Personnel state detection system and method based on wifi router

Abstract

The invention discloses a personnel state detection system and method based on a wifi router, and the method comprises the following steps: S1, deploying the wifi router, collecting subcarrier CSI which changes due to the disturbance of the movement of personnel, and forming a CSI signal sequence; s2, configuring a neural network classifier and loading parameters; s3, preprocessing the CSI signal sequence to obtain a stable CSI time sequence signal; s4, constructing a graph structure taking the subcarriers as graph nodes, and generating a spectrogram; s5, performing Laplace feature transformation on the spectrogram, and extracting a frequency domain feature vector of the spectrogram; s6, constructing a multi-scale Sheng differential equation model, and generating a state trajectory under each time scale; and S7, splicing the state tracks to form a state evolution sequence, inputting the state evolution sequence into a neural network classifier, and generating a state recognition result. According to the method, CSI is used for modeling personnel disturbance, multi-scale dynamic representation and accurate recognition are achieved, and the method is suitable for a non-contact personnel state sensing scene.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless sensing, and particularly to a personnel status detection system and method based on a wifi router. Background Art

[0002] With the continuous development of wireless communication and sensing technologies, non-contact personnel behavior recognition and status detection based on wifi signals have gradually become a research hotspot in the field of intelligent sensing. In an indoor environment, since people will cause occlusion, reflection, scattering, and multipath changes to the wifi signal propagation path during activities, wifi channel state information (CSI), as a fine-grained physical layer feature, can capture the perturbation changes caused by personnel activities to the wireless channel to a certain extent, providing a possible path for realizing non-visual and non-contact personnel status perception.

[0003] Traditional personnel status detection methods mostly rely on devices such as cameras, infrared sensors, and pressure pads. Although these methods have certain detection capabilities, they are often restricted by factors such as privacy protection, environmental occlusion, installation costs, and monitoring ranges. In contrast, the wifi-based detection method does not rely on additional sensing hardware and can use existing communication infrastructure to achieve wide-coverage passive status recognition, showing strong adaptability and practicality in scenarios such as home care, intelligent security, medical assistance, and behavior analysis.

[0004] Currently, some research works have attempted to use CSI signals for personnel status recognition. A relatively common approach is to directly use CSI amplitude or phase data as time series signals and input them into traditional convolutional neural networks (CNNs) or long short-term memory networks (LSTMs), and achieve state classification through feature learning. Although these methods have achieved acceptable recognition accuracies in certain scenarios, they have technical limitations in multiple aspects. First, directly using CSI time series signals ignores the signal spatial structure relationship between subcarriers and it is difficult to accurately model the influence of personnel activities on the cooperative perturbation of multi-channel signals. Second, traditional deep learning models generally adopt a fixed structure and a fixed time scale for feature extraction, and cannot effectively capture the multi-scale dynamic change laws during the evolution of personnel status. In addition, there are information redundancy and low-frequency information generalization problems in the feature modeling process of some models, and they cannot effectively distinguish the fine-grained differences between different states, especially showing weakness in boundary states such as falling, turning, and standing up.

[0005] Regarding the extraction of spatial features of CSI signals, there are also current studies that introduce graph neural networks (GNNs) to model CSI data. However, most methods are limited to constructing adjacency graphs based on amplitude changes and lack joint modeling of temporal perturbation features. In addition, in existing methods, spectral graph modeling usually does not combine frequency-domain energy distribution for effective feature compression, resulting in an increase in redundant features and low computational efficiency. On the other hand, in terms of dynamic modeling, neural ordinary differential equations, as a new type of continuous deep learning model, are used to represent the system state evolving over time. However, in the scenario of human state recognition, their ability to model dynamic structures at multiple time scales has not been fully explored, and there is also a lack of a fusion modeling mechanism with graph frequency-domain features.

[0006] In summary, there are still many problems in the existing methods for human state recognition based on wifi signals in aspects such as signal structure modeling, dynamic evolution expression, and state classification accuracy. Especially in dealing with the spatial dependence relationship between high-dimensional subcarriers, mining the frequency-domain perturbation features caused by human activities, and modeling the continuous time changes of state evolution, there is still a lack of a systematic method that integrates graph modeling, frequency-domain analysis, and multi-scale dynamic representation.

[0007] Therefore, how to provide a human state detection system and method based on a wifi router is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0008] An object of the present invention is to propose a human state detection system and method based on a wifi router. The present invention integrates CSI channel state information modeling, spectral graph frequency-domain feature extraction, multi-scale neural ordinary differential equation modeling, and deep neural network classification technologies. By constructing a stable CSI graph structure, performing spectral graph transformation and frequency-domain energy compression, and designing a multi-time scale state trajectory generation mechanism, continuous modeling and classification recognition of the dynamic evolution process of humans are realized. This method fully mines the channel perturbation features caused by human movement, avoids the privacy risks and deployment costs brought by cameras and wearable sensors, and has the advantages of fine structure modeling, continuous state representation, and high classification accuracy, and is suitable for non-contact indoor state perception and behavior recognition scenarios.

[0009] According to an embodiment of the present invention, a human state detection system and method based on a wifi router includes the following steps:

[0010] S1. Deploy a wifi router in the target area, and collect subcarrier CSI that changes due to the perturbation caused by human movement to form a CSI signal sequence containing amplitude information and phase information;

[0011] S2. Configure a neural network classifier and load the pre-constructed parameters;

[0012] S3. Preprocess the CSI signal sequence to generate a stable CSI timing signal;

[0013] S4. Based on the stable CSI timing signal, construct a graph structure, use each sub - carrier channel as a graph node, and set edge weights according to the correlation between signals to generate a spectral graph structure;

[0014] S5. Perform Laplace eigen - transformation on the spectral graph structure to extract graph frequency - domain feature vectors;

[0015] S6. Construct a multi - scale neural ordinary differential equation model, use the graph frequency - domain feature vectors as input, set micro - sub - structures with different time scales, and generate state trajectories representing the dynamic evolution of personnel at each time scale;

[0016] S7. Concatenate the state trajectories generated at each time scale to construct a state evolution sequence, and input the state evolution sequence into a neural network classifier. Use the loaded parameters to discriminate the state evolution sequence and generate a personnel state recognition result.

[0017] Optionally, the parameters include weight matrices, bias vectors, and activation function types of each connection structure from the input layer to the output layer.

[0018] Optionally, the preprocessing includes performing moving average filtering, band - pass filtering, and phase calibration operations.

[0019] Optionally, the personnel state recognition result includes at least one of a stationary state, a walking state, a falling state, a sitting state, a standing - up state, and a turning state.

[0020] Optionally, the specific content of S1 includes:

[0021] S11. Deploy at least two static wifi router nodes at and inside the boundary of the target area, and set the router operating frequency band and sub - carrier configuration parameters;

[0022] S12. Establish a data sending and receiving link for the wifi channel, set the CSI acquisition period, and obtain CSI measurement values of each sub - carrier in the physical layer down - link;

[0023] S13. Collect wifi signals under personnel activity conditions, record channel disturbances such as multipath fading, occlusion, and scattering caused by personnel movement, and form a sequence of time - series CSI raw data frames;

[0024] S14. Extract the amplitude information and phase information of each sub - carrier from each CSI data frame, and represent them as a complex modulus - value and argument pair;

[0025] S15. Combine the amplitude information and phase information in the order of acquisition time to form a CSI signal sequence containing timestamp marks.

[0026] Optionally, S4 specifically includes:

[0027] S41. Divide the stable CSI timing signal by sub - carrier channels, and map each sub - carrier channel to a graph node;

[0028] S42. Set a time window with a fixed length, perform statistical analysis on the CSI timing signal corresponding to each graph node, and extract a feature vector containing the mean amplitude, phase standard deviation, and signal energy;

[0029] S43. Calculate the edge connection weight between each pair of graph nodes through the Gaussian kernel function:

[0030]

[0031] where w ij represents the edge weight between the i - th node and the j - th node, v i represents the statistical feature vector of node i, v j represents the statistical feature vector of node j, ||·||2 represents the Euclidean distance function, and σ represents the bandwidth parameter;

[0032] S44. Combine all edge weights to form an adjacency matrix. The adjacency matrix is a symmetric real - number matrix, and the matrix elements are the edge weight values between each graph node;

[0033] S45. Calculate the degree value of the graph node according to the adjacency matrix, and construct a degree matrix. The degree matrix is a diagonal matrix, and the main diagonal elements are the sum of the edge weights connected to each node;

[0034] S46. Construct a symmetric normalized Laplacian matrix according to the adjacency matrix and the degree matrix:

[0035]

[0036] where L represents the symmetric normalized Laplacian matrix, I represents the identity matrix, D represents the degree matrix, and A represents the adjacency matrix;

[0037] S47. Use the symmetric normalized Laplacian matrix as the representation form of the spectral graph structure.

[0038] Optionally, S5 specifically includes:

[0039] S51. Combine the mean amplitude, phase standard deviation, and signal energy of each sub - carrier channel in the stable CSI timing signal to form a feature vector, and construct a graph signal vector. The dimension of the graph signal vector is the same as the number of sub - carrier channels;

[0040] S52. Construct a symmetric normalized Laplacian matrix based on the signal correlation between subcarrier channels, perform eigenvalue decomposition, extract all eigenvalues and corresponding eigenvectors, and construct a frequency-domain transformation basis matrix;

[0041] S53. Define a spectral convolution kernel function in the graph Laplacian frequency domain to obtain the graph frequency-domain eigenvector:

[0042]

[0043] where \(f\) represents the graph frequency-domain eigenvector, \(\alpha\) i represents the projection coefficient of the graph signal in the \(i\)-th frequency direction, \(u\) i represents the \(i\)-th eigenvector, \(\lambda\) i represents the corresponding eigenvalue, \(\lambda\) max represents the maximum value among all eigenvalues, \(\theta\) j represents the \(j\)-th order parameter of the spectral filter, \(T\) j represents the \(j\)-th order Chebyshev polynomial, \(K\) represents the order of the filter polynomial, and \(k\) represents the number of frequency truncations;

[0044] S54. Evaluate the spectral distribution of the graph frequency-domain eigenvector and construct a frequency energy concentration function:

[0045]

[0046] where \(\rho\) represents the weighted energy concentration of the frequency-domain feature, \(f\) i represents the \(i\)-th component of the graph frequency-domain eigenvector, \(w\) i represents the frequency weight corresponding to the \(i\)-th frequency;

[0047] S55. According to the calculation result of the spectral concentration function, extract the first \(k\) frequency components that meet the conditions from the graph frequency-domain eigenvector to form the final graph frequency-domain eigenvector.

[0048] Optionally, the specific content of S6 includes:

[0049] S61. Set a set of time scales, each set of time scales corresponds to a set of state trajectory generation branches, and each set of branches is divided using different time interval lengths and sampling resolutions;

[0050] S62. Construct a multi-scale neural ordinary differential equation model. The multi-scale neural ordinary differential equation model uses the graph frequency-domain eigenvector as the initial state and establishes a differential relationship between the states changing with time:

[0051]

[0052] where \(h\) (m) (t) represents the state representation vector at time \(t\) under the \(m\)-th time scale, denotes the output mapping weight matrix, denotes the input mapping weight matrix, b (m) denotes the bias vector, η (m) denotes the amplitude coefficient of the modulation term, ψ (m) denotes the angular frequency of the modulation term, and tanh denotes the hyperbolic tangent activation function;

[0053] S63. Set the start and end times of integration for each time scale, take the graph frequency domain feature vector as the initial state and input it into the differential model, and use the numerical integration method to solve the state within the specified time interval;

[0054] S64. Establish a state trajectory sampling mechanism, set a fixed sampling interval for each group of time scales, and perform non-linear integral combination on the state results at each sampling moment:

[0055]

[0056] where, denotes the state trajectory generated under the m-th time scale, denotes the r-th order non-linear combination weight matrix, σ denotes the activation function, denotes the state integral from the start time to the j-th sampling point, K (m) denotes the number of sampling points, and R denotes the number of combination terms;

[0057] S65. Generate a complete state trajectory sequence under each time scale. The state trajectory sequence is composed of the integral transformation results of each sampling point and has a fixed number of time steps and state vector dimensions;

[0058] S66. Save the state trajectories generated under each time scale separately to form the state representation results of multi-scale modeling.

[0059] Optionally, the S7 specifically includes:

[0060] S71. Set a neural network classifier. The structure of the neural network classifier consists of an input layer, a first hidden layer, a second hidden layer, and an output layer. The input layer receives the fixed-length vector converted from the state evolution sequence. The first and second hidden layers are fully connected structures, and the output layer uses normalized mapping to output the state probability distribution;

[0061] S72. Convert the state evolution sequence into a fixed-dimensional vector and input it into the neural network classifier, and sequentially perform feature mapping, bias superposition, and non-linear transformation operations to form a multi-layer feature representation;

[0062] S73. Calculate the hidden representation in the first hidden layer, combining the main weight path and the modulation branch:

[0063] z1 = σ(W1·x + b1 + A1·tanh(U1·x + c1));

[0064] Among them, z1 represents the output vector of the first hidden layer, x represents the input vector after the transformation of the state evolution sequence, W1 represents the weight matrix of the main mapping path, b1 represents the bias vector of the main path, U1 represents the weight matrix of the modulation path, c1 represents the bias vector of the modulation path, A1 represents the output scaling matrix of the modulation path, σ represents the activation function of the first hidden layer, and tanh represents the hyperbolic tangent function;

[0065] S74. Send the output of the first hidden layer to the second hidden layer and perform a double non-linear cross-mapping operation:

[0066] z2 = ReLU(W2·z1 + b2 + γ·sin(U2·z1 + c2));

[0067] Among them, z2 represents the output vector of the second hidden layer, W2 represents the main weight matrix of the second layer, b2 represents the bias vector, U2 represents the weight matrix of the modulation path, c2 represents the bias of the modulation path, γ represents the modulation ratio factor, ReLU represents the rectified linear unit function, and sin represents the modulation function;

[0068] S75. Send the output of the second hidden layer to the output layer, perform a normalization activation operation, and generate the probability distribution of the personnel state classification;

[0069] S76. Determine the final personnel state recognition result according to the maximum value index in the probability distribution.

[0070] A personnel state detection system based on a wifi router according to an embodiment of the present invention includes:

[0071] A wifi signal acquisition module, which is used to deploy a wifi router in the target area, collect the CSI of the subcarrier, and form a CSI signal sequence including amplitude information and phase information;

[0072] A neural network configuration module, which is used to configure a neural network classifier and load model parameters, and the model parameters include a weight matrix, a bias vector, and an activation function type;

[0073] A CSI signal preprocessing module, which is used to perform a moving average filter, a band-pass filter, and a phase calibration operation on the CSI signal sequence to generate a stable CSI time series signal;

[0074] A graph structure construction module, which is used to divide the stable CSI time series signal into graph nodes according to the subcarrier channels, calculate the feature vectors based on the amplitude mean, phase standard deviation, and signal energy of the signals corresponding to each node, calculate the edge weights using a Gaussian kernel function, generate an adjacency matrix and a degree matrix, and construct a symmetric normalized Laplacian matrix;

[0075] The frequency-domain feature extraction module is used to jointly model the graph signal vector and the symmetric normalized Laplacian matrix, perform frequency-domain transformation, construct a spectral convolution kernel function, calculate the spectral concentration function, and select several frequency components in the graph frequency domain to form a graph frequency-domain feature vector;

[0076] The neural ordinary differential modeling module is used to construct a multi-scale neural ordinary differential equation model, take the graph frequency-domain feature vector as the initial state, set different time scales and construct a differential equation system with modulation terms at each time scale, and use numerical integration methods to generate corresponding state trajectories to obtain a state trajectory sequence at each time scale;

[0077] The state recognition module is used to receive the state trajectories generated at each time scale, input them into the neural network classifier in turn, perform multi-layer mapping and non-linear function processing, output the state classification probability distribution, and generate a state recognition result according to the index corresponding to the maximum probability.

[0078] The beneficial effects of the present invention are:

[0079] Aiming at the problems of insufficient channel structure modeling, lag in dynamic change response, and limited state discrimination accuracy in the existing wifi-based personnel state recognition method, the present invention proposes a complete detection process that integrates graph modeling, frequency-domain feature extraction, and multi-scale neural ordinary differential modeling. By introducing the correlation between CSI subcarriers to construct a spectral graph structure, the spatial resolution ability of the channel changes caused by personnel movement disturbances is improved. Using the Laplacian eigen-transformation and spectral concentration evaluation mechanism, representative graph frequency-domain features are extracted, effectively compressing redundant feature information. Further, by constructing a multi-time-scale neural ordinary differential equation model, the evolution trajectory of the graph frequency-domain features in the continuous time domain is modeled, enhancing the model's expression ability for various state evolution modes. Finally, by setting the neural network classification structure, state category output is realized based on the complete state trajectory sequence, improving the discrimination accuracy for complex states such as falling, standing up, and turning. Compared with the existing methods, while retaining the advantages of non-contact perception, the present invention takes into account the modeling depth and discrimination accuracy, providing higher adaptability and stability for continuous personnel state detection in practical scenarios. Description of the Drawings

[0080] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0081] Figure 1 is a flowchart of a method for detecting personnel state based on a wifi router proposed by the present invention;

[0082] Figure 2 It is the processing structure diagram for constructing a graph structure and performing spectrogram modeling on the stable CSI signal in the present invention;

[0083] Figure 3 It is the schematic diagram of the modeling structure for generating the personnel state trajectory based on the multi-scale neural ordinary differential equation model in the present invention. Detailed implementation manners

[0084] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0085] Refer to Figures 1-3 , a personnel state detection system and method based on a wifi router, including the following steps:

[0086] S1. Deploy a wifi router in the target area, collect the subcarrier CSI that changes due to the disturbance caused by personnel movement, and form a CSI signal sequence containing amplitude information and phase information;

[0087] S2. Configure a neural network classifier and load the pre-constructed parameters;

[0088] S3. Preprocess the CSI signal sequence to generate a stable CSI time series signal;

[0089] S4. Based on the stable CSI time series signal, construct a graph structure, use each subcarrier channel as a graph node, and set the edge weights according to the correlation between signals to generate a spectrogram structure;

[0090] S5. Perform Laplace eigen-transformation on the spectrogram structure to extract the graph frequency domain feature vectors;

[0091] S6. Construct a multi-scale neural ordinary differential equation model, use the graph frequency domain feature vectors as the input, set the differential sub-structures with different time scales, and generate the state trajectories representing the dynamic evolution of personnel at each time scale;

[0092] S7. Concatenate the state trajectories generated at each time scale to construct a state evolution sequence, input the state evolution sequence into the neural network classifier, and use the loaded parameters to discriminate the state evolution sequence to generate the personnel state recognition result.

[0093] The present invention collects the CSI signal that changes due to the disturbance caused by personnel movement, constructs the whole process from data acquisition to state recognition, and can realize the non-contact detection of personnel state, avoiding the influence on privacy and use convenience caused by methods such as cameras and wearable devices.

[0094] In this embodiment, the parameters include the weight matrix, bias vector, and activation function type of each connection structure from the input layer to the output layer.

[0095] The present invention clarifies the types of parameters loaded by the neural network classifier, including the weight matrix, bias vector, and activation function type, making the classification model have a clear structure and controllable parameters in actual deployment, and facilitating the rapid construction of an inference model in different application environments.

[0096] In this embodiment, the preprocessing includes performing a moving average filter, a band-pass filter, and a phase calibration operation.

[0097] The present invention uses a moving average filter, a band-pass filter, and a phase calibration operation to effectively preprocess the original CSI signal, significantly suppressing environmental noise, frequency band spurs, and phase drift interference, and improving the stability and modeling quality of CSI data.

[0098] In this embodiment, the personnel status recognition result includes at least one of a stationary state, a walking state, a falling state, a sitting state, a standing-up state, and a turning state.

[0099] The present invention supports the recognition of typical human states such as stationary, walking, falling, sitting, standing up, and turning, covering the main activity patterns in common scenarios such as home, office, and medical care, and expanding the application scope of the status recognition system in the actual environment.

[0100] In this embodiment, S1 specifically includes:

[0101] S11. Deploy at least two static wifi router nodes at and inside the boundary of the target area, and set the router operating frequency band and subcarrier configuration parameters;

[0102] S12. Establish a data sending and receiving link for the wifi channel, set the CSI acquisition period, and obtain the CSI measurement values of each subcarrier in the physical layer downlink;

[0103] S13. Collect wifi signals under the condition of personnel activities, record channel disturbances such as multipath fading, occlusion, and scattering caused by personnel movement, and form a sequence of time-series CSI raw data frames;

[0104] S14. Extract the amplitude information and phase information of each subcarrier from each CSI data frame, and represent them as a complex modulus value and argument pair;

[0105] S15. Combine the amplitude information and phase information in the order of acquisition time to form a CSI signal sequence including timestamp marks.

[0106] The present invention establishes a complete CSI data acquisition process, covering router node layout, channel configuration, link establishment, data extraction, and time stamping, ensuring the physical layer integrity and the temporal correlation with human disturbances of each frame of data.

[0107] In this embodiment, step S4 specifically includes:

[0108] S41. Divide the stable CSI timing signal according to subcarrier channels, and map each subcarrier channel to a graph node;

[0109] S42. Set a time window with a fixed length, perform statistical analysis on the CSI timing signal corresponding to each graph node, and extract a feature vector including the amplitude mean, phase standard deviation, and signal energy;

[0110] S43. Calculate the edge connection weight between each pair of graph nodes through a Gaussian kernel function:

[0111]

[0112] where w ij represents the edge weight between the i-th node and the j-th node, v i represents the statistical feature vector of node i, v j represents the statistical feature vector of node j, ||·||2 represents the Euclidean distance function, and σ represents the bandwidth parameter;

[0113] S44. Combine all edge weights to form an adjacency matrix. The adjacency matrix is a symmetric real matrix, and the matrix elements are the edge weight values between each graph node;

[0114] S45. Calculate the degree value of the graph node according to the adjacency matrix and construct a degree matrix. The degree matrix is a diagonal matrix, and the main diagonal elements are the sum of the edge weights connected to each node;

[0115] S46. Construct a symmetric normalized Laplacian matrix according to the adjacency matrix and the degree matrix:

[0116]

[0117] where L represents the symmetric normalized Laplacian matrix, I represents the identity matrix, D represents the degree matrix, and A represents the adjacency matrix;

[0118] S47. Use the symmetric normalized Laplacian matrix as the representation form of the spectral graph structure.

[0119] The present invention constructs a spectral graph structure with CSI signals, combines statistical feature extraction and edge weight construction algorithms, and uses the symmetric normalized Laplacian matrix to express the topological relationship between subcarriers, accurately modeling the propagation mode of human disturbances in the spatial structure.

[0120] In this embodiment, S5 specifically includes:

[0121] S51. Form a feature vector by using the amplitude mean, phase standard deviation, and signal energy of each subcarrier channel in the stable CSI timing signal to constitute a graph signal vector. The dimension of the graph signal vector is the same as the number of subcarrier channels.

[0122] S52. Perform eigenvalue decomposition on the symmetric normalized Laplacian matrix constructed based on the signal correlation between subcarrier channels, extract all eigenvalues and corresponding eigenvectors, and construct a frequency-domain transformation basis matrix.

[0123] S53. Define a spectral convolution kernel function in the graph Laplacian frequency domain to obtain a graph frequency-domain feature vector:

[0124]

[0125] where f represents the graph frequency-domain feature vector, α i represents the projection coefficient of the graph signal in the i-th frequency direction, u i represents the i-th eigenvector, λ i represents the corresponding eigenvalue, λ max represents the maximum value among all eigenvalues, θ j represents the j-th order parameter of the spectral filter, T j represents the j-th order Chebyshev polynomial, K represents the order of the filter polynomial, and k represents the number of frequency cutoffs.

[0126] S54. Evaluate the spectral distribution of the graph frequency-domain feature vector and construct a frequency energy concentration function:

[0127]

[0128] where ρ represents the weighted energy concentration of the frequency-domain feature, f i represents the i-th component of the graph frequency-domain feature vector, w i represents the frequency weight corresponding to the i-th frequency.

[0129] S55. According to the calculation result of the spectral concentration function, extract the first k frequency components that meet the conditions from the graph frequency-domain feature vector to form the final graph frequency-domain feature vector.

[0130] The present invention designs a spectral convolution kernel function in the graph Laplacian frequency domain, and combines the spectral concentration function to extract finite high-energy frequency components, realizes the sparse construction of the graph frequency-domain feature vector, and improves the representation efficiency and discriminant ability of the frequency-domain modeling.

[0131] In this embodiment, S6 specifically includes:

[0132] S61. Set a set of time scales. Each set of time scales corresponds to a set of state trajectory generation branches, and each set of branches is divided using different time interval lengths and sampling resolutions;

[0133] S62. Construct a multi-scale neural ordinary differential equation model. The multi-scale neural ordinary differential equation model uses the graph frequency domain feature vector as the initial state and establishes a differential relationship for the state changing with time:

[0134]

[0135] where, h (m) (t) represents the state representation vector at time t under the m-th time scale, represents the output mapping weight matrix, represents the input mapping weight matrix, b (m) represents the bias vector, η (m) represents the amplitude coefficient of the modulation term, ψ (m) represents the angular frequency of the modulation term, and tanh represents the hyperbolic tangent activation function;

[0136] S63. Set the start and end times of integration for each time scale, input the graph frequency domain feature vector as the initial state into the differential model, and use the numerical integration method to solve for the state within the specified time interval;

[0137] S64. Establish a state trajectory sampling mechanism, set a fixed sampling interval for each set of time scales, and perform non-linear integral combination on the state results at each sampling moment:

[0138]

[0139] where, represents the state trajectory generated under the m-th time scale, represents the r-th order non-linear combination weight matrix, σ represents the activation function, represents the state integral from the start time to the j-th sampling point, K (m) represents the number of sampling points, and R represents the number of combination terms;

[0140] S65. Generate a complete state trajectory sequence under each time scale. The state trajectory sequence is composed of the integral transformation results of each sampling point and has a fixed number of time steps and state vector dimensions;

[0141] S66. Save the state trajectories generated under each time scale separately to form the state representation results of multi-scale modeling.

[0142] The present invention designs a multi-scale neural ordinary differential model to construct state trajectories at different time granularities, and uses time scale division, differential equation modeling, and integral sampling mechanisms to establish a continuous evolution expression of the state, enhancing the model's ability to describe the state change process.

[0143] In this embodiment, the S7 specifically includes:

[0144] S71. Set a neural network classifier. The structure of the neural network classifier consists of an input layer, a first hidden layer, a second hidden layer, and an output layer. The input layer receives the fixed-length vector converted from the state evolution sequence. The first and second hidden layers are fully connected structures, and the output layer uses normalized mapping to output the state probability distribution;

[0145] S72. After converting the state evolution sequence into a fixed-dimensional vector, input it into the neural network classifier, and sequentially perform feature mapping, bias superposition, and non-linear transformation operations to form multi-layer feature representations;

[0146] S73. Calculate the hidden representation in the first hidden layer, combining the main weight path and the modulation branch:

[0147] z1 = σ(W1·x + b1 + A1·tanh(U1·x + c1));

[0148] Among them, z1 represents the output vector of the first hidden layer, x represents the input vector after the state evolution sequence is converted, W1 represents the weight matrix of the main mapping path, b1 represents the bias vector of the main path, U1 represents the weight matrix of the modulation path, c1 represents the bias vector of the modulation path, A1 represents the output scaling matrix of the modulation path, σ represents the activation function of the first hidden layer, and tanh represents the hyperbolic tangent function;

[0149] S74. Send the output of the first hidden layer to the second hidden layer and perform a double non-linear cross mapping operation:

[0150] z2 = ReLU(W2·z1 + b2 + γ·sin(U2·z1 + c2));

[0151] Among them, z2 represents the output vector of the second hidden layer, W2 represents the main weight matrix of the second layer, b2 represents the bias vector, U2 represents the weight matrix of the modulation path, c2 represents the modulation path bias, γ represents the modulation ratio factor, ReLU represents the rectified linear unit function, and sin represents the modulation function;

[0152] S75. Send the output of the second hidden layer to the output layer, perform a normalization activation operation, and generate the personnel state classification probability distribution;

[0153] S76. Determine the final personnel state recognition result according to the maximum value index in the probability distribution.

[0154] The present invention constructs a neural network classification structure with a main path and a modulation branch. After inputting the state evolution sequence, it can achieve multi-level feature abstraction and cross-expression, improving the adaptability and recognition accuracy of the classifier for complex state patterns.

[0155] A personnel state detection system based on a wifi router, comprising:

[0156] A wifi signal acquisition module, used to deploy a wifi router in the target area, collect the CSI of subcarriers, and form a CSI signal sequence containing amplitude information and phase information;

[0157] A neural network configuration module, used to configure a neural network classifier and load model parameters, where the model parameters include a weight matrix, a bias vector, and an activation function type;

[0158] A CSI signal preprocessing module, used to perform moving average filtering, band-pass filtering, and phase calibration operations on the CSI signal sequence to generate a stable CSI time series signal;

[0159] A graph structure construction module, used to divide the stable CSI time series signal into graph nodes according to subcarrier channels, calculate feature vectors based on the amplitude mean, phase standard deviation, and signal energy of the signals corresponding to each node, calculate edge weights using a Gaussian kernel function, generate an adjacency matrix and a degree matrix, and construct a symmetric normalized Laplacian matrix;

[0160] A frequency domain feature extraction module, used to jointly model the graph signal vector and the symmetric normalized Laplacian matrix, perform a frequency domain transform, construct a spectral convolution kernel function, calculate a spectral concentration function, and select several frequency components in the graph frequency domain to form a graph frequency domain feature vector;

[0161] A neural ordinary differential modeling module, used to construct a multi-scale neural ordinary differential equation model, use the graph frequency domain feature vector as the initial state, set different time scales, and construct a differential equation system with a modulation term at each time scale, and use a numerical integration method to generate corresponding state trajectories to obtain a state trajectory sequence at each time scale;

[0162] A state recognition module, used to receive the state trajectories generated at each time scale, input them into the neural network classifier in turn, perform multi-layer mapping and non-linear function processing, output a state classification probability distribution, and generate a state recognition result according to the index corresponding to the maximum probability.

[0163] The present invention defines the connection relationships and input / output contents between the functional modules, realizes the modular implementation from signal perception, frequency domain modeling, dynamic trajectory generation to state recognition, and has good system integration and engineering deployment capabilities.

[0164] Example 1:

[0165] To verify the feasibility of the present invention in implementation, the present invention is applied to an actual intelligent elderly care scenario to non-contact monitor and identify the daily behavior status of the elderly in elderly care institutions. The purpose is to perceive in real time whether the elderly are in a stationary, walking, sitting, standing up or falling state without wearing devices or installing cameras, so as to assist the nursing staff in timely intervention and treatment.

[0166] In this embodiment, a standard single-person living room in a certain intelligent elderly care apartment in Chaoyang District, Beijing, with an area of about 20 square meters, is selected, which has an independent bathroom, a bed, a sofa and an activity space. Two wifi routers supporting CSI extraction (model: Intel 5300 + iwlwifi driver, frequency band set to 5GHz, number of subcarriers: 30) are arranged on the ceiling of the room. After the deployment is completed, a stable CSI receiving link is constructed, and 100 frames of CSI data are collected per second.

[0167] The system first models the empty room of the environment and records the CSI reference frames under the condition of no human activity. Then, it guides the monitored object to perform a series of standard actions, including long-term stillness, normal walking, sudden falling, standing up from a sitting position, turning around in place, etc. At the same time, taking the camera and manual observation as the label reference, the corresponding timestamps of the actions are synchronously recorded and compared with the output of the wifi system for verification.

[0168] In this scenario, after the CSI signal sequence is subjected to moving average filtering, band-pass filtering and phase calibration, the stability is improved by about 41.6%. The amplitude mean, phase standard deviation and energy value of the graph nodes are extracted from each group of CSI data within a fixed 5-second time window to construct a 30-node spectrogram. The edge weights are calculated through the Gaussian kernel function to generate the adjacency matrix and the symmetric normalized Laplacian matrix, and then the frequency domain feature extraction and convolutional mapping are performed.

[0169] In this embodiment, the neural ordinary differential model sets 3 groups of time scales: 0.5 seconds, 1 second and 2 seconds, and the Euler and Runge-Kutta fourth-order methods are respectively used for integral solution. A 16-dimensional state vector trajectory is generated for each scale, and they are spliced to form the final state evolution sequence. The classification module adopts a three-layer fully connected neural network, with the hidden layer dimensions of 64 and 32, and finally outputs the probability distribution of six types of states.

[0170] The system runs continuously for 72 hours, monitors 12 elderly people in total, collects about 25,920,000 frames of effective CSI data in total, and there are 3,927 segments of manually marked states in total. The accuracy rate of the recognition result compared with the manual marking is shown in Table 1:

[0171] Table 1 Comparison table of the recognition accuracy rate of personnel status

[0172]

[0173] In addition, to compare the effects of the present invention with those of traditional manual feature + machine learning classification methods, the support vector machine (SVM) and random forest (RF) methods were selected to train and test the same CSI data, and the results are shown in Table 2:

[0174] Table 2 Comparison of recognition accuracy of different methods

[0175]

[0176]

[0177] As can be seen from the table, the recognition accuracy of the present invention in short-term dynamic states such as personnel falling and turning is significantly better than that of traditional methods, and it has a lower inference latency after model compression and optimization, making it suitable for real-time monitoring scenarios.

[0178] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes should be covered within the protection scope of the present invention.

Claims

1. A method for detecting the status of personnel based on a wifi router, characterized in that, It includes the following steps: S1. Deploy a wifi router in the target area, collect the subcarrier CSI that changes due to the disturbance caused by personnel movement, and form a CSI signal sequence containing amplitude information and phase information; S2. Configure a neural network classifier and load the pre-constructed parameters; S3. Preprocess the CSI signal sequence to generate a stable CSI time series signal; S4. Construct a graph structure based on the stable CSI time series signal, use each subcarrier channel as a graph node, set the edge weights according to the correlation between signals, and generate a spectral graph structure; S5. Perform Laplace eigen-transformation on the spectral graph structure to extract the graph frequency domain feature vector; S6. Construct a multi-scale neural ordinary differential equation model, use the graph frequency domain feature vector as the input, set the microstructural components with different time scales, and generate the state trajectories representing the dynamic evolution of personnel at each time scale; S7. Concatenate the state trajectories generated at each time scale to construct a state evolution sequence, input the state evolution sequence into the neural network classifier, and use the loaded parameters to discriminate the state evolution sequence to generate the personnel state recognition result.

2. The method for detecting the status of a person based on a wifi router according to claim 1, wherein, The parameters include the weight matrix, bias vector, and activation function type of each connection structure from the input layer to the output layer.

3. A method for detecting the status of personnel based on a wifi router according to claim 1, characterized in that, The preprocessing includes performing moving average filtering, band-pass filtering, and phase calibration operations.

4. A method for detecting the status of personnel based on a wifi router according to claim 1, characterized in that, The personnel state recognition result includes at least one of a stationary state, a walking state, a falling state, a sitting state, a standing-up state, and a turning state.

5. A method for detecting the status of a person based on a wifi router according to claim 1, characterized in that, The specific content of S1 includes: S11. Arrange at least two static wifi router nodes at the boundary and inside of the target area, and set the router working frequency band and subcarrier configuration parameters; S12. Establish a data sending and receiving link for the wifi channel, set the CSI acquisition period, and obtain the CSI measurement values of each subcarrier in the physical layer downlink; S13. Collect wifi signals under the condition of personnel activities, record the channel disturbances such as multipath fading, occlusion, and scattering caused by personnel movement, and form a time series CSI raw data frame sequence; S14. Extract the amplitude information and phase information of each subcarrier from each CSI data frame, and represent them as a pair of complex modulus and argument; S15. Combine the amplitude information and phase information in the order of acquisition time to form a CSI signal sequence containing timestamp marks.

6. A method for detecting the status of a person based on a wifi router according to claim 1, characterized in that, The specific content of S4 includes: S41. Divide the stable CSI time series signal by subcarrier channel, and map each subcarrier channel to a graph node; S42. Set a time window with a fixed length, perform statistical analysis on the CSI time series signal corresponding to each graph node, and extract the feature vector containing the amplitude mean, phase standard deviation, and signal energy; S43. Calculate the edge connection weights between each pair of graph nodes through a Gaussian kernel function: Among them, w ij represents the edge weight between the $i$-th node and the $j$-th node, $v i represents the statistical feature vector of node $i$, $v j represents the statistical feature vector of node $j$, $\|\cdot\|_2$ represents the Euclidean distance function, and $\sigma$ represents the bandwidth parameter; S44. Combine all edge weights to form an adjacency matrix. The adjacency matrix is a symmetric real matrix, and the matrix elements are the edge weight values between each graph node; S45. Calculate the degree value of the graph node according to the adjacency matrix, and construct a degree matrix. The degree matrix is a diagonal matrix, and the main diagonal elements are the sum of the edge weights connected to each node; S46. Construct a symmetric normalized Laplacian matrix based on the adjacency matrix and the degree matrix: where L represents the symmetric normalized Laplacian matrix, I represents the identity matrix, D represents the degree matrix, and A represents the adjacency matrix; S47. Use the symmetric normalized Laplacian matrix as the representation form of the spectral graph structure.

7. A method for detecting the status of a person based on a wifi router according to claim 1, characterized in that, The specific steps of S5 are as follows: S51. Compose a feature vector from the amplitude mean, phase standard deviation, and signal energy of each subcarrier channel in the stable CSI time series signal to form a graph signal vector. The dimension of the graph signal vector is the same as the number of subcarrier channels; S52. Perform eigenvalue decomposition on the symmetric normalized Laplacian matrix constructed based on the signal correlation between subcarrier channels, extract all eigenvalues and corresponding eigenvectors, and construct a frequency domain transformation basis matrix; S53. Define a spectral convolution kernel function in the graph Laplacian frequency domain to obtain a graph frequency domain feature vector: Among them, f represents the graph frequency domain feature vector, and α i represents the projection coefficient of the graph signal in the i-th frequency direction, u i represents the i-th eigenvector, λ i represents the corresponding eigenvalue, λ max represents the maximum value among all eigenvalues, θ j represents the j-th order parameter of the spectral filter, T j represents the j-th order Chebyshev polynomial, K represents the filter polynomial order, and k represents the frequency truncation number; S54. Evaluate the spectral distribution of the graph frequency domain feature vector to construct a frequency energy concentration function: Among them, ρ represents the weighted energy concentration of frequency-domain features, and f i represents the i-th component of the graph frequency-domain feature vector, and w i represents the frequency weight corresponding to the i-th frequency; S55. According to the calculation result of the spectral concentration function, extract the first k frequency components that meet the conditions from the graph frequency domain feature vector to form the final graph frequency domain feature vector.

8. A method for detecting the status of personnel based on a wifi router according to claim 1, characterized in that, The specific steps of S6 are as follows: S61. Set a set of time scales. Each set of time scales corresponds to a set of state trajectory generation branches. Each set of branches is divided with different time interval lengths and sampling resolutions; S62. Construct a multi-scale neural ordinary differential equation model. The multi-scale neural ordinary differential equation model uses the graph frequency domain feature vector as the initial state and establishes a differential relationship between the states changing with time: where h (m) (t) represents the state representation vector at time t under the m-th time scale, represents the output mapping weight matrix, represents the input mapping weight matrix, b (m) represents the bias vector, η (m) represents the amplitude coefficient of the modulation term, ψ (m) represents the angular frequency of the modulation term, and tanh represents the hyperbolic tangent activation function; S63. Set the start and end times of integration for each time scale. Input the graph frequency domain feature vector as the initial state into the differential model, and use the numerical integration method to solve the state within the specified time interval; S64. Establish a state trajectory sampling mechanism. Set a fixed sampling interval for each set of time scales, and perform a non-linear integral combination on the state results at each sampling moment; Among them, represents the state trajectory generated at the m-th time scale, represents the r-th order non-linear combination weight matrix, and σ represents the activation function, represents the state integral from the starting moment to the j-th sampling point, and K (m) represents the number of sampling points, and R represents the number of combination terms; S65. Generate a complete state trajectory sequence under each time scale. The state trajectory sequence is composed of the integral transformation results of each sampling point, and has a fixed number of time steps and state vector dimensions; S66. Save the state trajectories generated under each time scale separately to form the state representation results of multi-scale modeling.

9. A method for detecting the status of a person based on a wifi router according to claim 1, characterized in that, The specific steps of S7 are as follows: S71. Set a neural network classifier. The structure of the neural network classifier consists of an input layer, a first hidden layer, a second hidden layer, and an output layer. The input layer receives the fixed-length vector converted from the state evolution sequence. The first and second hidden layers are fully connected structures, and the output layer uses a normalized mapping to output the state probability distribution; S72. Input the state evolution sequence into the neural network classifier after converting it into a fixed-dimensional vector, and sequentially perform feature mapping, bias superposition, and non-linear transformation operations to form a multi-layer feature representation; S73. Calculate the hidden representation in the first hidden layer, combining the main weight path and the modulation branch: z1 = σ(W1·x + b1 + A1·tanh(U1·x + c1)); Among them, z1 represents the output vector of the first hidden layer, x represents the input vector after the transformation of the state evolution sequence, W1 represents the weight matrix of the main mapping path, b1 represents the bias vector of the main path, U1 represents the weight matrix of the modulation path, c1 represents the bias vector of the modulation path, A1 represents the output scaling matrix of the modulation path, σ represents the activation function of the first hidden layer, and tanh represents the hyperbolic tangent function; S74. Send the output of the first hidden layer to the second hidden layer and perform a double non-linear cross mapping operation: z2 = ReLU(W2·z1 + b2 + γ·sin(U2·z1 + c2)); Among them, z2 represents the output vector of the second hidden layer, W2 represents the main weight matrix of the second layer, b2 represents the bias vector, U2 represents the weight matrix of the modulation path, c2 represents the bias of the modulation path, γ represents the modulation ratio factor, ReLU represents the rectified linear unit function, and sin represents the modulation function; S75. Send the output of the second hidden layer to the output layer, perform a normalization activation operation, and generate the probability distribution of the personnel state classification; S76. Determine the final personnel state recognition result according to the maximum value index in the probability distribution.

10. A personnel status detection system based on a Wi-Fi router, which executes a personnel status detection method based on a Wi-Fi router according to any one of claims 1 to 9, characterized in that, Including: A wifi signal acquisition module, which is used to deploy a wifi router in the target area, collect the CSI of the subcarrier, and form a CSI signal sequence containing amplitude information and phase information; A neural network configuration module, which is used to configure a neural network classifier and load model parameters, and the model parameters include a weight matrix, a bias vector, and an activation function type; A CSI signal preprocessing module, which is used to perform a moving average filter, a band-pass filter, and a phase calibration operation on the CSI signal sequence to generate a stable CSI time series signal; A graph structure construction module, which is used to divide the stable CSI time series signal into graph nodes according to the subcarrier channels, calculate the feature vector based on the amplitude mean, phase standard deviation, and signal energy of the signals corresponding to each node, calculate the edge weights using a Gaussian kernel function, generate an adjacency matrix and a degree matrix, and construct a symmetric normalized Laplacian matrix; A frequency domain feature extraction module, which is used to jointly model the graph signal vector and the symmetric normalized Laplacian matrix, perform a frequency domain transformation, construct a spectral convolution kernel function, calculate the spectral concentration function, and select several frequency components in the graph frequency domain to form a graph frequency domain feature vector; A neural ordinary differential modeling module, which is used to construct a multi-scale neural ordinary differential equation model, use the graph frequency domain feature vector as the initial state, set different time scales, and construct a differential equation system with a modulation term at each time scale, and use a numerical integration method to generate the corresponding state trajectory to obtain the state trajectory sequence at each time scale; A state recognition module, which is used to receive the state trajectories generated at each time scale, input them into the neural network classifier in turn, perform multi-layer mapping and non-linear function processing, output the state classification probability distribution, and generate the state recognition result according to the index corresponding to the maximum probability.

Images