An old people intelligent home equipment intelligent selection recommendation method based on multi-feature fusion

Abstract

The application discloses a kind of old people intelligent home equipment intelligent selection recommendation method based on multi-feature fusion, it is related to the field of intelligent home and old people auxiliary technology;Solve the problem that existing selection method cannot consider the physiological state change of old people demand level, function redundancy and physical conflict among multiple devices.This application obtains the multi-dimensional attribute of old people and device, calculates hierarchical static matching score in hyperbolic space, then constructs individual physiological time attenuation factor using maximum oxygen uptake decline trend and corrects to obtain dynamic matching score, combines the rigid demand dimension of old people and mutual exclusion constraint among devices, iteratively selects device combination with maximum demand net coverage gain as the goal, realizes hierarchical matching, physiological self-adaptation and collaborative conflict-free intelligent home equipment selection recommendation.

Description

Technical Field

[0001] This invention belongs to the field of smart home and elderly assistance technology, specifically a method for intelligent selection and recommendation of smart home devices for the elderly based on multi-feature fusion. Background Technology

[0002] As the population ages, assisting the elderly in living independently and safely through smart home devices has become an important technological direction. Currently, the variety of smart home devices for the elderly is increasingly rich, covering multiple categories such as fall detection radars, emergency call buttons, smart pillboxes, ambient temperature and humidity regulators, nighttime guiding lights, voice interaction terminals, and access control alarms. Selecting suitable devices or combinations of devices from a wide range of options for a specific elderly person requires not only examining the technical parameters and functional specifications of the devices themselves, but also comprehensively considering the elderly person's health status, behavioral habits, living environment, and other multi-dimensional characteristics.

[0003] Existing device recommendation methods for the elderly typically employ rule-based questionnaire matching or collaborative filtering-based score prediction. The former relies on manually pre-set simple mapping tables, such as directly assigning a type of device based on an elderly person's medical history tags, lacking a refined utilization of multi-dimensional features; the latter treats devices as ordinary commodities, recommending based on similarities between elderly individuals or between devices, ignoring the inherent hierarchical relationship between elderly needs and device functions. Furthermore, some methods introduce time decay mechanisms to handle demand drift caused by changes in the elderly's physical condition, but the decay rate often uses a fixed coefficient, failing to reflect individual differences in the rate of aging among different elderly individuals. When recommending multiple devices simultaneously, existing methods generally sort by individual device scores and then select the top few, failing to consider the potential functional redundancy and physical installation conflicts that may arise when multiple devices coexist, leading to usability issues in actual deployment.

[0004] The following problems exist in the existing technology: Existing methods use rule-based questionnaire matching or collaborative filtering approaches, treating devices as ordinary commodities and recommending them based on similarity among elderly people or between devices. This ignores the inherent hierarchical relationship between the needs of the elderly and the functions of the devices, fails to distinguish the priority differences of different need categories, and is prone to incorrectly recommending devices that are outside the category as high-scoring results. Existing methods use a fixed half-life or a fixed forgetting coefficient when introducing a time decay mechanism. The decay rate is linked to the passage of calendar time, which cannot reflect the individual differences in the aging speed of different elderly people. This results in the recommendation update being delayed for elderly people with rapid physiological decline, while the matching score of elderly people with stable physiological state is unnecessarily lowered. Existing methods, when recommending multiple devices simultaneously, generally adopt the approach of sorting by individual device scores and then selecting the top few. This approach does not consider the functional redundancy and physical installation conflicts that may arise when multiple devices coexist, resulting in issues such as functional overlap, wireless frequency interference, or mutually exclusive installation locations in the recommendation results, leading to insufficient usability during actual deployment. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a smart selection and recommendation method for smart home devices for the elderly based on multi-feature fusion, in order to solve the above-mentioned technical problem.

[0006] The first aspect of this invention provides a method for intelligent selection and recommendation of smart home devices for the elderly based on multi-feature fusion, comprising the following steps: S1: Obtain multidimensional characteristic data and daily health monitoring data of the elderly, as well as multidimensional attribute data of the selected smart home devices; among which, the multidimensional characteristic data of the elderly should at least include basic attributes, health indicators, behavioral abilities and environmental needs, and the multidimensional attribute data of the selected smart home devices should at least include functional categories, technical parameters and installation constraints. S2: Map the multidimensional feature data of the elderly and the multidimensional attribute data of the candidate smart home devices to the same hyperbolic space. Based on the geodesic distance between the vector corresponding to the elderly and the vector corresponding to the candidate smart home devices in the hyperbolic space, calculate the static matching score that reflects the degree of hierarchical matching between the two. S3: Based on the decreasing trend of maximum oxygen uptake extracted from the daily health monitoring data over time, an individualized physiological time decay factor is constructed; the individualized physiological time decay factor is used to perform time-series correction and normalization on the static matching score to obtain the final dynamic matching score. S4: Based on the final dynamic matching score, combined with the predefined set of rigid needs of the elderly and the mutual exclusion constraints between devices constructed according to the multi-dimensional attribute data of the candidate smart home devices, with the goal of maximizing the net coverage gain of the demand, the device combination without functional redundancy and physical conflict is iteratively selected from the candidate smart home devices, and the device combination and its demand coverage description are output as the final recommendation result.

[0007] Preferably, step S1 includes the following steps: The multidimensional feature data of the elderly is a cross-sectional feature vector composed of four sub-vectors: basic attributes, health indicators, behavioral abilities and environmental needs. The values ​​of each continuous health indicator are the values ​​after moving average and normalization. The daily health monitoring data of the elderly are stored independently in the form of time-series records with timestamps, and include at least two of the following continuous time series: resting heart rate, weight, body fat percentage, sleep duration and number of times to get up at night; The functional categories in the multi-dimensional attribute data of the candidate smart home devices are hierarchically coded using a three-level functional tree, and the installation constraints must include at least a binary identifier for the exclusivity of the device installation location.

[0008] Preferably, in step S2, mapping the multidimensional feature data of the elderly person and the multidimensional attribute data of the candidate smart home devices to the same hyperbolic space includes the following steps: Construct a heterogeneous graph containing elderly nodes, device nodes, and functional tag nodes. The functional tag nodes are taken from the three-level functional tree of the functional category and form a hierarchical connection according to the parent-child relationship. The elderly node is connected to the associated functional tag node according to the medical history tag in its basic attributes and the behavioral ability. The device node is connected to the corresponding three-level functional tag node according to its functional category. Perform graph convolution on the heterogeneous graph to embed each node into a hyperbolic space, which is a Lorentz manifold with constant negative cross-sectional curvature. The vector corresponding to the elderly person is obtained. and the vector corresponding to the candidate smart home devices , and All meet and ,in Let be the 0th dimension component of the vector.

[0009] Preferably, in step S2, the calculation of the static matching score, which reflects the degree of hierarchical matching between the vector corresponding to the elderly person and the vector corresponding to the candidate smart home device in hyperbolic space, includes the following steps: In the Lorentz manifold The above calculates the vector corresponding to the elderly. Vector corresponding to the candidate smart home devices The shortest path length between them is taken as the geodesic distance in the hyperbolic space; Then calculate the static matching score. Where u represents the elderly person, i represents the smart home device to be selected, and Lorentz product .

[0010] Preferably, in step S3, constructing an individualized physiological time decay factor based on the decreasing trend of maximum oxygen uptake extracted from the daily health monitoring data over time includes the following steps: From the resting heart rate, weight, and body fat percentage data in the daily health monitoring, the estimated maximum oxygen uptake is calculated on a time-by-time basis using a submaximal estimation model, and a smoothed maximum oxygen uptake sequence for the current time t is obtained by moving average. ; Calculate the time derivative of the smoothed VO2 max sequence, and then compare the derivative with... The product of these factors is determined as the effective descent rate. ,in, τ represents the value of 1 if the condition inside the square brackets is true and 0 otherwise, where τ is the integration variable; The effective descent rate at the initial assessment time The integral up to the current time t is multiplied by the preset sensitivity parameter γ and then negativeened. This negative value is used as the input to the exponential function to construct the individualized physiological time decay factor. , where u represents an elderly person.

[0011] Preferably, in step S3, the step of performing time-series correction and normalization on the static matching score using an individualized physiological time decay factor to obtain the final dynamic matching score includes the following steps: Dynamic matching score at current time t The calculation formula is: Where u represents the elderly person, i represents the smart home device to be selected, s(u,i) is the static matching score, and γ is the preset sensitivity parameter. For the smoothed maximum oxygen uptake sequence, This indicates that the value is 1 if the condition within the square brackets is true, and 0 otherwise. Let τ be the initial evaluation time, and τ be the integration variable; The dynamic matching score of each elderly person u across all the candidate smart home devices. The final dynamic matching score is obtained by normalization. .

[0012] Preferably, in step S4, the predefined set of rigid needs of the elderly and the inter-device mutual exclusion constraints constructed based on the multi-dimensional attribute data of the candidate smart home devices include the following steps: The predefined set of rigid needs for the elderly consists of multiple needs dimensions, each of which corresponds to a basic functional guarantee required for the elderly to live independently and safely. Based on the functional categories in the multidimensional attribute data of the candidate smart home devices, a binary coverage matrix is ​​constructed, where each element represents whether a candidate smart home device meets one of the requirement dimensions in the set of rigid requirements in terms of functional specifications. For candidate smart home devices that meet the requirements in terms of function but whose final dynamic matching score is lower than the preset validity threshold, their corresponding elements are set to uncovered. Based on the installation constraints and technical parameters in the multi-dimensional attribute data of the candidate smart home devices, a mutual exclusion degree between the devices is constructed: if the exclusive identification of the installation location of two candidate smart home devices is exclusive and points to the same location, the mutual exclusion degree is taken as a first preset value; if the communication protocols of two candidate smart home devices use the same wireless frequency band and are marked as having interference risk, the mutual exclusion degree is taken as a value within a preset range; if the number of overlapping function tags of two candidate smart home devices exceeds a preset overlap threshold, the mutual exclusion degree is taken as a second preset value; otherwise, the mutual exclusion degree is zero.

[0013] Preferably, in step S4, the step of iteratively selecting and generating a device combination without functional redundancy and physical conflicts from the candidate smart home devices with the goal of maximizing the net coverage gain of demand includes the following steps: Starting from an empty set S of selected smart home devices, an iterative selection process is executed; For each candidate smart home device that has not yet been selected, calculate its marginal net coverage gain after being added to the set S of selected smart home devices. The calculation formula is: Where C(S) is a subset of the set of rigid needs of the elderly covered by S. The number of the required dimensions newly covered after adding device i. Let η be the mutual exclusion degree between device i and device j, and η be the preset mutual exclusion penalty coefficient. In all satisfied Among the unselected smart home devices, select to use The largest equipment is added to S, of which To preset a global mutual exclusion tolerance threshold; When all unselected smart home devices The iteration stops at non-positive times, and it also stops when the total number of demand dimensions covered by S reaches the preset coverage target as a proportion of the total number of demand dimensions contained in the set of rigid demand dimensions for the elderly; after the iteration stops, S is a combination of devices without functional redundancy and physical conflict.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention maps multidimensional feature data of the elderly and multidimensional attribute data of candidate smart home devices to the same hyperbolic space, and calculates static matching scores based on geodesic distance on the Lorentz manifold. The negative curvature of the hyperbolic space can encode the hierarchical structure in the three-level functional tree without distortion, so that the needs of the elderly and the functions of the devices belonging to the same fine-grained functional branch are automatically brought closer, while nodes across major categories are pushed further apart, thereby achieving accurate measurement of the degree of hierarchical matching and avoiding misrecommendation across major categories. This invention extracts the decreasing trend of VO2 max over time from daily health monitoring data, constructs an individualized physiological time decay factor, and uses this factor to perform time-series correction and normalization on the static matching score to obtain the final dynamic matching score. The decay rate is determined by the individual's own VO2 max decline trajectory. The matching score of elderly people with stable physical function is maintained, while the matching score of elderly people with continuously declining function is promptly reduced. The recommendation results can adaptively follow the actual physiological change rhythm of each elderly person. This invention combines a predefined set of rigid needs for the elderly with mutual exclusion constraints between devices constructed based on multi-dimensional attribute data of candidate smart home devices. With the goal of maximizing net coverage gain, it iteratively selects and generates a combination of devices from candidate smart home devices that are free from functional redundancy and physical conflicts. The mutual exclusion constraints cover three types of situations: exclusive installation location, wireless frequency band interference, and overlapping functional labels. In each round of evaluation of marginal net coverage gain, the mutual exclusion cost is explicitly deducted during the iterative selection process, ensuring that the final selected combination of devices is functionally complementary and physically coexistent, and can be directly used for on-site deployment. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0016] 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.

[0017] Please see Figure 1 This invention provides a method for intelligent selection and recommendation of smart home devices for the elderly based on multi-feature fusion, comprising the following steps: S1: Obtain multidimensional characteristic data and daily health monitoring data of the elderly, as well as multidimensional attribute data of the selected smart home devices; among which, the multidimensional characteristic data of the elderly should at least include basic attributes, health indicators, behavioral abilities and environmental needs, and the multidimensional attribute data of the selected smart home devices should at least include functional categories, technical parameters and installation constraints. S2: Map the multidimensional feature data of the elderly and the multidimensional attribute data of the candidate smart home devices to the same hyperbolic space. Based on the geodesic distance between the vector corresponding to the elderly and the vector corresponding to the candidate smart home devices in the hyperbolic space, calculate the static matching score that reflects the degree of hierarchical matching between the two. S3: Based on the decreasing trend of maximum oxygen uptake extracted from the daily health monitoring data over time, an individualized physiological time decay factor is constructed; the individualized physiological time decay factor is used to perform time-series correction and normalization on the static matching score to obtain the final dynamic matching score. S4: Based on the final dynamic matching score, combined with the predefined set of rigid needs of the elderly and the mutual exclusion constraints between devices constructed according to the multi-dimensional attribute data of the candidate smart home devices, with the goal of maximizing the net coverage gain of the demand, the device combination without functional redundancy and physical conflict is iteratively selected from the candidate smart home devices, and the device combination and its demand coverage description are output as the final recommendation result.

[0018] Specifically, the process involves acquiring multidimensional characteristic data of the elderly (including basic attributes, health indicators, behavioral abilities, and environmental needs), daily health monitoring data, and multidimensional attribute data of candidate smart home devices (including functional categories, technical parameters, and installation constraints). The multidimensional characteristics of the elderly and the multidimensional attributes of the devices are mapped to the same hyperbolic space. Based on the geodesic distance between the corresponding vectors of the elderly and the corresponding vectors of the devices in the hyperbolic space, a static matching score reflecting the degree of hierarchical matching is calculated. Based on the decreasing trend of maximum oxygen uptake extracted from daily health monitoring data over time, an individualized physiological time decay factor is constructed. This factor is then used to perform time-series correction and normalization on the static matching score to obtain the final dynamic matching score. Based on the final dynamic matching score, combined with a predefined set of rigid needs for the elderly and mutual exclusion constraints between devices constructed based on the multidimensional attribute data, and with the goal of maximizing net demand coverage gain, a device combination without functional redundancy or physical conflict is iteratively selected from the candidate devices. This combination and its demand coverage description are then output as the final recommendation result.

[0019] In one embodiment of the present invention, step S1 includes the following steps: The multidimensional feature data of the elderly is a cross-sectional feature vector composed of four sub-vectors: basic attributes, health indicators, behavioral abilities and environmental needs. The values ​​of each continuous health indicator are the values ​​after moving average and normalization. The daily health monitoring data of the elderly are stored independently in the form of time-series records with timestamps, and include at least two of the following continuous time series: resting heart rate, weight, body fat percentage, sleep duration and number of times to get up at night; The functional categories in the multi-dimensional attribute data of the candidate smart home devices are hierarchically coded using a three-level functional tree, and the installation constraints must include at least a binary identifier for the exclusivity of the device installation location.

[0020] Specifically, the multidimensional characteristic data of the elderly includes at least basic attributes, health indicators, behavioral abilities, and environmental needs. The basic attributes are obtained by retrieving age, gender, medical history tags, and living status from the elderly's electronic health records or structured forms entered by their families. Living status is selected from living alone, living together, living with children, or receiving institutional care. Medical history tags are selected from a pre-defined list of common chronic diseases in the elderly, containing 20 entries such as hypertension, type 2 diabetes, and osteoarthritis. Each elderly person's medical history tag is represented as a binary vector of length 20, set to 1 if the corresponding entry exists, and 0 otherwise. Living status is encoded using one-hot encoding to generate a 4-dimensional vector. Age and gender are directly used as scalar features, forming the basic attribute sub-vector along with other encoded values. .

[0021] Health indicators were obtained based on the elderly person's daily physiological parameter records over the past 30 days, including at least five continuous numerical features: resting heart rate (beats / minute), weight (kg), body fat percentage (%), sleep duration (hours / day), and number of nighttime awakenings (times / night). Specifically, resting heart rate was taken as the lowest heart rate during sleep, averaged over 7 days; weight was the weekly average of the morning fasting weight; body fat percentage was the weekly average; sleep duration was the weekly average of the total actual sleep duration each night; and the number of nighttime awakenings was obtained by counting the number of times the person got out of bed at night as recorded by the mattress pressure sensor. Each of the five continuous values ​​was standardized using z-scores, which normalizes the individual mean and standard deviation over the past 90 days to obtain dimensionless standardized values. The five standardized values ​​were then concatenated to form a health indicator subvector. .

[0022] Behavioral abilities were assessed based on the elderly person's daily activity records within their residence, including at least four indicators: average daily steps, activity radius (meters), estimated gait speed (meters / second), and number of falls in the past three months. Average daily steps were calculated from 30 days of step count data from a smart bracelet; the activity radius was calculated by taking the square root of the convex hull area from the coordinates of the elderly person's daily activities recorded by indoor positioning beacons, and then converted to an equivalent radius; the estimated gait speed was estimated from the product of gait frequency and stride length provided by gait sensors, and the weekly average was used; the number of falls in the past three months was accumulated from fall detection logs. All four values ​​were also standardized using z-scores, based on the elderly person's 90-day historical mean and standard deviation, to form a behavioral ability sub-vector. .

[0023] Environmental requirements are obtained based on a residential environment questionnaire completed by family members and a list of existing smart home devices installed. This includes at least four categories of information: apartment type, main activity area area, existing smart home device protocol types, and network coverage. Apartment types are categorized into single-level apartments, duplexes, and villas, using one-hot encoding. The main activity area area is measured in square meters, with the maximum area of ​​the bedroom, living room, kitchen, and bathroom taken as a continuous feature. Existing smart home device protocol types include Wi-Fi 2.4GHz, Bluetooth 5.0, etc., using multi-hot encoding to represent existing protocol types. Network coverage is represented by the signal strength level of each room, with values ​​of strong, medium, and weak, encoded as 3, 2, and 1 respectively. The lowest signal strength level across all rooms is taken as the overall coverage score. All codes and continuous values ​​are concatenated to obtain the environmental requirement sub-vector. .

[0024] By concatenating the four sub-vectors in sequence, we obtain the multidimensional feature vector of the elderly person u. The total dimension n of the vector depends on the dimension of each subvector, with a typical value of about 80 dimensions.

[0025] Daily health monitoring data is stored separately as time-series records with timestamps. Each record contains a continuous time series of at least two of the following: resting heart rate, weight, body fat percentage, sleep duration, and number of nighttime urinations; typically, all five are included. The data source is the same as for health indicators, but raw daily or weekly samples are retained without moving averages or normalization. Each record is labeled with a date, and the time unit can be day or week. When using a weekly unit, resting heart rate is the average of daily resting heart rates within that week; weight and body fat percentage are the last measurement of that week; and sleep duration and number of nighttime urinations are the daily averages of that week. All time-series records are arranged chronologically.

[0026] The multi-dimensional attribute data of the candidate smart home devices includes at least three parts: functional category, technical parameters, and installation constraints. Each part is extracted from the specifications or open APIs provided by the device manufacturer and encoded as a structured vector. Functional categories are hierarchically encoded using a three-level functional tree. The top level of the three-level functional tree represents the major functional category, such as life safety or health management; the second level represents sub-functional domains, such as fall detection and emergency call under life safety; and the third level represents the specific device type or key technology, such as millimeter-wave radar and infrared detectors under fall detection. The functional category of each device is represented by the hierarchical path of its corresponding three-level label, and a multi-hot encoding method is used to generate the functional category sub-vector. During encoding, all possible third-level label nodes in the three-level function tree are arranged in a fixed order, with the position corresponding to the device's label set to 1, and the rest set to 0. For example, if the function category of a millimeter-wave fall radar is "Life Safety - Fall Detection - Millimeter-wave Radar", the three-level function tree contains 50 third-level label nodes. The encoding vector of this device has a 1 in the 17th bit (corresponding to the millimeter-wave radar), and the rest are 0.

[0027] Technical parameters include communication protocol, power supply method, power consumption, response delay, detection range, and false alarm rate, which are encoded as technical parameter sub-vectors. The communication protocol uses one-hot encoding, with candidate types including Wi-Fi 2.4GHz, Zigbee 3.0, etc.; the power supply method is divided into battery power, plug-in power, and Ethernet power supply, also using one-hot encoding; power consumption is expressed in watts as a continuous value of rated power; response delay is expressed in milliseconds as the nominal value; detection range is expressed in meters, valid only for sensor-type devices, fill in 0 for non-sensor devices; false alarm rate is expressed as a decimal, obtained from the nominal false alarm rate provided by the device manufacturer in the specifications, usually the ratio of the number of false alarms per unit time or unit event measured under standard test conditions to the total number of alarms, fill in -1 for devices without this parameter to distinguish them.

[0028] Installation constraints include installation location requirements, whether a neutral line is required (binary field), whether fixing holes are required (binary field), compatibility with existing gateways, and installation location exclusivity identifiers, which are encoded as installation constraint sub-vectors. Installation location requirements are categorized into four types: wall-mounted, ceiling-mounted, desktop placement, and sticker-mounted, with unique hot-coded properties. Compatibility is a category field with values ​​of compatible, requiring an additional gateway, or incompatible, also with unique hot-coded properties. The installation location exclusivity identifier is a binary field indicating whether the device must exclusively occupy a certain installation location and cannot share the same wall, ceiling, or desktop area with other devices.

[0029] Concatenate the three sub-vectors sequentially to obtain the multidimensional attribute vector of device i. The total dimension m of this vector is determined by the total number of functional labels, the quantification dimension of technical parameters, and the encoding dimension of installation constraints, with a typical value of about 120 dimensions.

[0030] In one embodiment of the present invention, step S2, mapping the multidimensional feature data of the elderly person and the multidimensional attribute data of the candidate smart home devices to the same hyperbolic space, includes the following steps: Construct a heterogeneous graph containing elderly nodes, device nodes, and functional tag nodes. The functional tag nodes are taken from the three-level functional tree of the functional category and form a hierarchical connection according to the parent-child relationship. The elderly node is connected to the associated functional tag node according to the medical history tag in its basic attributes and the behavioral ability. The device node is connected to the corresponding three-level functional tag node according to its functional category. Perform graph convolution on the heterogeneous graph to embed each node into a hyperbolic space, which is a Lorentz manifold with constant negative cross-sectional curvature. The vector corresponding to the elderly person is obtained. and the vector corresponding to the candidate smart home devices , and All meet and ,in Let be the 0th dimension component of the vector.

[0031] Specifically, a heterogeneous graph G is constructed, comprising three types of nodes: elderly nodes, device nodes, and function tag nodes. Each elderly person corresponds to an elderly node, each candidate smart home device corresponds to a device node, and function tag nodes are derived from all tags on the three-level function tree of the function categories in the multi-dimensional attribute data of the candidate smart home devices. First-level, second-level, and third-level tags all exist as independent nodes. Total number of nodes... ,in For the number of elderly people, The number of smart home devices to be selected. G represents the total number of all label nodes in the three-level functional tree. The adjacency relationships of G are represented by an adjacency matrix A, with dimensions |V| × |V|. Nodes with edges have a value of 1 for their corresponding element, otherwise 0. All edges are undirected, and the adjacency matrix is ​​symmetric. Edges between functional label nodes are established according to the parent-child relationships of the three-level functional tree; for example, an undirected edge is established between the first-level label "Life Safety" and the second-level label "Fall Detection," and another undirected edge is established between the second-level label "Fall Detection" and the third-level label "Millimeter-wave Radar." The weight of each edge is set to 1. Edges between elderly nodes and functional label nodes are established based on the elderly person's medical history labels and behavioral ability data in their basic attributes; specifically, a mapping table from medical history labels and behavioral ability status to functional labels is maintained in advance. The mapping table is constructed according to the following rules: Medical history tag mapping rules: 20 preset chronic disease history tags are divided into three categories based on their pathological impact. The first category, circulatory system tags (including hypertension and coronary heart disease), is fixedly mapped to the health monitoring and medication reminder function tag; the second category, musculoskeletal system tags (including osteoarthritis and Parkinson's disease), is fixedly mapped to the fall detection and voice interaction function tag; and the third category, metabolic system tags (including type 2 diabetes), is fixedly mapped to the environmental temperature and humidity regulation and medication reminder function tag. Behavioral ability mapping rules: for example, if the number of falls in the past 3 months is ≥1, it is unconditionally mapped to the fall detection function tag. Medical histories not explicitly listed are only connected to the emergency call function tag by default to ensure basic safety coverage. For each medical history tag and each behavioral ability status of an elderly node, the associated function tag is found in the mapping table, and an edge with a weight of 1 is established between the elderly node and these function tag nodes. The edges between device nodes and function tag nodes are established based on the function categories in the device's multidimensional attribute data. Each device's attribute data already contains its corresponding third-level function tag. An edge is established between the device node and the third-level function tag node, with an edge weight of 1.

[0032] Each node is assigned an initial feature vector; the initial feature vector for the elderly node is taken from the elderly's multidimensional feature vector. The initial feature vector of the device node is taken from the multi-dimensional attribute vector of the candidate smart home device. The initial feature vector of the function tag node is obtained by word embedding of its tag name text; the word embedding uses a 50-dimensional static word vector model trained on the Chinese Wikipedia corpus. Each function tag name is mapped to a 50-dimensional vector by this model, and the public vocabulary of this word vector model contains all tag names on the three-level function tree. The initial feature vectors of the elderly node and the device node have different dimensions (approximately 80 and 120 dimensions respectively). They are mapped to a unified dimension d+1 through a fully connected linear transformation, where d is the dimension of the hyperbolic space, typically 64 or 128. After mapping, the feature vector of the elderly node... and the feature vector of the device node All are d+1 dimensional vectors.

[0033] A Lorentz graph convolutional network is performed on a heterogeneous graph G, embedding each node into the same hyperbolic space. This hyperbolic space is a Lorentz manifold with constant negative cross-sectional curvature. , Defined as satisfying and A set of d+1 dimensional vectors.

[0034] In the initial training phase of a Lorentz graph convolutional network, the initial feature vectors in Euclidean space need to be mapped onto the hyperbolic space, serving as the embedding vectors for nodes at layer 0; this mapping is accomplished through exponential mapping. Taking the origin of the Lorentz manifold (i.e., vector (1,0,…,0)) as the base point, the feature vector of each node is linearly transformed... Treating the node as a vector in the tangent space, the initial embedding vector of the node in the hyperbolic space is obtained through exponential mapping. The message passing process for each layer of graph convolution is as follows: For the l-th layer, the neighbor set of node v is N(v), and the neighbor embedding vector of v is... The vector is projected onto the tangent space of v using a logarithmic map, where linear transformations and aggregations are performed. The aggregation result is then returned to the hyperbolic space via an exponential map to obtain the embedding vector of v at the (l+1)th layer. The neighbor embedding vectors are averaged when aggregated in the tangent space; the total number of layers in the graph convolution is 2.

[0035] The objective function for training is a loss function based on positive and negative sample pairs. Positive sample pairs are node pairs in a heterogeneous graph that are directly connected by an edge; for example, an elderly node and its connected "fall detection" label node constitute a positive sample pair. Negative sample pairs are randomly sampled from all node pairs without direct edge connections; for example, a randomly selected elderly node and a device node unrelated to that elderly node constitute a negative sample pair. The loss function uses a contrastive loss, aiming to minimize the squared geodesic distance between the embedding vectors of the two nodes in the hyperbolic space of the positive sample pair, while maximizing the geodesic distance between the embedding vectors of the two nodes in the negative sample pair, until it exceeds a preset threshold value (e.g., 2). The training optimizer uses the Riemann Adam optimizer, performing gradient updates directly on the Lorentz manifold. The training convergence criterion is that the decrease in the loss function over 10 consecutive iterations is less than a preset value of 0.001. After training convergence, the embedding vectors of the elderly nodes in the hyperbolic space are extracted from the final layer embedding matrix. Embedding vectors of candidate smart home device nodes in hyperbolic space All dimensions are d+1, satisfying and For example, when d is 64, each embedding vector has 65 dimensions, and its 0th dimension component... It is a positive value, and the result of the Lorentz inner product operation on all 65 components is exactly -1.

[0036] In one embodiment of the present invention, step S2, which involves calculating a static matching score reflecting the degree of hierarchical matching between the vector corresponding to the elderly person and the vector corresponding to the candidate smart home device in hyperbolic space, includes the following steps: In the Lorentz manifold The above calculates the vector corresponding to the elderly. Vector corresponding to the candidate smart home devices The shortest path length between them is taken as the geodesic distance in the hyperbolic space; Then calculate the static matching score. Where u represents the elderly person, i represents the smart home device to be selected, and Lorentz product .

[0037] Specifically, in a Lorentz manifold with constant negative cross-sectional curvature Above, the vector corresponding to the elderly Vector corresponding to the candidate smart home devices The shortest path length between them is determined by the Lorentz geodesic distance. Given, where the Lorentz inner product is... for . , Let be the 0th dimension component of the vector. , For the subsequent d spatial components. Because... and satisfy and The conditions guaranteed ,thereby It falls within the domain of the inverse hyperbolic cosine function. When and When the two elements completely coincide on a Lorentz manifold, the inner product is -1 and the geodesic distance is 0; the greater the distance between them on the manifold, the more negative the inner product becomes. The larger the value, the greater the geodetic distance.

[0038] The negative value of the geodesic distance is taken as the static matching score. The score is a nonlinear metric naturally derived from the geometry of the Lorentz manifold. Because the Lorentz manifold has negative curvature, the geodesic distance between nearby nodes in the embedding space is significantly reduced in the tree hierarchy, while nodes across large branches are pushed apart. Therefore, s(u,i) reflects the degree of hierarchical matching between the needs of the elderly and the functions of the devices on the three-level functional tree. The score ranges from (−∞,0], where 0 represents a perfect match.

[0039] For example, the embedding vector of the elderly u in hyperbolic space with dimension d=64. The embedding vector of a certain device i This is a 65-dimensional vector extracted after training the Lorentz diagram convolutional network; the first few dimensions are as follows: =(1.0321,0.2147,−0.0892,…), =(1.0289,0.2083,−0.0915,…); Calculate the Lorentz inner product. =−1.0321×1.0289+0.2147×0.2083+(−0.0892)×(−0.0915)+……, assuming the sum of contributions from other dimensions is −0.0156, then =−1.0246, taking the negative gives ≈1.0246; Calculating arcosh(1.0246)≈0.2213, then... ≈0.2213, s(u,i)≈−0.2213.

[0040] If the elderly person and the device are highly matched on the function tree (e.g., the elderly person's needs label and the device's function label belong to the same fine-grained branch), the embedding vectors will be very close, the inner product will be close to −1, and the score will be close to 0; conversely, if the matching degree is low, the score will be more negative. For example, another device that is not related to the elderly person's needs may get s(u,i)≈−3.5.

[0041] In one embodiment of the present invention, step S3, which involves constructing an individualized physiological time decay factor based on the decreasing trend of maximum oxygen uptake extracted from the daily health monitoring data over time, includes the following steps: From the resting heart rate, weight, and body fat percentage data in the daily health monitoring, the estimated maximum oxygen uptake is calculated on a time-by-time basis using a submaximal estimation model, and a smoothed maximum oxygen uptake sequence for the current time t is obtained by moving average. ; Calculate the time derivative of the smoothed VO2 max sequence, and then compare the derivative with... The product of these factors is determined as the effective descent rate. ,in, τ represents the value of 1 if the condition inside the square brackets is true and 0 otherwise, where τ is the integration variable; The effective descent rate at the initial assessment time The integral up to the current time t is multiplied by the preset sensitivity parameter γ and then negativeened. This negative value is used as the input to the exponential function to construct the individualized physiological time decay factor. , where u represents an elderly person.

[0042] Specifically, from daily health monitoring data, we obtain the elderly person's weight W(t) (in kilograms), body fat percentage B(t) (expressed as a decimal, e.g., 27.2% is expressed as 0.272), and resting heart rate at time t. (Unit: beats / minute), where t is the current time. In this embodiment, the time unit is a week, and the specific value of t is the current week number. Maximum heart rate (HRmax) is calculated based on age, i.e., HRmax = 220 - age. If the calculated result is lower than 160 beats / minute, this calculated value is used; otherwise, 160 beats / minute is used. For example, for a 78-year-old, the calculated HRmax is 142 beats / minute, which is lower than 160, so HRmax is taken as 142 beats / minute.

[0043] Maximum oxygen uptake The unit is .in Let be the correction function for body weight and body fat percentage; where, The ideal weight is obtained by subtracting 105 from the height. The standard body fat percentage is 0.25 for men and 0.35 for women; the coefficients α and β are 0.3 and 0.2 respectively. After a 7-day moving average, the daily resting heart rate is taken as the lowest value during sleep, and the arithmetic mean of these lowest values ​​over the past 7 days is calculated to obtain the daily moving average. If the time unit is a week, the weekly average of the daily moving averages within that week is taken as the weekly moving average. W(t) and B(t) are the last measurements of the week. The raw maximum oxygen uptake sequence is calculated week by week. A 30-week moving average is applied to this raw sequence to smooth the maximum oxygen uptake sequence, denoted as . The moving average is calculated as follows: for week t, take the arithmetic mean of VO2max values ​​of t and the previous 29 weeks.

[0044] right Find the time derivative in discrete cycles, approximating it with the difference between two adjacent cycles: Where τ is the integration variable, which iterates through each cycle in this embodiment; the time interval... =1 week. The effective rate of decline is defined to preserve only the monotonic decline in function caused by aging or deteriorating health. When the time derivative The derivative is taken when it is negative, otherwise it is taken as 0; the square brackets [⋅] indicate that the value is 1 when the condition is true, otherwise it is taken as 0.

[0045] In individualized physiological time decay factor In the calculation formula, The initial assessment week number is the time when the elderly person first participated in the recommendation or the last comprehensive health assessment; The sensitivity parameter is set to 1.0 in this embodiment. If it is necessary to adjust the attenuation response to physiological changes, a value can be selected within the range of 0.5 to 1.5. τ is the integration variable, iterating through... Each time unit up to t.

[0046] because The integral term takes a negative value only when the smooth sequence is decreasing, and a zero value in all other cases. Therefore, A(u,t)∈(0,1). When the elderly person's physical function is stable or even improves, the effective rate of decline is always 0, the integral term is 0, A(u,t)=1, and the historical matching score is completely preserved; when the function continues to decline, the larger the negative value of the integral term, the smaller the decay factor, the lower the retention rate of the matching score, and the decay rate depends entirely on the elderly person's own physiological change trajectory.

[0047] For example, an elderly person's initial assessment week The current week (t) is the 1st week of 2024, and the current week (t) is the 52nd week of 2025, for a total of 104 weeks. During this period, The value gradually decreased from 30.2 to 28.5; the derivative was calculated weekly and the negative values ​​were extracted to obtain the effective rate of decrease sequence; assuming that the derivative was negative in 60 out of 104 weeks, the cumulative decrease was... =−1.7. Taking γ=1.0, the integral term is −1.7, so A(u,t)=exp(−1.7)≈0.1827; indicating that the historical matching score at the current moment will retain about 18.3% weight, and the home recommendation will shift significantly to new devices adapted to the current physiological state of the elderly.

[0048] In one embodiment of the present invention, step S3, which involves using an individualized physiological time decay factor to perform temporal correction and normalization on the static matching score to obtain the final dynamic matching score, includes the following steps: Dynamic matching score at current time t The calculation formula is: Where u represents the elderly person, i represents the smart home device to be selected, s(u,i) is the static matching score, and γ is the preset sensitivity parameter. For the smoothed maximum oxygen uptake sequence, This indicates that the value is 1 if the condition within the square brackets is true, and 0 otherwise. Let τ be the initial evaluation time, and τ be the integration variable; The dynamic matching score of each elderly person u across all the candidate smart home devices. The final dynamic matching score is obtained by normalization. .

[0049] Specifically, from the dynamic matching score As can be seen from the calculation formula, when the elderly person's physical function is stable over a long period, A(u,t)=1, the dynamic matching score is the same as the static matching score, and the recommendation result remains unchanged. When the elderly person experiences a sustained decline in maximum oxygen uptake due to aging or disease, A(u,t)<1, and the static score is proportionally suppressed. The greater the actual decline and the longer the duration, the stronger the suppression effect. This mechanism causes the matching score of high-performance devices that rely on high self-care ability or complex operation (such as smart temperature control panels that require frequent manual settings) to decrease significantly as the elderly person's function declines, while the scores of assistive and easy-to-operate devices (such as voice emergency call devices, fall detection radars, and automatic night lights) increase relatively. The recommendation result is adaptively adjusted to better match the elderly person's current actual situation.

[0050] To eliminate the differences in the absolute value range of static scores among different elderly individuals, the dynamic matching score of each elderly individual u across all candidate devices is normalized using a min-max method, mapping the score to the [0,1] interval to obtain the final dynamic matching score. ,in, and These are the minimum and maximum values ​​of the dynamic matching score of the elderly user across all the candidate smart home devices.

[0051] We continuously acquire daily health monitoring data. Each time a new week's record is added, we update the smoothed VO2 max sequence and A(u,t), then re-execute the formula calculation and normalization calculation to obtain the updated result. The update cycle is generally once a week.

[0052] In one embodiment of the present invention, step S4, which involves the predefined set of rigid needs of the elderly and the inter-device mutual exclusion constraints constructed based on the multi-dimensional attribute data of the candidate smart home devices, includes the following steps: The predefined set of rigid needs for the elderly consists of multiple needs dimensions, each of which corresponds to a basic functional guarantee required for the elderly to live independently and safely. Based on the functional categories in the multidimensional attribute data of the candidate smart home devices, a binary coverage matrix is ​​constructed, where each element represents whether a candidate smart home device meets one of the requirement dimensions in the set of rigid requirements in terms of functional specifications. For candidate smart home devices that meet the requirements in terms of function but whose final dynamic matching score is lower than the preset validity threshold, their corresponding elements are set to uncovered. Based on the installation constraints and technical parameters in the multi-dimensional attribute data of the candidate smart home devices, a mutual exclusion degree between the devices is constructed: if the exclusive identification of the installation location of two candidate smart home devices is exclusive and points to the same location, the mutual exclusion degree is taken as a first preset value; if the communication protocols of two candidate smart home devices use the same wireless frequency band and are marked as having interference risk, the mutual exclusion degree is taken as a value within a preset range; if the number of overlapping function tags of two candidate smart home devices exceeds a preset overlap threshold, the mutual exclusion degree is taken as a second preset value; otherwise, the mutual exclusion degree is zero.

[0053] Specifically, the set of essential needs for the elderly, D, consists of multiple dimensions, each corresponding to a basic functional guarantee required for the elderly to live independently and safely; a typical D = {fall detection and early warning, emergency call, medication reminders and management, automatic adjustment of ambient temperature and humidity, nighttime lighting guidance, simple voice interaction, and access control anomaly alarm}. Each dimension... ∈D describes a fundamental function that cannot be replaced by other dimensions. For example, "emergency call" requires the elderly to be able to actively initiate help through voice or keypad. They are parallel items and have no preset priority or weight.

[0054] Based on the functional categories in the multidimensional attribute data of the candidate smart home devices, a binary covering matrix F is constructed. Each row of this matrix corresponds to one candidate smart home device, and each column corresponds to a... ∈D, matrix element This indicates whether the device i actually provides coverage according to its functional specifications and the current condition of the elderly person. For the candidate smart home devices i and The judgment rule is as follows: In the pre-established mapping table between device function labels and requirement dimensions, check whether the third-level function label of i matches... If a corresponding relationship exists, then device i can meet the functional specifications. The corresponding element in the matrix is ​​set to 1; if no correspondence exists, it is set to 0. The mapping table between device function labels and requirement dimensions is established before the method is executed; for example, the third-level function label "millimeter-wave radar" is mapped to the requirement dimension "fall detection and early warning", and the third-level function label "voice emergency call device" is mapped to both the requirement dimensions "emergency call" and "easy voice interaction"; when a device's function category contains multiple third-level function labels, the device can cover multiple requirement dimensions at the same time.

[0055] The preset validity threshold is A typical value is 0.3. This value is based on the dynamic matching score, which, after min-max normalization, is distributed in the [0,1] interval. 0.3 corresponds to a level of moderate to high satisfaction of the elderly's needs by the device. This allows for the selection of devices with a certain degree of applicability while excluding devices with low matching scores or insufficient practical value. For functionally satisfied... Given device i, if the elderly person u's final dynamic matching score for that device is... This indicates that, given the elderly person's current physical condition, the device is actually unable to effectively meet this requirement, and its corresponding matrix element is corrected from 1 to 0, meaning it is set to uncovered. If If the functions match, the corresponding element remains 1.

[0056] Based on the installation constraints and technical parameters in the multi-dimensional attribute data of the candidate smart home devices, a mutual exclusion degree between the devices is constructed. The mutual exclusion degree represents the degree of physical conflict or functional redundancy that may occur if two candidate smart home devices are selected into a device combination at the same time. The higher the mutual exclusion degree, the greater the adverse effects of their coexistence. There are four scenarios for assigning the mutual exclusion degree.

[0057] The first scenario is physical mutual exclusion. If both candidate smart home devices have exclusive installation location identifiers, and both point to the same installation location (such as the same wall, ceiling, or desktop area) in the installation constraints, then they are physically mutually exclusive, and the mutual exclusion degree is set to a first preset value. The typical value of the first preset value is 1.0, indicating that the two devices cannot coexist in the same location.

[0058] The second scenario involves wireless interference. If two candidate smart home devices use the same wireless frequency band for communication and are marked as having an interference risk, the mutual exclusivity is determined based on this marking. The method for determining whether an interference risk exists is to objectively classify them using the absolute value of the center frequency interval between the two devices' operating frequency bands. Specifically, when the center frequency interval is less than or equal to 5MHz, it is marked as severe interference, with a mutual exclusivity of 0.7; when the center frequency interval is greater than 5MHz and less than or equal to 20MHz, it is marked as moderate interference, with a mutual exclusivity of 0.5; and when the center frequency interval is greater than 20MHz, it is marked as slight interference, with a mutual exclusivity of 0.3. If the device specifications do not specify the exact center frequency value, the channel interval specified by the communication protocol standard (e.g., the standard channel interval for the Wi-Fi 2.4GHz protocol is 5MHz) is used by default and substituted into the above rules for marking.

[0059] The third scenario is functional redundancy. If the number of overlapping function tags between two candidate smart home devices exceeds a preset overlap threshold, the mutual exclusion is set to a second preset value, typically 0.2. The preset overlap threshold is typically 70% of the total number of function tags. The number of overlapping function tags is calculated by taking the intersection of the third-level function tag set of device i and the third-level function tag set of device j. When the intersection exceeds 70% of the tag set size of each device, it is considered functional redundancy, and the mutual exclusion is set to 0.2. For example, if the third-level function tag sets of two devices are {millimeter-wave radar, human presence detection} and {infrared detector, human presence detection}, respectively, the intersection is 1, each device has 2 tags, and the overlap ratio is 50%, which does not exceed 70%, so it does not constitute functional redundancy, and the mutual exclusion is not set according to this scenario.

[0060] The fourth scenario is the remaining scenarios; for any two devices not belonging to the first three scenarios, the mutual exclusion degree is zero, indicating that there is no adverse effect from co-selection. According to the four rules, the mutual exclusion degree is calculated for each pair of all candidate smart home devices, resulting in a mutual exclusion matrix R, whose elements... This represents the mutual exclusion between device i and device j.

[0061] In one embodiment of the present invention, step S4, which aims to maximize the net coverage gain of demand, involves iteratively selecting from candidate smart home devices to generate a combination of devices free from functional redundancy and physical conflicts, and includes the following steps: Starting from an empty set S of selected smart home devices, an iterative selection process is executed; For each candidate smart home device that has not yet been selected, calculate its marginal net coverage gain after being added to the set S of selected smart home devices. The calculation formula is: Where C(S) is a subset of the set of rigid needs of the elderly covered by S. The number of the required dimensions newly covered after adding device i. Let η be the mutual exclusion degree between device i and device j, and η be the preset mutual exclusion penalty coefficient. In all satisfied Among the unselected smart home devices, select to use The largest equipment is added to S, of which To preset a global mutual exclusion tolerance threshold; When all unselected smart home devices The iteration stops at non-positive times, and it also stops when the total number of demand dimensions covered by S reaches the preset coverage target as a proportion of the total number of demand dimensions contained in the set of rigid demand dimensions for the elderly; after the iteration stops, S is a combination of devices without functional redundancy and physical conflict.

[0062] Specifically, before the iterative selection process begins, the set of selected smart home devices S is initialized to an empty set. For S, the set of demand dimensions it covers... This definition indicates that as long as there is at least one device in S that covers a certain demand dimension in the coverage matrix, it will be sufficient. Marked as actual coverage, this dimension belongs to C(S). Initially, S is empty, so C(S) is also empty.

[0063] Assume the current set of selected smart home devices is S, and the remaining devices that have not yet been selected are called candidate devices. For each device i, calculate its marginal net coverage gain. ,exist In the calculation formula, It is the new coverage that can be achieved after device i is added. The number of elements is calculated by iterating through the set of rigid needs of the elderly, D, and identifying each element not yet covered by C(S). ,like =1 increments the counter by one; mutual exclusion penalty coefficient A typical value is 0.8. 0.8 means that the mutual exclusion cost is counted as 80% of the actual mutual exclusion value in the gain. This value can both preserve devices with some mutual exclusion but still significantly expand coverage and effectively suppress devices with high mutual exclusion.

[0064] Calculate all devices Then, apply mutual exclusion constraints. Device i must satisfy... Among them, the global mutual exclusion tolerance threshold A typical value is 0.6, which allows for a moderate level of cumulative mutual exclusion between a single device and the selected set. A device with a mutual exclusion degree of 0.5 can still pass the constraint, but if two devices with mutual exclusion degrees of 0.4 each exist simultaneously, reaching a cumulative mutual exclusion degree of 0.8, subsequent devices will be excluded. This threshold strikes a balance between ensuring overall compatibility and avoiding excessive elimination of usable devices. Devices that do not meet this condition are directly excluded. Among the candidate devices that pass the mutual exclusion constraint, the one that allows... Largest equipment Add S, then update C(S); if multiple devices obtain the same maximum Then select the device with the smallest total mutual exclusion degree with the devices in S; if no candidate device passes the mutual exclusion constraint, or all candidate devices pass the constraint... If none of the values ​​are positive, no device is selected in this round, and the entire iterative selection process terminates.

[0065] The iterative selection process terminates in two scenarios: Scenario 1, no candidate device is selected in a given iteration; Scenario 2, the total number of requirement dimensions covered by S reaches its limit. Represents the total number of demand dimensions in D The proportion reaches the preset coverage target proportion ρ, that is Here, ρ typically takes the value of 0.95, indicating that iteration can stop when 95% of the demand dimensions have been covered. This value allows individual dimensions that cannot be effectively covered under the current conditions to be left unmet, avoiding the introduction of inefficient or highly mutually exclusive devices in pursuit of complete coverage. Once this condition is met, iteration will terminate even if some devices are still available. After iteration stops, the current S is the selected combination of devices, which has no functional redundancy or physical conflicts and has maximized the net demand coverage gain.

[0066] For example, the rigid needs dimension set D for an elderly person contains 7 needs dimensions, and there are 5 smart home devices to choose from, numbered as follows: to Typical values ​​for the parameters are: η = 0.8, θ = 0.6, ρ = 0.95. Initially, S is empty, and C(S) is empty. In the first iteration, since there are no devices in S, all mutual exclusion terms are zero, and the devices... It is equal to the number of demand dimensions it can cover; assuming Covering 3 , Covering 2, Covering 2, Covering 1 Covering 1, all devices satisfy mutual exclusion constraints; Select The largest Add S, at this time In the second iteration, the remaining devices are recalculated. to of ; Assumption Able to cover 1 new ,and If the mutual exclusion degree is 0.4, then =0.68; Able to cover 1 new ,and No mutual exclusion, =1; Able to cover 1 new But with The mutual exclusivity is 0.7, which exceeds θ=0.6, so it is excluded; Unable to cover any new content , NOT positive; Choice Add S, at this time The third iteration, It can still cover 1 new area The total mutual exclusion degree between S and S remains 0.4. =0.68; Still excluded; In this round, it becomes possible to cover 1 new item. And it is not mutually exclusive with S. =1; Select Add S, at this time The fourth iteration, One new coverage The total mutual exclusion degree is 0.4. =0.68; Add The S covers 6 The coverage ratio is 6 / 7≈0.857, which has not yet reached ρ=0.95; [Selection] Join, at this time The fifth iteration, only one remains. ; and The mutual exclusivity of 0.7 still exceeds θ, so it is excluded, and it cannot newly cover any... , Non-positive; no candidate device was selected in this round, the stopping condition is met, and the iteration terminates. Final device combination. It covers 6 .

[0067] After iteration stops, the final recommendation result is output. The device combination list lists the name and model of each device in S; for example, the millimeter-wave fall radar FD-200, the voice emergency call device VC-100, etc. Installation recommendations provide specific installation locations and methods based on the installation constraints in the multi-dimensional attribute data of each device; for example, the millimeter-wave fall radar is recommended to be ceiling-mounted in the center of the bedroom, and the voice emergency call device is recommended to be wall-mounted at the head of the bed, 1.2 meters from the ground. Compatibility specifications provide adjustment measures to reduce actual interference for mutually exclusive device pairs in S whose sum of mutual exclusion does not exceed θ; for example, if the millimeter-wave fall radar and the automatic night light both use Wi-Fi 2.4GHz and have a mutual exclusion of 0.5, it is recommended to set their wireless router channels to 1 and 11 respectively, or maintain a distance of at least 3 meters between them. Requirement coverage specifications provide for each device in D... Provide the coverage status for each dimension, indicating which device(s) in S satisfy the dimension. For example, "fall detection and early warning" is satisfied by millimeter-wave fall radar, and "emergency call" is satisfied by voice emergency call device. For the uncovered demand dimensions, explain the reason for their non-coverage. For example, the access control abnormal alarm is not covered because the dynamic matching score of the device with this function in the candidate device pool is lower than the validity threshold, or no device in the candidate device pool meets the functional specifications of this dimension.

[0068] 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 selection and recommendation of smart home devices for the elderly based on multi-feature fusion, characterized in that, Includes the following steps: S1: Obtain multidimensional characteristic data and daily health monitoring data of the elderly, as well as multidimensional attribute data of the selected smart home devices; among which, the multidimensional characteristic data of the elderly should at least include basic attributes, health indicators, behavioral abilities and environmental needs, and the multidimensional attribute data of the selected smart home devices should at least include functional categories, technical parameters and installation constraints. S2: Map the multidimensional feature data of the elderly and the multidimensional attribute data of the candidate smart home devices to the same hyperbolic space. Based on the geodesic distance between the vector corresponding to the elderly and the vector corresponding to the candidate smart home devices in the hyperbolic space, calculate the static matching score that reflects the degree of hierarchical matching between the two. S3: Based on the decreasing trend of maximum oxygen uptake extracted from the daily health monitoring data over time, an individualized physiological time decay factor is constructed; the individualized physiological time decay factor is used to perform time-series correction and normalization on the static matching score to obtain the final dynamic matching score. S4: Based on the final dynamic matching score, combined with the predefined set of rigid needs of the elderly and the mutual exclusion constraints between devices constructed according to the multi-dimensional attribute data of the candidate smart home devices, with the goal of maximizing the net coverage gain of the demand, the device combination without functional redundancy and physical conflict is iteratively selected from the candidate smart home devices, and the device combination and its demand coverage description are output as the final recommendation result.

2. The intelligent selection and recommendation method for smart home devices for the elderly based on multi-feature fusion according to claim 1, characterized in that, S1 includes the following steps: The multidimensional feature data of the elderly is a cross-sectional feature vector composed of four sub-vectors: basic attributes, health indicators, behavioral abilities and environmental needs. The values ​​of each continuous health indicator are the values ​​after moving average and normalization. The daily health monitoring data of the elderly are stored independently in the form of time-series records with timestamps, and include at least two of the following continuous time series: resting heart rate, weight, body fat percentage, sleep duration and number of times to get up at night; The functional categories in the multi-dimensional attribute data of the candidate smart home devices are hierarchically coded using a three-level functional tree, and the installation constraints must include at least a binary identifier for the exclusivity of the device installation location.

3. The intelligent selection and recommendation method for smart home devices for the elderly based on multi-feature fusion according to claim 1, characterized in that, In step S2, mapping the multidimensional feature data of the elderly person and the multidimensional attribute data of the candidate smart home devices to the same hyperbolic space includes the following steps: Construct a heterogeneous graph containing elderly nodes, device nodes, and functional tag nodes. The functional tag nodes are taken from the three-level functional tree of the functional category and form a hierarchical connection according to the parent-child relationship. The elderly node is connected to the associated functional tag node according to the medical history tag in its basic attributes and the behavioral ability. The device node is connected to the corresponding three-level functional tag node according to its functional category. Perform graph convolution on the heterogeneous graph to embed each node into a hyperbolic space, which is a Lorentz manifold with constant negative cross-sectional curvature. The vector corresponding to the elderly person is obtained. and the vector corresponding to the candidate smart home devices , and All meet Moreover, among them Let be the 0th dimension component of the vector.

4. The intelligent selection and recommendation method for smart home devices for the elderly based on multi-feature fusion according to claim 3, characterized in that, In step S2, the calculation of the static matching score, which reflects the degree of hierarchical matching between the vector corresponding to the elderly person and the vector corresponding to the candidate smart home device in hyperbolic space, includes the following steps: In the Lorentz manifold The above calculates the vector corresponding to the elderly. Vector corresponding to the candidate smart home devices The shortest path length between them is taken as the geodesic distance in the hyperbolic space; Then calculate the static matching score. Where u represents the elderly person, i represents the smart home device to be selected, and Lorentz product .

5. The intelligent selection and recommendation method for smart home devices for the elderly based on multi-feature fusion according to claim 1, characterized in that, In step S3, constructing an individualized physiological time decay factor based on the decreasing trend of maximum oxygen uptake extracted from the daily health monitoring data over time includes the following steps: From the resting heart rate, weight, and body fat percentage data in the daily health monitoring, the estimated maximum oxygen uptake is calculated on a time-by-time basis using a submaximal estimation model, and a smoothed maximum oxygen uptake sequence for the current time t is obtained by moving average. ; Calculate the time derivative of the smoothed VO2 max sequence, and then compare the derivative with... The product of these factors is determined as the effective descent rate. ,in, τ represents the value of 1 if the condition inside the square brackets is true and 0 otherwise, where τ is the integration variable; The effective descent rate at the initial assessment time The integral up to the current time t is multiplied by the preset sensitivity parameter γ and then negativeened. This negative value is used as the input to the exponential function to construct the individualized physiological time decay factor. , where u represents an elderly person.

6. The intelligent selection and recommendation method for smart home devices for the elderly based on multi-feature fusion according to claim 5, characterized in that, In step S3, the step of using an individualized physiological time decay factor to perform temporal correction and normalization on the static matching score to obtain the final dynamic matching score includes the following steps: Dynamic matching score at current time t The calculation formula is: Where u represents the elderly person, i represents the smart home device to be selected, s(u,i) is the static matching score, and γ is the preset sensitivity parameter. For the smoothed maximum oxygen uptake sequence, This indicates that the value is 1 if the condition within the square brackets is true, and 0 otherwise. Let τ be the initial evaluation time, and τ be the integration variable; The dynamic matching score of each elderly person u across all the candidate smart home devices. The final dynamic matching score is obtained by normalization. .

7. The intelligent selection and recommendation method for smart home devices for the elderly based on multi-feature fusion according to claim 1, characterized in that, In step S4, the predefined set of rigid needs of the elderly and the inter-device mutual exclusion constraints constructed based on the multi-dimensional attribute data of the candidate smart home devices include the following steps: The predefined set of rigid needs for the elderly consists of multiple needs dimensions, each of which corresponds to a basic functional guarantee required for the elderly to live independently and safely. Based on the functional categories in the multidimensional attribute data of the candidate smart home devices, a binary coverage matrix is ​​constructed, where each element represents whether a candidate smart home device meets one of the requirement dimensions in the set of rigid requirements in terms of functional specifications. For candidate smart home devices that meet the requirements in terms of function but whose final dynamic matching score is lower than the preset validity threshold, their corresponding elements are set to uncovered. Based on the installation constraints and technical parameters in the multi-dimensional attribute data of the candidate smart home devices, a mutual exclusion degree between the devices is constructed: if the exclusive identification of the installation location of two candidate smart home devices is exclusive and points to the same location, the mutual exclusion degree is taken as a first preset value; if the communication protocols of two candidate smart home devices use the same wireless frequency band and are marked as having interference risk, the mutual exclusion degree is taken as a value within a preset range; if the number of overlapping function tags of two candidate smart home devices exceeds a preset overlap threshold, the mutual exclusion degree is taken as a second preset value; otherwise, the mutual exclusion degree is zero.

8. The intelligent selection and recommendation method for smart home devices for the elderly based on multi-feature fusion according to claim 7, characterized in that, In step S4, the step of iteratively selecting and generating a device combination without functional redundancy and physical conflicts from the candidate smart home devices with the goal of maximizing the net coverage gain of demand includes the following steps: Starting from an empty set S of selected smart home devices, an iterative selection process is executed; For each candidate smart home device that has not yet been selected, calculate its marginal net coverage gain after being added to the set S of selected smart home devices. The calculation formula is: Where C(S) is a subset of the set of rigid needs of the elderly covered by S. The number of the required dimensions newly covered after adding device i. Let η be the mutual exclusion degree between device i and device j, and η be the preset mutual exclusion penalty coefficient. In all satisfied Among the unselected smart home devices, select to use The largest equipment is added to S, of which To preset a global mutual exclusion tolerance threshold; When all unselected smart home devices The iteration stops at non-positive times, and it also stops when the total number of demand dimensions covered by S reaches the preset coverage target as a proportion of the total number of demand dimensions contained in the set of rigid demand dimensions for the elderly. After the iteration stops, S is a combination of devices without functional redundancy and physical conflict.

Images