An AI-based home-based elderly care abnormal behavior recognition and early warning method

Abstract

This invention discloses an AI-based method for identifying and warning of abnormal behaviors in home-based elderly care, relating to the field of intelligent elderly care technology. It collects home activity data of elderly people from the past week, preprocesses the data, and then uses a density clustering algorithm to divide the preprocessed data into several home scenarios. Feature extraction and analysis are performed on the preprocessed data in these scenarios to obtain a scenario baseline dataset. A detailed analysis is then performed on the scenario baseline dataset, setting multi-layered threshold indicators for multiple features in each home scenario. The key technical points are: achieving comprehensive, accurate, personalized, and privacy-preserving identification of abnormal behaviors in home-based elderly care, providing 24 / 7, non-intrusive, and highly reliable security for elderly people living at home. It possesses high practicality and applicability, is adaptable to various home-based elderly care scenarios, and can be widely applied to home safety monitoring for elderly people living alone or in empty nests, showing promising application prospects.

Description

Technical Field

[0001] This invention relates to the field of intelligent elderly care technology, specifically to an AI-based method for identifying and warning of abnormal behaviors in home-based elderly care. Background Technology

[0002] As my country's population continues to age rapidly, home-based elder care has become the mainstream model. Most elderly people prefer to spend their later years in the familiar environment of their home. Their home safety and health monitoring have become a core pain point of social concern. Due to the decline in physiological functions, the elderly are prone to abnormal situations such as falls, sudden cardiovascular and cerebrovascular diseases, and prolonged periods of silence. Moreover, most elderly people living at home lack real-time care. If abnormal situations are not detected and dealt with in a timely manner, they can easily lead to serious consequences. Therefore, it is of great significance to build an efficient and accurate home-based elder care abnormality identification and early warning system.

[0003] Current home-based elderly care monitoring technologies generally have significant limitations and are difficult to adapt to the needs of refined and personalized monitoring. Traditional monitoring methods are based on human companionship and emergency call devices. Among them, human companionship is costly and has limited coverage. Emergency call devices rely on the elderly to operate them actively. For the elderly who are confused or have difficulty moving, the warning mechanism cannot be triggered in time, which poses a serious safety hazard.

[0004] The existing patent application, publication number CN110349372B, entitled "Method and Device for Early Warning of Abnormal Activities of Home-based Elderly People," describes the following: collecting basic information of home-based elderly people; collecting multiple location information of home-based elderly people at multiple time points; determining the trajectory of home-based elderly people based on multiple location information, wherein the trajectory includes time information and dwell time information; determining behavioral events of home-based elderly people based on their trajectory, basic information, and map points of interest; and determining the behavioral events of home-based elderly people based on their trajectory and home-based elderly people's basic information. This invention identifies behavioral events of elderly people and establishes a network of related activity information for those living at home. Based on this network and the current movement of elderly people living at home, it determines whether their activities are abnormal. This invention can identify abnormal activity patterns of elderly people living at home in a very short time.

[0005] The solution described in the aforementioned patent has certain drawbacks. The solution mainly achieves non-contact detection in outdoor activity scenarios, but it suffers from problems such as a single information dimension and poor scenario adaptability. Therefore, it has the defects of single monitoring method, insufficient comfort, and inability to accurately determine anomalies in combination with the scenario. Moreover, it may result in a high rate of false alarms and false negatives.

[0006] In summary, existing methods and devices for early warning of abnormal activities among elderly people living at home do not meet market demands. Therefore, we propose an AI-based method for identifying and warning of abnormal behaviors in elderly people living at home. Summary of the Invention

[0007] To achieve the above objectives, the present invention provides the following technical solution: An AI-based method for identifying and issuing early warnings of abnormal behaviors in home-based elderly care includes: Data on elderly people's home activities over the past week was collected and preprocessed. Then, density clustering algorithm was used to divide the preprocessed data into several home scenarios. Feature extraction and analysis were performed on the preprocessed data in these scenarios to obtain a scenario baseline dataset. A detailed analysis was performed on the baseline dataset of the scenarios, and multi-level threshold indicators for multiple features under each home scenario were set. After summarizing, a scenario threshold system was constructed. Based on the scene threshold system and activity characteristics, a multi-dimensional anomaly judgment scheme is set up, and a routine monitoring strategy is implemented. The routine monitoring data is preprocessed, segmented and identified in sequence. Based on the identification results, the multi-dimensional anomaly judgment scheme is implemented to obtain a first-level comprehensive judgment result. A multi-dimensional analysis and evaluation system is constructed. When the first-level comprehensive judgment result is suspected anomaly, multi-dimensional behavior recognition and analysis are performed based on the pre-constructed multi-dimensional intelligent analysis and evaluation model to obtain the second-level comprehensive judgment result. The first-level comprehensive judgment result and the second-level comprehensive judgment result are fused together to obtain the fused judgment result; When the fusion judgment result is abnormal, an early warning message is generated and pushed to the warning object. The entire abnormal judgment process data is recorded in the historical database, and the most recent data in the historical database is periodically extracted to update the scenario threshold system and the multi-dimensional analysis and evaluation system.

[0008] Preferably, the home activity data of the elderly in the past week includes intelligent monitoring data, body posture data, and auxiliary data. Among them, intelligent monitoring data includes location data, posture change range, activity intensity, static duration, and body posture data collected by wearable sensors; body posture data includes the duration of static posture maintenance and the frequency of posture change; and auxiliary data includes daily routines and basic health data.

[0009] Preferably, the preprocessing of elderly people's home activity data includes: using a sliding window filtering algorithm to remove noise from outliers; using valid data from adjacent time periods to interpolate and complete missing values; and using the standard score method to normalize the interpolated and completed data. The preprocessed data on elderly people's home activities was divided using a density clustering algorithm. This included processing the location data using a trajectory segmentation algorithm based on minimum confidence interval and correlation coefficient to extract dwell points; performing cluster analysis on the dwell points using a density clustering algorithm; and combining this with the home space layout to perform scene segmentation, resulting in several home scenes.

[0010] Preferably, feature extraction of preprocessed data in several home scenarios includes: classifying the preprocessed data according to the home scenario to obtain a scenario dataset; refining the data in the scenario dataset using a combination of statistical analysis and machine learning feature selection to obtain a strongly correlated feature set; calculating the correlation between the features in the strongly correlated feature set and the normal behavior of the elderly using the mutual information method, removing redundant features to obtain a core feature set; and processing the core features in the core feature set and the scenario dataset using a behavioral baseline learning model to obtain the confidence intervals of the core features, which are then summarized to form a scenario baseline dataset.

[0011] Preferably, a detailed analysis is performed on the scenario baseline dataset, and multi-level threshold indicators for various home scenarios are set, including: establishing risk levels for home scenarios based on the probability of abnormal occurrence and risk level of each home scenario; retrieving the confidence intervals of each core feature in the scenario baseline dataset; and setting multi-level threshold indicators for multiple features based on the threshold setting strategy corresponding to the risk level of the home scenario. The multi-level threshold indicators for multiple features are the multi-level thresholds for core features. Core features include dwell time, activity intensity, body posture data, and low-activity periods. The multi-level thresholds include normal thresholds, suspected abnormal thresholds, and abnormal thresholds.

[0012] Preferably, the preprocessing, scene segmentation, and identification of routine monitoring data includes: processing the routine monitoring data using a sliding window filtering algorithm to remove noise; converting the triaxial acceleration data collected by the wearable sensor into activity intensity data using an activity intensity calculation formula; calibrating real-time posture data and removing outliers; establishing a correlation between location data, activity intensity data, and posture data and the home environment; drawing a movement trajectory based on the location data, and using a trajectory segmentation algorithm that integrates the minimum confidence interval and correlation coefficient to dynamically segment the elderly person's movement trajectory, while simultaneously determining stay and movement according to preset rules. Trajectory segmentation was completed to obtain the movement segment and dwell segment in the home scene; key dwell data was extracted from the movement segment and key movement data was extracted from the dwell segment; the key dwell data and key movement data were analyzed to perform anomaly detection and assess the anomaly confidence levels of dwell duration, activity intensity, and body posture data; the comprehensive confidence level of the first-level anomaly was calculated by weighted summation; and the anomaly confidence levels of dwell duration, activity intensity, body posture data, and the comprehensive confidence level of the first-level anomaly were analyzed using a multi-dimensional anomaly detection scheme to obtain the comprehensive first-level anomaly detection result.

[0013] Preferably, the first-level comprehensive judgment result includes normal, suspected abnormal and abnormal. When the first-level comprehensive judgment result is normal, no action is taken. When the first-level comprehensive judgment result is abnormal, an early warning message is generated and pushed to the warning object. At the same time, the camera in the corresponding home scene is activated to record in real time, and the data collection cycle of the wearable sensor is shortened.

[0014] Preferably, the multi-dimensional behavior recognition and analysis based on the pre-built multi-dimensional intelligent analysis and evaluation model includes: activating the camera corresponding to the home scene for real-time shooting and shortening the data collection cycle of wearable sensors; denoising, grayscale conversion, and target detection of the collected video to extract the elderly's human body contours and movement features, forming a visual feature vector; filtering body posture data to obtain a body posture feature set; aligning the visual feature vector, body posture feature set, and positioning data along the time axis, and then inputting them into the multi-dimensional intelligent analysis and evaluation model with the same level of anomaly comprehensive confidence; processing the input data using a combination of feature splicing and attention mechanisms to obtain a multi-dimensional comprehensive feature vector; and using a CNN-LSTM hybrid algorithm to analyze the multi-dimensional behavior based on a pre-built database of abnormal actions of elderly people living at home. The multi-dimensional comprehensive feature vector is used to perform continuous action sequence matching and feature comparison to obtain the action anomaly confidence and abnormal action type. A method combining threshold comparison and correlation analysis is used to process the multi-dimensional comprehensive feature vector, first-level anomaly comprehensive confidence, and abnormal action type to obtain the posture anomaly verification result and posture anomaly confidence. A method combining multi-dimensional threshold filtering and scene logic analysis is used to process the multi-dimensional comprehensive feature vector, action anomaly confidence, and posture anomaly confidence to obtain the false trigger confidence and corresponding cause. The false trigger confidence, action anomaly confidence, and posture anomaly confidence are weighted and calculated to obtain the second-level anomaly comprehensive confidence. The second-level anomaly comprehensive confidence is analyzed in conjunction with a preset anomaly judgment scheme for the corresponding home scenario to obtain the second-level comprehensive judgment result.

[0015] Preferably, the judgment fusion of the first-level comprehensive judgment result and the second-level comprehensive judgment result includes: weighting the second-level anomaly comprehensive confidence score and the first-level anomaly comprehensive confidence score, calculating the final confidence score, comparing the final confidence score with the preset anomaly classification threshold, and if the final confidence score is lower than the anomaly classification threshold, it is judged as a first-level false trigger, no warning is triggered, and anomaly files are recorded, and the scenario threshold system is optimized periodically; if the final confidence score is not lower than the anomaly classification threshold, it is judged as an anomaly.

[0016] Preferably, the periodic optimization of the scenario threshold system includes: filtering data from each home scenario in the anomaly archive within the past week, and counting the number of anomalies in each home scenario; obtaining the cause of the false trigger, and determining the category of the corresponding core feature based on the cause of the false trigger; calculating the adjustment range of the suspected anomaly threshold of the corresponding core feature based on the number of anomalies; and adjusting the suspected anomaly threshold of the core feature based on the adjustment range of the suspected anomaly threshold.

[0017] This invention provides an AI-based method for identifying and issuing early warnings of abnormal behaviors in home-based elderly care, which has the following beneficial effects: This invention combines positioning technology with sensor data acquisition technology to collect the data required for comprehensive judgment at the next level after conventional detection. It effectively solves the shortcomings of existing technologies that use a single detection method, collect only a single type of data, and cannot perform collaborative and comprehensive analysis. It realizes all-round monitoring of the elderly's home behavior and physical status. The collected monitoring data is more comprehensive in dimensions and has stronger data correlation, effectively avoiding the problem of anomaly omission in the single data monitoring mode, and providing solid data support for subsequent anomaly judgment work.

[0018] This invention combines scene segmentation and dynamic thresholding technologies to analyze behavioral and physical data of the elderly in different home scenarios, constructing a scenario-based baseline and dynamically setting three types of thresholds based on the risk level differences of each scenario. This effectively solves the problems of inconsistent thresholds and insufficient adaptability in existing technologies, achieving scenario-based adaptation of anomaly judgment standards. This method offers greater flexibility in threshold setting and more accurate anomaly judgment, enabling precise differentiation between normal and abnormal behaviors in different scenarios, effectively reducing false alarms and false negatives, and improving the reliability of anomaly judgment.

[0019] This invention employs a technology that combines primary comprehensive judgment of non-visual monitoring with precise video analysis in abnormal scenarios. This technology can effectively overcome the shortcomings of existing technologies where all-time visual monitoring infringes on the privacy of the elderly or lacks visual assistance in recognition, resulting in insufficient accuracy in anomaly judgment. It achieves dual optimization of privacy protection and anomaly recognition accuracy, avoiding the risk of privacy leakage caused by all-time visual collection, and relying on visual assistance in recognition to make up for the limitations of non-visual monitoring, thereby reducing the probability of misjudgment and missed judgment of anomalies.

[0020] This invention utilizes anomaly profile-oriented optimization technology and periodic update technology to effectively address the shortcomings of existing technologies, such as lack of self-learning and iterative capabilities, inability to adapt to dynamic changes in elderly people's behavioral habits and physical states, and decreased monitoring accuracy after long-term use. It enables autonomous iterative optimization of the solution, achieving continuous improvement in monitoring accuracy and enhanced adaptability. This ensures that the system can adapt to subtle changes in the behavior and physical state of the elderly for a long time, maintaining high monitoring accuracy even after long-term use, extending the system's lifespan, and reducing subsequent maintenance costs.

[0021] This invention comprehensively addresses the shortcomings of existing home-based elderly care monitoring patents, such as limited monitoring methods, insufficient privacy protection, poor scenario adaptability, and lack of self-learning ability. It achieves comprehensive, accurate, personalized, and privacy-preserving anomaly identification in home-based elderly care, providing 24 / 7, non-intrusive, and highly reliable security for seniors living at home. It has high practicality and applicability, is suitable for various home-based elderly care scenarios, and can be widely used for home safety monitoring of seniors living alone and empty nesters. It has good performance and promising application prospects. Attached Figure Description

[0022] Figure 1 This is a flowchart of an AI-based method for identifying and warning of abnormal behaviors in home-based elderly care according to the present invention. Figure 2 This is a flowchart of an AI-based method for identifying and warning of abnormal behaviors in home-based elderly care according to the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example: Please see Figure 1-2 This embodiment provides an AI-based method for identifying and issuing early warnings of abnormal behaviors in home-based elderly care, including: S1. Collect data on elderly people's home activities over the past week, perform preprocessing on the data, and then use density clustering algorithm to divide the preprocessed data into several home scenarios. Perform feature extraction and analysis on the preprocessed data in several home scenarios to obtain a scenario baseline dataset.

[0025] In actual use, the week can be adjusted according to the needs, such as 3 days or 30 days. 30 days will result in too much data to analyze and a long processing time, but the accuracy will be higher, while 3 days will result in a small amount of data to analyze and lower accuracy. Therefore, a week is used.

[0026] The data on elderly people's home activities in the past week includes smart monitoring data, posture data, and auxiliary data. Among them, smart monitoring data includes location data, posture change range, activity intensity, stillness duration, and posture data collected by wearable sensors; posture data includes the duration of posture maintenance and the frequency of posture changes in a still state; and auxiliary data includes daily routines and basic health data.

[0027] The attitude change amplitude is represented by triaxial acceleration data, and the conversion formula for activity intensity is: In the formula For activity intensity, The data includes triaxial acceleration data, where g is the acceleration due to gravity; body posture data includes blood pressure and pulse; and basic health data includes signs of discomfort in daily life, whether medications are being taken, etc., to facilitate more accurate setting of thresholds later.

[0028] The data on elderly people's home activities over the past week does not include periods of deep sleep at night when there is no significant activity. It only includes behaviors throughout the day, such as waking up, washing, eating, activities, and resting, to comprehensively capture the daily activity patterns of the elderly.

[0029] In the past week, data on the elderly’s home activities were obtained through a cloud platform or cloud-based system. All intelligent monitoring data was transmitted to the cloud platform via wearable sensors for easy access and backup.

[0030] Wearable sensors can utilize existing smart bracelets and other related devices.

[0031] Preprocessing of elderly people's home activity data includes: using a sliding window filtering algorithm to remove outliers and noise (with a window size of 5 seconds to retain valid data); using valid data from adjacent time periods to interpolate and complete missing values ​​(linear interpolation can also be used); and using standard scores to normalize the interpolated and completed data. Eliminating the influence of differences in the units of measurement across different data dimensions facilitates subsequent feature extraction.

[0032] The preprocessed data on elderly people's home activities was segmented using a density clustering algorithm. This included processing location data using a trajectory segmentation algorithm based on minimum confidence interval and correlation coefficient to extract dwell points. For example, if an elderly person stayed at a certain location for more than 3 minutes and the location fluctuation was less than 5 centimeters, the false dwell point was removed. The density clustering algorithm was then used to perform cluster analysis on the dwell points. Combined with the home space layout, scene segmentation was performed to obtain several home scenes, including bedrooms, living rooms, kitchens, bathrooms, corridors, and balconies.

[0033] Regarding the external environment, the first-level judgment method of this scheme is adopted. Since it is impossible to accurately collect video, the first-level judgment scheme can be directly adopted.

[0034] After dividing the home scene, the boundaries of the home scene can be calibrated and fine-tuned by combining the shooting range of the cameras installed in different scenes, which facilitates the subsequent secondary judgment.

[0035] Feature extraction for preprocessed data in several home scenarios includes: classifying the preprocessed data according to the home scenario to obtain a scenario dataset; this process removes transitional data during cross-scenario movement to reduce the impact of transitional data; refining the data in the scenario dataset using a combination of statistical analysis and machine learning feature selection to obtain a strongly correlated feature set; extracting features strongly correlated with the elderly's behavioral patterns in the scenario, specifically including dwell time features, activity features, movement pattern features, posture features, and posture-related features; calculating the correlation between the features in the strongly correlated feature set and the elderly's normal behavior using the mutual information method; removing redundant features; features with a correlation score lower than 0.3 are defined as redundant features to obtain a core feature set. The core feature set can be different for each home scenario. For example, in the bedroom, the core features include dwell time, posture maintenance time, and average pulse index; in the bathroom, the core features include dwell time, blood pressure fluctuation, and posture change frequency index. A behavioral baseline learning model is used to process the core features in the core feature set and the scenario dataset to obtain the confidence intervals of the core features. These are then summarized to form the scenario baseline dataset.

[0036] The stay characteristics cover the average daily stay duration, longest stay duration, shortest stay duration, and stay frequency in each home setting, reflecting the elderly's preference for different settings and stay patterns within each setting; the activity characteristics include the average daily activity intensity, variance of activity intensity variation, and distribution of active periods in each setting, reflecting the differences in the elderly's activity capacity and daily routines in different settings; the movement characteristics include the frequency of transitions between settings, transition time, and regular movement paths, reflecting the elderly's movement patterns between settings, and also relating to the entry and exit time characteristics of each setting; the posture characteristics include the average blood pressure, average pulse range, variance of blood pressure and pulse fluctuations, range of posture maintenance duration, and range of posture change frequency in each setting, combined with the functional characteristics of each setting, reflecting the focus of posture monitoring in each setting; the posture correlation characteristics refer to the differences in posture parameters of the same elderly person in different home settings.

[0037] Home settings can be designed individually based on the specific circumstances.

[0038] A behavioral baseline learning model is adopted, and the following processing flow is performed on the core features in the core feature set and the scene dataset: The scene baseline datasets of each home scene are input into the model input layer to construct an independent feature matrix for each scene; the data quality of each scene is further optimized through the model data preprocessing layer to remove potential redundant information and strengthen scene-specific features; relying on the shallow neural network of the feature extraction layer, the deep features of the elderly's activities in each scene are extracted, focusing on exploring the intrinsic relationship between body posture features, activity features, and dwell features in each scene; using the Softmax regression algorithm of the output layer, the deep features of each scene are classified and mapped, and the normal behavioral parameter range of the elderly in each scene is output; combined with the calculation formula of the normal range of the behavioral baseline, the 95% confidence interval of each behavioral parameter in each scene is calculated.

[0039] The behavioral baseline learning model includes an input layer, a temporal feature extraction layer, a clustering and distribution modeling layer, a threshold learning layer, and an adaptive output update layer. The input layer receives the scenario baseline datasets for each home scenario, performs normalization processing to eliminate differences in different data dimensions and units, and synchronously completes data axis alignment to ensure consistent data collection time. For missing data, linear interpolation is used to fill in the missing data to avoid the impact of missing data on subsequent modeling. The processed data is then input into the temporal feature extraction layer.

[0040] After receiving the data, the temporal feature extraction layer uses a one-dimensional convolutional neural network to extract local features from the temporal data, capturing key local features such as sudden changes in activity intensity, fluctuations in body parameters, and sudden changes in dwell time. Redundant and stable data are removed, and core local features related to the behavioral baseline are retained. The local features extracted by the one-dimensional convolutional neural network are then used to perform temporal association learning through a long short-term memory network to capture the temporal patterns of the elderly’s behavior and body posture. Multi-dimensional temporal core feature vectors are generated and associated with scene labels to clarify the home scenes corresponding to various features, and then output to the clustering and distribution modeling layer.

[0041] After receiving the data, the clustering and distribution modeling layer uses the K-Means clustering algorithm to cluster the time-series core feature vectors in the same scene, grouping similar behaviors and body posture features into the same category and removing extreme abnormal features. The Gaussian mixture method is used to model the distribution of each feature after clustering, analyze the distribution patterns of each feature, calculate the mean, variance and 95% confidence interval of each feature, and focus on constructing the normal distribution range of dwell time, activity intensity and body posture parameters to fit the behavioral habits of elderly individuals, thus obtaining a scene-specific feature distribution model, which is then fed into the threshold learning layer.

[0042] After receiving data, the threshold learning layer uses a linear regression algorithm to fit the calculation formulas for normal threshold, suspected abnormal threshold, and abnormal threshold based on the mean and variance of the following feature distributions, ensuring that the thresholds are highly matched with the individual feature distributions of the elderly. The strictness of the thresholds is adjusted according to the risk level of the scene. For example, the abnormal threshold for the length of stay in the bathroom is 20% lower than that in the bedroom. The softmax regression algorithm is used to normalize the thresholds to ensure the logical coherence of the three types of thresholds and avoid the judgment error caused by threshold overlap. This results in a set of three types of thresholds that are scene-specific and personalized. At the same time, the baseline of the behavior pattern for each scene is output to the adaptive update output layer.

[0043] The adaptive update output layer extracts features from newly added normal behavior data, updates the mean and variance of the feature distribution model, and gradually adapts to subtle changes in the elderly's behavioral habits. Based on false trigger data, it fine-tunes the suspected abnormal thresholds for corresponding scenarios to reduce the probability of false triggers. Based on real abnormal data, it verifies the rationality of the thresholds, and if there are any missed detections, it fine-tunes the strictness of the abnormal thresholds. A fixed update cycle is set to avoid the instability of the baseline caused by frequent updates, ensuring that the baseline can adapt to changes in the elderly's behavior in a timely manner. Finally, it generates and outputs the updated scenario-specific and personalized behavior baselines and threshold sets.

[0044] This refers to the updating of the scenario baseline dataset and the scenario threshold system, which will be described later.

[0045] The normal range of behavioral baseline [a, b] is calculated using the formula a = Xp - Zf × S / K 0.5 b = Xp + Zf × S / K 0.5 In the formula, a is the lower limit of the confidence interval of a certain core feature, b is the upper limit of the confidence interval of a certain core feature, Xp is the average value of a week's data for a certain core feature, S is the standard deviation of a week's data for a certain core feature, K is the number of samples for a certain core feature, and Zf is the standard normal distribution quantile corresponding to the 95% confidence interval.

[0046] After the baseline calculation is completed, the data from the last day of the scenario can be selected for verification. If the data fit is less than 90%, the feature selection criteria and model parameters for the scenario should be readjusted, and the modeling operation should be performed again.

[0047] This invention combines positioning technology with sensor data acquisition technology to collect the data required for comprehensive judgment at the next level after conventional detection. It effectively solves the shortcomings of existing technologies that use a single detection method, collect only a single type of data, and cannot perform collaborative and comprehensive analysis. It realizes all-round monitoring of the elderly's home behavior and physical status. The collected monitoring data is more comprehensive in dimensions and has stronger data correlation, effectively avoiding the problem of anomaly omission in the single data monitoring mode, and providing solid data support for subsequent anomaly judgment work.

[0048] S2. Perform detailed analysis on the scenario baseline dataset, set multi-level threshold indicators for multiple features under each home scenario, and then construct a scenario threshold system by summarizing the results.

[0049] A detailed analysis was performed on the baseline dataset of the scenarios, and multi-level threshold indicators for multiple features under each home scenario were set. These included: establishing risk levels for home scenarios based on the probability of anomalies and the degree of risk in each home scenario. For example, the bathroom is prone to falls, so the risk level is high, while other home scenarios have medium to low risk levels; retrieving the confidence intervals of each core feature in the baseline dataset of the scenarios; and setting multi-level threshold indicators for multiple features based on the threshold setting strategy corresponding to the risk level of the home scenario. The multi-level threshold indicators for multiple features are the multi-level thresholds for core features. Core features include dwell time, activity intensity, body posture data, and low-activity periods. The multi-level thresholds include normal thresholds, suspected abnormal thresholds, and abnormal thresholds.

[0050] For example, the normal threshold for dwell time is [at, bt], where [at, bt] represents the baseline normal range for dwell time. Here, t mainly corresponds to dwell time and has no other meaning. The abnormal threshold is (ct, +∞), and the suspected abnormal threshold is between bt and ct, including ct, where ct = bt + k1 × (xtp - zt). k1 is the abnormality coefficient, which is 0.45 for low and medium levels and 0.65 for high levels, and can be adjusted accordingly. xtp is the average dwell time of the elderly in a certain home scenario. zt is the minimum dwell time of the elderly in a certain home scenario. xtp and zt are different for elderly people in different home scenarios and are set based on the collected data.

[0051] Low-activity periods are the core features after processing. For example, the normal threshold for low-activity periods is the time period obtained from the analysis. For instance, the high-activity period for a bedroom is 22:00-7:00, and the corresponding low-activity period for a bedroom is 7:00-22:00. The suspected anomaly threshold for low-activity periods is the range of the number of times the scene is entered within the low-activity period and the range of the duration of a single stay within the low-activity period. For example, the range of the number of times the scene is entered within the low-activity period is 3 to 5 times per hour, and the range of the duration of a single stay within the low-activity period is 60 to 90 minutes. If either of the two conditions of the suspected anomaly threshold is met, it is judged as suspected anomaly (not meeting the anomaly threshold). The anomaly threshold is that the number of times the scene is entered within the low-activity period is more than 5 times per hour, and the duration of a single stay within the low-activity period is more than 90 minutes. If either of the two conditions of the anomaly threshold is met, it is judged as an anomaly. Meeting both conditions is the same as meeting only one condition.

[0052] Low-activity periods overlap with other core characteristics, therefore they are not included in the calculation of the first-level anomaly comprehensive confidence level. Their judgment effect is equivalent to that of the first-level comprehensive judgment, that is, if they are judged as normal, no action is taken; if they are judged as suspected anomalies, the second-level comprehensive judgment is triggered; if they are judged as anomalies, an early warning is triggered.

[0053] This key feature can serve as a reminder when the elderly are sleeping in other home settings besides the bedroom, preventing them from catching a cold and thus improving management effectiveness. In practical applications, this core feature can be removed.

[0054] For example, the normal threshold for activity intensity is [ah, bh], where [ah, bh] represents the baseline normal range for activity intensity. Here, h mainly corresponds to activity intensity and has no other meaning. The abnormal threshold is (ch, +∞), and the suspected abnormal threshold is between bh and ch, including ch = bh + k2 × (xhp - zh). k2 is the abnormality coefficient, which is 0.6 for low and medium levels and 0.8 for high levels, and can be adjusted accordingly. xhp is the average activity intensity of the elderly in a certain home scenario. zh is the minimum activity intensity set for the elderly in a certain home scenario. xhp and zh are different for different home scenarios for the elderly, and are set based on the collected data.

[0055] For example, the normal threshold for body posture data is [as, bs], where [as, bs] represents the baseline normal range of body posture data. Here, 's' mainly corresponds to body posture data and has no other meaning. The abnormal threshold is (cs, +∞), and the suspected abnormal threshold is between bs and cs, including cs = bs + k3 × (xsp - zs). k3 is the abnormality coefficient, which is 0.7 for low and medium levels and 0.9 for high levels, and can be adjusted accordingly. xsp is the average value of the elderly's body posture data in a certain home scenario. zs is the minimum body posture data of the elderly in a certain home scenario. The xsp and zs are different for different home scenarios of the elderly, and are set based on the collected data. However, the normal threshold, abnormal threshold, and suspected abnormal threshold for body posture data are all two items, namely blood pressure and pulse. If either of the two intervals of the suspected abnormal threshold is met, it is judged as suspected abnormal (not meeting the abnormal threshold). If either of the two intervals of the abnormal threshold is met, it is judged as abnormal. Meeting both is the same as meeting only one.

[0056] After summarizing the above thresholds, a scene threshold system is constructed.

[0057] This invention combines scene segmentation and dynamic thresholding technologies to analyze behavioral and physical data of the elderly in different home scenarios, constructing a scenario-based baseline and dynamically setting three types of thresholds based on the risk level differences of each scenario. This effectively solves the problems of inconsistent thresholds and insufficient adaptability in existing technologies, achieving scenario-based adaptation of anomaly judgment standards. This method offers greater flexibility in threshold setting and more accurate anomaly judgment, enabling precise differentiation between normal and abnormal behaviors in different scenarios, effectively reducing false alarms and false negatives, and improving the reliability of anomaly judgment.

[0058] S3. Based on the scene threshold system and activity characteristics, a multi-dimensional anomaly judgment scheme is set up, and a normal monitoring strategy is implemented. The data of normal monitoring is preprocessed, segmented and identified in sequence. Based on the identification results, the multi-dimensional anomaly judgment scheme is implemented to obtain the first-level comprehensive judgment result.

[0059] The data types for routine monitoring are the same as those analyzed in S2, but the collected data are real-time data. Specifically, the collected data include location data, attitude change amplitude, activity intensity, static duration, and body posture data. Based on the real-time coordinates and movement trajectory derived from the location data, edge computing technology is used to transmit the collected data to the edge computing terminal for easier calculation.

[0060] The preprocessing, scene segmentation, and identification of routine monitoring data include: processing the routine monitoring data using a sliding window filtering algorithm to remove noise (the sliding window is set to 3 seconds for rapid anomaly removal); converting the triaxial acceleration data collected by the wearable sensor into activity intensity data using the activity intensity calculation formula (see formula S1); calibrating real-time body posture data and removing outliers; establishing a correlation between location data, activity intensity data, and body posture data and the home scene, i.e., classifying the data according to the home scene; drawing movement trajectories based on location data, and using a trajectory segmentation algorithm that integrates minimum confidence interval and correlation coefficient (existing technology) to dynamically segment the elderly person's movement trajectory, while simultaneously completing trajectory segmentation according to preset stop and move judgment rules to obtain the movement segments and stop segments of the home scene. The pre-defined rules for determining dwell time and movement are as follows: if an elderly person stays at a certain location for more than 3 minutes and the fluctuation range of the location is less than 5 centimeters, that location is determined as a dwell point, and a dwell time segment is defined based on this. The remaining time periods are determined as movement segments. Key dwell time data is extracted from movement segments in the home environment, and key movement data is extracted from dwell time segments in the home environment. The key dwell time data and key movement data are analyzed to perform anomaly detection, and anomaly confidence levels for dwell time anomalies, activity intensity anomalies, and body posture data anomalies are assessed. The overall confidence level of the first-level anomaly is calculated by weighted summation. The anomaly confidence levels for dwell time anomalies, activity intensity anomalies, body posture data anomalies, and the overall confidence level of the first-level anomalies are analyzed in conjunction with a multi-dimensional anomaly detection scheme to obtain the overall first-level anomaly detection result.

[0061] Key data on stay include the start time, end time, home environment, and duration of stay, which are synchronously linked to the activity intensity and physical data within that home environment. Key data on movement includes the time of entry and exit from the environment, the starting and ending points of movement, and the time taken to move, which are synchronously linked to the movement period, the activity intensity and pulse data within the corresponding home environment. This data is used to determine abnormal movement patterns and compares the movement patterns with the elderly person's usual time for that movement pattern to determine if there are any deviations in movement patterns or abnormal physical characteristics.

[0062] To facilitate subsequent analysis, the trajectory segmentation results, related data, home environment, and corresponding body posture data of the elderly are recorded in real time to construct a daily scenario-based activity trajectory and body posture profile for the elderly.

[0063] The confidence scores for anomalies in dwell time, activity intensity, and body posture data were all calculated using a linear quantization algorithm. Specifically: Dwell time anomaly confidence score = (Tss - bt) / bt; Tss is the continuous dwell time at a single location. If the dwell time anomaly confidence score is < 0, it is set to 0; if the dwell time anomaly confidence score is > 1, it is set to 1. Activity intensity anomaly confidence score = (hqs - bh) / bh; hqs is the real-time activity intensity. If the activity intensity anomaly confidence score is < 0, it is set to 0; if the activity intensity anomaly confidence score is > 1, it is set to 1. If the activity intensity abnormality confidence score is >1, then it is set to 1. The abnormality confidence score of body posture data includes calculating the abnormality confidence scores of blood pressure and pulse separately. Similarly, if the abnormality confidence scores of blood pressure and pulse are less than 0, then it is set to 0; if the abnormality confidence scores of blood pressure and pulse are greater than 1, then it is set to 1. Then, the abnormality confidence score of body posture data is calculated by weighting. Under normal circumstances, the weights of abnormality confidence scores of blood pressure and pulse are both 0.5. If the elderly person has a corresponding defect, such as hypertension, then the weight of abnormality confidence score of blood pressure is increased.

[0064] The overall confidence level of the first-level anomaly is calculated by weighted summation: Overall Confidence Level of First-Level Anomaly = Confidence Level of Stay Duration Anomaly × Weight of Stay Duration Anomaly + Confidence Level of Activity Intensity Anomaly × Weight of Activity Intensity Anomaly + Confidence Level of Postural Data Anomaly × Weight of Postural Data Anomaly; where the weight of postural data anomaly is greater than the weight of activity intensity anomaly, which is greater than the weight of stay duration anomaly. For example, the weight of postural data anomaly is 0.6, the weight of activity intensity anomaly is 0.25, and the weight of stay duration anomaly is 0.15. This is combined with a multi-dimensional anomaly detection scheme to assess the confidence levels of stay duration anomaly, activity intensity anomaly, and postural data anomaly. The reliability and overall confidence level analysis for Level 1 anomalies are performed as follows: if the confidence levels for anomalies in duration, activity intensity, and body shape data are all 0, the condition is considered normal; if any one of these three indicators is 1, the condition is considered abnormal. The calculation of the overall confidence level for Level 1 anomalies is only initiated when all three indicators are between 0 and 1. If the overall confidence level for Level 1 anomalies is higher than the set anomaly threshold, the condition is considered abnormal; only when the overall confidence level for Level 1 anomalies is not higher than the set anomaly threshold is the condition considered suspected abnormal.

[0065] The first-level comprehensive judgment result includes normal, suspected abnormal, and abnormal. When the first-level comprehensive judgment result is normal, no action is taken. When the first-level comprehensive judgment result is abnormal, an early warning message is generated and pushed to the warning object. At the same time, the corresponding home scene camera is activated to record in real time, and the data collection cycle of wearable sensors is shortened.

[0066] The method for determining low-activity periods in S2 is the same as the method for processing the results of the first-level comprehensive determination.

[0067] S4. Construct a multi-dimensional analysis and evaluation system. When the first-level comprehensive judgment result is suspected abnormal, multi-dimensional behavior recognition and analysis are performed based on the pre-constructed multi-dimensional intelligent analysis and evaluation model to obtain the second-level comprehensive judgment result. Based on a pre-built multi-dimensional intelligent analysis and evaluation model, multi-dimensional behavior recognition and analysis are performed, including: waking up the camera corresponding to the home scene for real-time shooting and shortening the data collection cycle of wearable sensors; cameras in other home scenes are normally not woken up unless there is an overlap in the camera's shooting image. When the elderly person is located at the overlap in the camera's shooting image, cameras in other scenes will be activated. The data collection cycle of wearable sensors is shortened, that is, the collection frequency is shortened, adjusting the original 20 seconds-3 minutes of collecting first-level data to 1-10 seconds of collecting first-level data; noise reduction is performed on the collected video. Grayscale conversion and target detection are used to extract the elderly person's body contour and movement features to form a visual feature vector. To ensure accuracy, a step of calibrating the positioning data can be added to eliminate positioning deviations. The body posture data is filtered, mainly to remove outliers caused by device fluctuations, to obtain a body posture feature set. The visual feature vector, body posture feature set, and positioning data are aligned along the time axis, and then the comprehensive confidence score of the same level of anomalies is input into the multi-dimensional intelligent analysis and evaluation model. The multi-dimensional intelligent analysis and evaluation model is an AI model, which includes an input layer, a feature fusion layer, an anomaly recognition layer, an anomaly verification layer, a mis-touch exclusion layer, and an output layer.

[0068] The input layer is mainly used to input visual feature vectors, first-level anomaly comprehensive confidence scores, body posture feature sets, and localization data. The feature fusion layer uses a combination of feature concatenation and attention mechanisms to process the input data and obtain a multi-dimensional comprehensive feature vector. Specifically, the visual feature vectors, localization data, and body posture feature sets are first processed to unify the feature dimensions. Then, weights are allocated through the attention mechanism (with body posture features and visual action features having a higher weight ratio), strengthening the features corresponding to abnormal actions such as falling and sitting still, weakening environmental interference features, and finally generating a fused multi-dimensional comprehensive feature vector.

[0069] The anomaly recognition layer is based on a pre-set database of abnormal actions of elderly people living at home. It uses a CNN-LSTM hybrid algorithm to perform continuous action sequence matching and feature comparison on multi-dimensional comprehensive feature vectors to obtain the confidence level of the action anomaly and the type of the abnormal action. Specifically, after feature comparison, it combines the location data to judge the rationality of the scene in which the action occurs, and finally outputs the confidence level of the action anomaly and the type of the abnormal action.

[0070] The anomaly verification layer uses a combination of threshold comparison and correlation analysis to process multi-dimensional comprehensive feature vectors, first-level anomaly comprehensive confidence, and abnormal action types to obtain posture anomaly verification results and posture anomaly confidence. Specifically, the method of combining threshold comparison and correlation analysis compares the posture data in the comprehensive feature vector with the corresponding scene thresholds and posture data in the first-level judgment stage, while analyzing the correlation between posture parameters and abnormal action types to verify the authenticity of the first-level posture anomaly judgment results. Finally, it outputs posture anomaly verification results and posture anomaly confidence, where the posture anomaly verification results are divided into two categories: verification passed and verification failed.

[0071] The accidental triggering elimination layer uses a combination of multi-dimensional threshold filtering and scene logic analysis to process multi-dimensional comprehensive feature vectors, action anomaly confidence, and posture anomaly confidence to obtain the accidental trigger confidence and corresponding cause. Specifically, it uses a combination of multi-dimensional threshold filtering and scene logic analysis to compare visual action features with the elderly's normal activity action templates, posture data with the normal posture range, and positioning trajectory with the regular activity trajectory to identify accidental triggering scenarios, calculate the accidental trigger probability, and finally output the accidental trigger confidence and corresponding accidental trigger cause.

[0072] The output layer performs a weighted calculation of the confidence scores for false triggers, abnormal actions, and abnormal postures to obtain a comprehensive confidence score for secondary anomalies. The comprehensive confidence score for secondary anomalies is then analyzed in conjunction with a preset anomaly judgment scheme for the corresponding home scenario to obtain a comprehensive judgment result for secondary anomalies.

[0073] The confidence level of the secondary anomaly is analyzed by combining the preset anomaly judgment scheme for the corresponding home scenario to obtain the secondary comprehensive judgment result. That is, the confidence level of the secondary anomaly is compared with the preset secondary anomaly judgment threshold. If the confidence level of the secondary anomaly is greater than the preset secondary anomaly judgment threshold, the secondary comprehensive judgment result is judged to be abnormal. If the confidence level of the secondary anomaly is not greater than the preset secondary anomaly judgment threshold, the secondary comprehensive judgment result is judged to be normal.

[0074] The output layer also outputs the abnormal action type, abnormal posture verification result, and reason for false triggering, which facilitates subsequent analysis.

[0075] The camera can also take pictures in burst mode. In this case, the video processing can be changed to image processing.

[0076] This invention employs a technology that combines primary comprehensive judgment of non-visual monitoring with precise video analysis in abnormal scenarios. This technology can effectively overcome the shortcomings of existing technologies where all-time visual monitoring infringes on the privacy of the elderly or lacks visual assistance in recognition, resulting in insufficient accuracy in anomaly judgment. It achieves dual optimization of privacy protection and anomaly recognition accuracy, avoiding the risk of privacy leakage caused by all-time visual collection, and relying on visual assistance in recognition to make up for the limitations of non-visual monitoring, thereby reducing the probability of misjudgment and missed judgment of anomalies.

[0077] S5. Perform a judgment fusion on the first-level comprehensive judgment result and the second-level comprehensive judgment result to obtain the fused judgment result.

[0078] The judgment fusion process for the first-level and second-level comprehensive judgment results includes: weighting the comprehensive confidence scores of the second-level and first-level anomalies to calculate the final confidence score; comparing the final confidence score with the preset anomaly classification threshold; if the final confidence score is lower than the anomaly classification threshold, it is judged as a first-level false trigger, no warning is triggered, and anomaly files are recorded, and the scenario threshold system is optimized periodically; if the final confidence score is not lower than the anomaly classification threshold, it is judged as an anomaly.

[0079] Regularly optimizing the scenario threshold system includes: filtering data from each home scenario in the anomaly archive within the past week and counting the number of anomalies in each home scenario; obtaining the reasons for false triggers and determining the category of the corresponding core feature based on the reasons for false triggers; calculating the adjustment range of the suspected anomaly threshold for the corresponding core feature based on the number of anomalies; and adjusting the suspected anomaly threshold of the core feature based on the adjustment range of the suspected anomaly threshold.

[0080] For example: Based on the cause of the false trigger, the category of the corresponding core feature is determined to be activity intensity. For each abnormal occurrence, the adjustment range of the comprehensive confidence of the first-level abnormality increases by 1%. If there are 5 abnormal occurrences, the adjustment range of the suspected abnormality threshold of activity intensity is 5%, and then the suspected abnormality threshold of activity intensity is increased by 5%.

[0081] In actual use, if there is no camera, or the camera is obstructed or malfunctioning, the first-level comprehensive judgment result will be used directly, and the reason why the second-level comprehensive judgment result was not activated will be recorded simultaneously.

[0082] This invention utilizes anomaly profile-oriented optimization technology and periodic update technology to effectively address the shortcomings of existing technologies, such as lack of self-learning and iterative capabilities, inability to adapt to dynamic changes in elderly people's behavioral habits and physical states, and decreased monitoring accuracy after long-term use. It enables autonomous iterative optimization of the solution, achieving continuous improvement in monitoring accuracy and enhanced adaptability. This ensures that the system can adapt to subtle changes in the behavior and physical state of the elderly for a long time, maintaining high monitoring accuracy even after long-term use, extending the system's lifespan, and reducing subsequent maintenance costs.

[0083] S6. When the fusion judgment result is abnormal, generate early warning information and push it to the early warning object, record the entire abnormal judgment process data to the historical database, and periodically extract the most recent data from the historical database to perform scene threshold system updates and multi-dimensional analysis and evaluation system updates.

[0084] When in use, the secondary comprehensive judgment result is set to start periodically. It does not participate in real-time fusion decision-making, but only collects data and marks its validity. This data is used to adjust the scene threshold system and optimize the multi-dimensional intelligent analysis and evaluation model, and optimize the weight of the weighted calculation of the output layer. At this time, if the collected data shows that the elderly's behavior pattern has changed significantly, a prompt message is pushed to the family members, suggesting that they pay attention to the elderly's physical condition.

[0085] When the fusion judgment result is abnormal, an early warning information is generated and pushed to the warning target via SMS, APP push, or voice call to relevant personnel. Relevant personnel can be set, such as family members, community medical staff, and emergency contact personnel. The content pushed includes the time of the abnormality, the home scenario, the type of abnormality, the severity, and the current physical condition data of the elderly.

[0086] The full-process data refers to all data and source data generated in steps S3-S5.

[0087] Periodically extract the most recent data from the historical database, and perform scene threshold system updates and multi-dimensional analysis and evaluation system updates once a week. Statistically analyze the data of this week and compare it with the scene baseline dataset of last week to determine whether there have been significant changes in the behavior patterns of the elderly. If the change exceeds 20%, trigger the scene baseline dataset update. At this time, it is necessary to rebuild the scene threshold system based on the new scene baseline dataset.

[0088] In S5, the scenario threshold system optimization is performed periodically, while in S6, the most recent data is periodically extracted from the historical database. The scenario threshold system update and multi-dimensional analysis and evaluation system update are performed at intervals. For example, the scenario threshold system optimization is performed periodically on Mondays in S5, and on Thursdays in S6, to avoid duplication.

[0089] In addition, during use, the scenario baseline dataset, scenario threshold system and multi-dimensional intelligent analysis and evaluation model can be verified based on one month's data every month to ensure that the false alarm rate is reduced by more than 10% and the false negative rate is controlled within 5% after optimization; if the accuracy of anomaly judgment decreases after optimization, the parameters before optimization can be rolled back immediately.

[0090] This invention comprehensively addresses the shortcomings of existing home-based elderly care monitoring patents, such as limited monitoring methods, insufficient privacy protection, poor scenario adaptability, and lack of self-learning ability. It achieves comprehensive, accurate, personalized, and privacy-preserving anomaly identification in home-based elderly care, providing 24 / 7, non-intrusive, and highly reliable security for seniors living at home. It has high practicality and applicability, is suitable for various home-based elderly care scenarios, and can be widely used for home safety monitoring of seniors living alone and empty nesters. It has good performance and promising application prospects.

[0091] Example 2 Compared to Example 1, this example describes an AI-based method for identifying and warning of abnormal behaviors in home-based elderly care, including: Data on elderly people's home activities over the past week was collected and preprocessed. Then, density clustering algorithm was used to divide the preprocessed data into several home scenarios. Feature extraction and analysis were performed on the preprocessed data in these scenarios to obtain a scenario baseline dataset. A detailed analysis was performed on the baseline dataset of the scenarios, and multi-level threshold indicators for multiple features under each home scenario were set. After summarizing, a scenario threshold system was constructed. Based on the scene threshold system and activity characteristics, a multi-dimensional anomaly judgment scheme is set up, and a routine monitoring strategy is implemented. The routine monitoring data is preprocessed, segmented and identified in sequence. Based on the identification results, the multi-dimensional anomaly judgment scheme is implemented to obtain a first-level comprehensive judgment result. When the first-level comprehensive judgment result is abnormal, an early warning message is generated and pushed to the warning object. The entire abnormal judgment process data is recorded in the historical database, and the most recent data in the historical database is periodically extracted to update the scenario threshold system and the multi-dimensional analysis and evaluation system.

[0092] Compared to Example 1, this scheme reduces the number of secondary comprehensive judgments. Although the accuracy is slightly reduced, the amount of data processed is also reduced. Moreover, by adopting a scenario-based approach, the accuracy is still improved compared to existing technologies.

[0093] Example 3 Compared to Example 1, this example adds facial recognition technology to the secondary comprehensive judgment, while other steps remain unchanged; Specifically, the changes to the multi-dimensional intelligent analysis and evaluation model are as follows: the optimized multi-dimensional intelligent analysis and evaluation model includes an input layer, a feature fusion layer, an anomaly recognition layer, a face recognition layer, an anomaly verification layer, a false touch exclusion layer, and an output layer. The input layer is mainly used to input facial feature vectors, visual feature vectors, first-level anomaly comprehensive confidence, body feature set and localization data; the facial feature vector is extracted by sequentially performing denoising, face detection and feature extraction on the facial images of the elderly people collected, eliminating interference factors such as lighting and shooting angle. The feature fusion layer processes the input data using a combination of feature concatenation and attention mechanisms to obtain a multi-dimensional comprehensive feature vector. Specifically, it first performs unified feature dimension processing on facial feature vectors, visual feature vectors, localization data, and body posture feature sets. Then, it allocates weights through an attention mechanism (with body posture features, visual motion features, and facial features having higher weights), strengthening the features corresponding to abnormal actions such as falling and sitting still, as well as abnormal expressions such as pain, coma, and agitation, while weakening environmental interference features and irrelevant facial features, ultimately generating a fused multi-dimensional comprehensive feature vector. The anomaly recognition layer is based on a pre-set database of abnormal actions of elderly people living at home. It uses a CNN-LSTM hybrid algorithm to perform continuous action sequence matching and feature comparison on multi-dimensional comprehensive feature vectors to obtain the confidence level of the action anomaly and the type of the abnormal action. Specifically, after feature comparison, it combines the location data to judge the rationality of the scene in which the action occurs, and finally outputs the confidence level of the action anomaly and the type of the abnormal action. The facial recognition layer inputs multi-dimensional comprehensive feature vectors and abnormal action types to enhance facial recognition and expression analysis functions: Facial recognition uses a deep learning facial feature matching algorithm to compare the collected facial feature vectors with a preset facial template for the elderly, confirming that the monitored object is the elderly person and avoiding misjudgments caused by unauthorized personnel entering the premises; Expression analysis, based on facial feature vectors, uses a CNN expression recognition algorithm to identify the elderly person's current expression category, covering calm, pain, coma, agitation, and expressionlessness, distinguishing between normal and abnormal expressions, identifying expressions closely related to abnormal states such as coma and pain, and outputting the confidence level and specific expression type of the expression abnormality. Simultaneously, it correlates abnormal action types with body posture data to analyze the correlation between expressions and abnormal states. For example, a fall combined with a pain expression is judged as a true abnormality; sitting still combined with an expressionless face and an abnormal pulse is judged as a suspected abnormality.

[0094] The anomaly verification layer employs a combination of threshold comparison and correlation analysis to process multi-dimensional comprehensive feature vectors, first-level anomaly comprehensive confidence, and abnormal action types, thereby obtaining the posture anomaly verification results and posture anomaly confidence. Specifically, the method combines threshold comparison and correlation analysis to compare the posture data in the comprehensive feature vector with the corresponding scene thresholds and posture data in the first-level judgment stage. At the same time, it analyzes the correlation between posture parameters and abnormal action types and expression types to verify the authenticity of the first-level posture anomaly judgment results. Finally, it outputs the posture anomaly verification results and posture anomaly confidence, where the posture anomaly verification results are divided into two categories: verification passed and verification failed. The accidental trigger exclusion layer employs a method combining multi-dimensional threshold filtering and scene logic analysis to process multi-dimensional comprehensive feature vectors, action anomaly confidence, facial expression anomaly confidence, and body posture anomaly confidence, thereby obtaining the false trigger confidence and corresponding cause. Specifically, by comparing visual action features with the elderly's normal activity action template, facial expressions with the elderly's normal expression range, body posture data with the normal body posture range, and positioning trajectory with conventional activity trajectory, the layer identifies false trigger scenarios, calculates the false trigger probability, and finally outputs the false trigger confidence and corresponding false trigger cause. The output layer performs weighted calculations on the confidence scores of false triggers, facial expression anomalies, action anomalies, and body posture anomalies to obtain a comprehensive confidence score for secondary anomalies. The comprehensive confidence score for secondary anomalies is then analyzed in conjunction with a preset anomaly judgment scheme for the corresponding home scenario to obtain a comprehensive judgment result for secondary anomalies.

[0095] The above-mentioned method, which combines facial recognition, has greatly improved accuracy and is suitable for elderly people who are at higher risk and live alone. Its disadvantage is that it exposes the face and requires a lot of data to be analyzed, but its corresponding advantage is that it can accurately monitor the elderly person's condition and has high precision.

[0096] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0097] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0098] The above description is merely a specific embodiment of this application, but the scope of protection of this application 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 this application should be included within the scope of protection of this application.

Claims

1. An AI-based method for identifying and issuing early warnings of abnormal behaviors in home-based elderly care, characterized in that, include: We collected data on elderly people's home activities over the past week, preprocessed the data, and used a density clustering algorithm to divide the preprocessed data into several home scenarios. We then extracted and analyzed the preprocessed data for each scenario to obtain a baseline dataset. A detailed analysis was performed on the baseline dataset of the scenarios, and multi-level threshold indicators for multiple features under each home scenario were set. After summarizing, a scenario threshold system was constructed. Based on the scene threshold system and activity characteristics, a multi-dimensional anomaly judgment scheme is set up, and a routine monitoring strategy is implemented. The routine monitoring data is preprocessed, segmented and identified in sequence. Based on the identification results, the multi-dimensional anomaly judgment scheme is implemented to obtain a first-level comprehensive judgment result. A multi-dimensional analysis and evaluation system is constructed. When the first-level comprehensive judgment result is suspected anomaly, multi-dimensional behavior recognition and analysis are performed based on the pre-constructed multi-dimensional intelligent analysis and evaluation model to obtain the second-level comprehensive judgment result. The first-level comprehensive judgment result and the second-level comprehensive judgment result are fused together to obtain the fused judgment result; When the fusion judgment result is abnormal, an early warning message is generated and pushed to the warning object. The entire abnormal judgment process data is recorded in the historical database, and the most recent data in the historical database is periodically extracted to update the scenario threshold system and the multi-dimensional analysis and evaluation system.

2. The AI-based method for identifying and warning of abnormal behaviors in home-based elderly care according to claim 1, characterized in that: The data on elderly people's home activities in the past week includes smart monitoring data, posture data, and auxiliary data. Among them, smart monitoring data includes location data, posture change range, activity intensity, stillness duration, and posture data collected by wearable sensors; posture data includes the duration of posture maintenance and the frequency of posture changes in a still state; and auxiliary data includes daily routines and basic health data.

3. The AI-based method for identifying and warning of abnormal behaviors in home-based elderly care according to claim 1, characterized in that: Preprocessing of elderly people's home activity data includes: using a sliding window filtering algorithm to remove outliers and noise; using valid data from adjacent time periods to interpolate and complete missing values; and using the standard score method to normalize the interpolated and completed data. The preprocessed data on elderly people's home activities was divided using a density clustering algorithm. This included processing the location data using a trajectory segmentation algorithm based on minimum confidence interval and correlation coefficient to extract dwell points; performing cluster analysis on the dwell points using a density clustering algorithm; and combining this with the home space layout to perform scene segmentation, resulting in several home scenes.

4. The AI-based method for identifying and warning of abnormal behaviors in home-based elderly care according to claim 3, characterized in that: Feature extraction for preprocessed data in several home scenarios includes: classifying the preprocessed data according to the home scenario to obtain a scenario dataset; refining the data in the scenario dataset using a combination of statistical analysis and machine learning feature selection to obtain a strongly correlated feature set; calculating the correlation between the features in the strongly correlated feature set and the normal behavior of the elderly using the mutual information method, removing redundant features to obtain a core feature set; and using a behavioral baseline learning model to process the core features in the core feature set and the scenario dataset to obtain the confidence intervals of the core features, which are then summarized to form the scenario baseline dataset.

5. The AI-based method for identifying and warning of abnormal behaviors in home-based elderly care according to claim 1, characterized in that: A detailed analysis was performed on the baseline dataset of the scenarios, and multi-level threshold indicators for various home scenarios were set. These included: establishing risk levels for home scenarios based on the probability of anomalies and the degree of risk in each scenario; retrieving the confidence intervals of each core feature in the baseline dataset of the scenarios; and setting multi-level threshold indicators for multiple features based on the threshold setting strategy corresponding to the risk level of the home scenarios. The multi-level threshold indicators for multiple features are the multi-level thresholds for core features. Core features include dwell time, activity intensity, body posture data, and low-activity periods. The multi-level thresholds include normal thresholds, suspected abnormal thresholds, and abnormal thresholds.

6. The AI-based method for identifying and warning of abnormal behaviors in home-based elderly care according to claim 5, characterized in that: The preprocessing, scene segmentation, and identification of routine monitoring data include: processing routine monitoring data using a sliding window filtering algorithm to remove noise; converting triaxial acceleration data collected by wearable sensors into activity intensity data; calibrating posture data and removing outliers; establishing a correlation between location data, activity intensity data, and posture data and the home scene; drawing movement trajectories based on location data, and using a trajectory segmentation algorithm that integrates minimum confidence intervals and correlation coefficients to dynamically segment the elderly person's movement trajectory, while simultaneously completing trajectory segmentation according to preset stay and movement judgment rules to obtain home scene movement segments and stay segments; extracting key stay data from home scene movement segments and key movement data from home scene stay segments; performing anomaly judgment on key stay and key movement data, and evaluating the confidence levels of stay duration anomalies, activity intensity anomalies, and posture data anomalies; calculating the first-level anomaly comprehensive confidence level through weighted summation; and combining a multi-dimensional anomaly judgment scheme to analyze the confidence levels of stay duration anomalies, activity intensity anomalies, posture data anomalies, and the first-level anomaly comprehensive confidence level to obtain the first-level comprehensive judgment result.

7. The AI-based method for identifying and warning of abnormal behaviors in home-based elderly care according to claim 6, characterized in that: The first-level comprehensive judgment result includes normal, suspected abnormal, and abnormal. When the first-level comprehensive judgment result is normal, no action is taken. When the first-level comprehensive judgment result is abnormal, an early warning message is generated and pushed to the warning object. At the same time, the corresponding home scene camera is activated to record in real time, and the data collection cycle of wearable sensors is shortened.

8. The AI-based method for identifying and warning of abnormal behaviors in home-based elderly care according to claim 7, characterized in that: Multi-dimensional behavior recognition and analysis are performed based on a pre-built multi-dimensional intelligent analysis and evaluation model, including: It can activate cameras in the corresponding home environment for real-time shooting and shorten the data collection cycle of wearable sensors; The acquired video was denoised, converted to grayscale, and target detected. The human body contour and motion features of the elderly were extracted to form a visual feature vector. The body posture data is filtered to obtain a body posture feature set; The visual feature vectors, body feature sets, and positioning data are aligned on the time axis and the comprehensive confidence of the same level of anomalies is input into the multi-dimensional intelligent analysis and evaluation model. The input data is processed by a combination of feature concatenation and attention mechanism to obtain a multi-dimensional comprehensive feature vector. Based on a pre-set database of abnormal actions of elderly people living at home, a CNN-LSTM hybrid algorithm is used to perform continuous action sequence matching and feature comparison on multi-dimensional comprehensive feature vectors to obtain the confidence level of abnormal actions and the type of abnormal actions. A method combining threshold comparison and correlation analysis was used to process multi-dimensional comprehensive feature vectors, first-level anomaly comprehensive confidence, and abnormal action types to obtain the postural anomaly verification results and postural anomaly confidence. A method combining multi-dimensional threshold filtering and scene analysis is used to process multi-dimensional comprehensive feature vectors, action anomaly confidence, and posture anomaly confidence to obtain false trigger confidence and corresponding causes; The confidence scores for false triggering, abnormal movement, and abnormal posture are weighted and calculated to obtain the comprehensive confidence score for secondary abnormality. By combining the preset anomaly judgment scheme for the corresponding home scenario, the comprehensive confidence level of the second-level anomaly is analyzed to obtain the comprehensive judgment result of the second-level anomaly.

9. The AI-based method for identifying and warning of abnormal behaviors in home-based elderly care according to claim 1, characterized in that: The judgment fusion process for the first-level and second-level comprehensive judgment results includes: weighting the comprehensive confidence scores of the second-level and first-level anomalies to calculate the final confidence score; comparing the final confidence score with the preset anomaly classification threshold; if the final confidence score is lower than the anomaly classification threshold, it is judged as a first-level false trigger, no warning is triggered, and anomaly files are recorded, and the scenario threshold system is optimized periodically; if the final confidence score is not lower than the anomaly classification threshold, it is judged as an anomaly.

10. The AI-based method for identifying and warning of abnormal behaviors in home-based elderly care according to claim 9, characterized in that: Regularly optimizing the scenario threshold system includes: filtering data from each home scenario in the anomaly archive within the past week and counting the number of anomalies in each home scenario; obtaining the reasons for false triggers and determining the category of the corresponding core feature based on the reasons for false triggers; calculating the adjustment range of the suspected anomaly threshold for the corresponding core feature based on the number of anomalies; and adjusting the suspected anomaly threshold of the core feature based on the adjustment range of the suspected anomaly threshold.

Images