An AI-based dynamic risk early warning model construction method for elderly care

Abstract

This invention discloses a method for constructing a dynamic risk warning model based on AI-driven elderly care. The method obtains first data representing environmental activity through temporal convolution of graph nodes, and second data representing transient displacement based on the spatial alignment of radio frequency and inertial sequence features. It calculates fault-tolerant compensation scalars and threshold coefficients using the activity status reflected in the first data, and dynamically modulates the comparison time window and tolerance baseline of the second data. Risk weights are extracted based on the slope of the corrected sequence variance and encapsulated as a parameter set for deployment to a local transmitting device. This invention uses whole-house activity as the core of adaptive modulation, building an external anti-false alarm interception wall for daily activities, and internally triggering computing power redirection and single-mode degradation relay during disconnection, completely solving the problems of false alarms and missed alarms in complex home scenarios, and improving alarm robustness and all-weather monitoring efficiency.

Description

Technical Field

[0001] This invention relates to the field of risk warning technology, specifically to a method for constructing a dynamic risk warning model based on AI-driven elderly care. Background Technology

[0002] With the increasing aging population and the widespread adoption of home-based elderly care, life safety early warning systems based on IoT sensing and artificial intelligence have become crucial infrastructure. Traditional early warning architectures typically focus on capturing abnormal physical changes to trigger alarm commands. Their core challenge lies in minimizing false alarms in complex daily scenarios while reducing the false alarm rate, thereby ensuring the reliability and effectiveness of the underlying alarm signaling devices.

[0003] However, existing security early warning systems face irreconcilable engineering limitations in complex living environments: On the one hand, relying solely on wearable nodes such as smart bracelets for microscopic local posture monitoring (such as inertial fall impact detection) results in a limited sensing scale, easily generating numerous false positives when the person being monitored performs high-frequency household chores with impact characteristics, such as chopping vegetables or washing clothes. Furthermore, they are highly dependent on the individual's compliance with wearing the device and battery life; if the device is forgotten or the battery is depleted, it immediately falls into a monitoring blind spot. On the other hand, while non-intrusive solutions based on pure environmental radio frequency or infrastructure monitoring possess global situational awareness capabilities, they lack a mechanism for rapidly capturing and spatiotemporally verifying sudden physical events (such as millisecond-level fall impacts) of specific individuals, leading to slow response times in transient event identification and a high risk of missed detections.

[0004] Therefore, this invention provides a method for constructing a dynamic risk early warning model based on AI for elderly care. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a method for constructing a dynamic risk early warning model based on AI for elderly care, thereby solving the technical problems described in the background section.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: A method for constructing a dynamic risk early warning model for AI-based elderly care includes the following steps: S1. Obtain the first data of the whole-house equipment collaborative activity index, which represents the environmental correlation state within the target space. The first data is generated by graph node temporal convolution calculation based on the desensitized state jump log of the indoor environmental monitoring equipment. S2. Obtain the second data of feature spatial alignment difference that characterizes the transient displacement event of physical entity. The second data is generated based on the temporal matching similarity calculation between the indoor wireless radio frequency channel state sequence and the local spatial displacement vector sequence. S3. Based on the current environmental activity reflected in the first data, calculate the fault tolerance compensation scalar and threshold relaxation coefficient used to construct the dynamic judgment topology graph; use the fault tolerance compensation scalar and threshold relaxation coefficient to dynamically modulate the signal comparison time window width and matching tolerance baseline required for generating the second data, thereby obtaining the corrected alignment difference sequence containing context constraint logic. S4. Extract the risk weight factor that characterizes the true credibility of the event based on the slope of the variance change of the corrected alignment difference sequence. S5. Encapsulate the risk weighting factor, the comparison time window width, and the matching tolerance baseline to construct and generate an adaptive dynamic risk warning model parameter set; deploy the adaptive dynamic risk warning model parameter set to the local signal command transmitting device to update the local trigger rule library in the transmitting device used for intercepting false alarm signals and activating downgrade and substitute monitoring.

[0007] Further, the steps for acquiring the first data are as follows: extract the switching timestamps generated by the indoor environmental monitoring equipment and perform de-identification processing to generate de-identified status transition logs; construct a data association graph with different environmental monitoring equipment as nodes and the geographical topological location of the equipment as edges; perform temporal convolution calculation on the graph node states to output numerical features representing the frequency of use and linkage of each device; calculate the weighted average of the numerical features to obtain the whole-house device collaborative activity index.

[0008] Furthermore, the specific steps for calculating the numerical features representing the frequency of linkage in temporal convolution are as follows: Read the preset time sliding window width and spatial adjacency matrix from the local device file system; within the current time sliding window, extract the state transition frequency of each environmental monitoring node to construct the node feature vector at the current moment; perform matrix multiplication on the spatial adjacency matrix and the node feature vector to obtain the first intermediate aggregation vector, which represents the device state diffusion effect in the spatial dimension; extract the historical feature vector corresponding to the previous sliding window, and multiply it with the historical feature vector using a preset time decay coefficient to obtain the second intermediate time vector; perform addition on the first intermediate aggregation vector and the second intermediate time vector, and input the result of the addition operation into a nonlinear activation function. The output result is the numerical feature representing the frequency of linkage of each device.

[0009] Furthermore, the steps for acquiring the second data are as follows: extract the radio frequency channel state sequence from the underlying protocol of the indoor communication network; extract the extremely low-frequency local spatial displacement vector sequence reported by the wearable computing device; using canonical correlation analysis algorithm, extract the fluctuation component representing the displacement of large-volume objects in the radio frequency channel state sequence, and the component representing the fall of local limbs in the local spatial displacement vector sequence, thereby mapping the two sets of sequences to a feature alignment space of the same dimension; by calculating the dynamic time warping value of the time curves of the two sets of sequences in the feature alignment space, determine whether the spatial disturbance detected by radio frequency and the limb movement detected by inertial detection belong to the same transient physical event, and thus generate the second data representing the cross-confirmation result.

[0010] Furthermore, the calculation steps for constructing the fault-tolerant compensation scalar and threshold relaxation coefficient of the dynamic judgment topology are as follows: When the acquired first data is greater than the preset environmental active state benchmark value, it is confirmed that there is continuous housework interference in the current target space; based on the difference between the first data and the environmental active state benchmark value, the fault-tolerant compensation scalar used to delay the comparison judgment is calculated in the forward direction, and the threshold relaxation coefficient used to expand the difference tolerance interval is calculated simultaneously; using the fault-tolerant compensation scalar and threshold relaxation coefficient, the signal comparison time window width and matching tolerance benchmark required to generate the second data are dynamically modulated.

[0011] Furthermore, the dynamic modulation process is as follows: the width of the reference comparison time window is added to the fault tolerance compensation scalar to obtain the extended signal comparison time window width, and the reference matching tolerance threshold is multiplied by the threshold relaxation coefficient to obtain the amplified matching tolerance baseline.

[0012] Furthermore, the steps for constructing the risk weight factor are as follows: calculate the data discrete variance of the corrected alignment difference sequence within the preset sliding historical time interval; calculate the slope of the change of the data discrete variance at the current sampling time using the first-order difference function; when the slope of change is greater than the warning surge threshold, generate the risk weight factor with the highest priority value; when the slope of change is not greater than the warning surge threshold, but the absolute value of the discrete variance is greater than the safety baseline, generate the risk weight factor with the second-level value.

[0013] Further, the following steps are used to extract and calculate the discrete variance of the corrected alignment difference sequence: Extract all corrected alignment difference sequence data within a preset sliding observation window width, tracing back from the current sampling time, and count the total number of data points within the current sliding observation window width; Accumulate all data points within the sliding observation window width, and divide the sum by the total number of data points to obtain the local mean within the time interval; Traverse each data point within the sliding observation window width, calculate the difference between each data point and the local mean, and square each difference to generate the corresponding set of squared residuals; Summate all elements in the set of squared residuals to obtain the residual sum of squares; To obtain unbiased sample statistical characteristics, subtract a constant 1 from the total number of data points to obtain the correction divisor, and divide the residual sum of squares by the correction divisor; the resulting quotient is the discrete variance of the data at the current time.

[0014] Furthermore, the local trigger rule base is constructed as follows: the risk weight factor, the comparison time window width, and the matching tolerance baseline are written into the judgment logic layer of the false alarm signal interception as new signal release standards; resource redirection routing rules are written into the execution rules of the downgraded backup monitoring.

[0015] Furthermore, the resource redirection routing rules are configured as follows: monitor the underlying hardware data bus, and when no local spatial displacement vector sequence is received for a continuously preset silence period, and the first data is lower than the preset silence stagnation benchmark, generate a link interruption flag; based on the link interruption flag, send a forced sleep command to the process managing the analysis of radio frequency signals in non-associated areas, interrupting the corresponding polling operation; read the dimensional weight allocation vector used for multimodal fusion determination in the current working memory; perform an asymmetric parameter overlay operation on the dimensional weight allocation vector, assigning the inertial weight factor a value of 0, and simultaneously assigning the radio frequency weight factor a normalized maximum upper limit value of 1; through the absolute value redistribution logic, forcibly tilt and reorganize the model risk determination weights onto the fluctuation characteristics of a single radio frequency channel state sequence within the target stagnation area, thereby triggering the single-mode micro multipath variance extraction algorithm to determine the survival status of personnel.

[0016] (III) Beneficial Effects This invention provides a method for constructing a dynamic risk early warning model for AI-based elderly care, which has the following beneficial effects: This invention obtains the collaborative activity index of all devices in the house through temporal convolution of graph nodes, using it as a priori constraint of the macro environment. When high-frequency housework activities are detected, the fault tolerance compensation scalar and threshold relaxation coefficient are extracted in a forward calculation, and the micro signal comparison time window and matching tolerance baseline are dynamically relaxed. A false alarm interception wall with adaptive adjustment capability is constructed, which fundamentally eliminates the persistent problem of elderly people being misjudged as having high-risk falls due to micro-impact caused by daily housework, and greatly improves the accuracy of alarm signaling and the robustness against environmental noise disturbances. When the underlying data bus is detected to be silent and the environment is in a stagnant state, this invention generates a link interruption identifier and forcibly tilts and reorganizes the model risk judgment weights to the fluctuation characteristics of a single wireless radio frequency channel state sequence within the target stagnant area; under extreme working conditions with limited perception, it accurately allocates the computing power of the edge gateway, activates the single-mode micro multipath variance extraction algorithm to carry out vital sign monitoring relay, and realizes the dimensionality reduction of the security situation. This invention extracts the fluctuation components of the wireless radio frequency channel state sequence and the local limb fall components of the wearable device, maps them to a feature alignment space of a unified dimension using canonical correlation analysis algorithm, and calculates the cost of the time curve through dynamic time warping. The cross-verification mechanism effectively overcomes the limitation of slow response of pure environmental radio frequency schemes, and realizes the rapid capture and spatiotemporal verification of transient physical events in the multimodal physical field, reducing the false negative rate of high-risk events. This invention modularly encapsulates the extracted risk weighting factors, dynamic comparison time window width, and matching tolerance baseline to generate an adaptive dynamic risk warning model parameter set, which is then directly deployed to the local signal command transmission device. The core anti-false alarm interception and degradation trigger rule base is placed at the network edge node to avoid bandwidth congestion caused by uploading massive amounts of raw sensor data to the cloud. While ensuring ultimate privacy protection and low response latency, it achieves lightweight, decentralized closed-loop risk management. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall process flow of the present invention; Figure 2 This is a schematic diagram illustrating the specific steps of calculating and outputting numerical features representing the frequency of linkage in the temporal convolution of the present invention. Figure 3 This is a schematic diagram illustrating the calculation steps of the fault-tolerant compensation scalar and threshold relaxation coefficient for constructing the dynamic decision topology graph according to the present invention. Detailed Implementation

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

[0019] Please see Figure 1-3 This invention provides a method for constructing a dynamic risk early warning model for AI-based elderly care, comprising the following steps: S1. Obtain the first data of the whole-house equipment collaborative activity index, which represents the environmental correlation state within the target space. The first data is generated by graph node temporal convolution calculation based on the desensitized state jump log of the indoor environmental monitoring equipment. S2. Obtain the second data of feature spatial alignment difference that characterizes the transient displacement event of physical entity. The second data is generated based on the temporal matching similarity calculation between the indoor wireless radio frequency channel state sequence and the local spatial displacement vector sequence. S3. Based on the current environmental activity reflected in the first data, calculate the fault tolerance compensation scalar and threshold relaxation coefficient used to construct the dynamic judgment topology graph; use the fault tolerance compensation scalar and threshold relaxation coefficient to dynamically modulate the signal comparison time window width and matching tolerance baseline required for generating the second data, thereby obtaining the corrected alignment difference sequence containing context constraint logic. S4. Extract the risk weight factor that characterizes the true credibility of the event based on the slope of the variance change of the corrected alignment difference sequence. S5. Encapsulate the risk weighting factor, the comparison time window width, and the matching tolerance baseline to construct and generate an adaptive dynamic risk warning model parameter set; deploy the adaptive dynamic risk warning model parameter set to the local signal command transmitting device to update the local trigger rule library in the transmitting device used for intercepting false alarm signals and activating downgrade and substitute monitoring.

[0020] The steps for acquiring the first data are as follows: extract the switch status switching timestamps generated by the indoor environmental monitoring equipment and perform de-identification processing to generate de-identified status transition logs; construct a data association graph with different environmental monitoring equipment as nodes and the geographical topological location of the equipment as edges; perform temporal convolution calculation on the graph node status to output numerical features representing the frequency of use and linkage of each device; calculate the weighted average of the numerical features to obtain the whole-house device collaborative activity index.

[0021] The steps for constructing the data association graph are as follows: Extract the original operation logs of the indoor environmental monitoring equipment, and obtain the media access control address and switch status switching timestamp from the original operation logs; call the preset one-way hash algorithm (SHA-256 in this embodiment) to map the media access control address to a fixed-length anonymous device identifier, thereby erasing the time-residual features that may be associated with personal privacy, and outputting a desensitized status transition log containing only the anonymous device identifier and timestamp; read the absolute coordinates of each device in the three-dimensional coordinate system from the external space configuration file; calculate the Euclidean distance between the absolute coordinates of any two anonymous device identifiers to obtain the physical straight-line distance between the two nodes; perform the reciprocal operation on the physical straight-line distance, and add a very small smoothing constant to the reciprocal result to prevent the denominator from being zero, and use the final calculation result as the topological edge weight between the two device nodes, thereby constructing a data association graph containing node identity and edge weight values.

[0022] The specific steps for calculating and outputting numerical features representing the frequency of linkage in temporal convolution are as follows: Read the preset time sliding window width (configured to 60 seconds in this embodiment) and spatial adjacency matrix (generated by mapping the inverse of the physical absolute distance between each environmental monitoring device) from the local device file system; within the current time sliding window, extract the state transition frequency of each environmental monitoring node to construct the node feature vector at the current moment; perform matrix multiplication on the spatial adjacency matrix and the node feature vector to obtain the first intermediate aggregation vector, which represents the device state diffusion effect in the spatial dimension; extract the historical feature vector corresponding to the previous sliding window, and multiply it with the historical feature vector using a preset time decay coefficient (set to 0.8 in this embodiment) to obtain the second intermediate time vector; add the first intermediate aggregation vector and the second intermediate time vector, and input the result of the addition operation into a nonlinear activation function (using the Sigmoid function in this embodiment), the output of which is the numerical feature representing the frequency of linkage between each device.

[0023] In a preferred embodiment, the data processing after obtaining the first intermediate aggregation vector and the second intermediate time vector, specifically the spatial feature mapping weight matrix and the temporal feature mapping weight matrix, is a preferred implementation of the preprocessing before adding the two: Since the first intermediate aggregation vector in the spatial dimension of the graph nodes and the second intermediate time vector in the historical time dimension belong to different data structure dimension topological subspaces, to eliminate the mathematical dimension collapse caused by heterogeneous data merging, a feature transformation mapping relationship generated in the offline pre-configuration stage is introduced before performing the addition operation; specifically, the computation flow is executed based on the following linear mapping and aggregation mechanism: ;in, Numerical feature vectors representing the frequency of interaction of the fused devices are used in the state space. This represents the first intermediate aggregation vector of the current acquisition cycle; This represents the second intermediate time vector of the previous observation sliding window; This represents the spatial feature mapping weight matrix for graph nodes; This represents the weight matrix for mapping the time features of a time series. Represents the static bias compensation vector; This represents a nonlinear activation function; for the aforementioned spatial feature mapping weight matrix... Time feature mapping weight matrix and bias compensation vector This parameter array is based on a 30-day historical linkage status change log collected from the on-site survey equipment network (in this embodiment, the on-site includes a monitoring gateway group covering the floor restrooms and corridors) as the original offline training sample. Supervised training is performed with the global optimization objective of minimizing the cross-entropy error of the predicted device start-stop states. After meeting the convergence condition that the error loss rate during continuous verification cycles is less than 0.5%, the extracted and solidified deterministic engineering basic configuration carrier is frozen and stored in the non-volatile memory of edge computing. In this embodiment, to ensure the legality of the addition operator, the spatial feature mapping weight matrix... Mapping weight matrix with time features It must be strictly configured to have the same number of output rows; it should be noted that the embodiments of this application are not limited to the specific absolute dimension numerical parameters of the feature mapping space, and the upper limit of the multiply-accumulate operation computing power threshold of the visible edge chip is adaptively scaled and adjusted; by introducing the dimensionality reduction / dimensionality increase weight matrix extracted by offline training to perform explicit projection transformation on heterogeneous vectors, the risk of program abnormality and interruption caused by the mixing of objective physical meaning or misalignment of feature dimensions when spatiotemporal multidimensional feature flows are directly superimposed is avoided, thereby achieving robust convergence of heterogeneous spatiotemporal graph neural network parameters to unified fusion flow nodes.

[0024] The system continuously collects operational logs from devices such as smart sockets and water meters, removing time-dependent features that may be associated with personal privacy, and mapping discrete device start / stop states to a relational graph. In this embodiment, by calculating the temporal convolution of the impedance change of the associated lighting circuit after the bathroom water valve is opened, a whole-house device collaborative activity index that purely reflects the frequency of fluctuations in the physical state of the environment is generated in the working memory, serving as a priori benchmark for subsequent models. By abandoning visual or precise physiological sign tracking, the complex state changes of the indoor smart home network are abstracted into a quantifiable activity index. The activity index effectively decouples from the macroscopic situational dependence of wearable devices, providing objective and comprehensive environmental background reference data for subsequent false alarm arbitration.

[0025] The steps for obtaining the second data are as follows: extract the wireless radio frequency channel state sequence from the underlying protocol of the indoor communication network; extract the extremely low frequency local spatial displacement vector sequence reported by the wearable computing device; using canonical correlation analysis algorithm, extract the fluctuation component representing the displacement of large-volume objects in the wireless radio frequency channel state sequence, and the component representing the fall of local limbs in the local spatial displacement vector sequence, thereby mapping the two sets of sequences to a feature alignment space of the same dimension; by calculating the dynamic time warping value of the time curves of the two sets of sequences in the feature alignment space, determine whether the spatial disturbance detected by radio frequency and the limb movement detected by inertial detection belong to the same transient physical event, and then generate the second data (feature space alignment difference degree) representing the cross-confirmation result.

[0026] The steps for obtaining the dynamic time-warped cost value of the two sets of time curves are as follows: Obtain the mapped fluctuation component as the first sequence, and obtain the component of local limb descent as the second sequence; based on the time step lengths of the first and second sequences, initialize and construct a two-dimensional cumulative cost grid in memory; perform absolute value subtraction on each corresponding time point of the first and second sequences, and fill the subtraction result into the two-dimensional cumulative cost grid as the base distance; traverse the cost grid from the top left corner to the bottom right corner, and for the currently traversed grid node, extract the three historical cumulative cost values ​​of the adjacent nodes on the left, above, and diagonally opposite left of the sequentially traversed cost grid; perform a minimum value filtering operation on the above three historical cumulative cost values; add the filtered minimum value to the base distance of the current grid node, and use the addition result to update the value of the current grid node; after the entire two-dimensional cumulative cost grid has been traversed, extract the final value located at the absolute bottom right corner of the grid, and output the final value as the second data representing the cross-validation result (i.e., feature space alignment difference).

[0027] The specific steps of the canonical correlation analysis algorithm, which maps to a feature alignment space of a unified dimension, are as follows: First, obtain the radio frequency channel state sequence within the current decision period and define it as the first observation matrix. Simultaneously, obtain the local spatial displacement vector sequence and define it as the second observation matrix. The number of rows in the first and second observation matrices needs to be aligned to the same number of time-domain sampling points through resampling. Calculate the autocovariance matrix of the first and second observation matrices, as well as the cross-covariance matrix between them. Construct the generalized eigenvalue equation for canonical correlation analysis using these three covariance matrices, and perform eigenvalue decomposition on the generalized eigenvalue equation. Extract the first and second eigenvectors corresponding to the largest eigenvalues. Perform matrix multiplication on the first observation matrix and the first eigenvector to generate the mapped radio frequency fluctuation feature component. Perform matrix multiplication on the second observation matrix and the second eigenvector to generate the mapped inertial fall feature component. The output results constitute two sets of time curves within the feature alignment space of a unified dimension.

[0028] Furthermore, since the first observation matrix characterizes the amplitude and phase response of the microwave signal in the radio frequency channel, while the second observation matrix characterizes the inertial acceleration parameter of a human body's limbs falling, there is a significant numerical gap between the two during the initial field condition acquisition phase. To eliminate the dimensional offsetting defect, at this state transition level, independent time-domain feature standardization preprocessing is performed on the first and second observation matrices respectively: the time series of each physical antenna subcarrier channel in the first observation matrix is ​​extracted, and the local expected mean and fluctuation standard deviation amplitude of all discrete sampling points of a single channel within this time sliding period are calculated; each time-series sampling value point within the physical antenna subcarrier channel is traversed, the extracted local expected mean is subtracted from the time-series sampling value point, and the resulting first absolute difference component is divided by the fluctuation standard deviation amplitude to complete the standardized numerical reconstruction of the radio frequency mode within the current time slice; synchronously and in parallel, Using the same logical framework, for each three-dimensional coordinate axis displacement sequence in the second observation matrix, the mean and standard deviation of acceleration are independently extracted. Similarly, a step-by-step numerical replacement operation of subtracting the mean and dividing by the standard deviation is performed to complete the standardization and recombination of the inertial modes. The new dimensionless pure numerical array generated by the standardization and recombination is then used as the reference input source and fed into the subsequent autocovariance matrix operation flow step. By forcibly introducing a channel-by-channel standardization dimensionless processing method based on the statistical distribution characteristics of the same source at the multi-source data coupling entry point, the asymmetry of the underlying service dimensions caused by the direct multiplication and addition operation between radio frequency electromagnetic parameters and mechanical motion parameters is completely eliminated. This avoids the numerical swallowing effect on the weak inertial fall feature vector in the covariance analysis matrix due to the excessive absolute amplitude of the microwave signal, thereby ensuring that the constructed typical correlation feature extraction architecture can represent the synchronous coupling correlation of the two heterogeneous signals in a balanced and realistic manner.

[0029] The system accesses the physical layer interface of the underlying router to capture wireless radio frequency channel state information and receives the local inertial acceleration components of the wearable device. It performs absolute same-source, same-dimensional alignment physical common sense verification: instead of direct algebraic addition and subtraction, it uses canonical correlation analysis to eliminate dimensional differences, extracting the wave component representing the displacement of large-volume objects in the radio frequency data and the component representing the local limb fall in the inertial data. Then, it uses a dynamic time warping algorithm to calculate the morphological value of both in the time domain, thereby determining whether the spatial disturbance detected by the radio frequency and the limb movement detected by the inertial data belong to the same transient physical event. By extracting the feature spatial alignment difference between the radio frequency and inertial data, a cross-verification mechanism is established on the first line of underlying physical defense. The flow of multimodal heterogeneous features effectively eliminates device-level false positive alarms caused by a single wristband being accidentally tapped or dropped, improving the confidence in capturing real sudden movements.

[0030] The calculation steps for constructing the fault-tolerant compensation scalar and threshold relaxation coefficient of the dynamic judgment topology are as follows: When the acquired first data is greater than the preset environmental active state benchmark value, it is confirmed that there is continuous housework interference in the current target space; based on the difference between the first data and the environmental active state benchmark value, the fault-tolerant compensation scalar used to delay the comparison judgment is calculated in the forward direction, and the threshold relaxation coefficient used to expand the difference tolerance interval is calculated simultaneously; using the fault-tolerant compensation scalar and threshold relaxation coefficient, the signal comparison time window width and matching tolerance benchmark required to generate the second data are dynamically modulated.

[0031] The dynamic modulation process is as follows: the width of the reference comparison time window is added to the fault tolerance compensation scalar to obtain the extended signal comparison time window width, and the reference matching tolerance threshold is multiplied by the threshold relaxation coefficient to obtain the amplified matching tolerance baseline.

[0032] The steps for synchronously calculating the threshold relaxation coefficient are as follows: Extract the nonlinear growth rate configuration term, the relaxation limit approximation constant, and the preset basic offset constant from the pre-stored local external data carrier. In this embodiment, the basic offset constant is used to prevent mathematical collapse due to division by zero or the amplification factor exceeding the limit when the exponential decay term tends to zero. The value of the basic offset constant is strictly set to 1. Obtain the difference between the first data and the environmental active state benchmark value. Multiply the difference magnitude with the nonlinear growth rate configuration term and take the negative value. Perform a power operation with the base of the natural logarithm as the base and the negative value as the exponent to obtain the environmental inertia decay term. Add the environmental inertia decay term to the basic offset constant to construct the basic denominator. Use the relaxation limit approximation constant as the numerator and divide by the basic denominator to calculate the threshold relaxation coefficient controlled by the safe saturation boundary.

[0033] The steps for determining the environmental activity baseline value are as follows: Obtain desensitized state transition logs for a continuous preset period (in this embodiment, the continuous preset period is the past 30 natural days), and divide the time window according to 24 hours per day; use the K-Means unsupervised clustering algorithm to classify the data density within each time window, extract the data cluster with the lowest density centroid, and mark it as the basal silent sequence; calculate the first data (whole-house device collaborative activity index) at each sampling time within the basal silent sequence; sort all the obtained first data in ascending order according to their numerical values ​​to obtain an ordered activity set; perform 95th percentile extraction on the ordered activity set, and define the extracted specific value as the environmental activity baseline value specific to this space; the role of the environmental activity baseline value is to define whether there is continuous human activity in the environment; the consideration of setting the 95th percentile is to achieve a technical balance between filtering out occasional device self-starting noise and keenly capturing the weak features of the early stage of human activity.

[0034] The steps for determining the nonlinear growth rate configuration item are as follows: In a controlled laboratory environment, test subjects with different physical characteristics are recruited to perform a preset set of standard actions; the set of standard actions is forced to include 500 individual high-frequency housework actions and 100 dummy fall events that are suddenly interspersed during the housework actions; the first data and the second data when the action occurs are recorded throughout the process; a preset range of test parameter sequences is obtained (in this implementation, the range of test parameter sequences is from 0.01 to 0.1, with a step size of 0.01); for each discrete value in the test parameter sequence, the following verification operations are performed iteratively: the current test value is substituted into the nonlinear growth rate configuration item of S3, the number of times the high-priority risk weight factor is output under the test value is counted, and the corresponding number of false positives and false negatives is recorded; the statistical result sequence corresponding to different test values ​​is obtained; a subset of candidate test values ​​that satisfies the condition that the number of false negatives is strictly equal to 0 is selected; a specific value with the lowest number of false positives is extracted from the subset of candidate test values; the specific value is established and solidified as the nonlinear growth rate configuration item.

[0035] The steps for determining the relaxed limit approximation constant are as follows: Extract the pre-stored real physical fall standard feature set (in this embodiment, it is the data collected by an inertial sensing device with a sampling rate of not less than 100Hz and a range covering ±8g); extract the maximum interference feature set of extreme high-frequency housework activities captured under the same sensor network (in this embodiment, quickly chopping bones and spinning clothes); perform energy integration calculations on the extreme value data in the above two feature sets respectively; obtain the fall feature energy integration result and the maximum interference feature energy integration result, and perform division operation with the former as the dividend and the latter as the divisor to obtain the physical limit ratio; multiply the physical limit ratio by the preset sensor tolerance deviation coefficient (in this embodiment, the sensor tolerance deviation coefficient is set to 0.9) to obtain the final relaxed limit approximation constant (in this embodiment, the final derivation is 1.8).

[0036] The forward calculation extraction path of the fault tolerance compensation scalar is as follows: obtain the difference magnitude of the environmental activity state benchmark value; obtain the preset maximum tolerance upper limit time (in this embodiment, the maximum tolerance upper limit time is set to 2000 milliseconds); obtain the preset conversion ratio constant (in this embodiment, the conversion ratio constant represents the number of milliseconds of time extension corresponding to each unit of activity difference); multiply the difference magnitude of the environmental activity state benchmark value and the conversion ratio constant to obtain the unconstrained initial compensation time; perform a minimum value operation by comparing the unconstrained initial compensation time and the maximum tolerance upper limit time; output the final value filtered by the minimum value operation and establish it as the fault tolerance compensation scalar.

[0037] In practical engineering, the preset environmental activity baseline value is stored in a spreadsheet file. When the first data (the activity index of the whole house device collaboration) surges, i.e., it is in a continuous high-energy-consuming state such as washing vegetables or laundry, interference is judged. Based on the magnitude of the activity deviation from the baseline, a time-dimensioned fault-tolerant compensation scalar (+1500 milliseconds in this embodiment) and a dimensionless threshold relaxation coefficient (1.8 in this embodiment) are calculated. The original 2000 millisecond time window for dynamic time warping comparison is extended to 3500 milliseconds, and the fault-tolerant baseline for judging the mismatch between the two sets of features is increased by 1.8 times. Through the direct correction of physical parameters, a corrected alignment difference sequence is generated to suppress short-term violent fluctuations caused by normal high-frequency limb movements. The macroscopic intelligent device status is directly penetrated and the core parameters of microscopic motion capture are reshaped. The rigid single threshold setting of the traditional early warning model is broken, and the model is given a contextual causal judgment ability similar to human common sense.

[0038] The steps for constructing the risk weight factor are as follows: Calculate the data discrete variance of the corrected alignment difference sequence within the preset sliding historical time interval; calculate the slope of the change of the data discrete variance at the current sampling time using the first-order difference function; when the slope of change is greater than the warning surge threshold, generate the risk weight factor with the highest priority value; when the slope of change is not greater than the warning surge threshold, but the absolute value of the discrete variance is greater than the safety baseline, generate the risk weight factor with the second-level value.

[0039] The following steps are used to extract the discrete variance of the data in the corrected alignment difference sequence: Extract all corrected alignment difference sequence data within a preset sliding observation window width (in this embodiment, the preset sliding observation window width is set to include 100 consecutive sampling points) traversing backward from the current sampling time, and count the total number of data points within the current sliding observation window width; perform a cumulative operation on all data points within the sliding observation window width, and divide the cumulative sum by the total number of data points to obtain the local mean within the time interval; traverse each data point within the sliding observation window width, calculate the difference between each data point and the local mean, and square each difference to generate the corresponding set of squared residuals; sum all elements in the set of squared residuals to obtain the residual sum of squares; to obtain unbiased sample statistical characteristics, subtract a constant 1 from the total number of data points to obtain the correction divisor, and divide the residual sum of squares by the correction divisor; the resulting quotient is the discrete variance of the data at the current time.

[0040] The calculation steps for the slope change are as follows: First, read the first-order difference step size configuration parameter from the external structured carrier (in this embodiment, the first-order difference step size configuration parameter is preset to correspond to a sampling point interval of 500 milliseconds, in order to filter out the interference of high-frequency sampling noise on the slope calculation); obtain the data discrete variance of the corrected alignment difference sequence corresponding to the current sampling moment, and backtrack to obtain the data discrete variance of the historical moments corresponding to the first-order difference step size configuration parameter; subtract the variance of the current moment from the variance of the historical moments to obtain the variance increment; divide the variance increment by the actual time scale corresponding to the first-order difference step size, and output the result of the division operation as the slope change at the current sampling moment to identify the deterioration trend of the current physical space disturbance.

[0041] The steps for determining the warning surge threshold are as follows: Obtain a positive sample set of the slope of the variance change of the corrected alignment difference sequence of confirmed real fall events from the historical verification library, and obtain a negative sample set of the same slope for daily harmless activities; perform probability density distribution kernel estimation on the two sample sets respectively; calculate the coordinates of the minimum point in the overlapping interval of the two distribution curves; extract the slope value corresponding to the coordinates of the minimum point and set it as the warning surge threshold.

[0042] The deterministic calculation process for the safety baseline is as follows: obtain the variance values ​​of all corrected alignment difference sequences output under the high-frequency housework action conditions in the standard action set; extract the maximum envelope edge value of the corrected alignment difference sequence variance value (in this embodiment, the maximum extreme point in the distribution), multiply the maximum envelope edge value by a safety margin expansion factor of 1.1 to 1.3, and the calculated value is set as the safety baseline.

[0043] The continuous dissimilarity sequence is abstracted into a weight configuration that can be called downstream. The input corrected alignment dissimilarity sequence is not compared by absolute value, but the discrete variance within the sliding historical time interval is extracted. By calculating the first difference of the discrete variance, the deterioration trend of the current physical space disturbance is identified. The sharp rise in slope means that the abnormal matching degree between space and local area is rapidly solidifying into real fall events. The program generates risk weight factors of different discrete magnitudes accordingly. By using the slope of the variance change rather than the absolute value, the oscillations caused by edge samples and sampling noise are effectively suppressed, ensuring the smoothness and high reliability of the output of the constructed early warning model.

[0044] The local trigger rule base is constructed as follows: the risk weight factor, the comparison time window width, and the matching tolerance baseline are written into the judgment logic layer of the anti-false alarm signal interception as a new signal release standard; resource redirection routing rules are written into the execution rules of the downgraded backup monitoring.

[0045] The resource redirection routing rule is configured as follows: monitor the underlying hardware data bus, and when no local spatial displacement vector sequence is received for a continuously preset silent period, and the first data is lower than the preset silent stagnation benchmark, generate a link interruption flag; according to the link interruption flag, send a forced sleep command to the process managing the wireless radio frequency signal parsing of non-associated areas (in this embodiment, the rooms other than the toilet) to interrupt the corresponding polling operation; read the dimensional weight allocation vector used for multimodal fusion determination in the current working memory (in this embodiment, the dimensional weight allocation vector includes radio frequency weight factor and inertial weight factor); perform an asymmetric parameter overlay operation on the one-dimensional weight allocation vector: specifically, assign the inertial weight factor to the value 0, and assign the radio frequency weight factor to the normalized maximum upper limit value of 1; through the above absolute value redistribution logic, the model risk determination weight is essentially forcibly tilted and recombined to the single wireless radio frequency channel state sequence fluctuation characteristics in the target stagnation area, thereby triggering the single-mode micro multipath variance extraction algorithm to determine the survival status of personnel.

[0046] The steps for establishing a silent stagnation baseline are as follows: obtain the historical white noise level of the whole-house device collaborative activity index (in this embodiment, the average index of 72 consecutive hours when no one is in the house); add three times the standard deviation to the historical white noise level; and output the calculation result as the silent stagnation baseline.

[0047] For the security state degradation logic of triggering single-node modal survival authentication in S5, the extreme load drop protection and triggering single-mode security degradation bias compensation condition flow are the core fault-tolerant supplementary implementation methods under extreme game conditions of device link interruption; specifically, when the judgment model confidence weight pool is 100% redirected to the pure wireless radio frequency parameter domain by the main control logic, the original multi-mode cross-fusion adjudication benchmark immediately becomes invalid; if the existing joint judgment threshold is rigidly reused for unilateral micro multipath feature comparison, it is easy to cause a catastrophic false alarm storm under the interference of external air turbulence or small displacement of non-specific objects; in order to achieve dynamic system balance under extreme conditions, the following preset control jump mechanism must be nested before the sinking degradation: in response to the weight reorganization condition triggering instruction, the self-running state flag is flipped and updated from the joint normal flag bit to the single-mode degradation trigger discrete flag bit (where 0 value corresponds to normal and 1 value corresponds to degradation state); during the validity period of the single-mode degradation trigger discrete flag bit, the default joint comparator is prohibited from being called, and instead the security unidirectional degradation bias constant solidified in the file system is extracted; the original The basic multimodal alarm threshold and the safety one-way degradation bias constant are cumulatively aggregated and compensated to form an adaptive limit alarm scale that raises the level of false alarm prevention. An alarm pulse indicating that the target object has lost its ability to act is only output when the single-modal micro-multipath variance exceeds the adaptive limit alarm scale. For the introduced safety one-way degradation bias constant, in specific engineering simulations, the safety one-way degradation bias constant is obtained by reverse calibration based on the envelope of the highest noise floor fluctuation extreme value measured in pure RF deployment conditions. In this embodiment, its enhancement factor value is preferably... The range is [between 1.3 and 1.8 times the basic multimodal alarm threshold], with a preferred value of 1.5 times. By introducing an adaptive upward floating correction control mechanism for the threshold that is deeply coupled with the hardware disconnection and detachment state, the deadlock of high false alarms caused by the one-sided weight dependency architecture is broken. This avoids extreme downtime accidents such as the flood of invalid alarms or the complete loss of alarm capabilities during the shutdown of the external sensing subsystem. Thus, a robust flexible self-balancing game is achieved between the single-modal monitoring accuracy and the macro-global prevention and control of false alarms under harsh edge conditions.

[0048] The dynamic parameters generated by S3 and S4 are serialized and encapsulated into a structured parameter set. This parameter set, as the final adaptive dynamic risk warning model, is deployed to the edge alarm gateway's transmitting device. The edge alarm gateway's transmitting device parses the parameter set to update its local rule base. When extreme boundary conditions occur (in this embodiment, the elderly person forgets to wear their wristband, causing an interruption in the inertial link, and the first data indicates that the bathroom faucet has been turned on for a long time without any activity), the underlying hardware will trigger resource redirection routing rules based on the newly updated parameter set. The RF parsing process in other rooms is shut down, and all limited computing power is concentrated on extracting the multipath variance changes caused by weak breathing in the bathroom RF signal to determine the survival status of the person. This not only completes the transformation from algorithm to transmitting device configuration but also achieves system degradation in a multi-faceted way. It solves the task of issuing high-confidence alarm commands and proactively utilizes the underlying relay computing power allocation mechanism embedded in the same model parameter set when the hardware connection is lost, avoiding system-level disasters caused by the failure of a single device.

[0049] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing a dynamic risk early warning model for AI-based elderly care, characterized in that, Includes the following steps: S1. Obtain the first data of the whole-house equipment collaborative activity index, which represents the environmental correlation state within the target space. The first data is generated by graph node temporal convolution calculation based on the desensitized state jump log of the indoor environmental monitoring equipment. S2. Obtain the second data of feature spatial alignment difference that characterizes the transient displacement event of physical entity. The second data is generated based on the temporal matching similarity calculation between the indoor wireless radio frequency channel state sequence and the local spatial displacement vector sequence. S3. Based on the current environmental activity reflected in the first data, calculate the fault tolerance compensation scalar and threshold relaxation coefficient used to construct the dynamic judgment topology graph; use the fault tolerance compensation scalar and threshold relaxation coefficient to dynamically modulate the signal comparison time window width and matching tolerance baseline required for generating the second data, thereby obtaining the corrected alignment difference sequence containing context constraint logic. S4. Extract the risk weight factor that characterizes the true credibility of the event based on the slope of the variance change of the corrected alignment difference sequence. S5. Encapsulate the risk weighting factor, the comparison time window width, and the matching tolerance baseline to construct and generate an adaptive dynamic risk warning model parameter set; deploy the adaptive dynamic risk warning model parameter set to the local signal command transmitting device to update the local trigger rule library in the transmitting device used for intercepting false alarm signals and activating downgrade and substitute monitoring.

2. The method for constructing a dynamic risk early warning model for AI-based elderly care according to claim 1, characterized in that: The steps for acquiring the first data are as follows: extract the switch status switching timestamps generated by the indoor environmental monitoring equipment and perform de-identification processing to generate de-identified status transition logs; construct a data association graph with different environmental monitoring equipment as nodes and the geographical topological location of the equipment as edges; perform temporal convolution calculation on the graph node status to output numerical features representing the frequency of use and linkage of each device. The weighted average of the numerical features is used to obtain the whole-house device collaboration activity index.

3. The method for constructing a dynamic risk early warning model for AI-based elderly care according to claim 2, characterized in that: The specific steps for calculating and outputting numerical features representing the frequency of linkage in temporal convolution are as follows: Read the preset time sliding window width and spatial adjacency matrix from the local device file system; within the current time sliding window, extract the state transition frequency of each environmental monitoring node to construct the node feature vector at the current moment; perform matrix multiplication on the spatial adjacency matrix and the node feature vector to obtain the first intermediate aggregation vector, which represents the device state diffusion effect in the spatial dimension; extract the historical feature vector corresponding to the previous sliding window, and multiply it with the historical feature vector using a preset time decay coefficient to obtain the second intermediate time vector; perform addition on the first intermediate aggregation vector and the second intermediate time vector, and input the result of the addition operation into a nonlinear activation function. The output result is the numerical feature representing the frequency of linkage of each device.

4. The method for constructing a dynamic risk early warning model for AI-based elderly care according to claim 1, characterized in that: The steps for obtaining the second data are as follows: extract the radio frequency channel state sequence from the underlying protocol of the indoor communication network; extract the extremely low frequency local spatial displacement vector sequence reported by the wearable computing device; using canonical correlation analysis algorithm, extract the fluctuation component representing the displacement of large-volume objects in the radio frequency channel state sequence and the component representing the fall of local limbs in the local spatial displacement vector sequence, thereby mapping the two sets of sequences to a feature alignment space of the same dimension; by calculating the dynamic time warping value of the time curves of the two sets of sequences in the feature alignment space, determine whether the spatial disturbance detected by radio frequency and the limb movement detected by inertial detection belong to the same transient physical event, and then generate the second data representing the cross-confirmation result.

5. The method for constructing a dynamic risk early warning model for AI-based elderly care according to claim 1, characterized in that: The calculation steps for constructing the fault-tolerant compensation scalar and threshold relaxation coefficient of the dynamic judgment topology are as follows: When the acquired first data is greater than the preset environmental active state benchmark value, it is confirmed that there is continuous housework interference in the current target space; based on the difference between the first data and the environmental active state benchmark value, the fault-tolerant compensation scalar used to delay the comparison judgment is calculated in the forward direction, and the threshold relaxation coefficient used to expand the difference tolerance interval is calculated simultaneously; using the fault-tolerant compensation scalar and threshold relaxation coefficient, the signal comparison time window width and matching tolerance benchmark required to generate the second data are dynamically modulated.

6. The method for constructing a dynamic risk early warning model for AI-based elderly care according to claim 5, characterized in that: The dynamic modulation process is as follows: the width of the reference comparison time window is added to the fault tolerance compensation scalar to obtain the extended signal comparison time window width, and the reference matching tolerance threshold is multiplied by the threshold relaxation coefficient to obtain the amplified matching tolerance baseline.

7. The method for constructing a dynamic risk early warning model for AI-based elderly care according to claim 1, characterized in that: The steps for constructing the risk weight factor are as follows: Calculate the data discrete variance of the corrected alignment difference sequence within the preset sliding historical time interval; calculate the slope of the change of the data discrete variance at the current sampling time using the first-order difference function; when the slope of change is greater than the warning surge threshold, generate the risk weight factor with the highest priority value; when the slope of change is not greater than the warning surge threshold, but the absolute value of the discrete variance is greater than the safety baseline, generate the risk weight factor with the second-level value.

8. The method for constructing a dynamic risk early warning model for AI-based elderly care according to claim 7, characterized in that: The following steps are used to extract the discrete variance of the data in the corrected alignment difference sequence: extract all corrected alignment difference sequence data within a preset sliding observation window width that is traced back from the current sampling time, and count the total number of data points within the current sliding observation window width; The local mean within the time interval is obtained by summing all data points within the sliding observation window and dividing the sum by the total number of data points. Then, for each data point within the sliding observation window, the difference between each data point and the local mean is calculated, and each difference is squared to generate a set of squared residuals. All elements in the set of squared residuals are summed to obtain the residual sum of squares. To obtain unbiased sample statistical characteristics, a constant 1 is subtracted from the total number of data points to obtain a correction divisor. The residual sum of squares is then divided by the correction divisor, and the resulting quotient is the data variance at the current time point.

9. The method for constructing a dynamic risk early warning model for AI-based elderly care according to claim 1, characterized in that: The local trigger rule base is constructed as follows: the risk weight factor, the comparison time window width, and the matching tolerance baseline are written into the judgment logic layer of the anti-false alarm signal interception as a new signal release standard; resource redirection routing rules are written into the execution rules of the downgraded backup monitoring.

10. The method for constructing a dynamic risk early warning model for AI-based elderly care according to claim 9, characterized in that: The resource redirection routing rule is configured as follows: monitor the underlying hardware data bus, and when no local spatial displacement vector sequence is received for a continuous preset silence period, and the first data is lower than the preset silence stagnation benchmark, generate a link interruption flag; based on the link interruption flag, send a forced sleep command to the process that manages the parsing of wireless radio frequency signals in non-associated areas, and interrupt the corresponding polling operation; Read the one-dimensional weight allocation vector used for multimodal fusion determination from the current working memory; perform an asymmetric parameter overlay operation on the one-dimensional weight allocation vector, assign the inertial weight factor to the value 0, and assign the radio frequency weight factor to the normalized maximum upper limit value of 1; through the absolute value redistribution logic, forcibly tilt and reorganize the model risk determination weights to the fluctuation characteristics of the single wireless radio frequency channel state sequence in the target stagnant area, thereby triggering the single-mode micro multipath variance extraction algorithm to determine the survival status of personnel.

Images