A cognitive care data intelligent analysis method for an aged care information service system

Abstract

This invention discloses an intelligent data analysis method for cognitive care in elderly care information service systems, relating to the field of smart elderly care health monitoring technology. The method includes extracting a mixed dynamic monitoring data stream of elderly behavior and physiology, and separating it into time-series queues according to the source. Event boundary detection is performed on the behavioral activity data to identify the start and end points of independent activity units. Based on this, synchronized data segments are extracted from the physiological data queue and paired to generate a comprehensive snapshot data package of activity status. After consistency verification, the set of valid data packages is input into an adaptive analysis frequency scheduler to obtain a real-time updated demand scanning time interval. This interval is used to periodically traverse the data package set and extract features to generate a structured service demand feature vector. Finally, a specific elderly care service project suggestion instruction is output through an online service pattern matching engine. This invention achieves accurate data synchronization and on-demand adaptive intelligent analysis and scheduling centered on activity events.

Description

Technical Field

[0001] This invention belongs to the field of smart elderly care health monitoring technology, specifically a cognitive care data intelligent analysis method for elderly care information service systems. Background Technology

[0002] In the current field of smart elderly care and health monitoring, wearable devices and environmental sensors are commonly used to continuously collect behavioral and physiological data from the elderly, forming a mixed data stream. Existing technical solutions typically employ data alignment methods based on fixed time windows. This approach ignores the completeness and boundaries of specific activity events, resulting in rigid data correlations. Segmenting a complete activity into multiple time windows, or including irrelevant quiescent physiological data in the analysis, weakens the intrinsic connection between behavioral context and physiological response. The resulting composite data fragments lack quality and representativeness, directly impacting the accuracy of subsequent needs analysis.

[0003] Existing data analysis models often employ preset, fixed analysis frequencies. This static scheduling strategy lacks flexibility and cannot adjust resources according to changes in the elderly person's actual condition. When the elderly person's condition is stable, continuous high-frequency analysis consumes a large amount of computing resources with minimal benefits; however, when the elderly person's activity patterns are complex or abnormal signs appear, the fixed analysis cycle may fail to capture key changes in time, leading to service response delays. There is a mismatch between the allocation of system computing resources and the fluctuations in the elderly person's actual service needs.

[0004] Furthermore, existing methods only focus on the superficial correlation between behavior and physiological data, lacking a deep understanding of the cognitive state of older adults. For example, the same behavior may be accompanied by different physiological responses due to different cognitive needs. If the cognitive care dimension is ignored and analysis is based solely on fixed rules, it is difficult to distinguish between normal activities and abnormal patterns caused by cognitive impairment, resulting in a lack of targeted service recommendations. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes an intelligent analysis method for cognitive care data in elderly care information service systems, comprising: Extract a mixed dynamic monitoring data stream of behavioral activities and physiological indicators of the elderly; The hybrid dynamic monitoring data stream is separated according to the source attributes of the data to form a behavioral activity data queue and a physiological indicator data queue with time-series characteristics; Event boundary detection is performed on the behavioral activity data queue to identify the start and end times of each independent activity unit of the elderly person; Based on the start and end times, extract physiological indicator data segments synchronized with each independent activity unit from the physiological indicator data queue; Each independent activity unit is paired with its corresponding physiological indicator data segment to generate multiple comprehensive activity state snapshot data packages; An internal consistency check is performed on the set of comprehensive snapshot data packets of activity status, and valid comprehensive snapshot data packets of activity status that meet the consistency criteria are selected. The effective activity status comprehensive snapshot data packet set is input into a preset adaptive analysis frequency scheduler to obtain a real-time updated demand scan time interval; The effective activity status comprehensive snapshot data packet set is periodically traversed and features are extracted using the demand scanning time interval to generate a structured service demand feature vector; The structured service demand feature vector is delivered to an online service pattern matching engine for processing, and the final output is a series of specific elderly care service project suggestion instructions.

[0006] Furthermore, the event boundary detection includes the following steps: Starting from the first data point in the behavioral activity data queue, calculate the magnitude of change between adjacent data points one by one; A dynamic threshold is set for the magnitude of change. When the magnitude of change calculated in a certain instance exceeds the dynamic threshold, the current data point is marked as a potential event boundary point. After the potential event boundary point appears, observe a preset number of subsequent data points. If the values ​​of the subsequent data points tend to stabilize and fluctuate around a new value level, then the potential event boundary point is confirmed as a valid event starting point. The behavioral activity data queue is continuously monitored. When the change exceeds the dynamic threshold again, the next valid event start point is confirmed in the same way, and the time period between the previous valid event start point and the current valid event start point is defined as an independent activity unit. Based on all the identified valid event start points, determine the start and end times of all the independent activity units.

[0007] Furthermore, the dynamic threshold is determined based on the overall activity level of the behavioral activity data queue; Calculate the average energy value of the behavioral activity data queue within a full observation window; Multiply the average energy value by a preset basic sensitivity coefficient to obtain the initial threshold base; Simultaneously calculate the average fluctuation value of the behavioral activity data queue within a recent short time window; The average fluctuation value is used as an adjustment factor to weight and correct the initial threshold base, thereby generating a dynamic threshold for the current detection period. When new data enters the behavioral activity data queue, the average energy value within the complete observation window and the average fluctuation value within the recent short time window are recalculated, and the dynamic threshold is updated based on the recalculated average energy value and average fluctuation value.

[0008] Furthermore, the execution process of the internal consistency check is as follows: For each activity state comprehensive snapshot data packet, calculate the covariance of behavioral activity data and physiological indicator data segments in the time dimension; If the value of the covariance falls within the preset positive correlation range, it is determined that the internal data of the activity status comprehensive snapshot data packet is synchronized and consistent, and is marked as consistent. If the value of the covariance falls within the preset negative correlation range or zero correlation range, it is determined that the internal data of the activity state comprehensive snapshot data packet is asynchronous or contradictory, and is marked as inconsistent. All activity state composite snapshot data packets marked as having passed consistency are aggregated to form the set of valid activity state composite snapshot data packets.

[0009] Furthermore, the internal operations of the adaptive analysis frequency scheduler include: Receive the set of comprehensive snapshot data packets of the effective activity status; The density index of the data stream is obtained by analyzing the number of data packets generated per unit time in the comprehensive snapshot data packet set of the effective activity status; the anomaly index of the data stream is obtained by analyzing the average degree of deviation of the internal index of all data packets in the comprehensive snapshot data packet set of the effective activity status from the preset benchmark value. The density index and the anomaly index of the data stream are input into a pre-trained frequency decision network. The frequency decision network maps and outputs a specific demand scan time interval based on the implicit relationship between the density index and the anomaly index of the data stream.

[0010] Furthermore, the training process of the frequency decision network includes: Collect a comprehensive snapshot data set of effective activity status of multiple samples from historical periods and their corresponding manually set optimal required scanning time intervals to form training sample pairs; For each training sample pair, calculate the density index and the anomaly index of its sample data stream. The initial neural network model is trained under supervision using the density index and anomaly index of the sample data stream of all training sample pairs as input features and the manually set optimal demand scanning time interval as the target label. By iteratively adjusting the internal connection weights of the neural network model, the model can learn the mapping pattern from input features to target labels, ultimately forming the frequency decision network.

[0011] Furthermore, periodic traversal and feature extraction include: Using the required scanning time interval as a fixed time step, the effective activity status comprehensive snapshot data packet set is sampled uniformly over time. At each sampling time point, extract the pattern features of behavioral activity data from the comprehensive snapshot data packet of all activity states within the time window preceding the sampling time point; Simultaneously extract the statistical characteristics of physiological indicator data fragments from the comprehensive snapshot data package of all activity states within the same time window; The pattern features and the statistical features are combined into a multidimensional vector according to a preset vector concatenation rule; As the time step progresses, the multidimensional vectors generated at each sampling point are arranged in chronological order to form the structured service requirement feature vector.

[0012] Furthermore, the online service pattern matching engine processes the data as follows: Establish a predefined knowledge graph of elderly care service items, where each elderly care service item is associated with a set of ideal service demand feature vector patterns; The similarity between the received structured service demand feature vector and all ideal feature vector patterns in the elderly care service project knowledge graph is calculated. Select all ideal feature vector patterns whose similarity calculation results exceed the matching threshold; Find the elderly care service projects that correspond to the selected ideal feature vector pattern, sort them in descending order of similarity, and generate a series of specific elderly care service project suggestion instructions.

[0013] Furthermore, the extraction of the hybrid dynamic monitoring data stream is achieved by integrating multiple heterogeneous data sources, including a passive infrared sensor array deployed in the living space, a pressure distribution sensing system installed on the bed, a multi-parameter health monitoring wristband worn by the elderly, and a smart power consumption monitoring unit for the main electrical appliances in the room; the passive infrared sensor array is used to generate behavioral activity data, the multi-parameter health monitoring wristband and the pressure distribution sensing system are used to generate physiological indicator data, and the smart power consumption monitoring unit is used to provide auxiliary environmental interaction context information. All data are aggregated into the hybrid dynamic monitoring data stream after time synchronization.

[0014] Furthermore, the step of calculating the covariance of the behavioral activity data and physiological indicator data segments over time specifically includes: A unified covariance calculation time window is determined for the current activity status snapshot data packet to be verified. The starting point of the time window is the start time point of the independent activity unit, and the ending point is the end time point of the independent activity unit. Within the covariance calculation time window, the behavioral activity data and the physiological indicator data segments are time-stamp aligned and resampled to form a synchronized data sequence with the same number of time points and the same sampling interval. Calculate the mean of the behavioral activity data sequence and the mean of the physiological indicator data sequence in the synchronized data sequence, respectively. Based on the covariance calculation formula, each pair of synchronously sampled behavioral activity data values ​​and physiological indicator data values ​​in the synchronous data sequence are processed, and the average value of the product of their deviations from their respective means is calculated, thereby obtaining the covariance value in the time dimension.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By detecting event boundaries in behavioral activity data streams, the precise start and end times of independent activity units such as eating, washing, and exercising are identified. Based on this, physiological data segments completely corresponding to these time periods are extracted from continuous physiological indicator data streams. This process achieves dynamic and precise pairing and encapsulation of heterogeneous time-series data with meaningful "activity events" at its core. The resulting comprehensive snapshot data package of activity states has strictly synchronized behavioral information and physiological indicators in time and is highly correlated semantically, ensuring that each data package can reflect the coordinated state of mind and body within a complete activity cycle, thus improving the quality of basic data units and the effectiveness of subsequent feature extraction.

[0016] The validated set of valid data packets is input into an adaptive analysis frequency scheduler. This scheduler dynamically calculates and outputs the time interval for the next feature extraction and required scan based on the pattern, density, or implicit information of the input data. The analysis frequency is no longer a pre-set fixed value but fluctuates in real time according to the data performance of the monitored subject. When the data pattern tends to be stable or simple, the system automatically extends the scan interval to reduce unnecessary computational load; when the data reflects frequent pattern transitions or potential risk characteristics, the interval is automatically shortened to accelerate the analysis cycle. This allows the system's computing resources to focus on the most valuable moments of data change, optimizing overall resource utilization while ensuring timely service response. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of the cognitive care data intelligent analysis method for an elderly care information service system described in this invention. Figure 2 Flowchart for determining dynamic thresholds; Figure 3 A graph showing the trend of data stream density and anomaly over time; Figure 4 Match similarity radar charts for elderly care service projects at different times; Figure 5 A horizontal bar chart matching similarity among elderly care service projects. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1The present invention provides a cognitive care data intelligent analysis method for an elderly care information service system, the overall implementation scheme of which is as follows: A mixed dynamic monitoring data stream of elderly people's behavioral activities and physiological indicators is extracted from various sensors and monitoring devices deployed in the elderly's living environment. This mixed dynamic monitoring data stream is separated according to the data source attributes, forming a behavioral activity data queue and a physiological indicator data queue with time-series characteristics. Event boundary detection is performed on the behavioral activity data queue to identify the start and end times of each independent activity unit of the elderly. Based on these start and end times, physiological indicator data segments synchronized with each independent activity unit are extracted from the physiological indicator data queue. Each independent activity unit is paired with its corresponding physiological indicator data segment to generate multiple comprehensive activity status snapshot data packages. The comprehensive activity status snapshot data package set is then internally consistent, and valid comprehensive activity status snapshot data packages that meet the consistency criteria are selected. The valid comprehensive activity status snapshot data package set is input to a preset adaptive analysis frequency scheduler to obtain a real-time updated demand scanning time interval. This demand scanning time interval is used to periodically traverse and extract features from the valid comprehensive activity status snapshot data package set to generate a structured service demand feature vector. The structured service demand feature vector is delivered to the online service pattern matching engine for processing, and the final output is a series of specific elderly care service project suggestions.

[0020] See Figure 2 In one embodiment of the present invention, event boundary detection starts from the first data point in the behavioral activity data queue and calculates the change amplitude between adjacent data points one by one. A dynamic threshold is set for the change amplitude. When the calculated change amplitude exceeds the dynamic threshold, the current data point is marked as a potential event boundary point. After a potential event boundary point appears, a preset number of subsequent data points are observed. If the values ​​of the subsequent data points tend to stabilize and fluctuate around a new value level, the potential event boundary point is confirmed as a valid event start point. The behavioral activity data queue is continuously monitored. When a change amplitude exceeding the dynamic threshold appears again, the next valid event start point is confirmed using the same method, and the time period between the previous valid event start point and the current valid event start point is defined as an independent activity unit. Based on all identified valid event start points, the start and end times of all independent activity units are determined.

[0021] The dynamic threshold is determined based on the overall activity level of the behavioral activity data queue. The average energy value of the behavioral activity data queue within a complete observation window is calculated, and this average energy value is multiplied by a preset baseline sensitivity coefficient to obtain the initial threshold base. Simultaneously, the average fluctuation value of the behavioral activity data queue within a recent short-term window is calculated, and this average fluctuation value is used as a weighted adjustment factor to adjust the initial threshold base, generating the dynamic threshold for the current detection period. When new data enters the behavioral activity data queue, the average energy value within the complete observation window and the average fluctuation value within the recent short-term window are recalculated, and the dynamic threshold is updated based on these recalculated values. This adaptive adjustment mechanism of the dynamic threshold is particularly suitable for capturing behavioral patterns in older adults caused by changes in cognitive state, such as repetitive short-term activities triggered by anxiety or slowed movements due to cognitive decline. By identifying the boundaries of these abnormal events, the system can provide a data foundation that is closer to the needs of cognitive care for subsequent analysis.

[0022] In practice, the behavioral activity data queue originates from a passive infrared sensor array deployed in the living space. The data points in the queue represent the intensity of elderly people's movement at consecutive time stamps. For example, the behavioral activity data queue can contain a series of numerical sequences arranged chronologically, where abrupt changes in the values ​​may indicate the start or end of an activity event. Event boundary detection starts from the first data point in the behavioral activity data queue and calculates the magnitude of change between adjacent data points one by one. The magnitude of change is obtained by calculating the absolute difference between the values ​​of adjacent data points. For example, for data points at index i and index i+1 in the behavioral activity data queue, the magnitude of change is the absolute value of the difference between the values ​​of adjacent data points. A dynamic threshold is set for the magnitude of change. When the calculated magnitude of change exceeds the dynamic threshold, the current data point is marked as a potential event boundary point. The dynamic threshold is determined based on the overall activity level of the behavioral activity data queue.

[0023] In some embodiments, the dynamic threshold is calculated through the following steps: calculating the average energy value of the behavioral activity data queue within a complete observation window; multiplying the average energy value by a preset baseline sensitivity coefficient to obtain an initial threshold base; simultaneously calculating the average fluctuation value of the behavioral activity data queue within a recent short time window; and using the average fluctuation value as an adjustment factor to weight and correct the initial threshold base, generating a dynamic threshold for the current detection period. It can be understood that the complete observation window covers data over a relatively long time range in the behavioral activity data queue, such as all data points from the past hour, while the recent short time window covers data from a more recent period, such as data points from the past minute. The formula for calculating the dynamic threshold T is expressed as: in: This represents the average energy value within the entire observation window; This represents the preset basic sensitivity coefficient; This represents the average fluctuation value within a recent short time window; This represents the weighting coefficient of the adjustment factor. In practical implementation, the average energy value... The average fluctuation value is obtained by dividing the sum of the squares of all data point values ​​within the complete observation window by the number of data points. It is obtained by calculating the average value of the changes in adjacent data points within a recent short time window.

[0024] After a potential event boundary point appears, observe a predetermined number of subsequent data points. If the values ​​of these subsequent data points tend to stabilize and fluctuate around a new numerical level, the potential event boundary point is confirmed as a valid event start point. For example, when the change exceeds a dynamic threshold, observe the next five data points. If the values ​​of these data points fluctuate at a new, higher level, then this point is confirmed as a valid event start point. Continuously monitor the behavioral activity data queue. When a change exceeding the dynamic threshold occurs again, confirm the next valid event start point using the same method, and define the time period between the previous valid event start point and the current valid event start point as an independent activity unit. Based on all identified valid event start points, determine the start and end times of all independent activity units. The start and end times correspond to the timestamps of the valid event start points in the behavioral activity data queue.

[0025] When new data enters the behavioral activity data queue, the average energy value within the full observation window and the average fluctuation value within the recent short-term window are recalculated, and the dynamic threshold is updated based on the recalculated average energy value and average fluctuation value. Optionally, the size of the full observation window and the recent short-term window can be configured according to application requirements; for example, the full observation window can be set to 3600 data points, and the recent short-term window can be set to 60 data points. Regarding data comparison, during low-activity periods in the behavioral activity data queue, the fluctuation range is usually small; for example, the movement intensity value fluctuates between 10 and 20, and the calculated fluctuation range may be lower than the dynamic threshold. During high-activity periods in the behavioral activity data queue, the fluctuation range may increase significantly; for example, the movement intensity value may suddenly increase from 15 to 50, and the calculated fluctuation range exceeds the dynamic threshold, thus triggering event boundary detection. Through adaptive adjustment of the dynamic threshold, the system can adapt to different activity levels in the behavioral activity data queue, avoiding false triggering of event boundary detection during low-activity periods or missed triggering during high-activity periods.

[0026] It is understandable that the accuracy of event boundary detection depends on the reasonable setting of the dynamic threshold, which is adjusted based on the overall activity level and recent fluctuations of the behavioral activity data queue. In specific implementation, the preset base sensitivity coefficient C and the weighting coefficient W of the adjustment factor are set experimentally to optimize the event boundary detection performance. Optionally, the data point sampling frequency of the behavioral activity data queue is once per second, the complete observation window covers 3600 data points, the recent short time window covers 60 data points, the base sensitivity coefficient C is set to 0.5, the weighting coefficient W of the adjustment factor is set to 0.2, and the dynamic threshold is dynamically updated according to the real-time data of the behavioral activity data queue to ensure that the detection sensitivity matches the data stream characteristics.

[0027] In one embodiment of the present invention, the internal consistency check is performed as follows: for each activity state comprehensive snapshot data packet, the covariance of the behavioral activity data and physiological indicator data segments in the time dimension is calculated. If the value of the covariance falls within a preset positive correlation range, the internal data of the activity state comprehensive snapshot data packet is determined to be synchronous and consistent, and is marked as consistent. If the value of the covariance falls within a preset negative correlation range or zero correlation range, the internal data of the activity state comprehensive snapshot data packet is determined to be asynchronous or contradictory, and is marked as inconsistent. All activity state comprehensive snapshot data packets marked as consistent are summarized to form a valid activity state comprehensive snapshot data packet set. Consistency check not only ensures data synchronization, but also indirectly reflects the cognitive consistency between behavior and physiological response. For example, if high-intensity activity is accompanied by calm physiological indicators, it may indicate low cognitive load, such as unconscious behavior, requiring the introduction of cognitive care interventions in service recommendations.

[0028] The specific steps for calculating the covariance of behavioral activity data and physiological indicator data segments in the time dimension include: determining a unified covariance calculation time window for the current activity state snapshot data package to be verified; the start point of this time window is the start time point of the independent activity unit, and the end point is the end time point of the independent activity unit. Within the covariance calculation time window, the behavioral activity data and physiological indicator data segments are timestamped and resampled to form a synchronized data sequence with the same number of time points and the same sampling interval. The mean of the behavioral activity data sequence and the mean of the physiological indicator data sequence in the synchronized data sequence are calculated separately. According to the covariance calculation formula, each pair of synchronously sampled behavioral activity data values ​​and physiological indicator data values ​​in the synchronized data sequence are processed, and the average of the products of their deviations from their respective means is calculated to obtain the covariance value in the time dimension.

[0029] In practice, internal consistency verification is performed on each comprehensive snapshot data packet of activity states. This packet consists of behavioral activity data from independent activity units and corresponding physiological indicator data segments. For each snapshot data packet, the covariance of the behavioral activity data and physiological indicator data segments over time is calculated. Covariance calculation aims to quantify the synergy between the changes in the behavioral activity data sequence and the physiological indicator data sequence within the same time window. If the covariance value falls within a preset positive correlation range, the internal data of the snapshot data packet is considered synchronous and consistent, and is marked as consistent. If the covariance value falls within a preset negative correlation range or zero correlation range, the internal data of the snapshot data packet is considered asynchronous or contradictory, and is marked as inconsistent. All snapshot data packets marked as consistent are aggregated to form a valid set of snapshot data packets of activity states.

[0030] In some embodiments, the step of calculating the covariance of behavioral activity data and physiological indicator data segments in the time dimension specifically includes: determining a unified covariance calculation time window for the current activity state comprehensive snapshot data packet to be verified; the starting point of the covariance calculation time window is the start time point of the independent activity unit, and the ending point is the end time point of the independent activity unit. Within the covariance calculation time window, timestamp alignment and resampling processing are performed on the behavioral activity data and physiological indicator data segments to form a synchronized data sequence with the same number of time points and the same sampling interval. It can be understood that timestamp alignment and resampling processing ensure that the two sequences are strictly corresponding in time. For example, the sampling frequency of the behavioral activity data sequence is once per second, while the sampling frequency of the physiological indicator data segment is once every five seconds. The resampling processing can interpolate or downsample the physiological indicator data segment to once per second to match the time points of the behavioral activity data sequence. The mean of the behavioral activity data sequence and the mean of the physiological indicator data sequence in the synchronized data sequence are calculated separately.

[0031] Based on the covariance calculation formula, each pair of synchronously sampled behavioral activity data values ​​and physiological indicator data values ​​in the synchronous data sequence are processed, and the average of the products of their deviations from their respective means is calculated, thus obtaining the covariance value in the time dimension. The formula for calculating covariance (COV) is expressed as: in: This represents the sequence of synchronized behavioral activity data; This represents the synchronized physiological indicator data sequence. and Representing sequences respectively and At the point of time The value; Representative behavioral activity data sequence The mean; Representative physiological indicator data sequence The mean; This represents the total number of data points in the synchronized data sequence. Optionally, the preset positive correlation range can be a threshold greater than zero, such as a covariance greater than... The negative correlation range can be a threshold less than zero, for example, a covariance less than 0. Zero correlation can be an interval where the absolute value is close to zero, for example, the absolute value of the covariance is less than or equal to... .

[0032] In practical implementation, an activity status snapshot data packet may contain data from an independent activity unit lasting ten minutes, such as walking activity from 9:00 AM to 9:10 AM. The corresponding physiological indicator data segment is heart rate data within the same time period. A covariance value is finally obtained through covariance calculation, time window determination, timestamp alignment, resampling, and covariance calculation. For data comparison, a consistent data packet example may show a positive correlation between behavioral activity data (such as movement intensity) and physiological indicator data (such as heart rate), meaning that heart rate increases with activity intensity, and the calculated covariance is positive and greater than the set lower limit of the positive correlation range. A inconsistent data packet example may show high behavioral activity data (such as movement intensity) while physiological indicator data (such as heart rate) is abnormally low or remains unchanged, and the calculated covariance is negative or close to zero, potentially indicating sensor data asynchrony or abnormal physiological state. In some embodiments, the preset thresholds for positive correlation range, negative correlation range, and zero correlation range can be set based on specific application scenarios and sensor characteristics through prior knowledge or experiments. It is understandable that internal consistency verification effectively filters out erroneous data pairings caused by sensor delays, data loss, or accidental interference by calculating the covariance in the time dimension, thus ensuring the credibility of the activity status comprehensive snapshot data package on which subsequent analysis is based in terms of time correlation.

[0033] Optionally, timestamp alignment and resampling can employ linear interpolation or nearest-neighbor interpolation. For cases where there are minor discrepancies between the time base of behavioral activity data sequences and physiological indicator data segments, clock synchronization correction is required in the timestamp alignment step. In practice, all data streams from monitoring devices undergo preliminary time synchronization before being aggregated into a hybrid dynamic monitoring data stream. The timestamp alignment process in the internal consistency check is used to further refine the alignment of data belonging to the same activity unit. After summarizing all consistent activity status snapshot data packets, a valid activity status snapshot data packet set is formed, which will serve as the input to the subsequent adaptive analysis frequency scheduler.

[0034] In one embodiment of the present invention, the internal operation of the adaptive analysis frequency scheduler includes receiving a set of comprehensive snapshot data packets of effective activity status. The number of data packets generated in the comprehensive snapshot data packet set of effective activity status within a unit time is analyzed to obtain a data flow density index; the average degree to which the internal indices of all data packets in the comprehensive snapshot data packet set deviate from a preset benchmark value is analyzed to obtain a data flow anomaly index. The data flow density index and the data flow anomaly index are input into a pre-trained frequency decision network. Based on the implicit relationship between the data flow density index and the data flow anomaly index, the frequency decision network maps and outputs a specific demand scan time interval.

[0035] The training process of the frequency decision network includes collecting a comprehensive snapshot data set of effective activity states from multiple historical periods, along with their corresponding manually set optimal demand scan intervals, to form training sample pairs. For each training sample pair, the density index and anomaly index of its sample data stream are calculated. Using the density index and anomaly index of the sample data streams of all training sample pairs as input features, and the manually set optimal demand scan interval as the target label, the initialized neural network model is subjected to supervised training. By iteratively adjusting the internal connection weights of the neural network model, the model learns the mapping pattern from input features to target labels, ultimately forming the frequency decision network.

[0036] In specific implementation, the adaptive analysis frequency scheduler receives a set of valid activity state comprehensive snapshot data packets from the preceding steps. The adaptive analysis frequency scheduler analyzes the number of data packets generated within the valid activity state comprehensive snapshot data packet set per unit time, which can be set to one minute or one hour. Dividing the number of data packets generated by the unit time length yields the data stream density index. The adaptive analysis frequency scheduler analyzes the average degree of deviation of the internal indices of all data packets in the valid activity state comprehensive snapshot data packet set from a preset baseline value to obtain the data stream anomaly index. In some embodiments, the preset baseline value can be a set of reference values ​​for physiological indicators and behavioral activity data under standard conditions. The degree of deviation can be obtained by calculating the absolute value of the difference between the key indicator value within each data packet and the corresponding baseline value, and then averaging the absolute values ​​of all data packets. The data stream density index and the data stream anomaly index are input into a pre-trained frequency decision network. Based on the implicit relationship between the data stream density index and the data stream anomaly index, the frequency decision network maps and outputs a specific demand scan time interval.

[0037] The training process of the frequency decision network includes collecting a comprehensive snapshot data set of effective activity states from multiple historical periods, along with their corresponding manually set optimal demand scan intervals, to form training sample pairs. For each training sample pair, the density index and anomaly index of its sample data stream are calculated. Using the density index and anomaly index of the sample data streams of all training sample pairs as input features, and the manually set optimal demand scan interval as the target label, the initialized neural network model is subjected to supervised training. By iteratively adjusting the internal connection weights of the neural network model, the model learns the mapping pattern from input features to target labels, ultimately forming the frequency decision network. In specific implementations, the neural network model can adopt a multilayer perceptron structure, the training algorithm can use the backpropagation algorithm, the loss function can use the mean squared error loss function, and the optimizer can use the stochastic gradient descent optimizer. The training process continues until the prediction error of the model on the validation set converges to a preset range.

[0038] It can be understood that the frequency decision network realizes a nonlinear mapping from data flow characteristics to analysis frequency. The mapping relationship can be formally expressed as: Where: symbol Indicates the required scan time interval of the frequency decision network output; symbol A density index representing the input data stream; symbol An anomaly index representing the anomaly rate of the input data stream; symbol This represents the nonlinear mapping function carried by the frequency decision network obtained through training. Regarding data comparison, an example of a scenario with dense data flow and high anomaly is: a large number of comprehensive activity state snapshot data packets are generated per unit time, and the physiological indicators in these packets, such as heart rate, are generally significantly higher than the baseline value, with a large density index D value and a large anomaly index A value. In this case, the frequency decision network may map out a shorter demand scan interval. An example of a scenario with sparse data flow and low anomaly is: a small number of comprehensive activity state snapshot data packets are generated per unit time, and the indicators within these packets are close to the baseline value, with a small density index D value and a small anomaly index A value. In this case, the frequency decision network may map out a longer demand scan interval.

[0039] In some embodiments, the manually set optimal demand scan interval is a suitable analysis frequency labeled by domain experts for different data stream states based on historical experience. Optionally, the input layer of the neural network model has two neurons, corresponding to the density index and the anomaly index, respectively, and the output layer has one neuron, corresponding to the demand scan interval. The number of hidden layers and neurons is determined through cross-validation. It can be understood that the trained frequency decision network encapsulates decision knowledge from data stream state to analysis frequency, enabling the adaptive analysis frequency scheduler to adaptively adjust the rhythm of subsequent periodic traversals and feature extraction based on the dynamic characteristics of the real-time input effective activity state comprehensive snapshot data packet set. When the effective activity state comprehensive snapshot data packet set is generated quickly or its content is abnormal, the demand scan interval is shortened for more intensive analysis; when the effective activity state comprehensive snapshot data packet set is generated slowly and its content is normal, the demand scan interval is extended to reduce the system's computational load.

[0040] In practice, the unit of time is set to one hour, and the density index is calculated as the number of valid activity state snapshot data packets received in the past hour. The anomaly index is calculated for each data packet by first calculating the absolute difference between its average heart rate and the baseline resting heart rate, then calculating the absolute difference between its average behavioral activity intensity and the baseline activity intensity. These two absolute differences are added together as the deviation value for that data packet. Finally, the arithmetic mean of the deviation values ​​of all data packets in the past hour is calculated to obtain the anomaly index. The training samples are derived from data accumulated from the system's operation over the past three months and expert-annotated records. Optionally, the neural network model includes a hidden layer with ten neurons, using the ReLU activation function, and the training iterations are set to 1000. After the frequency decision network is trained, it is deployed in the adaptive analysis frequency scheduler, receiving the set of valid activity state snapshot data packets from the internal consistency verification module in real time, calculating the real-time density index and anomaly index, and outputting the real-time required scan interval. This interval value will be directly used to control the triggering frequency of subsequent periodic traversal and feature extraction modules.

[0041] See Figure 3This chart shows the trend of data flow density and anomaly over time in the elderly care information service system. It illustrates the dynamic changes in data flow density (data packets / hour) and anomaly (deviation value) over 11 hours. The two are highly positively correlated, reflecting the system's operational status at different times. The positive correlation between density and anomaly verifies the design rationality of the adaptive analysis frequency scheduler. When both increase simultaneously, the system automatically shortens the scan interval; when they decrease simultaneously, the interval is extended, achieving efficient resource utilization. The peak values ​​at hours 3 and 8 serve as potential risk warning signals, prompting the system to pay close attention to the health status of elderly people during these periods and prepare medical or emergency service resources in advance. The low-load data during off-peak hours provides a time window for system maintenance, model updates, and other operations during inactive periods, avoiding disruption to normal services.

[0042] In one embodiment of the present invention, the periodic traversal and feature extraction includes uniformly sampling the set of comprehensive snapshot data packets of effective activity states over time, using the demand scanning time interval as a fixed time step. At each sampling point, pattern features of behavioral activity data in all comprehensive snapshot data packets of activity states within the previous time window are extracted, and statistical features of physiological indicator data segments in all comprehensive snapshot data packets of activity states within the same time window are also extracted. The pattern features and statistical features are combined into a multidimensional vector according to a preset vector concatenation rule. As the time step progresses, the multidimensional vectors generated at each sampling point are arranged in chronological order to form a structured service demand feature vector.

[0043] In practical implementation, the periodic traversal and feature extraction module receives the demand scan interval from the adaptive analysis frequency scheduler. The demand scan interval is a value in seconds or minutes. Using the demand scan interval as a fixed time step, the system performs uniform temporal sampling on the set of effective activity state comprehensive snapshot data packets. For example, if the demand scan interval is set to five minutes, the system generates a sampling point every five minutes. At each sampling point, the system extracts the pattern features of behavioral activity data from all activity state comprehensive snapshot data packets within the time window preceding the sampling point. The length of the time window can be set to be equal to the demand scan interval, for example, both being five minutes, or it can be set independently.

[0044] Pattern features of behavioral activity data can include the frequency of occurrence of independent activity units within a time window, the average duration of independent activity units, and the mean energy of the behavioral activity data sequence. Simultaneously, statistical features of physiological indicator data segments are extracted from the comprehensive snapshot data package of all activity states within the same time window. These statistical features can include the mean and variance of heart rate data, the mean of blood oxygen saturation data, and the maximum value of body temperature data within the time window. In some embodiments, the specific items of the pattern features and statistical features are predefined, with each feature corresponding to a numerical calculation result. The pattern features and statistical features are combined into a multidimensional vector according to a preset vector concatenation rule. This rule specifies the order in which the features are arranged in the vector; for example, the first position of the vector stores the frequency of occurrence of independent activity units, the second position stores the mean energy of the behavioral activity data sequence, and the third position stores the mean heart rate data.

[0045] As the time step progresses, the system generates the next sampling point and repeats the feature extraction and vector combination operations. The multi-dimensional vectors generated at each sampling point are arranged in chronological order to form a structured service requirement feature vector. Essentially, the structured service requirement feature vector is a sequence of feature vectors indexed by timestamps. It can be understood that the requirement scanning interval determines the temporal resolution of feature extraction; shorter intervals capture more detailed requirement changes, while longer intervals provide a more macroscopic trend view. In practical implementation, a half-day analysis process with a requirement scanning interval of ten minutes will generate dozens of chronologically arranged feature vectors, which together constitute the structured service requirement feature vector for that period.

[0046] The data comparison is illustrated through a specific scenario example. Referring to Table 1, assuming the required scanning interval is set to five minutes and the time window length is also set to five minutes, the system performs two samplings at 9:00 AM and 9:05 AM respectively. For the 9:00 AM sampling time point, features of the comprehensive snapshot data packet of all activity states within the five-minute time window before 9:00 AM are extracted; for the 9:05 AM sampling time point, features of the comprehensive snapshot data packet of all activity states within the five-minute time window before 9:05 AM are extracted. The activity content and physiological states within these two adjacent time windows may differ, leading to differences in the extracted feature vectors.

[0047] Table 1: Feature vector data at two example sampling time points The process of combining pattern features and statistical features into a multidimensional vector can be represented by a feature combination function. The formula for combining pattern features and statistical features is as follows: Where: symbol This represents the multidimensional feature vector generated at sampling time point t; symbol This represents the pattern feature vector extracted at sampling time point t; symbol This represents the statistical feature vector extracted at sampling time point t; symbol This indicates a predefined vector concatenation operation, which concatenates subvectors. and Connect the beginning and end to form a longer vector. Optionally, vector concatenation can also be performed by stacking vectors along the feature dimension.

[0048] In some embodiments, the length of the time window can be configured independently of the demand scanning interval. For example, the demand scanning interval may be ten minutes, but the time window length for feature extraction may be set to fifteen minutes. This means that the data covered when extracting features at adjacent sampling points overlaps. It can be understood that the periodic traversal and feature extraction process transforms a series of temporally continuous, valid activity state snapshot data packets into a structured numerical feature sequence sampled at equal temporal intervals, i.e., a structured service demand feature vector. This transformation enables the subsequent online service pattern matching engine to perform pattern recognition and matching based on fixed-dimensional, semantically rich feature vectors, rather than directly processing the original heterogeneous, variable-length time-series data packet set. The generation of structured service demand feature vectors is a crucial transformation step in shifting data from unstructured monitoring streams to structured service demand understanding.

[0049] See Figure 4 This is a radar chart showing the similarity of elderly care service items across different time periods, displaying the demand intensity distribution of the five core elderly care service items in the morning, afternoon, and evening. The items with the highest demand intensity in all three time periods (approximately 90-100) reflect the core need of the elderly for continuous health monitoring. The relatively stable demand intensity across the three time periods (approximately 60-70) indicates that social interaction is an important need for the elderly throughout the day. Based on the differences in demand across time periods, the scheduling of service personnel and equipment can be optimized. For example, additional rehabilitation training specialists can be added in the morning, and additional daily living care staff can be added in the evening. Specialized service packages can be developed for high-demand items in different time periods, such as "morning rehabilitation packages" and "evening sleep aid packages." The focus of health monitoring can be adjusted based on the characteristics of demand across time periods. For example, the focus can be on monitoring heart rate and sleep apnea in the evening, and on blood pressure and exercise capacity in the morning.

[0050] In one embodiment of the present invention, the online service pattern matching engine processes the following steps: A predefined knowledge graph of elderly care service projects is established, where each elderly care service project is associated with a set of ideal service demand feature vector patterns. The received structured service demand feature vectors are compared with all ideal feature vector patterns in the elderly care service project knowledge graph for similarity calculation. All ideal feature vector patterns whose similarity calculation results exceed a matching threshold are selected. The elderly care service projects corresponding to the selected ideal feature vector patterns are then identified and sorted according to their similarity from high to low, generating a series of specific elderly care service project suggestion instructions.

[0051] Extracting the hybrid dynamic monitoring data stream is achieved by integrating multiple heterogeneous data sources, including a passive infrared sensor array deployed in the living space, a pressure distribution sensing system installed on the bed, a multi-parameter health monitoring wristband worn by the elderly, and a smart power consumption monitoring unit for the main indoor electrical appliances. The passive infrared sensor array is used to generate behavioral activity data, the multi-parameter health monitoring wristband and pressure distribution sensing system are used to generate physiological indicator data, and the smart power consumption monitoring unit is used to provide auxiliary environmental interaction context information. All data are synchronized in time and then converged into a hybrid dynamic monitoring data stream.

[0052] In its implementation, the online service pattern matching engine establishes a predefined knowledge graph of elderly care service projects. This knowledge graph contains preset service project patterns, including specialized services for cognitive care, such as cognitive training activities or emotional support interventions. When feature vectors show fragmented behavior or abnormal fluctuations in physiological indicators, the pattern matching engine can prioritize recommending such services, achieving precise care at the cognitive level. The elderly care service project knowledge graph is organized in a graph structure, with nodes representing specific elderly care service projects, such as meal delivery, health checks, emergency calls, or accompanying walks. Each elderly care service project node is associated with a set of ideal service demand feature vector patterns. These ideal service demand feature vector patterns are predefined feature vector templates based on historical service records and expert knowledge, used to describe the typical data characteristics that trigger the corresponding elderly care service project. The online service pattern matching engine calculates the similarity between the received structured service demand feature vectors and all ideal service demand feature vector patterns in the elderly care service project knowledge graph. This similarity calculation measures the closeness between the real-time generated structured service demand feature vectors and each predefined pattern.

[0053] In some embodiments, the similarity calculation uses the cosine similarity method, and the formula for calculating cosine similarity is expressed as: Where: symbol This represents the feature vector of real-time structured service requests received by the online service pattern matching engine; (symbol) This represents the k-th ideal service demand feature vector pattern in the knowledge graph of elderly care services; symbol This represents the calculated cosine similarity value, ranging from -1 to 1; [symbol] Represents the dot product operation of vectors; symbol Representing vectors Euclidean norm; symbol Representing vectors The Euclidean norm of the function. The online service pattern matching engine selects all ideal service demand feature vector patterns whose similarity calculation results exceed a preset matching threshold. The preset matching threshold is a constant between 0 and 1, for example, 0.7. The online service pattern matching engine finds elderly care service projects corresponding to the selected ideal service demand feature vector patterns, sorts them in descending order of similarity, and generates a series of specific elderly care service project suggestion instructions. The instructions can include the elderly care service project name and execution priority.

[0054] The data comparison is illustrated through a specific example scenario, assuming a structured service requirement feature vector is generated in real time. The ideal model for the "emergency medical examination" item exists in the knowledge graph of elderly care services, characterized by high activity frequency, high average heart rate, and low blood oxygen saturation. The ideal model for "daily activity monitoring" projects .calculate and The similarity result was 0.85. (Calculation) and The similarity result was 0.60. If the matching threshold is set to 0.70, only the ideal pattern for the "Emergency Medical Examination" item will be found. If selected, the online service pattern matching engine outputs "emergency medical examination" as a suggested instruction for elderly care services. In another scenario, if the structured service demand feature vector shows low activity frequency and normal physiological indicators, its similarity with the "daily activity monitoring" pattern may be 0.75, and its similarity with the "emergency medical examination" pattern may be 0.30, then "daily activity monitoring" will be output as a suggested instruction.

[0055] Extracting the hybrid dynamic monitoring data stream is achieved by integrating multiple heterogeneous data sources. In practice, these heterogeneous data sources include a passive infrared sensor array deployed in the living space, a pressure distribution sensing system installed on the bed, a multi-parameter health monitoring wristband worn by the elderly, and a smart power monitoring unit for the main indoor electrical appliances. The passive infrared sensor array generates behavioral activity data, including the elderly's movement trajectories and activity intensity time series in different areas. The multi-parameter health monitoring wristband and pressure distribution sensing system generate physiological indicator data; the wristband collects signals such as heart rate, blood oxygen saturation, and body temperature, while the pressure distribution sensing system monitors respiratory rate and body movement information while lying in bed. The smart power monitoring unit provides auxiliary environmental interaction context information, such as the start-stop times and energy consumption data of kitchen appliances, televisions, or air conditioners. The raw data collected from all data sources are synchronized with timestamps using a unified time synchronization service and then aggregated into a continuous hybrid dynamic monitoring data stream. The time synchronization service ensures that data events from different sensors have comparable time bases, which is the basis for subsequent data separation and pairing.

[0056] Optionally, the time synchronization service employs a network time protocol or hardware trigger signal to ensure that the time deviation of all data sources is within the millisecond range. In some embodiments, the multi-parameter health monitoring wristband transmits data to a centralized gateway in real time via Bluetooth, the passive infrared sensor array transmits data via a Zigbee network, and the pressure distribution sensing system and smart power monitoring unit access the network via wired or wireless means. All data is transmitted to a central data aggregation server, which performs timestamp alignment and data stream merging operations to form the final hybrid dynamic monitoring data stream. Regarding data comparison, in an example of a hybrid dynamic monitoring data stream without time synchronization, the behavioral activity data sequence shows that the elderly person left the bedroom at 10:00:00, but the timestamp of the event showing an increase in the central rate in the physiological indicator data sequence is 10:00:05. This may lead to deviations in subsequent event boundary detection and data pairing. After strict time synchronization, the timestamps of all data events are based on the same clock source, and the changes in behavioral activities and physiological indicators can be precisely aligned, ensuring the internal time consistency of the activity state comprehensive snapshot data packet.

[0057] It is understandable that extracting the mixed dynamic monitoring data stream is the initial data preparation stage of the overall analysis process. It provides raw, multimodal input for the subsequent separation of behavioral activity data queues and physiological indicator data queues. The online service pattern matching engine relies on the high-quality structured service demand feature vectors generated in the previous steps. By matching predefined knowledge graph patterns, it transforms abstract data features into executable elderly care service project suggestion instructions, thus completing the closed loop from multi-source monitoring data to specific service recommendations.

[0058] See Figure 5 This is a horizontal bar chart showing the similarity between elderly care service projects. Cosine similarity (0-1) quantifies the degree of matching between service demand characteristics and preset service models. The dashed line marks the matching threshold (0.7). Red bars indicate a matching degree that meets the standard, while gray bars indicate a non-matching degree. Projects marked in red should be prioritized for service execution; emergency medical examinations require immediate response, and health follow-ups should be included in the near-term plan. Projects marked in gray can be postponed or have their service frequency adjusted, concentrating resources on high-matching projects to improve service efficiency. This chart serves as a "real-time snapshot" of service demand; combined with changes over time, it can analyze the evolution of elderly people's needs, providing a basis for long-term service planning.

[0059] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for intelligent analysis of cognitive care data in an elderly care information service system, characterized in that, include: Extract a mixed dynamic monitoring data stream of behavioral activities and physiological indicators of the elderly; The hybrid dynamic monitoring data stream is separated according to the source attributes of the data to form a behavioral activity data queue and a physiological indicator data queue with time-series characteristics; Event boundary detection is performed on the behavioral activity data queue to identify the start and end times of each independent activity unit of the elderly person; Based on the start and end times, extract physiological indicator data segments synchronized with each independent activity unit from the physiological indicator data queue; Each independent activity unit is paired with its corresponding physiological indicator data segment to generate multiple comprehensive activity state snapshot data packages; An internal consistency check is performed on the set of comprehensive snapshot data packets of activity status, and valid comprehensive snapshot data packets of activity status that meet the consistency criteria are selected. The effective activity status comprehensive snapshot data packet set is input into a preset adaptive analysis frequency scheduler to obtain a real-time updated demand scan time interval; The effective activity status comprehensive snapshot data packet set is periodically traversed and features are extracted using the demand scanning time interval to generate a structured service demand feature vector; The structured service demand feature vector is delivered to an online service pattern matching engine for processing, and the final output is a series of specific elderly care service project suggestion instructions.

2. The method for intelligent analysis of cognitive care data in an elderly care information service system as described in claim 1, characterized in that, The event boundary detection includes the following steps: Starting from the first data point in the behavioral activity data queue, calculate the magnitude of change between adjacent data points one by one; A dynamic threshold is set for the magnitude of change. When the magnitude of change calculated in a certain instance exceeds the dynamic threshold, the current data point is marked as a potential event boundary point. After the potential event boundary point appears, observe a preset number of subsequent data points. If the values ​​of the subsequent data points tend to stabilize and fluctuate around a new value level, then the potential event boundary point is confirmed as a valid event starting point. The behavioral activity data queue is continuously monitored. When the change exceeds the dynamic threshold again, the next valid event start point is confirmed in the same way, and the time period between the previous valid event start point and the current valid event start point is defined as an independent activity unit. Based on all the identified valid event start points, determine the start and end times of all the independent activity units.

3. The method for intelligent analysis of cognitive care data in an elderly care information service system as described in claim 2, characterized in that, The dynamic threshold is determined based on the overall activity level of the behavioral activity data queue; Calculate the average energy value of the behavioral activity data queue within a full observation window; Multiply the average energy value by a preset basic sensitivity coefficient to obtain the initial threshold base; Simultaneously calculate the average fluctuation value of the behavioral activity data queue within a recent short time window; The average fluctuation value is used as an adjustment factor to weight and correct the initial threshold base, thereby generating a dynamic threshold for the current detection period. When new data enters the behavioral activity data queue, the average energy value within the complete observation window and the average fluctuation value within the recent short time window are recalculated, and the dynamic threshold is updated based on the recalculated average energy value and average fluctuation value.

4. The method for intelligent analysis of cognitive care data in an elderly care information service system as described in claim 1, characterized in that, The execution process of the internal consistency check is as follows: For each activity state comprehensive snapshot data packet, calculate the covariance of behavioral activity data and physiological indicator data segments in the time dimension; If the value of the covariance falls within the preset positive correlation range, it is determined that the internal data of the activity status comprehensive snapshot data packet is synchronized and consistent, and is marked as consistent. If the value of the covariance falls within the preset negative correlation range or zero correlation range, it is determined that the internal data of the activity state comprehensive snapshot data packet is asynchronous or contradictory, and is marked as inconsistent. All activity state composite snapshot data packets marked as having passed consistency are aggregated to form the set of valid activity state composite snapshot data packets.

5. The method for intelligent analysis of cognitive care data in an elderly care information service system as described in claim 1, characterized in that, The internal operations of the adaptive analysis frequency scheduler include: Receive the set of comprehensive snapshot data packets of the effective activity status; The density index of the data stream is obtained by analyzing the number of data packets generated per unit time in the comprehensive snapshot data packet set of the effective activity status; the anomaly index of the data stream is obtained by analyzing the average degree of deviation of the internal index of all data packets in the comprehensive snapshot data packet set of the effective activity status from the preset benchmark value. The density index and the anomaly index of the data stream are input into a pre-trained frequency decision network. The frequency decision network maps and outputs a specific demand scan time interval based on the implicit relationship between the density index and the anomaly index of the data stream.

6. The method for intelligent analysis of cognitive care data in an elderly care information service system as described in claim 5, characterized in that, The training process of the frequency decision network includes: Collect a comprehensive snapshot data set of effective activity status of multiple samples from historical periods and their corresponding manually set optimal required scanning time intervals to form training sample pairs; For each training sample pair, calculate the density index and the anomaly index of its sample data stream. The initial neural network model is trained under supervision using the density index and anomaly index of the sample data stream of all training sample pairs as input features and the manually set optimal demand scanning time interval as the target label. By iteratively adjusting the internal connection weights of the neural network model, the model can learn the mapping pattern from input features to target labels, ultimately forming the frequency decision network.

7. The method for intelligent analysis of cognitive care data in an elderly care information service system as described in claim 1, characterized in that, Periodic traversal and feature extraction include: Using the required scanning time interval as a fixed time step, the effective activity status comprehensive snapshot data packet set is uniformly sampled over time. At each sampling time point, extract the pattern features of behavioral activity data from the comprehensive snapshot data packet of all activity states within the time window preceding the sampling time point; Simultaneously extract the statistical characteristics of physiological indicator data fragments from the comprehensive snapshot data package of all activity states within the same time window; The pattern features and the statistical features are combined into a multidimensional vector according to a preset vector concatenation rule; As the time step progresses, the multidimensional vectors generated at each sampling point are arranged in chronological order to form the structured service requirement feature vector.

8. The method for intelligent analysis of cognitive care data in an elderly care information service system as described in claim 1, characterized in that, The online service pattern matching engine processes the data as follows: Establish a predefined knowledge graph of elderly care service items, where each elderly care service item is associated with a set of ideal service demand feature vector patterns; The similarity between the received structured service demand feature vector and all ideal feature vector patterns in the elderly care service project knowledge graph is calculated. Select all ideal feature vector patterns whose similarity calculation results exceed the matching threshold; Find the elderly care service projects that correspond to the selected ideal feature vector pattern, sort them in descending order of similarity, and generate a series of specific elderly care service project suggestion instructions.

9. The method for intelligent analysis of cognitive care data in an elderly care information service system as described in claim 1, characterized in that, Extracting the hybrid dynamic monitoring data stream is achieved by integrating multiple heterogeneous data sources, including a passive infrared sensor array deployed in the living space, a pressure distribution sensing system installed on the bed, a multi-parameter health monitoring wristband worn by the elderly, and a smart power consumption monitoring unit for the main indoor electrical appliances. The passive infrared sensor array is used to generate behavioral activity data, the multi-parameter health monitoring wristband and the pressure distribution sensing system are used to generate physiological indicator data, and the smart power consumption monitoring unit is used to provide auxiliary environmental interaction context information. All data are aggregated into the hybrid dynamic monitoring data stream after time synchronization.

10. The method for intelligent analysis of cognitive care data in an elderly care information service system as described in claim 4, characterized in that, The step of calculating the covariance of the behavioral activity data and physiological indicator data segments over time specifically includes: A unified covariance calculation time window is determined for the current activity status snapshot data packet to be verified. The starting point of the time window is the start time point of the independent activity unit, and the ending point is the end time point of the independent activity unit. Within the covariance calculation time window, the behavioral activity data and the physiological indicator data segments are time-stamp aligned and resampled to form a synchronized data sequence with the same number of time points and the same sampling interval. Calculate the mean of the behavioral activity data sequence and the mean of the physiological indicator data sequence in the synchronized data sequence, respectively. Based on the covariance calculation formula, each pair of synchronously sampled behavioral activity data values ​​and physiological indicator data values ​​in the synchronous data sequence are processed, and the average value of the product of their deviations from their respective means is calculated, thereby obtaining the covariance value in the time dimension.

Images