A home security intelligent early warning device, system and method
By using multi-physics field collaborative sensing through electrostatic field, vibration, and airflow analysis, the problem of single sensors in existing home security systems has been solved, enabling all-weather stealth detection and proactive early warning, thus improving identification accuracy and protection effectiveness.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENZHEN HONGJINXIN DECORATION ENGINEERING CO LTD
- Filing Date
- 2026-03-25
- Publication Date
- 2026-05-29
Smart Images
Figure CN122116535A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent security technology, specifically relating to a home security intelligent early warning device, system and method. Background Technology
[0002] With social development and improved living standards, the demand for home security is increasing. Existing home security systems mainly employ traditional technologies such as video surveillance and infrared detection; however, they have the following shortcomings in practical applications: Firstly, existing security systems mostly rely on a single sensing mode, especially visual detection, which is easily affected by environmental factors such as lighting conditions and obstructions, leading to frequent missed or false alarms. Secondly, the sensors of traditional security equipment are usually exposed and easily discovered by intruders, who can then take evasive measures, thus reducing the actual protective effect of the security system. Third, existing systems lack the ability to deeply identify targets, making it difficult to distinguish between family members, visitors, and intruders, and thus unable to accurately determine behavioral intent. Fourth, most security systems rely on passive recording methods and lack proactive early warning and intelligent intervention mechanisms, making it impossible to effectively prevent danger before it occurs.
[0003] Therefore, there is an urgent need for a home security solution that can achieve invisible detection, 24 / 7 operation, intelligent identification, and proactive early warning. Summary of the Invention
[0004] To address the aforementioned technical problems, this invention provides a smart home security early warning device, system, and method based on multi-physics field collaborative sensing. By integrating three physical detection methods—electrostatic field sensing, structural vibration detection, and airflow disturbance analysis—it achieves comprehensive, invisible, and intelligent security protection for the home environment.
[0005] In a first aspect, the present invention provides a smart home security early warning device, characterized in that it comprises: The electrostatic field sensing module is used to detect changes in the spatial electrostatic field distribution through a distributed high-impedance antenna array, extract time-domain, frequency-domain, and statistical domain features, and generate an electrostatic field feature vector. The vibration detection module is used to collect vibration signals transmitted through the building structure using piezoelectric sensors, analyze time-frequency characteristics and energy distribution, and generate vibration feature vectors. The airflow analysis module is used to monitor indoor airflow field disturbances through a thermal flow sensor, invert airflow patterns based on fluid dynamics principles, and generate airflow feature vectors. The feature fusion module is connected to the three perception modules mentioned above. It is used to perform spatiotemporal alignment, normalization and adaptive weighted fusion of multi-physics field features to generate a multi-dimensional fusion feature vector. The target recognition module, connected to the feature fusion module, is used to perform identity matching and behavior analysis based on fused features to determine the target category. The risk assessment module, connected to the target identification module, is used to calculate the threat risk value by comprehensively considering identity, behavior, time, and environmental factors. The early warning control module, connected to the risk assessment module, is used to generate graded early warning instructions based on the risk level. The response execution module, connected to the early warning control module, is used to execute local alarms, remote notifications, and device linkage responses.
[0006] Furthermore, the antenna of the electrostatic field sensing module is embedded in the ceiling light fixture, the sensor of the vibration detection module is installed in the door and window frame, stairs and walls, and the sensor of the airflow analysis module is integrated into the air conditioner return air vent and smart switch.
[0007] Furthermore, the feature fusion module uses a network time protocol to achieve clock synchronization, and the fusion weights are dynamically adjusted based on the signal-to-noise ratio and signal quality.
[0008] Furthermore, the target recognition module performs identity matching using Mahalanobis distance and uses a Hidden Markov Model for behavior analysis.
[0009] Furthermore, the risk assessment module calculates the comprehensive risk value through weighted summation, and the early warning control module sets multi-level thresholds to achieve tiered early warning.
[0010] Secondly, the present invention provides a smart home security early warning system, characterized in that it includes: The sensing layer includes an electrostatic field sensing module, a vibration detection module, and an airflow analysis module, which are deployed in a distributed manner and transmit data through different wireless protocols; The transport layer includes a protocol gateway module and a data routing module, which implement multi-protocol conversion and priority routing; The processing layer includes a signal preprocessing module, a feature extraction module, and a data fusion module, which perform signal processing and evidence fusion. The analysis layer includes a target recognition module, a behavior analysis module, and a risk assessment module, which uses machine learning algorithms for intelligent analysis. The decision-making layer, including the early warning strategy module and the linkage control module, maintains the strategy rules and executes control commands. The application layer includes a user interaction module, a data management module, and a system configuration module, providing user interface and system management functions.
[0011] Furthermore, the perception layer communicates with the transport layer via WiFi, Zigbee, and Bluetooth protocols, while the transport layer uses JSON format for a unified data structure.
[0012] Furthermore, the processing layer uses DS evidence theory to fuse multi-source information, the analysis layer uses support vector machines for target recognition, and long short-term memory networks for behavior analysis.
[0013] Furthermore, the application layer adopts a hybrid storage architecture, and the system also includes a cloud service platform to provide model training and remote operation and maintenance services.
[0014] Thirdly, the present invention provides a smart early warning method for home security, characterized by comprising the following steps: S1: Multiphysics environment baseline establishment steps: collect electrostatic field, vibration and airflow signals, calculate statistical characteristics according to time period, and establish a baseline for normal activity mode; S2: Three-dimensional physical field signal synchronous acquisition steps, which acquire three physical field signals in parallel and achieve time synchronization through a precise time protocol; S3: Preprocessing steps for multi-source heterogeneous signals, employing filtering, denoising, and trend elimination methods for signal preprocessing respectively; S4: Multidimensional physical feature extraction step, extracting amplitude, frequency, energy and mode features; S5: Spatiotemporal alignment feature fusion step, unifying time resolution, eliminating dimensional differences, and dynamically adjusting fusion weights; S6: Dual-mechanism collaborative anomaly detection step, combining statistical detection and machine learning detection to determine anomalies; S7: Intelligent identity and behavior recognition steps, which identify identity through feature matching and judge behavior through time-series analysis; S8: Multi-factor risk assessment steps, calculating risk values by integrating multiple dimensions of factors; S9: Tiered early warning intelligent response steps, implementing differentiated response strategies based on risk level; S10: Online adaptive model update step, using incremental learning to optimize model parameters.
[0015] Beneficial effects:
[0016] 1. Significantly improved detection accuracy: Through the coordinated sensing and cross-verification of three physical fields—electrostatic field, vibration, and airflow—the detection blind spots of a single sensor are effectively eliminated, reducing the false alarm and missed alarm rates.
[0017] 2. Strong ability to be concealed and evaded: The sensors are integrated into everyday facilities such as lights, switches, and air conditioners, achieving complete concealment, making them undetectable and unavoidable for intruders.
[0018] 3. Intelligent recognition and behavior analysis: Accurately distinguish family members, visitors and intruders through machine learning algorithms, judge behavioral intentions, and avoid misjudging normal activities.
[0019] 4. Proactive early warning and intelligent intervention: Establish a risk assessment system to make predictions in the early stages of threat formation, and achieve proactive defense through tiered response and equipment linkage.
[0020] 5. 24 / 7 operation and privacy protection: It does not rely on lighting conditions and can work normally in dark, smoky, and other environments; it does not collect images or videos, thus fundamentally protecting user privacy. Attached Figure Description
[0021] Figure 1 A schematic diagram of the structure of the device described in this invention is shown; Figure 2 A schematic diagram of the system architecture described in this invention is shown; Figure 3 A flowchart illustrating the steps of the method described in this invention is shown. Detailed Implementation
[0022] Exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0023] Combination Figure 1 In a first aspect, the present invention provides a smart home security early warning device, comprising an electrostatic field sensing module, a vibration detection module, an airflow analysis module, a feature fusion module, a target recognition module, a risk assessment module, an early warning control module, and a response execution module. The electrostatic field sensing module, vibration detection module, and airflow analysis module operate in parallel, respectively collecting electrostatic field signals, vibration signals, and airflow signals from the home environment. These three signals are transmitted to a data aggregation gateway via their respective wireless communication protocols, and then simultaneously transmitted to the feature fusion module via an internal bus. The feature fusion module performs spatiotemporal alignment and feature-level fusion on the received multi-source heterogeneous signals, generating a fused feature vector which is output to the target recognition module. The target recognition module performs identity recognition and behavior classification based on the fused features, and the recognition result is transmitted to the risk assessment module. The risk assessment module comprehensively evaluates the risk and outputs a risk level to the early warning control module. The early warning control module generates tiered early warning commands, which are then executed by the response execution module.
[0024] The electrostatic field sensing module achieves non-contact personnel sensing based on the principle of human electrostatic field detection. During daily activities, the human body accumulates static charge due to friction from clothing and contact between the soles of shoes and the ground. The electric field distribution generated by these charges in space is individual-specific. The electrostatic field sensing module uses a distributed high-impedance antenna array as the sensing front end. Each antenna consists of a planar electrode made of conductive material, with an input impedance designed to be above 10MΩ. This high impedance characteristic ensures that the antenna can sense minute changes in the spatial electric field without causing a load effect on the electric field itself. The antennas are embedded inside ceiling light fixtures and deployed in four main activity areas: the living room, master bedroom, secondary bedroom, and study. The sensing ranges of each antenna overlap, forming a three-dimensional detection network without blind spots. The antennas are installed at a height of 2.2 to 2.5 meters above the ground, a height that effectively detects the human electrostatic field without causing signal saturation due to excessive distance.
[0025] The output signal of the antenna array is connected to the signal acquisition circuit via a shielded cable. Since the current signal induced by the electrostatic field is extremely weak, typically in the pA to nA range, the signal acquisition circuit employs a special low-noise design. The preamplifier is an ultra-low-noise operational amplifier with an input bias current of less than 1 pA, and the amplification gain is adjustable from 20 dB to 60 dB, with an appropriate gain value selected based on the actual signal strength. The amplified signal enters a bandpass filter, which is a fourth-order Butterworth filter with a passband range of 0.1 Hz to 100 Hz. This frequency range covers the frequency of electrostatic field changes generated by human activity and effectively suppresses 50 Hz power frequency interference and its harmonics. The filtered analog signal is converted into a digital signal by a 16-bit analog-to-digital converter at a 1 kHz sampling rate. The digital signal is then transmitted to a digital signal processor via an SPI interface for feature extraction.
[0026] The feature extraction process of the electrostatic field sensing module adopts a sliding window approach, with a window length of 1 second and a sliding step size of 0.5 seconds, ensuring a 50% overlap rate. For the signal within each time window, time-domain, frequency-domain, and statistical domain features are calculated separately. The time-domain features include the maximum field strength. Minimum value mean and standard deviation These fundamental statistics reflect the intensity level and fluctuation characteristics of the electrostatic field. Frequency domain characteristics were obtained through a 512-point Fast Fourier Transform to extract the dominant frequencies. and the ratio of main frequency energy The dominant frequency is defined as the frequency component with the highest energy in the spectrum, and the dominant frequency-energy ratio is calculated as the proportion of energy within ±0.5 Hz of the dominant frequency to the total energy. Statistical domain characteristics include the signal skewness. and kurtosis These reflect the asymmetry and sharpness of the signal distribution, respectively. Furthermore, the cross-correlation matrix of signals between different antennas is calculated. This is used to analyze the target's position and direction of movement in space. All features are combined to form the electrostatic field eigenvector: in: This represents the maximum value of the electrostatic field strength within the time window; This represents the minimum electrostatic field strength within the time window. This represents the average value of the electrostatic field strength. The standard deviation of the electrostatic field strength; The dominant frequency component; The proportion of the main frequency energy to the total energy; The skewness of the signal; The kurtosis of the signal; This is the cross-correlation matrix of signals between different antennas.
[0027] The vibration detection module utilizes the building structure as the vibration propagation medium, identifying various activities by detecting structural vibrations. Vibrations propagate rapidly and attenuate minimally in solid structures, carrying rich event information. The vibration detection module employs a piezoelectric accelerometer with a sensitivity of 100mV / g and a frequency response range of 0.5Hz to 10kHz. The sensor is installed in three key locations via a magnetic base: the first is at door and window frames, including entrance doors, balcony doors, living room windows, and bedroom windows, to detect opening and closing and prying actions; its output signal is connected to a charge amplifier via a coaxial cable. The second is at stair treads, installed below the first and fifth steps, to identify gait characteristics of people going up and down stairs, transmitting data wirelessly to a receiver. The third is at load-bearing walls, selected as the main load-bearing walls of the living room and bedrooms, to monitor drilling and hammering damage; the sensor is connected to a signal conditioning circuit via a wired connection.
[0028] The vibration signal conditioning process includes three stages: charge-to-voltage conversion, filtering, and gain adjustment. The charge amplifier converts the high-impedance charge signal output from the piezoelectric sensor into a low-impedance voltage signal, with a conversion sensitivity of 10mV / pC. The high-pass filter has a cutoff frequency of 0.5Hz to remove DC bias and low-frequency drift. The anti-aliasing low-pass filter has a cutoff frequency of 5kHz to prevent high-frequency noise from affecting subsequent digital processing. The programmable gain amplifier automatically adjusts its gain according to the vibration intensity, with a gain range of 1 to 100 times, ensuring the signal amplitude is within the optimal input range of the analog-to-digital converter. The conditioned signal is then digitized by a 16-bit precision, 16kHz sampling rate analog-to-digital converter.
[0029] The vibration characteristic analysis employs a time-frequency joint analysis method. First, the peak value of the vibration acceleration is calculated. and root mean square value The peak value reflects the maximum intensity of the vibration, while the root mean square value characterizes the average energy of the vibration. The peak factor is defined as: in: Peak factor; This represents the peak value of the vibration acceleration. This represents the root mean square value of the vibration acceleration.
[0030] The peak factor is used to distinguish between impact vibration and steady-state vibration; the peak factor for impact vibration is typically greater than 3. The time-varying spectrum is obtained through short-time Fourier transform and divided into four frequency bands: 0-50Hz, 50-200Hz, 200-1000Hz, and 1000-5000Hz. The energy distribution of each frequency band is then calculated. Energy center of gravity frequency Defined as: in: The energy center of gravity frequency; For the first One frequency component; This corresponds to the energy value.
[0031] effective bandwidth The dispersion of vibration frequency components is characterized by calculating the second-order central moment of the spectrum. Vibrations from normal activities such as walking and opening doors are mainly concentrated in the low-frequency range, while actions like picking locks and drilling into walls produce significant high-frequency components. The vibration eigenvector is represented as: The airflow analysis module senses object movement and human activity by monitoring the disturbance patterns of indoor airflow. In a stable state, indoor air forms specific convection patterns; any object movement will disturb the airflow field. The airflow analysis module uses a MEMS thermal mass flow sensor with a measurement range of -200 slm to +200 slm and a response time of less than 100 ms. The sensor directly outputs flow velocity values via an I2C digital interface, avoiding interference issues in analog signal transmission. The sensors are deployed in two ways: the first is integrated into the air conditioner return air vents, with two sensors located at the return air vents of the indoor air conditioner units in the living room and bedroom, respectively, utilizing the airflow circulation of the air conditioning system to monitor a wide range of airflow; the second is embedded in a smart switch panel, with four sensors distributed on the walls of each room at an installation height of 1.2 meters, close to the main height range of human activity.
[0032] Since changes in indoor temperature can affect the accuracy of airflow measurements, the airflow analysis module implements a temperature compensation algorithm. Each sensor has a built-in temperature measurement function to acquire the ambient temperature in real time. According to the ideal gas law, gas density is inversely proportional to temperature, and flow rate measurements need to be corrected for temperature. The temperature compensation formula is: in: The compensated flow rate value; These are the original measured values; This is the temperature compensation coefficient, with a value of 0.003 / ℃; The current temperature; The reference temperature is set to 25℃.
[0033] The compensated flow velocity data undergoes differential processing to eliminate slow changes in ambient airflow and extracts the rapid disturbance components caused by object movement. The airflow analysis module, based on computational fluid dynamics principles, inverts the airflow field distribution across the entire space using limited sensor measurements. A simplified CFD model of the indoor space is established, dividing the room into 1m×1m×1m grid cells, with each sensor location serving as a boundary condition input. The flow velocity vector at each grid point is estimated by solving the simplified Navier-Stokes equations. When abnormal disturbances are detected, a particle tracking algorithm is used to infer the location and trajectory of the disturbance source. The airflow disturbance velocity generated by human walking ranges from 0.1 to 0.5 m / s, while rapid running can reach 1 to 2 m / s; the movement speed can be inferred from the disturbance intensity.
[0034] The feature fusion module is a key component in achieving multi-physics collaborative sensing, responsible for integrating heterogeneous data from the three sensing modules. First, time alignment is performed. Although the sampling rates of the three sensors differ (1kHz for electrostatic field, 16kHz for vibration, and 10Hz for airflow), they all need to be unified to the same time base. The system uses a network time protocol to synchronize the clocks of each sensor node, achieving millisecond-level accuracy. Specifically, the gateway device acts as the master clock, and each sensor node acts as a slave clock, synchronizing via the IEEE 1588v2 precise time protocol. The synchronization process includes four message exchanges: the master clock sends a Sync message and records the sending time. Receive and record the reception time from the clock. The master clock sends a Follow_Up message containing the precise transmission time of the Sync message; the slave clock sends a Delay_Req message and records the transmission time. The master clock receives and records the reception time. Then, it replies with a Delay_Resp message. Using these four timestamps, it calculates the clock skew and path delay from the clock: in: This is due to clock skew; For path delay; , , , These are the timestamps of the four messages.
[0035] For airflow sensor data with a low sampling rate, cubic spline interpolation is used to upsample to the same temporal resolution as other sensors. Spatial alignment is achieved by mapping measurements from different spatial locations to a unified three-dimensional coordinate system using pre-calibrated sensor position coordinates. The origin of the coordinate system is set at the geometric center of the building, with the X-axis pointing east, the Y-axis pointing north, and the Z-axis pointing vertically upward.
[0036] Before feature fusion, different physical quantities need to be normalized to eliminate dimensional differences. The z-score standardization method is used to convert each feature into a standard distribution with a mean of 0 and a standard deviation of 1. in: These are the standardized eigenvalues; These are the original eigenvalues; The characteristic mean; The characteristic standard deviation is denoted as .
[0037] The fusion weights are dynamically adjusted based on the real-time status of each sensor. Confidence coefficient. Signal-to-noise ratio (SNR) is calculated based on the ratio of signal power to noise power, expressed in decibels (dB). in: Signal-to-noise ratio; Signal power; This represents noise power.
[0038] Signal quality indicators The system comprehensively considers three aspects: signal stability, continuity, and reasonableness. Stability is assessed by calculating the signal's variance coefficient; continuity is judged by the data loss rate; and reasonableness is measured by the deviation from historical data. When an abnormal jump or deviation from a reasonable range is detected in a sensor signal, its quality index is reduced. The weighting formula is as follows: in: For the first The fusion weights of the individual sensors; The confidence level coefficient ranges from 0 to 1. This is a signal quality indicator, ranging from 0 to 1; For the summation index, values 1, 2, and 3 correspond to the electrostatic field, vibration, and airflow sensors, respectively.
[0039] The final fused feature vector is obtained by weighted summation: in: The fused feature vector; The normalized electrostatic field eigenvector; The normalized vibration eigenvector; This is a normalized airflow feature vector; , , This represents the corresponding fusion weight.
[0040] The target recognition module receives the fused feature vector and performs processing at two levels: identity recognition and behavior recognition. Identity recognition is achieved by matching against a pre-established identity feature database. During the system initialization phase, each family member needs to collect multiple sets of feature samples at different times and in different states. The specific collection process is as follows: each person moves around in different rooms during three time periods: morning, noon, and evening, in three states: normal walking, fast walking, and standing still, respectively. Each state lasts for 2 minutes, for a total of 54 minutes of data collection. Representative feature templates are extracted using the K-means clustering algorithm, with the number of clusters set to 5, resulting in 5 cluster centers that serve as the identity feature templates for that person and are stored in the identity database.
[0041] During recognition, the Mahalanobis distance between the feature to be tested and each template is calculated: in: The Mahalanobis distance; The feature vector to be identified; As a reference feature vector; The covariance matrix of the training data; superscript Indicates matrix transpose; superscript This represents finding the inverse of a matrix.
[0042] When the minimum distance is less than the preset threshold If the visitor is identified as such, they are identified as a stranger; otherwise, they are marked as such. For registered visitors, the system temporarily stores their characteristic template, which is valid for 24 hours.
[0043] Behavior recognition employs a time-series analysis method based on Hidden Markov Models (HMMs). Behavior is modeled as a sequence of states, defining 10 basic states: stationary, slow walking, normal walking, fast walking, running, going upstairs, going downstairs, opening a door, closing a door, and abnormal action. Each state corresponds to a specific feature distribution. The observation sequence is a sequence of feature vectors spanning 20 consecutive time windows. Model parameters include the state transition probability matrix. Observation probability distribution and initial state probability These parameters were learned on labeled training data using the Baum-Welch algorithm. The training data contained 1000 labeled behavioral sequences, covering various normal and abnormal behavioral patterns.
[0044] The Viterbi algorithm is used to identify the most likely state sequence during the identification process. Let the observed sequence be... The state sequence is The Viterbi algorithm solves the problem using dynamic programming: in: For a moment In state The maximum probability; These are the parameters of the HMM model.
[0045] The risk assessment module comprehensively evaluates the threat level based on multi-dimensional information. A four-dimensional risk assessment model is established, considering four factors: identity, behavior, time, and environment. Identity risk is also addressed. Assign a value based on the recognition result: Behavioral risk Assessment based on behavior type: Normal behaviors include walking, sitting, and opening doors; suspicious behaviors include lingering for a long time, repeatedly probing, and running quickly; and clear intrusion behaviors include picking locks, breaking windows, and drilling through walls.
[0046] Time risk Calculated based on the current time period: Environmental risks Considering the current environmental conditions: The overall risk value is calculated by weighted summation: in: This is the overall risk value; , , , For the weighting coefficients, satisfying .
[0047] The early warning control module generates tiered early warning commands based on a comparison between risk values and preset thresholds. Three thresholds are set. , , The risks are divided into four levels: when At the normal level, the system remains in silent monitoring mode, only storing the data of the most recent 24 hours locally in a loop. when When the event is classified as suspicious, the system stores complete event data locally, including timestamps, feature vectors, and identification results. The data is retained for 30 days. when When the event is at the warning level, the system pushes a notification to the user's mobile phone, including the event type, location, and risk assessment results, and simultaneously pre-starts the audio and video recording equipment. when The situation is classified as dangerous, and the system immediately executes a full response.
[0048] The response execution module performs corresponding actions based on the warning command. Local audible and visual alarms control alarms installed in the living room and hallway, emitting a 110dB alarm sound and flashing lights. Remote notifications send preset SMS messages to three emergency contact phones via the GSM module and push detailed information to the user's app via the 4G network. Smart device linkage includes: automatically locking all smart door locks and escalating the locking level to the highest level; turning on all smart lighting in the house and adjusting the brightness to maximum; starting network camera recording, with the video stream simultaneously stored locally and in the cloud; closing electric curtains to prevent external observation; and playing warning voice messages through a smart speaker. All modules work collaboratively, with the entire process from detection to response time controlled within 500 milliseconds.
[0049] Combination Figure 2 Secondly, this invention provides a smart home security early warning system. The system adopts a six-layer architecture, including a perception layer, a transmission layer, a processing layer, an analysis layer, a decision-making layer, and an application layer. Each layer interacts with the others through standardized interfaces. Data flows upwards from the perception layer, while control commands are sent downwards from the decision-making layer, forming a bidirectional information flow path.
[0050] The sensing layer, serving as the system's data source, is responsible for real-time acquisition of multi-physics field signals from the home environment. This layer includes a distributed electrostatic field sensing module, a vibration detection module, and an airflow analysis module. The electrostatic field sensing module comprises multiple sensing nodes deployed in the ceiling locations of the living room, master bedroom, secondary bedroom, and study. Each node has an effective sensing radius of several meters, with overlapping sensing ranges between adjacent nodes to ensure continuous spatial coverage. The vibration detection module contains multiple sensing nodes deployed in key locations such as door and window frames, stair treads, and load-bearing walls, forming a structural vibration monitoring network. The airflow analysis module's sensing nodes are integrated into air conditioning return air vents and smart switch panels, enabling comprehensive monitoring of indoor airflow circulation.
[0051] Each sensing module uses a different wireless communication protocol to connect to the transmission layer. The electrostatic field sensing module uses the WiFi protocol, operates in a general frequency band, and supports high-speed data transmission. Each electrostatic field sensing node is equipped with an independent WiFi module, connecting to the home wireless network via encryption. The vibration detection module uses the Zigbee protocol, forming a self-organizing mesh topology. Each node can act as both a terminal device and a router, forwarding data to other nodes. When a node fails, the network can automatically reroute, improving system reliability. The airflow analysis module uses the Bluetooth Low Energy protocol, reducing power consumption while maintaining transmission distance.
[0052] The transport layer is responsible for protocol conversion and data routing, including a protocol gateway module and a data routing module. The protocol gateway module enables mutual conversion between different wireless protocols. The gateway adopts a multi-core processor architecture, is equipped with ample memory and storage space, and runs a Linux-based embedded operating system. The protocol conversion process includes: receiving data packets from the perception layer, parsing the packet header to identify the source protocol type, extracting the payload according to predefined mapping rules, and re-encapsulating it into a unified internal data format. The internal data format uses a JSON structure and includes fields such as timestamp, sensor identifier, data type, measurement value, signal quality, and priority.
[0053] The data routing module selects the transmission path based on the data type and priority. The system defines three priorities: high, medium, and low, corresponding to abnormal event data, normal monitoring data, and system status data, respectively. The routing strategy employs priority-based queue management: high-priority data enters the fast channel to ensure minimal transmission latency; medium-priority data enters the ordinary queue and is processed in order; low-priority data is transmitted when the network is idle. When network congestion is detected, the routing module initiates a flow control mechanism, using a token bucket algorithm to dynamically adjust the data transmission rate.
[0054] The processing layer preprocesses and extracts features from the received raw data, including a signal preprocessing module, a feature extraction module, and a data fusion module. The signal preprocessing module performs data quality assessment, checking the data's completeness, timeliness, and reasonableness. Cyclic redundancy check ensures error-free data transmission, verifies the validity of data timestamps, and confirms that data values are within a reasonable range. Missing data points are filled using interpolation methods; outliers are removed using median filtering; and noisy signals are processed using wavelet denoising.
[0055] The feature extraction module extracts multidimensional features for different types of signals. For electrostatic field signals, it extracts time-domain statistical features, frequency-domain energy features, and spatial distribution features; for vibration signals, it extracts time-domain peak features, frequency-domain spectral features, and time-frequency wavelet features; and for airflow signals, it extracts velocity amplitude features, direction vector features, and disturbance mode features. All features are normalized to eliminate dimensional differences.
[0056] The data fusion module employs Dempster's evidence theory to achieve decision-level fusion. Each sensor independently makes preliminary judgments and generates basic probability assignment functions. Multi-source evidence is fused using Dempster's combination rules. When conflicts exist between pieces of evidence, a similarity-weighted method is used for preprocessing to reduce conflicts before fusion. The fusion process considers the reliability and environmental adaptability of each sensor, dynamically adjusting the fusion strategy.
[0057] The analysis layer performs in-depth analysis based on the feature data provided by the processing layer, including a target recognition module, a behavior analysis module, and a risk assessment module. The target recognition module employs a support vector machine classifier, selects a radial basis function kernel function, and optimizes hyperparameters through grid search and cross-validation. The training data contains feature samples from multiple family members and typical visitors, covering different time periods and states.
[0058] The behavior analysis module uses a Long Short-Term Memory (LSTM) network for temporal pattern recognition. The network employs a multi-layer architecture containing multiple LSTM units, each controlling information flow through a gating mechanism. The network input is a feature sequence, and the output is the probability distribution of behavior categories. Training utilizes the backpropagation algorithm with an adaptive learning rate optimizer and an early stopping strategy to prevent overfitting.
[0059] The risk assessment module combines rule-based reasoning and probabilistic calculation methods. The rule base contains judgment rules summarized from expert knowledge and historical experience, employing a forward reasoning mechanism to derive conclusions from known facts. Probabilistic calculations are based on Bayesian networks, where nodes represent risk factors and directed edges represent causal relationships. Through probabilistic reasoning, the posterior probabilities of various threat scenarios are calculated, and the overall risk level is comprehensively assessed.
[0060] The decision-making layer formulates response strategies based on the analysis results, including an early warning strategy module and a linkage control module. The early warning strategy module maintains a set of strategy rules, each rule containing triggering conditions and execution actions. Rules are represented using production rules, supporting logical combinations of conditions. Based on the risk level, corresponding rules are matched to generate a sequence of early warning instructions.
[0061] The linkage control module is responsible for interacting with smart home devices. The system supports multiple smart home communication protocols, converting internal control commands into device control commands through a protocol adapter. Multiple preset operating modes are available, including security mode, alert mode, and emergency mode. Security mode implements basic protective measures; alert mode enhances monitoring; and emergency mode initiates a comprehensive response, including device linkage, alarms, and notifications.
[0062] The application layer provides user interaction and system management functions, including a user interaction module, a data management module, and a system configuration module. The user interaction module provides a mobile application and a web management interface. The mobile application supports cross-platform operation and provides functions such as real-time monitoring, historical query, and remote control. The web interface adopts a responsive design to adapt to different terminal devices.
[0063] The data management module is responsible for the storage, retrieval, and maintenance of system data. It employs a hybrid storage architecture: real-time data is stored in an in-memory database to maintain fast read / write performance; historical data is stored in a time-series database to support efficient time-series queries; and key event records are stored in a relational database for easy structured management. A data lifecycle management strategy is implemented, setting different retention periods based on data importance and timeliness.
[0064] The system configuration module allows users to adjust system parameters according to actual needs. Configurable parameters include detection sensitivity, warning threshold, response delay, and operating mode. A graphical configuration interface is provided, allowing users to intuitively adjust parameters and view the effects in real time. Importing and exporting configurations is supported for easy backup and recovery.
[0065] The system also includes a cloud service platform, providing value-added service support. The cloud platform deploys model training services, continuously optimizing the recognition model using large-scale computing resources and massive amounts of data. Data backup services employ an incremental backup strategy to ensure the security of critical data. Remote operation and maintenance services allow technicians to perform remote diagnostics and maintenance with user authorization. The cloud platform communicates with the local system through an encrypted channel, ensuring secure data transmission.
[0066] Combination Figure 3 Thirdly, this invention provides a smart early warning method for home security. The method achieves high-precision intrusion detection and smart early warning through multi-physics field collaborative sensing.
[0067] Step S1: Multiphysics Environment Baseline Establishment. Upon initial installation, the system automatically enters baseline learning mode, simultaneously collecting signals from three physical fields: electrostatics, vibration, and airflow, to establish a database of normal activity patterns in the home environment. The data collection process lasts for a sufficiently long period, covering the complete cycle of weekdays and weekends. The time resolution of data collection is set to the second level, generating tens of thousands of data points daily. The day is divided into multiple time periods, such as early morning, morning, afternoon, evening, and night, with statistical characteristics calculated independently for each period.
[0068] For the data in each time period, calculate the mean and standard deviation: in: For time period Internal sensor The signal mean; The corresponding variance; For time period Number of samples within; For the first Each sampled value; subscript Indicates a time period index; subscript Indicates the sensor index.
[0069] The baseline model is defined as a set of statistical characteristics for each time period: in: Baseline model; The confidence coefficient is usually set to 3. For time periods; This is the collection of all sensors.
[0070] In addition to statistical features, the baseline establishment process also includes behavioral pattern learning. The system records the daily activity patterns of family members, including wake-up time, departure time, return time, and bedtime. Typical activity patterns, such as weekday patterns, weekend patterns, and holiday patterns, are identified through clustering algorithms. Each pattern corresponds to a different distribution of physical field characteristics.
[0071] Step S2: Synchronous Acquisition of 3D Physical Field Signals. After the system enters normal operating mode, the three physical field sensors work in parallel to acquire environmental signals in real time. Electrostatic field signals, vibration signals, and airflow signals are acquired at their respective sampling frequencies. Signals with different sampling rates are synchronized through a timestamp alignment mechanism. Each data point contains a precise timestamp, sensor identification code, and measurement value.
[0072] Time synchronization employs a master-slave architecture with a precise time protocol. The gateway device acts as the master clock, and each sensor node acts as a slave clock. The synchronization process is achieved through multiple message exchanges: the master clock sends a synchronization message, and the slave clocks record the reception time; the master clock sends subsequent messages containing the precise transmission time of the synchronization messages; the slave clocks send delay request messages; and the master clock replies with delay response messages. Clock skew and path delay are calculated and compensated using the timestamp information of the messages.
[0073] Spatial calibration is achieved by establishing a unified coordinate system. A three-dimensional Cartesian coordinate system is established with the geometric center of the building as the origin. The position of each sensor is precisely measured using a laser rangefinder, and the three-dimensional coordinates and installation attitude angles are recorded. Spatial calibration information is stored in a configuration file for subsequent spatial feature analysis.
[0074] Step S3: Preprocessing of Multi-Source Heterogeneous Signals. Customized preprocessing strategies are adopted based on the different characteristics of the three physical field signals. The electrostatic field signal is bandpass filtered using a zero-phase filter to remove DC bias and high-frequency noise while maintaining phase integrity. A moving average method is used for smoothing to reduce random disturbances.
[0075] The vibration signal preprocessing employs wavelet packet denoising. The signal is decomposed into multiple wavelet packets, and threshold denoising is applied to each frequency band. The threshold selection is adaptively determined based on the signal's statistical characteristics. The denoised signal is then reconstructed to obtain a clean vibration signal. For vibrations containing impact components, envelope detection is used to extract impact features.
[0076] The airflow signal undergoes empirical mode decomposition to remove trend terms, retaining high-frequency components reflecting rapid disturbances. Outliers are identified using statistical methods, and data points exceeding reasonable ranges are replaced using interpolation. Temperature compensation corrects the airflow velocity based on real-time temperature measurements, eliminating the effects of temperature variations.
[0077] Step S4: Multidimensional Physical Feature Extraction. Multidimensional feature vectors characterizing the target's properties are extracted from the preprocessed signal. Electrostatic field features include instantaneous amplitude, rate of change, and spatial gradient. The instantaneous amplitude is obtained through Hilbert transform, reflecting the instantaneous strength of the electrostatic field. The rate of change is calculated through differential operations, characterizing the dynamic change in field strength. The spatial gradient is obtained through the difference in field strength between sensors at different locations, reflecting the target's spatial position.
[0078] Vibration features are extracted using joint time-frequency analysis. A time-frequency representation is obtained through continuous wavelet transform, and the distribution of vibration energy in the time and frequency domains is analyzed. Energy-concentrated time-frequency regions are extracted as features, and instantaneous frequency and instantaneous energy are calculated. The time-varying characteristics of vibration frequency are traced using wavelet ridges. Wavelet entropy is calculated to assess signal complexity.
[0079] Airflow characteristics are obtained through fluid dynamics analysis. The velocity gradient tensor is calculated, and the flow field structure is identified through eigenvalue decomposition. Vorticity characterizes the rotational properties of the flow field, while divergence reflects the source and sink distribution of the flow field. The existence of vortices is determined using the Q-criterion, where the Q-value is defined as the relationship between the eigenvalues of the vorticity tensor and the strain rate tensor. The intensity, size, and velocity of the vortices are extracted as characteristic parameters.
[0080] Step S5: Spatiotemporal Alignment Feature Fusion. The three heterogeneous features are spatiotemporally aligned and then fused. Temporal alignment is achieved through interpolation, unifying data from different sampling rates onto the same time grid. Spatial alignment is achieved through coordinate transformation, mapping measurements from different locations to the target location.
[0081] Feature normalization employs a standardization method to eliminate dimensional differences between various physical quantities. Each feature is converted to a standard normal distribution by subtracting its mean and dividing by its standard deviation. The fusion weights are dynamically adjusted based on signal quality, which comprehensively considers signal-to-noise ratio, stability, and continuity. Signals with high signal-to-noise ratio, good stability, and strong continuity are assigned higher weights.
[0082] The fusion strategy employs an adaptive weighting method. Based on environmental conditions and the characteristics of the target being detected, the contribution of each physical field is dynamically adjusted. For example, the weight of vibration detection is increased in quiet environments, while the weight of airflow detection is decreased when air conditioning is running. The fused feature vector contains complementary information from the three physical fields, improving the accuracy of target recognition.
[0083] Step S6: Dual-Mechanism Collaborative Anomaly Detection. A dual-mechanism approach combining statistical detection and machine learning detection is employed. Statistical detection is based on the deviation of the fused features from the baseline. The distance metric between the current feature and the baseline is calculated; commonly used metrics include Euclidean distance, Mahalanobis distance, and Chebyshev distance. Mahalanobis distance considers the correlation between features and is more suitable for anomaly detection of multi-dimensional features. in: The Mahalanobis distance; For current features; The baseline mean; is the baseline covariance matrix.
[0084] Mahalanobis distance follows a chi-square distribution with degrees of freedom equal to the feature dimension. A detection threshold is determined based on a set significance level. When the Mahalanobis distance exceeds the threshold, it is judged as a statistical anomaly.
[0085] Machine learning detection employs a type of support vector machine, trained only on normal data, to learn the support domain boundaries of normal data. The decision function is defined as: in: Let it be the decision function; For Lagrange multipliers; For kernel functions; As support vectors; This is a bias term.
[0086] When the decision function value is negative, the system is considered an anomaly. The two detection results are combined through logical operations; an "AND" operation can be used to improve accuracy, while an "OR" operation can be used to improve recall. The appropriate combination strategy should be selected based on the specific application requirements.
[0087] Step S7: Intelligent Identity and Behavior Recognition. This step involves identifying and analyzing the behavior of detected abnormal targets. Identity recognition is achieved through feature matching. The similarity between the current feature and each template in the identity database is calculated. Common similarity metrics include cosine similarity, Pearson correlation coefficient, and dynamic time-warped distance. Cosine similarity is simple to calculate and insensitive to feature scale.
[0088] in: Similarity; , The numerator is the eigenvector; the numerator is the inner product of the vectors; the denominator is the product of the vector magnitudes.
[0089] A successful match is determined when the similarity score is the highest and exceeds a preset threshold, and the corresponding identity is identified. If all similarities are below the threshold, the user is marked as a stranger.
[0090] Behavioral analysis employs deep learning methods. A gated recurrent unit (GRU) network is used to process temporal features; the network selectively memorizes and forgets information through a gating mechanism. Update gates control the reception of new information, while reset gates determine the utilization of historical information. The network outputs a probability distribution of behavioral categories, selecting the category with the highest probability as the recognition result.
[0091] Step S8: Multi-factor Comprehensive Risk Assessment. Establish a multi-dimensional risk assessment model that comprehensively considers factors such as identity, behavior, time, and environment. Identity risk is assigned a value based on personnel category, with family members having the lowest risk and strangers the highest. Behavioral risk is determined based on behavioral intent, with normal activities carrying low risk and abnormal behavior carrying high risk. Time risk considers the time period in which the event occurs, with higher risk at night than during the day. Environmental risk assesses the current environmental state, with increased risk when no one is home.
[0092] The risk value is calculated through a weighted summation, with weighting coefficients determined based on expert experience and historical data statistics. The weights meet normalization conditions to ensure the risk value remains within a reasonable range. Risk assessment also considers dynamic changes in risk, judging the threat's development trend by accumulating risk over a time window.
[0093] Step S9: Tiered Early Warning and Intelligent Response. Based on the comparison between the risk value and the preset threshold, a tiered response strategy is executed. The risk is divided into multiple levels, each corresponding to different response measures.
[0094] At the normal level, the system maintains silent monitoring, recording only necessary log information to avoid disturbing the user. At the suspicious level, monitoring density is increased, data sampling intervals are shortened, and detailed event data is stored for subsequent analysis. At the warning level, notifications are sent to the user via mobile app push notifications, SMS, or telephone. Simultaneously, response devices are pre-activated, such as enabling camera previews and preparing alarms. At the danger level, a full response is immediately executed, including triggering audible and visual alarms, automatically dialing emergency numbers, starting video recording, and implementing device linkage.
[0095] The response strategy supports user-defined configuration. Users can adjust the response actions at different levels according to their individual needs, such as setting do-not-disturb periods, specifying emergency contacts, and selecting linked devices. The system also supports scenario-based configuration, with preset modes such as outing mode, sleep mode, and vacation mode, allowing users to switch between different protection strategies with a single click.
[0096] Step S10: Online Adaptive Model Update. An incremental learning strategy is used to continuously optimize the model. When new data arrives, the entire model does not need to be retrained; instead, parameters are fine-tuned based on the existing model. Parameter updates use stochastic gradient descent. in: For a moment Model parameters; Adaptive learning rate; The gradient of the loss function; Input data; For tags.
[0097] The learning rate employs an adaptive adjustment strategy, dynamically updating based on historical gradient information. Commonly used adaptive optimization algorithms include Adam and RMSprop, which can accelerate convergence and avoid local optima.
[0098] Model performance is monitored and evaluated online. Performance metrics are set, including detection accuracy, false positive rate, and false negative rate. When a performance metric drops below a threshold, model retraining is triggered. Retraining uses recent data to ensure the model adapts to changes in the environment.
[0099] The system also implements a concept drift detection mechanism. It uses statistical tests to detect changes in data distribution, and when a significant change in distribution is detected, it adjusts model parameters or updates the model structure in a timely manner. This adaptive mechanism ensures that the system maintains stable performance during long-term operation.
[0100] Through the above-described embodiments, this invention achieves intelligent early warning for home security based on multi-physics field collaborative sensing. Compared with traditional single-sensor methods, it significantly improves detection accuracy and system reliability, providing an innovative technical solution for home security. The technical solution of this invention is not only applicable to ordinary home environments but can also be extended to various scenarios such as offices, shops, and warehouses, demonstrating broad application prospects.
[0101] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A smart home security early warning device, characterized in that, include: The electrostatic field sensing module is used to detect changes in the spatial electrostatic field distribution through a distributed high-impedance antenna array embedded in the ceiling light fixture, extract the time-domain, frequency-domain, and statistical domain features of the electrostatic field, and generate an electrostatic field feature vector. The vibration detection module is used to collect vibration signals transmitted by the building structure through piezoelectric sensors installed on door and window frames, stair treads and load-bearing walls. It extracts vibration energy distribution and spectral characteristics through time-frequency joint analysis and generates vibration feature vectors. The airflow analysis module is used to monitor indoor airflow field disturbances through thermal flow sensors integrated in the air conditioning return air vent and smart switch panel, and to invert the airflow vortex structure based on the principle of computational fluid dynamics to generate airflow feature vectors. The feature fusion module is connected to the electrostatic field sensing module, vibration detection module and airflow analysis module respectively. It is used to perform spatiotemporal alignment and normalization processing on the three physical field features, dynamically adjust the fusion weights according to the signal-to-noise ratio and signal quality of each sensor, and generate a multi-dimensional fusion feature vector. The target recognition module, connected to the feature fusion module, is used to perform identity matching by calculating the Mahalanobis distance between the fused feature vector and the pre-built identity database, and to perform behavioral analysis on the temporal features using a hidden Markov model. The risk assessment module, connected to the target identification module, is used to calculate the comprehensive risk value by weighted summation of four factors: identity risk, behavioral risk, time risk, and environmental risk. The early warning control module, connected to the risk assessment module, is used to generate graded early warning instructions based on the comparison relationship between risk values and multi-level thresholds. The response execution module is connected to the early warning control module and is used to execute local audible and visual alarms, remote notification pushes, and intelligent device linkage control according to the early warning instructions.
2. The home security intelligent early warning device according to claim 1, characterized in that, The antenna array of the electrostatic field sensing module uses conductive planar electrodes with high input impedance. The signal acquisition circuit includes a preamplifier with ultra-low input bias current, a low-frequency bandpass filter, and a high-precision analog-to-digital converter. Feature extraction adopts a sliding window method with partial overlap between windows.
3. The home security intelligent early warning device according to claim 1, characterized in that, The vibration detection module performs multi-level decomposition of the vibration signal through wavelet packet decomposition, and applies soft threshold denoising to each frequency band. The threshold is adaptively determined by the unbiased risk estimation criterion. The extracted features include vibration amplitude features, frequency domain energy features, and time-frequency distribution features.
4. The home security intelligent early warning device according to claim 1, characterized in that, The airflow analysis module uses a temperature compensation algorithm to correct the influence of ambient temperature on flow velocity measurement, identifies vortex structures through velocity gradient tensor analysis, and determines the existence of vortices based on the relative magnitudes of vorticity and strain rate.
5. The home security intelligent early warning device according to claim 1, characterized in that, The fusion weights of the feature fusion module are calculated based on the confidence coefficients and signal quality indicators of each sensor, ensuring that the sum of the weights is 1. The fused feature vector is obtained by a weighted linear combination of the normalized feature vectors.
6. A smart home security early warning system, characterized in that, include: The sensing layer includes a distributed electrostatic field sensing module, a vibration detection module, and an airflow analysis module. The electrostatic field sensing module, vibration detection module, and airflow analysis module respectively collect corresponding physical field signals and transmit them through their respective wireless communication protocols. The transport layer, connected to the perception layer, is used to receive data packets of different protocols, parse and extract the payload, re-encapsulate them into a unified internal data format, and select the transmission path according to the data priority. The processing layer, connected to the transmission layer, is used to perform quality assessment, noise filtering, and feature extraction on the received multi-source signals, and to fuse the judgment results of each sensor using evidence theory. The analysis layer, connected to the processing layer, includes a target recognition module, a behavior analysis module, and a risk assessment module, which are used for identity classification based on support vector machines, behavior recognition based on long short-term memory networks, and risk reasoning based on Bayesian networks, respectively. The decision-making layer, connected to the analysis layer, is used to match preset strategy rules based on the analysis results and generate corresponding control commands. The application layer, connected to the decision layer, is used to provide users with system status display, parameter configuration, and remote control functions.
7. A smart early warning method for home security, characterized in that, Includes the following steps: S1: During the system initialization phase, electrostatic field signals, vibration signals, and airflow signals are continuously collected within a set period. The collection period is divided into multiple intervals, and the mean and variance of each signal in each interval are calculated to form a baseline model that characterizes the normal activity mode. S2: When the system is working normally, the electrostatic field sensing module, vibration detection module and airflow analysis module collect signals at their respective sampling frequencies, and ensure the time consistency of the data through timestamp marking and clock synchronization protocol. S3: Perform bandpass filtering and moving average processing on the collected electrostatic field signal, wavelet decomposition and threshold denoising on the vibration signal, and modal decomposition and trend separation on the airflow signal. S4: Extract features from the preprocessed signal, including amplitude changes and spatial distribution of the electrostatic field, spectral energy and time-frequency patterns of vibration, velocity gradient and vortex characteristics of airflow. S5: Unify the feature data with different sampling rates to the same time base through interpolation, eliminate the dimensional differences of physical quantities by normalization, calculate the fusion weights based on the signal quality of each sensor and perform weighted combination. S6: Compare the fused features with the baseline model, calculate the degree of deviation, and use a pre-trained classifier to detect anomalies. Combine the two detection results to determine whether anomalies exist. S7: When an anomaly is detected, the current feature is matched with the identity feature database to identify the target identity, and the behavior type is determined by analyzing the temporal pattern of the feature sequence. S8: Assign corresponding risk scores based on the identified identity category, behavior type, current time period, and environmental status, and obtain the comprehensive risk assessment value by weighted summation; S9: Compare the risk assessment value with preset multi-level thresholds, and execute corresponding response measures according to the risk range, including record storage, message push, audible and visual alarms, and device linkage.