Home security model cooperation method and system based on wireless communication

Abstract

The application relates to a wireless communication-based home security model cooperation method and system, which comprises the following steps: collecting steady-state I / Q wireless signals, extracting and fusing feature vectors through double-channel feature fusion and SE attention mechanism; based on an integrated classifier and an improved OpenMax function, performing open set intrusion detection, generating a first alarm signal for unknown intrusion equipment; collecting environmental parameters through a ZigBee sensor network and comparing the environmental parameters with threshold values to generate a second alarm signal; extracting a speech mel-frequency cepstral coefficient and generating a user intention signal by matching and generating the speech mel-frequency cepstral coefficient through an optimized dynamic time warping algorithm; a cooperative decision engine receives the above signals, generates a security linkage instruction after time-space correlation verification and rule matching, and issues the security linkage instruction to an executor. The application fuses radio frequency fingerprint recognition and environmental perception, realizes joint defense from a wireless signal layer to a physical environment layer through multi-model cooperation, and effectively improves the accuracy and reliability of home security.

Description

Technical Field

[0001] This application relates to the field of home security technology, and in particular, to a collaborative method and system for a home security model based on wireless communication. Background Art

[0002] With the rapid development of Internet of Things technology and the wide popularization of smart home devices, home security systems are gradually evolving from traditional independent alarm devices towards networking and intelligence. In a modern home environment, various security devices such as door magnetic sensors, infrared detectors, smoke alarms, and smart cameras are usually deployed. These devices are networked through wireless communication protocols, can real-time sense abnormal conditions in the home physical environment, and send warning messages to users or execute preset linkage actions.

[0003] Currently, a typical home security system adopts the following architecture: using an embedded microcontroller as the main control core, connecting sensor terminal nodes distributed in each room through a ZigBee wireless sensor network, including temperature and humidity sensors, smoke concentration sensors, combustible gas sensors, and passive infrared sensors. Each sensor terminal node periodically collects environmental parameters and transmits the data to the main controller through a ZigBee coordinator. The main controller compares the collected data with preset thresholds. When detecting parameter abnormalities, it drives a buzzer to give a local alarm, and at the same time uploads the warning message to the cloud platform through a Wi-Fi module. Users can receive push notifications and conduct remote viewing through a mobile terminal application.

[0004] Regarding the above technology, the above security system relies on MAC address filtering or key authentication, and these upper-layer logical identifiers are extremely easy to be tampered with or cracked by software. The sensor network cannot sense such illegal access behaviors, resulting in the risk that user privacy data is stolen.

[0005] Based on this, this application provides a collaborative method and system for a home security model based on wireless communication. Summary of the Invention

[0006] In order to improve the problem that the above security system relies on MAC address filtering or key authentication, and these upper-layer logical identifiers are extremely easy to be tampered with or cracked by software. The sensor network cannot sense such illegal access behaviors, resulting in the risk that user privacy data is stolen, this application provides a collaborative method and system for a home security model based on wireless communication.

[0007] In a first aspect, a collaborative method for a home security model based on wireless communication provided by this application adopts the following technical solutions: including: The system collects steady-state I / Q wireless signals from target devices in a home environment, extracts multi-dimensional handcrafted features through an expert feature channel, extracts time-frequency domain depth features through a one-dimensional convolutional neural network channel, and then splices the features together and recalibrates them through an SE attention module to output a fused feature vector. The fused feature vector is input into an ensemble classifier based on the AutoGluon framework, and the output of the penultimate fully connected layer is extracted as the activation vector. The average activation vector is calculated based on the activation vectors of known device categories. The tail distribution of the distance from the activation vector of each category to the corresponding average activation vector is fitted based on the Weibull distribution to establish a boundary probability model. A decision threshold is set independently for each category of device, and the belonging probability is calculated using the improved OpenMax function. If the maximum probability is less than the corresponding decision threshold, a first alarm signal is generated. Real-time parameters in the home physical environment are collected through a ZigBee wireless sensor network. The real-time parameters are compared with preset safety thresholds. When any of the real-time parameters exceeds the corresponding preset safety threshold, a second alarm signal is generated. The user's voice signal is acquired and preprocessed by pre-emphasis, frame windowing and endpoint detection. Mel frequency cepstral coefficients are extracted as voice features. The optimized dynamic time warping algorithm is used to match the voice features with the pre-stored instruction template to generate the corresponding user intent signal. The system receives the first alarm signal, the second alarm signal, and the user intent signal, performs verification and matching based on preset spatiotemporal association rules, generates a security linkage command, and sends it to the actuator terminal through the ZigBee network to execute the response action.

[0008] Preferably, the step of extracting multi-dimensional handcrafted features through an expert feature channel, extracting time-frequency domain depth features through a one-dimensional convolutional neural network channel, concatenating them, recalibrating them through an SE attention module, and outputting a fused feature vector includes: The acquired steady-state I / Q wireless signal is divided into two paths: the first path is input to the expert feature channel, and the second path is input to the one-dimensional convolutional neural network channel. In the expert feature channel, the in-phase component I and quadrature component Q of the steady-state I / Q wireless signal are calculated respectively to extract the multidimensional handcrafted features, which include: amplitude, phase, real part of mean, imaginary part of mean, real part of standard deviation, imaginary part of standard deviation, maximum value, minimum value, mean absolute deviation, skewness, kurtosis, information entropy, energy, and root mean square. In a one-dimensional convolutional neural network channel, a network structure containing two convolutional layers and two pooling layers is used to perform convolution and pooling processing on steady-state I / Q wireless signals to extract the time-frequency domain depth features, and the extraction results are flattened. The extracted multidimensional handcrafted features are concatenated with the flattened time-frequency domain depth features to obtain initial concatenated features; the initial concatenated features are then input into the SE attention module. The SE attention module compresses the spatial dimension through a global average pooling layer, and then generates channel weights by passing them sequentially through a first fully connected layer, a ReLU activation function, a second fully connected layer, and a Sigmoid activation function. Finally, each channel weight is multiplied by the initial concatenated feature channel by channel to output the fused feature vector after weight calibration.

[0009] Preferably, the step of inputting the fused feature vector into an ensemble classifier based on the AutoGluon framework, extracting the output of the penultimate fully connected layer as the activation vector; calculating the average activation vector based on the activation vectors of known device categories; and establishing a boundary probability model by fitting the tail distribution of the distances from the activation vectors of each category to the corresponding average activation vector based on the Weibull distribution, including: The ensemble classifier based on the AutoGluon framework adopts a bagging ensemble model with a lightweight gradient boosting machine. It integrates and trains multiple base learners using the bagging method, and L2 regularization is used during the training process. The fused feature vector is input into the trained ensemble classifier, and the output value of the second-to-last fully connected layer in the ensemble classifier is extracted as the activation vector of the corresponding sample. For each known device category, collect the activation vectors of all correctly classified samples in the corresponding category, calculate the arithmetic mean of the activation vectors of all samples in that category, and use the arithmetic mean as the average activation vector of that category; For each known device category, calculate the Euclidean distance from the activation vector of each sample within the category to the corresponding average activation vector to obtain the distance set for that category; select a predetermined number of sample points with the largest distance values ​​from the distance set as tail samples; The distance distribution of the tail samples is fitted using a two-parameter Weibull distribution to obtain the Weibull probability density function and cumulative distribution function of the category, which serve as the boundary probability model corresponding to the known device category. The two-parameter Weibull distribution includes a shape parameter that controls the shape of the distribution and a scale parameter that controls the scaling of the distribution.

[0010] Preferably, the step of fitting the distance distribution of the tail samples using a two-parameter Weibull distribution to obtain the Weibull probability density function and cumulative distribution function of the category, as the boundary probability model corresponding to the known device category, includes: Obtain the Euclidean distance values ​​corresponding to each sample point in the tail samples of the target device category to form a distance dataset to be fitted; The maximum likelihood estimation method is used to estimate the shape parameter and scale parameter of the two-parameter Weibull distribution based on the distance dataset to be fitted. The shape parameter determines the morphological characteristics of the distribution curve, and the scale parameter determines the broadening degree of the distribution curve. Substituting the estimated shape and scale parameters into the probability density function expression of the Weibull distribution, we obtain the Weibull probability density function corresponding to the target equipment category. The probability density function is used to describe the relative probability of different Euclidean distance values ​​occurring within the target equipment category. Integrating the probability density function yields the cumulative distribution function corresponding to the target device category. The cumulative distribution function is used to calculate the cumulative probability of any given Euclidean distance value in that category distribution. The probability density function and the cumulative distribution function are used together as the boundary probability model corresponding to the target device category, which is then used to calculate the confidence probability that the test sample belongs to the target device category.

[0011] Preferably, the decision thresholds set independently for each type of device are used to calculate the attribution probability using an improved OpenMax function, including: Obtain a validation set sample containing labeling information for each known device category, wherein the validation set sample was not involved in the training process of the ensemble classifier; The validation set samples are input into the ensemble classifier to extract the corresponding validation set activation vectors. Calculate the Euclidean distance from the activation vectors of the validation set to the average activation vectors of each known device category; Substituting the Euclidean distance of the validation set samples into the cumulative distribution function of the corresponding known device categories, we obtain the preliminary confidence probability that the validation set samples belong to each known device category; For each known device category, within the preset threshold search range, each candidate threshold is traversed with a preset step size; for each candidate threshold, the precision and recall of the target device category on the validation set samples are calculated, and the F1 score is further calculated, which is the harmonic mean of precision and recall. Select the candidate threshold that maximizes the F1 score as the decision threshold corresponding to the target device category; repeat the search process until the decision threshold corresponding to each of the known device categories is determined.

[0012] Preferably, the calculation of the attribution probability using the improved OpenMax function includes: The steady-state I / Q wireless signal of the target device to be identified is collected, and after feature extraction and fusion, it is input into the ensemble classifier to extract the corresponding test sample activation vector; Calculate the Euclidean distance from the activation vector of the test sample to the average activation vector of each known device category to obtain a set of test distance values; Substitute the test distance values ​​into the cumulative distribution function of the corresponding known device category to obtain the reduction coefficient for each known device category; multiply the probability of each category output by the original SoftMax function by the corresponding reduction coefficient to obtain the reduced probability of the known category; Calculate the difference between the probability of each category output by the original SoftMax function and the reduced probability of the known categories, and sum the differences of each known category to obtain the probability of the unknown device belonging to a category; The reduced known category probabilities and the unknown device category attribution probabilities are combined to form an extended probability vector. The extended probability vector is then normalized so that the sum of all category probabilities is one. The maximum probability value is found in the normalized extended probability vector; if the category corresponding to the maximum probability value is a known device category and the maximum probability value is not less than the decision threshold corresponding to the known device category, then the target device to be identified is determined to be the known device category; otherwise, the target device to be identified is determined to be an unknown intrusion device, and the first alarm signal is generated.

[0013] Preferably, the step of acquiring user speech signals and performing pre-emphasis, frame-segmentation windowing, and endpoint detection preprocessing, extracting Mel-frequency cepstral coefficients as speech features; and using an optimized dynamic time warping algorithm to match the speech features with a pre-stored instruction template to generate a corresponding user intent signal, including: After acquiring the user's voice signal, the user's voice signal is pre-emphasized through a first-order finite impulse response high-pass filter to compensate for the high-frequency energy loss caused by the lip radiation effect and improve the signal-to-noise ratio in the high-frequency band. The pre-emphasized user speech signal is divided into short time frames, and each frame signal is multiplied by a Hamming window function to reduce signal discontinuity at frame boundaries, thus obtaining windowed frame signals. Calculate the short-time energy and short-time zero-crossing rate of each windowed frame signal, use the dual-threshold detection method to distinguish speech segments from silence segments, remove noise and non-speech regions, and extract effective speech segments; The effective speech segment is subjected to a fast Fourier transform to obtain an amplitude spectrum. The amplitude spectrum is then filtered through a Mel filter bank. The energy of each frequency band of the filtered output is taken as a logarithm and then subjected to a discrete cosine transform. The Mel frequency cepstral coefficients are extracted as the speech features. The speech features are matched with pre-stored instruction templates. The optimized dynamic time warping algorithm constrains the matching search path to two parallelogram regions with slopes of 1 / 2 and 2 respectively. The cumulative distance is calculated only for pre-stored instruction templates that meet the frame number constraint, thereby reducing the amount of invalid matching calculations. The frame number constraint is that the ratio of the number of input speech feature frames to the number of template frames is between 1 / 2 and 2. Calculate the cumulative matching distance between the speech features and each pre-stored instruction template, select the pre-stored instruction template with the smallest cumulative matching distance that is lower than a preset similarity threshold as the matching result, and generate the user intent signal corresponding to the matching result.

[0014] Preferably, the step of receiving the first alarm signal, the second alarm signal, and the user intent signal, and performing verification and matching based on preset spatiotemporal association rules to generate a security linkage command includes: Receive the first alarm signal, the second alarm signal, and the user control intent signal, and add a timestamp of the time of reception to each signal; When both the first alarm signal and the second alarm signal exist simultaneously, it is determined whether the spatial regions corresponding to the two belong to the coverage area of ​​the same ZigBee subnet, and whether the difference between their timestamps is less than a preset time window threshold. If the spatial regions are consistent and the time difference is less than the preset time window threshold, the spatiotemporal correlation verification is deemed to have passed, and the response level is improved. If only a single alarm signal exists, an independent response is performed according to the corresponding preset rules. The alarm signal combination or independent alarm signal after being verified by spatiotemporal correlation will be compared with the rule conditions in the preset collaborative rule base; the preset collaborative rule base predefines the logical combination relationship and corresponding linkage action mapping between different alarm signal types, user control intention signals and the current system arming status. Based on the successfully matched collaboration rules, a corresponding security linkage instruction is generated, which includes the target actuator terminal identifier and the action parameters to be executed. The security linkage command is wirelessly transmitted to the corresponding actuator terminal via a ZigBee network, driving the actuator terminal to perform the corresponding security response action.

[0015] Secondly, this application discloses a home security model collaboration device based on wireless communication, which adopts the following technical solution, including: The feature extraction module is used to collect steady-state I / Q wireless signals of target devices in the home environment, extract multi-dimensional handmade features through the expert feature channel, extract time-frequency domain depth features through the one-dimensional convolutional neural network channel, and after splicing, recalibrate through the SE attention module to output a fused feature vector. The classification probability module is used to input the fused feature vector into an ensemble classifier based on the AutoGluon framework, extract the output of the penultimate fully connected layer as the activation vector; calculate the average activation vector based on the activation vectors of known device categories; fit the tail distribution of the distance from each category's activation vector to the corresponding average activation vector based on the Weibull distribution to establish a boundary probability model; calculate the attribution probability using an improved OpenMax function based on a decision threshold set independently for each category of device; and generate a first alarm signal if the maximum probability is less than the corresponding decision threshold. The real-time alarm module is used to collect real-time parameters in the home physical environment through a ZigBee wireless sensor network, compare the real-time parameters with a preset safety threshold, and generate a second alarm signal when any of the real-time parameters exceeds the corresponding preset safety threshold. The speech signal module is used to acquire user speech signals and perform pre-emphasis, frame windowing and endpoint detection preprocessing, and extract Mel frequency cepstral coefficients as speech features; the speech features are matched with pre-stored instruction templates using an optimized dynamic time warping algorithm to generate corresponding user intent signals; The security linkage module is used to receive the first alarm signal, the second alarm signal and the user intent signal, perform verification and matching based on preset spatiotemporal association rules, generate security linkage instructions, and send them to the actuator terminal through the ZigBee network to execute the response action.

[0016] Thirdly, this application also provides a control device, the device comprising: It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed, such as the wireless communication-based home security model collaboration method described above.

[0017] Fourthly, this application also provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above regarding the collaborative method for a home security model based on wireless communication.

[0018] In summary, this application's system collects steady-state I / Q wireless signals from target devices in a home environment, extracts multi-dimensional handcrafted features through an expert feature channel, and simultaneously extracts time-frequency domain depth features through a one-dimensional convolutional neural network channel. These features are then concatenated and recalibrated by an SE attention module to output a fused feature vector. This fused feature vector is input into an ensemble classifier based on the AutoGluon framework to extract activation vectors. The average activation vector for each known device category is calculated, and a boundary probability model is established based on the tail distribution of the distance from the activation vector to the average activation vector using a Weibull distribution. Independent decision thresholds are set for each category, and an improved OpenMax function is used to calculate the attribution probability. When the maximum probability is less than the corresponding independent decision threshold, a first alarm signal is generated. Simultaneously, real-time parameters in the home physical environment are collected via a ZigBee wireless sensor network and compared with preset safety thresholds to generate a second alarm signal. User voice signals are collected, and Mel-frequency cepstral coefficients are extracted. An optimized dynamic time warping algorithm is used to match these parameters with a pre-stored instruction template to generate a user control intent signal. The collaborative decision engine receives these three signals, performs verification and matching based on preset spatiotemporal association rules, generates security linkage instructions, and sends them to the actuator terminal via the ZigBee network. This enables cross-domain collaboration between wireless signal layer intrusion detection and physical environment perception, effectively identifying unknown intrusion devices that are disguised, reducing the false alarm rate and missed alarm risk of traditional security systems, and significantly improving the overall reliability of home security. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a collaborative method for home security models based on wireless communication.

[0020] Figure 2 This is a structural block diagram of a home security model collaborative device based on wireless communication. Detailed Implementation

[0021] The following combination Figure 1 - Figure 2 This application will be described in further detail.

[0022] With the rapid increase in the number of wireless devices in the home environment, traditional security systems often only focus on monitoring physical environmental parameters, while neglecting the ability to detect unauthorized device access at the wireless signal level; at the same time, various security modules operate relatively independently and lack an effective collaborative judgment mechanism.

[0023] To address the aforementioned issues, this application proposes a collaborative home security model solution based on wireless communication, with the collaborative model's control system as the executing entity. The system comprises four core subsystems: the first subsystem is an RF fingerprint recognition module, responsible for collecting steady-state I / Q signals from wireless devices in the home environment, extracting fused feature vectors through a dual-channel deep feature fusion network, and determining whether an access device is an unknown intrusion device based on an improved OpenMax open-set recognition algorithm; the second subsystem is a multi-source environmental perception module, which collects physical environmental parameters such as temperature, humidity, smoke concentration, combustible gas concentration, and human infrared sensing in real time through a ZigBee wireless sensor network, compares them with preset safety thresholds, and generates an environmental anomaly alarm; the third subsystem is a voice interaction module, which uses an optimized dynamic time warping algorithm to recognize user voice commands and generate user control intent signals; the fourth subsystem is a collaborative decision engine, which receives alarm signals and intent signals from the above three modules, performs comprehensive verification and logical matching based on preset spatiotemporal association rules, generates the final security linkage command, and drives the actuator terminal to complete the response action.

[0024] Through the collaboration of the above four subsystems, a joint defense system from the wireless signal layer to the physical environment layer is constructed, which effectively solves the technical defects of the existing technology, such as single perception dimension and insufficient module collaboration, and greatly increases the security and privacy of smart home life.

[0025] Reference Figure 1 The embodiments of this application include at least steps S10 to S50.

[0026] S10 collects steady-state I / Q wireless signals from target devices in a home environment, extracts multi-dimensional handcrafted features through the expert feature channel, extracts time-frequency domain depth features through the one-dimensional convolutional neural network channel, splices them together, recalibrates them through the SE attention module, and outputs a fused feature vector.

[0027] S20: Input the fused feature vector into an ensemble classifier based on the AutoGluon framework, and extract the output of the penultimate fully connected layer as the activation vector; calculate the average activation vector based on the activation vectors of known device categories; fit the tail distribution of the distance from each category's activation vector to the corresponding average activation vector based on the Weibull distribution to establish a boundary probability model; set a decision threshold independently for each category of device, and calculate the attribution probability using the improved OpenMax function; if the maximum probability is less than the corresponding decision threshold, generate the first alarm signal.

[0028] S30 collects real-time parameters in the home physical environment through a ZigBee wireless sensor network, compares the real-time parameters with preset safety thresholds, and generates a second alarm signal when any real-time parameter exceeds the corresponding preset safety threshold.

[0029] S40: Acquire user speech signals and perform pre-emphasis, frame windowing and endpoint detection preprocessing; extract Mel frequency cepstral coefficients as speech features; use an optimized dynamic time warping algorithm to match speech features with pre-stored instruction templates to generate corresponding user intent signals.

[0030] S50 receives the first alarm signal, the second alarm signal, and the user intent signal. Based on preset spatiotemporal association rules, it performs verification and matching, generates security linkage instructions, and sends them to the actuator terminal via the ZigBee network to execute the response action.

[0031] It should be understood that, in this application, the device category refers to the types of legally registered and authorized wireless devices in the home, including but not limited to: smartphones, computers, smartwatches or wireless earphones, smart home control terminals and various sensors used by family members.

[0032] Specifically, a dual-channel feature extraction process is performed on the steady-state I / Q signals of home wireless devices. One channel calculates manually statistical features such as amplitude and phase, while the other channel utilizes a one-dimensional convolutional network to mine deep features in the time-frequency domain. The concatenated features are then automatically weighted by the SE attention module to form a highly discriminative fused feature vector. This vector is then fed into an ensemble classifier, where the activation values ​​of the penultimate fully connected layer are extracted as device fingerprints. The average center of activation values ​​for each class of legitimate devices is calculated, and a boundary probability model is obtained by modeling the intra-class distance tail using a Weibull distribution. An improved open set recognition function is then used to calculate the attribution confidence of the device under test. If the confidence is lower than the dynamic threshold for independent classification of that class, a wireless intrusion alarm is triggered. Simultaneously, a ZigBee sensor network continuously collects environmental parameters such as temperature, humidity, smoke, gas concentration, and human infrared radiation. Exceeding these limits generates an environmental anomaly alarm. The voice module optimizes dynamic time warping to match user commands and outputs control intent signals. The collaborative model integrates the above alarms and intentions, performs logical analysis based on spatiotemporal correlation rules, and finally generates linkage commands to drive the actuator response, thereby forming a joint defense mechanism from the wireless signal layer to the physical environment layer.

[0033] Furthermore, this application explicitly employs an RF fingerprint feature extraction technology that fuses expert features with deep features. This technology extracts subtle differences in the inherent physical layer RF fingerprint of the transmitter hardware components. This fingerprint is unique and unforgeable. Even if an attacker modifies the MAC address or cracks the Wi-Fi password, its hardware fingerprint features will still not match the features in the legitimate device database, thus facilitating the identification of unknown intrusion devices.

[0034] In some embodiments, step S10 specifically includes the following steps: the acquired steady-state I / Q wireless signal is divided into two paths, the first path is input to the expert feature channel, and the second path is input to the one-dimensional convolutional neural network channel; in the expert feature channel, the in-phase component I path and the quadrature component Q path of the steady-state I / Q wireless signal are calculated respectively to extract multi-dimensional handcrafted features, which include: amplitude, phase, real part of mean, imaginary part of mean, real part of standard deviation, imaginary part of standard deviation, maximum value, minimum value, mean absolute deviation, skewness, kurtosis, information entropy, energy, and root mean square; For I / Q signals, let the amplitude of the in-phase component be I and the amplitude of the quadrature component be Q, then the formula for calculating the amplitude A is: Phase The calculation formula is: For a signal sequence containing n samples Where n represents the total number of samples, the formula for calculating the mean is: , where k is the sample index variable. The standard deviation is calculated using the following formula: ; Skewness The formula used to measure the asymmetry of the distribution is: ; Kudo The formula used to measure the sharpness of the peaks in a distribution is: ; energy The calculation formula is: ; Root Mean Square The calculation formula is: ; Mean absolute deviation The calculation formula is: ; In a one-dimensional convolutional neural network channel, a network structure containing two convolutional layers and two pooling layers is used to perform convolution and pooling processing on steady-state I / Q wireless signals to extract time-frequency domain depth features. The extracted results are then flattened. The extracted multi-dimensional handcrafted features are concatenated with the flattened time-frequency domain depth features to obtain the initial concatenated features. The initial concatenated features are then input into the SE attention module. In the SE attention module, let the height of the input feature map U be H, the width be W, and the number of channels be C. For the c-th channel in spatial position... The value at that location is denoted as Where c ranges from 1 to C, i ranges from 1 to H, and j ranges from 1 to W. The global average pooling layer compresses the spatial dimension of each channel into a scalar, calculated as follows: ; income This is the global descriptor value for the c-th channel. Subsequently, this global descriptor value is passed sequentially through the first fully connected layer, the ReLU activation function, the second fully connected layer, and the Sigmoid activation function to generate the channel weight vector s, calculated using the following formula: ; in, This is the weight matrix of the first fully connected layer. This is the weight matrix of the second fully connected layer. This represents the Sigmoid activation function, whose output value is between zero and one. Finally, the channel weights are multiplied channel by channel by the initial concatenation features, i.e. , where the symbol This represents element-wise multiplication, and the output is the fused feature vector after weight calibration. .

[0035] Specifically, a dual-channel parallel architecture is employed to extract features from steady-state I / Q wireless signals. The first expert feature channel performs statistical calculations on the in-phase and quadrature components of the signal, extracting fourteen types of handcrafted features, including amplitude, phase, mean, standard deviation, skewness, kurtosis, information entropy, energy, and root mean square, to characterize the overall distribution and morphology of the signal. The second one-dimensional convolutional neural network channel automatically learns the local detail patterns of the signal in the time-frequency domain through two layers of convolution and two layers of pooling operations, obtaining deep features after flattening. The two types of features are concatenated to form an initial fused representation, which is then fed into the SE attention module for channel-level weight recalibration. This module first compresses the spatial information of each channel into a scalar using global average pooling, then uses two fully connected layers with activation functions to learn the importance of each channel. Finally, the learned weight coefficients are multiplied back into the original concatenated features channel by channel, enhancing key features and suppressing redundant features, outputting a more discriminative fused feature vector.

[0036] In some embodiments, step S20 specifically includes the following steps: the ensemble classifier based on the AutoGluon framework adopts a lightweight gradient boosting machine bagging ensemble model, and performs ensemble training on multiple base learners using the bagging method, with L2 regularization applied during training; the fused feature vector is input into the trained ensemble classifier, and the output value of the second-to-last fully connected layer in the ensemble classifier is extracted as the activation vector of the corresponding sample; for each known device category, the activation vectors of all correctly classified samples of the corresponding category are collected, the arithmetic mean of the activation vectors of all samples in that category is calculated, and the arithmetic mean is used as the average activation vector of that category; For each known device category, calculate the Euclidean distance from the activation vector of each sample within the category to the corresponding average activation vector to obtain the distance set for that category.

[0037] Let the activation vector be: ; The average activation vector is: Where n is the vector dimension, the Euclidean distance is calculated as follows: ; in, Let d be the dimension index variable. A predetermined number of sample points with the largest distance values ​​are selected from the distance set as tail samples. A two-parameter Weibull distribution is used to fit the distance distribution of the tail samples to obtain the Weibull probability density function and cumulative distribution function for this category, which serve as the boundary probability model corresponding to the known device category. Let the distance variable be d, and the expression for the probability density function of the Weibull distribution is: When the distance d is greater than or equal to zero When the distance d is less than zero, the probability density is zero.

[0038] Wherein, parameter k is a shape parameter that determines the steepness of the distribution curve; parameter The scale parameter and A value greater than zero controls the broadening of the distribution curve; parameter Here, is the location parameter, representing the initial offset of the distribution curve. The corresponding cumulative distribution function expression is: ; This function gives the cumulative probability value corresponding to the distance variable d.

[0039] Specifically, a lightweight gradient boosting machine-based bagging ensemble model is used to classify and train the fused features. This model uses multiple base learners to vote and employs L2 regularization to prevent overfitting. After training, the output value of the penultimate layer is extracted as the activation vector of the sample, representing the deep abstract representation of the device in the feature space. For each known device class, the mean of the activation vectors of all correctly classified samples is calculated as the class center, i.e., the average activation vector. Subsequently, the Euclidean distance from each sample within the class to the class center is calculated, and a small number of tail samples with the largest distance are selected. A two-parameter Weibull distribution is used to fit this tail distribution to obtain the boundary probability model for each device class. This model controls the distribution steepness through a shape parameter and the distribution width through a scale parameter, which is used to subsequently determine whether the device under test falls within the reasonable distribution range of the known class.

[0040] Furthermore, considering the establishment of a boundary probability model for the target device category based on Euclidean distance data from tail samples, the corresponding processing steps are as follows: Obtain the Euclidean distance values ​​corresponding to each sample point in the tail samples of the target device category, forming a distance dataset to be fitted; use the maximum likelihood estimation method to estimate the shape and scale parameters of the two-parameter Weibull distribution based on the distance dataset to be fitted. The shape parameter determines the morphological characteristics of the distribution curve, and the scale parameter determines the broadening degree of the distribution curve; substitute the estimated shape and scale parameters into the probability density function expression of the Weibull distribution to obtain the Weibull probability density function corresponding to the target device category. The probability density function describes the relative probability of different Euclidean distance values ​​occurring within the target device category; perform an integral operation on the probability density function to obtain the cumulative distribution function corresponding to the target device category. The cumulative distribution function is used to calculate the cumulative probability of any given Euclidean distance value in this distribution; use the probability density function and the cumulative distribution function together as the boundary probability model corresponding to the target device category for subsequent calculation of the confidence probability of test samples belonging to the target device category.

[0041] Specifically, the Euclidean distance values ​​corresponding to each sample point in the tail samples are extracted to form the dataset to be fitted. Then, the maximum likelihood estimation method is used to estimate the parameters of this dataset, solving for the shape parameter (which determines the curve shape) and the scale parameter (which determines the distribution width) in the two-parameter Weibull distribution. The estimated parameters are then substituted into the Weibull probability density function to obtain the probability density curve describing the likelihood of different distance values ​​occurring within the class. Integrating this density function yields the cumulative distribution function, which is used to calculate the cumulative probability of any given distance value in this class distribution. These two functions together constitute the boundary probability model for this class, providing a probabilistic benchmark for calculating the confidence level of subsequent test samples.

[0042] Furthermore, considering the use of validation sets not involved in training to independently determine the optimal decision threshold for each known device class, the corresponding processing steps are as follows: Obtain validation set samples containing annotation information for each known device category. These validation set samples were not involved in the training process of the ensemble classifier. Input the validation set samples into the ensemble classifier and extract the corresponding validation set activation vectors. Calculate the Euclidean distance between the validation set activation vectors and the average activation vectors of each known device category. Substitute the Euclidean distance of the validation set samples into the cumulative distribution function of the corresponding known device category to obtain the preliminary confidence probability that the validation set samples belong to each known device category. For each known device category, within the preset threshold search range, traverse each candidate threshold with a preset step size. For each candidate threshold, calculate the precision and recall of the target device category on the validation set samples.

[0043] Let TP represent the number of samples that are actually positive and correctly predicted as positive, FP represent the number of samples that are actually negative but incorrectly predicted as positive, and FN represent the number of samples that are actually positive but incorrectly predicted as negative. Then the precision calculation formula is: The recall rate is calculated using the following formula: Further calculate the F1 score, which is the harmonic mean of precision and recall. The formula is: ; Select the candidate threshold that maximizes the F1 score as the decision threshold for the target device category; repeat the search process until a decision threshold is determined for each of the known device categories.

[0044] Specifically, the activation vectors of the validation samples are extracted, and their Euclidean distances to the average activation vectors of each class are calculated. These distances are then substituted into the corresponding cumulative distribution function to obtain the initial confidence probability. Subsequently, for each device class, candidate thresholds are iterated within a preset threshold range with a fixed step size. Precision and recall are calculated for each candidate value, and the harmonic mean of these two values, the F1 score, is used as the comprehensive evaluation metric. Finally, the candidate threshold that maximizes the F1 score is selected as the decision threshold specific to that class. This process is repeated until all class thresholds are determined. This ensures that each threshold can adapt to differences in feature distribution within the classification, balancing false negatives and false positives.

[0045] Furthermore, considering the use of the improved OpenMax function to calculate the attribution probability, the corresponding processing steps are as follows: Collect the steady-state I / Q wireless signal of the target device to be identified, and after feature extraction and fusion, input it into the ensemble classifier to extract the corresponding test sample activation vector; calculate the Euclidean distance from the test sample activation vector to the average activation vector of each known device category to obtain a set of test distance values; substitute the test distance values ​​into the cumulative distribution function of the corresponding known device category to obtain the reduction coefficient of each known device category; multiply the probability of each category output by the original SoftMax function by the corresponding reduction coefficient to obtain the reduced probability of the known category. Calculate the difference between the probability of each class output by the original SoftMax function and the reduced probability of the known classes, and sum the differences for each known class to obtain the probability of the unknown device being assigned to a class. Let the total number of known device classes be C, and the original SoftMax output probability for class c be... The reduced probability is The formula for calculating the probability of belonging to an unknown device category is as follows: ; The reduced probabilities of known classes and the probabilities of belonging to unknown device classes are combined to form an expanded probability vector. This expanded probability vector is then normalized so that the sum of all class probabilities is one. The maximum probability value is then found in the normalized expanded probability vector. Let the probability of the k-th class after normalization be... The independent decision threshold corresponding to the k-th class is The final decision rule is: if the highest probability corresponds to a known category and satisfies... If the target device is identified as a known device of type k, then the target device is determined to be an unknown intrusion device, and a first alarm signal is generated.

[0046] Specifically, after feature extraction and fusion of the signal from the device under test, the data is input into a classifier. The activation vector is extracted, and the Euclidean distance to the average activation vector of each class is calculated. Substituting each distance into the Weibull cumulative distribution function of the corresponding class yields a reduction coefficient used to lower the known class probability of the original SoftMax output. The difference between the reduced probability and the original probability is accumulated to obtain the probability of the unknown device. The reduced probabilities of the known class and the unknown class probabilities are merged and normalized to form an expanded probability vector with a sum of one. The class corresponding to the maximum value in the vector is selected. If it is a known class and the probability is not lower than the class-specific decision threshold, it is determined to be a legitimate device; otherwise, it is considered an unknown intrusion device, and the first alarm signal is triggered. Thus, by adjusting the confidence level through distance, known devices and disguised illegal access can be effectively distinguished.

[0047] In some embodiments, step S40 specifically includes the following steps: after acquiring the user's voice signal, pre-emphasis processing is performed on the user's voice signal through a first-order finite impulse response high-pass filter, and the transfer function of the pre-emphasis filter is: , The pre-emphasis coefficient typically ranges from 0.9 to 0.99, where z is a complex frequency domain variable. This indicates the unit delay. This processing is used to compensate for high-frequency energy loss caused by lip radiation effects, thereby improving the signal-to-noise ratio in the high-frequency band.

[0048] The pre-emphasized user speech signal is segmented into short-time frames, and each frame is multiplied by a Hamming window function. Let the frame length be N, and n be the sampling point index within the frame, ranging from zero to N minus 1. Then the expression for the Hamming window function is: When the value exceeds the specified range, the window function value is zero. This process is used to reduce signal discontinuities at frame boundaries, resulting in a windowed frame signal.

[0049] Calculate the short-time energy and short-time zero-crossing rate of the signal in each windowed frame. Assume the speech signal sampling sequence is... , m is the sampling point number, and the window function is If the current frame endpoint is n and the window length is N, then the short-time energy calculation formula is: ; The formula for calculating the short-time zero crossing rate is: ; in, The sign function is defined as: when the independent variable x is less than zero. The value is zero when x is greater than or equal to zero. The value is set to one. A dual-threshold detection method is used to distinguish speech segments from silence segments, remove noise and non-speech regions, and extract the effective speech segments.

[0050] The amplitude spectrum is obtained by performing a Fast Fourier Transform on the effective speech segment. Let the time-domain signal be... If the number of transform points is N and the frequency domain index is k, then the formula for calculating the Fast Fourier Transform is: ; Where k ranges from zero to N minus 1, and j is the imaginary unit. The amplitude spectrum is filtered through a Mel filter bank, and the Mel frequency... The conversion relationship with the linear frequency f is as follows: ; The frequency response functions of each filter in the Mel filter bank exhibit a triangular distribution in the frequency domain. Taking the logarithm of the energy of each frequency band of the filtered output and then performing a discrete cosine transform, let the th... The logarithmic energy output of the Mel filter is If the total number of filters is M, and the required order of the Mel-frequency cepstral coefficients is L (L is usually taken as twelve to sixteen), then the formula for calculating the discrete cosine transform is: ; Where n ranges from one to L. The extracted... These are the Mel-frequency cepstral coefficients, used as speech features.

[0051] The speech features are matched with pre-stored instruction templates. The optimized dynamic time warping algorithm constrains the matching search path to a parallelogram region with slopes of 1 / 2 and 2 respectively. Let the number of input speech feature frames be N and the number of reference template frames be M. When the constraints are satisfied... If the ratio of the number of frames between the two templates is within a reasonable range, otherwise the difference is considered too large and they will not be matched.

[0052] In the cumulative distance calculation, let This represents the local matching distance between the features of the input speech in frame x and the features of the reference template in frame y. Let x represent the cumulative minimum distance from the starting point (1,1) to the current point (x,y). Then the dynamic programming recurrence relation is: ; Cumulative distance calculation is performed only on pre-stored instruction templates that meet the frame count constraint, reducing the amount of computation for invalid matches. The cumulative matching distance between speech features and each pre-stored instruction template is calculated, and the pre-stored instruction template with the smallest cumulative matching distance that is lower than a preset similarity threshold is selected as the matching result, generating the user intent signal corresponding to the matching result.

[0053] Specifically, the speech signal is first pre-emphasized using a high-pass filter to compensate for high-frequency attenuation and improve clarity. Then, it is framed and Hamming windows are applied to smooth the boundaries and prevent spectral leakage. By calculating the short-time energy and zero-crossing rate of each frame, a dual-threshold method is used to eliminate silent segments and extract the effective speech. After performing a Fast Fourier Transform on the effective segments, a Mel filter bank is used to simulate human hearing characteristics. The logarithmic energy is taken and a Discrete Cosine Transform is performed to obtain the Mel frequency cepstral coefficients as compact speech features. In the recognition stage, an optimized dynamic time warping algorithm is used to compare the input features with pre-stored instruction templates. By restricting the search path to a parallelogram interval with a slope of half to half, invalid matching operations are significantly reduced, and the cumulative distance is calculated only for templates whose frame ratio meets the constraints. Finally, the template with the smallest distance and below the similarity threshold is selected, and the corresponding control intention signal is output.

[0054] In some embodiments, step S50 specifically includes the following steps: receiving a first alarm signal, a second alarm signal, and a user control intent signal, and attaching a timestamp to each signal at the time of reception; when both the first alarm signal and the second alarm signal exist simultaneously, determining whether the spatial regions corresponding to the two belong to the same ZigBee subnet coverage area, and whether the difference between their timestamps is less than a preset time window threshold; if the spatial regions are consistent and the time difference is less than the preset time window threshold, then the spatiotemporal correlation verification is deemed successful, and the response level is improved; if only a single alarm signal exists, then an independent response is performed according to the corresponding preset rules. The alarm signal combination or independent alarm signal after spatiotemporal correlation verification is compared with the rule conditions in the preset collaborative rule base. The preset collaborative rule base predefines the logical combination relationship and corresponding linkage action mapping between different alarm signal types, user control intention signals and the current system arming status. According to the successfully matched collaborative rule, the corresponding security linkage instruction is generated. The security linkage instruction includes the target actuator terminal identifier and the action parameters to be executed. The security linkage instruction is wirelessly transmitted to the corresponding actuator terminal through the ZigBee network, driving the actuator terminal to execute the corresponding security response action.

[0055] Specifically, after receiving RF intrusion alarms, environmental anomaly alarms, and voice intent signals, the system first adds a reception timestamp to each signal. When two types of alarms exist simultaneously, the engine verifies whether their spatial sources belong to the same ZigBee subnet and whether the time difference is within a preset window. If they match in both time and space, the association is confirmed and the response level is upgraded; otherwise, it is handled according to the single alarm rule. The verified alarm combination or independent signal is matched against a preset rule base. The rule base defines the logical combination of alarm type, voice intent, and system arming status, as well as the corresponding linkage action. Upon successful matching, a linkage instruction containing the target actuator identifier and action parameters is generated and wirelessly transmitted to the corresponding terminal via the ZigBee network for execution. This reduces false alarms through time-space correlation and achieves multi-source information fusion for security linkage.

[0056] In some embodiments, considering model optimization based on execution feedback, the corresponding processing steps are as follows: After the executor terminal executes the security response action, execution result information is collected, including execution success status, execution failure status, and user confirmation feedback on the execution result; a model optimization feedback signal is generated based on the execution result information; when the execution result information is execution success and user confirmation, a positive feedback signal is generated; when the execution result information is execution failure or user cancellation, a negative feedback signal is generated. When a negative feedback signal related to a known device category is received, a threshold adjustment command is sent to the open-set intrusion detection module. The open-set intrusion detection module lowers the independent decision threshold corresponding to the known device category according to a preset step size to improve the sensitivity of subsequent detection. When positive feedback signals are received multiple times in a row, the independent decision threshold corresponding to the known device category is appropriately increased according to a preset step size to reduce the false alarm rate. When a negative feedback signal related to voice control is received, a weight update instruction is sent to the voice recognition module to reduce the confidence weight of the pre-stored instruction template that caused misrecognition in the matching process; when a positive feedback signal is received, the confidence weight of the corresponding pre-stored instruction template is increased; the adjusted independent decision threshold and the updated confidence weight of the pre-stored instruction template are stored locally as optimized model parameters and loaded for use the next time the system starts.

[0057] Specifically, after the actuator completes the linkage action, the system collects the execution result and user confirmation status, and generates positive and negative feedback signals accordingly. For the open-set recognition module, if negative feedback is received indicating that a legitimate device has been falsely rejected, the system lowers the specific decision threshold for that type by a step size to improve sensitivity; if continuous positive feedback is received indicating stable recognition, the threshold is appropriately increased to suppress false alarms. For the speech recognition module, negative feedback triggers a reduction in the matching confidence weight of the misidentified template, while positive feedback enhances the weight of the correct template. The adjusted threshold and weight parameters are stored locally for loading on the next startup. This mechanism enables the system to continuously learn from operational experience, achieving a synergistic improvement in detection accuracy and user experience.

[0058] In some embodiments, considering the issues of intrusion path prediction and early warning, the corresponding processing steps are as follows: After generating a security linkage command and confirming the occurrence of an intrusion event, record the event attribute information of this intrusion event. The event attribute information includes: intrusion event type, occurrence timestamp, ZigBee subnet area identifier that triggered the alarm, RF fingerprint feature summary of the device identified as an unknown intrusion device, and associated environmental parameter anomaly type. Store the event attribute information in the home security event history database. The home security event history database stores the event attribute information of each security event in chronological order, and each event record contains complete event attribute fields.

[0059] Once a new intrusion event is confirmed, the system searches the home security event history database for historical event records with the same or similar event attributes as the current intrusion event. Search matching criteria include: same intrusion event type, adjacent or identical ZigBee subnet area identifiers triggering alarms, similar occurrence timestamps, or similarity of RFID fingerprint feature summaries exceeding a preset feature similarity threshold. If a matching historical event record is found, information on subsequent related events is extracted from that record. Related event information includes: alarm types and alarm order triggered successively in other ZigBee subnet areas within a preset time window after the historical intrusion event. Based on the extracted related event information, an intrusion path prediction result is generated. The intrusion path prediction result includes: the predicted next ZigBee subnet area identifier to be intruded upon, the predicted intrusion time window, and a list of recommended pre-deployed actuator terminals.

[0060] Based on the intrusion path prediction results, pre-response instructions are sent to the actuator terminals in the next predicted ZigBee subnet area to be intruded upon. These pre-response instructions include: increasing the sampling frequency of the human infrared sensors in the corresponding area; lowering the independent decision thresholds for each type of device in the open-set intrusion detection module of the corresponding area to improve detection sensitivity; and pre-activating the camera modules in the corresponding area to enter a ready-to-trigger state. When the predicted next ZigBee subnet area actually triggers an alarm within the prediction time window, the prediction is confirmed as a hit, and the subsequent linkage response delay is shortened. When the prediction fails, the false alarm information is fed back to the collaborative rule base to optimize the association judgment parameters in the preset spatiotemporal association rules.

[0061] Specifically, this embodiment constructs a historical database of intrusion behavior within the home by recording the spatiotemporal attributes and radio frequency fingerprint characteristics of each intrusion event. When a new intrusion event occurs, the system searches the historical database for similar event patterns to identify the intruder's behavioral patterns and habitual paths. For example, if historical records show that an unknown device first triggered the radio frequency intrusion alarm of the living room ZigBee subnet at night, and then triggered the infrared intrusion alarm of the master bedroom subnet within five minutes, then when the system detects the unknown device again in the living room, it can predict in advance that it will move to the master bedroom next, and proactively increase the sensor sampling frequency and detection sensitivity in the master bedroom area, achieving a technological leap from passive response to proactive early warning. This solution fully utilizes the unforgeability of radio frequency fingerprints to track intrusion devices across areas, and combined with the spatiotemporal correlation capabilities of the environmental perception module, enables the system to learn and predict intrusion behavior based on historical data, significantly enhancing the overall protection effect of home security.

[0062] The implementation principle of the collaborative method for home security models based on wireless communication in this application embodiment is as follows: The system collects steady-state I / Q wireless signals of target devices in the home environment, extracts multi-dimensional handcrafted features through an expert feature channel, and simultaneously extracts time-frequency domain depth features through a one-dimensional convolutional neural network channel. After concatenation, the features are recalibrated by the SE attention module to output a fused feature vector. This fused feature vector is input into an ensemble classifier based on the AutoGluon framework to extract activation vectors. The average activation vector for each known device category is calculated. A boundary probability model is established based on the tail distribution of the distance from the activation vector to the average activation vector by fitting the Weibull distribution. The system independently sets decision thresholds and uses an improved OpenMax function to calculate the attribution probability. When the maximum probability is less than the corresponding independent decision threshold, a first alarm signal is generated. Simultaneously, real-time parameters of the home physical environment are collected through a ZigBee wireless sensor network and compared with a preset security threshold to generate a second alarm signal. User voice signals are collected to extract Mel-frequency cepstral coefficients, and an optimized dynamic time warping algorithm is used to match them with a pre-stored instruction template to generate a user control intent signal. The collaborative decision engine receives these three signals, performs verification and matching based on preset spatiotemporal association rules, generates security linkage instructions, and sends them to the actuator terminal via the ZigBee network. This achieves cross-domain collaboration between wireless signal layer intrusion detection and physical environment perception, effectively identifying unknown intrusion devices disguised as their own, reducing the false alarm rate and missed alarm risk of traditional security systems, and significantly improving the overall reliability of home security.

[0063] Figure 1 This is a flowchart illustrating a collaborative method for a home security model based on wireless communication in one embodiment. It should be understood that, although... Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows; unless explicitly stated otherwise, there is no strict order requirement for the execution of these steps, and they can be executed in other orders; and Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0064] Based on the same technical concept, referring to Figure 2 This application also provides a home security model collaboration device based on wireless communication, which adopts the following technical solution: The device includes: The feature extraction module is used to collect steady-state I / Q wireless signals of target devices in the home environment, extract multi-dimensional handmade features through the expert feature channel, extract time-frequency domain depth features through the one-dimensional convolutional neural network channel, and after splicing, recalibrate through the SE attention module to output a fused feature vector. The classification probability module is used to input the fused feature vector into an ensemble classifier based on the AutoGluon framework, extract the output of the penultimate fully connected layer as the activation vector; calculate the average activation vector based on the activation vectors of known device categories; fit the tail distribution of the distance from each category's activation vector to the corresponding average activation vector based on the Weibull distribution to establish a boundary probability model; calculate the attribution probability using an improved OpenMax function based on the decision threshold set independently for each category of device; if the maximum probability is less than the corresponding decision threshold, generate the first alarm signal. The real-time alarm module is used to collect real-time parameters in the home physical environment through the ZigBee wireless sensor network, compare the real-time parameters with preset safety thresholds, and generate a second alarm signal when any real-time parameter exceeds the corresponding preset safety threshold. The speech signal module is used to acquire user speech signals and perform pre-emphasis, frame windowing and endpoint detection preprocessing, and extract Mel frequency cepstral coefficients as speech features; an optimized dynamic time warping algorithm is used to match the speech features with pre-stored instruction templates to generate corresponding user intent signals; The security linkage module is used to receive the first alarm signal, the second alarm signal, and the user intent signal, perform verification and matching based on preset spatiotemporal association rules, generate security linkage instructions, and send them to the actuator terminal through the ZigBee network to execute the response action.

[0065] In some embodiments, the feature extraction module is specifically used to divide the acquired steady-state I / Q wireless signal into two paths, with the first path input to the expert feature channel and the second path input to the one-dimensional convolutional neural network channel; In the expert feature channel, the in-phase component I and quadrature component Q of the steady-state I / Q wireless signal are calculated separately to extract multi-dimensional handcrafted features, which include: amplitude, phase, real part of mean, imaginary part of mean, real part of standard deviation, imaginary part of standard deviation, maximum value, minimum value, mean absolute deviation, skewness, kurtosis, information entropy, energy, and root mean square. In a one-dimensional convolutional neural network channel, a network structure containing two convolutional layers and two pooling layers is used to perform convolution and pooling processing on steady-state I / Q wireless signals to extract time-frequency domain depth features, and the extraction results are flattened. The extracted multidimensional handcrafted features are concatenated with the flattened time-frequency domain depth features to obtain the initial concatenated features; the initial concatenated features are then input into the SE attention module. The SE attention module compresses the spatial dimension through a global average pooling layer, and then generates channel weights by passing them sequentially through a first fully connected layer, a ReLU activation function, a second fully connected layer, and a Sigmoid activation function. Finally, each channel weight is multiplied by the initial concatenated features channel by channel to output the fused feature vector after weight calibration.

[0066] In some embodiments, the classification probability module is specifically used for the bagging ensemble model of the ensemble classifier based on the AutoGluon framework, which adopts a lightweight gradient boosting machine, to integrate and train multiple base learners through the bagging method, and L2 regularization is used during the training process. The fused feature vector is input into the trained ensemble classifier, and the output value of the second-to-last fully connected layer in the ensemble classifier is extracted as the activation vector of the corresponding sample. For each known device category, collect the activation vectors of all correctly classified samples in the corresponding category, calculate the arithmetic mean of the activation vectors of all samples in that category, and use the arithmetic mean as the average activation vector of that category; For each known device category, calculate the Euclidean distance from the activation vector of each sample within the category to the corresponding average activation vector to obtain the distance set for that category; select a predetermined number of sample points with the largest distance values ​​from the distance set as tail samples; The distance distribution of the tail samples is fitted using a two-parameter Weibull distribution to obtain the Weibull probability density function and cumulative distribution function of the category, which serve as the boundary probability model corresponding to the known device category. The two-parameter Weibull distribution includes a shape parameter that controls the shape of the distribution and a scale parameter that controls the scaling of the distribution.

[0067] In some embodiments, the classification probability module is specifically used to obtain the Euclidean distance values ​​corresponding to each sample point in the tail samples of the target device category, forming a distance dataset to be fitted. The maximum likelihood estimation method is used to estimate the shape and scale parameters of the two-parameter Weibull distribution based on the distance dataset to be fitted. The shape parameter determines the morphological characteristics of the distribution curve, and the scale parameter determines the broadening of the distribution curve. Substituting the estimated shape and scale parameters into the probability density function expression of the Weibull distribution, we obtain the Weibull probability density function corresponding to the target equipment category. The probability density function is used to describe the relative probability of different Euclidean distance values ​​occurring within the target equipment category. Integrating the probability density function yields the cumulative distribution function corresponding to the target device category. The cumulative distribution function is used to calculate the cumulative probability of any given Euclidean distance value in that category distribution. The probability density function and cumulative distribution function are used together as the boundary probability model corresponding to the target device category, which is then used to calculate the confidence probability of the test sample belonging to the target device category.

[0068] In some embodiments, the classification probability module is specifically used to obtain a validation set sample containing labeling information of each known device category, and the validation set sample is not involved in the training process of the ensemble classifier. Input the validation set samples into the ensemble classifier and extract the corresponding validation set activation vectors; Calculate the Euclidean distance from the activation vectors of the validation set to the average activation vectors of each known device class; Substituting the Euclidean distance of the validation set samples into the cumulative distribution function of the corresponding known device categories, we obtain the preliminary confidence probability that the validation set samples belong to each known device category; For each known device category, within the preset threshold search range, each candidate threshold is traversed with a preset step size; for each candidate threshold, the precision and recall of the target device category on the validation set samples are calculated, and the F1 score is further calculated. The F1 score is the harmonic mean of precision and recall. Select the candidate threshold that maximizes the F1 score as the decision threshold for the target device category; repeat the search process until a decision threshold is determined for each of the known device categories.

[0069] In some embodiments, the classification probability module is specifically used to collect the steady-state I / Q wireless signal of the target device to be identified, and input it into the ensemble classifier after feature extraction and fusion to extract the corresponding test sample activation vector; Calculate the Euclidean distance from the activation vector of the test sample to the average activation vector of each known device category to obtain a set of test distance values; Substitute the test distance values ​​into the cumulative distribution function of the corresponding known device category to obtain the reduction coefficient for each known device category; multiply the probability of each category output by the original SoftMax function by the corresponding reduction coefficient to obtain the reduced probability of the known category; Calculate the difference between the probability of each category output by the original SoftMax function and the reduced probability of the known categories, and sum the differences of each known category to obtain the probability of the unknown device belonging to a category; The reduced known category probabilities and the unknown device category attribution probabilities are combined to form an extended probability vector. The extended probability vector is then normalized so that the sum of all category probabilities is one. The maximum probability value is found in the normalized extended probability vector. If the category corresponding to the maximum probability value is a known device category and the maximum probability value is not less than the decision threshold corresponding to the known device category, then the target device to be identified is determined to be the known device category. Otherwise, the target device to be identified is determined to be an unknown intrusion device, and a first alarm signal is generated.

[0070] In some embodiments, the speech signal module is specifically used to collect the user's speech signal and then pre-emphasize the user's speech signal through a first-order finite impulse response high-pass filter to compensate for the high-frequency energy loss caused by the lip radiation effect and improve the high-frequency signal-to-noise ratio. The pre-emphasized user speech signal is divided into short time frames, and each frame signal is multiplied by a Hamming window function to reduce signal discontinuity at frame boundaries, thus obtaining windowed frame signals. Calculate the short-time energy and short-time zero-crossing rate of the signal in each windowed frame, use the dual-threshold detection method to distinguish speech segments from silence segments, remove noise and non-speech regions, and extract the effective speech segments; The effective speech segment is subjected to fast Fourier transform to obtain the amplitude spectrum. The amplitude spectrum is then filtered by a Mel filter bank. The energy of each frequency band of the filtered output is taken as logarithm and then subjected to discrete cosine transform. The Mel frequency cepstral coefficients are extracted as speech features. Speech features are matched with pre-stored instruction templates. The dynamic time warping algorithm is optimized to constrain the matching search path within a parallelogram region with two hypotenuses having slopes of 1 / 2 and 2 respectively. Cumulative distance calculation is performed only on pre-stored instruction templates that meet the frame number constraint, reducing the amount of invalid matching calculation. The frame number constraint is: the ratio of the number of input speech feature frames to the number of template frames is between 1 / 2 and 2. Calculate the cumulative matching distance between the speech features and each pre-stored instruction template, select the pre-stored instruction template with the smallest cumulative matching distance that is lower than the preset similarity threshold as the matching result, and generate the user intent signal corresponding to the matching result.

[0071] In some embodiments, the security linkage module is specifically used to receive a first alarm signal, a second alarm signal, and a user control intent signal, and to add a timestamp information of the time of reception to each signal; When both the first alarm signal and the second alarm signal exist simultaneously, it is determined whether the spatial regions corresponding to the two signals belong to the same ZigBee subnet coverage area, and whether the difference between their timestamps is less than a preset time window threshold. If the spatial regions are the same and the time difference is less than the preset time window threshold, the spatiotemporal correlation verification is deemed to have passed, and the response level is upgraded. If only a single alarm signal exists, an independent response is performed according to the corresponding preset rules. The alarm signal combination or independent alarm signal after being verified by spatiotemporal correlation will be compared with the rule conditions in the preset collaborative rule base; the preset collaborative rule base predefines the logical combination relationship and corresponding linkage action mapping between different alarm signal types, user control intention signals and the current system arming status. Based on the successfully matched collaboration rules, a corresponding security linkage instruction is generated. The security linkage instruction includes the target actuator terminal identifier and the parameters of the action to be executed. Security linkage commands are wirelessly transmitted to the corresponding actuator terminals via the ZigBee network, driving the actuator terminals to perform the corresponding security response actions.

[0072] This application also discloses a control device.

[0073] Specifically, the control device includes a memory and a processor, the memory storing a computer program that can be loaded by the processor and executed to perform the aforementioned wireless communication-based home security model collaboration method.

[0074] This application also discloses a computer-readable storage medium.

[0075] Specifically, the computer-readable storage medium stores a computer program that can be loaded by a processor and executed, such as the aforementioned collaborative method for a home security model based on wireless communication. The computer-readable storage medium includes, for example, various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0076] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A collaborative method for home security models based on wireless communication, characterized in that, include: The system collects steady-state I / Q wireless signals from target devices in a home environment, extracts multi-dimensional handcrafted features through an expert feature channel, extracts time-frequency domain depth features through a one-dimensional convolutional neural network channel, and then splices the features together and recalibrates them through an SE attention module to output a fused feature vector. The fused feature vector is input into an ensemble classifier based on the AutoGluon framework, and the output of the penultimate fully connected layer is extracted as the activation vector; the average activation vector is calculated based on the activation vectors of known device categories. A boundary probability model is established by fitting the tail distribution of the distance from the activation vector of each category to the corresponding average activation vector based on the Weibull distribution; a decision threshold is set independently for each category of device, and the belonging probability is calculated using the improved OpenMax function. If the maximum probability is less than the corresponding decision threshold, a first alarm signal is generated. Real-time parameters in the home physical environment are collected through a ZigBee wireless sensor network. The real-time parameters are compared with preset safety thresholds. When any of the real-time parameters exceeds the corresponding preset safety threshold, a second alarm signal is generated. The user's speech signal is acquired and preprocessed by pre-emphasis, frame windowing and endpoint detection, and Mel frequency cepstral coefficients are extracted as speech features. An optimized dynamic time warping algorithm is used to match the speech features with a pre-stored instruction template to generate a corresponding user intent signal; The system receives the first alarm signal, the second alarm signal, and the user intent signal, performs verification and matching based on preset spatiotemporal association rules, generates a security linkage command, and sends it to the actuator terminal through the ZigBee network to execute the response action.

2. The collaborative method for a home security model based on wireless communication according to claim 1, characterized in that, The process involves extracting multidimensional handcrafted features through an expert feature channel, extracting time-frequency domain depth features through a one-dimensional convolutional neural network channel, concatenating them, recalibrating them via an SE attention module, and outputting a fused feature vector, including: The acquired steady-state I / Q wireless signal is divided into two paths: the first path is input to the expert feature channel, and the second path is input to the one-dimensional convolutional neural network channel. In the expert feature channel, the in-phase component I and quadrature component Q of the steady-state I / Q wireless signal are calculated respectively to extract the multidimensional handcrafted features, which include: amplitude, phase, real part of mean, imaginary part of mean, real part of standard deviation, imaginary part of standard deviation, maximum value, minimum value, mean absolute deviation, skewness, kurtosis, information entropy, energy, and root mean square. In a one-dimensional convolutional neural network channel, a network structure containing two convolutional layers and two pooling layers is used to perform convolution and pooling processing on steady-state I / Q wireless signals to extract the time-frequency domain depth features, and the extraction results are flattened. The extracted multidimensional handcrafted features are concatenated with the flattened time-frequency domain depth features to obtain initial concatenated features; the initial concatenated features are then input into the SE attention module. The SE attention module compresses the spatial dimension through a global average pooling layer, and then generates channel weights by passing them sequentially through a first fully connected layer, a ReLU activation function, a second fully connected layer, and a Sigmoid activation function. Finally, each channel weight is multiplied by the initial concatenated feature channel by channel to output the fused feature vector after weight calibration.

3. The collaborative method for a home security model based on wireless communication according to claim 1, characterized in that, The fused feature vector is input into an ensemble classifier based on the AutoGluon framework, and the output of the penultimate fully connected layer is extracted as the activation vector; the average activation vector is calculated based on the activation vectors of known device categories. A boundary probability model is established based on the tail distribution of the distances from each class's activation vector to its corresponding average activation vector, fitted using the Weibull distribution. This model includes: The ensemble classifier based on the AutoGluon framework adopts a bagging ensemble model with a lightweight gradient boosting machine. It integrates and trains multiple base learners using the bagging method, and L2 regularization is used during the training process. The fused feature vector is input into the trained ensemble classifier, and the output value of the second-to-last fully connected layer in the ensemble classifier is extracted as the activation vector of the corresponding sample. For each known device category, collect the activation vectors of all correctly classified samples in the corresponding category, calculate the arithmetic mean of the activation vectors of all samples in that category, and use the arithmetic mean as the average activation vector of that category; For each known device category, calculate the Euclidean distance from the activation vector of each sample within the category to the corresponding average activation vector to obtain the distance set for that category; select a predetermined number of sample points with the largest distance values ​​from the distance set as tail samples; The distance distribution of the tail samples is fitted using a two-parameter Weibull distribution to obtain the Weibull probability density function and cumulative distribution function of the category, which serve as the boundary probability model corresponding to the known device category. The two-parameter Weibull distribution includes a shape parameter that controls the shape of the distribution and a scale parameter that controls the scaling of the distribution.

4. The collaborative method for a home security model based on wireless communication according to claim 3, characterized in that, The method of fitting the distance distribution of the tail samples using a two-parameter Weibull distribution to obtain the Weibull probability density function and cumulative distribution function of the category serves as the boundary probability model corresponding to the known device category, including: Obtain the Euclidean distance values ​​corresponding to each sample point in the tail samples of the target device category to form a distance dataset to be fitted; The maximum likelihood estimation method is used to estimate the shape parameter and scale parameter of the two-parameter Weibull distribution based on the distance dataset to be fitted. The shape parameter determines the morphological characteristics of the distribution curve, and the scale parameter determines the broadening degree of the distribution curve. Substituting the estimated shape and scale parameters into the probability density function expression of the Weibull distribution, we obtain the Weibull probability density function corresponding to the target equipment category. The probability density function is used to describe the relative probability of different Euclidean distance values ​​occurring within the target equipment category. Integrating the probability density function yields the cumulative distribution function corresponding to the target device category. The cumulative distribution function is used to calculate the cumulative probability of any given Euclidean distance value in that category distribution. The probability density function and the cumulative distribution function are used together as the boundary probability model corresponding to the target device category, which is then used to calculate the confidence probability that the test sample belongs to the target device category.

5. The collaborative method for a home security model based on wireless communication according to claim 4, characterized in that, The decision thresholds set independently for each type of device include: Obtain a validation set sample containing labeling information for each known device category, wherein the validation set sample was not involved in the training process of the ensemble classifier; The validation set samples are input into the ensemble classifier to extract the corresponding validation set activation vectors. Calculate the Euclidean distance from the activation vectors of the validation set to the average activation vectors of each known device category; Substituting the Euclidean distance of the validation set samples into the cumulative distribution function of the corresponding known device categories, we obtain the preliminary confidence probability that the validation set samples belong to each known device category; For each known device category, within the preset threshold search range, each candidate threshold is traversed with a preset step size; for each candidate threshold, the precision and recall of the target device category on the validation set samples are calculated, and the F1 score is further calculated, which is the harmonic mean of precision and recall. Select the candidate threshold that maximizes the F1 score as the decision threshold corresponding to the target device category; repeat the search process until the decision threshold corresponding to each of the known device categories is determined.

6. The collaborative method for a home security model based on wireless communication according to claim 5, characterized in that, The calculation of the attribution probability using the improved OpenMax function includes: The steady-state I / Q wireless signal of the target device to be identified is collected, and after feature extraction and fusion, it is input into the ensemble classifier to extract the corresponding test sample activation vector; Calculate the Euclidean distance from the activation vector of the test sample to the average activation vector of each known device category to obtain a set of test distance values; Substitute the test distance values ​​into the cumulative distribution function of the corresponding known device category to obtain the reduction coefficient for each known device category; multiply the probability of each category output by the original SoftMax function by the corresponding reduction coefficient to obtain the reduced probability of the known category; Calculate the difference between the probability of each category output by the original SoftMax function and the reduced probability of the known categories, and sum the differences of each known category to obtain the probability of the unknown device belonging to a category; The reduced known category probabilities and the unknown device category attribution probabilities are combined to form an extended probability vector. The extended probability vector is then normalized so that the sum of all category probabilities is one. The maximum probability value is found in the normalized extended probability vector; if the category corresponding to the maximum probability value is a known device category and the maximum probability value is not less than the decision threshold corresponding to the known device category, then the target device to be identified is determined to be the known device category; otherwise, the target device to be identified is determined to be an unknown intrusion device, and the first alarm signal is generated.

7. The collaborative method for a home security model based on wireless communication according to claim 1, characterized in that, The process involves collecting user speech signals and performing pre-emphasis, frame windowing, and endpoint detection preprocessing, and extracting Mel frequency cepstral coefficients as speech features. An optimized dynamic time warping algorithm is used to match the speech features with a pre-stored instruction template to generate a corresponding user intent signal, including: After acquiring the user's voice signal, the user's voice signal is pre-emphasized through a first-order finite impulse response high-pass filter to compensate for the high-frequency energy loss caused by the lip radiation effect and improve the signal-to-noise ratio in the high-frequency band. The pre-emphasized user speech signal is divided into short time frames, and each frame signal is multiplied by a Hamming window function to reduce signal discontinuity at frame boundaries, thus obtaining windowed frame signals. Calculate the short-time energy and short-time zero-crossing rate of each windowed frame signal, use the dual-threshold detection method to distinguish speech segments from silence segments, remove noise and non-speech regions, and extract effective speech segments; The effective speech segment is subjected to a fast Fourier transform to obtain an amplitude spectrum. The amplitude spectrum is then filtered through a Mel filter bank. The energy of each frequency band of the filtered output is taken as a logarithm and then subjected to a discrete cosine transform. The Mel frequency cepstral coefficients are extracted as the speech features. The speech features are matched with pre-stored instruction templates. The optimized dynamic time warping algorithm constrains the matching search path to two parallelogram regions with slopes of 1 / 2 and 2 respectively. The cumulative distance is calculated only for pre-stored instruction templates that meet the frame number constraint, thereby reducing the amount of invalid matching calculations. The frame number constraint is that the ratio of the number of input speech feature frames to the number of template frames is between 1 / 2 and 2. Calculate the cumulative matching distance between the speech features and each pre-stored instruction template, select the pre-stored instruction template with the smallest cumulative matching distance that is lower than a preset similarity threshold as the matching result, and generate the user intent signal corresponding to the matching result.

8. The collaborative method for a home security model based on wireless communication according to claim 7, characterized in that, The process of receiving the first alarm signal, the second alarm signal, and the user intent signal, performing verification and matching based on preset spatiotemporal association rules, and generating security linkage instructions includes: Receive the first alarm signal, the second alarm signal, and the user control intent signal, and add a timestamp of the time of reception to each signal; When both the first alarm signal and the second alarm signal exist simultaneously, it is determined whether the spatial regions corresponding to the two belong to the coverage area of ​​the same ZigBee subnet, and whether the difference between their timestamps is less than a preset time window threshold. If the spatial regions are consistent and the time difference is less than the preset time window threshold, the spatiotemporal correlation verification is deemed to have passed, and the response level is improved. If only a single alarm signal exists, an independent response is performed according to the corresponding preset rules. The alarm signal combination or independent alarm signal after being verified by spatiotemporal correlation will be compared with the rule conditions in the preset collaborative rule base; the preset collaborative rule base predefines the logical combination relationship and corresponding linkage action mapping between different alarm signal types, user control intention signals and the current system arming status. Based on the successfully matched collaboration rules, a corresponding security linkage instruction is generated, which includes the target actuator terminal identifier and the action parameters to be executed. The security linkage command is wirelessly transmitted to the corresponding actuator terminal via a ZigBee network, driving the actuator terminal to perform the corresponding security response action.

9. A collaborative device for a home security model based on wireless communication, characterized in that, The device includes: The feature extraction module is used to collect steady-state I / Q wireless signals of target devices in the home environment, extract multi-dimensional handmade features through the expert feature channel, extract time-frequency domain depth features through the one-dimensional convolutional neural network channel, and after splicing, recalibrate through the SE attention module to output a fused feature vector. The classification probability module is used to input the fused feature vector into an ensemble classifier based on the AutoGluon framework, extract the output of the penultimate fully connected layer as the activation vector; calculate the average activation vector based on the activation vectors of known device categories; fit the tail distribution of the distance from each category's activation vector to the corresponding average activation vector based on the Weibull distribution to establish a boundary probability model; calculate the attribution probability using an improved OpenMax function based on a decision threshold set independently for each category of device; and generate a first alarm signal if the maximum probability is less than the corresponding decision threshold. The real-time alarm module is used to collect real-time parameters in the home physical environment through a ZigBee wireless sensor network, compare the real-time parameters with a preset safety threshold, and generate a second alarm signal when any of the real-time parameters exceeds the corresponding preset safety threshold. The speech signal module is used to acquire user speech signals and perform pre-emphasis, frame windowing and endpoint detection preprocessing, and extract Mel frequency cepstral coefficients as speech features; the speech features are matched with pre-stored instruction templates using an optimized dynamic time warping algorithm to generate corresponding user intent signals; The security linkage module is used to receive the first alarm signal, the second alarm signal and the user intent signal, perform verification and matching based on preset spatiotemporal association rules, generate security linkage instructions, and send them to the actuator terminal through the ZigBee network to execute the response action.

10. A control device, characterized in that, The device includes: A memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in any one of claims 1 to 8.

Images