Science popularization content propaganda and education method and system based on intelligent interaction

Abstract

The invention relates to the technical field of hospital intelligent science popularization, in particular to a science popularization content propaganda and education method and system based on intelligent interaction, and the method comprises the steps: constructing an image-text propaganda and education content library and a rehabilitation training video library which cover multiple departments, multiple disease categories and multiple stages, and carrying out the structural processing and labeling modeling of the content, time mapping and node sorting are executed based on the hospitalization cycle of the patient, and a daily personalized propaganda and education task is automatically generated. Furthermore, a propaganda and education plan is executed through the robot, the robot is controlled to go to a ward according to the bed sequence, the identity of the patient is verified through visual recognition and voice confirmation, matched image-text propaganda and education, voice explanation and training videos are automatically played, and the manual propaganda and education workload is effectively reduced. Training actions of the patient are collected through camera equipment, whether action execution is standard or not is judged in real time, training feedback data are generated, standardized supervision of rehabilitation actions is achieved, and the risk that the rehabilitation progress is affected due to improper training is reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent science popularization technology in hospitals, and more particularly to a method and system for disseminating science popularization content based on intelligent interaction. Background Technology

[0002] With the continuous development of the healthcare system, the number of hospitalized patients is constantly increasing. Various diseases require extensive, time-sensitive health education at admission, early stages, recovery, and discharge, covering topics such as disease awareness, precautions, medication knowledge, rehabilitation training methods, and discharge guidance. Currently, clinical education mainly relies on medical staff delivering information orally, distributing printed materials, or simply showing videos. However, traditional education methods generally suffer from the following problems: First, the content of patient education relies on manual delivery, which is labor-intensive and repetitive. Hospitalized patients require multiple rounds of educational information with significantly varying content depending on their condition and surgical stage. Medical staff, in addition to daily ward rounds, surgical scheduling, and nursing duties, find it difficult to promptly and comprehensively meet the educational needs of all patients, resulting in low efficiency. Second, standardized execution of rehabilitation training and movement demonstrations cannot be guaranteed. For example, functional training, breathing exercises, and limb exercises during the rehabilitation period should be practiced strictly according to standard movement videos. However, without sufficient human supervision or when patients practice on their own, problems such as incorrect movements and inaccurate rhythm are prone to occur, affecting rehabilitation progress and even posing risks. Summary of the Invention

[0003] To address the aforementioned problems, this invention provides a method and system for popular science content dissemination based on intelligent interaction.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for popularizing science content based on intelligent interaction, comprising: S1. Based on the hospital's clinical departments and corresponding diseases, collect and structure the content of the education standards for admission, early stage, rehabilitation period and discharge stage to obtain graphic education data and training video data, and construct the initial knowledge dataset. S2. Based on the initial knowledge dataset, the graphic education data and training video data therein are labeled according to department, disease, physical condition, surgical type, anesthesia type, gender and time node to obtain labeled knowledge data; S3. Based on the tagged knowledge data, perform content matching on the patient's department, disease, bed number, surgery type, anesthesia type, gender, and hospitalization period to obtain matched data. Then, sort the matched content data according to time nodes to obtain a personalized science popularization and education dataset. S4. Based on the personalized science popularization and education dataset, perform task splitting and bed number ordering processing on the content to obtain the science popularization and education dataset for the day. S5. Based on the daily science education dataset, control the robot to go to the corresponding bed in sequence, perform visual recognition and voice interaction to confirm the identity of the patient, obtain the identity confirmation data, perform educational graphic display, voice broadcast and video playback based on the daily science education dataset and the identity confirmation data, and perform visual comparison processing on the patient's interactive actions to obtain training feedback data.

[0005] Further, S1 includes the following steps: Based on the disease catalogs of various clinical departments in the hospital, the field extraction and processing of the educational texts at the admission, early stage, recovery stage and discharge stage were performed. Symptom descriptions, precautions, contraindications, drug information and nursing requirements were parsed into paragraph-level structured fields to obtain graphic educational data. Based on the rehabilitation training requirements of different diseases, keyframe extraction processing was performed on the collected standard training videos, and the action posture, target limb parts and action rhythm of each keyframe were annotated frame by frame to obtain training video data. Based on the aforementioned text and image educational data and training video data, format unification, field mapping, and content denoising processing are performed on them, and the text fields, action label fields, and video frame sequences are integrated into an initial knowledge dataset.

[0006] Further, S2 includes the following steps: Based on the graphic education data in the initial knowledge dataset, the data is compared by department, disease, physical condition, surgical type, anesthesia type and gender. The text content is then mapped to the corresponding tag fields to obtain graphic tag data. Based on the training video data in the initial knowledge dataset, the action sequence is compared according to the time node stage. The key frame sequence is labeled with the label fields of admission, early stage, rehabilitation period and discharge stage to obtain video label data. Based on the image and text tag data and video tag data, the multimodal tag fields are integrated to merge text tags, action tags, and stage tags into a unified tag index structure, generating tagged knowledge data.

[0007] Further, S3 includes the following steps: Based on the patient's department, disease, physical condition, type of surgery, type of anesthesia, and gender, conditional matching processing is performed on the tagged knowledge data, and image and text tag data and video tag data that meet all conditions are selected as candidate matching content; Based on the inpatient cycle, time mapping processing is performed on the stage tags in the candidate matching content. The admission, early stage, recovery period and discharge tags are mapped according to the specific date range of the inpatient cycle to obtain stage content data with date attributes. Based on the content data of the aforementioned stage, sorting is performed on it according to the date attribute, and the text and image content of the same date and the labeled content of the training video are combined into content units of the same time node to form a personalized science popularization and education dataset.

[0008] Further, S4 includes the following steps: Based on the time node field in the personalized science popularization and education dataset, the text and image content and training video content corresponding to each time node are split into units, and the daily education content is parsed into independent task units to obtain daily task data. Based on the daily task data, the data entries containing the bed number field are sorted by bed number, and the task entries on the same date are arranged in numerical order of the bed number to obtain bed-sequential data. Based on the bed sequence data, the task items belonging to the same date are aggregated, and the sorted task items are recombined into a task set corresponding to the date to obtain the science popularization and education dataset for that day.

[0009] Furthermore, the step of controlling the robot to sequentially proceed to the corresponding beds based on the daily science popularization and education dataset includes the following steps: Based on the bed number field in the daily science popularization and education dataset, spatial mapping processing is performed on the task set corresponding to each bed to map the bed number to target coordinates within the ward area, thus obtaining bed location data. Using the bed location data, path planning is performed on the robot's current position and the target bed coordinates to generate navigation path data. Based on the navigation path data, the robot's movement control is performed to enable the robot to reach the corresponding bed and obtain a bed arrival tag. Based on the bed arrival tag, the robot performs visual and voice acquisition processing on the bed area to obtain the patient's facial image and voice response content. The acquired facial image is then processed for identity verification to obtain identity confirmation data.

[0010] Furthermore, the educational graphic display, voice broadcast, and video playback are performed through physical and audio devices installed on the robot itself.

[0011] Furthermore, the visual comparison processing of the patient's interactive actions includes the following steps: The robot acquires real-time images of the patient using a camera mounted on the robot body, performs posture key point extraction processing on the real-time images of the patient, and parses the patient's limb movements into motion feature data including joint positions and movement angles. The motion feature data is compared with the keyframe motion features of the training video to calculate the degree of difference between the two in terms of motion trajectory and motion amplitude, thereby obtaining training feedback data.

[0012] Furthermore, when the robot's battery level is lower than a preset level, the robot is controlled to return to its charging position and continue performing tasks after charging is completed.

[0013] A science popularization content dissemination system based on intelligent interaction, applied to any of the aforementioned science popularization content dissemination methods based on intelligent interaction, includes: The knowledge database module is used to collect and structure the content of the education standards for admission, early stage, rehabilitation period and discharge stage according to the hospital's clinical departments and corresponding diseases, so as to obtain graphic education data and training video data and build an initial knowledge dataset. The tagging module is used to perform tagging processing on the graphic education data and training video data in the initial knowledge dataset according to department, disease, physical condition, surgical type, anesthesia type, gender and time node to obtain tagged knowledge data; The health education data integration module is used to perform content matching based on the labeled knowledge data for patients’ departments, diseases, physical conditions, bed numbers, surgical types, anesthesia types, genders and hospitalization cycles to obtain matched data, and to sort the matched content data according to time nodes to obtain personalized science education datasets. The science education task generation module is used to perform task splitting and bed number ordering processing on the content according to the personalized science education dataset to obtain the science education dataset for the day. The education execution module is used to control the robot to go to the corresponding bed in sequence based on the daily science education dataset, perform visual recognition and voice interaction to confirm the identity of the patient, obtain the identity confirmation data, perform education graphic display, voice broadcast and video playback based on the daily science education dataset and the identity confirmation data, and perform visual comparison processing on the patient's interactive actions to obtain training feedback data.

[0014] The beneficial effects of this invention are as follows: By constructing a comprehensive library of textual and graphic educational content and a rehabilitation training video library covering multiple departments, diseases, and stages, and using structured processing and tag-based modeling, this invention digitizes and standardizes educational materials that previously relied on manual compilation. This allows educational content to be matched according to department, disease, surgical type, gender, and hospitalization period, thereby avoiding the high workload of repetitive explanations by medical staff. Furthermore, by mapping time and sorting nodes according to the patient's specific hospitalization period, it automatically generates personalized daily educational tasks, ensuring that different patients receive consistent, well-paced, and precisely timed educational arrangements at different stages, significantly improving the coverage and consistency of educational content. Further, by utilizing service robots to execute the educational plan, the robot can automatically go to the ward according to bed number order, perform visual recognition and voice confirmation of patients to ensure accurate targeting, and automatically play personalized textual and graphic content, voice content, and training videos through the robot terminal, thus automating the educational process. Compared to traditional paper-based education and verbal explanations, this invention enables continuous, high-standard education tasks to be carried out unattended, significantly reducing the workload of medical staff in routine, repetitive education. Simultaneously, by using cameras mounted on the robot to capture patient movements in real time, and comparing the extracted posture key points with key frames from training videos, the system can determine in real time whether the training movements meet the standards, generating training feedback data. This achieves standardized and quantifiable supervision of movement instruction, reducing the risk of patients' recovery progress being affected by improper movements. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the steps of a science popularization content dissemination method based on intelligent interaction in this invention.

[0016] Figure 2 This is a flowchart of the steps in this invention to control the robot to go to the corresponding beds in sequence based on the daily science popularization and education dataset. Detailed Implementation

[0017] Please see Figures 1-2 As shown, this invention relates to a method for popularizing science content based on intelligent interaction, comprising: S1. Based on the hospital's clinical departments and corresponding diseases, collect and structure the content of the education standards for admission, early stage, rehabilitation period and discharge stage to obtain graphic education data and training video data, and construct the initial knowledge dataset. S2. Based on the initial knowledge dataset, the graphic education data and training video data therein are labeled according to department, disease, physical condition, surgical type, anesthesia type, gender and time node to obtain labeled knowledge data; S3. Based on the tagged knowledge data, perform content matching on the patient's department, disease, physical condition, bed number, surgery type, anesthesia type, gender, and hospitalization period to obtain matched data. Then, sort the matched content data according to time nodes to obtain a personalized science popularization and education dataset. S4. Based on the personalized science popularization and education dataset, perform task splitting and bed number ordering processing on the content to obtain the science popularization and education dataset for the day. S5. Based on the daily science education dataset, control the robot to go to the corresponding bed in sequence, perform visual recognition and voice interaction to confirm the identity of the patient, obtain the identity confirmation data, perform educational graphic display, voice broadcast and video playback based on the daily science education dataset and the identity confirmation data, and perform visual comparison processing on the patient's interactive actions to obtain training feedback data.

[0018] In some embodiments, standardized educational materials from multiple clinical departments of the hospital are first collected uniformly, including key information that departments such as pulmonology, cardiac surgery, and orthopedics need to convey to patients during admission assessment, early preparation, rehabilitation, and discharge guidance. In the original materials, educational content is typically stored in the form of text descriptions, illustrative images, lists of precautions, equipment usage instructions, and rehabilitation exercise demonstration videos, which suffer from inconsistent formats, inconsistent granularity, and subjective expression. Therefore, this embodiment performs paragraph segmentation, entity recognition, and field extraction processing on the original text, reconstructing symptom descriptions, medication use, behavioral contraindications, and rehabilitation requirements into structured fields; and performs keyframe extraction, target limb recognition, and time-series segmentation on the video data, breaking down rehabilitation training movements into standardized video data with clear movement numbers, joint angle parameters, and rhythm cycles, thereby constructing a multimodal fusion initial knowledge dataset. After constructing the initial knowledge dataset, a labeling modeling mechanism is further introduced to finely label all structured content according to department category, disease type, surgical method (e.g., thoracoscopic surgery, open surgery), anesthesia method, gender, and hospitalization stage. This multi-dimensional tagging system enables the system to achieve a three-layer association of "disease – stage – individual characteristics" in subsequent matching, thereby overcoming the limitations of existing technologies that can only provide coarse-grained education at the departmental or disease level, and achieving more precise content matching capabilities. For example, for the same lung cancer patient population, since some patients undergo open-chest surgery and some undergo minimally invasive surgery, and the recovery rhythm and training content differ significantly, this invention uses tags based on surgical type and anesthesia method to enable different types of patients to be automatically matched with corresponding rehabilitation training video sequences, achieving a high degree of content personalization. After the patient is admitted to the hospital, this embodiment performs matching calculations on the tagged knowledge data based on their hospital registration information (such as department, disease, bed number, gender, surgical plan, and hospitalization period), automatically filters text and video content that meets all conditions, and performs time node mapping according to the patient's expected hospitalization period, generating an executable education node sequence for the corresponding stage content on a daily basis. For example, for patients undergoing pulmonary rehabilitation, the system prioritizes cough training and pain management information on the first day of rehabilitation; on the third day, it pushes information on getting out of bed and deep breathing exercises; and before discharge, it pushes information on inhaler use and home care instructions. The daily educational milestones are further broken down into specific task units and sorted by bed number to generate a daily science education dataset. During the education execution phase, service robots automatically go to the corresponding beds in sequence. The robots use visual sensors to collect facial images of patients and compare them with pre-stored patient profiles to confirm their identities and avoid mistakenly pushing educational content to the wrong recipients.After confirming the patient's identity, the robot automatically retrieves the corresponding text, illustrations, and video content based on the day's science education dataset. It then presents multimodal educational information through its display screen and voice module, achieving a fully automated educational process. During the rehabilitation training demonstration phase, a visual-motion comparison mechanism is introduced. The robot captures the patient's training movements using a camera and generates motion feature vectors containing joint positions and movement angles using a human posture keypoint extraction algorithm. These vectors are then compared frame-by-frame with keyframes from the training video. Quantitative feedback is generated based on the degree of movement deviation, prompting the patient to adjust their posture or range of motion, thus transforming the training from "demonstration-imitation" to "demonstration-monitoring-correction."

[0019] Further, S1 includes the following steps: Based on the disease catalogs of various clinical departments in the hospital, the field extraction and processing of the educational texts at the admission, early stage, recovery stage and discharge stage were performed. Symptom descriptions, precautions, contraindications, drug information and nursing requirements were parsed into paragraph-level structured fields to obtain graphic educational data. Based on the rehabilitation training requirements of different diseases, keyframe extraction processing was performed on the collected standard training videos, and the action posture, target limb parts and action rhythm of each keyframe were annotated frame by frame to obtain training video data. Based on the aforementioned text and image educational data and training video data, format unification, field mapping, and content denoising processing are performed on them, and the text fields, action label fields, and video frame sequences are integrated into an initial knowledge dataset.

[0020] In some embodiments, an entry point for collecting educational data is first constructed based on the disease catalogs of various clinical departments in the hospital, and hierarchical processing is performed on educational texts at different stages (admission, early stage, rehabilitation period, and discharge). To achieve standardized expression and computer readability, this embodiment uses a field extraction algorithm based on Natural Language Processing (NLP) to perform structured parsing of the educational texts at the paragraph level. Specifically, by performing syntactic dependency analysis and named entity recognition on the original educational texts, text fragments such as "symptom description," "precautions," "contraindications," "drug information," and "nursing requirements" are extracted from natural language sentences as independent fields and organized using a unified field template. For example, when parsing educational texts for the pulmonary rehabilitation period, the system identifies key fields such as "rehabilitation cough training," "activity tolerance assessment," and "residual anesthesia risk" based on part-of-speech tagging and a domain dictionary, making the text and image educational data searchable, indexable, and matchable. Furthermore, action recognition and time series analysis algorithms are used to perform keyframe extraction processing on the collected standard rehabilitation training videos. First, optical flow estimation and inter-frame difference algorithms are used to detect regions of motion change, filtering motion inflection points and rhythm nodes from consecutive video frames to maximize the motion information density of keyframes. Then, a human pose recognition network (such as a keypoint detection model based on HRNet or OpenPose) extracts skeletal keypoints from each keyframe, obtaining the joint position information of 17 or 25 points on the human body. To further meet the rhythm dependence of rehabilitation training, this embodiment performs rhythm annotation processing on the time intervals between keyframes, including the duration of the motion, the frequency of the motion, and the angle change of the target limb. For example, when processing knee flexion and extension training videos, the range of knee angle change, peak occurrence time, and motion rhythm cycle corresponding to each flexion and extension motion are recorded and annotated as a standard sequence of training motions. After completing the structured representation of text and video, an initial knowledge dataset is constructed through format unification, field mapping, and content denoising algorithms. Format unification mainly maps text fields and video motion fields to a consistent data structure through field standardization rules, such as uniformly organizing information such as "field name—field value—tag set—time series" in JSON format. The field mapping process associates the parsed natural language fields with action label fields to disease dictionaries and education stage dictionaries, enabling subsequent matching algorithms to quickly locate content based on rule indexing and label retrieval. Content denoising employs text stop word filtering, video redundant frame removal, and erroneous keypoint correction algorithms to improve the effectiveness and reliability of the initial dataset. For example, it removes tone-related words and redundant explanations from the text content, and in video skeleton extraction, it uses a keypoint confidence judgment mechanism to eliminate unstable outputs from the pose recognition model, thereby ensuring that all information in the initial knowledge dataset is high-quality, standardized data that can be directly used for subsequent matching and generation.

[0021] Further, S2 includes the following steps: Based on the graphic and textual educational data in the initial knowledge dataset, the data is processed by field comparison according to department, disease, surgical type, anesthesia type and gender. The text content is mapped with the corresponding tag fields to obtain graphic and textual tag data. Based on the training video data in the initial knowledge dataset, the action sequence is compared according to the time node stage. The key frame sequence is labeled with the label fields of admission, preoperative / early stage, postoperative / rehabilitation stage and discharge stage to obtain video label data. Based on the image and text tag data and video tag data, the multimodal tag fields are integrated to merge text tags, action tags, and stage tags into a unified tag index structure, generating tagged knowledge data.

[0022] In some embodiments, firstly, for text and image educational data, a field comparison algorithm based on a multi-dimensional label space is used to automatically map the text content to label fields such as department, disease, surgical type, anesthesia method, and gender. This comparison process not only relies on the traditional keyword matching mechanism but also introduces a text semantic similarity calculation model based on domain word vectors, measuring the relevance between text fields and label fields through vector space distance. For example, for the text "Precautions for breathing training during rehabilitation," the system converts its semantic embedding into feature vectors based on department and disease dictionaries and calculates similarity with the label space "thoracic surgery – lung cancer – rehabilitation recovery," automatically classifying the text under the corresponding label. This processing method differs from traditional manual classification and can identify implicit relationships and cross-paragraph semantic associations. In the labeling processing of training video data, this embodiment uses an action sequence comparison algorithm to label keyframe sequences with four stages: admission, pre-operative / early stage, post-operative / rehabilitation stage, and discharge. This process, based on the action feature vectors of video keyframes, identifies changes in action rhythm and intention through action time series models (such as LSTM or temporal convolutional networks), and compares them with the execution sequences of typical action templates corresponding to stage labels. For example, in a cardiac surgery rehabilitation training video, the deep breathing training action has high temporal consistency and key point trajectory similarity with the template sequence of the early rehabilitation stage, so the system can automatically classify the keyframe sequence into the "rehabilitation stage". By establishing a mapping between action features and stage labels, this embodiment realizes the staged structure organization of rehabilitation training videos, so that each group of keyframes not only has action labels, but also has clear time node attributes. After obtaining text label data and video label data, this embodiment further introduces a multimodal label fusion algorithm to integrate the two types of label fields into a unified label index structure. This fusion process combines a hierarchical label tree with a multimodal feature alignment mechanism, and by constructing a cross-modal label association graph, a one-to-many and many-to-many mapping relationship is formed between text labels, action labels, and stage labels. For example, in the education of patients in the pulmonary rehabilitation period, a text field of "deep breathing training" corresponds to multiple key frame sequences of training videos. These video sequences belong to different stages of the rehabilitation period, namely day 1 and day 2. Through the index structure, complex query capabilities such as aggregation by stage, retrieval by action type, and cross-modal matching by disease can be achieved.

[0023] Further, S3 includes the following steps: Based on the patient's department, disease, physical condition, type of surgery, type of anesthesia, and gender, conditional matching processing is performed on the tagged knowledge data, and image and text tag data and video tag data that meet all conditions are selected as candidate matching content; Based on the inpatient cycle, time mapping processing is performed on the stage tags in the candidate matching content. The admission, early stage, recovery period and discharge tags are mapped according to the specific date range of the inpatient cycle to obtain stage content data with date attributes. Based on the content data of the aforementioned stage, sorting is performed on it according to the date attribute, and the text and image content of the same date and the labeled content of the training video are combined into content units of the same time node to form a personalized science popularization and education dataset.

[0024] In some embodiments, conditional matching is first performed from tagged knowledge data based on the patient's department, disease, physical condition, surgical type, anesthesia method, and gender information. This processing employs a multi-condition joint filtering algorithm, using the above six dimensions as constraints. Through Boolean logic combination and tag vector intersection operations, text and image tag data and video tag data that simultaneously satisfy all tag conditions are selected. Unlike existing coarse-grained matching based only on department or disease, the conditional matching process in this embodiment utilizes a tag weight matrix and vectorized index structure, maintaining efficient filtering capabilities even with complex tag combinations. After obtaining candidate matching content, time mapping processing is further performed on the stage tags in the candidate content based on the hospitalization cycle filled in by the patient. This process constructs a time stage mapping model, mapping abstract stage tags such as "admission," "early stage," "recovery period," and "discharge" to the patient's actual hospitalization date range. Based on key time nodes such as "hospitalization start date," "surgery date," and "expected discharge date" in the hospitalization information, a time period segmentation algorithm is used to map the stage tags to specific dates. For example, when the hospitalization period for a patient in the pulmonary rehabilitation period is 7 days, the system can correspond "early rehabilitation period" to the first 1-2 days of rehabilitation, "functional recovery during rehabilitation" to the third 3-5 days of rehabilitation, and "discharge preparation" to the sixth 6-7 days of hospitalization.

[0025] Further, S4 includes the following steps: Based on the time node field in the personalized science popularization and education dataset, the text and image content and training video content corresponding to each time node are split into units, and the daily education content is parsed into independent task units to obtain daily task data. Based on the daily task data, the data entries containing the bed number field are sorted by bed number, and the task entries on the same date are arranged in numerical order of the bed number to obtain bed-sequential data. Based on the bed sequence data, the task items belonging to the same date are aggregated, and the sorted task items are recombined into a task set corresponding to the date to obtain the science popularization and education dataset for that day.

[0026] Specifically, firstly, based on the time node field in the personalized science popularization and education dataset, the text and image content and training video content corresponding to each time node are processed into execution units. This process involves parsing the multimodal content units for each date, independently splitting the text and image education paragraphs, action training video sequences, and precaution lists into several task units with execution attributes. For example, for the education node on the second day of a certain rehabilitation period, "explanation of breathing training knowledge," "playback of lower limb activity training videos," and "display of dietary precautions during the rehabilitation period" are extracted independently from the multimodal content package, giving them the execution granularity to be individually distributed to the robot. After obtaining the daily task data with execution granularity, the system further performs bed number sorting processing on the task items that include the bed number field. This embodiment uses a "bed number ordered scheduling algorithm," which achieves stable sorting of task items on the same date by numerically processing the bed number field of the task items. The sorting process not only considers the numerical order of the bed numbers but also adds auxiliary parameters such as ward zoning information and distance matrices in the implementation environment, thereby achieving physical sequence optimization that conforms to the ward spatial layout. For example, in a thoracic surgery ward with 20 beds, when the system generates daily task data for the third day of the recovery period, the algorithm can organize the tasks in the order of 01, 02, 03...20, while avoiding cross-room skipping visits, significantly reducing the complexity of the robot's movement path and improving execution efficiency. After obtaining the sequential bed data, the task items belonging to the same date are aggregated based on the sorting results to generate a structured daily science education dataset.

[0027] Furthermore, the step of controlling the robot to sequentially proceed to the corresponding beds based on the daily science popularization and education dataset includes the following steps: Based on the bed number field in the daily science popularization and education dataset, spatial mapping processing is performed on the task set corresponding to each bed to map the bed number to target coordinates within the ward area, thus obtaining bed location data. Using the bed location data, path planning is performed on the robot's current position and the target bed coordinates to generate navigation path data. Based on the navigation path data, the robot's movement control is performed to enable the robot to reach the corresponding bed and obtain a bed arrival tag. Based on the bed arrival tag, the robot performs visual and voice acquisition processing on the bed area to obtain the patient's facial image and voice response content. The acquired facial image is then processed for identity verification to obtain identity confirmation data.

[0028] Specifically, based on the bed number field in the daily science education dataset, spatial mapping processing is performed on the task set corresponding to each bed. This mapping process calls the "bed spatial coordinate modeling algorithm," which maps the bed number to the actual physical coordinates within the ward area by reading the ward building floor plan, bed arrangement table, and spatial calibration points. The system uses a homogeneous coordinate system transformation to map the bed number sequence to a unified two-dimensional or three-dimensional spatial coordinate system, obtaining information-based bed location data with position vectors. For example, for a patient in bed 15 of the thoracic surgery department, the bed number can generate coordinates (x=3.2, y=7.8) after spatial mapping, along with the spatial area label "Left Ward A Zone," thus providing a precise navigation target for path planning. Unlike existing technologies that rely on manual input or fixed template coordinates, the spatial mapping algorithm in this embodiment can adapt to different departments and ward layouts, achieving dynamic and configurable spatial positioning. After obtaining the bed location data, path planning processing is performed based on the robot's current position and target coordinates. This embodiment employs a cost function-based pathfinding algorithm (such as A*, Dijkstra's algorithm, or an improved heuristic planning model) to perform global path planning for the robot based on the corridor structure, obstacle distribution, and passage rules within the ward. During the planning process, the system assigns a "travel cost" to each segment of the path, including distance cost, turning cost, and congestion risk cost. Optimal or suboptimal navigation path data is generated through cost accumulation calculations. Furthermore, to adapt to the dynamic nature of the hospital environment, this embodiment incorporates a local obstacle avoidance algorithm (such as DWA or VFH), allowing the robot to fine-tune its navigation path based on real-time sensor data, ensuring the continuity and stability of path tracking. The robot executes movement control based on the navigation path data and automatically generates a bed arrival tag upon reaching the target coordinate area, serving as a trigger signal for subsequent identity verification procedures. After the robot successfully reaches the target bed, visual and voice data acquisition and processing are performed using the robot's onboard sensors based on the bed arrival tag. The visual acquisition section uses a depth camera or RGB camera to capture patient facial images and extracts facial feature vectors using a face detection model. The voice acquisition section uses a microphone array to capture the patient's response audio and verifies the validity of the response using voice activity detection (VAD) and keyword recognition algorithms. In the identity verification stage, facial feature vectors are matched against patient registration features stored in the electronic medical record system. The comparison algorithm can use cosine similarity or a feature distance function based on metric learning. When the similarity value exceeds a preset threshold, the system confirms the patient's identity and generates identity verification data, ensuring the accuracy and security of bed access.

[0029] Furthermore, the educational graphic display, voice broadcast, and video playback are performed through physical and audio devices installed on the robot itself.

[0030] Furthermore, the visual comparison processing of the patient's interactive actions includes the following steps: The robot acquires real-time images of the patient using a camera mounted on the robot body, performs posture key point extraction processing on the real-time images of the patient, and parses the patient's limb movements into motion feature data including joint positions and movement angles. The motion feature data is compared with the keyframe motion features of the training video to calculate the degree of difference between the two in terms of motion trajectory and motion amplitude, thereby obtaining training feedback data.

[0031] In some embodiments, the robot's onboard camera first captures real-time images of the patient's current interactive movements. To obtain reliable motion structure information in complex ward environments, the system performs pose keypoint extraction processing on the acquired image sequences, employing human pose estimation algorithms (e.g., keypoint detection models based on HRNet, OpenPose, or lightweight Transformer) to locate keypoints in the patient's head, torso, and limb joints. By fitting two-dimensional or three-dimensional spatial coordinates to the keypoints, the system parses the patient's movements into motion feature data including joint position coordinates, angle changes, limb motion vectors, and movement amplitude, thereby forming a feature sequence that can be used for temporal analysis. For example, during "shoulder abduction training," the system continuously tracks the spatial trajectories of the patient's shoulder, elbow, and wrist joints, calculating their joint angles, range of motion, and movement rhythm to characterize the completeness and standardization of the movements. After obtaining the patient's motion features, a sequence comparison process is further performed between these features and the pre-annotated keyframe motion features in the training video. The comparison process employs sequence alignment models based on Dynamic Temporal Warping (DTW), Temporal Convolutional Networks (TCN), or Transformer to calculate similarity and align the patient's action sequences with standard action sequences along the timeline. The comparison process considers not only the keypoint similarity of individual frames but also the continuity of joint angle changes, the curvature of the movement trajectory, the degree of rhythmic synchronization, and the amplitude of movement. For example, in knee flexion and extension training, the patient's knee joint angle sequence is matched with the angle sequence in a standard video, and a comprehensive score is given based on indicators such as curve fitting error, peak amplitude difference, and movement cycle consistency. Using these temporal comparison results, the system calculates the degree of difference between the patient's movements and the standard movements, generating training feedback data, including structured feedback information such as the magnitude of deviation, the location of insufficient movement, the rhythmic deviation range, and movement abnormalities that may pose a rehabilitation risk.

[0032] Furthermore, when the robot's battery level is lower than a preset level, the robot is controlled to return to its charging position and continue performing tasks after charging is completed.

[0033] It should be noted that while the robot is performing navigation, identity recognition, or educational playback, its energy management module continuously monitors the remaining battery power and compares it in real time with a preset battery power threshold. Once the battery power is detected to be below the preset threshold, the system immediately triggers the battery management process. The scheduling control algorithm records the status of the ongoing task, including the current bed number, task progress marker, remaining distance on the navigation path, and the current playback position of the multimodal educational content, forming recoverable task context data. Subsequently, the automatic return path planning model is invoked, enabling the robot to plan the shortest or safest path from its current position to the charging station and automatically return to its original position based on the planned path. During this process, the robot maintains environmental awareness and local obstacle avoidance capabilities, ensuring a safe return to the charging point in complex spaces such as wards. Upon reaching the charging station, the charging process is initiated wirelessly or via mechanical contact, and the charging progress is monitored in real time. When the battery power reaches the preset recovery threshold or charging is complete, the system automatically resumes the educational process based on the previously recorded task context data. The task recovery mechanism employs a state backtracking algorithm to restore the robot to its position in the task chain before the interruption. For example, if the robot pauses due to insufficient power while explaining a rehabilitation training video to the patient in bed 12, the system can automatically navigate the robot to bed 12 after recharging and resume execution from the paused task point without manual intervention. In continuous education scenarios, this mechanism can avoid task skipping or repetition caused by power fluctuations, improving the integrity and continuity of task execution.

[0034] This invention also includes a science popularization content dissemination system based on intelligent interaction, applied to any of the aforementioned science popularization content dissemination methods based on intelligent interaction, comprising: The knowledge database module is used to collect and structure the content of the education standards for admission, early stage, rehabilitation period and discharge stage according to the hospital's clinical departments and corresponding diseases, so as to obtain graphic education data and training video data and build an initial knowledge dataset. The tagging module is used to perform tagging processing on the graphic education data and training video data in the initial knowledge dataset according to department, disease, physical condition, surgical type, anesthesia type, gender and time node to obtain tagged knowledge data; The health education data integration module is used to perform content matching based on the labeled knowledge data for patients’ departments, diseases, physical conditions, bed numbers, surgical types, anesthesia types, genders and hospitalization cycles to obtain matched data, and to sort the matched content data according to time nodes to obtain personalized science education datasets. The science education task generation module is used to perform task splitting and bed number ordering processing on the content according to the personalized science education dataset to obtain the science education dataset for the day. The education execution module is used to control the robot to go to the corresponding bed in sequence based on the daily science education dataset, perform visual recognition and voice interaction to confirm the identity of the patient, obtain the identity confirmation data, perform education graphic display, voice broadcast and video playback based on the daily science education dataset and the identity confirmation data, and perform visual comparison processing on the patient's interactive actions to obtain training feedback data.

[0035] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for popularizing science content based on intelligent interaction, characterized in that, include: S1. Based on the hospital's clinical departments and corresponding diseases, collect and structure the content of the education standards for admission, early stage, rehabilitation period and discharge stage to obtain graphic education data and training video data, and construct the initial knowledge dataset. S2. Based on the initial knowledge dataset, the graphic education data and training video data therein are labeled according to department, disease, physical condition, surgical type, anesthesia type, gender and time node to obtain labeled knowledge data; S3. Based on the tagged knowledge data, perform content matching on the patient's department, disease, physical condition, bed number, surgery type, anesthesia type, gender, and hospitalization period to obtain matched data. Then, sort the matched content data according to time nodes to obtain a personalized science popularization and education dataset. S4. Based on the personalized science popularization and education dataset, perform task splitting and bed number ordering processing on the content to obtain the science popularization and education dataset for the day. S5. Based on the daily science education dataset, control the robot to go to the corresponding bed in sequence, perform visual recognition and voice interaction to confirm the identity of the patient, obtain the identity confirmation data, perform educational graphic display, voice broadcast and video playback based on the daily science education dataset and the identity confirmation data, and perform visual comparison processing on the patient's interactive actions to obtain training feedback data.

2. The method for popularizing science content based on intelligent interaction according to claim 1, characterized in that, S1 includes the following steps: Based on the disease catalogs of various clinical departments in the hospital, the field extraction and processing of the educational texts at the admission, early stage, recovery stage and discharge stage were performed. Symptom descriptions, precautions, contraindications, drug information and nursing requirements were parsed into paragraph-level structured fields to obtain graphic educational data. Based on the rehabilitation training requirements of different diseases, keyframe extraction processing was performed on the collected standard training videos, and the action posture, target limb parts and action rhythm of each keyframe were annotated frame by frame to obtain training video data. Based on the aforementioned text and image educational data and training video data, format unification, field mapping, and content denoising processing are performed on them, and the text fields, action label fields, and video frame sequences are integrated into an initial knowledge dataset.

3. The method for popularizing science content based on intelligent interaction according to claim 1, characterized in that, S2 includes the following steps: Based on the graphic education data in the initial knowledge dataset, the data is compared by department, disease, physical condition, surgical type, anesthesia type and gender. The text content is then mapped to the corresponding tag fields to obtain graphic tag data. Based on the training video data in the initial knowledge dataset, the action sequence is compared according to the time node stage. The key frame sequence is labeled with the label fields of admission, early stage, rehabilitation period and discharge stage to obtain video label data. Based on the image and text tag data and video tag data, the multimodal tag fields are integrated to merge text tags, action tags, and stage tags into a unified tag index structure, generating tagged knowledge data.

4. The method for popularizing science content based on intelligent interaction according to claim 1, characterized in that, S3 includes the following steps: Based on the patient's department, disease, physical condition, type of surgery, type of anesthesia, and gender, conditional matching processing is performed on the tagged knowledge data, and image and text tag data and video tag data that meet all conditions are selected as candidate matching content; Based on the inpatient cycle, time mapping processing is performed on the stage tags in the candidate matching content. The admission, early stage, recovery period and discharge tags are mapped according to the specific date range of the inpatient cycle to obtain stage content data with date attributes. Based on the content data of the aforementioned stage, sorting is performed on it according to the date attribute, and the text and image content of the same date and the labeled content of the training video are combined into content units of the same time node to form a personalized science popularization and education dataset.

5. A method for popularizing science content based on intelligent interaction according to claim 1, characterized in that, S4 includes the following steps: Based on the time node field in the personalized science popularization and education dataset, the text and image content and training video content corresponding to each time node are split into units, and the daily education content is parsed into independent task units to obtain daily task data. Based on the daily task data, the data entries containing the bed number field are sorted by bed number, and the task entries on the same date are arranged in numerical order of the bed number to obtain bed-sequential data. Based on the bed sequence data, the task items belonging to the same date are aggregated, and the sorted task items are recombined into a task set corresponding to the date to obtain the science popularization and education dataset for that day.

6. The method for popularizing science content based on intelligent interaction according to claim 1, characterized in that, The process of controlling the robot to sequentially proceed to the corresponding beds based on the daily science education dataset includes the following steps: Based on the bed number field in the daily science popularization and education dataset, spatial mapping processing is performed on the task set corresponding to each bed to map the bed number to target coordinates within the ward area, thus obtaining bed location data. Using the bed location data, path planning is performed on the robot's current position and the target bed coordinates to generate navigation path data. Based on the navigation path data, the robot's movement control is performed to enable the robot to reach the corresponding bed and obtain a bed arrival tag. Based on the bed arrival tag, the robot performs visual and voice acquisition processing on the bed area to obtain the patient's facial image and voice response content. The acquired facial image is then processed for identity verification to obtain identity confirmation data.

7. The method for popularizing science content based on intelligent interaction according to claim 1, characterized in that, The educational graphic displays, voice broadcasts, and video playback are performed through physical and audio devices installed on the robot itself.

8. The method for popularizing science content based on intelligent interaction according to claim 1, characterized in that, The visual comparison processing of the patient's interactive actions includes the following steps: The robot acquires real-time images of the patient using a camera mounted on the robot body, performs posture key point extraction processing on the real-time images of the patient, and parses the patient's limb movements into motion feature data including joint positions and movement angles. The motion feature data is compared with the keyframe motion features of the training video to calculate the degree of difference between the two in terms of motion trajectory and motion amplitude, thereby obtaining training feedback data.

9. A method for popularizing science content based on intelligent interaction according to claim 1, characterized in that, When the robot's battery level is lower than a preset level, the robot is controlled to return to its charging position and continue performing tasks after charging is completed.

10. A science popularization content dissemination system based on intelligent interaction, applied to the science popularization content dissemination method based on intelligent interaction as described in any one of claims 1-9, characterized in that, include: The knowledge database module is used to collect and structure the content of the education standards for admission, early stage, rehabilitation period and discharge stage according to the hospital's clinical departments and corresponding diseases, so as to obtain graphic education data and training video data and build an initial knowledge dataset. The tagging module is used to perform tagging processing on the graphic education data and training video data in the initial knowledge dataset according to department, disease, physical condition, surgical type, anesthesia type, gender and time node to obtain tagged knowledge data; The health education data integration module is used to perform content matching based on the labeled knowledge data for patients’ departments, diseases, physical conditions, bed numbers, surgical types, anesthesia types, genders and hospitalization cycles to obtain matched data, and to sort the matched content data according to time nodes to obtain personalized science education datasets. The science education task generation module is used to perform task splitting and bed number ordering processing on the content according to the personalized science education dataset to obtain the science education dataset for the day. The education execution module is used to control the robot to go to the corresponding bed in sequence based on the daily science education dataset, perform visual recognition and voice interaction to confirm the identity of the patient, obtain the identity confirmation data, perform education graphic display, voice broadcast and video playback based on the daily science education dataset and the identity confirmation data, and perform visual comparison processing on the patient's interactive actions to obtain training feedback data.

Images