Intelligent health monitoring alarm system

Abstract

The invention discloses an intelligent health monitoring alarm system, and the system comprises an environment behavior monitoring module which is used for collecting environment and behavior data related to daily activities of a user, and the environment and behavior data comprises water data, power data, gas data, image data, environment sound data and refrigerator door state change data; wherein the environment sound data particularly refers to a playing sound source signal which is acquired by an indoor microphone and is generated by a television or audio playing equipment; the data processing center is used for receiving and storing the environment and behavior data; and the user habit learning module is used for establishing and dynamically updating a personalized daily behavior baseline model of the user through a machine learning algorithm based on historical data. According to the invention, through multi-dimensional non-contact monitoring, the influence of contact equipment on the self-respect of the elderly can be avoided, and more comprehensive and intelligent health monitoring alarm of the elderly can be monitored at the same time.

Description

Technical Field

[0001] This invention relates to the field of elderly care, and more specifically to an intelligent health monitoring alarm system. Background Technology

[0002] As the population ages, the number of elderly people living alone continues to expand, posing significant challenges to their daily care. On the one hand, elderly people living alone are prone to two core risks—safety and health—due to the absence of their children and declining mobility or memory. These risks are often delayed due to the elderly's reluctance to disclose information or their inability to express it clearly. On the other hand, traditional care methods have obvious limitations: manual care is costly and lacks real-time capability, making it difficult to cover 24-hour needs. While contact monitoring can collect some data, it can easily create a psychological burden of "being managed" for the elderly, damaging their self-esteem regarding independent living and leading to resistance to its use, thus limiting the actual care effectiveness.

[0003] Existing health monitoring alarm systems have been initially applied, but they still have many shortcomings and are difficult to meet the needs of accurate and humanized care. Specifically, they have problems such as single monitoring dimensions, incomplete risk coverage, rigid judgment standards, high false alarm rate, independent and non-integrated models, inaccurate risk judgment, and lack of coverage of non-visual scenes. Therefore, this invention proposes an intelligent health monitoring alarm system. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent health monitoring alarm system, comprising:

[0005] The environmental behavior monitoring module is used to collect environmental and behavioral data related to users' daily activities. The environmental and behavioral data include water consumption data, electricity consumption data, gas consumption data, video data, environmental sound data, and refrigerator door status change data.

[0006] Among them, ambient sound data specifically refers to the playback sound source signal generated by a television or audio playback device and collected through an indoor microphone;

[0007] The data processing center is used to receive and store environmental and behavioral data;

[0008] The user habit learning module is used to build and dynamically update a user's personalized daily behavior baseline model based on historical data and machine learning algorithms.

[0009] The data processing center is configured to call the baseline model generated by the user habit learning module and run multiple independent analysis models, including: a first analysis model, a second analysis model, a third analysis model, and a fourth analysis model;

[0010] The alarm generation module is used to generate and send corresponding monitoring alarm information when any analysis model outputs an anomaly determination result.

[0011] Furthermore, the first analysis model is used to perform correlation analysis on the time-series characteristics of electricity consumption data and water consumption data, calculate their behavioral coupling degree, and compare it with the first baseline threshold to determine whether the user's basic life activity pattern has deviated abnormally.

[0012] The specific method by which the first analysis model calculates the behavioral coupling degree is as follows:

[0013] Establish a power consumption sequence sampled at fixed time intervals and water flow sequence ;

[0014] Calculate the behavioral coupling degree CD of two sequences within a preset time window:

[0015] ;

[0016] in, and Let P and W be the means of sequences P and W within the time window T, respectively. and Let n be its standard deviation, where n is the total number of sampling points within the preset time window T;

[0017] The first baseline threshold is the dynamic coupling threshold. It is calculated by the user habit learning module based on the mean and standard deviation of historical behavior coupling data;

[0018] When CD remains below If the time exceeds the preset duration, the first analysis model will determine it as abnormal.

[0019] Furthermore, the user habit learning module determines the dynamic coupling threshold in the following manner. :

[0020] During the system learning period, the behavioral coupling degree at the same time each day was recorded. ;

[0021] Calculate all during the learning period mean and standard deviation ;

[0022] Then the dynamic coupling threshold Set to: , where f1 is a sensitivity coefficient adjusted based on the stability of user behavior.

[0023] Furthermore, the second analysis model is used to analyze image data to identify the duration of a user's time in the bathroom, and to perform cross-validation by combining the eating patterns reflected by the refrigerator door status change data, calculate the toilet-eating abnormality index, and compare it with the second baseline threshold to determine whether the user's physiological health condition is abnormal.

[0024] The specific calculation process of the second analytical model is as follows:

[0025] The system identifies user entry events into the restroom from video data, records the duration of each event, and calculates the average duration of each user entry into the restroom for the current day. ;

[0026] Extract the average opening and closing frequency during the current day's key eating periods from the refrigerator door status change data. ;

[0027] Calculate the toilet-eating abnormality index AEI :

[0028] ;

[0029] in, The current daily average duration and the historical baseline mean deviation, The standard deviation of historical duration data;

[0030] The current daily average switching frequency and the historical baseline average. deviation, The standard deviation of historical frequency data;

[0031] α and β As a weighting factor, and ;

[0032] Will AEI With the second baseline threshold In comparison, when When this happens, it is judged as abnormal;

[0033] Second baseline threshold ,in For history AEI mean f1 represents the historical AEI standard deviation, and f2 is the sensitivity coefficient adjusted based on the stability of the user's toilet-eating behavior.

[0034] Furthermore, the third analysis model is used to perform correlation analysis on the continuous opening status of gas usage data and the change data of refrigerator door status, calculate the risk coefficient, and compare it with the third baseline threshold to determine whether there is any abnormality in the safety of the user's cooking activities.

[0035] The specific calculation method of the third analytical model is as follows:

[0036] Real-time monitoring of gas flow data ,when When the stove is detected as being turned on, the safety timer is activated. , To determine the minimum flow threshold for turning on the stove;

[0037] exist If the refrigerator door is opened during the timer, the timer will reset. ;

[0038] Calculate the risk factor D:

[0039] ;

[0040] in, For the continuous, unreset timer, The average cooking time for users, a and b These are weighting coefficients;

[0041] Compare D with the third baseline threshold In comparison, when When this happens, it is judged as abnormal;

[0042] Third baseline threshold The value is calculated by the user habit learning module based on the mean and standard deviation of historical risk coefficient data. ,in The historical mean of D, f3 represents the historical standard deviation of D, and f3 is the sensitivity coefficient adjusted according to the cooking safety risk level.

[0043] Furthermore, the fourth analysis model is used to analyze the playback audio source signal. By analyzing its sound pressure level timing pattern and the consistency of the audio track content, it can determine whether the user's physical condition is abnormal.

[0044] The specific method for performing the analysis using the fourth analytical model is as follows:

[0045] Step S61: Preprocess the playback audio source signal to obtain the preprocessed signal, including noise reduction and filtering of non-media playback audio sources;

[0046] Step S62: Extract the sound pressure level time series data of the preprocessed signal. and audio spectrum time-series data;

[0047] Step S63: Process the sound pressure level time series data Perform a first-order difference operation to obtain the sound pressure level change sequence. and to Perform Z-score standardization to identify those with absolute values ​​greater than a threshold. The points are used as candidate mutation points;

[0048] Step S64: Perform parallel analysis on the consistency between the sound pressure level timing pattern and the audio track content, specifically including:

[0049] Mode Analysis A, Sound Pressure Level Timing Mode Analysis: Calculation of Sound Pressure Level linear regression slope within the sliding time window k ,like k It remains positive and its absolute value exceeds the slope threshold. If the time exceeds the preset duration, it is determined to be a mode where the volume continues to rise abnormally.

[0050] Mode Analysis A, Sound Pressure Level Timing Mode Analysis: Calculation of Sound Pressure Level linear regression slope within the sliding time window k ,like k It remains positive and its absolute value exceeds the slope threshold. If the time exceeds the preset duration, it is determined to be a mode where the volume continues to rise abnormally.

[0051] Step S65: Match the judgment result of the above pattern analysis A or pattern analysis B with the pre-stored risk feature library to obtain the matching degree. The risk feature library contains typical sound patterns associated with the user's physical discomfort.

[0052] Step S66: When the matching degree exceeds the risk feature threshold At that time, the fourth analysis model determines that the user's physical condition is abnormal;

[0053] Risk characteristic threshold The user habit learning module is trained based on historical matching data and matching samples corresponding to user physical discomfort events. ,in The mean matching score for each physical discomfort event. f1 represents the standard deviation of the matching degree corresponding to physical discomfort events, and f4 is the adjustment coefficient that balances the false negative rate and false positive rate of user physical status determination.

[0054] Furthermore, the alarm generation module is configured to execute a tiered alarm strategy, wherein the alarm level L in the tiered alarm strategy is determined by the following formula:

[0055] ;

[0056] in, These are the normalized anomaly values ​​output by the first, second, third, and fourth analysis models, respectively.

[0057] These are the weighting factors corresponding to each model;

[0058] Depending on the different ranges of alarm level L, the system triggers different levels of response measures, ranging from sending reminder messages to activating local alarms and even requesting external assistance.

[0059] Furthermore, the user habit learning module also includes a feedback optimization unit;

[0060] The feedback optimization unit receives feedback on the effectiveness of historical alarms;

[0061] The first baseline threshold is dynamically adjusted based on the feedback statistics. Second baseline threshold Third baseline threshold Risk characteristic threshold and weighting factors By continuously optimizing the above parameters, the false alarm rate of the system can be reduced.

[0062] The beneficial effects of this invention are reflected in:

[0063] The environmental behavior monitoring module collects various data, including water, electricity, gas, video, sound sources from indoor audio playback devices, and refrigerator door status, comprehensively covering core daily scenarios for elderly people living alone, such as cooking, washing, resting, and physiological activities. It can monitor the activity dynamics of elderly people living alone in real time, avoiding the omission of key safety or health hazards by a single monitoring dimension, and providing data support for subsequent accurate judgment. The user habit learning module is based on machine learning algorithms and combines the historical behavior data of elderly people living alone to establish and dynamically update a personalized daily behavior baseline model, rather than using a uniform standard. This design adapts to the diverse lifestyles of elderly people living alone, effectively reducing false alarms caused by standardized judgments and improving monitoring accuracy. The system addresses core risks for elderly people living alone through four independent analysis models: The first analysis model links electricity and water usage data, judging whether basic living patterns are abnormal through behavioral coupling, and can promptly detect reduced daily activities due to physical discomfort; the second analysis model combines toilet time and refrigerator opening / closing frequency to calculate the "toilet-eating" abnormality index, enabling early identification of physiological health problems and meeting the monitoring needs of common health risks among the elderly; the third analysis model monitors gas usage and links it to the refrigerator door, quickly determining cooking safety risks if the gas is on but the elderly have not operated the refrigerator for a long time, preventing accidents such as fires and gas leaks caused by poor memory in elderly people living alone; the fourth analysis model uses changes in sound pressure level of audio equipment to assist in judging the elderly's physical condition, supplementing risk identification in non-visual scenarios, and non-contact monitoring does not require the elderly to... People can change their existing lifestyles to cooperate with the monitoring work. The elderly do not need to wear, charge, or adjust the monitoring equipment regularly, nor do they need to actively report their health status or activity information. Daily activities such as washing, cooking, resting, and using television are not affected by the monitoring. Compared to traditional contact monitoring, this system does not require direct physical contact with the elderly, avoiding any feeling of physical constraint or psychological pressure of being forcibly managed. Especially for the elderly who have difficulty moving or are sensitive to physical contact, non-contact monitoring significantly reduces their resistance to the monitoring system and increases its long-term acceptance. Attached Figure Description

[0064] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0065] Picture 1 This is a system block diagram of the present invention. Detailed Implementation

[0066] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0067] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0068] like Picture 1 As shown, an intelligent health monitoring alarm system includes:

[0069] (1) Environmental behavior monitoring module, used to collect environmental and behavioral data related to users’ daily activities. The environmental and behavioral data include water consumption data, electricity consumption data, gas consumption data, video data, environmental sound data and refrigerator door status change data.

[0070] Among them, ambient sound data specifically refers to the playback sound source signal generated by a television or audio playback device and collected through an indoor microphone;

[0071] (2) Data processing center, used to receive and store environmental and behavioral data;

[0072] (3) User habit learning module, which is used to build and dynamically update a user's personalized daily behavior baseline model based on historical data and machine learning algorithms;

[0073] (4) The data processing center is configured to call the baseline model generated by the user habit learning module and run multiple independent analysis models, including: a first analysis model, a second analysis model, a third analysis model and a fourth analysis model;

[0074] (5) Alarm generation module, used to generate and send corresponding monitoring alarm information when any analysis model outputs an abnormal judgment result.

[0075] The first analysis model is used to perform correlation analysis on the time-series characteristics of electricity consumption data and water consumption data, calculate their behavioral coupling degree, and compare it with the first baseline threshold to determine whether there is an abnormal deviation in the user's basic life activity pattern.

[0076] The specific method by which the first analysis model calculates the behavioral coupling degree is as follows:

[0077] Establish a power consumption sequence sampled at fixed time intervals and water flow sequence ;

[0078] Calculate the behavioral coupling degree CD of two sequences within a preset time window:

[0079] ;

[0080] in, and Let P and W be the means of sequences P and W within the time window T, respectively. and Let n be its standard deviation, where n is the total number of sampling points within the preset time window T;

[0081] The first baseline threshold is the dynamic coupling threshold. It is calculated by the user habit learning module based on the mean and standard deviation of historical behavior coupling data;

[0082] When CD remains below If the time exceeds the preset duration, the first analysis model determines it to be abnormal;

[0083] Traditional monitoring often relies on single data points, which can easily lead to misjudgments of user activity status due to chance. However, the first analysis model performs correlation analysis on the time-series characteristics of electricity and water consumption data, calculates the behavioral coupling degree, and can capture the inherent relationship between the two types of data. This avoids misjudgments caused by the bias of a single data point, making the judgment of anomalies in basic life activity patterns more accurate.

[0084] Dynamic baseline thresholds adapt to individualized habits, reducing standardization interference. The first baseline threshold is not a fixed value, but is calculated by the user habit learning module based on the mean and standard deviation of historical behavior coupling data. For example, some users are accustomed to using electricity and water in the early morning (cooking and washing up), while others are accustomed to evening activities. Dynamic thresholds can fit these individualized patterns, avoiding the use of a uniform standard to measure all users, reducing the occurrence of normal habits being judged as abnormal, and improving user acceptance.

[0085] A stable pattern of basic daily activities is an important signal of good physical condition. Any abnormal deviation from this pattern is often an early sign of discomfort. The First Analysis Model, by monitoring these latent changes, can identify potential health risks before users explicitly express discomfort, allowing time for subsequent intervention.

[0086] Taking Grandma Li, an elderly woman living alone, as an example, the specific application process is as follows: During the system learning period, a personalized baseline is established. The system records Grandma Li's daily electricity consumption sequence from 7:00 to 8:00 AM (breakfast time) during a one-month learning period. (such as power data for rice cookers and electric kettles) and water flow rate sequences (e.g., water flow data for washing vegetables and brushing teeth), and calculate the daily behavioral coupling degree for that period. According to statistics, during the learning period mean =0.75 and standard deviation =0.12.

[0087] Because Grandma Li's daily behavior is highly stable, with a sensitivity coefficient f1 = 0.8, according to the dynamic coupling threshold formula: Calculated .

[0088] Anomaly Monitoring and Judgment: From Tuesday to Thursday of a certain week, Grandma Li was feeling weak due to a cold. Between 7:00 and 8:00 AM, she only occasionally used water (for simple rinsing) and did not cook breakfast. The system collected data during this period and calculated the daily behavioral coupling degrees as CD1=0.43, CD2=0.39, and CD3=0.41, all below the dynamic coupling threshold. =0.654, and the duration exceeds the preset 24 hours;

[0089] The first analytical model triggered the early warning and response, determining that Grandma Li's basic daily living activity pattern had deviated abnormally, and sent the result to the alarm generation module. Based on the tiered alarm strategy, the system sent an alert message to Grandma Li's children regarding the abnormality in her basic activities. Her children promptly visited her and discovered that Grandma Li had a cold but had not sought medical attention, thus preventing her condition from worsening.

[0090] The user habit learning module determines the dynamic coupling threshold in the following way. :

[0091] During the system learning period, the behavioral coupling degree at the same time each day was recorded. ;

[0092] Calculate all during the learning period mean and standard deviation ;

[0093] Then the dynamic coupling threshold Set to: , where f1 is a sensitivity coefficient adjusted based on the stability of user behavior;

[0094] The user habit learning module determines dynamic coupling thresholds by recording behavioral coupling at the same time of day during the system learning period and calculating the mean and standard deviation of behavioral coupling during the learning period, rather than using a fixed threshold. This design can accurately capture users' own behavioral patterns (such as some elderly people's habit of using electricity and water in the morning, and others' habit of evening activities), making the threshold highly compatible with users' actual life rhythm, avoiding the use of a uniform standard to measure the normal behavior of different users, and providing a personalized benchmark for subsequent anomaly detection;

[0095] The above example of monitoring the morning activities of Grandma Li, an elderly woman living alone, is combined with the threshold determination method;

[0096] Scenario: Grandma Li has a regular daily routine, engaging in breakfast-related activities (cooking porridge, boiling water in an electric kettle, and washing vegetables) every day from 7:00 to 8:00. The system enters a one-month (30-day) learning period, and the core task is to determine the dynamic coupling threshold during this period. .

[0097] The process of determining the execution threshold in this case:

[0098] Step 1: Record historical behavior data. The system collects Grandma Li's electricity consumption sequence daily from 7:00 to 8:00. P and water flow sequence W Calculate the behavioral coupling degree for the day. (e.g., within 30 days) The values ​​were 0.78, 0.72, 0.76, 0.74, 0.77, etc., all fluctuating around a fixed range.

[0099] Step 2: Calculate the coupling degree of statistical parameters with 30-day behavior. Statistical analysis revealed:

[0100] mean =0.75 (reflects the routine coupling level of Grandma Li's morning activities);

[0101] Standard deviation =0.12 (reflecting the normal fluctuation range of Grandma Li's morning activity coupling degree).

[0102] Step 3: Determine the sensitivity coefficient f1:

[0103] Because Grandma Li has maintained a stable morning routine for a long time and has not shown any significant behavioral fluctuations in the past year, her behavior is highly stable. Therefore, f1=0.8 (a relatively small value to avoid false alarms caused by an overly strict threshold).

[0104] Step 4: Calculate the dynamic threshold According to the formula Substituting the data, we get: ;

[0105] For example, Grandma Li had a mild cold, so her morning activities were slightly slower. =0.62, dynamic threshold 0.654. ​​At this time, 0.62 < 0.654, and the duration exceeds the preset time, accurately judging the abnormality and triggering a reminder to the children.

[0106] The second analysis model is used to analyze image data to identify the duration of a user's time in the bathroom, and to perform cross-validation by combining the eating patterns reflected by the refrigerator door status change data, calculate the toilet-eating abnormality index, and compare it with the second baseline threshold to determine whether the user's physiological health status is abnormal.

[0107] The specific calculation process of the second analytical model is as follows:

[0108] The system identifies user entry events into the restroom from video data, records the duration of each event, and calculates the average duration of each user entry into the restroom for the current day. ;

[0109] Extract the average opening and closing frequency during the current day's key eating periods from the refrigerator door status change data. ;

[0110] Calculate the toilet-eating abnormality index AEI :

[0111] ;

[0112] in, The current daily average duration and the historical baseline mean deviation, The standard deviation of historical duration data;

[0113] The current daily average switching frequency and the historical baseline average. deviation, The standard deviation of historical frequency data;

[0114] α and β As a weighting factor, and ;

[0115] Will AEI With the second baseline threshold In comparison, when When this happens, it is judged as abnormal;

[0116] Second baseline threshold ,in For history AEI mean f1 represents the historical AEI standard deviation, and f2 is the sensitivity coefficient adjusted based on the stability of the user's toilet-eating behavior.

[0117] The second analytical model does not rely on a single data point to determine health status. Instead, it cross-validates eating patterns reflected by changes in bathroom duration and refrigerator door status through image data. Abnormal toilet time is often associated with physiological problems such as gastrointestinal and urinary issues, while abnormal eating frequency reflects changes in appetite and digestive function. Combining the two can avoid misjudgments caused by single data bias (e.g., an occasional long toilet time ≠ a health problem), making the determination of physiological health abnormalities more reliable. Common physiological health problems of elderly people living alone (such as gastrointestinal discomfort, indigestion, and decreased appetite) often do not immediately manifest as severe symptoms, but will first be reflected in toilet behavior (e.g., changes in toilet time due to diarrhea or constipation) and eating behavior (e.g., reduced refrigerator opening and closing due to poor appetite). The second analytical model calculates the toilet-eating anomaly index (AEI), which can identify these hidden health risks early, even before the elderly explicitly express discomfort, thus buying time for timely intervention. The calculation of AEI relies on historical baseline data and balances the influence of toilet and eating data through α and β weighting factors. Elderly people who are more sensitive to toilet behavior (such as those with chronic gastrointestinal diseases) can have their α weight increased, while those who need more attention to changes in eating (such as diabetic patients) can have their β weight increased. At the same time, the second baseline threshold is calculated based on the historical AEI mean and standard deviation, avoiding the use of a uniform standard to measure all elderly people and adapting to individual needs.

[0118] Take Grandma Li, an elderly woman living alone, as an example:

[0119] Systematic learning period: Establishing a personalized physiological behavior baseline. Grandma Li's daily physiological behavior patterns will be systematically followed by a 1-month (30-day) learning period.

[0120] The environmental behavior monitoring module records the duration of Grandma Li's daily bathroom visits using video data, and calculates the historical average duration of her bathroom visits over 30 days. =12 minutes, standard deviation of historical duration data =3 minutes;

[0121] Simultaneously, data on refrigerator door status changes were extracted, and the opening and closing frequencies during 11:00-13:00 (lunchtime) and 17:00-19:00 (dinnertime) were statistically analyzed to obtain the historical average of the average opening and closing frequency during key eating periods. =4 times / day, standard deviation of historical frequency data =1 time / day;

[0122] Calculate the historical average of daily AEI over 30 days. =0.35, historical AEI standard deviation =0.12; Because Grandma Li's "toilet-eating" behavior has moderate stability, the sensitivity coefficient f2 = 0.9, according to the second baseline threshold formula: Calculated ;

[0123] Considering Grandma Li's weak gastrointestinal function, weighting factors were set as α=0.6 (emphasizing toilet data) and β=0.4 (emphasizing food intake data).

[0124] Routine monitoring: Based on AEI to determine abnormal physiological health. One Thursday, Grandma Li experienced diarrhea and decreased appetite due to acute gastroenteritis.

[0125] The system uses video data to identify that Grandma Li entered the bathroom four times that day, with durations of 20 minutes, 18 minutes, 15 minutes, and 16 minutes respectively. The average duration of her stay in the bathroom that day was calculated as follows:

[0126] minute;

[0127] Extract the refrigerator's on / off frequency during key mealtimes of the day: once during lunchtime and once during dinnertime. Calculate the current daily average on / off frequency. = 2 times / day;

[0128] Calculate the deviation value:

[0129] minute,

[0130] Once / day;

[0131] Substitute into the AEI formula to calculate:

[0132]

[0133] Comparing AEI=1.85 with =0.242, AEI is much greater than the threshold, the second analysis model determines that Grandma Li's physiological health condition is abnormal, and sends the result to the alarm generation module. The system sends a physiological health abnormality reminder to her children, and the children take Grandma Li to see a doctor in time to avoid the condition from worsening.

[0134] The third analysis model is used to perform correlation analysis between the continuous opening status of gas usage data and the change data of refrigerator door status, calculate the risk coefficient, and compare it with the third baseline threshold to determine whether there is any abnormality in the safety of the user's cooking activities.

[0135] The specific calculation method of the third analytical model is as follows:

[0136] Real-time monitoring of gas flow data ,when When the stove is detected as being turned on, the safety timer is activated. , To determine the minimum flow threshold for turning on the stove;

[0137] exist If the refrigerator door is opened during the timer, the timer will reset. ;

[0138] Calculate the risk factor D:

[0139] ;

[0140] in, For the continuous, unreset timer, The average cooking time for users, a and b These are weighting coefficients;

[0141] Compare D with the third baseline threshold In comparison, when When this happens, it is judged as abnormal;

[0142] Third baseline threshold The value is calculated by the user habit learning module based on the mean and standard deviation of historical risk coefficient data. ,in The historical mean of D, f3 represents the historical standard deviation of D, and f3 is the sensitivity coefficient adjusted according to the cooking safety risk level.

[0143] The third analytical model does not simply monitor gas usage; instead, it analyzes gas usage data in conjunction with changes in the refrigerator door's status. Simply turning on the gas does not equate to risk. Risk is only assessed when the refrigerator door remains inactive for a preset time after the gas is turned on (prompting the elderly person to leave the kitchen). This avoids accidental alarms triggered by the elderly person's normal cooking activities (such as chopping vegetables without touching the refrigerator after turning on the gas), making the identification of cooking safety risks more accurate. Elderly people living alone, due to poor memory and mobility issues, are prone to forgetting to turn off the gas after turning it on or leaving temporarily while cooking, potentially leading to serious accidents such as gas leaks and fires. The third analytical model specifically monitors these scenarios, identifying risks in their early stages and preventing escalation, providing crucial protection for the safety of elderly people at home.

[0144] Grandma Li lives alone and has a regular cooking routine (preparing lunch from 11:00-12:00 and dinner from 17:30-18:30 daily). There was a previous safety hazard of "forgetting to turn off the gas after turning it on." The logic of this case is to first enter a one-month (30-day) learning period to determine the core parameters, and then carry out daily cooking safety monitoring.

[0145] During the learning period, determine the key parameters:

[0146] 1. Set the minimum flow threshold for the stove to start. The system uses indoor gas consumption monitoring equipment to collect real-time gas flow data of Grandma Li's gas stove (for cooking and making soup) during the learning period, and calculates that the minimum stable flow rate for normal ignition of the stove is 0.02m³ / h. 3 / h, therefore set =0.02m 3 / h (i.e., the real-time gas flow rate) (At that time, it is determined that the stove is turned on).

[0147] 2. Calculate the average cooking time for users. Record the duration of each cooking session (from start to stop) during the learning period, collecting 60 sets of valid data (twice a day). The results show:

[0148] Average cooking time =25 minutes (e.g., lunch averages 22 minutes, dinner averages 28 minutes, overall average is 25 minutes).

[0149] 3. Determine the third baseline threshold Calculate the statistical parameters of the historical risk factor D:

[0150] Calculate the risk factor D for occasional cooking time-overrun scenarios during the learning period (such as cooking soup for 5 minutes overtime), and obtain the historical mean of D. =1.8, standard deviation =0.5;

[0151] Sensitivity coefficient f3 is set as follows: Since gas leaks are a high-risk hazard requiring strict monitoring, f3 = 1.0; according to the formula... calculate: (i.e., risk factor) (At that time, the cooking safety was deemed abnormal).

[0152] Set weighting coefficients a , b Considering the time that has not been reset ( The impact on safety risks is more direct (the longer the timeout, the higher the risk), so a weighting coefficient should be set. a =0.6、 b =0.4 (satisfies the calculation logic of the risk factor).

[0153] Routine monitoring, anomaly detection and response:

[0154] 1. Real-time monitoring and timer start: On a Tuesday at 11:05, Grandma Li turned on the gas stove to cook soup (real-time gas flow rate). The system detects that the stove is on and immediately starts the safety timer. The timer starts from 0 minutes.

[0155] 2. Timer not reset triggers risk calculation. At 11:08, Grandma Li received a phone call from her children and left the kitchen to answer it in the living room, forgetting to turn off the gas stove; as of 11:35, the system continued to monitor... Still 0.08m 3 / h (stove not turned off), and no refrigerator door was opened during this period (Grandma Li did not return to the kitchen to get / put in food, so the condition of "refrigerator door opening resets timer" is not met), safety timer If it remains unreset, at this time =11:30-11:05=30 minutes.

[0156] 3. Calculate the risk factor D:

[0157] Substitute data ( a =0.6、 b =0.4、 =30 minutes =25): .

[0158] 4. Anomaly detection and alarm triggering: Compare the risk factor D=18.48 with the third baseline threshold. =1.3, because The third analysis model determined that "Grandma Li's cooking activities are not safe" (there is a risk of leakage / fire due to the gas being continuously turned on and unattended).

[0159] The system triggers a response according to a preset strategy: activates the local kitchen audible and visual alarm (volume 60dB, red warning light flashing, alerting Grandma Li);

[0160] An emergency alert was sent to Grandma Li's children via SMS and app push notification: "Gas has been on for more than 30 minutes, and no refrigerator operation has been detected. There may be a safety risk. Please check immediately."

[0161] After the risk was eliminated, Grandma Li's children contacted her immediately by phone after receiving the alarm, reminding her to turn off the gas. Grandma Li hung up and returned to the kitchen, turning off the gas stove at 11:38 AM. The safety timer was activated. Stop, risk eliminated.

[0162] The specific method for performing the analysis using the fourth analytical model is as follows:

[0163] Step S61: Preprocess the playback audio source signal to obtain the preprocessed signal, including noise reduction and filtering of non-media playback audio sources;

[0164] Step S62: Extract the sound pressure level time series data of the preprocessed signal. and audio spectrum time-series data;

[0165] Step S63: Process the sound pressure level time series data Perform a first-order difference operation to obtain the sound pressure level change sequence. and to Perform Z-score standardization to identify those with absolute values ​​greater than a threshold. The points are used as candidate mutation points;

[0166] Step S64: Perform parallel analysis on the consistency between the sound pressure level timing pattern and the audio track content, specifically including:

[0167] Mode Analysis A, Sound Pressure Level Timing Mode Analysis: Calculation of Sound Pressure Level linear regression slope within the sliding time window k ,like k It remains positive and its absolute value exceeds the slope threshold. If the time exceeds the preset duration, it is determined to be a mode where the volume continues to rise abnormally.

[0168] Pattern Analysis B, Audio Track Content Consistency Analysis: For the candidate mutation points identified in step S63, analyze the audio spectrum characteristics within the time window before and after them. If the spectrum characteristics indicate that the audio track content has not changed but the sound pressure level has changed abruptly, it is determined to be a sudden volume change mode.

[0169] Step S65: Match the judgment result of the above pattern analysis A or pattern analysis B with the pre-stored risk feature library to obtain the matching degree. The risk feature library contains typical sound patterns associated with the user's physical discomfort.

[0170] Step S66: When the matching degree exceeds the risk feature threshold At that time, the fourth analysis model determines that the user's physical condition is abnormal;

[0171] Risk characteristic threshold The user habit learning module is trained based on historical matching data and matching samples corresponding to user physical discomfort events. ,in The mean matching score for each physical discomfort event. f4 is the standard deviation of the matching degree corresponding to the physical discomfort event, and f4 is the adjustment coefficient to balance the false negative rate and false positive rate of user physical status judgment.

[0172] The fourth analytical model uses sound pressure level data from indoor audio playback devices as its core monitoring dimension, filling the gaps in previous monitoring of vision (images) and behavior (water and electricity usage, refrigerator / gas operation). When elderly people are in blind spots of visual monitoring (such as sofa corners or bedrooms), or when changes in visual / behavioral data cannot be triggered due to accidents (such as falls), abnormal sound pressure levels of audio devices (such as sudden increases or decreases in volume, or sustained excessively high volumes) can serve as key risk signals. This avoids missing potential hazards due to missing monitoring dimensions. Some changes in the condition of elderly people living alone, such as hearing loss leading to continuously increasing volume, accidental falls causing sudden changes in volume due to knocking over remote controls, or accidental volume button presses due to confusion, are difficult to identify through traditional behavioral data, but can directly reflect their health (hearing) or safety (fall) status. By calculating sound pressure level deviation values, the fourth analytical model can identify these hidden changes early and specifically cover the atypical risk monitoring needs of the elderly.

[0173] For example, Grandma Li, an elderly woman living alone:

[0174] System Learning Period: Establishing a Personalized Sound Pressure Level Baseline. Grandma Li's daily habit is to play traditional opera on the living room TV between 2:00 PM and 4:00 PM. The system enters a 1-month (30-day) learning period.

[0175] The environmental behavior monitoring module records television sound pressure level data daily from 2:00 PM to 4:00 PM using audio acquisition equipment, and calculates the average sound pressure level over 30 days. =65 dB Sound pressure level standard deviation =5 dB ;

[0176] Because Grandma Li's daily TV operation is routine (volume adjustment is stable), the sensitivity coefficient f4 is set to 1.5. The third baseline threshold is calculated using the threshold formula for this case.

[0177] This means that a risk assessment is triggered when the TV sound pressure level exceeds 72.5dB for more than 10 seconds, or suddenly drops to 45dB (the preset low volume abnormal threshold).

[0178] Routine monitoring: Risk assessment based on abnormal sound pressure level. At 15:10 on a certain Friday, Grandma Li got up to get a water glass while watching TV in the living room. She slipped and fell, knocking over the TV remote control. This caused the volume button on the remote control to be pressed and held down for a long time, and the TV sound pressure level suddenly increased from 63dB to 85dB and did not drop back for 15 seconds.

[0179] The system collects sound pressure level data in real time and calculates the current sound pressure level S=85dB, which exceeds the third baseline threshold. =72.5dB, and the duration of the abnormality (15 seconds) exceeds the preset judgment time (10 seconds).

[0180] The fourth analysis model determines that the abnormally high sound pressure level may indicate an accidental risk, and immediately sends a signal to the alarm generation module: first triggering a gentle indoor sound and light alert (to avoid startling the elderly), and at the same time sending a reminder to the children that the TV volume has increased abnormally and suggesting that they check the elderly's safety.

[0181] When her children tried to contact Grandma Li by phone but couldn't get through, they immediately went to her home to check on her and found that she had suffered a minor fall. They promptly took her to the hospital for treatment to prevent the injury from worsening.

[0182] The alarm generation module is configured to execute a hierarchical alarm strategy, wherein the alarm level L in the hierarchical alarm strategy is determined by the following formula:

[0183] ;

[0184] in, These are the normalized anomaly values ​​output by the first, second, third, and fourth analysis models, respectively.

[0185] These are the weighting factors corresponding to each model;

[0186] Depending on the different ranges of alarm level L, the system triggers different levels of response measures, ranging from sending reminder messages to activating local alarms and even requesting external assistance.

[0187] By using min-max normalization, the output values ​​of the first to fourth analysis models (behavioral coupling deviation, "toilet-eating" abnormality index AEI, gas risk duration deviation, and sound pressure level deviation) are uniformly mapped to the [0,1] interval. The original output dimensions of different models differ significantly (e.g., sound pressure level deviation is in dB, AEI is a dimensionless exponent). Without normalization, an imbalance can easily occur where larger output values ​​dominate risk assessment. After normalization, the data from each model can participate in the fusion calculation on the same dimension, avoiding dimensional interference, ensuring fairness in the assessment, and the weighting factors can be flexibly adjusted according to the health and safety priorities of the elderly. For example, for elderly people with weak gastrointestinal function or a history of forgetting about gas, the weighting factor can be increased. W 2 (Physiological health monitoring) and W 3. Weighting of (Cooking Safety Monitoring);

[0188] During the system learning period (e.g., the normalization parameters and weights are determined within one month). In addition to establishing baseline thresholds for each analysis model, the following key parameters are also determined to meet the requirements of multi-model fusion:

[0189] The minimum / maximum range of each model's output values ​​(based on historical data during the learning period, used for subsequent normalization calculations):

[0190] Model 1 (Behavioral Coupling Deviation) ,Right now ), ;

[0191] The second model ("Toileting-Eating" Abnormality Index (AEI)): (AEI has no negative value; the maximum value is the AEI value under extreme abnormal conditions during the learning period.)

[0192] Third Model (Gas Risk Duration Bias) ,Right now ;

[0193] (Unit: minutes, no negative values);

[0194] Fourth model (sound pressure level deviation) ,Right now ):

[0195] (Unit: dB, can be positive or negative);

[0196] Personalized weighting: Considering Grandma Li's weak digestive system (requiring close attention to her physical health) and her history of forgetting to turn off the gas (requiring close attention to cooking safety), the weighting factor is set as follows: (First Model) (Second Model) (Third Model) (Fourth Model);

[0197] Overall risk threshold: set The normalized composite risk value A level 2 alarm will be triggered (if it is not an emergency, a reminder will be sent to family members). The system will trigger a Level 1 alarm (emergency, activate local audible and visual alarms and coordinate with family members / rescue personnel).

[0198] Routine Monitoring: Multi-model Output Normalization and Fusion Calculation One Wednesday, Grandma Li experienced a chain of physiological and behavioral abnormalities due to mild gastroenteritis. The real-time output and normalization process of each analysis model are as follows: Model 1 (Behavioral Coupling Deviation) Real-time behavior coupling degree CD =0.62, the previously established dynamic coupling threshold Therefore .

[0199] Substituting into the min-max normalization formula ,have to ;

[0200] The weighted result is: .

[0201] The second model (“Toilet-Eating” Abnormality Index (AEI)): Real-time monitoring showed that Grandma Li’s average toilet time that day was 17 minutes (historical average). =12 minutes), the refrigerator is opened and closed twice a day during key eating periods (historical average). =4 times / day), and the real-time AEI was calculated to be 1.2.

[0202] Substituting into the normalization formula, we get ;

[0203] The weighted result is .

[0204] Third Model (Gas Risk Duration Bias) ):

[0205] During lunchtime that day, Grandma Li turned on the gas but did not operate the refrigerator for 9 minutes (preset linkage time threshold). (minutes), therefore minute.

[0206] Substituting into the normalization formula, we get ;

[0207] The weighted result is .

[0208] Fourth model (sound pressure level deviation) That afternoon, when Grandma Li was watching TV, the real-time sound pressure level was... S =70 d B (Previously established sound pressure level threshold) Therefore .

[0209] Substituting into the normalization formula, we get The weighted result is .

[0210] Comprehensive risk assessment and alarm response calculation of the comprehensive risk value: The weighted results of the four models are summed to obtain... ; Threshold comparison determination: because The alarm was not triggered, but the system detected a single point of anomaly in the second model (AEI). Therefore, it sent a friendly reminder to Grandma Li's children that "Grandma Li's physiological behavior (toileting, eating) was slightly abnormal today. We suggest you pay attention to her diet and physical condition." Subsequent intervention: The children visited Grandma Li the next day and found that she had mild gastroenteritis symptoms. They promptly provided her with symptomatic medication and a light diet to prevent her condition from worsening.

[0211] The user habit learning module also includes a feedback optimization unit;

[0212] The feedback optimization unit receives feedback on the effectiveness of historical alarms;

[0213] The first baseline threshold is dynamically adjusted based on the feedback statistics. Second baseline threshold Third baseline threshold Risk characteristic threshold and weighting factors By continuously optimizing the above parameters, the false alarm rate of the system can be reduced.

[0214] The entire process from false alarms to parameter optimization:

[0215] Historical False Alarm Cases and Feedback Collection: After Grandma Li used the system for one month, the system triggered a false alarm once using the second analysis model ("toilet-eating anomaly"). The specific scenario is as follows:

[0216] False alarm trigger scenario: One Wednesday, Grandma Li helped her neighbor pick up a package, resulting in the refrigerator being opened and closed only 3 times during the key mealtimes of the day (11:00-13:00, 17:00-19:00) (historical average). =4 times), average toilet time 13 minutes (historical average) =12 minutes).

[0217] Original parameter calculation: According to the second model formula, the "toileting-eating" abnormality index is:

[0218] ;

[0219] At that time, the second baseline threshold =0.242 (from (Calculated) because The system detects an anomaly and sends an alert to the children.

[0220] Feedback collection: After the children visit to confirm, they mark the alarm as a false alarm through the system APP (reason: caused by temporary matters, not a physiological health problem), and the feedback optimization unit receives the valid feedback.

[0221] Feedback Analysis: To pinpoint the cause of the false alarm, the feedback optimization unit performs data backtracking and analysis on the false alarm case to determine the core cause;

[0222] Second baseline threshold Too strict: The original f2=0.9 (set based on the moderate stability of Grandma Li's "toilet-eating" behavior) resulted in an excessively high threshold. =0.242 is too low; even slight fluctuations (such as opening and closing the refrigerator one less time or taking one minute longer to use the toilet) will trigger an alarm.

[0223] Second model weights W 2. Too high: OriginalW 2=0.3 (because Grandma Li has a weak stomach, the focus was initially on physiological health monitoring), which caused the abnormality value of the second model to have too much impact on the "comprehensive alarm level". Even a slight abnormality in a single model can raise the overall risk.

[0224] Dynamic parameter adjustment: The optimization unit, based on the above analysis and combined with historical feedback statistics of 1 false alarm and 9 valid alarms, adjusts key parameters as follows:

[0225] Adjust the second baseline threshold Reduce the sensitivity coefficient f2 (from 0.9 to 0.6), as per this case. Formula recalculation:

[0226] (Thresholds are raised, and the criteria for anomaly detection are relaxed.)

[0227] Adjusting the weights of the second model W 2: Reduced from 0.3 to 0.25 (reducing the influence weight of the physiological health monitoring model on the overall alarm to avoid triggering alarms by a single minor abnormality);

[0228] Simultaneous validation of other parameters: Historical feedback showed no significant false positives / false negatives in models one, three, and four; therefore, they were maintained. =0.654、 =1.3、 =0.7 and W 1 =0.2 、W 3 =0.4 、W 4 =0.2 remains unchanged.

[0229] Post-treatment effect verification: After avoiding similar false alarms for one week, similar behavioral fluctuations occurred: the average toilet time was 13 minutes. ), refrigerator open and close 3 times ( The system operates according to the adjusted parameters;

[0230] Second model calculation: Comparison of the adjusted ,although However, the anomaly value of the second model (After normalization);

[0231] Comprehensive Alert Level Calculation: According to the formula:

[0232] After substitution and adjustment ,have to ;

[0233] Result determination: Since 0.15 did not reach the comprehensive risk threshold for triggering an alarm (the preset value is 0.2), the system did not trigger an alarm, successfully avoiding similar false alarms, which is consistent with the goal of reducing the system's false alarm rate.

[0234] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. An intelligent health monitoring alarm system, characterized in that... include: The environmental behavior monitoring module is used to collect environmental and behavioral data related to users' daily activities. The environmental and behavioral data include water consumption data, electricity consumption data, gas consumption data, video data, environmental sound data, and refrigerator door status change data. Among them, ambient sound data specifically refers to the playback sound source signal generated by a television or audio playback device and collected through an indoor microphone; The data processing center is used to receive and store environmental and behavioral data; The user habit learning module is used to build and dynamically update a user's personalized daily behavior baseline model based on historical data and machine learning algorithms. The data processing center is configured to call the baseline model generated by the user habit learning module and run multiple independent analysis models, including: a first analysis model, a second analysis model, a third analysis model, and a fourth analysis model; The alarm generation module is used to generate and send corresponding monitoring alarm information when any analysis model outputs an anomaly determination result.

2. The intelligent health monitoring alarm system according to claim 1, characterized in that: The first analysis model is used to perform correlation analysis on the time-series characteristics of electricity consumption data and water consumption data, calculate their behavioral coupling degree, and compare it with the first baseline threshold to determine whether there is an abnormal deviation in the user's basic life activity pattern. The specific method by which the first analysis model calculates the behavioral coupling degree is as follows: Establish a power consumption sequence sampled at fixed time intervals and water flow sequence ; Calculate the behavioral coupling degree CD of two sequences within a preset time window; The first baseline threshold is the dynamic coupling threshold. When CD remains below If the time exceeds the preset duration, the first analysis model will determine it as abnormal.

3. The intelligent health monitoring alarm system according to claim 2, characterized in that: The user habit learning module determines the dynamic coupling threshold in the following way. : During the system learning period, the behavioral coupling degree at the same time each day was recorded. ; Calculate all during the learning period mean and standard deviation ; Then the dynamic coupling threshold Set to: , where f1 is a sensitivity coefficient adjusted based on the stability of user behavior.

4. The intelligent health monitoring alarm system according to claim 1, characterized in that: The second analysis model is used to analyze image data to identify the duration of a user's time in the bathroom, and to perform cross-validation by combining the eating patterns reflected by the refrigerator door status change data, calculate the "toilet-eating" abnormality index, and compare it with the second baseline threshold to determine whether the user's physiological health condition is abnormal. The specific calculation process of the second analytical model is as follows: The system identifies user entry events into the restroom from video data, records the duration of each event, and calculates the average duration of each user entry into the restroom for the current day. ; Extract the average opening and closing frequency during the current day's key eating periods from the refrigerator door status change data. ; Calculate the "toileting-eating" abnormality index AEI ,Will AEI With the second baseline threshold In comparison, when When this occurs, it is considered abnormal.

5. The intelligent health monitoring alarm system according to claim 1, characterized in that: The third analysis model is used to perform correlation analysis between the continuous opening status of gas usage data and the change data of refrigerator door status, calculate the risk coefficient, and compare it with the third baseline threshold to determine whether there is any abnormality in the safety of the user's cooking activities. The specific calculation method of the third analytical model is as follows: Real-time monitoring of gas flow data ,when When the stove is detected as being turned on, the safety timer is activated. , To determine the minimum flow threshold for turning on the stove; exist If the refrigerator door is opened during the timer, the timer will reset. ; Calculate the hazard factor D, and compare D with the third baseline threshold. In comparison, when When this occurs, it is considered abnormal.

6. The intelligent health monitoring alarm system according to claim 1, characterized in that: The fourth analysis model is used to analyze the playback audio source signal. By analyzing its sound pressure level timing pattern and the consistency of the audio track content, it can determine whether the user's physical condition is abnormal. The specific method for performing the analysis using the fourth analytical model is as follows: Step S61: Preprocess the playback audio source signal to obtain the preprocessed signal, including noise reduction and filtering of non-media playback audio sources; Step S62: Extract the sound pressure level time series data of the preprocessed signal. and audio spectrum time-series data; Step S63: Process the sound pressure level time series data Perform a first-order difference operation to obtain the sound pressure level change sequence. and to Perform Z-score standardization to identify those with absolute values ​​greater than a threshold. The points are used as candidate mutation points; Step S64: Perform parallel analysis on the consistency between the sound pressure level timing pattern and the audio track content, specifically including: Mode Analysis A, Sound Pressure Level Timing Mode Analysis: Calculation of Sound Pressure Level linear regression slope within the sliding time window k ,like k It remains positive and its absolute value exceeds the slope threshold. If the time exceeds the preset duration, it is determined to be a mode where the volume continues to rise abnormally. Pattern Analysis B, Audio Track Content Consistency Analysis: For the candidate mutation points identified in step S63, analyze the audio spectrum characteristics within the time window before and after them. If the spectrum characteristics indicate that the audio track content has not changed but the sound pressure level has changed abruptly, it is determined to be a sudden volume change mode. Step S65: Match the judgment result of the above pattern analysis A or pattern analysis B with the pre-stored risk feature library to obtain the matching degree. The risk feature library contains typical sound patterns associated with the user's physical discomfort. Step S66: When the matching degree exceeds the risk feature threshold At that time, the fourth analysis model determines that the user's physical condition is abnormal; Risk characteristic threshold It is derived by the user habit learning module based on historical matching data and matching samples corresponding to user physical discomfort events.

7. The intelligent health monitoring alarm system according to claim 1, characterized in that: The alarm generation module is configured to execute a hierarchical alarm strategy, wherein the alarm level L in the hierarchical alarm strategy is determined by the following formula: ； in, These are the normalized anomaly values ​​output by the first, second, third, and fourth analysis models, respectively. These are the weighting factors corresponding to each model; Depending on the different ranges of alarm level L, the system triggers different levels of response measures, ranging from sending reminder messages to activating local alarms and even requesting external assistance.

8. An intelligent health monitoring alarm system according to any one of claims 1-7, characterized in that: The user habit learning module also includes a feedback optimization unit; The feedback optimization unit receives feedback on the effectiveness of historical alarms; The first baseline threshold is dynamically adjusted based on the feedback statistics. Second baseline threshold Third baseline threshold Risk characteristic threshold and weighting factors .

Images