Smart home system for identifying abnormal behaviors of old people

Abstract

The invention, which relates to the technical field of smart home and elderly behavior monitoring, discloses an elderly abnormal behavior identification smart home system comprising a collaborative sensing layer, a dynamic causal reasoning layer, a self-adaptive decision early warning layer, and an event trigger type data bus connecting the collaborative sensing layer, the dynamic causal reasoning layer and the self-adaptive decision early warning layer. According to the invention, behavior data and environmental parameters of the elderly are synchronously collected through a distributed sensor network, and a data island is broken through an event-triggered data bus; adapting a non-preset scene through incremental learning and time decay weight depending on a dynamic knowledge graph containing 300 + scene rules, and constructing environment-behavior association in combination with dual-threshold data fusion; a contextual dynamic threshold value and a three-level early warning mechanism are adopted, and a misinformation self-correction function is matched to optimize a judgment standard, so that the problems of environmental semantic deficiency and static baseline stiffness are solved, the scene generalization ability is improved, and accurate and humanized abnormal behavior monitoring is realized.

Description

Technical Field

[0001] This invention relates to the field of smart home and elderly behavior monitoring technology, specifically a smart home system for recognizing abnormal behavior in the elderly. Background Technology

[0002] With the accelerating aging of the population, the need for safety monitoring of elderly people living alone is becoming increasingly prominent. Smart home systems that recognize abnormal behaviors in the elderly have become a key technological means to ensure the safety of elderly people living at home. In existing technologies, such systems typically collect the movement characteristics of the elderly (such as gait, frequency of limb movements, and duration of stay) through devices such as cameras, millimeter-wave radar, and infrared sensors deployed indoors. Combined with preset behavioral baselines or threshold models, these systems can identify and issue warnings for abnormal behaviors such as falls, prolonged periods of stillness, and aimless wandering.

[0003] Currently, technologies for recognizing abnormal behavior in the elderly generally employ a "dual-module independent operation" architecture: the behavior recognition module extracts data such as human skeletal features and movement trajectories, compares them with a preset "normal behavior template," and determines an anomaly when the deviation exceeds a threshold; the environmental perception module collects environmental parameters through temperature and humidity sensors, light sensors, smoke sensors, etc., primarily used to trigger warnings for specific scenarios (such as fires, gas leaks, etc.). The two modules lack deep collaboration in data transmission, analysis, and decision-making, forming a typical "data silo" architecture. This architectural flaw directly prevents the system from establishing a correlation mechanism between "dynamic environmental changes and semantic interpretation of behavior."

[0004] This modular architecture leads to a series of limitations in practical applications. Its core contradiction lies in the inability to dynamically calibrate the judgment criteria for behavioral semantics using environmental parameters, specifically manifested as follows:

[0005] Behavioral recognition suffers from semantic misinterpretation due to "lack of environmental context": behavioral adaptive adjustments caused by sudden environmental changes lose their interpretable basis. For example, when an elderly person fumbles for objects during a sudden power outage, the behavior is simply judged as aimless wandering because the light sensor data is not synchronized; when an elderly person frequently checks the window during a rainstorm, the behavior is misjudged as repetitive abnormal behavior because it is not associated with rain sensor data.

[0006] Static behavioral baselines are ill-suited to adapting to dynamic environmental changes: due to the lack of a causal relationship model between environmental parameters and behavioral motivations, the system judges anomalies solely based on action feature thresholds, ignoring the reversal effect of sudden environmental changes on behavioral semantics. For example, the behavior of an elderly person pacing to warm up when the indoor temperature suddenly drops, or the action of an elderly person rushing to ventilate when a smoke alarm falsely alarms, are both misjudged as abnormal because they are not linked to environmental data such as sudden temperature drops and sudden increases in smoke concentration.

[0007] Insufficient generalization ability for non-preset scenarios: Due to the lack of a universal environment-behavior correlation framework, the system cannot cope with sudden environmental changes such as heating failures, abnormal pet behavior, and burst water pipes. It also cannot dynamically update the contextual weights of behavior judgments, resulting in a persistently high false alarm rate. This not only increases the burden of ineffective intervention for caregivers but also may lead to missed opportunities for optimal rescue due to misjudgments of critical anomalies.

[0008] In summary, the core technological bottleneck of existing smart home systems for recognizing abnormal behavior in the elderly lies in the "data silo" architecture of the behavior recognition module and the environmental perception module. This prevents the system from addressing the fundamental issue of "semantic reversal of behavior caused by sudden environmental changes," making it difficult to meet the precise and humanized monitoring needs of elderly people living alone in complex and dynamic environments. Therefore, there is an urgent need to develop an anomaly recognition technology solution that can achieve deep correlation between behavior and environment, fundamentally breaking through the limitations of the existing modular and fragmented approach.

[0009] Therefore, a smart home system for recognizing abnormal behavior in the elderly is provided to overcome the above problems. Summary of the Invention

[0010] The purpose of this invention is to provide a smart home system for recognizing abnormal behavior in the elderly, so as to solve the problems mentioned in the background art.

[0011] To address the aforementioned technical problems, this invention provides a smart home system for recognizing abnormal behavior in the elderly, comprising: a collaborative perception layer, a dynamic causal reasoning layer, an adaptive decision-making and early warning layer, and an event-triggered data bus connecting the collaborative perception layer, the dynamic causal reasoning layer, and the adaptive decision-making and early warning layer;

[0012] The collaborative sensing layer is used to synchronously collect behavioral data and indoor environmental parameters of the elderly, and to perform time-stamped association tagging on the behavioral data and environmental parameters;

[0013] The dynamic causal reasoning layer is used to establish a real-time correlation model between environmental parameters and behavioral semantics, and to construct a dynamic knowledge graph containing three-element nodes of environmental parameters, behavioral features and semantic labels.

[0014] The adaptive decision warning layer is used to determine abnormal behavior with context based on a dynamic correlation model and to dynamically adjust the abnormal judgment threshold.

[0015] The event-triggered data bus is used to enable real-time data interaction between the collaborative perception layer, the dynamic causal reasoning layer, and the adaptive decision-making and early warning layer, ensuring that each frame of behavioral data is bound to a synchronized snapshot of environmental parameters.

[0016] Furthermore, the behavioral data acquisition of the collaborative perception layer adopts a distributed contact point sensor network, including a flexible thin-film pressure sensor array integrated into furniture and environmental carriers that the elderly frequently come into contact with daily. The flexible thin-film pressure sensor array forms a behavioral feature triangular network composed of 8 key sensing points. The furniture includes sofa cushions, mattress surfaces, dining table edges, and headboards, while the environmental carriers include the inside of door frames, bathroom handrails, and the edges of kitchen countertops. The sampling frequency for behavioral data acquisition is 50Hz, and one behavioral data frame is generated every 20ms.

[0017] Furthermore, the environmental data acquisition of the collaborative perception layer adopts a multimodal scene perception node network. Each perception node integrates a temperature and humidity sensor, a light intensity sensor, an environmental noise sensor, a raindrop vibration sensor, a door and window micro-vibration sensor, and an air refractive index sensor. The nodes adopt ZigBee wireless networking, and the distance between nodes does not exceed 3 meters. Each perception node has a built-in environmental change detection submodule. When the rate of change of environmental parameters exceeds a preset threshold, it triggers an environmental event marker and sends a synchronization signal through an event-triggered data bus.

[0018] Furthermore, the dynamic causal knowledge graph of the dynamic causal reasoning layer initially contains 300+ common scenario rules and 200+ furniture association rules. The knowledge graph adopts an incremental causal learning algorithm. When an unpreset scenario is detected, it automatically marks potential causal pairs between environmental parameters and behavioral characteristics. After manual confirmation, the graph weights are updated, and similar events more than 3 times are automatically solidified into new rules. The knowledge graph also adopts a time decay weight mechanism, and the weights of recently frequently occurring environment-behavior associations are automatically increased.

[0019] Furthermore, the dynamic causal inference layer also includes an event-triggered data fusion algorithm. The algorithm adopts a dual threshold triggering mechanism of environmental mutation and behavioral response: when the rate of change of environmental parameters exceeds the first threshold or the behavioral characteristics deviate from the baseline by more than the second threshold, cross-modal data fusion is initiated. The fusion process adopts a contact-trajectory dual anchoring strategy, using the environmental mutation event as the anchor point, backtracking the behavioral data of the previous 3 seconds and continuously monitoring the behavioral data of the next 10 seconds to construct the behavioral sequence before, during and after the environmental mutation.

[0020] Furthermore, the contextualized anomaly detection model of the adaptive decision-making early warning layer adopts a dynamic threshold adjustment mechanism.

[0021] Furthermore, the adaptive decision-making early warning layer includes a multi-level early warning mechanism, with three levels of warnings: Level 1 warnings are for low-risk behaviors requiring attention, recorded only locally; Level 2 warnings are for medium-risk suspected abnormalities, pushed to the guardian's APP; and Level 3 warnings are for high-risk confirmed abnormalities, triggering audible and visual alarms and emergency contact. The adaptive decision-making early warning layer also includes a false alarm self-correction module. After the guardian marks a Level 2 warning as a false alarm, the system automatically extracts scene features to optimize the knowledge graph weights, and updates the threshold parameters of the corresponding rules after three false alarms.

[0022] Furthermore, the flexible thin-film pressure sensing array is a 5×5 array of piezoresistive sensors with a sensitivity of 1g pressure detection threshold, a thickness of 0.5mm, and a size of 10×10cm; the strip pressure sensors installed on the inside of the door frame, the bathroom handrail, and the edge of the kitchen countertop are 30cm long; the collaborative sensing layer also includes a miniature millimeter-wave radar installed in the center of the indoor ceiling with a transmission power of <10mW, used for non-contact capture of the human torso movement trajectory.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] I. Addressing semantic misinterpretations caused by a lack of context:

[0025] 1. Synchronous Association Mechanism of the Collaborative Perception Layer: Through the synchronous acquisition of data from a distributed contact point sensor network and multimodal environmental perception nodes, combined with timestamped association tags, it is ensured that each frame of behavioral data (such as getting up, moving, touching furniture, etc.) is bound to a synchronous snapshot of environmental parameters (temperature, humidity, light, noise, etc.). For example, during a sudden power outage, the environmental event of a sudden drop in light intensity is correlated in real time with the behavioral data of concentrated hand pressure, providing key context for subsequent semantic interpretation and avoiding misjudging groping for objects as "aimless wandering".

[0026] 2. Event-triggered data fusion algorithm: Employing a dual-threshold triggering mechanism of "environmental mutation - behavioral response," it uses an environmental mutation event as an anchor point to backtrack behavioral data for the previous 3 seconds and monitor the following 10 seconds to construct a complete behavioral sequence. For example, during a sudden rainstorm, the sequence characteristics of "raindrop vibration sensor trigger → window pressure stay + light hand touch" are combined with the "rainfall - window viewing" rule in the knowledge graph to determine frequent window viewing behavior as normal, avoiding misjudgment as "repetitive anomalies."

[0027] II. Solving the problem that static behavioral baselines are difficult to adapt to dynamic environmental changes:

[0028] 1. Contextualized Dynamic Threshold Adjustment Mechanism: The adaptive decision-making and early warning layer adjusts the anomaly judgment threshold in real time based on environmental parameters. The standard is dynamically adapted using the formula "Dynamic Threshold = Basic Threshold × (1 + Environmental Parameter Influence Coefficient × Contextual Weight)". For example, when the indoor temperature drops by more than 5°C, the system automatically reduces the strictness of the pacing frequency threshold (from 20 times / minute to 30 times / minute), excluding adaptive behaviors such as "pacing for warmth" from anomaly judgment and solving the problem of static baseline rigidity.

[0029] 2. Real-time optimization of environmental parameter influence coefficients: Combining the contextual weights of the dynamic causal knowledge graph, judgment criteria are dynamically assigned for different environmental scenarios. For example, the behavior of "frequent turning over at night" combined with the environmental data of "sudden drop in room temperature" is judged as "cold discomfort" rather than abnormal, improving the humanization and accuracy of anomaly identification.

[0030] III. Addressing the issue of insufficient generalization capability in non-preset scenarios:

[0031] 1. Incremental learning capability of dynamic causal knowledge graph: Initially, it has 300+ common scenario rules and 200+ furniture association rules. Through incremental causal learning algorithm, it automatically marks potential environment-behavior causal pairs for unpreset scenarios (such as burst water pipes, abnormal pets, etc.). After manual confirmation of similar events more than 3 times, it is solidified into new rules, realizing the system's autonomous adaptation to unknown scenarios.

[0032] 2. Time decay and dynamic weight optimization mechanism: The knowledge graph adopts a time decay weight strategy. The weight of frequently occurring environmental-behavioral associations (such as "rain vibration - staying by the window" during the rainy season) is automatically increased to ensure the system's adaptability to seasonal and phased behavioral patterns. At the same time, through the false alarm self-correction module, the corresponding rule threshold is automatically updated after the guardian marks 3 false alarms, continuously optimizing the accuracy of judgment in non-preset scenarios.

[0033] IV. Other core advantages:

[0034] 1. Enhanced privacy protection and deployment compatibility: The use of a distributed contact point sensor network (flexible thin film pressure sensor, strip pressure sensor) combined with non-contact millimeter-wave radar avoids the risk of privacy leakage from optical imaging devices; the sensor installation only requires adhesive or screw fixing, which is highly compatible with existing homes and does not require large-scale modification.

[0035] 2. Multi-level early warning and efficient resource utilization: Through a multi-level early warning mechanism of Level 1 (local recording), Level 2 (pushing to guardians), and Level 3 (audio-visual alarm + emergency contact), risk levels are accurately classified; the false alarm self-correction module optimizes the knowledge graph weight through guardian feedback, reduces the burden of ineffective intervention, and ensures that key anomalies (such as falls or prolonged stillness) are responded to in a timely manner. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of a smart home system for recognizing abnormal behavior in the elderly, as per the present invention. Detailed Implementation

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0038] Please see Figure 1 , the present invention provides a technical solution:

[0039] See Figure 1 As shown, an embodiment of a smart home system for recognizing abnormal behavior in the elderly is provided:

[0040] I. System Architecture:

[0041] This system breaks through the traditional "dual-module independent operation" architecture and innovatively constructs a "behavior-environment real-time coupled perception and dynamic semantic reasoning" architecture. The core of this architecture is a three-layer linkage mechanism: the bottom layer, the collaborative perception layer, is responsible for the synchronous collection and association labeling of behavior and environmental data; the middle layer, the dynamic causal reasoning layer, establishes a real-time association model between environmental parameters and behavioral semantics; and the top layer, the adaptive decision-making and early warning layer, realizes context-based anomaly detection and dynamic threshold adjustment. These three layers achieve real-time data interaction through a self-developed "event-triggered data bus." The bus uses a timestamped association labeling protocol to ensure that each frame of behavior data is bound to a synchronized environmental parameter snapshot, breaking down "data silos" at the architectural level.

[0042] II. Collaborative Sensing Layer: Synchronous Data Acquisition

[0043] 1. Behavioral data acquisition: Distributed contact point sensor network:

[0044] A hybrid data acquisition strategy combining "key contact points + non-contact supplementation" is employed. The core technology is a flexible thin-film pressure sensor array, which is integrated into furniture and environmental surfaces that elderly people frequently come into contact with daily.

[0045] By employing a contact point distribution strategy, eight key sensor points are formed into a "behavioral feature triangular network." This network, combined with a spatiotemporal correlation algorithm, fuses multi-source data—for example, by following a time-series chain of "sofa pressure disappears (getting up) → door frame sensor triggers (movement) → dining table pressure appears (sitting down)"—to accurately reconstruct the activity path, generating a behavioral data frame every 20ms. Installation requires only adhesive or screw mounting of the sensors, offering high compatibility with already renovated homes and avoiding privacy issues associated with optical imaging.

[0046] It should be noted here that:

[0047] formula:

[0048] f 采样≥f 行为特征max ;

[0049] in:

[0050] f 采样 f is the sampling frequency for behavioral data (50Hz, i.e., 20ms / frame). 行为特征max The frequency of the most significant characteristic change in typical behavior of the elderly;

[0051] Statistically, the highest frequency of characteristic changes in daily activities of the elderly, such as getting up and moving, is 35Hz (e.g., the rate of pressure change when getting up quickly). Since 50Hz ≥ 35Hz satisfies the Nyquist sampling theorem, ensuring that the behavioral path such as the time chain of "sofa → door frame → dining table" can be accurately reconstructed, a 20ms sampling interval is reasonable.

[0052] Furniture Integration: Flexible pressure sensing modules (only 0.5mm thick) of 10×10cm are embedded in sofa cushions, mattress surfaces, dining table edges, headboards, and other locations. The modules contain a 5×5 array of piezoresistive sensors with a sensitivity reaching a 1g pressure detection threshold. When elderly individuals sit, lie down, lean against, or hold onto something, the sensors capture changes in pressure distribution in real time and extract six static behavioral characteristics, including "sitting duration, standing force, and leaning stability," through a contact posture analysis algorithm.

[0053] Environmental carrier integration: Strip pressure sensors (30cm in length) are installed on the inside of door frames, bathroom handrails, and the edges of kitchen countertops. When elderly people pass by or come into contact with these sensors, dynamic characteristics such as "frequency of passage, contact force, and duration of stay" are recorded. Simultaneously, a miniature millimeter-wave radar (transmitting power <10mW, meeting safety standards) is installed in the center of the indoor ceiling. Using micro-Doppler effect analysis technology, it non-contactly captures the trajectory of the human torso, supplementing behavioral data in areas not covered by contact points.

[0054] 2. Environmental Data Acquisition: Multimodal Scene Perception Node Network:

[0055] The distributed scene sensing nodes integrate six types of sensors: temperature and humidity, light intensity, environmental noise, raindrop vibration, door and window micro-vibration, and air refractive index (for indirect detection of smoke / water vapor). They adopt ZigBee wireless networking, with the node spacing not exceeding 3 meters to ensure no blind spots in indoor sensing.

[0056] Each sensing node uses a dual fixing method of 3M adhesive and magnetic attachment, allowing direct adhesion to walls, window frames, and appliance surfaces, reducing installation time to 5 minutes per node. Each node incorporates an environmental change detection submodule and features a preset scene-adaptive threshold adjustment function. For example, it automatically lowers the light change detection threshold in bedroom areas. When the rate of change of a certain environmental parameter exceeds a preset threshold, an "environmental event marker" is immediately triggered, sending a synchronization signal to the behavior acquisition module via the data bus to ensure accurate binding of behavior data frames to environmental change events. For instance, during a sudden power outage, the light sensor triggers an "environmental change - sudden drop in light" marker, simultaneously binding the pressure sensor data at that moment, providing crucial context for subsequent semantic interpretation.

[0057] III. Dynamic Causal Inference Layer: Environment-Behavior Association Mechanism

[0058] 1. Construction of dynamic causal knowledge graph:

[0059] A dynamic knowledge graph is constructed, comprising three-element nodes: "environmental parameters - behavioral features - semantic tags," with nodes connected by context-weighted edges. Unlike traditional static knowledge graphs, this graph possesses real-time learning capabilities.

[0060] The initial map contains 300+ common scene rules (such as "sudden drop in light + concentrated hand pressure → groping behavior" and "sudden drop in temperature + shortened gait cycle → pacing to warm up"), and a preset contact point behavior rule library, including 200+ furniture association rules such as "getting up from the sofa + increased pressure on the armrest → unstable standing warning" and "staying on the kitchen counter for more than 10 minutes + no pressure change → abnormal stillness".

[0061] An incremental causal learning algorithm is introduced. When the system detects an unpreset scenario (such as a sudden increase in ground humidity due to a burst water pipe and rapid movement), it automatically marks "sudden increase in humidity" and "rapid movement" as potential causal pairs. After manual confirmation, the graph weights are updated, and three or more similar events are automatically solidified as new rules.

[0062] The knowledge graph uses a time decay weighting mechanism, which automatically increases the weight of frequently occurring environment-behavior associations (such as "rain vibration + standing by the window" during the plum rain season) to solve the problem of insufficient generalization ability in non-preset scenarios.

[0063] It should be noted here that:

[0064] formula:

[0065] w 新 =w 历史 ×αt+Δw×β;

[0066] in:

[0067] w 历史 Historical weighting;

[0068] α is the time decay coefficient;

[0069] t is the time interval;

[0070] Δw represents the weight of the newly added event;

[0071] β is the learning rate;

[0072] Taking the correlation between "rainfall vibration and lingering by the window" during the plum rain season as an example:

[0073] ① Initial w 历史 =0.3; 3 more similar events occur after 10 days, α=0.95 (decaying by 5% daily), t=10, Δw=0.2, β=0.6, then:

[0074] w 新 =0.3 × 0.95 10 +0.2×0.6≈0.3×0.5987+0.12≈0.3;

[0075] ② If the three events occur consecutively within three days, and t = 3, then:

[0076] w 新 =0.3 × 0.95 3 +0.2×0.6≈0.3×0.857+0.12≈0.377;

[0077] The significant increase in weight proves that the recent high-frequency event weighting mechanism is effective.

[0078] 2. Event-triggered data fusion algorithm:

[0079] A dual-threshold triggering mechanism of "environmental mutation - behavioral response" is designed: when the rate of change of environmental parameters exceeds the threshold (e.g., temperature drops by >8°C in 5 minutes), or when behavioral characteristics deviate from the baseline (e.g., gait speed increases by 30%), cross-modal data fusion is immediately initiated. The fusion process is optimized to a contact-trajectory dual-anchoring strategy: using the environmental mutation event as the anchor point, behavioral data from the previous 3 seconds is retrospectively analyzed, and then continuously monitored for the following 10 seconds to construct a behavioral sequence of "before environmental mutation - during mutation - after mutation". When the door frame sensor triggers a "movement" event, the motion trajectory captured by the radar is automatically associated with the corresponding environmental parameters to construct a more complete behavioral context chain.

[0080] For example, when a sudden rainstorm hits, the raindrop vibration sensor triggers the "environmental change - rainfall" anchor point. The system automatically extracts the pressure sensing data before and after the anchor point. If the sequence feature of "pressure stay in the window area + light hand touch" is detected, combined with the causal rule of "rainfall - window viewing" in the knowledge graph, it is determined to be normal behavior, avoiding misjudgment as repetitive anomalies.

[0081] IV. Adaptive Decision Early Warning Layer: Context-Based Dynamic Thresholds

[0082] 1. Contextualized Anomaly Detection Model:

[0083] Instead of static thresholds, a context-weighted dynamic allocation algorithm is adopted to assign a judgment threshold that changes with the environment to each behavioral feature. The threshold calculation formula is: Dynamic threshold = Base threshold × (1 + Environmental parameter influence coefficient × Context weight). For contact point data features, a preset dynamic threshold for contact frequency and duration is established.

[0084] Specific implementation: When environmental parameters are normal, the system uses a basic threshold (e.g., pacing frequency > 20 times / minute is considered abnormal); when a "sudden temperature drop > 5℃" is detected, the environmental parameter influence coefficient is automatically adjusted to -0.4 (reducing the threshold strictness), and the pacing frequency threshold is relaxed to 30 times / minute to avoid misjudging heating-related pacing as abnormal. For example, "bedside sensor triggers > 5 times in 1 hour at night (frequent turning over)" combined with "sudden drop in room temperature" environmental data is judged as "discomfort behavior caused by cold" rather than abnormal; "pressure on the edge of the dining table remains unchanged for 30 minutes" combined with "no activity record in the kitchen" triggers a level two warning.

[0085] For non-preset scenarios (such as a pet spilling a water cup causing a localized slippery surface), the system automatically assigns a tolerance coefficient of 0.3 to the "unsteady gait" feature through the temporary rule of "sudden increase in local humidity - avoidance movement" in the knowledge graph, temporarily increasing the threshold.

[0086] It should be noted here that:

[0087] formula:

[0088] T 动态 =T 基础 ×(1+k 环境 ×w 语境 );

[0089] in:

[0090] T 基础 The preset base threshold;

[0091] k 环境 For environmental parameter influence coefficients;

[0092] w 语境 Context weight (0≤w) 语境 ≤1);

[0093] Taking the pacing frequency threshold as an example, T 基础 = 20 times / minute; when "sudden temperature drop > 5℃" is detected, the confidence level of the "sudden temperature drop - pacing to warm up" rule in the knowledge graph is 0.75, therefore w 语境=0.75; Due to the need to relax the judgment threshold due to this environmental change, the revised formula is:

[0094] T 动态 =T 基础 ×(1+k 环境 |×w 语境 );

[0095] in:

[0096] k 环境 =0.5 (a positive value indicates the amplification threshold), substituting into the calculation, we get:

[0097] T 动态 =20×(1+0.5×0.75)=20×1.375≈27.5≈30 times / minute; therefore, the dynamic threshold adjustment is reasonable.

[0098] 2. Multi-level early warning and false alarm self-correction mechanism:

[0099] The warning is divided into three levels: Level 1 (low risk) is "behavior requiring attention" (such as slow groping after a sudden drop in light), which is only recorded locally; Level 2 (medium risk) is "suspected abnormality" (such as prolonged stillness without environmental changes), which is pushed to the guardian's APP; Level 3 (high risk) is "confirmed abnormality" (such as pressure distribution matching the characteristics of a fall + no environmental changes), which immediately triggers an audible and visual alarm and emergency contact.

[0100] The system includes a false alarm feedback learning module and a preset contact point weight adjustment function. After a guardian marks a level 2 warning as a "false alarm," the system automatically extracts the environmental and behavioral features of the scene, reverse-optimizes the knowledge graph weights, automatically adjusts the feature weights of the corresponding furniture sensor points, and automatically lowers the warning level for the next similar scene. After 3 false alarms, the system updates the threshold parameters of the corresponding rules.

[0101] V. Core System Operation Process:

[0102] Environmental change scenario handling process: Taking a sudden power outage as an example, the light sensor triggers the "environmental change - sudden drop in light" flag → the data bus synchronously binds the pressure sensor data at this time → the dynamic causal inference layer calls the "sudden drop in light - groping behavior" rule → the adaptive decision layer increases the "aimless wandering" judgment threshold by 40% → if the pressure data shows concentrated hand pressure + slow movement trajectory, it is judged as normal groping behavior and no warning is triggered.

[0103] Static baseline adaptation process: When the indoor temperature drops suddenly by 8°C, the environmental node sends a "temperature drop" event → the dynamic causal inference layer activates the "temperature-walking for warmth" causal pair → the adaptive decision layer automatically adjusts the pacing frequency threshold (from 20 times / minute to 35 times / minute) → when pacing behavior is detected, it is determined to be a normal physiological adaptation behavior because it does not exceed the dynamic threshold.

[0104] Unpreset scenario response process: A burst water pipe causes a sudden increase in ground humidity → an environmental node triggers an "unpreset environmental change" flag → the dynamic causal reasoning layer associates the "sudden increase in humidity" with the "rapid movement trajectory" of the pressure sensor → temporarily assigns a "risk avoidance behavior" semantic label → the adaptive decision layer increases the movement speed threshold → if the movement trajectory points to the bathroom (where water tools are stored), combined with historical behavior data, it is determined to be a normal handling behavior and no false alarm is triggered.

[0105] Summarize:

[0106] Distributed contact point sensor network: It solves the problems of privacy leakage and light dependence of traditional devices. By "distributing key contact points + fusion of multi-source data", it breaks the traditional single device acquisition mode and achieves high-sensitivity behavioral feature acquisition.

[0107] Event-triggered data bus: By associating tags with timestamps, it fundamentally breaks down "data silos" and ensures that behavioral data is always bound to the context.

[0108] Dynamic causal knowledge graph: It uses incremental learning and weight decay mechanisms, combined with a touchpoint behavior rule base, to solve the problem of insufficient generalization ability in non-preset scenarios;

[0109] Contextualized dynamic thresholds: The judgment criteria are adjusted in real time by environmental parameters, and the dynamic thresholds for contact frequency and duration are preset to solve the shortcomings of static baselines that cannot adapt to environmental changes.

[0110] The system is easy to install, and all its technologies are based on mature pressure sensing, wireless networking, and machine learning technologies. Its core innovation lies in the ingenious coupling and correlation mechanism design of these technologies. Ordinary electronics factories can mass-produce it with slight adjustments to their existing production lines. It can effectively solve existing technical bottlenecks such as lack of environmental semantics, rigid static baselines, and insufficient generalization ability.

Claims

1. A smart home system for recognizing abnormal behavior in the elderly, characterized in that, include: The system consists of a collaborative perception layer, a dynamic causal reasoning layer, an adaptive decision-making and early warning layer, and an event-triggered data bus that connects the collaborative perception layer, the dynamic causal reasoning layer, and the adaptive decision-making and early warning layer. The collaborative sensing layer is used to synchronously collect behavioral data and indoor environmental parameters of the elderly, and to perform time-stamped association tagging on the behavioral data and environmental parameters; The dynamic causal reasoning layer is used to establish a real-time correlation model between environmental parameters and behavioral semantics, and to construct a dynamic knowledge graph containing three-element nodes of environmental parameters, behavioral features and semantic labels. The adaptive decision warning layer is used to determine abnormal behavior with context based on a dynamic correlation model and to dynamically adjust the abnormal judgment threshold. The event-triggered data bus is used to enable real-time data interaction between the collaborative perception layer, the dynamic causal reasoning layer, and the adaptive decision-making and early warning layer, ensuring that each frame of behavioral data is bound to a synchronized snapshot of environmental parameters.

2. The smart home system for recognizing abnormal behavior in the elderly as described in claim 1, characterized in that: The behavioral data acquisition of the collaborative perception layer adopts a distributed contact point sensor network, including a flexible thin-film pressure sensor array integrated into furniture and environmental carriers that the elderly frequently come into contact with in their daily lives. The flexible thin-film pressure sensor array forms a behavioral feature triangular network composed of 8 key sensing points. The furniture includes sofa cushions, mattress surfaces, dining table edges, and headboards. The environmental carriers include the inside of door frames, bathroom handrails, and kitchen countertop edges. The sampling frequency of behavioral data acquisition is 50Hz, and one behavioral data frame is generated every 20ms.

3. The smart home system for recognizing abnormal behavior in the elderly as described in claim 1, characterized in that: The environmental data acquisition of the collaborative perception layer adopts a multimodal scene perception node network. Each perception node integrates a temperature and humidity sensor, a light intensity sensor, an environmental noise sensor, a raindrop vibration sensor, a door and window micro-vibration sensor, and an air refractive index sensor. The nodes adopt ZigBee wireless networking, and the distance between nodes does not exceed 3 meters. Each perception node has a built-in environmental change detection submodule. When the rate of change of environmental parameters exceeds a preset threshold, it triggers an environmental event marker and sends a synchronization signal through an event-triggered data bus.

4. The smart home system for recognizing abnormal behavior in the elderly as described in claim 1, characterized in that: The dynamic causal knowledge graph of the dynamic causal reasoning layer initially contains 300+ common scenario rules and 200+ furniture association rules. The knowledge graph adopts an incremental causal learning algorithm. When an unpreset scenario is detected, it automatically marks potential causal pairs between environmental parameters and behavioral characteristics. After manual confirmation, the graph weights are updated, and similar events more than 3 times are automatically solidified into new rules. The knowledge graph also adopts a time decay weight mechanism, and the weight of environmental-behavior associations that have appeared frequently recently is automatically increased.

5. The smart home system for recognizing abnormal behavior in the elderly as described in claim 1, characterized in that: The dynamic causal inference layer also includes an event-triggered data fusion algorithm. The algorithm adopts a dual threshold triggering mechanism of environmental mutation and behavioral response: when the rate of change of environmental parameters exceeds the first threshold or the behavioral characteristics deviate from the baseline by more than the second threshold, cross-modal data fusion is initiated. The fusion process adopts a contact-trajectory dual anchoring strategy, using the environmental mutation event as the anchor point, backtracking the behavioral data of the previous 3 seconds and continuously monitoring the behavioral data of the next 10 seconds to construct the behavioral sequence before, during and after the environmental mutation.

6. The smart home system for recognizing abnormal behavior in the elderly as described in claim 1, characterized in that: The contextualized anomaly detection model of the adaptive decision-making early warning layer adopts a dynamic threshold adjustment mechanism.

7. The smart home system for recognizing abnormal behavior in the elderly as described in claim 1, characterized in that: The adaptive decision-making early warning layer includes a multi-level early warning mechanism, with three levels of early warning: Level 1 warning is for low-risk behaviors that require attention and is only recorded locally; Level 2 warning is for medium-risk suspected abnormalities and is pushed to the guardian's APP. A Level 3 warning indicates a high-risk confirmed anomaly, triggering audible and visual alarms and emergency contact. The adaptive decision-making early warning layer also includes a false alarm self-correction module. After the guardian marks the level 2 warning as a false alarm, the system automatically extracts scene features to optimize the knowledge graph weights, and updates the threshold parameters of the corresponding rules after 3 false alarms.

8. The smart home system for recognizing abnormal behavior in the elderly as described in claim 2, characterized in that: The flexible thin-film pressure sensing array is a 5×5 array of piezoresistive sensors with a sensitivity of 1g pressure detection threshold, a thickness of 0.5mm, and a size of 10×10cm; the strip pressure sensors installed on the inside of the door frame, bathroom handrail, and edge of the kitchen countertop are 30cm long; the collaborative sensing layer also includes a miniature millimeter-wave radar installed in the center of the indoor ceiling with a transmission power of <10mW, used for non-contact capture of human torso movement trajectory.

Images