale priority experience replay buffer, updating parameters of the small-time-scale strategy network and the Q network, and obtaining an updated small-scale network; The updated large-scale network and small-scale network are combined to construct a hierarchical adjustment strategy tree, and a hierarchical bed adjustment strategy is output.

9. The artificial intelligence-assisted adaptive bed control method according to claim 1, characterized in that: Based on the stratified bed adjustment strategy, the model predictive control and the adaptive PID controller are combined to perform real-time bed adjustment and online learning update, and output optimized bed control instructions, including: Analyze the stratified bed adjustment strategy, extract large-scale adjustment targets and small-scale fine-tuning parameters, and obtain stratified control targets; Based on the hierarchical control objectives, a dynamic model of the hospital bed is constructed, including kinematic equations of the back, legs, and height adjustment mechanisms, to obtain a state space representation of the hospital bed; The model predictive controller is designed by using the state space representation of the hospital bed, the prediction time domain is defined as 10 steps, the control time domain is defined as 3 steps, and the optimal control sequence is obtained by solving the quadratic programming problem; Applying an adaptive PID controller to the first control quantity of the optimal control sequence, automatically adjusting PID parameters according to the control error, and obtaining an initial execution instruction; Inputting the initial execution instruction into the bed actuator, collecting actual adjustment effects and environmental feedback, constructing an online learning data set, and updating the bed dynamics model parameters using an incremental learning algorithm based on the online learning data set to obtain a revised state space representation; The parameters of the model predictive controller and the adaptive PID controller are recalculated based on the modified state space representation to obtain an updated controller, and the updated controller is applied to the next bed adjustment to output an optimized bed control instruction.

10. An artificial intelligence-assisted adaptive bed control system, characterized in that: A system for executing the artificial intelligence-assisted adaptive bed control method according to any one of claims 1 to 9, the system comprising: A preprocessing module is used to perform noise reduction and normalization preprocessing on the original physiological data collected by the multimodal sensor to obtain standardized multimodal data; A fusion module is used to perform time series fusion on the standardized multimodal data and perform dimensionality reduction through principal component analysis to obtain low-dimensional fusion features; An extraction module, used for inputting the low-dimensional fusion features into a multi-branch convolutional neural network for feature extraction, and realizing domain adaptation through adversarial training to obtain a generalized feature representation; A prediction module, used for taking the generalized feature representation as input, predicting the patient's optimal posture requirement through a long short-term memory network model with an attention mechanism, and obtaining an optimal posture requirement prediction result; A generation module, for generating a hierarchical bed adjustment strategy based on the optimal posture demand prediction result through a dual-time scale reinforcement learning network including a priority experience playback mechanism; The output module is used to perform real-time bed adjustment and online learning update based on the stratified bed adjustment strategy, combined with model predictive control and adaptive PID controller, and output optimized bed control instructions.

Images