A method and device for monitoring intensive care rehabilitation training

Abstract

The present invention discloses a monitoring method and device for intensive care rehabilitation training, which relates to the technical field of intensive care rehabilitation. It includes collecting the movement data of patients in real time, preliminarily analyzing the current movement state of patients, performing movement pattern recognition on patients, and outputting preliminary movement pattern labels; inputting the movement pattern labels into a generative adversarial network, training different stages of movement patterns and movement postures through a generator to generate virtual data with high similarity; initializing a movement data set with the virtual data, inputting the movement data set into a multi-level convolutional neural network architecture for feature extraction to generate the movement rehabilitation state of patients; uploading the training feedback data to a cloud platform for storage and analysis, generating a warning signal according to the analysis results of the cloud data, and automatically issuing a warning when the patient's training is abnormal. Through the multi-level deep learning architecture and virtual data generation method of the present invention, the accuracy of movement pattern recognition and the evaluation effect of the rehabilitation state can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of critical care rehabilitation, and particularly to a method and device for monitoring critical care rehabilitation training. Background Art

[0002] In recent years, with the development of rehabilitation medicine, especially in the exercise rehabilitation training of critically ill patients, the real-time monitoring and evaluation of patients' exercise states have gradually become a research hotspot. Traditional rehabilitation monitoring methods mostly rely on manual intervention or simple sensor devices, and cannot accurately reflect the subtle posture deviations and exercise load changes during the patients' exercise process, resulting in the difficulty of maximizing the rehabilitation effect. With the progress of intelligent sensors, wearable devices, and deep learning technologies, more and more intelligent motion monitoring methods have been proposed and began to be applied to the field of rehabilitation training. However, the existing technologies generally have problems such as insufficient accuracy of motion pattern recognition, difficulty in annotating motion data, and inaccurate real-time feedback, which all restrict the improvement of rehabilitation training effects and the realization of personalized treatment plans.

[0003] Currently, although the motion monitoring methods based on deep learning have made certain progress, they still face challenges such as data loss, inconsistent time series, and difficulty in capturing long-term dependencies between motion patterns when processing complex motion data. In addition, the existing technologies are difficult to accurately correct the patients' motion postures and exercise loads at different stages, and lack sufficient attention to the real-time and accuracy of fatigue assessment. Therefore, how to efficiently integrate motion patterns, exercise loads, and physiological state data, and combine virtual data to enhance the diversity of the training set has become the key to solving these problems. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a method for monitoring critical care rehabilitation training to solve the problems of how to accurately and real-time monitor the motion state of critically ill patients, correct the motion pattern, and dynamically adjust the exercise load in the training state.

[0006] To solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a method for monitoring critical care rehabilitation training, which includes: collecting the motion data of patients in real time, preliminarily analyzing the current motion state of patients, performing motion pattern recognition on patients, and outputting preliminary motion pattern labels;

[0008] Inputting the motion pattern labels into a generative adversarial network, and training the motion patterns and motion postures at different stages through a generator to generate virtual data with high similarity;

[0009] Initialize the motion data set with virtual data, input the motion data set into a multi-level convolutional neural network architecture for feature extraction, and generate the patient's motion rehabilitation status;

[0010] Based on the patient's motion rehabilitation status, learn the features of different motion patterns through a deep neural network, integrate the features of different motion patterns into a unified training network using a multi-task learning framework, and establish the correlation between different tasks by using the feature layer in the shared model;

[0011] Real-time monitor the patient's training status between different tasks, perform posture correction on the change of the motion pattern, and evaluate the fatigue degree of the motion at the same time to form training feedback data;

[0012] Upload the training feedback data to the cloud platform for storage and analysis, generate a warning signal according to the analysis result of the cloud data, and automatically issue a warning when the patient's training is abnormal.

[0013] As a preferred scheme of the intensive rehabilitation training monitoring method described in the present invention, wherein: the motion data includes motion posture data, motion load data, and motion state evaluation data;

[0014] Preliminarily analyze the patient's current motion state, perform motion pattern recognition on the patient, and output a preliminary motion pattern label. The specific steps are as follows:

[0015] For the motion posture data, motion load data, and motion state evaluation data, perform timestamp alignment and Kalman filter denoising;

[0016] Use a temporal convolutional neural network to extract dynamic features from the processed motion posture data and extract load features from the processed motion load data;

[0017] Use a deep neural network to extract rehabilitation features from the processed motion state evaluation data;

[0018] Input the extracted dynamic features, load features, and rehabilitation features into a long short-term memory network to capture temporal features and identify the patient's motion pattern;

[0019] Combine the real-time motion state evaluation data to correct the patient's motion pattern and output a preliminary motion pattern label.

[0020] As a preferred scheme of the intensive rehabilitation training monitoring method described in the present invention, wherein: input the motion pattern label into a generative adversarial network, and generate virtual data with high similarity by training the motion patterns and motion postures at different stages through a generator. The specific steps are as follows:

[0021] Use the motion pattern label as the input of the generator, where each label corresponds to the motion state of a time period;

[0022] The generator adopts a multi-layer structure that combines a bidirectional long short-term memory network and a self-attention mechanism to capture long-term and short-term dependencies in motion data;

[0023] Through a cross-modal fusion layer, different modal data in the motion data are mapped into a shared feature space, and using the multi-layer structure of the generator, virtual data of different stages of motion patterns are generated;

[0024] A discriminator is constructed using a deep convolutional neural network. The discriminator detects the authenticity of the virtual data and applies a temporal consistency loss to ensure the temporal continuity of the virtual data;

[0025] A virtual data generation model is used to supplement the training dataset, and the generator is pre-trained in combination with a transfer learning method;

[0026] The motion pattern label is input into the pre-trained generator to generate virtual data, and through the adversarial training of the discriminator, virtual data with high similarity is output.

[0027] As a preferred solution of the intensive rehabilitation training monitoring method described in the present invention, wherein: the motion dataset is initialized with virtual data, and the motion dataset is input into a multi-level convolutional neural network architecture for feature extraction to generate the patient's motion rehabilitation state. The specific steps are as follows:

[0028] The generated virtual data is normalized and amplified to construct a motion dataset, and it is labeled based on an automatic label generation method of deep learning to form a motion training set;

[0029] The motion training set is input into a multi-level convolutional neural network, and the primary and high-level representations of dynamic features, load features, and rehabilitation features are gradually extracted through convolutional layers;

[0030] The extracted primary and high-level representations are input into a fully connected layer for fusion, and the patient's motion rehabilitation state is generated through a classification and regression network.

[0031] As a preferred solution of the intensive rehabilitation training monitoring method described in the present invention, wherein: based on the patient's motion rehabilitation state, the features of different motion patterns are learned through a deep neural network, a multi-task learning framework is used to integrate the features of different motion patterns into a unified training network, and the feature layer in the shared model is adopted to establish the relevance between different tasks. The specific steps are as follows:

[0032] According to the patient's motion rehabilitation state, task one is established as identifying the change pattern of the motion posture, task two is established as evaluating the change trend of the motion load, and task three is established as identifying the abnormal state of the patient's rehabilitation state;

[0033] After clarifying the task objectives, a joint feature extraction layer is constructed jointly by a convolutional neural network and a deep neural network. In the joint feature extraction layer, features of different motion patterns will be extracted through the convolutional layer in the initial stage, and then further fused through the fully connected layer to form a joint feature space;

[0034] In the joint feature space, the joint feature extraction layer adjusts the feature sharing ratio between each task in real time through a weighted attention mechanism, thereby establishing a dynamic correlation between different tasks;

[0035] Based on the data features of each task, the correlation degree between each task and the joint feature extraction layer is adaptively adjusted. In the multi-task learning framework, a weighted loss function is used to adjust the loss weight of each task, and the contributions between tasks are optimized according to the feedback during the training process;

[0036] Combining the adjusted loss weights between tasks to establish a dynamic correlation, constructing a correlation matrix between tasks, and dynamically adjusting the information flow between tasks accordingly;

[0037] Based on the information flow between tasks after dynamic adjustment, a graph convolutional network is used for modeling;

[0038] According to the training process of each task, the weights of different tasks in the joint feature extraction layer are adjusted through real-time feedback, and the real-time feedback during the training process is combined with the feedback signal generated by virtual data generation to further optimize the joint feature extraction layer of the tasks.

[0039] As a preferred solution of the intensive rehabilitation training monitoring method described in the present invention, wherein: the training status of the patient between different tasks is monitored in real time, the posture is corrected for changes in the motion pattern, and the degree of exercise fatigue is evaluated simultaneously to form training feedback data. The specific steps are as follows:

[0040] The motion posture, motion load, and physiological state data in the training status of the patient between different tasks are monitored in real time, and denoising and time series alignment are performed through Kalman filtering to output standardized training status data;

[0041] Use a temporal convolutional neural network and a long short-term memory network to analyze the standardized training status data, and identify the motion pattern and training task type in real time;

[0042] Based on the motion pattern and training task type, detect the posture deviation in the motion pattern and correct it in real time;

[0043] Extract heart rate, electromyogram, respiratory rate, and blood oxygen saturation in the physiological state data as physiological signals using frequency domain analysis;

[0044] Input the physiological signal into a deep neural network to evaluate the fatigue state, and adjust the motion load in the training status according to the fatigue evaluation;

[0045] Record the adjusted training data in real time and integrate it to form training feedback data.

[0046] As a preferred solution of the intensive rehabilitation training monitoring method described in the present invention, wherein: uploading the training feedback data to the cloud platform for storage and analysis, generating a warning signal according to the analysis result of the cloud data, and automatically issuing a warning when the patient's training is abnormal. The specific steps are as follows:

[0047] Upload the training feedback data to the cloud platform through an encrypted channel for classified storage;

[0048] Clean and denoise the classified and stored training feedback data, and extract motion posture, motion load, and heart rate features through deep learning analysis;

[0049] The cloud platform identifies and marks abnormal patterns in the training feedback data based on motion posture, motion load, and heart rate features. Based on the abnormal patterns, the cloud generates a warning signal and divides the warning level according to the severity;

[0050] According to the warning level, issue notifications through multiple channels for real-time warning.

[0051] In a second aspect, the present invention provides an intensive rehabilitation training monitoring device, including a data collection module, a pattern generation module, a feature extraction module, a feature integration module, a status monitoring module, and a cloud warning module;

[0052] The data collection module is used to collect the patient's motion data in real time, preliminarily analyze the patient's current motion state, identify the motion pattern of the patient, and output a preliminary motion pattern label;

[0053] The pattern generation module is used to input the motion pattern label into the generative adversarial network, train the motion patterns and postures at different stages through the generator, and generate highly similar virtual data;

[0054] The feature extraction module is used to initialize the motion data set with virtual data, input the motion data set into a multi-level convolutional neural network architecture for feature extraction, and generate the patient's motion rehabilitation status;

[0055] The feature integration module is used to learn the features of different motion patterns based on the patient's motion rehabilitation status, integrate the features of different motion patterns into a unified training network using a multi-task learning framework, and establish the correlation between different tasks by adopting the feature layer in the shared model;

[0056] The state monitoring module is used to monitor the training state of the patient among different tasks in real time, perform posture correction on the change of the motion pattern, and evaluate the degree of exercise fatigue at the same time, so as to form training feedback data;

[0057] The cloud warning module is used to upload the training feedback data to the cloud platform for storage and analysis, generate a warning signal according to the analysis result of the cloud data, and automatically issue a warning when the patient's training is abnormal.

[0058] In a third aspect, the present invention provides a computer device, including a memory and a processor, where the memory stores a computer program, and: when the computer program is executed by the processor, any step of the intensive rehabilitation training monitoring method as described in the first aspect of the present invention is implemented.

[0059] In a fourth aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and: when the computer program is executed by the processor, any step of the intensive rehabilitation training monitoring method as described in the first aspect of the present invention is implemented.

[0060] The beneficial effects of the present invention are as follows: through a multi-level deep learning architecture and virtual data generation method, the accuracy of motion pattern recognition and the evaluation effect of the rehabilitation state can be effectively improved. First, the temporal convolutional neural network and the long short-term memory network are used to extract the dynamic features and temporal features in the motion data, which can accurately capture the subtle changes in the motion posture, motion load and rehabilitation state, so as to provide more accurate motion pattern labels. Secondly, by generating virtual data through the generative adversarial network and combining it with a multi-level convolutional neural network for training, the training data set can be greatly enriched, the generalization ability can be improved, and the problem of data scarcity that may occur in the actual data collection process can be overcome. In addition, based on the multi-task learning framework, the weights between different motion pattern features can be dynamically adjusted, the synergy between tasks can be enhanced, and thus the correction accuracy of the motion pattern and the real-time performance of fatigue evaluation can be further improved. Finally, through the intelligent warning function of the cloud platform, a warning signal can be sent in time when the patient's motion state is abnormal, which helps to realize a personalized rehabilitation training plan. Description of the Drawings

[0061] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative work, other drawings can be obtained according to these drawings.

[0062] Figure 1 It is a flowchart of the intensive rehabilitation training monitoring method in Embodiment 1.

[0063] Figure 2 It is a module diagram of the intensive care rehabilitation training monitoring device in Embodiment 1. Detailed implementation manners

[0064] To make the above objects, features and advantages of the present invention more obvious and understandable, the following detailed description of the specific implementation manners of the present invention will be given in conjunction with the accompanying drawings of the specification.

[0065] In the following description, many specific details are set forth to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar generalizations without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0066] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure or characteristic that can be included in at least one implementation manner of the present invention. The appearances of "in one embodiment" in different places in this specification do not all refer to the same embodiment, nor are they separate or alternative embodiments that exclude each other with other embodiments.

[0067] Embodiment 1, referring to Figure 1 and Figure 2 is the first embodiment of the present invention. This embodiment provides a method for monitoring intensive care rehabilitation training, including the following steps:

[0068] S1: Collect the patient's motion data in real time, preliminarily analyze the patient's current motion state, perform motion pattern recognition on the patient, and output a preliminary motion pattern label.

[0069] Furthermore, the motion data includes motion posture data, motion load data, and motion state evaluation data;

[0070] Specifically, the motion posture data, motion load data, and motion state evaluation data are collected in real time through sensors worn on the patient (such as inertial measurement unit (IMU) sensors, force sensors, etc.). Furthermore, the IMU sensor is used to collect the patient's three-dimensional acceleration, angular velocity, and magnetic field data to obtain the patient's motion posture information in real time; the motion load data is obtained in real time through an electromyogram (EMG) sensor to evaluate the load change of the patient during training; the motion state evaluation data is obtained in real time through a physiological signal monitoring device (such as a heart rate, respiratory rate, blood oxygen saturation, etc. monitoring device) to evaluate the patient's physiological state.

[0071] Preliminarily analyze the patient's current motion state, perform motion pattern recognition on the patient, and output a preliminary motion pattern label. The specific steps are as follows:

[0072] For motion posture data, motion load data, and motion state evaluation data, through timestamp alignment and Kalman filter denoising, the high quality and consistency of the data can be ensured, thus accurately reflecting the rehabilitation needs of patients;

[0073] Use a temporal convolutional neural network (TCN) to extract dynamic features from the processed motion posture data and load features from the processed motion load data;

[0074] Preferably, the use of a temporal convolutional neural network (TCN) effectively solves the limitations of traditional neural networks in processing long-time series data. Compared with traditional methods, the temporal convolutional neural network has stronger temporal feature capture ability, can accurately identify the minute changes in patients' movements, reduce motion deviation and errors, thus reducing the risks during the training process.

[0075] Use a deep neural network to extract rehabilitation features from the processed motion state evaluation data;

[0076] Input the extracted dynamic features, load features, and rehabilitation features into a long short-term memory network (LSTM) to capture temporal features and identify the patient's motion pattern;

[0077] Specifically, first arrange the dynamic features, load features, and rehabilitation features in a time series manner and construct corresponding time steps. Before inputting into the LSTM network, standardize the dynamic features, load features, and rehabilitation features to ensure that each feature dimension is within the same magnitude range, so as to avoid unnecessary impacts caused by excessive differences between features during the network training process. Then, through the gating mechanism of the LSTM, weight the input of each time step, capture the temporal dependence relationships of the dynamic features, load features, and rehabilitation features, and at the same time retain the long-term memory information to avoid the problem of gradient vanishing or explosion. Finally, the results output by the LSTM network will identify the patient's motion pattern by analyzing the temporal features.

[0078] Combine the real-time motion state evaluation data to correct the patient's motion pattern and output a preliminary motion pattern label.

[0079] S2: Input the motion pattern label into a generative adversarial network, and through the generator, train the motion patterns and motion postures at different stages to generate highly similar virtual data.

[0080] Furthermore, use the motion pattern label as the input of the generator, where each label corresponds to the motion state of a time period;

[0081] Among them, these motion pattern tags carry key information related to the patient's motion patterns, such as the type of motion, load, posture, etc. Inputting these motion pattern tags into the generator provides the generator with guiding information about the motion state, helping the generator generate virtual data that matches it.

[0082] The generator adopts a multi-layer structure that combines a bidirectional long short-term memory network and a self-attention mechanism to capture long-term and short-term dependencies in the motion data;

[0083] It should be noted that the motion pattern tags provide structural guidance for the motion data of the generator, while the complex network structure of the generator (the multi-layer structure of the bidirectional long short-term memory network and the self-attention mechanism) ensures the temporal characteristics of the data and the true simulation of the motion state. The two complement each other and jointly promote the generation of virtual data and the accurate reproduction of motion patterns.

[0084] Through the cross-modal fusion layer, different modal data in the motion data are mapped into a shared feature space, and using the multi-layer structure of the generator, virtual data of motion patterns at different stages are generated;

[0085] Among them, the cross-modal fusion layer (Cross-modal Fusion Layer) is a key component in deep learning for integrating data from different modalities; cross-modal fusion is widely used in multi-modal learning tasks, especially in the fields of medical and motion monitoring. It can map multiple types of data into a shared feature space, thereby learning the potential associations between different data sources.

[0086] Furthermore, the shared feature space refers to mapping different modal data (such as motion posture data, motion load data, and motion state assessment data) in the motion data into a unified representation space through the cross-modal fusion layer. This shared feature space is a high-dimensional feature space that contains comprehensive information from multiple data modalities, enabling different modal data to be correlated and compared with each other.

[0087] Preferably, the cross-modal fusion layer maps different modal motion data (such as motion posture data, motion load data, and motion state assessment data) into the shared feature space, enhancing the synergistic effect of multi-task learning, improving the processing ability and the temporal consistency of the generated data, thereby optimizing the accuracy and feedback effect of patient motion rehabilitation monitoring.

[0088] A discriminator is constructed using a deep convolutional neural network to detect the authenticity of the virtual data through the discriminator, and the temporal consistency loss is applied to ensure the temporal continuity of the virtual data;

[0089] It should be noted that the temporal consistency loss is a loss function used in temporal data generation, aiming to ensure the smoothness and consistency of the generated data in the time dimension. Specifically, the temporal consistency loss measures the continuity of the generated data between time steps and penalizes data that exhibits sudden changes or unnatural variations in the time series. It is typically applied to tasks dealing with continuous time series, such as video generation, speech synthesis, and motion data modeling, to ensure that the data generated at each frame or time step is temporally coherent and does not show obvious inconsistencies or jumps.

[0090] Preferably, the temporal consistency loss calculates the changes between adjacent time steps of the generated data, ensuring a reasonable transition in the changes between each time point of the virtual data and its adjacent time points, avoiding abrupt fluctuations or temporal breaks in the generated data, and enhancing the naturalness and temporal consistency of the generated data. As a result, the virtual data generated during the training process is more in line with the characteristics of real data in terms of long-term dependencies and temporal continuity.

[0091] Use a virtual data generation model to supplement the training dataset and pre-train the generator in combination with transfer learning methods;

[0092] Specifically, by collecting and annotating a certain amount of real patient motion data, a preliminary training set is constructed, including multi-modal data such as motion postures, changes in exercise loads, and motion state assessments. Then, this multi-modal data is input into a deep learning model, and a temporal convolutional neural network (TCN) and a long short-term memory network (LSTM) are used to extract the temporal features and dynamic features of the motion data. Subsequently, based on these extracted temporal and dynamic features, virtual data is generated through the generator of a generative adversarial network. However, the initial generator may not be able to produce high-quality virtual data, so transfer learning methods are adopted.

[0093] Furthermore, first use an existing large-scale motion dataset (such as a publicly available motion pattern or physiological signal dataset) to pre-train the generator to ensure that it can understand the basic laws and temporal characteristics of motion patterns. During the pre-training process, the data patterns learned by the generator will be transferred as prior knowledge to the new task, which can significantly shorten the training time and improve the quality of the generated virtual data.

[0094] Input the motion pattern labels into the pre-trained generator to generate virtual data, and through the adversarial training of the discriminator, output highly similar virtual data;

[0095] Specifically, these pre-trained generators are used to virtually generate patient personalized data (i.e., generate virtual data), and the authenticity of the generated data is further optimized through adversarial training of the discriminator. At this time, the generator can already generate motion data at multiple different stages based on the knowledge learned from pre-training, and ensure the temporal continuity of the virtual data through temporal consistency loss. Finally, these virtual data are combined with the actual data to form a more abundant training dataset (i.e., highly similar virtual data) for the training of subsequent motion pattern recognition and rehabilitation status assessment models.

[0096] S3: Initialize the motion dataset with virtual data, and input the motion dataset into a multi-level convolutional neural network architecture for feature extraction to generate the patient's motion rehabilitation status.

[0097] Furthermore, the generated virtual data is normalized and augmented to construct a motion dataset, and it is labeled based on a deep learning-based automatic label generation method (such as a deep learning method based on convolutional neural network CNN and long short-term memory network LSTM) to form a motion training set;

[0098] Specifically, the normalization process includes normalizing and scaling the data so that different types of motion data (such as motion posture data, motion load data, and motion state assessment data) can be processed within a unified range to ensure that it meets the input requirements for subsequent training.

[0099] Furthermore, data augmentation techniques are adopted to further improve the diversity of the dataset. This process includes operations such as deforming, translating, and rotating the virtual data to simulate different motion conditions and scenarios (such as posture transformation, load increase, training intensity change, etc.), thereby generating more diverse and representative training samples to ensure that the dataset covers all possible situations in actual training.

[0100] Input the motion training set into a multi-level convolutional neural network, and gradually extract the primary representations and high-level representations of dynamic features, load features, and rehabilitation features through the convolutional layer;

[0101] Among them, the primary representation is usually low-level and specific pattern information, such as the simple shape of the motion posture, the basic fluctuations of the load, etc.; while the high-level representation is more complex abstract patterns, such as the overall trend of the motion state, the changes in the rehabilitation progress, etc. These primary representations and high-level representations capture various information levels in the patient's motion data and provide a rich semantic basis for subsequent analysis.

[0102] Input the extracted primary representations and high-level representations into the fully connected layer for fusion, and generate the patient's motion rehabilitation status through the classification and regression network.

[0103] Preferably, by inputting the extracted primary representation and high-level representation into the fully connected layer for fusion, the low-level and high-level feature information can be effectively combined, enhancing the comprehensive understanding and analysis ability of the patient's movement rehabilitation status. The fusion operation of the fully connected layer not only improves the information interaction between different feature levels but also optimizes the recognition ability of complex movement patterns, thereby achieving a more accurate rehabilitation status assessment.

[0104] Furthermore, by jointly using the classification and regression networks, the classification of the patient's movement state (such as identifying different stages of rehabilitation) and the regression prediction of continuous values (such as assessing the rehabilitation progress or fatigue level) can be carried out simultaneously. This multi-task learning method combining classification and regression can not only achieve more accurate movement pattern recognition but also effectively predict the patient's rehabilitation effect and potential risks, providing more personalized and dynamic training feedback.

[0105] S4: Based on the patient's movement rehabilitation status, learn the characteristics of different movement patterns through a deep neural network, integrate the characteristics of different movement patterns into a unified training network using a multi-task learning framework, and establish the correlation between different tasks by using the feature layer in the shared model.

[0106] Furthermore, according to the patient's movement rehabilitation status, Task 1 is established as identifying the change pattern of movement postures, Task 2 is to evaluate the change trend of movement load, and Task 3 is to identify the abnormal status of the patient's rehabilitation status;

[0107] Preferably, the setting of different tasks reflects the multi-dimensional requirements in the patient's rehabilitation process, ensuring the comprehensiveness and accuracy of rehabilitation monitoring.

[0108] After clarifying the task objectives, jointly construct a joint feature extraction layer through a convolutional neural network and a deep neural network. In the joint feature extraction layer, the features of different movement patterns will be extracted through the convolutional layer in the initial stage and then further fused through the fully connected layer to form a joint feature space;

[0109] Specifically, use the convolutional layer to extract local features from the input data, usually using a convolutional kernel to perform a sliding operation on the input data. Let the input data be , and the output of the convolutional layer be . The convolutional operation is specifically represented by the following formula:

[0110] ;

[0111] Among them, represents the feature map output by the convolutional layer, represents the convolutional kernel, represents the bias term;

[0112] Furthermore, the fully connected layer flattens the feature map output by the convolutional layer into a one-dimensional vector, multiplies it with the weight matrix, and then adds the bias term, which is specifically represented by the following formula:

[0113] ;

[0114] where, represents the weight matrix of the fully connected layer ; represents flattening the feature map output by the convolutional layer into a one-dimensional vector, represents the bias term of the fully connected layer ; represents the fused feature output by the fully connected layer, represents the non-linear activation function;

[0115] All the fused features extracted and fused through the convolutional and fully connected layers will ultimately be integrated into the joint feature space, which contains all the features from different tasks.

[0116] Among them, the features of the motion pattern include the features of various motion behaviors and physiological responses shown by the patient during the training process, such as motion postures, motion loads, motion amplitudes, motion frequencies, and physiological states during the motion (such as fatigue level, heart rate changes, etc.). These features of the motion pattern can reflect the patient's motion mode, training intensity, and rehabilitation progress during the rehabilitation training. Through analysis by the deep neural network, it can identify changes in different motion patterns and further guide the adjustment of the motion load and the optimization of the training plan.

[0117] It should be noted that the joint feature extraction layer refers to the layer in the multi-task learning framework where different tasks extract features of the input data through shared neural network layers. These joint feature extraction layers are usually located at the front part of the network and serve multiple different tasks. For example, tasks such as motion posture recognition, load assessment, and rehabilitation state recognition can share this part of the network layer, thereby extracting low-level and high-level features of the data in the same feature space. Through the joint feature extraction layer, it is possible to better capture the common features in the data, improve the synergy effect between different tasks, and avoid redundant feature learning when each task is trained separately.

[0118] In the joint feature space, the joint feature extraction layer adjusts the feature sharing ratio between each task in real time through the weighted attention mechanism, thereby establishing a dynamic correlation between different tasks;

[0119] Among them, dynamic relevance refers to the fact that the dependency relationships between tasks at the shared feature layer can be automatically adjusted according to the training process and the requirements of different tasks. For example, Task A may be more sensitive to certain features in the shared feature space at a certain stage, while Task B may be more dependent on these features at other stages.

[0120] Specifically, through the joint feature extraction of a convolutional neural network (CNN) and a deep neural network (DNN), the preliminary features of multiple tasks are extracted and fused within the shared feature space. Then, based on the relevance of the current training status of each task, the weighted attention mechanism dynamically adjusts the weights of each task in the shared feature layer. The feature sharing ratio of each task is not fixed, but continuously changes according to the importance between tasks, the training effect, and the relative requirements of the tasks during the training process. Furthermore, through this dynamic adjustment mechanism, the feature allocation between tasks can be optimized according to real-time feedback, ensuring that each task obtains the optimal feature information in the shared feature space. Finally, this adjustment mechanism makes the relationship between tasks no longer static, but flexibly adjusts with the changes in data characteristics and task objectives during the training process, thereby establishing the dynamic relevance between tasks and optimizing the cooperative effect and overall performance between tasks.

[0121] Preferably, the introduction of the weighted attention mechanism solves the problem of uneven feature sharing ratios between tasks in multi-task learning. By dynamically adjusting the feature sharing ratio between tasks, resources can be allocated according to the importance of tasks and the current training status, thereby improving learning efficiency and ensuring that the optimization objectives of each task do not interfere with each other.

[0122] Based on the data characteristics of each task, the degree of association between each task and the joint feature extraction layer is adaptively adjusted. In the multi-task learning framework, a weighted loss function is used to adjust the loss weight of each task and optimize the loss contribution between tasks;

[0123] It should be noted that dynamic relevance adjusts the dependency relationships between tasks at the feature level, while adaptive adjustment is based on the feedback of training data to adjust the attention degree of each task to the shared feature layer. The former determines which features are more important for which tasks during the training process through the weighted attention mechanism, while the latter further finely adjusts the feature extraction strategy of each task through the changes in task features. Specifically, dynamic relevance provides a basic framework for subsequent adjustments, and adaptive adjustment is to further refine and optimize according to the feedback of each task within this framework.

[0124] Specifically, first, based on the training data characteristics of each task, different feature information required by each task is analyzed to understand the requirements of each task in the joint feature extraction layer. Then, by calculating the correlation between the data features of each task and the feature space output by the joint feature extraction layer, the importance of task features in the shared feature space is evaluated. On this basis, a weighted attention mechanism is used to adaptively adjust the association degree between each task and the joint feature extraction layer, so that the features of each task can dynamically allocate more or less resources according to their importance. This adjustment is completed through the backpropagation mechanism, in which the loss function of each task is weighted according to its weight in the feature extraction layer, so that during the training process, the feature contributions of each task are optimized, and finally the maximization of feature sharing and synergy effects between tasks is achieved.

[0125] Furthermore, according to the feature importance and training progress of each task, the loss values of each task are calculated, and the losses of each task are weighted. The determination of the weighting coefficient is based on the priority of the task, the performance in the current training stage, and the relative importance between tasks. Then, the weighted loss function is used to adjust the loss value of each task according to its corresponding weight, so that the contributions of tasks during the training process are reasonably allocated. During the training process, as the losses of each task change dynamically, the loss weights are adjusted adaptively, and the weight allocation between tasks is optimized through backpropagation. Finally, the optimized loss weights not only improve the training effect of each task, but also ensure the balance of loss contributions between tasks, thereby enhancing the overall synergy effect of the multi-task learning framework.

[0126] For example, in a critical care rehabilitation monitoring task, Task A (motion posture recognition) may be more dependent on fine motion features, while Task B (load assessment) needs to focus on the load change trend. By calculating the feature importance of each task, the weighted attention mechanism is used to adaptively adjust the association degree between Task A and Task B in the shared feature space, so as to ensure that Task A pays more attention to motion features and Task B pays more attention to load features. On this basis, in the multi-task learning framework, a weighted loss function is used to adjust the loss weights of each task. In the initial stage of training, Task A may face a higher loss, so a higher weight is given to it. As the training progresses and the performance of Task B improves, its loss weight is adjusted to maintain the balance between tasks. In this way, the loss contributions between tasks are optimized, ensuring that each task receives appropriate attention and optimization at different stages during the model training process, and finally improving the overall performance.

[0127] Establish dynamic associations by combining the adjusted loss weights between tasks, construct an inter-task association matrix, and dynamically adjust the information flow between tasks accordingly;

[0128] Specifically, during the training process, the weights between tasks are optimized through real-time feedback to adjust the degree of dependence of each task on shared features, ensuring that the contribution degrees between tasks meet the current training requirements. At this time, low-level features of each task are extracted through the convolutional layer, and the self-attention mechanism is used to capture the similarities and dependencies between tasks, establishing dynamic correlations between tasks. Combining these dynamic correlations of the optimized task weights, a task correlation matrix is constructed in a weighted manner, and this task correlation matrix is used to represent the mutual influence and feature sharing ratio between tasks. According to the task correlation matrix, the information flow between tasks can be dynamically adjusted, enabling different tasks to flexibly adjust the intensity and direction of feature sharing based on real-time training feedback, thereby maximizing the synergy between different tasks and improving the overall training performance.

[0129] Based on the information flow between tasks after dynamic adjustment, a graph convolutional network is used for modeling to improve the synergy between different tasks;

[0130] It should be noted that by optimizing the synergy between tasks, each task can cooperate more efficiently when sharing features, avoiding the situations of excessive independence or excessive dependence between tasks. Specifically, first, the convolutional layer and the self-attention mechanism dynamically adjust the degree of dependence of each task on shared features by capturing the correlations and dependencies between tasks, enabling each task to flexibly adjust its feature sharing strategy according to real-time feedback. Subsequently, the graph convolutional network (GCN) further accurately models the information flow and interaction between tasks by constructing a dependency graph between tasks, enhancing the efficiency of information transmission between tasks. This collaborative optimization enables different tasks to cooperate more closely during the training process, thereby improving the overall learning efficiency, maximizing the collaborative effect during the training process, and further promoting the improvement of the overall training performance.

[0131] According to the training process of each task, the weights of different tasks in the joint feature extraction layer are adjusted through real-time feedback, combining the real-time feedback during the training process with the feedback signal of virtual data generation to further optimize the joint feature extraction layer of the tasks.

[0132] S5: Real-time monitor the training status of patients between different tasks, perform posture correction on changes in movement patterns, and at the same time evaluate the degree of exercise fatigue to form training feedback data.

[0133] Furthermore, real-time monitor the movement postures, movement loads, and physiological state data of patients between different tasks, and perform denoising and time series alignment through Kalman filtering to output standardized training status data;

[0134] Use a temporal convolutional neural network and a long short-term memory network to analyze the standardized training status data to real-time identify movement patterns and training task types;

[0135] It should be noted that the motion mode is defined as the combination of different motion manners and postures shown by the patient during the rehabilitation training process. For example, the patient may perform actions such as knee flexion, standing, gait, etc. Different motion modes can represent different stages or action types in the patient's rehabilitation process. The training task type refers to different training goals or task types set during the rehabilitation process. These task types are usually closely related to the patient's rehabilitation goals. Each task is trained for specific rehabilitation needs with the aim of helping the patient recover motor function.

[0136] Based on the motion mode and training task type, detect the posture deviation in the motion mode and correct it in real time;

[0137] Specifically, according to the posture characteristics in the motion mode and the training task objectives, first establish the desired posture standard, compare it with the actually collected motion posture data, and calculate the posture deviation. Then, transmit the posture deviation through the backpropagation algorithm, and use the real-time correction mechanism to adjust parameters such as the exercise load and exercise angle in the training state, and dynamically correct the posture deviation. During this process, the adjustment measures are matched with the patient's training task type to avoid adverse effects on the patient due to excessive load or overcorrection of the posture. Finally, combine the corrected posture data with the updated training task type, adjust the training content, ensure that the patient maintains the best posture at each rehabilitation training stage, and optimize the training effect.

[0138] Preferably, in the traditional rehabilitation training monitoring method, the detection and correction of posture deviation often rely on manual judgment or simple sensor feedback. This method has problems such as low recognition accuracy and poor real-time performance, and it is difficult to achieve precise personalized rehabilitation training. Especially during the intensive rehabilitation process, the patient's physiological state is complex and has significant individual differences, and the traditional method is difficult to meet the needs of this high complexity.

[0139] Furthermore, by combining the motion mode and training task type, the intelligent and real-time detection and correction of posture deviation are realized, and the real-time problem and the personalized problem are specifically solved as follows:

[0140] Real-time problem: Through deep learning methods such as the Temporal Convolutional Network (TCN) and the Long Short-Term Memory Network (LSTM), dynamically capture the dynamic features in the motion data, so as to immediately detect and correct the posture deviation during the training process. Personalized problem: The correction is combined with the patient's motion mode and task type, avoiding the limitations of the "one-size-fits-all" in the traditional solution, so that the rehabilitation training of each patient can be precisely adjusted according to their actual state.

[0141] Extract the heart rate, electromyogram, respiratory rate, and blood oxygen saturation in the physiological state data as physiological signals by using frequency domain analysis;

[0142] Input physiological signals into a deep neural network to evaluate the fatigue state, and adjust the exercise load in the training state according to the fatigue evaluation;

[0143] It should be noted that the deep neural network extracts features from the patient's physiological signals, and uses a multi-layer neural network model (such as convolutional neural network, long short-term memory network, etc.) to analyze the time-domain and frequency-domain features in the physiological signals to identify the patient's fatigue level. Specifically, the deep neural network analyzes the change trend of physiological signals by comparing historical data and real-time data, and automatically evaluates the patient's current fatigue state. On the basis of fatigue state evaluation, the exercise load in the training state is adjusted through an algorithm. The specific methods include automatically reducing or increasing the exercise intensity according to the fatigue level, adjusting the training duration or frequency, and optimizing the rest cycle, so as to ensure that the exercise load in the training state is within the patient's tolerance range. The whole process will provide real-time feedback on the fatigue state, and continuously adjust the training plan through continuous monitoring and evaluation to optimize the rehabilitation effect.

[0144] Record the adjusted training data in real time and integrate it to form training feedback data.

[0145] Among them, the integration step can be to normalize the adjusted training data recorded in real time, and then use a data fusion algorithm (such as weighted average, feature splicing or principal component analysis, etc.) to merge multiple data types into unified training feedback data for subsequent analysis and cloud storage.

[0146] S6: Upload the training feedback data to the cloud platform for storage and analysis, generate a warning signal according to the analysis result of the cloud data, and automatically issue a warning when the patient's training is abnormal.

[0147] Furthermore, upload the training feedback data to the cloud platform through an encrypted channel for classified storage;

[0148] Clean and denoise the classified and stored training feedback data, and extract motion posture, exercise load and heart rate features through deep learning analysis;

[0149] Among them, the Kalman filter algorithm is used for cleaning and denoising to remove noise in the training feedback data and perform time series alignment to ensure the smoothness and accuracy of the data.

[0150] The cloud platform identifies and marks abnormal patterns in the training feedback data based on motion posture, exercise load and heart rate features. Based on the abnormal patterns, the cloud generates a warning signal and classifies the warning level according to the severity;

[0151] Preferably, through cloud analysis combined with deep learning methods, potential dangers such as excessive exercise load, posture deviation, or abnormal physiological signals can be identified in real time. Especially the comprehensive consideration of exercise load and physiological state can timely detect potential risks during the training of patients, so as to issue early warnings in the early stage and prevent excessive exercise or improper postures from damaging the health of patients.

[0152] According to the warning level, notifications are sent through multiple channels for real-time warning.

[0153] It should be noted that the warning level division is to solve the problem of different emergency degrees of different types of abnormal events. On this basis, real-time warning is realized by using multiple channels of notification (such as APP push, SMS, etc.) to ensure that different levels of emergencies are timely feedback. Warnings in the prior art usually rely on a single channel, while the present invention can more flexibly handle different warning levels through the combination of multiple notification methods, enabling relevant personnel (such as patients, trainers, medical staff) to take effective measures in the shortest time. This multi-level and timely notification mechanism greatly improves the practicability and response speed, playing a key role in protecting the safety of patients.

[0154] This embodiment also provides a severe rehabilitation training monitoring device, including: a data collection module, a mode generation module, a feature extraction module, a feature integration module, a state monitoring module, and a cloud warning module; the data collection module is used to collect the exercise data of patients in real time, preliminarily analyze the current exercise state of patients, identify the exercise mode of patients, and output preliminary exercise mode labels; the mode generation module is used to input the exercise mode labels into the generative adversarial network, train the exercise modes and postures at different stages through the generator, and generate highly similar virtual data; the feature extraction module is used to initialize the exercise data set through the virtual data, input the exercise data set into a multi-level convolutional neural network architecture for feature extraction, and generate the exercise rehabilitation state of patients; the feature integration module is used to, based on the exercise rehabilitation state of patients, learn the features of different exercise modes through a deep neural network, adopt a multi-task learning framework to integrate the features of different exercise modes into a unified training network, and use the feature layer in the shared model to establish the correlation between different tasks; the state monitoring module is used to monitor the training state of patients between different tasks in real time, correct the posture for changes in the exercise mode, and at the same time evaluate the fatigue degree of the exercise to form training feedback data; the cloud warning module is used to upload the training feedback data to the cloud platform for storage and analysis, generate a warning signal according to the cloud data analysis result, and automatically issue a warning when an abnormality occurs in the patient's training.

[0155] This embodiment also provides a computer device applicable to the situation of the intensive rehabilitation training monitoring method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the intensive rehabilitation training monitoring method proposed in the above embodiment.

[0156] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, a carrier network, NFC (Near Field Communication), or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covered on the display screen, or a button, a trackball, or a touchpad provided on the outer shell of the computer device, or an external keyboard, a touchpad, or a mouse, etc.

[0157] This embodiment also provides a storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements the intensive rehabilitation training monitoring method proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (Static Random Access Memory, abbreviated as SRAM), an electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, abbreviated as EEPROM), an erasable programmable read-only memory (Erasable Programmable Read Only Memory, abbreviated as EPROM), a programmable read-only memory (Programmable Red-Only Memory, abbreviated as PROM), a read-only memory (Read-Only Memory, abbreviated as ROM), a magnetic memory, a flash memory, a magnetic disk, or an optical disc.

[0158] In summary, through the multi-level deep learning architecture and virtual data generation method of the present invention, the accuracy of motion pattern recognition and the evaluation effect of the rehabilitation state can be effectively improved. First, by using the temporal convolutional neural network and the long short-term memory network to extract the dynamic features and temporal features in the motion data, the subtle changes in the motion posture, motion load, and rehabilitation state can be accurately captured, thus providing more accurate motion pattern labels. Second, by generating virtual data through the generative adversarial network and combining it with the multi-level convolutional neural network for training, the training data set can be greatly enriched, the generalization ability can be improved, and the problem of data scarcity that may occur in the actual data collection process can be overcome. In addition, based on the multi-task learning framework, the weights between different motion pattern features can be dynamically adjusted, the synergy between tasks can be enhanced, and thus the calibration accuracy of the motion pattern and the real-time performance of fatigue evaluation can be further improved. Finally, through the intelligent warning function of the cloud platform, a warning signal can be sent in time when the patient's motion state is abnormal, which helps to implement a personalized rehabilitation training plan.

[0159] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A monitoring method for intensive care rehabilitation training, characterized in that: Including: Collecting the patient's movement data in real time, preliminarily analyzing the patient's current movement state, performing movement pattern recognition on the patient, and outputting preliminary movement pattern labels; the movement data includes movement posture data, movement load data, and movement state assessment data; The steps of preliminarily analyzing the patient's current movement state, performing movement pattern recognition on the patient, and outputting preliminary movement pattern labels are as follows: Aligning the movement posture data, movement load data, and movement state assessment data by timestamp and denoising them through Kalman filtering; Using a temporal convolutional neural network to extract dynamic features from the processed movement posture data and extract load features from the processed movement load data; Using a deep neural network to extract rehabilitation features from the processed movement state assessment data; Inputting the extracted dynamic features, load features, and rehabilitation features into a long short-term memory network to capture temporal features and identify the patient's movement pattern; Correcting the patient's movement pattern in combination with real-time movement state assessment data and outputting preliminary movement pattern labels; Inputting the movement pattern labels into a generative adversarial network, and training the movement patterns and postures at different stages through the generator to generate highly similar virtual data; Initializing the movement dataset with the virtual data, inputting the movement dataset into a multi-layer convolutional neural network architecture for feature extraction, and generating the patient's movement rehabilitation state; Based on the patient's movement rehabilitation state, learning the features of different movement patterns through a deep neural network, integrating the features of different movement patterns into a unified training network using a multi-task learning framework, and establishing the correlation between different tasks by sharing the feature layer in the model; Real-time monitoring of the patient's training state between different tasks, performing posture correction on the changes in the movement pattern, and at the same time evaluating the degree of fatigue of the movement to form training feedback data; Uploading the training feedback data to the cloud platform for storage and analysis, generating a warning signal according to the analysis results of the cloud data, and automatically issuing a warning when the patient's training is abnormal.

2. The intensive rehabilitation training monitoring method according to claim 1, characterized in that: The steps of inputting the movement pattern labels into a generative adversarial network, and training the movement patterns and postures at different stages through the generator to generate highly similar virtual data are as follows: Taking the movement pattern labels as the input of the generator, where each label corresponds to the movement state of a time period; The generator adopts a multi-layer structure combining a bidirectional long short-term memory network and a self-attention mechanism to capture the long-term and short-term dependencies in the movement data; Mapping different modal data in the movement data to a shared feature space through a cross-modal fusion layer, and using the multi-layer structure of the generator to generate virtual data of different stages of movement patterns; Constructing a discriminator using a deep convolutional neural network, detecting the authenticity of the virtual data through the discriminator, and applying a temporal consistency loss to ensure the temporal continuity of the virtual data; Using a virtual data generation model to supplement the training dataset and pre-training the generator in combination with a transfer learning method; Inputting the movement pattern labels into the pre-trained generator, generating virtual data, and performing adversarial training through the discriminator to output highly similar virtual data.

3. The intensive rehabilitation training monitoring method according to claim 2, characterized in that: Initialize the motion dataset with virtual data, input the motion dataset into a multi-level convolutional neural network architecture for feature extraction, and generate the patient's motion rehabilitation status. The specific steps are as follows: Standardize and augment the generated virtual data, construct a motion dataset, and label it based on the deep learning-based automatic label generation method to form a motion training set; Input the motion training set into a multi-level convolutional neural network, and gradually extract the primary and high-level representations of dynamic features, load features, and rehabilitation features through the convolutional layer; Input the extracted primary and high-level representations into the fully connected layer for fusion, and generate the patient's motion rehabilitation status through the classification and regression network.

4. The intensive rehabilitation training monitoring method according to claim 3, wherein: Based on the patient's motion rehabilitation status, learn the features of different motion patterns through a deep neural network, adopt a multi-task learning framework to integrate the features of different motion patterns into a unified training network, and use the feature layer in the shared model to establish the correlation between different tasks. The specific steps are as follows: According to the patient's motion rehabilitation status, establish Task 1 as identifying the change pattern of motion postures, Task 2 as evaluating the change trend of motion load, and Task 3 as identifying the abnormal status of the patient's rehabilitation status; After clarifying the task objectives, jointly construct a joint feature extraction layer through a convolutional neural network and a deep neural network. In the joint feature extraction layer, the features of different motion patterns will be extracted through the convolutional layer in the initial stage, and then further fused through the fully connected layer to form a joint feature space; In the joint feature space, the joint feature extraction layer adjusts the feature sharing ratio between each task in real time through a weighted attention mechanism, thereby establishing a dynamic correlation between different tasks; Based on the data features of each task, adaptively adjust the degree of association between each task and the joint feature extraction layer. In the multi-task learning framework, use a weighted loss function to adjust the loss weight of each task, and optimize the contribution between tasks according to the feedback during the training process; Combine the adjusted loss weights between tasks to establish a dynamic correlation, construct a correlation matrix between tasks, and dynamically adjust the information flow between tasks accordingly; Based on the dynamically adjusted information flow between tasks, use a graph convolutional network for modeling; According to the training process of each task, adjust the weights of different tasks in the joint feature extraction layer through real-time feedback, and combine the real-time feedback during the training process with the feedback signal generated by virtual data to further optimize the joint feature extraction layer of the task.

5. The intensive rehabilitation training monitoring method according to claim 4, wherein: Monitor the patient's training status between different tasks in real time, correct the posture of the motion pattern change, and at the same time evaluate the fatigue degree of the motion to form training feedback data. The specific steps are as follows: Monitor the motion posture, motion load, and physiological state data in the patient's training status between different tasks in real time, and perform denoising and time series alignment through Kalman filtering to output standardized training status data; Use a temporal convolutional neural network and a long short-term memory network to analyze the standardized training status data to identify the motion pattern and training task type in real time; Based on the motion pattern and training task type, detect the posture deviation in the motion pattern and correct it in real time; Extract heart rate, electromyogram, respiratory rate, and blood oxygen saturation in physiological state data as physiological signals using frequency domain analysis; Input the physiological signals into a deep neural network to evaluate the fatigue state, and adjust the exercise load in the training state according to the fatigue assessment; Record the adjusted training data in real time and integrate it to form training feedback data.

6. The intensive rehabilitation training monitoring method according to claim 5, characterized in that: Upload the training feedback data to the cloud platform for storage and analysis, generate a warning signal based on the cloud data analysis results, and automatically issue a warning when the patient's training is abnormal. The specific steps are as follows: Upload the training feedback data to the cloud platform through an encrypted channel for classified storage; Clean and denoise the classified and stored training feedback data, and extract motion posture, exercise load, and heart rate characteristics through deep learning analysis; The cloud platform identifies and marks abnormal patterns in the training feedback data based on motion posture, exercise load, and heart rate characteristics. Based on the abnormal patterns, the cloud generates a warning signal and classifies the warning level according to the severity; According to the warning level, issue notifications through multiple channels for real-time warning.

7. A monitoring device for intensive care rehabilitation training, based on the intensive care rehabilitation training monitoring method according to any one of claims 1 to 6, characterized in that: It includes a data collection module, a pattern generation module, a feature extraction module, a feature integration module, a status monitoring module, and a cloud warning module; The data collection module is used to collect the patient's motion data in real time, preliminarily analyze the patient's current motion state, identify the motion pattern of the patient, and output the preliminary motion pattern label; The pattern generation module is used to input the motion pattern label into the generative adversarial network, and train the motion patterns and motion postures at different stages through the generator to generate highly similar virtual data; The feature extraction module is used to initialize the motion data set with virtual data, input the motion data set into a multi-level convolutional neural network architecture for feature extraction, and generate the patient's motion rehabilitation state; The feature integration module is used to learn the characteristics of different motion patterns through a deep neural network based on the patient's motion rehabilitation state, adopt a multi-task learning framework to integrate the characteristics of different motion patterns into a unified training network, and establish the correlation between different tasks by using the feature layer in the shared model; The status monitoring module is used to monitor the patient's training status between different tasks in real time, correct the posture of the motion pattern change, and evaluate the fatigue degree of the motion at the same time, forming training feedback data; The cloud warning module is used to upload the training feedback data to the cloud platform for storage and analysis, generate a warning signal based on the cloud data analysis results, and automatically issue a warning when the patient's training is abnormal.

Images