Intelligent walking stick early warning method and system based on multi-parameter physiological monitoring

Abstract

The application discloses an intelligent walking stick early warning method and system based on multi-parameter physiological monitoring, relates to the technical field of intelligent walking sticks, and comprises the following steps: collecting physiological parameters and motion data of an old user through various sensors, and extracting physiological features and motion features; collecting normal state monitoring data of the old user to construct a normal behavior model for multi-level anomaly detection, classifying the multi-level anomaly detection results to obtain an anomaly category; performing hierarchical early warning according to the anomaly category, generating corresponding early warning information, triggering fuzzy voice interaction, obtaining an abnormal state and real-time demand of the old user according to a fuzzy voice recognition result; obtaining a response strategy through the early warning information and the abnormal state and real-time demand of the old user, recording actual conditions after each early warning information to perform model feedback adjustment. Through the closed-loop architecture of multi-modal perception, intelligent decision and hierarchical response, the safety of daily life of the old user and the health monitoring efficiency are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of smart cane technology, and more specifically, to a smart cane early warning method and system based on multi-parameter physiological monitoring. Background Technology

[0002] With the accelerating global aging process, the needs for health monitoring and safety protection among the elderly are becoming increasingly prominent. Traditional canes only provide simple support and cannot meet the needs of the elderly for health monitoring, emergency assistance, and intelligent interaction. In recent years, smart canes have gradually become a research hotspot. By integrating multiple sensors and communication technologies, they can monitor the user's physiological parameters, track location, detect falls, and make emergency calls, thereby improving the elderly's ability to live independently and their level of safety.

[0003] In terms of physiological parameter monitoring, heart rate and blood oxygen monitoring based on photoplethysmography (PPG) technology has become one of the important functions of smart canes. PPG calculates heart rate and blood oxygen saturation by detecting changes in blood volume. When the data is abnormal, it can trigger an alarm, providing early warning for patients with cardiovascular or respiratory diseases. The positioning and trajectory recording function usually relies on GPS or Beidou satellite positioning systems, combined with inertial sensor data, to track the user's location in real time and provide navigation assistance when lost or wandering. In terms of emergency assistance, smart canes are usually equipped with an SOS one-button call function and support voice calls, enabling the elderly to quickly seek help in emergencies. The data cloud synchronization function enables long-term storage and analysis of health data. Physiological parameters, activity trajectories, and other information are uploaded to the cloud platform through wireless transmission technologies (such as 4G / 5G, Wi-Fi, or NB-IoT), making it convenient for family members or medical staff to remotely monitor health status and providing data support for chronic disease management.

[0004] In summary, the intelligent cane early warning method based on multi-parameter physiological monitoring integrates PPG, ECG, accelerometer, and positioning technologies to construct a real-time and accurate health and safety monitoring system. Therefore, there is an urgent need for a high-precision intelligent cane early warning method based on multi-parameter physiological monitoring to provide a more comprehensive solution for elderly health management. Summary of the Invention

[0005] To address the aforementioned technical issues, this invention proposes an intelligent cane early warning method and system based on multi-parameter physiological monitoring. This system enables real-time monitoring of the health status and safety of elderly users and triggers an early warning mechanism in abnormal situations, providing efficient and accurate health and safety monitoring for the elderly.

[0006] The first aspect of this invention provides an intelligent cane early warning method based on multi-parameter physiological monitoring, comprising the following steps:

[0007] Physiological parameters and movement data of elderly users are collected through multiple sensors. The physiological parameters and movement data are preprocessed and physiological and movement features are extracted.

[0008] Collect monitoring data of normal status of elderly users to construct a normal behavior model, import the physiological and motor characteristics into the normal behavior model to perform multi-level anomaly detection, and classify the multi-level anomaly detection results to obtain the anomaly category.

[0009] Based on the abnormality category, a graded early warning is issued, corresponding early warning information is generated, the early warning information is verified and fuzzy voice interaction is triggered, and the abnormal status and real-time needs of elderly users are obtained based on the fuzzy voice recognition results.

[0010] Response strategies are obtained by analyzing the warning information, the abnormal status of elderly users, and their real-time needs. The actual situation after each warning is recorded, and the model is adjusted based on the actual situation.

[0011] In this solution, physiological parameters and movement data of elderly users are collected through multiple sensors. The physiological parameters and movement data are then preprocessed, and physiological and movement features are extracted. Specifically:

[0012] Based on the sensor group integrated in the handle of the smart cane and the sensor group embedded in the cane body to sense the user's physiological parameters and motion data, edge computing preprocessing is used, the sliding window Z-score method is used to identify and remove abnormal values ​​of physiological parameters, time series linear difference is used to fill missing values, and Kalman filtering is applied to eliminate sensor noise.

[0013] Demand information is generated based on blood pressure monitoring, blood oxygen assessment and fall detection. Based on the demand information, similarity calculation is used for retrieval, historical instances that meet the similarity requirements are extracted, and the physiological and motor features involved in the historical instances are used to construct an initial feature pool.

[0014] The initial feature pool is mixed and encoded to initialize the population of the genetic algorithm. A multi-objective fitness function is constructed based on the blood pressure prediction accuracy, blood oxygen abnormality recognition rate and fall detection timeliness. During the iteration process of the genetic algorithm, ten-fold cross-validation is used to calculate the three-dimensional fitness of each individual.

[0015] A predetermined number of individuals are selected as parents using a tournament selection method. The physiological and motor characteristic segments of the selected parents are subjected to independent crossover and position flip mutations, and the parameter weight segments are perturbed by Gaussian. After iterative training, the optimal individual is selected after the predetermined termination condition is met.

[0016] Based on the optimal individual, key physiological and motor feature indicators are extracted. The importance of the features is ranked using learned parameter weights to obtain the final physiological and motor feature indicators. Based on the final physiological and motor feature indicators, physiological and motor features are extracted from the preprocessed physiological parameters and motor data.

[0017] In this solution, data on the normal state of elderly users is collected to construct a normal behavior model, specifically as follows:

[0018] Obtain physiological parameters and motion data within a preset historical time step, extract normal state monitoring data of elderly users, and extract physiological and motion characteristics under normal state conditions from the normal state monitoring data;

[0019] The physiological and motor characteristics under normal conditions are standardized and nonlinear dimensionality is reduced by t-SNE to form the final feature vector space. In the feature vector space, the K nearest neighbor set is calculated for each data point, and the mutual proximity degree is calculated based on the K nearest neighbor set.

[0020] Draw a density-distance decision map, identify cluster centers, cluster data points in the feature vector space based on the cluster centers, find the unassigned data point with the highest proximity to the assigned data point, and assign the point to the cluster to which the assigned data point belongs. Set a dynamic threshold to dynamically adjust the cluster centers when the data distribution changes.

[0021] A dynamic memory matrix is ​​constructed based on the number of dynamic cluster centers, feature dimensions, and time sliding window. A weighted covariance matrix is ​​constructed by introducing time decay weight, cluster correlation weight, and feature importance weight. Multivariate state estimation modeling is performed on the dynamic memory matrix based on the weighted covariance matrix to construct a normal behavior model of elderly users under normal conditions.

[0022] In this scheme, the physiological and motor characteristics are imported into the normal behavior model for multi-level anomaly detection, specifically as follows:

[0023] The real-time extracted physiological and motion features are standardized, and an observation vector is constructed based on the standardized feature vector. The observation vector is then used to find the best match in the dynamic memory matrix.

[0024] Calculate the Mahalanobis distance between the observation vector and each cluster center, select the clusters corresponding to the top preset number of cluster centers with the smallest distance as the candidate set, and match the dynamic memory matrix based on the candidate set;

[0025] Load the pre-trained weighted covariance matrix, weight the dynamic memory matrix to obtain the estimated vector, calculate the residual vector between the observed vector and the estimated vector according to the feature dimension, obtain the standardized residual vector, and obtain the comprehensive anomaly score by combining the feature importance weights.

[0026] A multi-level anomaly detection mechanism is set up. In the primary detection, an independent threshold is set to mark the dimensions whose comprehensive anomaly scores exceed the threshold. In the intermediate detection, anomaly combinations of different dimensions are identified and the combination of dimensions whose comprehensive anomaly scores exceed the threshold is marked.

[0027] In advanced detection, a sliding window is constructed to calculate the linear change trend and mutation point generation characteristic trend in different dimensions, assess the persistence of anomalies, and identify anomaly patterns based on the anomaly persistence assessment results.

[0028] In this scheme, the multi-level anomaly detection results are classified to obtain anomaly categories, and graded early warnings are generated based on the anomaly categories to produce corresponding early warning information. Specifically:

[0029] The standardized anomaly scores of single-dimensional anomaly features in the primary detection, the multi-dimensional combined features in the intermediate detection, and the temporal features in the advanced detection are integrated. The integrated features are then combined with environmental parameters, elderly users’ prior health knowledge, and smart cane status information to perform contextual feature fusion.

[0030] An anomaly classification model is constructed by integrating a multimodal classifier. Adversarial training is introduced to train the anomaly classification model. XGBoost is used to process structured numerical features, 1D-CNN is combined to process temporal anomaly patterns, and an attention mechanism is used to fuse multimodal features. Anomaly categories are identified based on the multimodal features.

[0031] Based on the abnormality category and the degree of abnormality, a graded early warning is issued. Based on the graded early warning results, corresponding multi-channel early warning information is generated. The changes in physiological parameters after the early warning are continuously tracked, and the recovery of movement patterns is analyzed to verify the early warning.

[0032] In this solution, fuzzy voice interaction is triggered, and the abnormal state and real-time needs of elderly users are obtained based on the fuzzy voice recognition results. Specifically:

[0033] When an anomaly is detected, a fuzzy voice interaction will be triggered. A preset dialogue template will be loaded according to the current anomaly category and the corresponding warning information. The current ambient noise level will be assessed and the speaker volume will be automatically adjusted.

[0034] Multi-level voice prompts are generated based on the preset dialogue template. A voice feature library for elderly users is constructed by analyzing their acoustic characteristics, pronunciation variations, and communication habits. Sensitive keywords for different abnormal categories are set, and sensitive keywords are prioritized for matching during voice interaction.

[0035] A semantic parsing model is constructed based on U-Net. The speech feature library and sensitive keywords of different abnormal categories are used for enhanced training. The fuzzy speech interaction sequence of elderly users is used as the model input. Symptom descriptions are extracted from the speech through sensitive keyword matching, and the degree of urgency is judged through speech features. The abnormal state of elderly users is generated according to the symptom descriptions and the degree of urgency.

[0036] The abnormal state is fused with sensor data and enhanced with context interpretation. The enhanced abnormal state is then obtained and mapped to a standard demand category to identify the real-time needs of elderly users.

[0037] In this solution, a response strategy is obtained based on the aforementioned early warning information, the abnormal status of elderly users, and their real-time needs. Specifically:

[0038] A multidimensional emergency resource knowledge graph is constructed based on the multidimensional emergency resource nodes and the relationships between various emergency resources. In the multidimensional emergency resource knowledge graph, the early warning information, abnormal status and real-time needs of elderly users are located. With the goal of minimizing the overall response time and maximizing resource adaptability, Dijkstra's critical path identification is used to obtain the optimal response path for different modalities.

[0039] The response strategy is obtained based on the optimal response path for different modalities, the execution of the response strategy is tracked, and dynamic updates of arrival time are provided for elderly users.

[0040] The second aspect of the present invention provides an intelligent cane early warning system based on multi-parameter physiological monitoring. The system includes a multi-parameter sensing and preprocessing unit, a multi-level anomaly detection unit, an anomaly early warning classification unit, a fuzzy voice interaction unit, and an emergency response unit.

[0041] The multi-parameter sensing and preprocessing unit collects physiological parameters and motion data of elderly users through various sensors, preprocesses the physiological parameters and motion data, and extracts physiological and motion features.

[0042] The multi-level anomaly detection unit collects normal state monitoring data of elderly users to construct a normal behavior model, and imports the physiological and motor characteristics into the normal behavior model for multi-level anomaly detection.

[0043] When an abnormal situation exists, the anomaly warning classification unit classifies the multi-level anomaly detection results in combination with the context to obtain the anomaly category, performs hierarchical warnings according to the anomaly category, and generates corresponding warning information.

[0044] The fuzzy voice interaction unit triggers fuzzy voice interaction when an early warning message is generated, and obtains the abnormal status and real-time needs of elderly users based on the fuzzy voice recognition results.

[0045] The emergency response unit acquires the early warning information, as well as the abnormal status and real-time needs of elderly users, and obtains a response strategy based on the spatiotemporal context. Based on the response strategy, it generates multi-party linkage information.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] This invention enables intelligent health monitoring for elderly users, significantly improving the safety and efficiency of their daily lives. It overcomes the limitations of traditional walking sticks with only single-function assistance, achieving collaborative monitoring of multimodal physiological parameters. A multi-level verification architecture significantly reduces false alarm rates, and a personal health baseline is established through multivariate state estimation model initialization, enabling accurate anomaly identification. A multi-level early warning system optimizes the emergency response process, and fuzzy voice interaction improves emergency response efficiency, greatly shortening the average time from an anomaly occurrence to receiving assistance. Furthermore, physiological parameters and activity trajectories are uploaded to a cloud platform, facilitating remote monitoring of health status by family members or medical personnel and providing data support for chronic disease management. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0049] Figure 1 A flowchart of an intelligent cane early warning method based on multi-parameter physiological monitoring is shown;

[0050] Figure 2 The flowchart illustrates the process of collecting monitoring data on the normal state of elderly users to construct a model of their normal behavior.

[0051] Figure 3 A flowchart is shown that obtains the abnormal status and real-time needs of elderly users based on fuzzy speech recognition results.

[0052] Figure 4 A block diagram of an intelligent cane early warning system based on multi-parameter physiological monitoring is shown. Detailed Implementation

[0053] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0054] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0055] Figure 1 A flowchart of an intelligent cane early warning method based on multi-parameter physiological monitoring is shown.

[0056] like Figure 1 As shown, this embodiment provides a smart cane early warning method based on multi-parameter physiological monitoring, including:

[0057] S102, collect physiological parameters and movement data of elderly users through multiple sensors, preprocess the physiological parameters and movement data, and extract physiological and movement features;

[0058] S104, collect monitoring data of normal status of elderly users to construct a normal behavior model, import the physiological characteristics and motor characteristics into the normal behavior model to perform multi-level anomaly detection, and classify the multi-level anomaly detection results to obtain the anomaly category.

[0059] S106, perform graded early warning according to the abnormality category, generate corresponding early warning information, verify the early warning information and trigger fuzzy voice interaction, and obtain the abnormal status and real-time needs of elderly users according to the fuzzy voice recognition results.

[0060] S108, obtain response strategies based on the warning information and the abnormal status and real-time needs of elderly users, record the actual situation after each warning information, and adjust the model feedback according to the actual situation.

[0061] It should be noted that the smart cane uses a sensor array integrated into the handle and embedded in the cane to sense the user's physiological parameters and motion data. This includes a photoplethysmography (PPG) sensor for real-time monitoring of heart rate and blood oxygen saturation; a bioimpedance sensor for detecting grip status and skin conductance; a temperature sensor for measuring hand and ambient temperature; a three-axis accelerometer combined with a gyroscope for detecting gait characteristics and fall detection; a barometer for monitoring altitude changes and providing fall prevention warnings; an ultrasonic sensor for detecting obstacles; and preferably, a multi-functional sensor embedded in the tip of the smart cane for environmental perception, including: ground humidity detection, a vibration sensor (for road surface smoothness analysis), and an ambient light sensor (for day / night mode switching).

[0062] Edge computing preprocessing is used, and a sliding window Z-score method is employed to identify and remove outliers in physiological parameters. Missing values ​​are filled using time series linear interpolation, and Kalman filtering is applied to eliminate sensor noise. Demand information is generated based on blood pressure monitoring, blood oxygen assessment, and fall detection. Based on this demand information, similarity calculation is used for retrieval, and historical instances that meet the similarity requirements are extracted. The physiological and motor features involved in the historical instances are used to construct an initial feature pool. For example, the candidate set of physiological features includes: 24-hour blood pressure variability coefficient, blood oxygen saturation decline slope, time-domain indices of heart rate variability (SDNN, RMSSD), and frequency-domain indices (LF, HF). The candidate set of motor features includes: gait symmetry index, gait variability, trunk tilt angle, peak fall impact force, and posture transition frequency.

[0063] The initial feature pool is hybrid-encoded, with each chromosome containing physiological feature segments, motor feature segments, and parameter weight segments. The genetic algorithm population is initialized, and a multi-objective fitness function is constructed based on the blood pressure prediction accuracy, blood oxygen anomaly detection rate, and fall detection timeliness. The blood pressure prediction accuracy is calculated using the mean absolute error after training the SVR model with selected features; the blood oxygen anomaly detection rate is obtained by calculating the F1-score on the validation set; and the average delay time from event occurrence to alarm is read as the fall detection timeliness. During the iteration process of the genetic algorithm, ten-fold cross-validation is used to calculate the three-dimensional fitness of each individual.

[0064] A tournament selection method is used to select a predetermined number of individuals as parents. Independent crossover and position-flipping mutations are performed on the physiological and motor characteristic segments of the selected parents, and Gaussian perturbation is applied to the parameter weights. After iterative training, the optimal individual is selected after meeting a predetermined termination condition. In the mutation process, forced mutations are performed on gene loci that have not improved for three consecutive generations. Key physiological and motor characteristic indicators are extracted from the optimal individual. The learned parameter weights are used to rank the importance of features, obtaining the final physiological and motor characteristic indicators. Based on these final indicators, physiological and motor features are extracted from preprocessed physiological parameters and motor data. An improved genetic algorithm achieves efficient feature construction for multi-parameter monitoring, meeting real-time requirements while ensuring accuracy, providing reliable feature support for the early warning function of the smart cane.

[0065] Figure 2 The flowchart illustrates the process of collecting monitoring data on the normal state of elderly users to construct a model of their normal behavior.

[0066] According to an embodiment of the present invention, a normal behavior model is constructed by collecting monitoring data on the normal state of elderly users, specifically as follows:

[0067] S202, acquire physiological parameters and motion data within a preset historical time step, extract normal state monitoring data of elderly users, and extract physiological and motion characteristics under normal state conditions from the normal state monitoring data;

[0068] S204, standardize the physiological and motor characteristics under normal conditions, perform nonlinear dimensionality reduction through t-SNE to form the final feature vector space, calculate the K nearest neighbor set for each data point in the feature vector space, and calculate the mutual proximity degree based on the K nearest neighbor set;

[0069] S206, Draw a density-distance decision map, identify cluster center points, cluster data points in the feature vector space based on the cluster center points, find the unassigned data point with the highest proximity to the assigned data points, assign the point to the cluster where the assigned data points are located, set a dynamic threshold, and dynamically adjust the cluster center when the data distribution changes;

[0070] S208. A dynamic memory matrix is ​​constructed based on the number of dynamic cluster centers, feature dimensions, and time sliding window. A weighted covariance matrix is ​​constructed by introducing time decay weight, cluster correlation weight, and feature importance weight. Multivariate state estimation modeling is performed on the dynamic memory matrix based on the weighted covariance matrix to construct a normal behavior model of elderly users under normal conditions.

[0071] It should be noted that the Euclidean distance between data points in the feature vector space is calculated, and the local density is calculated based on this Euclidean distance. After calculating the local density, the local density of each data point is sorted, and the relative distance is given to the data point with the highest density. The density peak needs to satisfy two conditions simultaneously: high local density and large relative distance. Therefore, the density-distance decision graph finds the density peak and identifies the cluster center. For each data point, the proximity relationship of its k nearest neighbors is calculated. Only points that are bidirectional nearest neighbors are included in the density calculation. The mutual proximity degree is calculated based on the K nearest neighbor set. The calculation formula is: in Representing data points The set of K nearest neighbors, Representing data points With data points European distance, The written test covers local scaling functions. Indicates an indicator function if and only if the data points Data points K nearest neighbors and data points Data points The K-nearest neighbor (KNN) value is 1 when the KNN is active and 0 otherwise. Preferably, a time decay factor can be introduced into the neighbor calculation to give recent data a higher density weight. An adaptive kernel bandwidth is used, and the scale parameter for density calculation is adjusted according to the local data distribution. During clustering, the nearest neighbor of the density peak is assigned to its corresponding cluster. The unassigned data point with the highest proximity to the assigned data point is found and assigned to the cluster of the assigned data point. If an unassigned data point exists, it is assigned to the nearest and highest-density assigned data point; otherwise, clustering ends. A triple weighting is introduced based on traditional multivariate state estimation: the time decay weight uses an exponential decay function to give recent data a higher weight; the cluster correlation weight is assigned based on the similarity between the data point and the cluster center; and the feature importance weight, combined with clinical prior knowledge, assigns different features different importance based on weights obtained from a genetic algorithm. A dynamic threshold for Mahalanobis distance is calculated based on the distribution of each cluster center, generating a dual judgment standard: a strict threshold for acute anomaly detection and a lenient threshold for monitoring chronic change trends.

[0072] The real-time extracted physiological and motion features are standardized. An observation vector is constructed based on the standardized feature vector. This observation vector is used to find the best match in the dynamic memory matrix. During the matching process, the center vectors of each candidate cluster are weighted by similarity and a time decay factor is added to enhance the influence of recent clusters. The Mahalanobis distance between the observation vector and each cluster center is calculated. The clusters corresponding to the top preset number of cluster centers with the smallest distance are selected as the candidate set. The dynamic memory matrix is ​​matched based on this candidate set. A pre-trained weighted covariance matrix is ​​loaded, and the dynamic memory matrix is ​​weighted to obtain the estimated vector. The residual vector between the observed vector and the estimated vector is calculated according to the feature dimension, and the standardized residual vector is obtained. A comprehensive anomaly score is obtained by combining the feature importance weights. A multi-level anomaly detection mechanism is set. In the primary detection, independent thresholds (such as heart rate, blood oxygen, gait variation) are set to mark dimensions whose comprehensive anomaly scores exceed the thresholds, and an instant voice prompt is triggered for the corresponding dimension's anomaly. In the intermediate detection, anomaly combinations of different dimensions are identified, such as increased heart rate + gait instability + A decrease in blood oxygen level corresponds to a risk of cardiac involvement. The combination of dimensions with an overall abnormal score exceeding the threshold is marked. In advanced detection, a sliding window is constructed to calculate the linear change trend and mutation point generation characteristic trend of different dimensions, assess the persistence of abnormalities such as transient fluctuations or continuous abnormalities, and identify abnormal patterns such as progressive deterioration and periodic abnormalities based on the abnormality persistence assessment results.

[0073] It should be noted that the standardized abnormal scores of single-dimensional abnormal features (such as heart rate exceeding the threshold) in the primary detection, the multi-dimensional combined features (such as the combined abnormal patterns of heart rate + blood oxygen + gait) in the intermediate detection, and the temporal features (such as dynamic characteristics such as duration and gradient of change) in the advanced detection are integrated. The integrated features are combined with environmental parameters (based on indoor / outdoor, day / night) to label the abnormal occurrence scenario, the elderly user's prior health knowledge (such as chronic disease history, commonly used medications), and the status information of the smart cane for contextual feature fusion. An abnormal classification model is constructed by integrating a multimodal classifier. The first layer of the abnormal classification model is used to identify the major abnormal categories (physiological / behavioral / environmental / equipment failure), the second layer is used to classify clinical subcategories (such as cardiovascular, respiratory, neurological, etc.), and the third layer is used to identify specific abnormal types (such as atrial fibrillation, hypoxemia, falls, etc.). Adversarial training is introduced to train the anomaly classification model. XGBoost is used to process structured numerical features, combined with 1D-CNN to process temporal anomaly patterns, and an attention mechanism is used to fuse multimodal features. Anomaly categories are identified based on these multimodal features. Graded warnings are then issued based on the anomaly category and the degree of anomaly (e.g., the degree of persistent tachycardia). Corresponding multi-channel warning information is generated based on the graded warning results. The changes in physiological parameters after the warning are continuously tracked, and the recovery of movement patterns is analyzed to verify the warning. Furthermore, spatiotemporal adaptive adjustments are used in the anomaly classification model, such as increasing fall detection sensitivity and reducing voice volume at night, enhancing positioning functionality outdoors, and extending the warning response timeout. Preferably, the multi-channel warning information includes: an emergency channel: automatically triggering an emergency call, including GPS location and key vital signs; a medical channel: generating a structured warning report and pushing it to contracted medical institutions; a family channel: sending graded alert information to designated relatives; and a user interface: conveying the risk level through the cane's LED color ring and vibration mode.

[0074] Figure 3 A flowchart is shown to obtain the abnormal status and real-time needs of elderly users based on the results of fuzzy speech recognition.

[0075] According to an embodiment of the present invention, fuzzy voice interaction is triggered, and the abnormal state and real-time needs of the elderly user are obtained based on the fuzzy voice recognition result, specifically as follows:

[0076] S302 will trigger fuzzy voice interaction when an abnormal situation is detected, load a preset dialogue template according to the current abnormality category and the corresponding warning information, and assess the current ambient noise level to automatically adjust the speaker volume;

[0077] S304, Generate multi-level voice prompts according to the preset dialogue template, construct a voice feature library of elderly users through the acoustic characteristics, pronunciation variations and communication habits of elderly users, set sensitive keywords of different abnormal categories, and prioritize matching sensitive keywords in voice interaction;

[0078] S306. Construct a semantic parsing model based on U-Net, use the speech feature library and sensitive keywords of different abnormal categories for enhanced training, take the fuzzy speech interaction sequence of elderly users as the model input, extract symptom descriptions from speech through sensitive keyword matching, judge the degree of urgency through speech features, and generate the abnormal state of elderly users according to the symptom descriptions and degree of urgency.

[0079] S308, the abnormal state is fused with sensor data and enhanced with context interpretation to obtain the context-enhanced abnormal state and map it to a standard demand category to identify the real-time needs of elderly users.

[0080] It should be noted that, based on the type of abnormality, possible dialogue templates are pre-loaded (cardiovascular abnormalities, fall risk, etc.), and user historical interaction records are retrieved to optimize the speech output speed and word choice habits. In the semantic parsing model, a speech feature library for the elderly is constructed, including typical pronunciation variations (such as dentalization and vowel confusion). Context-sensitive recognition technology is used to prioritize matching keywords such as dizziness and discomfort when blood pressure is abnormal, achieving anti-interference processing and effectively separating speech from environmental noise (such as television sound and water sound). Clinical elements such as pain description (location / nature / degree) and discomfort are extracted from the speech. The severity of the emergency is judged by speech features (speech speed, volume, and clarity). Sensor data is fused to interpret the speech content, mapping ambiguous expressions to standard demand categories (medical assistance, daily living assistance, and emotional comfort).

[0081] It should be noted that a multidimensional emergency resource knowledge graph is constructed based on multidimensional emergency resource nodes and the relationships between various emergency resources. The medical emergency system, community service network, and kinship are incorporated into the unified graph. Nodes include medical nodes, community service nodes, kinship nodes, environmental nodes, etc. The dynamic edge weights of the edge structure include: time dimension weights: different response time coefficients for weekdays / nighttimes; spatial dimension weights: arrival time prediction under the influence of real-time traffic conditions; capability dimension weights: medical institution specialty suitability (e.g., cardiology connecting with cardiovascular abnormalities); and relationship dimension weights: kinship historical response timeliness scores.

[0082] The system locates early warning information, abnormal states, and real-time needs of elderly users within the multi-dimensional emergency resource knowledge graph. With the goal of minimizing overall response time (discovery → decision-making → arrival) and maximizing resource suitability (specialist doctor matching), Dijkstra's critical path identification is used to screen feasible nodes within a 3km radius. Specialized resources are filtered by anomaly type, and the optimal path is calculated based on real-time status to obtain optimal response paths for different modalities. Response strategies are then derived based on these optimal response paths, such as those for clinical needs, daily living needs, and emotional needs. The execution of these response strategies is tracked, and dynamic updates on arrival times are provided to elderly users. Through intelligent resource scheduling and precise strategy generation, an emergency support system for the elderly covering the entire chain of "monitoring-early warning-response-review" is constructed. In actual deployment, the system particularly emphasizes the coordinated cooperation between medical resources and community services, avoiding over-reliance on emergency resources while ensuring professional handling of critical situations, achieving optimal allocation of limited care resources.

[0083] Record the actual situation after each warning, calculate the detection delay of various anomalies, statistically analyze the accuracy of each level of warning, record user response time and handling results to evaluate the warning effect, adjust the model feedback based on the actual situation, reassess the importance of features and include newly confirmed anomaly cases.

[0084] Figure 4 A block diagram of an intelligent cane early warning system based on multi-parameter physiological monitoring is shown.

[0085] The second embodiment of the present invention provides an intelligent cane early warning system 4 based on multi-parameter physiological monitoring, including: a multi-parameter sensing and preprocessing unit 401, a multi-level anomaly detection unit 402, an anomaly early warning classification unit 403, a fuzzy voice interaction unit 404, and an emergency response unit 405;

[0086] The multi-parameter sensing and preprocessing unit collects physiological parameters and motion data of elderly users through various sensors, preprocesses the physiological parameters and motion data, and extracts physiological and motion features.

[0087] The multi-level anomaly detection unit collects normal state monitoring data of elderly users to construct a normal behavior model, and imports the physiological and motor characteristics into the normal behavior model for multi-level anomaly detection.

[0088] When an abnormal situation exists, the anomaly warning classification unit classifies the multi-level anomaly detection results in combination with the context to obtain the anomaly category, performs hierarchical warnings according to the anomaly category, and generates corresponding warning information.

[0089] The fuzzy voice interaction unit triggers fuzzy voice interaction when an early warning message is generated, and obtains the abnormal status and real-time needs of elderly users based on the fuzzy voice recognition results.

[0090] The emergency response unit acquires the early warning information, as well as the abnormal status and real-time needs of elderly users, and obtains a response strategy based on the spatiotemporal context. Based on the response strategy, it generates multi-party linkage information.

[0091] A third aspect of the present invention provides a computer-readable storage medium including a smart cane early warning method program based on multi-parameter physiological monitoring. When the smart cane early warning method program based on multi-parameter physiological monitoring is executed by a processor, it implements the steps of the smart cane early warning method based on multi-parameter physiological monitoring.

[0092] In the several embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, and can be electrical, mechanical, or other forms. Furthermore, in the various embodiments of the present invention, all functional units can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0093] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0094] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent walking stick early warning method based on multi-parameter physiological monitoring, characterized in that, Includes the following steps: Physiological parameters and movement data of elderly users are collected through multiple sensors. The physiological parameters and movement data are preprocessed and physiological and movement features are extracted. Collect monitoring data of normal status of elderly users to construct a normal behavior model, import the physiological and motor characteristics into the normal behavior model to perform multi-level anomaly detection, and classify the multi-level anomaly detection results to obtain the anomaly category. Based on the abnormality category, a graded early warning is issued, corresponding early warning information is generated, the early warning information is verified and fuzzy voice interaction is triggered, and the abnormal status and real-time needs of elderly users are obtained based on the fuzzy voice recognition results. Response strategies are obtained by using the warning information and the abnormal status and real-time needs of elderly users. The actual situation after each warning information is recorded, and the model is adjusted based on the actual situation. Collect monitoring data on the normal state of elderly users to construct a normal behavior model, specifically: Obtain physiological parameters and motion data within a preset historical time step, extract normal state monitoring data of elderly users, and extract physiological and motion characteristics under normal state conditions from the normal state monitoring data; The physiological and motor characteristics under normal conditions are standardized and nonlinear dimensionality is reduced by t-SNE to form the final feature vector space. In the feature vector space, the K nearest neighbor set is calculated for each data point, and the mutual proximity degree is calculated based on the K nearest neighbor set. Draw a density-distance decision map, identify cluster centers, cluster data points in the feature vector space based on the cluster centers, find the unassigned data points with the highest proximity to the assigned data points, and assign the unassigned data points to the clusters where the assigned data points are located. Set a dynamic threshold to dynamically adjust the cluster centers when the data distribution changes. A dynamic memory matrix is ​​constructed based on the number of dynamic cluster centers, feature dimensions, and time sliding window. A weighted covariance matrix is ​​constructed by introducing time decay weight, cluster correlation weight, and feature importance weight. Multivariate state estimation modeling is performed on the dynamic memory matrix based on the weighted covariance matrix to construct a normal behavior model of elderly users under normal conditions. The physiological and motor characteristics are imported into the normal behavior model for multi-level anomaly detection, specifically as follows: The real-time extracted physiological and motion features are standardized, and an observation vector is constructed based on the standardized feature vector. The observation vector is then used to find the best match in the dynamic memory matrix. Calculate the Mahalanobis distance between the observation vector and each cluster center, select the clusters corresponding to the top preset number of cluster centers with the smallest distance as the candidate set, and match the dynamic memory matrix based on the candidate set; Load the pre-trained weighted covariance matrix, weight the dynamic memory matrix to obtain the estimated vector, calculate the residual vector between the observed vector and the estimated vector according to the feature dimension, obtain the standardized residual vector, and obtain the comprehensive anomaly score by combining the feature importance weights. A multi-level anomaly detection mechanism is set up. In the primary detection, an independent threshold is set to mark the dimensions whose comprehensive anomaly scores exceed the threshold. In the intermediate detection, anomaly combinations of different dimensions are identified and the combination of dimensions whose comprehensive anomaly scores exceed the threshold is marked. In advanced detection, a sliding window is constructed to calculate the linear change trend and mutation point generation characteristic trend in different dimensions, assess the persistence of anomalies, and identify anomaly patterns based on the anomaly persistence assessment results.

2. The multi-parameter physiology monitoring based smart cane early warning method of claim 1, wherein, Physiological parameters and movement data of elderly users are collected through multiple sensors. The physiological parameters and movement data are preprocessed, and physiological and movement features are extracted, specifically as follows: Based on the sensor group integrated in the handle of the smart cane and the sensor group embedded in the cane body to sense the user's physiological parameters and motion data, edge computing preprocessing is used, the sliding window Z-score method is used to identify and remove abnormal values ​​of physiological parameters, time series linear difference is used to fill missing values, and Kalman filtering is applied to eliminate sensor noise. Demand information is generated based on heart rate monitoring, blood oxygen assessment and fall detection. Based on the demand information, similarity calculation is used for retrieval, historical instances that meet the similarity requirements are extracted, and the physiological and motor features involved in the historical instances are used to construct an initial feature pool. The initial feature pool is mixed and encoded to initialize the population of the genetic algorithm. A multi-objective fitness function is constructed based on the heart rate prediction accuracy, blood oxygen abnormality recognition rate and fall detection timeliness. During the iteration process of the genetic algorithm, ten-fold cross-validation is used to calculate the three-dimensional fitness of each individual. A predetermined number of individuals are selected as parents using a tournament selection method. The physiological and motor characteristic segments of the selected parents are subjected to independent crossover and position flip mutations, and the parameter weight segments are perturbed by Gaussian. After iterative training, the optimal individual is selected after the predetermined termination condition is met. Based on the optimal individual, key physiological and motor feature indicators are extracted. The importance of the features is ranked using learned parameter weights to obtain the final physiological and motor feature indicators. Based on the final physiological and motor feature indicators, physiological and motor features are extracted from the preprocessed physiological parameters and motor data.

3. The multi-parameter physiology monitoring based smart cane early warning method of claim 1, wherein, The results of multi-level anomaly detection are classified to obtain anomaly categories. Based on these anomaly categories, graded early warnings are issued, and corresponding early warning information is generated. Specifically: The standardized anomaly scores of single-dimensional anomaly features in the primary detection, the multi-dimensional combined features in the intermediate detection, and the temporal features in the advanced detection are integrated. The integrated features are then combined with environmental parameters, elderly users’ prior health knowledge, and smart cane status information to perform contextual feature fusion. An anomaly classification model is constructed by integrating a multimodal classifier. Adversarial training is introduced to train the anomaly classification model. XGBoost is used to process structured numerical features, 1D-CNN is combined to process temporal anomaly patterns, and an attention mechanism is used to fuse multimodal features. Anomaly categories are identified based on the multimodal features. Based on the abnormality category and the degree of abnormality, a graded early warning is issued. Based on the graded early warning results, corresponding multi-channel early warning information is generated. The change trend of physiological parameters after the early warning is continuously tracked, and the recovery of movement patterns is analyzed to verify the early warning.

4. The multi-parameter physiology monitoring based smart cane early warning method of claim 1, wherein, Trigger fuzzy voice interaction, and obtain the elderly user's abnormal state and real-time needs based on the fuzzy voice recognition results, specifically: When an anomaly is detected, a fuzzy voice interaction will be triggered. Based on the current anomaly category and the corresponding warning information, a preset dialogue template will be loaded, and the current ambient noise level will be assessed to automatically adjust the speaker volume. Multi-level voice prompts are generated based on the preset dialogue template. A voice feature library for elderly users is constructed by analyzing their acoustic characteristics, pronunciation variations, and communication habits. Sensitive keywords for different abnormal categories are set, and sensitive keywords are prioritized for matching during voice interaction. A semantic parsing model is constructed based on U-Net. The speech feature library and sensitive keywords of different abnormal categories are used for enhanced training. The fuzzy speech interaction sequence of elderly users is used as the model input. Symptom descriptions are extracted from the speech through sensitive keyword matching, and the degree of urgency is judged through speech features. The abnormal state of elderly users is generated according to the symptom descriptions and the degree of urgency. The abnormal state is fused with sensor data and enhanced with context interpretation. The enhanced abnormal state is then obtained and mapped to a standard demand category to identify the real-time needs of elderly users.

5. The multi-parameter physiology monitoring based smart cane early warning method of claim 1, wherein, The response strategy is derived from the aforementioned early warning information and the abnormal status and real-time needs of elderly users, specifically as follows: A multidimensional emergency resource knowledge graph is constructed based on the multidimensional emergency resource nodes and the relationships between various emergency resources. The warning information, abnormal status and real-time needs of elderly users are located in the multidimensional emergency resource knowledge graph. With the goal of minimizing the overall response time and maximizing resource adaptability, Dijkstra's critical path identification is used to obtain the optimal response path for different modalities. The response strategy is obtained based on the optimal response path for different modalities, the execution of the response strategy is tracked, and dynamic updates of arrival time are provided for elderly users.

6. An intelligent walking stick early warning system based on multi-parameter physiological monitoring, characterized in that, The intelligent cane early warning method based on multi-parameter physiological monitoring as described in any one of claims 1-5 includes a multi-parameter sensing and preprocessing unit, a multi-level anomaly detection unit, an anomaly early warning classification unit, a fuzzy voice interaction unit, and an emergency response unit. The multi-parameter sensing and preprocessing unit collects physiological parameters and motion data of elderly users through various sensors, preprocesses the physiological parameters and motion data, and extracts physiological and motion features. The multi-level anomaly detection unit collects normal state monitoring data of elderly users to construct a normal behavior model, and imports the physiological and motor characteristics into the normal behavior model for multi-level anomaly detection. When an abnormal situation exists, the anomaly warning classification unit classifies the multi-level anomaly detection results in combination with the context to obtain the anomaly category, performs hierarchical warnings according to the anomaly category, and generates corresponding warning information. The fuzzy voice interaction unit triggers fuzzy voice interaction when an early warning message is generated, and obtains the abnormal status and real-time needs of elderly users based on the fuzzy voice recognition results. The emergency response unit acquires the early warning information, as well as the abnormal status and real-time needs of elderly users, and obtains a response strategy based on the spatiotemporal context. Based on the response strategy, it generates multi-party linkage information.

Images