Personal health abnormal behavior monitoring system fused with unsupervised learning

Abstract

The invention provides a personal health abnormal behavior monitoring system fused with unsupervised learning, and relates to the technical field of data processing, and the system is used for carrying out the clustering processing of historical behavior data, dividing the historical behavior data into a plurality of behavior tags, extracting the feature vector of each behavior tag, building an individual behavior physical model according to the feature vector, and carrying out the analysis of the individual behavior physical model. Inputting the target feature data into the individual behavior physical model, calculating the deviation degree of the recent behavior of the user, monitoring and recording the deviation degree to obtain an observation label group, judging whether the observation label group meets the absorption condition of the individual behavior physical model or not, if not, judging that the behavior label is an abnormal behavior, and if not, judging that the behavior label is an abnormal behavior. And according to the deviation degree and the duration of the abnormal behavior, calculating a behavior inhibition factor, and according to the behavior inhibition factor, dynamically adjusting the individual behavior physical model to limit the abnormal behavior from being identified as a normal behavior. According to the invention, the system can be prevented from mistakenly considering periodic abnormal behaviors as normal behaviors.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly to a personal health abnormal behavior monitoring system integrating unsupervised learning. Background Art

[0002] Currently, personal health monitoring systems based on machine learning are gradually replacing traditional rule-based warning methods. In particular, the method of identifying behavior patterns through unsupervised learning and then establishing a personalized baseline model with the assistance of supervised learning has become a common practice. Such methods usually first use the K-means clustering algorithm to automatically identify behavior labels such as bedtime, toilet frequency, and indoor activity interval distribution from the behavior data within 7 to 15 days, and then train a model with supervised algorithms such as support vector machines to establish the health behavior baseline of an individual. After the model inputs new data, it realizes health warnings by comparing the deviation between the current behavior and the historical pattern. Such methods have good adaptability in constructing personalized models and are particularly suitable for long-term monitoring in chronic disease management and home care scenarios.

[0003] However, in actual applications, there are still key technical defects, that is, the model is not sensitive enough to potential behavior variations and may not be able to identify abnormal patterns gradually formed over consecutive days in a timely manner. For example, in the scenario of nighttime behavior monitoring, if an elderly person has frequent toilet use due to urinary system problems recently, their behavior may still be within the normal fluctuation range in the previous few days, and the model regards it as a reasonable content for updating the health baseline. However, if this abnormality continues to evolve into a new behavior habit, due to the lack of a mechanism for identifying the intention of abnormal patterns in the model, this behavior may be classified as normal, resulting in missed reporting problems. The current technology may not be able to clearly distinguish between deviated but reasonable behaviors, and there is a risk that abnormal behaviors are absorbed by the model as a new normal. Summary of the Invention

[0004] The purpose of the present invention is to provide a personal health abnormal behavior monitoring system integrating unsupervised learning, aiming to solve the problems mentioned in the background art.

[0005] To solve the above technical problems, the technical solution of the present invention is as follows:

[0006] A personal health abnormal behavior monitoring system integrating unsupervised learning, the system includes:

[0007] An unsupervised module, configured to obtain historical behavior data of a user, perform clustering processing on the historical behavior data, divide the historical behavior data into multiple behavior labels, and obtain behavior label data;

[0008] An individual model module, configured to extract feature vectors of each behavior label according to the behavior label data, obtain sample feature data, and establish an individual behavior physical model according to the sample feature data;

[0009] A target behavior module, which is used to obtain the target behavior data of a user in the near future, cluster and extract features from it, and obtain target feature data;

[0010] A deviation calculation module, which is used to input the target feature data into an individual behavior physical model, calculate the deviation degree of the user's recent behavior, and monitor and record the abnormal behavior labels whose deviation degree exceeds a preset deviation threshold to obtain an observation label group;

[0011] An abnormal detection module, which is used to judge whether the observation label group meets the conditions for absorption by the individual behavior physical model. When the result is negative, the behavior label is determined to be an abnormal behavior;

[0012] A behavior inhibition module, which is used to calculate a behavior inhibition factor according to the deviation degree and duration of the abnormal behavior, and dynamically adjust the individual behavior physical model according to it to prevent the abnormal behavior from being recognized as a normal behavior.

[0013] Furthermore, the unsupervised module includes:

[0014] A time division unit, which is used to perform division processing on historical behavior data, split the historical behavior data into multiple time segments according to a preset time window, and obtain a time segment data set;

[0015] A feature extraction unit, which is used to extract features from the time segment data set, extract the stay ratio, spatial transfer frequency, and associated physiological values of the user in different spatial regions within each time segment, and obtain a feature data set;

[0016] A clustering unit, which is used to perform clustering processing on the preliminary feature data set, classify the time segments with similar features into the same clustering cluster, and obtain clustering result data;

[0017] A behavior mapping unit, which is used to perform behavior analysis on each clustering cluster in the clustering result data, identify the significant features of the clustering cluster in terms of spatial behavior and physiological parameters, and map the significant features to behavior labels to obtain behavior label data.

[0018] Furthermore, the behavior mapping unit includes:

[0019] A statistical analysis unit, which is used to perform statistics on the time segment data included in each clustering cluster in the clustering result data, calculate the spatial behavior parameters of the user in each spatial region and the physiological parameters under each spatial region within the clustering cluster respectively, and obtain a clustering cluster feature index set;

[0020] A significant feature identification unit, which is used to perform standardization processing on the spatial behavior parameters and physiological parameters in the clustering cluster feature index set, and screen out the spatial regions and physiological value intervals with significant features according to a preset behavior feature threshold to obtain a significant feature set;

[0021] A label classification unit, which is used to match the set of significant features with a preset behavior pattern, select labels according to the spatial region and the physiological value range, determine the behavior label type that best matches the clustering cluster, and bind the behavior label type to the clustering cluster to obtain behavior label data.

[0022] Furthermore, the individual model module includes:

[0023] A behavior decomposition unit, which is used to perform sample analysis on each behavior label in the behavior label data, extract the behavior occurrence time period, behavior duration, behavior spatial region, behavior occurrence frequency, and behavior time concentration corresponding to the label, and obtain a set of behavior features;

[0024] A behavior structure unit, which is used to construct ternary structure data for each behavior label in the set of behavior feature units. The ternary structure data includes a time period field, a spatial region field, and a behavior feature field;

[0025] A behavior sequence unit, which is used to combine all the ternary structure data in chronological order, identify the behavior patterns of the user in different time periods and different spatial regions, and obtain a behavior structure sequence;

[0026] A model construction unit, which is used to define the ternary structures with repeated patterns in the behavior structure sequence as periodic behaviors according to the behavior structure sequence, and convert them into a rule expression form to generate an individual behavior physical model.

[0027] Furthermore, the model construction unit includes:

[0028] A candidate behavior unit, which is used to perform repetitive matching on the ternary structure data in the behavior structure sequence, and classify the ternary structures with time periods and spatial regions that exceed the preset coincidence rate in consecutive days and have behavior features that exceed the preset coincidence rate into the same candidate behavior segment to obtain a set of candidate behavior segments;

[0029] A period recognition unit, which is used to statistically calculate the occurrence frequency and time concentration of each candidate behavior segment according to the set of candidate behavior segments. When its occurrence frequency is greater than the preset frequency threshold and the time concentration is less than the preset concentration threshold, the behavior segment is recognized as a periodic behavior;

[0030] A rule generation unit, which is used to reconstruct the patterns of the time period field, spatial region field, and behavior feature field in the periodic behavior, extract the average occurrence time, main region coordinates, and stable behavior features, and obtain a structured behavior set;

[0031] A physical modeling unit, which is used to determine the behavior pattern of the user in a specific time period and spatial region according to the structured behavior set to obtain an individual behavior physical model.

[0032] Further, the deviation calculation module includes:

[0033] A first calculation unit, configured to compare the target feature data with the time field, space field, and behavior field of the corresponding behavior label in the individual behavior physical model, calculate the differences of the behavior feature sets respectively to obtain a behavior deviation index set, and calculate the deviation degree according to it;

[0034] An anomaly recognition unit, configured to compare the deviation degree with a preset deviation threshold. When the deviation degree is greater than the preset deviation threshold, the behavior label corresponding to the target feature data is determined to be a suspicious behavior label;

[0035] A behavior recording unit, configured to monitor the suspicious behavior label, record the first occurrence time, last occurrence time, continuous occurrence times, and spatial distribution range of the suspicious behavior label to obtain an observation label group.

[0036] Further, the first calculation unit includes:

[0037] A second calculation unit, configured to obtain a first deviation index term according to the product relationship among the behavior time period, behavior frequency, and behavior concentration of the user's recent behavior, combined with the inhibitory effect of the behavior concentration on the time-frequency offset; obtain a second deviation index term according to the interaction between the behavior duration and the behavior spatial position of the user's recent behavior, combined with the adjustment function of the spatial offset on the duration fluctuation; obtain a third deviation index term according to the joint relationship among the behavior time period, behavior concentration, and behavior duration of the user's recent behavior, combined with the interference degree of the frequency offset on the joint relationship;

[0038] Fuse the first deviation index term, the second deviation index term, and the third deviation index term to obtain the deviation degree of the user's recent behavior relative to the individual behavior physical model.

[0039] Further, the anomaly detection module includes:

[0040] A trend extraction unit, configured to extract the continuous occurrence records of the suspicious behavior label in the time dimension according to the suspicious behavior label in the observation label group to obtain a trend sample set;

[0041] A trend analysis unit, configured to calculate the average frequency and average deviation degree of the suspicious behavior label in the continuous records according to the trend sample set to obtain a trend feature index set;

[0042] A judgment rule unit, configured to judge whether the average frequency in the trend feature index set exceeds the behavior occurrence frequency of the same type of label in the individual behavior physical model, and at the same time judge whether the average deviation degree exceeds the preset deviation threshold. When any one of the judgment results is yes, it is determined that the behavior label does not meet the absorption condition;

[0043] An anomaly judgment unit, configured to judge whether there is a trend reversal of a suspicious behavior label within a continuous time window. When there is no trend reversal, it is confirmed that the suspicious behavior label is an abnormal behavior.

[0044] Furthermore, the behavior inhibition module includes:

[0045] An inhibition factor unit, configured to analyze the minimum confidence intensity required for an abnormal behavior to be absorbed by an individual behavior physical model according to the deviation degree and duration of the abnormal behavior, and obtain a behavior inhibition factor;

[0046] An absorption judgment unit, configured to compare the behavior inhibition factor with the absorption trigger intensity of the individual behavior physical model. When the behavior inhibition factor is greater than the absorption trigger intensity, the abnormal behavior does not meet the absorption condition;

[0047] An absorption adjustment unit, configured to dynamically limit the behavior label related to the abnormal behavior in the individual behavior physical model, and increase the time length of the update window of the behavior label in the individual behavior physical model;

[0048] A behavior blocking unit, configured to add a blocking status identifier to the abnormal behavior label that does not meet the absorption condition, and store the behavior label, its deviation degree, and the behavior inhibition factor in the behavior adjustment data.

[0049] Furthermore, the inhibition factor unit includes:

[0050] An inhibition factor calculation unit, configured to calculate the change impact of the abnormal behavior in terms of time distribution and trust according to the ratio of the time offset value of the abnormal behavior to its coverage ratio in the individual behavior physical model, and obtain a main control factor term; calculate the immediate interference ability of the short-term abnormal behavior on the individual behavior physical model according to the duration difference of the abnormal behavior, and obtain a linear adjustment factor term; calculate the overall inhibition amplitude according to the ratio of the concentration offset value to the model coverage and the ratio of the frequency offset value to the behavior duration, and obtain a balance adjustment term;

[0051] Fuse the main control factor term, the linear adjustment factor term, and the balance adjustment term to obtain the behavior inhibition factor of the abnormal behavior.

[0052] The above solution of the present invention has at least the following beneficial effects:

[0053] By establishing a linkage mechanism between the anomaly detection module and the behavior inhibition module, the present invention realizes the dynamic identification and screening of potential abnormal behaviors during the process of updating the individual behavior physical model, and can effectively prevent the technical defect that anomalies are normalized in traditional models. Especially in scenarios where the user's behavior gradually deviates from the original pattern, such as an increase in the frequency of night-time toilet use due to physical health problems, the system can judge whether it meets the condition of being absorbed by the model according to the deviation degree and duration of the abnormal behavior. When it is determined that the absorption condition is not met, the update logic of the behavior model will be intervened based on the behavior inhibition factor, dynamically restricting the update window time of the behavior label in the behavior model, and adding a blocking state to the behavior label to ensure that this anomaly is not regarded as a new behavior baseline. This pre-blocking mechanism based on the behavior evolution trend effectively prevents behavior deviations from penetrating into the model structure, maintains the long-term stability and prediction ability of the model, and has significant value in the practical application of individual health management.

[0054] The present invention judges the behavior evolution trend and conducts trend feature analysis on the continuity and deviation degree of the user's recent behavior through observing the label group, enabling the system to have the ability to predict progressive anomalies in advance. Different from the traditional model's outlier detection method based only on single-point deviation, this system pays more attention to the evolution direction and stability of behavior in the time dimension. By extracting target behavior data and constructing a trend sample set, and further combining trend indicators such as average frequency and deviation degree, intervention can be carried out before the behavior has not completely formed a significant deviation. The system can make a judgment based on its continuous change trend and the comparison result with similar labels in the individual behavior model to achieve early warning. This trend judgment mechanism improves the stability and forward-looking of the system in long-term behavior monitoring and helps to provide timely health intervention suggestions.

[0055] The present invention can comprehensively judge the deviation index from three aspects: behavior time period, spatial position, and behavior concentration, and conduct overall deviation quantification. Different from the traditional single-dimensional deviation identification method that only uses activity frequency as a criterion, the system fully considers the multi-variable interaction relationship between behavior data and realizes the quantitative description of multi-dimensional different behaviors. Through the deviation index set composed of these calculation results, the system can identify complex behavior differences more precisely, thereby reducing the false alarm and missed alarm rates. It is especially suitable for scenarios with strong behavior concentration and high occasionality such as toilet use and sleep, effectively improving the recognition rate of the system for atypical abnormal behaviors and enhancing the credibility and adaptability of the actual health monitoring system.

[0056] The present invention generates an individual behavior physical model in the form of rules by constructing a behavior feature set, a ternary structure sequence, and a behavior structure sequence. Compared with existing rule-based or statistical modeling methods, this solution breaks through the traditional single-variable modeling limitation in design, introduces the ability to express composite features such as behavior time periods, spatial regions, behavior frequencies, and concentration degrees, realizes the dynamic characterization of behavior patterns. This modeling method has strong personalized expression ability, can adapt to the behavior differences among different users, ensures that the model has high adaptability and interpretability, and at the same time provides structured support for multi-dimensional comparison, provides a data basis for the system to identify anomalies, and improves the long-term stability and behavior prediction ability of the model.

[0057] The present invention introduces the calculation logic of a behavior inhibition factor, comprehensively judges its credibility intensity according to multiple factors such as the deviation degree, time coverage, behavior frequency, and concentration degree change of abnormal behaviors, dynamically generates an inhibition factor, and compares it with the absorption threshold of the behavior model to form an absorption constraint mechanism. If the inhibition factor is greater than the model trigger intensity, the system will prevent the model from being updated, further enhancing the conservativeness and controllability of the model. In the scenario of the dynamic evolution of health status, this mechanism can ensure that the individual behavior model is not disturbed by short-term abnormal behaviors, effectively extends the model life cycle and improves the prediction consistency, and has practical value for maintaining the accuracy of the long-term monitoring system, reflecting high engineering practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 is a flowchart of a personal health abnormal behavior monitoring system integrating unsupervised learning provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] Hereinafter, exemplary embodiments of the present disclosure will be described in more detail with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that this disclosure will be more thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0060] As Figure 1 shown, an embodiment of the present invention proposes a personal health abnormal behavior monitoring system integrating unsupervised learning, and the system includes:

[0061] An unsupervised module, configured to obtain historical behavior data of a user, perform clustering processing on the historical behavior data, divide the historical behavior data into multiple behavior labels, and obtain behavior label data;

[0062] An individual model module, configured to extract feature vectors of each behavior label according to the behavior label data to obtain sample feature data, and establish an individual behavior physical model according to the sample feature data;

[0063] A target behavior module, configured to obtain the target behavior data of a user in the near future, cluster and extract features therefrom, so as to obtain target feature data;

[0064] A deviation calculation module, configured to input the target feature data into an individual behavior physical model, calculate the deviation degree of the user's recent behavior, and monitor and record the abnormal behavior labels whose deviation degree exceeds a preset deviation threshold, so as to obtain an observation label group;

[0065] An abnormal detection module, configured to determine whether the observation label group meets the conditions for absorption by the individual behavior physical model. When the result is negative, the behavior label is determined to be an abnormal behavior;

[0066] A behavior inhibition module, configured to calculate a behavior inhibition factor according to the deviation degree and duration of the abnormal behavior, and dynamically adjust the individual behavior physical model according to the behavior inhibition factor, so as to limit the abnormal behavior from being recognized as a normal behavior.

[0067] In an embodiment of the present invention, an unsupervised module is configured to obtain the historical behavior data of a user, perform clustering processing thereon, divide the historical behavior data into multiple behavior labels to obtain behavior label data, and by introducing the joint analysis of spatial behavior parameters and physiological parameters, make the clustering result more physiologically meaningful and behaviorally interpretable, improving the accuracy of the subsequent model and the stability of behavior recognition; an individual model module is configured to extract the feature vectors of each behavior label according to the behavior label data to obtain sample feature data, and establish an individual behavior physical model according to the sample feature data. By means of structured modeling, not only the typical behavior pattern of the user is completely expressed, but also the model's recognition ability for the behavior sequence and regularity is enhanced; a target behavior module is configured to obtain the target behavior data of a user in the near future, cluster and extract features therefrom to obtain target feature data, realizing the rapid response and feature extraction of the system for recent behaviors, supporting continuous tracking of the user's state, and being able to detect abnormal fluctuations in behavior in a timely manner.

[0068] A deviation calculation module is used to input target feature data into the individual behavior physical model, calculate the deviation degree of the user's recent behavior, monitor and record the abnormal behavior labels whose deviation degree exceeds the preset deviation threshold, and obtain an observation label group, enabling the system to make accurate responses to the mutual influences between different behavior attributes, greatly reducing the false negative rate and false positive rate, and improving the accuracy and robustness of health anomaly detection; an anomaly detection module is used to determine whether the observation label group meets the absorption conditions of the individual behavior physical model. When the result is negative, the behavior label is determined as an abnormal behavior. By extracting and comparing trend features, fluctuating behaviors and progressive abnormal behaviors can be distinguished, effectively improving the system's ability to judge the abnormal evolution trend; a behavior inhibition module is used to calculate a behavior inhibition factor according to the deviation degree and duration of the abnormal behavior, and dynamically adjust the individual behavior physical model according to it, restricting the abnormal behavior from being recognized as a normal behavior. By quantifying and dynamically adjusting the model absorption conditions, the evolution process of the model structure can be effectively controlled, avoiding the abnormal behavior from being mislearned by the model due to frequent occurrence, and improving the safety and stability of the entire system in the actual health monitoring process.

[0069] Among them, a target behavior module is used to obtain the user's recent target behavior data, cluster and extract features from it, and obtain target feature data, specifically including:

[0070] The system first accesses the user's raw behavior data within the recent 1 to 3 days from the data acquisition end. This data can be obtained through environmental sensors, wearable devices or smart terminals deployed in the user's residence, specifically including the user's activity trajectories, spatial positions, stay times, behavior frequencies, physiological parameters such as heart rate and body temperature, and environmental parameters such as illuminance and noise. The system performs time synchronization and cleaning processing on the collected data, removing invalid values, missing segments and abnormal noises to ensure the reliability of subsequent analysis.

[0071] Subsequently, the system uses a preset time window strategy to divide the raw data in terms of time. Usually, with 30 minutes or 60 minutes as a unit, the data of a whole day is cut into several continuous time segments. In each time segment, the system calculates multi-dimensional features such as the activity stay ratio, spatial transfer frequency, average value and fluctuation range of physiological parameters in the corresponding time period of the user in each spatial area, and according to the set feature template, uniformly outputs a vectorized feature representation. Then, the system uses an unsupervised clustering algorithm consistent with the historical data processing to cluster these feature vectors, and classifies the behavior segments with similar features into the same clustering cluster, thus forming new behavior labels.

[0072] After the clustering is completed, the system will conduct a preliminary analysis on the segments within each clustering cluster, identify their significant manifestations in terms of temporal continuity, spatial concentration, and physiological parameters, and confirm whether they belong to the evolutionary manifestations of a certain behavior pattern. Finally, the system will generate corresponding target behavior labels for each clustering cluster, and extract data such as the central eigenvalue, standard deviation, spatial behavior hot zone, and concentration fluctuation index of each cluster as target feature data. This target feature data will then be input into the deviation calculation module as the basic data source for comparison with the individual behavior physical model, thereby realizing the quantitative analysis and judgment of the differences between recent behaviors and existing behavior models.

[0073] In a preferred embodiment of the present invention, the unsupervised module includes:

[0074] A time division unit for dividing the historical behavior data, splitting the historical behavior data into multiple time segments according to a preset time window to obtain a time segment data set;

[0075] A feature extraction unit for extracting features from the time segment data set, extracting the residence ratio, spatial transfer frequency, and associated physiological values of the user in different spatial regions within each time segment to obtain a feature data set;

[0076] A clustering unit for clustering the preliminary feature data set, classifying time segments with similar features into the same clustering cluster to obtain clustering result data;

[0077] A behavior mapping unit for conducting behavior analysis on each clustering cluster in the clustering result data, identifying the significant features of the clustering cluster in terms of spatial behavior and physiological parameters, and mapping the significant features into behavior labels to obtain behavior label data.

[0078] In an embodiment of the present invention, a time division unit is configured to perform division processing on historical behavior data, split the historical behavior data into multiple time segments according to a preset time window, and obtain a time segment data set, so as to implement standardized preprocessing of unstructured continuous behavior data, and divide complex time series into behavior observation units with a unified granularity; a feature extraction unit is configured to perform feature extraction on the time segment data set, extract the residence ratio, spatial transfer frequency, and associated physiological values of the user in different spatial regions within each time segment, and obtain a feature data set, realizing the mapping from the original perception data to a structured behavior feature vector, and solving the problems of fuzzy behavior representation and inconsistent data dimensions; a clustering unit is configured to perform clustering processing on the preliminary feature data set, classify time segments with similar features into the same clustering cluster, and obtain clustering result data, realizing the structured classification operation of the system for non-labeled behavior data, enabling the system to identify the inherent similarity and potential categories between behavior patterns in an unsupervised context; a behavior mapping unit is configured to perform behavior analysis on each clustering cluster in the clustering result data, identify the significant features of the clustering cluster in terms of spatial behavior and physiological parameters, and map the significant features to behavior labels to obtain behavior label data, semantically interpreting the abstract clustering result, that is, endowing each clustering cluster corresponding to a time segment with a clear behavior meaning, and enabling the system to transition from low-level feature processing to high-level behavior understanding stage.

[0079] Among them, the clustering unit is configured to perform clustering processing on the preliminary feature data set, classify time segments with similar features into the same clustering cluster, and obtain clustering result data, specifically including:

[0080] The system first receives the feature data set generated by the feature extraction unit. The feature data set is composed of feature vectors of multiple time segments. Each feature vector contains multi-dimensional information such as the residence ratio, spatial transfer frequency, and related physiological parameters of the user in different spatial regions within the time segment. In actual operation, the system first performs standardization processing on all feature vectors to eliminate the order-of-magnitude differences between different feature dimensions and avoid the phenomenon that a certain type of index dominates the clustering distance due to a large numerical range in the subsequent clustering process.

[0081] The standardized feature data will be subjected to clustering calculations. The clustering algorithm can select K-means. The system divides the feature space according to the preset or dynamically adjusted clustering number K value. Before starting the clustering, the system first uses the silhouette coefficient index to evaluate the stability of the clustering structure under different K values and automatically selects the optimal K value. During the clustering process, the system continuously adjusts the belonging category of each feature vector through an iterative method to minimize the Euclidean distance between samples within the class and maximize the distance between samples in different classes, so as to reach the convergence judgment condition. To cope with scenarios with blurred boundaries and unclear behavior transitions, the system can also introduce the GMM soft clustering strategy, enabling some time segments to have membership degrees of multiple clustering clusters simultaneously and enhancing the recognition ability of behavior crossover states.

[0082] After clustering is completed, the system will assign time segments with similar feature patterns to the same clustering cluster and output clustering result data including the clustering cluster number, the corresponding time segment ID, and its central feature vector. This clustering result is used as the input for the next stage to identify the specific behavior label types represented by each clustering cluster, thus completing the transition from the original behavior segment to the semantic label. This process improves the system's automatic recognition ability of user behavior patterns.

[0083] In a preferred embodiment of the present invention, the behavior mapping unit includes:

[0084] A statistical analysis unit for statistically analyzing the time segment data included in each clustering cluster in the clustering result data, respectively calculating the spatial behavior parameters of the user in each spatial region and the physiological parameters under each spatial region within the clustering cluster to obtain a clustering cluster feature index set;

[0085] A significant feature recognition unit for standardizing the spatial behavior parameters and physiological parameters in the clustering cluster feature index set and screening out the spatial regions and physiological value intervals with significant features according to the preset behavior feature threshold to obtain a significant feature set;

[0086] A label classification unit for matching the significant feature set with the preset behavior patterns, selecting labels according to the spatial regions and physiological value intervals, determining the behavior label type most matching the clustering cluster, and binding the behavior label type to the clustering cluster to obtain behavior label data.

[0087] In the embodiment of the present invention, a statistical analysis unit is configured to statistically analyze the time segment data included in each clustering cluster in the clustering result data, calculate the spatial behavior parameters of the user in each spatial region and the physiological parameters in each spatial region respectively, obtain a clustering cluster feature index set, and comprehensively reflect the changes in the behavior state and physiological state of the user in a specific environment through the coupled statistics of the spatial behavior parameters and the physiological parameters; a significant feature recognition unit is configured to perform standardization processing on the spatial behavior parameters and physiological parameters in the clustering cluster feature index set, and screen out the spatial regions and physiological value intervals with significant features according to a preset behavior feature threshold to obtain a significant feature set, effectively removing redundant features and reducing noise interference, and improving the effectiveness and accuracy of behavior feature extraction in the clustering cluster; a label classification unit is configured to match the significant feature set with a preset behavior pattern, select a label according to the spatial region and the physiological value interval, determine the behavior label type that best matches the clustering cluster, and bind the behavior label type to the clustering cluster to obtain behavior label data. Through the label classification mechanism, the system can automatically identify the living state of the user within a specific time period, improve the expression ability of the health monitoring system for behavior semantics, and enhance the interpretability of subsequent abnormal behavior recognition and intervention.

[0088] Among them, the label classification unit is configured to match the significant feature set with a preset behavior pattern, select a label according to the spatial region and the physiological value interval, determine the behavior label type that best matches the clustering cluster, and bind the behavior label type to the clustering cluster to obtain behavior label data, which specifically includes:

[0089] First, a structured behavior pattern library is constructed based on the long-term accumulated data of the user's behavior by the system. Each behavior pattern is composed of typical spatial region identifiers, physiological parameter intervals, and behavior time periods, serving as a reference for label determination. When the label classification unit operates, it first receives the significant feature set output by the significant feature recognition unit. This set usually includes the high-frequency activity behavior characteristics of the user in several spatial regions in the clustering cluster and the corresponding physiological index intervals.

[0090] The system then vectorizes each set of significant features and calls a matching algorithm to calculate the similarity with each behavior pattern in the behavior pattern library one by one. Specifically, the vector space algorithm can be used to calculate the Euclidean distance between the set of significant features and the preset pattern in the spatial dimension and physiological parameter dimension to obtain a set of similarity scores. In the similarity scores, the system sets a pre-screening threshold to eliminate candidate patterns with low matching degrees, and then selects the behavior pattern with the highest similarity in the candidate set as the optimal matching result. If the set of significant features of a certain clustering cluster is close to a certain preset behavior pattern in multiple dimensions, for example, staying in the kitchen area for more than 15 minutes, while the heart rate fluctuates between 90 and 110, and the occurrence time is concentrated at noon, the system will match this clustering cluster as the dining behavior label.

[0091] When the matching is successful, the system assigns this label to the current clustering cluster to realize the mapping from low-level behavior data to high-level semantic labels. The label classification unit binds this behavior label to the clustering cluster in the system to form a structured behavior label data entry, which not only contains the label type, but also the spatial, physiological and time features based on the label. Through this label classification process, the system can semanticize the original behavior clustering result, making the subsequent individual behavior physical model more interpretable and facilitating the system to display the behavior classification result to users or professionals in human-computer interaction or clinical analysis.

[0092] In a preferred embodiment of the present invention, the individual model module includes:

[0093] A behavior decomposition unit for parsing each behavior label in the behavior label data to extract the behavior occurrence time period, behavior duration, behavior space area, behavior occurrence frequency and behavior time concentration corresponding to the label, and obtaining a set of behavior features;

[0094] A behavior structure unit for constructing ternary structure data for each behavior label in the set of behavior feature units. The ternary structure data includes a time period field, a spatial area field and a behavior feature field;

[0095] A behavior sequence unit for combining all the ternary structure data in chronological order to identify the behavior patterns of the user in different time periods and different spatial areas, and obtaining a behavior structure sequence;

[0096] A model construction unit for defining the ternary structure with repeated patterns in the behavior structure sequence as periodic behaviors according to the behavior structure sequence, and converting it into a rule expression form to generate an individual behavior physical model.

[0097] In the embodiment of the present invention, the behavior decomposition unit is used to perform sample parsing on each behavior label in the behavior label data, extract the behavior occurrence time period, behavior duration, behavior space area, behavior occurrence frequency, and behavior time concentration corresponding to the label, obtain a behavior feature set, and extract key feature dimensions through high-dimensional deconstruction, which helps to improve the accuracy and diversity of the model in behavior feature representation and avoid the problems of blurred or overlapping behavior patterns caused by low-dimensional descriptions; the behavior structure unit is used to construct ternary structure data for each behavior label in the behavior feature unit set. The ternary structure data includes a time period field, a space area field, and a behavior feature field, and converts the unstructured behavior feature results into a standardized structure body to ensure the consistency of behavior data expression and facilitate subsequent processing; the behavior sequence unit is used to combine all ternary structure data in chronological order, identify the behavior patterns of the user in different time periods and different space areas, obtain a behavior structure sequence, organize the discrete behavior labels into a continuous sequence, and mine the temporal rules between behaviors, providing a complete semantic basis for subsequent periodicity recognition; the model construction unit is used to define the ternary structures with repeated patterns in the behavior structure sequence as periodic behaviors according to the behavior structure sequence, convert them into rule expression forms, and generate an individual behavior physical model, convert the abstract behavior chain into a periodic rule expression, realize the conversion from data to knowledge, and enable the system to have the long-term memory function of the user's specific behaviors.

[0098] Among them, the behavior sequence unit is used to combine all ternary structure data in chronological order, identify the behavior patterns of the user in different time periods and different space areas, obtain a behavior structure sequence, and specifically includes:

[0099] The system first receives multiple ternary structure data output from the behavior structure unit. Each ternary structure includes a time period field, a space area field, and a behavior feature field of the user's behavior. The system sorts these structure data according to the time information contained in the time period field to ensure that the arrangement of the behavior data conforms to the actual occurrence order of the user's daily activities. When sorting, the system uses the start time of the time period of all behavior labels as the main sequence index, linearly arranges multiple behavior labels, and constructs an initial behavior sequence covering the complete daily cycle.

[0100] After completing the time sorting, the system further identifies the continuity of behaviors in terms of space and behavior features in this ordered sequence. Specifically, the system determines whether the time interval between two consecutive behaviors is less than a set behavior switching window threshold, such as 10 minutes. If the interval is short and there is a logical connection in the space area, such as from the bedroom to the bathroom, it is considered that the two behaviors have a behavior linkage relationship. The system combines such consecutive behaviors into a behavior paragraph, and at the same time records their order of appearance and space transfer path in a day.

[0101] After all ternary structures are organized into intraday behavior segments, the system counts the user's behavior sequences over multiple consecutive days and extracts stable and recurring behavior patterns. For example, if the user shows a sequence of "getting up at 07:00 - entering the bathroom at 07:10 - entering the kitchen at 07:25" on multiple weekdays, the system will identify this sequence as a typical morning behavior pattern and record it as part of the behavior structure sequence. The construction of the behavior structure sequence not only includes the structural attributes of each behavior but also its relative position and the dependency relationships before and after in the full-day behavior, finally forming a set of sequence expression results with temporality, spatial distribution, and behavior concentration to support subsequent cycle recognition and individual model construction tasks. This process significantly enhances the system's ability to understand the user's complex behavior flow and provides sufficient context for identifying fine-grained behavior variations.

[0102] In a preferred embodiment of the present invention, the model construction unit includes:

[0103] A candidate behavior unit for performing repetitive matching on the ternary structure data in the behavior structure sequence, classifying the time periods and spatial regions with a preset coincidence rate exceeding a certain value and having behavior characteristics with a preset coincidence rate exceeding a certain value in multiple consecutive days into the same candidate behavior segment to obtain a candidate behavior segment set;

[0104] A cycle recognition unit for statistically calculating the occurrence frequency and time concentration of each candidate behavior segment based on the candidate behavior segment set. When its occurrence frequency is greater than a preset frequency threshold and the time concentration is less than a preset concentration threshold, the behavior segment is identified as a periodic behavior;

[0105] A rule generation unit for reconstructing the patterns of the time period field, spatial region field, and behavior characteristic field in the periodic behavior, extracting the average occurrence time, main region coordinates, and stable behavior characteristics to obtain a structured behavior set;

[0106] A physical modeling unit for determining the user's behavior pattern within a specific time period and spatial region based on the structured behavior set to obtain an individual behavior physical model.

[0107] In an embodiment of the present invention, a candidate behavior unit is configured to perform repetitive matching on the ternary structure data in the behavior structure sequence, and classify the time periods and spatial regions that exceed a preset coincidence rate in consecutive multiple days and the ternary structures with behavior characteristics that exceed the preset coincidence rate into the same candidate behavior segment, obtaining a candidate behavior segment set. By identifying the segments with high similarity in the behavior sequence, a stable behavior pattern in the user's daily life is determined; a period recognition unit is configured to, according to the candidate behavior segment set, statistically calculate the occurrence frequency and time concentration of each candidate behavior segment. When its occurrence frequency is greater than a preset frequency threshold and the time concentration is less than a preset concentration threshold, the behavior segment is identified as a periodic behavior. By setting the dual conditions of frequency and concentration, the system effectively avoids misjudgment caused by accidental behavior fluctuations and improves the scientificity and stability of periodic behavior recognition; a rule generation unit is configured to perform pattern reconstruction on the time period field, spatial region field, and behavior characteristic field in the periodic behavior, extract the average occurrence time, main region coordinates, and stable behavior characteristics, obtaining a structured behavior set, realizing the abstraction and standardization of periodic behaviors, converting complex and variable behavior sequences into general rule templates, facilitating the quick call and matching of the system; a physical modeling unit is configured to, according to the structured behavior set, determine the behavior pattern of the user in a specific time period and spatial region, obtaining an individual behavior physical model. By constructing a complete and dynamically updated individual behavior physical model, the problem that traditional models rely on static behavior templates and lack personalized expression is solved.

[0108] Among them, the rule generation unit is configured to perform pattern reconstruction on the time period field, spatial region field, and behavior characteristic field in the periodic behavior, extract the average occurrence time, main region coordinates, and stable behavior characteristics, obtaining a structured behavior set, specifically including:

[0109] After receiving the periodic behavior output by the period recognition unit, first perform a parsing operation on the time period field in the periodic behavior. In multiple periodic behavior segments, the system extracts the start time and duration of all occurrence time periods, statistically averages the start times to form a unified average occurrence time parameter, and at the same time calculates the standard deviation of the fluctuation range of the duration and sets this value as the time tolerance for setting the time matching threshold in subsequent behavior determination. In this process, the system uses a sliding window mechanism to densely sample the behavior characteristics in different time periods to ensure that the obtained average value is representative and avoid deviating from the user's true behavior concentration due to extreme value interference.

[0110] Next, the system performs an aggregation operation on the spatial region fields involved in the periodic behavior segments. By analyzing the spatial location information associated with different behavior segments, the system identifies the region with the highest frequency of occurrence as the main region coordinates of the periodic behavior. If there may be switches or parallel operations in multiple spatial regions during the same periodic behavior, such as brief movements between the kitchen and the living room, the system will use the position density function to determine the core center of gravity of the behavior activities and define this region as the main behavior region, rather than simply based on frequency statistics, so as to enhance the robustness of spatial recognition. During the process of determining the main region coordinates, the system also maps the sensor numbers or geographical tags to spatial coordinate identifiers to ensure the consistency of the model structure and the callability of subsequent algorithm modules.

[0111] In terms of the behavior feature fields, the system focuses on extracting the behavior stability parameters reflected in the periodic behavior, including but not limited to feature dimensions such as action intensity, behavior frequency, and physiological indicators. For these numerical features, the system uses multiple rounds of aggregation and filtering algorithms to eliminate noise and sporadic deviation data, and retains the set of behavior features with high confidence. Then, through methods such as mean fusion or principal component analysis, the most representative and least variable stability parameters of the behavior in the periodic sequence are extracted as the behavior stability features.

[0112] Finally, the processing results of the above three dimensions, the average occurrence time, the main region coordinates, and the stable behavior features, are structurally encapsulated to form a structured behavior set in a unified format. This not only can standardize the expression of the user's periodic behavior, but also provides a matching template and reference benchmark for subsequent deviation recognition, trend judgment, and anomaly detection, thus greatly improving the reliability and adaptability of the entire monitoring system in practical applications.

[0113] Among them, the physical modeling unit is used to determine the behavior pattern of the user within a specific time period and spatial region according to the structured behavior set, and obtain the individual behavior physical model, specifically including:

[0114] After receiving the structured behavior set, first, a behavior templatization process is performed on each structured rule in it. The system will map the average occurrence time and its tolerance range to a unified time axis according to the time period field in the structured behavior set, and construct the daily behavior time framework of the user. The system sorts all rule entries in chronological order to form a complete time period sequence, and constructs the intraday concentration map of the user's behavior based on this. This map can clearly describe the activity distribution of the user in each time period of a day and is an important basis for subsequent model matching and deviation calculation.

[0115] Subsequently, the system performs spatial layout on the spatial region fields in the structured behavior set, maps the main region coordinates to the physical locations in the actual monitoring environment, and constructs a two-dimensional time-space behavior scenario model. If multiple behavior rules are continuous or overlapping, the system will represent them in the form of a graph structure, where each node represents a specific behavior scenario, such as kitchen activities within a certain time period, and the edges represent the transition relationships between behaviors, such as the movement path from the bedroom to the kitchen. Based on this, the system identifies the common spatial movement paths of users and their time series patterns through graph traversal algorithms, enhancing the understanding ability of dynamic behavior flows.

[0116] For the behavior feature field part, the system injects stable behavior feature parameters into each time-space node and performs clustering analysis according to the correlation between parameter features to identify typical behavior pattern combinations. For example, medium-intensity activities and high-frequency physiological parameter fluctuations within a certain time period can be associated with the exercise behavior pattern. The system will label them with specific behavior tags and perform unified coding. The system will also perform frequency statistics and concentration fitting on the combinations of these behavior tags to construct multiple typical behavior units, and on this basis, construct a periodic behavior template library.

[0117] Finally, the above time series, spatial trajectories, and behavior parameters are fused to form a unified individual behavior physical model. This individual behavior physical model is essentially a multi-dimensional vector structure with multiple dimensions such as time dimension, space dimension, behavior dimension, and feature dimension, and is used to represent the standardized behavior performance of users in specific space-time scenarios. This individual behavior physical model can not only be used for the comparison and deviation judgment of target behaviors, but also supports trend prediction, anomaly assessment, and model dynamic update during the behavior evolution process, so as to achieve high reliability, strong adaptability, and interpretability of individual behavior modeling in long-term monitoring scenarios.

[0118] In a preferred embodiment of the present invention, the deviation calculation module includes:

[0119] The first calculation unit is used to compare the target feature data with the time field, space field, and behavior field of the corresponding behavior tag in the individual behavior physical model, calculate the differences of the behavior feature sets respectively to obtain a behavior deviation index set, and calculate the deviation degree according to it;

[0120] The anomaly recognition unit is used to compare the deviation degree with a preset deviation threshold. When the deviation degree is greater than the preset deviation threshold, it is determined that the behavior tag of the corresponding target feature data is a suspicious behavior tag;

[0121] The behavior recording unit is used to monitor the suspicious behavior tags, record the first occurrence time, the last occurrence time, the continuous occurrence times, and the spatial distribution range of the suspicious behavior tags to obtain an observation tag group.

[0122] In an embodiment of the present invention, a first calculation unit is configured to compare target feature data with the time field, space field, and behavior field corresponding to the behavior label in the individual behavior physical model, calculate the differences of the behavior feature sets respectively to obtain a behavior deviation index set, and calculate the deviation degree according to it, which can simply and accurately reflect whether the behavior is essentially deviated from the historical model, effectively enhancing the judgment efficiency and response speed of the system to behavior variations; an anomaly recognition unit is configured to compare the deviation degree with a preset deviation threshold, and when the deviation degree is greater than the preset deviation threshold, determine that the behavior label corresponding to the target feature data is a suspicious behavior label, which can quickly respond to behaviors with significant deviation features and screen out high-risk behavior events, thereby realizing the preliminary filtering and extraction from normal behaviors to abnormal behaviors; a behavior recording unit is configured to monitor the suspicious behavior label, record the first occurrence time, the last occurrence time, the continuous occurrence times, and the spatial distribution range of the suspicious behavior label to obtain an observation label group, which can systematically save the evolution path and spatio-temporal distribution of the target behavior, thus providing solid data support for trend behavior analysis and abnormal evolution pattern recognition.

[0123] In a preferred embodiment of the present invention, the first calculation unit includes:

[0124] A second calculation unit is configured to obtain a first deviation index term according to the product relationship among the behavior time period, behavior frequency, and behavior concentration of the user's recent behavior, combined with the inhibitory effect of behavior concentration on time-frequency offset; obtain a second deviation index term according to the interaction between the behavior duration and behavior spatial position of the user's recent behavior, combined with the adjustment function of spatial offset on the duration fluctuation; obtain a third deviation index term according to the joint relationship among the behavior time period, behavior concentration, and behavior duration of the user's recent behavior, combined with the interference degree of frequency offset on the joint relationship;

[0125] Fuse the first deviation index term, the second deviation index term, and the third deviation index term to obtain the deviation degree of the user's recent behavior relative to the individual behavior physical model.

[0126] In the embodiments of the present invention, according to the product relationship among the behavior time period, behavior frequency, and behavior concentration of the user's recent behaviors, and in combination with the inhibitory effect of behavior concentration on time-frequency offset, a first deviation index term is obtained to effectively identify the concentration misalignment abnormality caused by physiological changes; according to the interaction between the behavior duration and behavior spatial position of the user's recent behaviors, and in combination with the adjustment function of spatial offset on the duration fluctuation, a second deviation index term is obtained, and through the coupling analysis of behavior duration and spatial distribution, the fineness of the system in identifying spatial behavior abnormalities is improved; according to the joint relationship among the behavior time period, behavior concentration, and behavior duration of the user's recent behaviors, and in combination with the interference degree of frequency offset on this joint relationship, a third deviation index term is obtained to effectively reveal the behavior characteristics with stable concentration but strong interference, and the timeliness and accuracy of the system in identifying mutant behaviors are improved; the first deviation index term, the second deviation index term, and the third deviation index term are fused to obtain the deviation degree of the user's recent behaviors relative to the individual behavior physical model, eliminating the judgment distortion problem caused by single-dimensional deviation, and providing a stable and comprehensive abnormal assessment reference in the case of the coexistence of multiple behavior changes.

[0127] Among them, the calculation formula for the deviation degree is: ,

[0128] Among them, is the deviation degree of the user's recent behaviors, is the difference between the behavior occurrence time period of the user's recent behaviors and the individual behavior physical model, , is the behavior occurrence time period of the user's recent behaviors, is the behavior occurrence time period of the individual behavior physical model, is the difference between the behavior occurrence frequency of the user's recent behaviors and the individual behavior physical model, , is the behavior occurrence frequency of the user's recent behaviors, is the behavior occurrence frequency of the individual behavior physical model, is the difference between the behavior time concentration of the user's recent behaviors and the individual behavior physical model, , is the behavior time concentration of the user's recent behaviors, is the behavior time concentration of the individual behavior physical model, is the difference between the behavior duration of the user's recent behaviors and the individual behavior physical model, , is the behavior duration of the user's recent behaviors, is the behavior duration of the individual behavior physical model, It is the difference between the coordinates of the behavior space region of the user's recent behavior and the physical model of individual behavior. , is the coordinate of the behavior space region of the user's recent behavior, and is the coordinate of the behavior space region of the physical model of individual behavior.

[0129] Among them, when calculating the deviation degree of the behavior, the differences of each behavior field are uniformly normalized, that is, the normalization is carried out by using the reference value of the corresponding data in the physical model of individual behavior. The differences in time, frequency, space, etc. between the current behavior and the physical model of individual behavior are respectively divided by the standard value of this field in the physical model of individual behavior, so as to be converted into a dimensionless offset ratio with consistent units and dimensions.

[0130] In a preferred embodiment of the present invention, the anomaly detection module includes:

[0131] A trend extraction unit, which is used to extract its continuous occurrence records in the time dimension according to the suspicious behavior labels in the observation label group to obtain a trend sample set;

[0132] A trend analysis unit, which is used to calculate the average frequency and average deviation degree of the suspicious behavior labels in the continuous records according to the trend sample set to obtain a trend feature index set;

[0133] A judgment rule unit, which is used to judge whether the average frequency in the trend feature index set exceeds the behavior occurrence frequency of the same type of labels in the physical model of individual behavior, and at the same time judge whether the average deviation degree exceeds the preset deviation threshold. When any judgment result is yes, it is determined that the behavior label does not meet the absorption condition;

[0134] An anomaly judgment unit, which is used to judge whether there is a trend reversal of the suspicious behavior label in the continuous time window. When there is no trend reversal, it is confirmed that the suspicious behavior label is an abnormal behavior.

[0135] In an embodiment of the present invention, a trend extraction unit is configured to extract consecutive occurrence records of suspicious behavior tags in the time dimension according to the suspicious behavior tags in the observation tag group, so as to obtain a trend sample set, which can screen out trend behaviors with time evolution characteristics from a large number of suspicious behavior tags and avoid misjudgment of short-term occasional behaviors; a trend analysis unit is configured to calculate the average frequency and average deviation degree of the suspicious behavior tags in the consecutive records according to the trend sample set, so as to obtain a trend feature index set, and convert the trend behavior from a perceptual observation into a set of indicators that can be numerically processed through a quantification operation, so that the subsequent judgment logic has a quantitative basis; a judgment rule unit is configured to judge whether the average frequency in the trend feature index set exceeds the behavior occurrence frequency of the same type of tag within the individual behavior physical model, and at the same time judge whether the average deviation degree exceeds a preset deviation threshold. When any one of the judgment results is yes, it is determined that the behavior tag does not meet the absorption condition. Through a double-threshold judgment mechanism, it is possible to avoid misidentifying a behavior that temporarily increases due to environmental changes as abnormal, and at the same time, it can also identify a trend behavior that changes slowly but deviates from stability; an anomaly judgment unit is configured to judge whether there is a trend reversal of the suspicious behavior tag within a consecutive time window. When there is no trend reversal, it is confirmed that the suspicious behavior tag is an abnormal behavior, which effectively increases the buffer for the system to judge abnormal behaviors and avoids false alarms caused by premature determination of abnormalities due to momentary deviations, thereby improving the judgment stability.

[0136] Among them, the anomaly judgment unit is configured to judge whether there is a trend reversal of the suspicious behavior tag within a consecutive time window. When there is no trend reversal, it is confirmed that the suspicious behavior tag is an abnormal behavior, and specifically includes:

[0137] After receiving the result from the judgment rule unit, that is, a certain suspicious behavior tag is determined not to meet the absorption condition, it enters the trend reversal analysis stage of this tag. This operation process depends on systematically tracking and statistically analyzing the evolution trend of this behavior within a consecutive time window. The goal is to judge whether this behavior has a self-correction trend, that is, whether its behavior performance returns to the normal fluctuation range of the individual behavior model in subsequent time periods, so as to avoid misjudgment caused by short-term behavior disturbances.

[0138] The system first sets a trend monitoring time window, which generally covers a consecutive number of days after the user's behavior occurs, such as 7 days, 10 days, or 14 days. All records of the suspicious behavior label are extracted within this window, and a trend trajectory is constructed in chronological order. For each record, the system obtains its corresponding behavior frequency and deviation degree value, and compares and analyzes them with the average value in the trend sample set. If within this window, the system monitors that the frequency of this behavior shows a continuous downward trend, for example, the behavior frequency decreases day by day within three days, or is lower than 50% of the trend average frequency at most time points, then the system considers that the behavior intensity is weakening; at the same time, if the behavior deviation degree also shows a decreasing trend in subsequent records, especially being lower than the deviation threshold set by the individual behavior physical model for several consecutive times, it indicates that the behavior deviation state has gradually returned to the model baseline range.

[0139] Furthermore, the system performs trend fitting processing on the frequency curve and deviation curve, calculates their change slopes and fluctuation intervals. If the overall trend slope is negative and the fluctuation amplitude is lower than the set change tolerance value, it is determined that the behavior has a trend reversal; at this time, the system retains this label in the observation state and does not perform abnormal processing. However, if any of the above trends does not meet the reversal standard, and the system confirms that this behavior has stably maintained a high frequency or continuous deviation degree within a certain period and shows no self-repair trend, then this behavior label is confirmed as an abnormal behavior and marked as an object that needs to be suppressed or intervened, and is transmitted to the behavior suppression module for subsequent processing. This operation strategy ensures that the system has a certain flexible judgment space in the process of abnormal behavior recognition, avoids misclassifying behavior fluctuations into the abnormal category, and improves the judgment stability of the health monitoring system and the reliability of clinical intervention.

[0140] In a preferred embodiment of the present invention, the behavior suppression module includes:

[0141] A suppression factor unit, which is used to analyze the minimum confidence intensity required for an abnormal behavior to be absorbed by the individual behavior physical model according to the deviation degree and duration of the abnormal behavior, and obtain a behavior suppression factor;

[0142] An absorption judgment unit, which is used to compare the behavior suppression factor with the absorption trigger intensity of the individual behavior physical model. When the behavior suppression factor is greater than the absorption trigger intensity, then this abnormal behavior does not meet the absorption condition;

[0143] An absorption adjustment unit, which is used to dynamically limit the behavior label related to this abnormal behavior in the individual behavior physical model, and increase the time length of the update window of this behavior label in the individual behavior physical model;

[0144] A behavior blocking unit, which is used to add a blocking status identifier to the abnormal behavior label that does not meet the absorption condition, and store this behavior label, its deviation degree, and the behavior suppression factor in the behavior adjustment data.

[0145] In an embodiment of the present invention, an inhibition factor unit is configured to analyze the minimum confidence intensity required for an abnormal behavior to be absorbed by an individual behavior physical model according to the deviation degree and duration of the abnormal behavior, so as to obtain a behavior inhibition factor, realizing a quantitative expression of the confidence degree of the abnormal behavior, enabling the system to make a judgment basis on whether the behavior should be absorbed by the model through a quantitative index; an absorption judgment unit is configured to compare the behavior inhibition factor with the absorption trigger intensity of the individual behavior physical model. When the behavior inhibition factor is greater than the absorption trigger intensity, the abnormal behavior does not meet the absorption condition, ensuring that only behaviors with small deviations and stability will have a structural impact on the individual behavior physical model; an absorption adjustment unit is configured to dynamically limit the behavior label related to the abnormal behavior in the individual behavior physical model, and increase the time length of the update window of the behavior label in the individual behavior physical model. By restricting the premature acceptance of potential abnormal behaviors by the model, the observation period of behavior evolution is increased, the probability of mis-absorption is reduced, and at the same time, the tolerance for sudden behaviors is retained; a behavior blocking unit is configured to add a blocking status identifier to the abnormal behavior label that does not meet the absorption condition, and store the behavior label, its deviation degree, and the behavior inhibition factor in the behavior adjustment data, constructing a blacklist mechanism for abnormal behaviors in the system to ensure that the behavior will not repeatedly trigger model mis-update in the short term.

[0146] Among them, the absorption adjustment unit is configured to dynamically limit the behavior label related to the abnormal behavior in the individual behavior physical model, and increase the time length of the update window of the behavior label in the individual behavior physical model, specifically including:

[0147] The system will first identify the abnormal behavior label based on the previous module, and combine the current status information of the label, such as the behavior occurrence frequency, deviation degree, duration, etc., to retrieve the behavior structure unit in the individual behavior physical model that has a structural mapping relationship with the behavior label. This behavior structure unit usually contains three - field information, namely a time - period field, a space - region field, and a behavior - feature field. After establishing a correspondence relationship with the historical behavior sequence in the model through the three - element structure, the system will determine whether the label already has the identity of a candidate segment with periodic behavior characteristics, and consult the behavior update window parameters formed during the model construction stage.

[0148] Next, the system will compare the suppression factor strength previously calculated by the suppression factor unit with the default absorption update strategy of the current behavior label in the individual model; if the suppression factor is greater than the model absorption trigger threshold, the system will execute the window parameter adjustment mechanism, that is, extend the observation period of this behavior label in the individual model. This observation period is defined by the system as the time range in which this label must meet the conditions of "low deviation - high persistence - stable concentration". In actual operation, the system will overwrite the update time threshold of this behavior label in the behavior model. For example, if the originally set update period is 3 days, it will currently be adjusted to 7 days, 10 days or even longer. The specific extended value is dynamically calculated proportionally according to the suppression factor strength. The higher the suppression factor, the longer the window, thereby reducing the risk of it being quickly updated to the model structure.

[0149] While the update window is extended, the system will also synchronously reduce the active weight of this label in the model structure sequence, that is, by adding dynamic scaling factors to the behavior frequency term and concentration distribution term of this label, making the weight it occupies in the next round of periodic behavior recognition decrease, avoiding false triggering of the update mechanism due to frequent occurrences in the short term. The entire process does not modify the original label content, but realizes the regulation of the behavior absorption path through the dynamic restriction of parameters, thus forming a gentle and continuous behavior structure restriction method, which not only retains the possibility of the label to continue to be observed, but also effectively curbs the potential risk of it misleading the model evolution, ensuring the long-term stability of the individual behavior model and the accuracy of behavior discrimination.

[0150] In a preferred embodiment of the present invention, the suppression factor unit includes:

[0151] A suppression factor calculation unit, which is used to calculate the change impact of the abnormal behavior on the time distribution and trust based on the ratio of the time offset pair value of the abnormal behavior to its coverage ratio in the individual behavior physical model, and obtain the main control factor term; calculate the immediate interference ability of the short-term abnormal behavior on the individual behavior physical model based on the duration difference of the abnormal behavior, and obtain the linear adjustment factor term; calculate the overall suppression amplitude based on the ratio of the concentration offset value to the model coverage and the ratio of the frequency offset value to the behavior duration, and obtain the balance adjustment term;

[0152] Fuse the main control factor term, the linear adjustment factor term and the balance adjustment term to obtain the behavior suppression factor of the abnormal behavior.

[0153] In the embodiment of the present invention, the inhibition factor calculation unit is configured to calculate the change impact of abnormal behavior on time distribution and confidence according to the ratio of the time offset pair value of the abnormal behavior to its coverage ratio in the individual behavior physical model, so as to obtain the main control factor item, which can avoid the problem of imbalance in inhibition intensity caused by extreme time differences and improve the discrimination accuracy of time-sensitive abnormal behaviors; calculate the immediate interference ability of short-term abnormal behaviors on the individual behavior physical model according to the duration difference of the abnormal behavior, so as to obtain the linear adjustment factor item, and convert the time difference into a quantitative index of interference ability, which can effectively evaluate short-term or overly long abnormalities; calculate the overall inhibition amplitude according to the ratio of the concentration offset value to the model coverage and the ratio of the frequency offset value to the behavior duration, so as to obtain the balance adjustment item, and form a quantitative deviation ratio through concentration and frequency, so as to accurately evaluate the impact of abnormal behavior on the structural stability of the model; fuse the main control factor item, the linear adjustment factor item and the balance adjustment item to obtain the behavior inhibition factor of the abnormal behavior, and establish a unified evaluation scale in the behavior space, time dimension and concentration characteristics, which overcomes the problem of excessive dependence of traditional systems on one-dimensional data.

[0154] Among them, the calculation formula of the behavior inhibition factor is: ,

[0155] Among them, is the behavior inhibition factor of the abnormal behavior, is the deviation degree of the abnormal behavior, is the coverage rate of the behavior label of the same type as the abnormal behavior in the individual behavior physical model, is the duration of the abnormal behavior, is a coefficient.

[0156] Among them, , , and are the same as the data in the calculation formula of the deviation degree, is the difference between the time period of the user's abnormal behavior and the time period of the behavior occurrence in the individual behavior physical model, , is the time period of the user's abnormal behavior occurrence, is the time period of the behavior occurrence in the individual behavior physical model, is the difference between the behavior occurrence frequency of the user's abnormal behavior and the behavior occurrence frequency in the individual behavior physical model, , is the behavior occurrence frequency of the user's abnormal behavior, is the behavior occurrence frequency in the individual behavior physical model, is the difference between the behavior time concentration of the user's abnormal behavior and the behavior time concentration in the individual behavior physical model, , is the concentration of behavior time of user abnormal behavior, is the concentration of behavior time of the individual behavior physical model, is the difference between the duration of user abnormal behavior and the duration of the individual behavior physical model, , is the duration of user abnormal behavior, is the duration of the individual behavior physical model.

[0157] The above are the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A personal health abnormal behavior monitoring system integrating unsupervised learning, characterized in that, The system includes: An unsupervised module, which is used to obtain historical behavior data of a user, perform clustering processing on the historical behavior data, divide the historical behavior data into multiple behavior labels, and obtain behavior label data; An individual model module, which is used to extract feature vectors of each behavior label according to the behavior label data, obtain sample feature data, and establish an individual behavior physical model according to the sample feature data; A target behavior module, which is used to obtain the target behavior data of the user in the near future, perform clustering and feature extraction on the target behavior data, and obtain target feature data; A deviation calculation module, which is used to input the target feature data into the individual behavior physical model, calculate the deviation degree of the user's recent behavior, and monitor and record the abnormal behavior labels whose deviation degree exceeds a preset deviation threshold, and obtain an observation label group; An anomaly detection module, which is used to determine whether the observation label group meets the conditions for absorption by the individual behavior physical model. When the result is negative, the behavior label is determined to be an abnormal behavior; A behavior inhibition module, which is used to calculate a behavior inhibition factor according to the deviation degree and duration of the abnormal behavior, and dynamically adjust the individual behavior physical model according to the behavior inhibition factor to limit the abnormal behavior from being recognized as a normal behavior.

2. The personal health abnormal behavior monitoring system integrating unsupervised learning according to claim 1, characterized in that, The unsupervised module includes: A time division unit, which is used to perform division processing on the historical behavior data, split the historical behavior data into multiple time segments according to a preset time window, and obtain a time segment data set; A feature extraction unit, which is used to perform feature extraction on the time segment data set, extract the residence ratio, spatial transfer frequency, and associated physiological values of the user in different spatial regions within each time segment, and obtain a feature data set; A clustering unit, which is used to perform clustering processing on the preliminary feature data set, classify time segments with similar features into the same clustering cluster, and obtain clustering result data; A behavior mapping unit, which is used to perform behavior analysis on each clustering cluster in the clustering result data, identify the significant features of the clustering cluster in terms of spatial behavior and physiological parameters, and map the significant features to behavior labels to obtain behavior label data.

3. The personal health abnormal behavior monitoring system integrating unsupervised learning according to claim 2, characterized in that, The behavior mapping unit includes: A statistical analysis unit, which is used to perform statistics on the time segment data included in each clustering cluster in the clustering result data, calculate the spatial behavior parameters of the user in each spatial region and the physiological parameters under each spatial region within the clustering cluster respectively, and obtain a clustering cluster feature index set; A significant feature identification unit, which is used to perform standardization processing on the spatial behavior parameters and physiological parameters in the clustering cluster feature index set, and screen out the spatial regions and physiological value intervals with significant features according to a preset behavior feature threshold, and obtain a significant feature set; A label classification unit, which is used to match the significant feature set with a preset behavior pattern, select a label according to the spatial region and physiological value interval, determine the behavior label type that best matches the clustering cluster, and bind the behavior label type to the clustering cluster to obtain behavior label data.

4. The personal health abnormal behavior monitoring system integrating unsupervised learning according to claim 3, characterized in that, The individual model module includes: A behavior decomposition unit, which is used to perform sample analysis on each behavior label in the behavior label data, extract the behavior occurrence time period, behavior duration, behavior spatial region, behavior occurrence frequency, and behavior time concentration corresponding to the label, and obtain a behavior feature set; A behavior structure unit, which is used to construct ternary structure data for each behavior label in the behavior feature unit set. The ternary structure data includes a time period field, a spatial region field, and a behavior feature field; A behavior sequence unit, which is used to combine all the ternary structure data in chronological order, identify the behavior patterns of the user in different time periods and different spatial regions, and obtain a behavior structure sequence; A model construction unit, which is used to define the ternary structures with repeated patterns in the behavior structure sequence as periodic behaviors according to the behavior structure sequence, and convert them into a regular expression form to generate an individual behavior physical model.

5. The personal health abnormal behavior monitoring system integrating unsupervised learning according to claim 4, characterized in that, The model construction unit includes: A candidate behavior unit, which is used to perform repetitive matching on the ternary structure data in the behavior structure sequence, and classify the ternary structures with time periods and spatial regions with a preset coincidence rate exceeding the preset value for multiple consecutive days and having behavior features with a preset coincidence rate exceeding the preset value into the same candidate behavior segment to obtain a candidate behavior segment set; A period recognition unit, which is used to statistically calculate the occurrence frequency and time concentration of each candidate behavior segment according to the candidate behavior segment set. When its occurrence frequency is greater than a preset frequency threshold and the time concentration is less than a preset concentration threshold, the behavior segment is recognized as a periodic behavior; A rule generation unit, which is used to reconstruct the patterns of the time period field, the spatial region field, and the behavior feature field in the periodic behavior, extract the average occurrence time, the main region coordinates, and the stable behavior features to obtain a structured behavior set; A physical modeling unit, which is used to determine the behavior pattern of the user in a specific time period and spatial region according to the structured behavior set to obtain an individual behavior physical model.

6. The personal health abnormal behavior monitoring system integrating unsupervised learning according to claim 5, characterized in that, The deviation calculation module includes: A first calculation unit, which is used to compare the target feature data with the time field, the spatial field, and the behavior field of the corresponding behavior label in the individual behavior physical model, calculate the differences of the behavior feature sets respectively to obtain a behavior deviation index set, and calculate the deviation degree according to it; An anomaly recognition unit, which is used to compare the deviation degree with a preset deviation threshold. When the deviation degree is greater than the preset deviation threshold, the behavior label of the corresponding target feature data is determined as a suspicious behavior label; A behavior recording unit, which is used to monitor the suspicious behavior label, record the first occurrence time, the last occurrence time, the consecutive occurrence times, and the spatial distribution range of the suspicious behavior label to obtain an observation label group.

7. The personal health abnormal behavior monitoring system integrating unsupervised learning according to claim 6, characterized in that, The first calculation unit includes: A second calculation unit, which is used to obtain a first deviation index term according to the product relationship among the behavior time period, the behavior frequency, and the behavior concentration of the user's recent behavior, combined with the inhibitory effect of the behavior concentration on the time-frequency offset; obtain a second deviation index term according to the interaction between the behavior duration and the behavior spatial position of the user's recent behavior, combined with the adjustment function of the spatial offset on the duration fluctuation; obtain a third deviation index term according to the joint relationship among the behavior time period, the behavior concentration, and the behavior duration of the user's recent behavior, combined with the interference degree of the frequency offset on the joint relationship; Fuse the first deviation index item, the second deviation index item, and the third deviation index item to obtain the degree of deviation of the user's recent behavior from the individual behavior physical model.

8. The personal health abnormal behavior monitoring system integrating unsupervised learning according to claim 7, characterized in that, The anomaly detection module includes: A trend extraction unit, configured to extract its consecutive occurrence records in the time dimension according to the suspicious behavior labels in the observation label group to obtain a trend sample set; A trend analysis unit, configured to calculate the average frequency and average deviation degree of the suspicious behavior labels in the consecutive records according to the trend sample set to obtain a trend feature index set; A judgment rule unit, configured to judge whether the average frequency in the trend feature index set exceeds the behavior occurrence frequency of the same type of labels in the individual behavior physical model, and at the same time judge whether the average deviation degree exceeds a preset deviation threshold. When any judgment result is yes, it is determined that the behavior label does not meet the absorption condition; An anomaly judgment unit, configured to judge whether there is a trend reversal of the suspicious behavior labels within a consecutive time window. When there is no trend reversal, it is confirmed that the suspicious behavior labels are abnormal behaviors.

9. The personal health abnormal behavior monitoring system integrating unsupervised learning according to claim 8, characterized in that, The behavior inhibition module includes: An inhibition factor unit, configured to analyze the minimum confidence intensity required for the abnormal behavior to be absorbed by the individual behavior physical model according to the deviation degree and duration of the abnormal behavior to obtain a behavior inhibition factor; An absorption judgment unit, configured to compare the behavior inhibition factor with the absorption trigger intensity of the individual behavior physical model. When the behavior inhibition factor is greater than the absorption trigger intensity, the abnormal behavior does not meet the absorption condition; An absorption adjustment unit, configured to dynamically limit the behavior labels related to the abnormal behavior in the individual behavior physical model and increase the time length of the update window of the behavior labels in the individual behavior physical model; A behavior blocking unit, configured to add a blocking status flag to the abnormal behavior labels that do not meet the absorption condition, and store the behavior labels, their deviation degrees, and behavior inhibition factors in the behavior adjustment data.

10. The personal health abnormal behavior monitoring system integrating unsupervised learning according to claim 9, characterized in that, The inhibition factor unit includes: An inhibition factor calculation unit, configured to calculate the change impact of the abnormal behavior in terms of time distribution and trust according to the numerical value of the time offset of the abnormal behavior and its coverage ratio in the individual behavior physical model to obtain a main control factor item; calculate the immediate interference ability of the short-term abnormal behavior on the individual behavior physical model according to the duration difference of the abnormal behavior to obtain a linear adjustment factor item; calculate the overall inhibition amplitude according to the ratio of the concentration offset value to the model coverage and the ratio of the frequency offset value to the behavior duration to obtain a balance adjustment item; Fuse the main control factor item, the linear adjustment factor item, and the balance adjustment item to obtain the behavior inhibition factor of the abnormal behavior.

Images