Intelligent refrigerator grading early warning method and system for solitary old people based on multi-dimensional behavior analysis and intention reasoning

Abstract

The application discloses a kind of old age intelligent refrigerator grading early warning system and method based on multidimensional behavior monitoring and intention reasoning, the method includes: real-time acquisition is monitored object to the time sequence behavior data of refrigerator;Based on historical data, the dynamic behavior baseline of each behavior dimension is established, and is updated adaptively with time;Current behavior data is compared with corresponding dynamic behavior baseline, and abnormal level is generated according to the degree of deviation;Behavior chain is input into trained health state evaluation model, and the abnormal interrupt behavior chain caused by transient health event is identified, and potential health risk level is reasoned out;Abnormal level and potential health risk level are fused, and final risk level is generated, and grading early warning is carried out according to final risk level.The application innovatively introduces behavior intention reasoning capability, identifies the sudden disability event hidden behind abnormal interrupt behavior chain, realizes the prevention upgrade from post-alarm to intervention in progress.

Description

Technical Field

[0001] This invention relates to the field of smart home technology, and in particular to a graded early warning method and system for age-friendly smart refrigerators based on multi-dimensional behavior monitoring and intent reasoning. Background Technology

[0002] With the aging population, the number of elderly people living alone is increasing, and their home safety is receiving growing attention. Refrigerators are an essential appliance for the elderly every day, and their usage data naturally reflects their living conditions. Therefore, monitoring refrigerator usage behavior to assess the safety of the elderly has become a research hotspot.

[0003] Existing technologies already contain relevant solutions. For example, Japanese Patent Application Publication No. 2003-185315 discloses a monitoring device and method that determines whether an elderly person is living a normal life by monitoring the frequency of refrigerator door openings, and allows setting a "monitoring period" for the caregiver to receive notifications. However, this type of solution has the following shortcomings: (1) Single monitoring dimension: Relying solely on the frequency of door opening to judge the elderly's condition can easily lead to misjudgment. For example, the elderly may experience a decrease in the frequency of door opening due to short-term travel, temporary changes in their daily routine, or occasional forgetting to open the door, but this does not necessarily mean they are in a dangerous situation.

[0004] (2) No dynamic behavioral baseline was established: Using fixed thresholds or simple averaging methods cannot adapt to the natural fluctuations in the elderly’s lifestyle (such as seasonal changes, changes in health status, etc.), resulting in a high false alarm rate.

[0005] (3) Insufficient privacy protection: Most solutions do not fully consider the elderly’s concerns about privacy, and uploading image data to the cloud for processing may lead to privacy leakage risks.

[0006] (4) Lack of age-friendly interactive design: The existing solutions are mainly aimed at caregivers, and the elderly themselves find it difficult to use the system on their own. Moreover, the reminder methods are limited (such as only APP push), which is not convenient for the elderly who are not good at using smartphones to receive information.

[0007] (5) Single function: It only has the function of monitoring and has failed to be combined with daily needs such as food management and health intervention, making it difficult to form a complete care loop.

[0008] To address the aforementioned issues, providing an early warning method for an age-friendly smart refrigerator that can balance accurate monitoring and privacy security is a direction that those skilled in the art are dedicated to researching. Summary of the Invention

[0009] The first objective of this invention is to provide a graded early warning method for age-friendly smart refrigerators based on multi-dimensional behavior monitoring and intent reasoning, so as to solve the problems mentioned in the background art.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a graded early warning method for an age-friendly smart refrigerator based on multi-dimensional behavior monitoring and intent reasoning, comprising the following steps: Step S1: Real-time collection of the monitored object's temporal behavior data on the refrigerator. The temporal behavior data includes at least two of the following: door opening time point, door closing time point, duration of each door opening, and shelf weight continuously changing over time. Step S2: Based on the various time-series behavioral data collected in history, establish dynamic behavioral baselines corresponding to each behavioral dimension, and the dynamic behavioral baselines are adaptively updated over time. Step S3: Compare the currently collected multiple time-series behavioral data with the corresponding dynamic behavioral baselines, and generate an anomaly level R1 based on the degree of deviation of each behavioral dimension. The value of R1 ranges from 0 to 10. Step S4, Construction of Behavioral Chains and Intent Reasoning: The continuously collected temporal behavioral data is segmented according to a preset fixed time window, and behavioral chains with causal correlation are extracted. Then, the behavioral chains are input into the trained health status assessment model to identify abnormal interrupted behavioral chains, and a potential health risk level R2 is generated based on the matching results of the behavioral chains and the health event feature library. The value of R2 ranges from 0 to 10. Each complete behavioral chain includes at least the door opening action, the door closing action, and the object retrieval action that occurs within the door opening to closing time window. Behavioral chains in which no valid object retrieval action is detected within the time window are defined as abnormal interrupted behavioral chains. The health status assessment model adopts a temporal neural network model. Step S5: The abnormality level and the potential health risk level are fused to generate the final risk level, and the final comprehensive risk score S is calculated according to the following formula: S = α·(R1 / 10) + β·(R2 / 10), Where α and β are dynamic weight coefficients and satisfy α+β=1, (R1 / 10) and (R2 / 10) are the normalized values ​​of R1 and R2 respectively, with a value range of 0 to 1; Step S6: Map the final comprehensive risk score S to the final risk level. Specifically: when 0 ≤ S < 0.3, it is mapped to low risk; when 0.3 ≤ S < 0.6, it is mapped to medium risk; when 0.6 ≤ S ≤ 1.0, it is mapped to high risk. Step S7: Based on the final risk level, a graded warning is issued. Specifically, low risk triggers the first-level warning, medium risk triggers the second-level warning, and high risk triggers the third-level warning.

[0011] As a specific implementation method, in step S2, the establishment of the dynamic behavior baseline adopts the following steps: A sliding time window is used to statistically analyze various time-series behavioral data from the previous N days, and the average value and standard deviation of each behavioral dimension are calculated. The average value plus or minus a preset multiple of the standard deviation is used as the normal fluctuation range. The value of N ranges from 7 to 30 days and is automatically adjusted as the usage time increases.

[0012] As a specific implementation method, step S3, the process of generating the anomaly level R1 includes the following steps: Calculate the deviation between the current data for each behavioral dimension and the corresponding baseline value. The deviation is the ratio of the absolute value of the difference to the baseline value, expressed as a percentage. Preset a first threshold T1 and a second threshold T2 for the range of deviation, where T1 ranges from 30% to 70% and T2 ranges from 70% to 30%. For each behavioral dimension, a contribution value C is assigned based on its deviation D: if D ≤ T1, then C = 0; if T1 ≤ T1, then C = 0. <d ≤t2，则c="1；若D">If T2, then C=2; Sum the contribution values ​​C of all behavioral dimensions to obtain the total contribution value S. Linearly map S to the interval between 0 and 10 to obtain the anomaly level R1, where R1 = (S / (2 × number of dimensions)) × 10. Through the above steps, the degree of deviation of each dimension is quantified into a continuous R1 value. The larger the R1 value, the higher the degree of anomaly.

[0013] As a specific implementation method, the health status assessment model in step S4 adopts a temporal neural network model. The model input layer receives the behavioral chain sequence within a fixed time window, and the output layer uses the Softmax activation function to output the probability distribution of health risk level and behavioral intention classification results.

[0014] As a specific implementation method, step S4, the process of generating the potential health risk level includes the following steps: Step 41, Behavioral chain feature extraction: Extract the feature vector of each behavioral chain from the time-series behavioral data. The feature vector includes at least the duration of door opening, the waiting time for retrieving the item, the change in the weight of the retrieved item, and the integrity identifier of the behavioral chain. Step 42, Abnormal Interruption Behavior Chain Identification: Behavior chains that exceed the set normal retrieval time threshold for the duration of door opening, exceed the set normal reaction time threshold for the retrieval waiting time, and have zero or less than the set minimum effective retrieval threshold for the weight of the retrieved item are identified as abnormal interruption behavior chains. Step 43, Behavioral Pattern-Health Event Association Reasoning: Match the behavioral chain features with a preset health event feature database, which includes the following mapping relationships: The first mapping relationship is as follows: when at least one abnormal interruption behavior chain is detected, and at least one normal behavior chain exists within the same day, it is matched as a "suspected transient loss of consciousness event," and then further judgment is made: If a normal behavior chain exists in the next time window after the occurrence of the behavior chain, the behavior chain in the current time window will maintain the original matching result; if there is no normal behavior chain in the next time window and the total number of normal behavior chains on that day is less than 50% of the historical daily average, the behavior chain in the current time window will be downgraded to "non-specific behavior interruption". The second mapping relationship is: when no behavioral chain is detected for two or more consecutive days, and the monitored person has no record of going out or no calendar notification, it is matched as "suspected event of persistent disability"; The third mapping relationship: When the completeness of the behavior chain is detected to show a monotonically decreasing trend for 7 consecutive days or more, and the cumulative decrease exceeds 20%, it is matched as "suspected event of progressive functional decline"; Fourth mapping relationship: When none of the above first to third mapping relationships are satisfied, the match is "no clear health event indication".

[0015] As a specific implementation method, in step S7, the graded early warning is divided into three levels: low risk, medium risk, and high risk. Specifically: Level 1 warning: When the final risk level is low, a reminder message will be broadcast through the refrigerator's local voice module. The reminder message will use a preset caring tone. Level 2 warning: When the final risk level is medium risk, a notification will be pushed to the client via the mobile terminal APP, along with a summary of recent behavioral data; Level 3 warning: When the final risk level is high, the system will automatically dial the preset emergency contact number and announce the identity information and address of the monitored person.

[0016] As one specific implementation, the time-series behavioral data also includes food images captured by the refrigerator's built-in camera. The food images are immediately deleted after being identified and processed on the local device, and only anonymized text information is stored. The text information includes the food name, quantity, and storage time. All usage behavior data is stored locally, and anonymized behavioral statistics are uploaded only after obtaining authorization from the monitored individual or guardian.

[0017] The second objective of this invention is to provide a graded early warning system for an age-friendly smart refrigerator based on multi-dimensional behavior monitoring and intent reasoning, for implementing the aforementioned graded early warning method. The graded early warning system includes: The multi-dimensional behavior monitoring module includes a door magnetic sensor, a timer, a weight sensor, and a camera, which are used to collect real-time behavioral data of the monitored object during its interaction with the refrigerator. The dynamic baseline establishment module is connected to the multi-dimensional behavior monitoring module, and establishes and dynamically updates the dynamic behavior baseline corresponding to each behavior dimension based on historical data. The multi-dimensional anomaly determination module is connected to the multi-dimensional behavior monitoring module and the dynamic baseline establishment module, respectively. It is used to compare the current behavior data with the corresponding dynamic behavior baseline and generate an anomaly level based on the degree of deviation of each dimension. The graded early warning module is connected to the multi-dimensional anomaly determination module and is used to trigger different levels of alerts according to different anomaly levels. The local processing unit, deployed at the refrigerator, performs image recognition and data processing, and performs privacy protection processing on the images to ensure that the original image data is not uploaded to the cloud. The behavior chain construction and intent reasoning module is connected to the local processing unit. It is used to extract causally related behavior chains from aligned time-series behavior data and output the potential health risk level R2 and the matched health event type by matching them with a preset health event feature library.

[0018] The Behavior Chain Construction and Intent Reasoning module upgrades "abnormal behavior" to "health risk assessment": by assembling fragmented actions of opening, retrieving, and closing doors into causal behavior chains, it identifies abnormal interruption patterns, including "not retrieving items for a long time after opening the door," thereby accurately distinguishing risk levels and realizing preventive care from "post-event alarm" to "in-event intervention."

[0019] As one specific implementation, the tiered early warning system also includes a food ingredient management module and a health intervention module. The food ingredient management module is connected to the local processing unit and is used to record the type, quantity, and storage time of the food ingredients, and to trigger an expiration reminder based on the preset shelf life. The health intervention module is connected to the food management module and is used to store the health data and prohibited food information of the monitored object. When a prohibited food is detected to be removed, a voice reminder is triggered.

[0020] As a specific implementation, the graded early warning system also includes an age-friendly interactive unit placed on the refrigerator door. The age-friendly interactive unit includes a voice recognition module, a voice synthesis module, and a display screen. The voice recognition module is used to receive voice query commands from the monitored object, and the voice recognition module supports dialect recognition and slow speech speed adaptation. The voice synthesis module is used to broadcast food inventory information and expiration reminder information. The display screen is equipped with an interactive interface.

[0021] Compared with the prior art, the beneficial effects of this invention are as follows: 1) The age-friendly smart refrigerator graded early warning method of the present invention upgrades from "behavioral monitoring" to "health reasoning". By segmenting time-series behavioral data into behavioral chains and matching them with a health event feature database, the early warning information has clinical interpretability and can clearly point out specific risk types such as "suspected transient loss of consciousness events". 2) This invention adopts a two-level fusion architecture of "preliminary anomaly level + potential health risk level", which takes into account both statistical sensitivity and medical specificity. It increases the statistical weight in the cold start stage and increases the medical weight when an abnormal interruption of the behavioral chain is detected, effectively reducing the false alarm rate. 3) This invention establishes a dynamic behavioral baseline to adapt to individual differences, avoiding the one-size-fits-all false alarms caused by fixed thresholds. Tests show that the false alarm rate is reduced by approximately 80% compared to traditional solutions. 4) This invention introduces cross-time period behavioral chain correlation analysis, which identifies a gradual decline trend by tracking the "behavioral chain integrity change curve" and issues a preventive care warning 3-5 days in advance; 5) This invention adopts a multimodal privacy protection architecture of local identification, de-identified storage, and edge computing, ensuring that all original image and behavioral data are not uploaded to the cloud; 6) This invention is designed for the elderly, featuring dialect voice recognition, a large-font interface (36pt), and four-button one-click operation, allowing the elderly to use it independently. Attached Figure Description

[0022] Figure 1 This is a flowchart of the aging-friendly smart refrigerator grading and early warning system based on multi-dimensional behavior monitoring and intent reasoning as described in this invention. Figure 2 The flowchart below illustrates the multi-dimensional anomaly determination process used in the age-friendly smart refrigerator grading and early warning method based on multi-dimensional behavior monitoring and intent reasoning described in this invention. Figure 3 This is a flowchart of the behavior chain construction and intent reasoning process in the age-friendly smart refrigerator graded early warning method based on multi-dimensional behavior monitoring and intent reasoning described in this invention. Figure 4 This is a flowchart of the integrated risk fusion and graded early warning process in the age-friendly smart refrigerator graded early warning method based on multi-dimensional behavior monitoring and intent reasoning described in this invention. Detailed Implementation

[0023] The following will refer to the appendices in the embodiments of the present invention. Figure 1-4 The technical solutions in the embodiments of the present invention are clearly and completely described herein. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0024] See Figure 1 As shown, this invention provides a method for designing an age-friendly smart refrigerator based on multi-dimensional behavior monitoring, comprising the following steps: Step S1: Real-time collection of the monitored object's temporal behavior data on the refrigerator. The temporal behavior data includes at least two of the following: door opening time point, door closing time point, duration of each door opening, and shelf weight continuously changing over time. Step S2: Based on the various time-series behavioral data collected in history, establish dynamic behavioral baselines corresponding to each behavioral dimension, and the dynamic behavioral baselines are adaptively updated over time. Step S3: Compare the currently collected multiple time-series behavioral data with the corresponding dynamic behavioral baselines, and generate an anomaly level R1 based on the degree of deviation of each behavioral dimension. The value of R1 ranges from 0 to 10. Step S4, Construction of the Behavior Chain and Intent Reasoning, see [link / reference] Figure 3 As shown: Continuously collected time-series behavioral data is segmented according to a preset fixed time window, and behavioral chains with causal relationships are extracted. Then, the behavioral chains are input into a trained health status assessment model to identify abnormal interrupted behavioral chains. Based on the matching results of the behavioral chains and the health event feature library, a potential health risk level R2 is generated, where the value of R2 ranges from 0 to 10. Each complete behavioral chain includes at least the actions of opening and closing a door, as well as the action of retrieving an object within the time window from opening to closing the door. Behavioral chains in which no effective retrieving action is detected within the time window are defined as abnormal interrupted behavioral chains. The health status assessment model adopts a time-series neural network model. Step S5, see Figure 4 As shown, the abnormality level and the potential health risk level are fused to generate the final risk level, and the final comprehensive risk score S is calculated using the following formula: S = α·(R1 / 10) + β·(R2 / 10), Where α and β are dynamic weight coefficients and satisfy α+β=1, (R1 / 10) and (R2 / 10) are the normalized values ​​of R1 and R2 respectively, with a value range of 0 to 1; Step S6: Map the final comprehensive risk score S to the final risk level. Specifically: when 0 ≤ S < 0.3, it is mapped to low risk; when 0.3 ≤ S < 0.6, it is mapped to medium risk; when 0.6 ≤ S ≤ 1.0, it is mapped to high risk. Step S7: Based on the final risk level, a graded warning is issued. Specifically, low risk triggers the first-level warning, medium risk triggers the second-level warning, and high risk triggers the third-level warning.

[0025] Specifically, in step S1, the temporal behavioral data refers to the raw data stream with precise timestamps collected sequentially by various sensors deployed on the refrigerator, recording the complete temporal evolution trajectory of the monitored object's interaction with the refrigerator. This includes: (1) Door opening and closing event timing data: collected by door magnetic sensor, recording the precise time point of each door opening and closing action, represented as a triple sequence (specifically including action type and timestamp), where the action type includes "open door" and "close door"; (2) Door opening duration sequence: Calculated by the timer module, the time interval between each "door opening-closing" action pair is recorded, that is, the duration of a single door opening, which is represented as a binary sequence (specifically including door opening event ID and duration); (3) Weight change time series data: The weight sensor at the bottom of the shelf collects data at a fixed sampling frequency (e.g., every 10 seconds) and records the continuous change curve of the weight of each shelf over time. The time point of retrieval and the change in weight of the retrieved item can be extracted from this data.

[0026] When these three types of data are aligned on the timeline, they constitute complete time-series behavioral data. Unlike traditional isolated statistical indicators (such as "number of times the door is opened per day"), time-series behavioral data retains key causal information such as the sequence of events, time intervals, and durations, laying the foundation for constructing "behavioral chains" in the future.

[0027] In step S2, the establishment of the dynamic behavior baseline adopts the following steps: A sliding time window is used to collect various usage behavior data from the previous N days, and the average value and standard deviation of each behavior dimension are calculated. The average value plus or minus a preset multiple of the standard deviation is used as the normal fluctuation range. The value of N ranges from 7 to 30 days and is automatically adjusted as the usage time increases.

[0028] In step S3, see Figure 2 As shown, the process of generating anomaly level R1 includes the following steps: Calculate the deviation between the current data for each behavioral dimension and the corresponding baseline value. The deviation is the ratio of the absolute value of the difference to the baseline value, expressed as a percentage. Preset a first threshold T1 and a second threshold T2 for the range of deviation, where T1 ranges from 30% to 70% and T2 ranges from 70% to 30%. For each behavioral dimension, a contribution value C is assigned based on its deviation D: if D ≤ T1, then C = 0; if T1 ≤ T1, then C = 0. <d ≤t2，则c="1；若D">If T2, then C=2; Sum the contribution values ​​C of all behavioral dimensions to obtain the total contribution value S. Linearly map S to the interval between 0 and 10 to obtain the anomaly level R1, where R1 = (S / (2 × number of dimensions)) × 10. Through the above steps, the degree of deviation of each dimension is quantified into a continuous R1 value. The larger the R1 value, the higher the degree of anomaly.

[0029] In this example, the first threshold T1 is used to determine whether a single dimension has a mild anomaly, and its value ranges from 30% to 70%; the second threshold T2 is used to determine whether a single dimension has a severe anomaly, and its value ranges from 70% to 90%.

[0030] Specifically, when the deviation exceeds T1 but does not exceed T2, it is marked as a mild one-dimensional anomaly; when the deviation exceeds T2, it is marked as a severe one-dimensional anomaly.

[0031] Preferably, T1 is set to 50% and T2 to 80%. These preferred values ​​are determined based on the analysis of behavioral statistics of normal populations: the diurnal fluctuation range of each behavioral dimension of normal populations is usually within 30% to 50% of the mean. Setting T1 to 50% can effectively filter out normal physiological fluctuations; setting T2 to 80% means that behaviors with a deviation of more than 80% have significantly deviated from the individual's normal pattern and have a high value in indicating abnormalities.

[0032] It should be noted that the specific values ​​of T1 and T2 can be adaptively adjusted based on the individual differences of the monitored individuals and the actual usage scenarios. For example, for users with strong behavioral patterns, T1 and T2 can be appropriately reduced to improve sensitivity; for users with large behavioral fluctuations, T1 and T2 can be appropriately increased to avoid false alarms.

[0033] The health status assessment model in step S4 adopts a temporal neural network model. The input layer of the model receives the behavioral chain sequence within a fixed time window, and the output layer uses the Softmax activation function to output the probability distribution of health risk level and the behavioral intention classification result.

[0034] Specifically, the model is trained using supervised learning. Training data is collected from refrigerator usage data of no fewer than 500 elderly people living alone, with each person monitored continuously for at least 30 days. The raw time-series data is then processed through behavioral chain construction and health status labeling to form a labeled sample set. Label categories include: normal behavioral chains, abnormally interrupted behavioral chains, and abnormally incomplete behavioral chains, along with corresponding health status labels (healthy, mild discomfort, transient health event, sudden disability).

[0035] During training, the cross-entropy loss function and Adam optimizer were used, and the training, validation, and test sets were divided in a 6:2:2 ratio. On the test set, the model achieved an overall accuracy of over 85% in judging health risk levels.

[0036] This embodiment adopts a layered deployment architecture of "cloud-based general pre-training + edge-based personalized fine-tuning". The general model is pre-trained on large-scale anonymized group data and then pre-placed in the local processing unit on the device. When the locally accumulated behavioral data of the monitored individuals reaches more than 14 days, the system automatically triggers the personalized fine-tuning process. Fine-tuning only uses locally stored time-series behavioral data, behavioral chain construction results, and guardian feedback tags; the original data is not uploaded to the cloud. Fine-tuning adopts an incremental learning strategy. Specifically, it uses an elastic weight consolidation algorithm to prevent the forgetting of old knowledge due to new training, and a layered fine-tuning strategy to freeze the first few layers of the model and only update the parameters of the last few layers. To balance the distribution differences between normal and abnormal samples in the local data, a combination of undersampling and oversampling is used for data augmentation. With user authorization, the system can participate in federated learning, uploading only the noisy model gradient parameters. All local data is stored using AES-256 encryption, and the key is stored in the device's secure area. After personalized fine-tuning is completed, the system uses a locally reserved validation set for evaluation. After confirming that the model performance improvement meets the standards, the personalized model is switched to be used for subsequent health status assessments.

[0037] The model takes a sequence of behavioral chains within a fixed time window as input and outputs a potential health risk level (low / medium / high) and / or a classification result of behavioral intent (normal eating / rummaging / transient health event / sudden disability). Trained with a large number of labeled samples, the model learns the mapping relationship between the temporal features of the behavioral chain and health status.

[0038] The process of generating a potential health risk level includes the following steps: Step 41, Behavioral Chain Feature Extraction: Extract feature vectors for each behavioral chain from the time-series behavioral data. The feature vectors include at least the duration of door opening, the waiting time for retrieving the item, the change in the weight of the retrieved item, and a behavioral chain integrity identifier. The behavioral chain integrity identifier is a binary parameter, denoted as I_complete, used to characterize whether a behavioral chain contains a complete "door opening-item retrieval-door closing" process. When I_complete=1, it indicates that the behavioral chain simultaneously detected a door opening event, a valid item retrieval action, and a door closing event within the time window. When I_complete=0, it indicates that the behavioral chain is interrupted, including situations where the door was opened but not closed, or where there were both door opening and closing actions but no item retrieval action between them. This identifier is used to distinguish between normal behavioral chains and abnormally interrupted behavioral chains and is one of the core judgment criteria for health event matching. Step 42, Abnormal Interruption Behavior Chain Identification: Behavior chains that exceed the set normal retrieval time threshold for the duration of door opening, exceed the set normal reaction time threshold for the retrieval waiting time, and have zero or less than the set minimum effective retrieval threshold for the weight of the retrieved item are identified as abnormal interruption behavior chains. Step 43, Behavioral Pattern-Health Event Association Reasoning: Match the behavioral chain features with a preset health event feature database, which includes the following mapping relationships: The first mapping relationship is as follows: when at least one abnormal interruption behavior chain is detected, and at least one normal behavior chain exists within the same day, it is matched as a "suspected transient loss of consciousness event," and then further judgment is made: If a normal behavior chain exists in the next time window after the occurrence of the behavior chain, the behavior chain in the current time window will maintain the original matching result; if there is no normal behavior chain in the next time window and the total number of normal behavior chains on that day is less than 50% of the historical daily average, the behavior chain in the current time window will be downgraded to "non-specific behavior interruption". The second mapping relationship is: when no behavioral chain is detected for two or more consecutive days, and the monitored person has no record of going out or no calendar notification, it is matched as "suspected event of persistent disability"; The third mapping relationship: When the completeness of the behavior chain is detected to show a monotonically decreasing trend for 7 consecutive days or more, and the cumulative decrease exceeds 20%, it is matched as "suspected event of progressive functional decline"; Fourth mapping relationship: When none of the above first to third mapping relationships are satisfied, the match is "no clear health event indication".

[0039] Risk level fusion: The type of health event is determined based on the matching results of the health event feature database. The health event types include E1 (suspected transient loss of consciousness), E2 (suspected persistent disability), E3 (suspected progressive functional decline), and E4 (no clear health event indication). Assign a base risk score B according to the following rules: If E = E², then B = 9.0. If E = E1, then B = 7.5. If E = E3, then B = 4.5. If E = E4, then B = 1.0; Calculate the matching confidence C according to the following rules: If E=E1, then C=min(1.0, (N_normal / (N_normal+N_interrupt))×2); If E = E², then C = min(1.0, D_continuous / 2); If E = E3, then C = min(1.0, |ΔIntegrity| / 15%); If E = E4, then C = 1.0; Wherein, N_normal represents the number of behavioral chains identified as normal within the day; N_interrupt represents the number of behavioral chains identified as abnormally interrupted within the day; D_continuous represents the number of consecutive natural days without any behavioral chains generated; and ΔIntegrity represents the cumulative decrease in behavioral chain integrity over multiple consecutive days, expressed as a percentage.

[0040] The confidence calculation formula described above applies only when the corresponding triggering conditions are met. For E1, the triggering condition requires at least one abnormal interruption behavior chain and at least one normal behavior chain on that day; therefore, the denominator N_normal + N_interrupt ≥ 2, and division by zero will not occur. For other mapping relationships, the denominator is also guaranteed to be non-zero under the corresponding triggering conditions.

[0041] It should be noted that the constant parameters in the above formula (such as ×2, divide by 2, 15%) are optimized based on verification experiments of no less than 500 normal elderly people and 100 clinically confirmed cases. In practical applications, they can be adjusted according to the characteristics of the user group and clinical feedback, and are not intended to limit the present invention.

[0042] Calculate the preliminary risk score R2' = B × C; The initial risk score R2' is mapped to the range of 0 to 10 to obtain the potential health risk level R2.

[0043] In step S5, the values ​​of the dynamic weighting coefficients α and β are determined according to the following rules: When the system has been running for less than 7 days (cold start phase), take α=0.8 and β=0.2; When the system runtime reaches or exceeds 14 days (normal monitoring phase) and there are no special scenarios, take α=0.5 and β=0.5. When at least one abnormal interruption behavior chain is detected, α = 0.3 and β = 0.7 are taken; When no behavioral chain is detected for two or more consecutive days, α=0.4 and β=0.6 are used. When a record of an outing or calendar notification is detected for the monitored individual, α=0.7 and β=0.3 are used. When multiple conditions are met simultaneously, the arithmetic mean of the weight coefficients corresponding to each condition is taken.

[0044] In step S7, the tiered early warning system is set up with three levels: low risk, medium risk, and high risk. Specifically: Level 1 warning: When the final risk level is low, a reminder message will be broadcast through the refrigerator's local voice module. The reminder message will use a preset caring tone. Level 2 warning: When the final risk level is medium risk, a notification will be pushed to the client via the mobile terminal APP, along with a summary of recent behavioral data; Level 3 warning: When the final risk level is high, the system will automatically dial the preset emergency contact number and announce the identity information and address of the monitored person.

[0045] To implement the above-mentioned tiered early warning method, the tiered early warning system includes: a multi-dimensional behavior monitoring module, a dynamic baseline establishment module, a multi-dimensional anomaly judgment module, a behavior chain construction and intent reasoning module, a health event matching module, a comprehensive risk fusion module, a tiered early warning module, and a local processing unit.

[0046] This multi-dimensional behavior monitoring module includes a door magnetic sensor, a timer, and a camera. The door magnetic sensor uses a Hall effect sensor to determine the open / closed state of the refrigerator door by detecting changes in the magnetic field. The timer records the start and end times of each door opening. The camera uses a 5-megapixel CMOS sensor with infrared illumination and only triggers recording when the door magnetic sensor detects an opening event, reducing energy consumption and privacy risks.

[0047] The multi-dimensional behavior monitoring module also includes a weight sensor, which is installed at the bottom of each shelf in the refrigerator to monitor the weight changes of food on each shelf, so as to help judge the behavior of taking out items and the consumption of food.

[0048] The dynamic baseline establishment module connects to the multi-dimensional behavior monitoring module, establishing and dynamically updating dynamic behavior baselines corresponding to each behavior dimension based on historically collected usage behavior data. Specifically, a sliding time window is used to statistically analyze historical data from the previous N days, calculating the mean μ and standard deviation σ for each behavior dimension, with μ ± kσ as the normal fluctuation range. Here, N ranges from 7 to 30 days, and k ranges from 1.5 to 3; preferably, N is 14 days, and k is 2, i.e., the normal range is μ ± 2σ. The baseline is automatically updated daily, and the sliding window always maintains data from the most recent N days.

[0049] The multi-dimensional anomaly detection module is connected to both the multi-dimensional behavior monitoring module and the dynamic baseline establishment module. It compares the currently collected behavior data with the corresponding baseline and generates an anomaly level based on the degree of deviation in each dimension. See also... Figure 2 As shown, specifically, the formula for calculating the deviation is: Deviation = |Current value - Baseline mean| / Baseline mean × 100%; Set the first threshold T1 (single-dimensional anomaly threshold) to 30%-70%; The second threshold T2 (severe abnormality threshold) is 70%-90%; In this example, T1 is 50% and T2 is 80%.

[0050] The exception detection logic is as follows: When the deviation of any behavioral dimension exceeds T1, it is marked as a mild anomaly in a single dimension; When the deviation of two or more behavioral dimensions exceeds T1 simultaneously, or the deviation of any behavioral dimension exceeds T2, it is judged as moderate abnormality. If no door opening event is detected for 48 consecutive hours, it is considered a severe anomaly.

[0051] The behavior chain construction module is connected to the local processing unit and is used to extract causally related behavior chains from aligned temporal behavior data. This module performs the following steps: First, starting with each door opening event and ending with the corresponding door closing event, all relevant data within the time window is encapsulated into a single initial behavior chain unit. This relevant data includes the door opening timestamp, door closing timestamp, door opening duration, item retrieval waiting time, and shelf weight change sequence. For abnormal situations where no door closing event is detected, a preset timeout period (preferably 5 minutes) after door opening is used as a forced termination point, and the behavior chain is marked as incomplete.

[0052] Secondly, for each behavior chain, the following feature vectors are extracted: door opening duration, which is the difference between the door opening timestamp and the door closing timestamp, in seconds; item retrieval waiting time, which is the difference between the door opening timestamp and the first detected weight change, or the door opening duration if there is no weight change, in seconds; item retrieval weight change, which is the maximum absolute value of the shelf weight change within the behavior chain time window, in grams; behavior chain integrity flag, with a value of 1 indicating that a complete door opening-item retrieval-door closing process was detected, and a value of 0 indicating that the process was interrupted; behavior chain occurrence timestamp, used to determine the time period to which the behavior chain belongs; and item retrieval count, which is the number of valid item retrieval actions detected within the behavior chain time window.

[0053] Furthermore, based on the aforementioned feature vectors, the behavior chains are classified into the following types: Normal behavior chain, which satisfies that the duration of door opening is within the normal range, the waiting time for retrieving the item is within the normal range, the change in the weight of the retrieved item is greater than or equal to the minimum effective retrieval threshold, and the integrity flag is 1; Abnormal interruption behavior chain, which satisfies that the duration of door opening is greater than or equal to the abnormal judgment threshold, the waiting time for retrieving the item is greater than or equal to the abnormal judgment threshold, the change in the weight of the retrieved item is less than or equal to the abnormal judgment threshold, and the integrity flag is 1; Abnormal incomplete behavior chain, which satisfies that the integrity flag is 0; Operational negligence behavior chain, which satisfies that the duration of door opening is greater than or equal to the abnormal judgment threshold, but the change in the weight of the retrieved item is greater than or equal to the minimum effective retrieval threshold.

[0054] The thresholds mentioned above are determined based on behavioral statistics of a normal population (with a statistically significant sample size, such as no fewer than 500 elderly people aged 65-85 living alone, monitored continuously for no fewer than 14 days). For each behavioral indicator, its empirical distribution is calculated, with the 5th percentile as the lower limit of the normal range, the 95th percentile as the upper limit of the normal range, and the 95th percentile as the anomaly threshold. It should be noted that the above percentile thresholds can be adjusted according to the actual application scenario and user group characteristics. For example, the anomaly threshold can be appropriately relaxed for elderly people (over 80 years old). Finally, this module outputs the following data to the health event matching module: total number of behavioral chains, number of normal behavioral chains, number of abnormally interrupted behavioral chains, number of operational negligence behavioral chains, number of incomplete behavioral chains, and the complete feature vector of each behavioral chain.

[0055] The health event matching module connects to the behavior chain construction module and the local processing unit. It matches behavior chain features with a pre-defined health event feature library and outputs the potential health risk level. This module executes the following steps: First, the module has a built-in health event feature library, which stores the following preset mapping relationships.

[0056] First mapping relationship: When at least one abnormal interruption behavior chain is detected and at least one normal behavior chain exists on the same day, it is matched as a suspected event of transient consciousness impairment and assigned a basic risk score of 7.5.

[0057] The second mapping relationship: When it is detected that no behavioral chain has been generated for two consecutive days or more and the monitored person has no record of going out or notified of calendar schedules, it is matched as a suspected event of persistent disability and assigned a basic risk score of 9.0.

[0058] The third mapping relationship: When the completeness of the behavior chain is detected to show a monotonically decreasing trend for 7 consecutive days or more and the cumulative decrease is greater than or equal to 20%, it is matched as a suspected event of progressive functional decline and assigned a basic risk score of 4.5.

[0059] Fourth mapping relationship: When none of the above first to third mapping relationships are satisfied, the match is no clear indication of health event, and a basic risk score of 1.0 is assigned.

[0060] Among them, the completeness of the behavioral chain is calculated by multiplying the ratio of the number of normal behavioral chains to the total number of behavioral chains by 100%. When the total number of behavioral chains is zero, the completeness is defined as 0%.

[0061] Secondly, the module calculates the matching confidence of each mapping relationship according to the following rules.

[0062] For the first mapping relationship, the matching confidence is calculated as follows: C1 = min(1.0, (number of normal behavior chains / (number of normal behavior chains + number of abnormal interruption behavior chains)) × 2).

[0063] For the second mapping relationship, the matching confidence is calculated as follows: C2 = min(1.0, number of consecutive days without behavior / 2).

[0064] For the third mapping relationship, the matching confidence is calculated as follows: C3 = min(1.0, cumulative decrease in completeness / 15%).

[0065] For the fourth mapping relationship, the matching confidence is always 1.0.

[0066] Next, the module calculates the potential health risk level using the following formula: R² = B × C, where B is the base risk score corresponding to the matched health event type, and C is the corresponding matching confidence level. The calculation result is truncated to the range of 0 to 10; if R² is greater than 10, it is set to 10; if R² is less than 0, it is set to 0.

[0067] Then, the module performs special scenario forced corrections: if the match is a suspected event of persistent disability and the number of consecutive days without behavioral chains is greater than or equal to 5, then R2 is forced to be 10; if the match is a suspected event of transient loss of consciousness and the number of abnormal interrupted behavioral chains is greater than or equal to 3, then R2 is forced to be 9.5; if the match is a suspected event of progressive functional decline and the cumulative decline in integrity is greater than or equal to 40%, then R2 is forced to be 8.0.

[0068] Finally, the module outputs the following data to the comprehensive risk fusion module: potential health risk level R2, matching health event type label, matching confidence level, and specific feature parameters that trigger the judgment.

[0069] The tiered early warning module connects to the comprehensive risk fusion module to trigger different levels of alerts based on the final risk level. Level 1 Warning (Mild Abnormality): A reminder message will be broadcast through the refrigerator's local voice module. The broadcast content will use a preset caring tone, such as "You haven't opened the refrigerator today. Remember to eat something." The broadcast time will be selected between 18:00 and 20:00 in the evening to avoid disturbing people at night.

[0070] Level 2 warning (moderate abnormality): A notification will be sent via mobile app, along with a summary of recent behavioral data, such as "Abnormal behavior of your family member has been detected in the past 2 days. We recommend contacting them by phone to confirm."

[0071] Level 3 warning (severe anomaly): Automatically dials the preset emergency contact number and announces the identity information and address of the monitored person via voice, redialing once per hour until confirmed.

[0072] A local processing unit is deployed within the refrigerator to perform image recognition and data processing. Specifically, a camera is installed inside the refrigerator. Images captured by the camera are processed locally using a lightweight deep learning model (such as YOLOv8) to identify the type and quantity of food items. After recognition, the original image is immediately deleted, and only anonymized text information (such as food name, quantity, and storage time) is stored. All behavioral data is stored locally, and anonymized behavioral statistics are uploaded only with the authorization of the monitored individual or guardian.

[0073] The behavior chain construction and intent reasoning module is connected to the local processing unit to extract causally related behavior chains from aligned time-series behavior data, and outputs the potential health risk level R2 and the matched health event type by matching them with a preset health event feature library.

[0074] The graded early warning system also includes an age-friendly interactive unit placed on the refrigerator door, which includes a voice recognition module, a voice synthesis module, and a display screen.

[0075] The voice recognition module supports dialect recognition (Sichuanese, Northeastern Mandarin, Shanghainese, Cantonese, Henan dialect, etc.) and slow speech speed adaptation. The speech synthesis module's playback speed is adjustable, with a default setting of 120 words per minute. The display screen is located on the refrigerator door and features a large-font interactive interface. The font size of the large-font interface is more than 1.5 times that of the regular interface (36pt), and the operation hierarchy is simplified. The main interface displays only four large icons: "Check Ingredients" to check the refrigerator's food inventory and expiration information; "Listen to Reminders" to announce unread warnings or expiration reminders; "Find Recipes" to recommend simple recipes based on available ingredients; and "Call Family" to dial a preset emergency contact number with one click. Through these four core function entry points, the elderly can intuitively and conveniently complete daily operations without a complicated learning curve.

[0076] The tiered management system also includes a food ingredient management module, connected to the local processing unit, used to record the type, quantity, and storage time of ingredients. Specifically, the system presets the shelf life of various ingredients (e.g., 21 days for eggs, 7 days for milk, and 5 days for vegetables), and starts reminding users daily 3 days before the ingredients are due to expire. The reminders also use a tiered mechanism: a pop-up window on the refrigerator screen on day 1, a voice announcement on day 2, and a push notification via the app on day 3.

[0077] The hierarchical management system also includes a health intervention module, which is connected to the food management module to store the health data and prohibited food information of the monitored individuals. When the camera recognizes that an elderly person has taken out a prohibited food, a voice reminder is triggered. The reminder is given in a suggestive rather than a commanding tone, such as, "Your blood pressure is a bit high, and this cured meat is a bit high in salt. Would you like to eat less of it today?"

[0078] The following is a specific example: Hardware configuration: The main control chip is a Raspberry Pi 4B (4GB memory), a 5-megapixel USB camera (with infrared fill light function), a door magnetic sensor is a reed switch sensor, the weight sensor uses 4 HX711 modules (installed at the bottom of each shelf of the refrigerator), the voice module uses an iFlytek offline voice synthesis module, the communication module uses a 4G CAT1 module, and the display is a 7-inch IPS touch screen (resolution of 1024×600).

[0079] Behavioral data collection: The door magnetic sensor monitors the refrigerator door's opening and closing status in real time, recording the start time t_start and end time t_end of each opening, and calculates accordingly: (1) Door opening frequency f = total number of times the door is opened per day; (2) Door opening duration d_avg =Σ(t_end -t_start) / f; (3) The time to retrieve the item is t_total = Σ(t_end - t_start); The weight sensor collects weight data for each shelf every 10 seconds. When a weight change of more than 50 grams is detected, the change time t_change and the change amount Δw are recorded as auxiliary criteria for judging the use of food.

[0080] The camera is installed inside the refrigerator door and takes a picture only when the door magnetic sensor detects an opening event. Local image recognition is performed immediately after the picture is taken.

[0081] Example of establishing a dynamic baseline: The system continuously records behavioral data for the previous 14 days and calculates the daily average and standard deviation of various indicators. For example, a user's door opening frequency averages 8.3 times / day, with a standard deviation of 3.2, and a normal fluctuation range of 5.1 to 11.5 times / day; the door opening duration averages 15.2 seconds / time, with a standard deviation of 5.6, and a normal fluctuation range of 9.6 to 20.8 seconds / time; the item retrieval time averages 125 seconds / day, with a standard deviation of 45, and a normal fluctuation range of 80 to 170 seconds / day.

[0082] Behavioral chain construction and health event matching: The system segments time-series behavioral data into behavioral chains based on time windows. Each behavioral chain starts with a door opening event and ends with a corresponding door closing event. Feature vectors are extracted from each behavioral chain, including door opening duration, item retrieval waiting time, item weight change, and an integrity flag. Behavioral chains where the door opening duration exceeds the upper limit of the normal range, the item retrieval waiting time exceeds the upper limit of the normal range, and the item weight change is below the minimum valid threshold are identified as abnormally interrupted behavioral chains.

[0083] The system has a built-in health event feature database that stores the following mapping relationships: when at least one abnormally interrupted behavioral chain is detected and at least one normal behavioral chain exists on the same day, it is matched as a suspected event of transient loss of consciousness; when no behavioral chain is generated for two or more consecutive days and there is no record of going out, it is matched as a suspected event of persistent disability; when the integrity of the behavioral chain shows a monotonically decreasing trend for seven or more consecutive days and the cumulative decrease exceeds 20%, it is matched as a suspected event of progressive functional decline. Based on the matching results, the potential health risk level R2 is output.

[0084] Integrated Risk Management: The system weights and fuses the preliminary anomaly level R1 output by the multi-dimensional anomaly judgment module with the potential health risk level R2 output by the health event matching module, and calculates the final comprehensive risk score using the following formula: S = α·(R1 / 10) +β·(R2 / 10), where α and β are dynamically adjusted weighting coefficients that satisfy α+β=1. The final risk level is mapped based on the S value: S less than 0.3 indicates low risk, S greater than or equal to 0.3 and less than 0.6 indicates medium risk, and S greater than or equal to 0.6 indicates high risk.

[0085] Tiered early warning: The corresponding warning is triggered based on the final risk level: low risk triggers the first level warning, which is a care reminder broadcast locally on the refrigerator; medium risk triggers the second level warning, which pushes a behavioral summary to the guardian through the APP; high risk triggers the third level warning, which automatically dials the emergency contact number and broadcasts a description of the abnormal behavior.

[0086] Example of Anomaly Detection: To more clearly illustrate the collaborative application of multi-dimensional anomaly detection and behavioral chain intent reasoning in this embodiment, the following describes the collaborative application of multi-dimensional anomaly detection and behavioral chain intent reasoning in this embodiment using specific scenarios. All behavioral dimension data are statistically analyzed using a natural day (24 hours) as the statistical period, and the behavioral chain is based on a single "open door-retrieve item-close door" process as the basic unit. The data for each scenario below are all based on the normal behavioral baseline of a 78-year-old elderly person living alone (living alone for 3 years). The normal behavioral baseline of this elderly person is: door opening frequency of 6 to 10 times per day, average door opening duration of 12 to 18 seconds, average item retrieval time of 90 to 150 seconds per day, average number of normal behavioral chains of 6 to 10 per day, and normal behavioral chain integrity range of 85% to 100%.

[0087] Scenario 1: Elderly people traveling: Behavioral data: For two consecutive days, the daily door opening frequency was 0 times, the door opening duration was 0 seconds, the item retrieval time was 0 seconds, the total number of behavioral chains was 0, and the completeness of the behavioral chain was 0%.

[0088] Feature extraction: No behavioral chains were generated for two consecutive days. The system searched historical behavior records and found that the elderly person had two similar outing records in the past 6 months (3 days and 4 days respectively). The system had already learned of the elderly person's itinerary through calendar synchronization one day before this outing, and the outing record was tagged "yes".

[0089] Health event matching: The second mapping relationship was triggered (no behavioral chain for ≥2 consecutive days), but due to the existence of an outing record, the exclusion rule took effect, resulting in a mismatch as a suspected event of persistent disability. The match was downgraded to the fourth mapping relationship (no clear indication of a health event). The basic risk score B=1.0, confidence level C4=1.0, and potential health risk level R2=1.0×1.0=1.0, which falls within the low-risk range (0-3.0).

[0090] Preliminary anomaly assessment: Door opening frequency deviation 100%, door opening duration deviation 100%, item retrieval time deviation 100%, multi-dimensional anomaly assessment module outputs preliminary anomaly level R1=6.0 (moderate anomaly range).

[0091] Comprehensive Risk Fusion: The system detected an outing record, set the weights α=0.7 and β=0.3. The fusion score S=0.7×(6.0 / 10)+0.3×(1.0 / 10)=0.42+0.03=0.45, which is mapped to medium risk.

[0092] Tiered alert: When the second-level alert is triggered, a message will be sent to the guardian's app: "Your family member has not used the refrigerator (out-of-town mode) for two consecutive days and is expected to return tomorrow. Please update your information promptly if there are any changes to the itinerary." If there is still no behavioral chain for 3 consecutive days, it exceeds the maximum number of days of historical outing records, the exclusion rule is invalid, and the second mapping relationship is rematched. R2 is upgraded to 9.0×1.0=9.0. After fusion, S=0.3×0.6+0.7×0.9=0.18+0.63=0.81, which is mapped to high risk and triggers the third level warning.

[0093] Scenario 2: An elderly person is bedridden due to a cold. Behavioral data: Day 1, door opening frequency was 0 times, door opening duration was 0 seconds, item retrieval time was 0 seconds, total number of behavioral chains was 0, and behavioral chain completeness was 0%. Day 0 (previous day) behavior was normal (door opening frequency was 8 times, behavioral chain completeness was 100%).

[0094] Feature extraction: No behavioral chains were generated on that day, and the consecutive number of days without behavioral chains was 1. System retrieval revealed that the elderly person generated 6 to 10 behavioral chains daily over the past 30 days, never experiencing a day without any behavioral chains, and had no record of going out or having any calendar appointments on that day. Compared to the conditions in Scenario 1: the consecutive number of days was only 1, which did not reach the trigger threshold for the second mapping relationship.

[0095] Health event matching: If the E2 trigger condition is not met (requires ≥2 consecutive days), and the E1 and E3 trigger conditions are not met, the match is the fourth mapping relationship (no explicit health event indication). The base risk score B = 1.0, confidence level C4 = 1.0, and potential health risk level R2 = 1.0. However, the system simultaneously outputs a "missing unexpected behavior" flag for weight adjustment.

[0096] Preliminary anomaly assessment: Door opening frequency deviation 100%, door opening duration deviation 100%, item retrieval time deviation 100%, multi-dimensional anomaly assessment module outputs preliminary anomaly level R1=6.0 (moderate anomaly).

[0097] Comprehensive Risk Fusion: The system detected a "missing unexpected behavior" flag and no outing records. Weights were set α=0.4 and β=0.6. The fusion score S=0.4×0.6+0.6×0.1=0.24+0.06=0.30, which is mapped to medium risk (critical value).

[0098] Tiered warning: When the second-level warning is triggered, a message will be sent to the guardian's APP: "Your family member did not use the refrigerator today, which is inconsistent with the daily routine (8 times per day). It is recommended to contact them by phone to confirm." If there is no behavioral chain for two consecutive days, the E2 trigger condition is met (no behavioral chain for ≥2 consecutive days and no record of going out), matching as a suspected event of persistent disability, B=9.0, confidence level C2=min(1.0, 2 / 2)=1.0, R2=9.0. The fusion score S=0.4×0.6+0.6×0.9=0.24+0.54=0.78, mapping to high risk, triggering the third-level warning.

[0099] Scenario 3: The elderly person forgets to close the refrigerator door: Behavioral data: Door opening frequency was 9 times on the day (normal range is 6 to 10 times), average door opening duration was 45 seconds (baseline 15.2 seconds, deviation +196%), cumulative item retrieval time was 405 seconds (baseline 125 seconds, deviation +224%), and total item retrieval amount was 1.2 kg (historical daily average was 1.1 kg, deviation +9%).

[0100] Behavioral chain analysis: The system identified 9 behavioral chains. Seven of these were normal behavioral chains, characterized by: opening the door - a short pause of 3-8 seconds - retrieving an item weighing 80-250g - closing the door, with the door opening duration being 12-18 seconds. The other two were abnormal behavioral chains, characterized by: opening the door - a long pause of 42 seconds - retrieving an item weighing 120-180g - closing the door, with the door opening duration being 40-48 seconds. No abnormally interrupted behavioral chains were identified (because the changes in the weight of the retrieved item all exceeded the minimum effective threshold).

[0101] Health event matching: The E1 trigger condition is not met (requires an abnormal interruption of the behavioral chain), and the E2 and E3 trigger conditions are not met. The matching is the fourth mapping relationship, but it is further subdivided into the "operational negligence subtype" based on behavioral characteristics. The basic risk score B=1.0, the confidence level C4=1.0, and the potential health risk level R2=1.0.

[0102] Preliminary anomaly assessment: Door opening frequency deviation 8% (normal), door opening duration deviation 196% (severe anomaly), item retrieval time deviation 224% (severe anomaly). The multi-dimensional anomaly assessment module outputs a preliminary anomaly level R1=4.0 (mild anomaly).

[0103] Comprehensive Risk Fusion: The system detected an operational negligence pattern, setting weights α=0.5 and β=0.5. The fusion score S=0.5×0.4+0.5×0.1=0.20+0.05=0.25, which is mapped to low risk.

[0104] Tiered warning: When the first level warning is triggered, the refrigerator will announce in a local voice: "The refrigerator door has not been closed for a long time. Please check." Scene 4: The Elderly's Normal Life Behavioral data: Door opening frequency was 8 times on the day (normal range is 6 to 10 times), average door opening duration was 16 seconds (baseline 15.2 seconds, deviation +5.2%), cumulative item retrieval time was 130 seconds (baseline 125 seconds, deviation +4%), and total item retrieval amount was 1.15 kg (historical daily average was 1.1 kg, deviation +4.5%).

[0105] Behavioral chain analysis: The system identified 8 complete behavioral chains, each characterized as follows: opening the door - a short pause of 5.2 seconds on average - retrieving an object weighing 80 to 200g - closing the door, with the door opening duration ranging from 14 to 18 seconds. There were no abnormal behavioral chains, no abnormally interrupted behavioral chains, and the completeness of the behavioral chains was 100%.

[0106] Health event matching: If none of the triggering conditions E1, E2, or E3 are met, the match is the fourth mapping relationship (no explicit health event indication). Base risk score B=1.0, confidence level C4=1.0, potential health risk level R2=1.0.

[0107] Preliminary anomaly assessment: Door opening frequency deviation of 3.6%, door opening duration deviation of 5.2%, and item retrieval time deviation of 4%, all within the normal range. The multi-dimensional anomaly assessment module outputs a preliminary anomaly level R1=0 (normal).

[0108] Comprehensive Risk Fusion: During the normal monitoring period of the system, the weights are set as α=0.5 and β=0.5. The fusion score S=0.5×0+0.5×0.1=0.05, which is mapped to low risk (but close to 0).

[0109] Tiered warning: If the situation is determined to be normal, no warning will be triggered.

[0110] Scenario 5: An elderly person suddenly experiences dizziness: Behavioral data: The door was opened twice that day (below the normal lower limit of 6 times), with the duration of the two openings being 8 seconds and 52 seconds respectively, and the cumulative retrieval time being 60 seconds (below the normal lower limit of 90 seconds). Weight sensor data showed that the weight of the retrieved item decreased by 250g after the first door opening (normal retrieval), and the weight of the retrieved item changed to 0g after the second door opening.

[0111] Behavioral Chain Analysis: The system identified two behavioral chains. Behavioral Chain 1 (9:15 AM): Door open for 8 seconds - Item retrieved (250g) - Door closed. This is a normal behavioral chain. The door opening duration is 8 seconds (normal range 9.6 to 20.8 seconds, slightly below the lower limit but close), the item retrieval waiting time is 5 seconds, and the retrieved weight is 250g. Behavioral Chain 2 (3:42 PM): Door open for 52 seconds - No item retrieved (weight change 0g) - Door closed. This is an abnormally interrupted behavioral chain. The door opening duration is 52 seconds (exceeding the normal upper limit of 20.8 seconds, deviation +150%), the item retrieval waiting time is 52 seconds (no item retrieved), and the retrieved weight is 0g (below the minimum effective threshold of 50g). The integrity flag is 1.

[0112] Health event matching: Triggering the first mapping relationship (number of abnormal interruption behavior chains ≥ 1 and number of normal behavior chains ≥ 1 on the same day). Base risk score B = 7.5. Confidence calculation: C1 = min(1.0, (N_normal / (N_normal+N_interrupt)) × 2) = min(1.0, (1 / (1+1)) × 2) = min(1.0, 1.0) = 1.0. Potential health risk level R2 = 7.5 × 1.0 = 7.5, belonging to the high-risk range (6.0-10.0). The system outputs the matched event type as "Suspected transient loss of consciousness event".

[0113] Preliminary anomaly assessment: Door opening frequency deviation 75% (below the lower limit of the normal range by 6 times, 2 / 8=0.25, deviation 75%), door opening duration deviation +150% (behavioral chain 2), item retrieval time deviation 52% (60 / 125=0.48, deviation 52%). The multi-dimensional anomaly assessment module outputs a preliminary anomaly level R1=6.5 (moderately high anomaly).

[0114] Comprehensive Risk Fusion: The system detected an abnormal interruption behavior chain, setting weights α=0.3 and β=0.7. The fusion score S=0.3×(6.5 / 10)+0.7×(7.5 / 10)=0.3×0.65+0.7×0.75=0.195+0.525=0.72, which is mapped to high risk.

[0115] Tiered Early Warning: Upon triggering a Level 3 warning, the system automatically dials a pre-set emergency contact number and announces via voice: "Your family member (address: Room XX, Building XX, XX Community) exhibited an abnormal interruption in their behavioral chain at 15:42 today—they did not retrieve the item 52 seconds after opening the door. This matches a suspected transient loss of consciousness event, possibly indicating sudden dizziness. Please call to confirm or visit immediately." Simultaneously, the system reports the complete feature vector of behavioral chain 2 (door opening timestamp, duration, and change in item weight) to the guardian's app. The following is an optimization scheme provided for rural areas or scenarios with unstable networks, based on the above implementation.

[0116] Specific improvements include: increasing local storage capacity to 64GB, which can store at least one year of behavioral data; prioritizing the use of 4G CAT1 network (good coverage and low power consumption) for the early warning module, with local records being retransmitted after network recovery when there is no network; adopting a fully offline solution for the voice module, supporting basic voice recognition and synthesis; and further simplifying the large-font interface, adding dual prompts of icons and text to suit elderly people with low literacy levels.

[0117] This embodiment also provides an implementation plan that links with a community-based elderly care service center.

[0118] In addition to sending alerts to children, refrigerator warnings are also simultaneously pushed to the monitoring platform of the community elderly care service center. Community staff can arrange home visits based on the warning level: a phone call for mild abnormalities, a home visit within 3 days for moderate abnormalities, and an immediate home visit for severe abnormalities. The system regularly generates reports on the elderly person's living status, including trends in door opening frequency and food consumption, for reference in the community health record. Emergency contact settings include the phone numbers of community grid workers and the community health station.

[0119] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.< / d> < / d>

Claims

1. A graded early warning method for age-friendly smart refrigerators based on multi-dimensional behavior monitoring and intent reasoning, characterized in that, Includes the following steps: Step S1: Real-time collection of the monitored object's temporal behavior data on the refrigerator. The temporal behavior data includes at least two of the following: door opening time point, door closing time point, duration of each door opening, and shelf weight continuously changing over time. Step S2: Based on the various time-series behavioral data collected in history, establish dynamic behavioral baselines corresponding to each behavioral dimension, and the dynamic behavioral baselines are adaptively updated over time. Step S3: Compare the currently collected multiple time-series behavioral data with the corresponding dynamic behavioral baselines, and generate an anomaly level R1 based on the degree of deviation of each behavioral dimension. The value of R1 ranges from 0 to 10. Step S4, Construction of Behavioral Chains and Intent Reasoning: The continuously collected temporal behavioral data is segmented according to a preset fixed time window, and behavioral chains with causal correlation are extracted. Then, the behavioral chains are input into the trained health status assessment model to identify abnormal interrupted behavioral chains, and a potential health risk level R2 is generated based on the matching results of the behavioral chains and the health event feature library. The value of R2 ranges from 0 to 10. Each complete behavioral chain includes at least the door opening action, the door closing action, and the object retrieval action that occurs within the door opening to closing time window. Behavioral chains in which no valid object retrieval action is detected within the time window are defined as abnormal interrupted behavioral chains. The health status assessment model adopts a temporal neural network model. Step S5: The abnormality level and the potential health risk level are fused to generate the final risk level, and the final comprehensive risk score S is calculated according to the following formula: S = α·(R1 / 10) + β·(R2 / 10), Where α and β are dynamic weight coefficients and satisfy α+β=1, (R1 / 10) and (R2 / 10) are the normalized values ​​of R1 and R2 respectively, with a value range of 0 to 1; Step S6: Map the final comprehensive risk score S to the final risk level. Specifically: when 0 ≤ S < 0.3, it is mapped to low risk; when 0.3 ≤ S < 0.6, it is mapped to medium risk; when 0.6 ≤ S ≤ 1.0, it is mapped to high risk. Step S7: Based on the final risk level, a graded warning is issued. Specifically, low risk triggers the first-level warning, medium risk triggers the second-level warning, and high risk triggers the third-level warning.

2. The age-friendly smart refrigerator tiered early warning method based on multi-dimensional behavior monitoring and intent reasoning according to claim 1, characterized in that, In step S2, the establishment of the dynamic behavior baseline adopts the following steps: A sliding time window is used to statistically analyze various time-series behavioral data from the previous N days, and the average value and standard deviation of each behavioral dimension are calculated. The average value plus or minus a preset multiple of the standard deviation is used as the normal fluctuation range. The value of N ranges from 7 to 30 days and is automatically adjusted as the usage time increases.

3. The age-friendly smart refrigerator tiered early warning method based on multi-dimensional behavior monitoring and intent reasoning according to claim 1, characterized in that, In step S3, the process of generating the anomaly level R1 includes the following steps: Calculate the deviation between the current data for each behavioral dimension and the corresponding baseline value. The deviation is the ratio of the absolute value of the difference to the baseline value, expressed as a percentage. Preset a first threshold T1 and a second threshold T2 for the range of deviation, where T1 ranges from 30% to 70% and T2 ranges from 70% to 30%. For each behavioral dimension, a contribution value C is assigned based on its deviation D: if D ≤ T1, then C = 0; if T1 ≤ T1, then C = 0. <d ≤t2，则c="1；若D"> If T2, then C=2;< / d> Sum the contribution values ​​C of all behavioral dimensions to obtain the total contribution value S. Linearly map S to the interval between 0 and 10 to obtain the anomaly level R1, where R1 = (S / (2 × number of dimensions)) × 10. Through the above steps, the degree of deviation of each dimension is quantified into a continuous R1 value. The larger the R1 value, the higher the degree of anomaly.

4. The age-friendly smart refrigerator tiered early warning method based on multi-dimensional behavior monitoring and intent reasoning according to claim 1, characterized in that, In the health status assessment model in step S4, the model input layer receives the behavioral chain sequence within a fixed time window, and the output layer uses the Softmax activation function to output the probability distribution of health risk level and the behavioral intent classification result.

5. A graded early warning method for an age-friendly smart refrigerator based on multi-dimensional behavior monitoring and intent reasoning, as described in claim 1, is characterized in that... In step S4, the process of generating the potential health risk level R2 includes the following steps: Step 41, Behavioral chain feature extraction: Extract the feature vector of each behavioral chain from the time-series behavioral data. The feature vector includes at least the duration of door opening, the waiting time for retrieving the item, the change in the weight of the retrieved item, and the integrity identifier of the behavioral chain. Step 42, Abnormal Interruption Behavior Chain Identification: Behavior chains that exceed the set normal retrieval time threshold for the duration of door opening, exceed the set normal reaction time threshold for the retrieval waiting time, and have zero or less than the set minimum effective retrieval threshold for the weight of the retrieved item are identified as abnormal interruption behavior chains. Step 43, Behavioral Pattern-Health Event Association Reasoning: Match the behavioral chain features with a preset health event feature database, which includes the following mapping relationships: The first mapping relationship is as follows: when at least one abnormal interruption behavior chain is detected, and at least one normal behavior chain exists within the same day, it is matched as a "suspected transient loss of consciousness event," and then further judgment is made: If a normal behavior chain exists in the next time window after the occurrence of the behavior chain, the behavior chain in the current time window will maintain the original matching result; if there is no normal behavior chain in the next time window and the total number of normal behavior chains on that day is less than 50% of the historical daily average, the behavior chain in the current time window will be downgraded to "non-specific behavior interruption". The second mapping relationship is: when no behavioral chain is detected for two or more consecutive days, and the monitored person has no record of going out or no calendar notification, it is matched as "suspected event of persistent disability"; The third mapping relationship: When the completeness of the behavior chain is detected to show a monotonically decreasing trend for 7 consecutive days or more, and the cumulative decrease exceeds 20%, it is matched as "suspected event of progressive functional decline"; Fourth mapping relationship: When none of the above first to third mapping relationships are satisfied, the match is "no clear health event indication".

6. A graded early warning method for an age-friendly smart refrigerator based on multi-dimensional behavior monitoring and intent reasoning, as described in claim 1, is characterized in that... In step S7, the graded early warning is divided into three levels: low risk, medium risk, and high risk. Specifically: Level 1 warning: When the final risk level is low, a reminder message will be broadcast through the refrigerator's local voice module. The reminder message will use a preset caring tone. Level 2 warning: When the final risk level is medium risk, a notification will be pushed to the client via the mobile terminal APP, along with a summary of recent behavioral data; Level 3 warning: When the final risk level is high, the system will automatically dial the preset emergency contact number and announce the identity information and address of the monitored person.

7. A graded early warning method for an age-friendly smart refrigerator based on multi-dimensional behavior monitoring and intent reasoning, as described in claim 1, is characterized in that... The time-series behavioral data also includes food images captured by the refrigerator's built-in camera. The food images are immediately deleted after being recognized and processed on the local device, and only the desensitized text information is stored. The text information includes the food name, quantity, and storage time. All behavioral data is stored locally, and anonymized behavioral statistics are uploaded only with the authorization of the ward or guardian.

8. A graded early warning system for an age-friendly smart refrigerator based on multi-dimensional behavior monitoring and intent reasoning, used to implement the graded early warning method as described in any one of claims 1 to 7, characterized in that, The tiered early warning system includes: The multi-dimensional behavior monitoring module includes a door magnetic sensor, a timer, a weight sensor, and a camera, which are used to collect real-time behavioral data of the monitored object during its interaction with the refrigerator. The dynamic baseline establishment module is connected to the multi-dimensional behavior monitoring module, and establishes and dynamically updates the dynamic behavior baseline corresponding to each behavior dimension based on historical data. The multi-dimensional anomaly determination module is connected to the multi-dimensional behavior monitoring module and the dynamic baseline establishment module, respectively. It is used to compare the current behavior data with the corresponding dynamic behavior baseline and generate an anomaly level based on the degree of deviation of each dimension. The graded early warning module is connected to the multi-dimensional anomaly determination module and is used to trigger different levels of alerts according to different anomaly levels. The local processing unit, deployed at the refrigerator, performs image recognition and data processing, and performs privacy protection processing on the images to ensure that the original image data is not uploaded to the cloud. The behavior chain construction and intent reasoning module is connected to the local processing unit. It is used to extract causally related behavior chains from aligned time-series behavior data and output the potential health risk level R2 and the matched health event type by matching them with a preset health event feature library.

9. A graded early warning system for an age-friendly smart refrigerator based on multi-dimensional behavior monitoring and intent reasoning, as described in claim 8, is characterized in that... The tiered early warning system also includes a food ingredient management module and a health intervention module. The food ingredient management module is connected to the local processing unit and is used to record the type, quantity, and storage time of the food ingredients, and to trigger an expiration reminder based on the preset shelf life. The health intervention module is connected to the food management module and is used to store the health data and prohibited food information of the monitored object. When a prohibited food is detected to be removed, a voice reminder is triggered.

10. A graded early warning system for an age-friendly smart refrigerator based on multi-dimensional behavior monitoring and intent reasoning, as described in claim 8, is characterized in that... The graded early warning system also includes an age-friendly interactive unit placed on the refrigerator door. The age-friendly interactive unit includes a voice recognition module, a voice synthesis module, and a display screen. The voice recognition module is used to receive voice query commands from the monitored object. The voice recognition module supports dialect recognition and slow speech speed adaptation. The voice synthesis module is used to broadcast food inventory information and expiration reminder information. The display screen is equipped with an interactive interface.

Images