A csi human fall recognition method in a bathroom shower interference environment

Abstract

The present application relates to a kind of CSI human body fall identification method under bathroom shower interference environment, belong to wireless communication technical field.The method includes: S1: in the bathroom area under shower interference environment, obtain fall data, calculate space channel constant R by CSI amplitude and CSI phase, calculate shower signal energy ratio SSER by power spectral density PSD, interference discrimination is carried out by combining space channel parameter R and shower signal energy ratio SSER to calculate shower interference index;S2: on the basis of step S1, the CSI data of the same action collected is divided, the data without shower interference is divided into source domain data set, and the data with shower interference is divided into target domain data set, a neighborhood adaptive network model is constructed and trained, and a fall activity recognition result is obtained.The present application can effectively improve the human body fall activity recognition accuracy under shower interference environment.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology and relates to a CSI human fall detection method in a bathroom shower interference environment. Background Technology

[0002] Reports indicate that many people with mobility impairments frequently fall in bathrooms, a problem that has drawn widespread attention. The slippery and confined environment of bathrooms, coupled with the inherent mobility limitations of some elderly or disabled individuals, contributes to the frequent occurrence of accidental falls.

[0003] Currently, fall detection methods, from a signal acquisition perspective, mainly include vision-based fall detection, acoustic-based fall detection, and wearable fall detection based on inertial sensors. Vision-based fall detection uses a camera to capture images of human movement and extracts feature frames indicating a fall using image processing algorithms. It boasts high accuracy, but this method is prone to privacy leaks, is susceptible to environmental influences, and is costly, making it unsuitable for bathroom scenarios. Acoustic-based fall detection uses the sound waves generated after a fall as evidence, but this method is more susceptible to noise from water and showers, leading to a high false alarm rate. Wearable inertial sensor-based fall detection uses miniature devices made of micro-sensors and controllers for fall recognition, processing the data with algorithms to trigger alarms. However, this method is inconvenient to wear in a bathroom, easily affected by environmental factors, and has low detection accuracy.

[0004] Wi-Fi sensing detects falls by analyzing changes in Wi-Fi Channel State Information (CSI). It boasts advantages such as being contactless, easy to deploy, unaffected by light, and having a wider sensing range, making it an emerging research direction in human fall detection and recognition. The recognition process mainly consists of two steps: data preprocessing and a sensing algorithm. Data preprocessing primarily involves denoising the CSI data and removing additive interference, while the sensing algorithm mainly identifies the processed CSI data. Existing research largely focuses on removing static environmental interference, such as the effects of bathroom walls and objects, employing only general filtering algorithms and noise-resistant deep learning algorithms to tolerate interference. Because this method does not specifically consider the dynamic environmental interference of flowing shower water, its recognition performance in shower environments is less than ideal. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a CSI human fall recognition method in a bathroom shower environment with interference, so as to solve the problem of low human fall recognition rate in a bathroom shower environment.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A CSI (Continuous Sensor) method for human fall detection in a bathroom shower environment with interference, specifically including the following steps:

[0008] S1: Acquire data related to activities such as falls in the bathroom area under shower interference environment, calculate the spatial channel constant R by CSI amplitude and CSI phase, calculate the shower signal energy ratio SSER by power spectral density PSD, and calculate the shower interference index SII by combining spatial channel parameter R and shower signal energy ratio SSER for interference discrimination;

[0009] S2: Based on step S1, the collected CSI data of the same action are divided into two datasets: data without shower interference is divided into the source domain dataset, and data with shower interference is divided into the target domain dataset. A neighborhood adaptive network model is constructed and trained. Specifically, features are extracted from the source and target datasets using a feature extractor and then shared. Then, a feature weighting algorithm is combined with a convolutional layer to select the CSI data features that best represent the human fall action. The feature vectors of the two domains are input into a feature aligner, and the distance between the second-order statistics of the source and target features is used as the aligner. C The difference between the two domains is calculated; finally, the output is obtained by gradient inversion and network parameter optimization, which identifies the fall action.

[0010] Furthermore, step S1 specifically includes the following steps:

[0011] S11: Data was collected in the bathroom area under shower interference environment according to the types of activities: no activity, falling, and normal bathing. The amplitude, phase, and power spectral density of the CSI data were extracted according to the activity type and time sequence.

[0012] S12: Calculate the spatial equivalent dielectric constant R using the amplitude and phase of the CSI, perform frequency domain analysis on the CSI to obtain the power spectral density PSD of the CSI, and further calculate the shower signal energy ratio SSER based on the characteristics of the shower interference frequency energy distribution. Finally, calculate the shower interference index SII using the spatial equivalent dielectric constant R and the shower signal energy ratio SSER. This process includes the following steps:

[0013] S121: Due to packet boundary delay, sampling frequency offset, and carrier frequency offset during transmission, Wi-Fi transceivers using Intel 5300 network cards experience CSI phase shift. A linear transformation scheme can be used for CSI phase calibration. After calibration, the CSI amplitude and CSI phase are extracted to calculate the spatial equivalent dielectric constant R.

[0014]

[0015] in, and These are the CSI amplitudes of the k-th subcarrier received by adjacent antennas on the same network card. and These are the CSI phases of the k-th subcarriers received by adjacent antennas on the same network card;

[0016] By combining different antenna pairs, calculations are performed. The values ​​also differ; experimental observations show that different antenna pairs... The values ​​differ. Furthermore, because the Wi-Fi signal transmission paths between different receiving and transmitting antennas are different, the measured amplitude and phase values ​​on different receiving antennas also differ, ultimately leading to variations in the calculated values. The values ​​are different; therefore, the calculated values ​​for different subcarriers on the same antenna pair are different. The values ​​also differ. To fully describe the shower water information in space, the SEP (Risk Identification) representation method is adopted. Assuming the receiver has m antennas, the number of antenna pairs formed is p = m(m-1) / 2. When the collected CSI data packets contain k subcarriers, a p*k dimensional R vector can be calculated. SEP is defined as:

[0017]

[0018] Calculate the mean of the SEP matrix.

[0019] The degree of shower interference is characterized by fusing the CSI amplitude and phase information of k subcarriers on different receiving antennas in the bathroom space;

[0020] S122: Calculate the shower signal energy ratio (SSER) index. Use Fast Fourier Transform (FFT) to convert the acquired time-domain CSI signal to obtain the power spectral density (PSD) in the frequency domain: PSD = |FFT(H). 2 , representing the power distribution of the CSI signal in the frequency domain, and statistical analysis determines the energy distribution of the shower signal;

[0021] S123: Combination The Shower Interference Index (SII) is calculated using the SSER parameter to quantitatively represent the impact of shower interference on wireless signals.

[0022]

[0023] in, It is the vacuum permittivity, typically 8.854 × 10⁻⁶. -12 F / m.

[0024] The Shower Interference Index (SII) can effectively reflect the degree of shower interference in the bathroom. Based on the interference index, interference detection and interference degree classification are performed on the data. The CSI data with interference and the CSI data without interference are labeled according to activity type, time sequence, and interference degree. The CSI dataset without shower interference is added to the source dataset, and the CSI dataset with shower interference is added to the target domain dataset.

[0025] Furthermore, step S122 specifically includes the following steps:

[0026] (3) First, assume the experimental data collection environment. Data collection is carried out in a real scene of a residential apartment bathroom with dimensions of 2.05m×1.89m×2.23m. A 3×3 transmit and receive CSI device equipped with an Intel 5300 network card is deployed at a height of 1.5m in the bathroom. The Wi-Fi working frequency is 5GHz, the sampling frequency is 1000Hz, the shower head in the bathroom is round with a diameter of 19cm, the water flow rate is 10L / min, and the vertical distance from the ground is 2.2m.

[0027] (4) Data collection process and fitting ratio definition: 10 data points were collected in a no-shower environment and 20 data points were collected with the shower on, each lasting 1 minute; the CSI transmitting device sent 1000 data packets per second, splitting the 1-minute CSI file into 6 10-second CSI files, each file generating one spectrogram; the shower area contained in each spectrogram can be used for KS distribution test; the Kolmogorov-Smirnov test was used to test the distribution of the size of each sample dataset. If the test does not show the data following the specified distribution at a certain significance level, the result is 1; otherwise, it is 0. Seven distributions were selected, and the KS test was performed on each sample dataset. The following fitting ratio was used to evaluate the fitting effect; the fitting ratio of the distribution is defined as FP:

[0028]

[0029] Where, N ks＝0 N represents the number of times the KS test result is 0. total This indicates the total number of KS tests performed. The higher the fit of the distribution, the better the background interference matches the distribution.

[0030] Based on the fitting ratios of the seven distributions, it can be concluded that the Rayleigh distribution is no longer suitable for this shower environment, while the Weibull distribution is the expected distribution and can be suitable for most shower scenarios.

[0031] The probability density function (PDF) of a Weibull random variable x is given by the following equation:

[0032]

[0033] Where f(x; λ, k) is the probability density function following a Weibull distribution, k>0 is the shape parameter, and λ>0 is the scaling factor of the distribution;

[0034] The energy information of N shower background areas can be represented as follows:

[0035] The reciprocal of k is given by the following formula:

[0036]

[0037] Therefore, the estimate of k is:

[0038]

[0039] Therefore, the estimated value of the shape parameter of the Weibull distribution is obtained, which is regarded as a known value k; the maximum likelihood estimation method is used to estimate λ as:

[0040]

[0041] Assume the frequency range [f1, f2] represents the frequency range of the shower interference signal. This frequency range signal follows a Weibull distribution, and the total energy E of the shower signal is... shower It can be calculated using the following formula:

[0042]

[0043] Where f is the frequency point, and λ and k are the Weibull distribution parameters obtained by fitting the data;

[0044] Let E all The shower signal energy ratio (SSER) can be expressed as the energy at each window frequency point of the PSD:

[0045]

[0046] Among them, E all This represents the estimated frequency energy caused by shower interference in the power spectral density (PSD) spectrum. i This represents the energy value at each frequency point i in the power density spectrum (PSD).

[0047] Furthermore, step S2 specifically includes the following steps:

[0048] S21: The segmented CSI data is fed into a deep transfer network model based on domain-invariant features. First, motion-related features in the CSI data are extracted and selected. The CSI sample data includes time-domain features and frequency-domain features. Time-domain features include mean, maximum, minimum, skewness, kurtosis, variance, and average line throughput, which can capture CSI waveform patterns. Frequency-domain features include normalized entropy, normalized energy, and FFT peak value. The ReliefF algorithm is used to calculate the weights of different feature categories, and a voting mechanism is used to select the most effective features to identify human motion.

[0049] The feature projection neural network layer uses the weights obtained by ReliefF as prior knowledge and automatically focuses on effective features as continuously trained recognition models.

[0050] S22: Based on the shower interference detection model, the source domain dataset and target domain dataset are divided. Features of the source and target datasets are extracted through step S21, and weights are shared to classify human actions, obtaining the predicted result y. The loss between the predicted result and the true result y is calculated. M :

[0051]

[0052] Where C is the number of action data and x is the input data;

[0053] Assume the source dataset is labeled L s ={y i}, i∈{1,...,L}, where L is the total number of labeled samples in the source domain dataset, y i D is the label of the i-th sample. s ={x i},x∈R d D s x is the feature set of the source domain dataset. i R is the feature vector of the i-th sample. d These feature vectors represent points in a d-dimensional real vector space, and the unlabeled target dataset is D. T ={u i},u∈R d D T Let D be the set of feature vectors of the target domain dataset, and let D be the unlabeled target dataset. T ={u i},u∈R d The source and target datasets contain n data points. S x and u are learnable d-dimensional deep activations φ(I) of input I. It is the j-th dimension of the i-th source. C is the j-th dimension of the i-th target. S C T Let the source data and target data represent the covariance matrices, respectively, and the loss be... C The loss is the distance between the second-order statistics of the source features and the target features, as shown below:

[0054]

[0055] In the above formula, The Frobenius norm of the square matrix is ​​shown below; the covariance matrices of the source and target data are as follows:

[0056]

[0057] Where l is a column vector with all elements equal to 1; the gradient of the input feature is calculated using the chain rule:

[0058]

[0059] The feature vectors from the two domains are then input into the feature aligner to calculate the loss. C This is used to calculate the degree of difference between two domains; the output is optimized through gradient inversion and network parameters to reduce the loss. M and Loss C The parameters that minimize the loss are then shared between the training and testing networks to obtain a domain-adaptive network model. This model is then trained to ultimately obtain the fall recognition result.

[0060] Furthermore, in step S21, the specific steps of the ReliefF algorithm are as follows:

[0061] (1) Initialize the weight of each feature to 0, and randomly select a sample R from the dataset;

[0062] (2) Find the nearest neighbor sample: find the nearest neighbor sample H that is in the same class as sample R, and for each class, find the nearest neighbor sample that is in a different class than sample R;

[0063] (3) Update feature weights: For each feature A, update the weight W(A):

[0064]

[0065] Where, diff(A,R,H) represents the difference between feature A and R and H, and diff(A,R,M) represents the difference between feature A and R and M (nearest neighbor samples of different classes); m is the total number of samples, used for normalization;

[0066] (4) Feature selection: If a feature is greater than the preset threshold δ, i.e. W(A)≥δ, then feature A is added to the result set.

[0067] (5) Repeat the above operation m times and select the feature set T.

[0068] The beneficial effects of this invention are as follows:

[0069] (1) This invention addresses the CSI (Continuous Sensor Intrusion) human fall recognition scenario in bathrooms where shower interference may exist. It employs a linear transformation to correct the CSI phase, obtaining an accurate CSI phase. Furthermore, it proposes a shower interference detection method to determine whether shower interference exists in the CSI data. A neighborhood-adaptive neural network model is constructed to learn the differences between the source and target domain data. In summary, the method of this invention can significantly improve the accuracy of human fall recognition in shower environments.

[0070] (2) This invention can be extended to other areas of CSI for application, which helps to improve its work efficiency.

[0071] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0072] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0073] Figure 1 This is a flowchart of the CSI human fall detection method in a bathroom shower interference environment proposed in this invention;

[0074] Figure 2 This invention presents a human fall recognition model under shower interference conditions.

[0075] Figure 3 This invention presents the neighborhood adaptive network model structure for a shower interference environment. Detailed Implementation

[0076] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0077] Please see Figures 1-3 This invention provides a method for human fall recognition based on WIFI CSI neighborhood adaptation in a shower interference environment, such as... Figure 1 As shown, this method mainly consists of the following steps:

[0078] S1: Acquire data related to activities such as falls in the bathroom area under shower interference environment, calculate the spatial channel constant R by CSI amplitude and CSI phase, calculate the shower signal energy ratio SSER by power spectral density PSD, and calculate the shower interference index SII by combining spatial channel parameter R and shower signal energy ratio for interference discrimination.

[0079] Step S1 specifically includes the following steps:

[0080] Step 1: First, data was collected in the bathroom area under shower interference environment according to the activity types of no activity, falling, and normal bathing. Then, the amplitude, phase and power spectral density of the CSI data were extracted according to the activity type and time sequence.

[0081] Step 2: Calculate the spatial equivalent dielectric constant R using the amplitude and phase of the CSI, perform frequency domain analysis on the CSI to obtain the power spectral density PSD of the CSI, and further calculate the shower signal energy ratio SSER by combining the characteristics of the shower interference frequency energy distribution. Calculate the shower interference index by combining the spatial equivalent dielectric constant R and the shower signal energy ratio SSER.

[0082] First, Wi-Fi transceivers using Intel 5300 network cards may experience packet boundary delay, sampling frequency offset, and carrier frequency offset during data transmission, which can cause CSI phase shift. To address this issue, a linear transformation method can be used to calibrate the phase. After phase calibration, the amplitude and phase of the CSI are further extracted, and the spatial equivalent dielectric constant is calculated based on this.

[0083]

[0084] and These are the CSI amplitudes of the k-th subcarrier received by adjacent antennas on the same network card. and These are the CSI phases of the k-th subcarriers received by adjacent antennas on the same network card.

[0085] By combining different antenna pairs, calculations are performed. The values ​​are also different; this was observed through implementation, and different antenna pairs... The values ​​differ. Furthermore, because the Wi-Fi signal transmission paths between different receiving and transmitting antennas are different, the measured amplitude and phase values ​​on different receiving antennas also differ, ultimately leading to variations in the calculated values. The values ​​are different; therefore, the calculated values ​​for different subcarriers on the same antenna pair are different. The values ​​also differ. To fully describe the shower water information in space, a SEP representation method is proposed. Assuming the receiver has m antennas, the number of antenna pairs formed is p = m(m-1) / 2. Therefore, when the collected CSI data packets contain k subcarriers, a p*k dimensional R vector can be calculated, and SEP is defined as:

[0086]

[0087] Furthermore, the mean of the SEP matrix is ​​obtained.

[0088] The degree of shower interference is characterized by fusing the CSI amplitude and phase information of k subcarriers on different receiving antennas in the bathroom space.

[0089] Secondly, the shower signal energy ratio (SSER) index is calculated. The acquired time-domain CSI signal is converted using Fast Fourier Transform (FFT) to obtain the power spectral density in the frequency domain, PSD = |FFT(H). 2 This represents the power distribution of the CSI signal in the frequency domain. Statistical analysis determines the energy distribution of the shower signal.

[0090] (5) First, assume the experimental data collection environment. Data collection is carried out in a real scene of a residential apartment bathroom with dimensions of 2.05m×1.89m×2.23m. A 3×3 transmit and receive CSI device equipped with an Intel 5300 network card is deployed at a height of 1.5m in the bathroom. The Wi-Fi operating frequency is 5GHz, the sampling frequency is 1000Hz, the shower head in the bathroom is round with a diameter of 19cm, the water flow rate is 10L / min, and the vertical distance from the ground is 2.2m.

[0091] (1) Data collection process and fitting ratio definition: 10 data points were collected under no-shower conditions and 20 data points were collected under shower conditions, with each data point lasting 1 minute. The CSI transmitting device sent 1000 data packets per second, splitting the 1-minute CSI file into 6 10-second CSI files, each of which could generate one spectrogram. The shower area contained in each spectrogram could be used for the KS distribution test. The Kolmogorov-Smirnov test was used to test the distribution of the size of each sample dataset. If the test did not show the data following the specified distribution at a certain significance level, the result was 1; otherwise, it was 0. Seven distributions were selected, and the KS test was performed on each sample dataset. The following fitting ratio was used to evaluate the fit. The fitting ratio of the distribution is defined as FP:

[0092]

[0093] Among them, Nk s＝0 N represents the number of times the KS test result is 0. total This indicates the total number of KS tests performed. The higher the fit of the distribution, the better the background interference matches the distribution.

[0094] Based on the fitting ratios of the seven distributions, it can be concluded that the Rayleigh distribution is no longer suitable for this shower environment, while the Weibull distribution is the expected distribution and can be suitable for most shower scenarios.

[0095] The PDF of a Weibull random variable x is given by the following formula:

[0096]

[0097] Where k>0 is the shape parameter and λ>0 is the scaling factor of the distribution.

[0098] The energy information of N shower background areas can be represented as:

[0099] The reciprocal of k is given by the following formula:

[0100]

[0101] Therefore, the estimate of k is:

[0102]

[0103] Therefore, the estimated shape parameter of the Weibull distribution is obtained, which is considered a known value k. The maximum likelihood estimation method is used to estimate λ as:

[0104]

[0105] Assuming the frequency range [f1, f2] represents the frequency range of the shower interference signal, and the signal in this frequency band follows a Weibull distribution, the total energy of the shower signal can be calculated using the following formula:

[0106]

[0107] Where f is the frequency point, and λ and k are Weibull distribution parameters obtained by fitting data.

[0108] Let E all The shower signal energy ratio (SSER) can be expressed as the energy at each window frequency point of the PSD:

[0109]

[0110] Finally, combining The Shower Interference Index (SII) is calculated using the SSER parameter to quantitatively represent the impact of shower interference on wireless signals.

[0111]

[0112] in, It is the vacuum permittivity, typically 8.854 × 10⁻⁶. -12 F / m.

[0113] The Shower Interference Index (SII) can effectively reflect the degree of shower interference in the bathroom. Based on the interference index, interference detection and interference degree classification are performed on the data. The CSI data with interference and the CSI data without interference are labeled according to activity type, time sequence, and interference degree. The CSI dataset without shower interference is added to the source dataset, and the CSI dataset with shower interference is added to the target domain dataset.

[0114] S2: Based on S1, the dataset is divided into source domain datasets (data without shower interference) and target domain datasets (data with shower interference), and a neighborhood-adaptive fall recognition method is constructed.

[0115] Step S2 specifically includes the following steps:

[0116] Step 1: Input the segmented CSI data into a deep transfer learning network model based on domain-invariant features. First, extract and select motion-related CSI features. CSI sample data includes time-domain and frequency-domain features. Time-domain features include mean, maximum, minimum, skewness, kurtosis, variance, and average line throughput, which can capture the pattern information of the CSI waveform. Frequency-domain features include normalized entropy, normalized energy, and FFT peak value, which can reflect the frequency distribution of the CSI signal. To further improve the effectiveness of the features, the ReliefF algorithm is used to calculate the weights of different feature categories, and a voting mechanism is used to select the most representative features to accurately identify human motion. The steps of the ReliefF algorithm are as follows:

[0117] (1) Initialize the weight of each feature to 0, and randomly select a sample R from the dataset.

[0118] (2) Find the nearest neighbor sample: find the nearest neighbor sample H that is in the same category as sample R. For each category, find the nearest neighbor sample that is in a different category than sample R.

[0119] (3) Update feature weights: For each feature A, update the weight W(A):

[0120]

[0121] Where, diff(A,R,H) represents the difference between feature A and R and H, diff(A,R,M) represents the difference between feature A and R and M (nearest neighbor samples of different classes), and m is the total number of samples used for normalization.

[0122] (4) Feature selection: If a feature is greater than a preset threshold, i.e. W(A)≥δ, then feature A is added to the result set.

[0123] (5) Repeat the above operation m times and select the feature set T.

[0124] The pseudocode for Relief is as follows:

[0125]

[0126] The feature projection neural network layer uses the weights obtained from ReliefF as prior knowledge and automatically focuses on effective features as continuously trained recognition models.

[0127] Step 2: Based on the shower interference detection model, the source and target datasets are divided. Features are extracted from both datasets using the method described in Step 1, and weights are shared to classify human actions, obtaining the predicted result y. The loss between the predicted and actual results y is then calculated. M :

[0128]

[0129] Where C is the number of action data and x is the input data.

[0130] Assume source dataset label L s ={y i}, i∈{1,...,L} is labeled as D s ={x i},x∈R d The unlabeled target dataset is D. T ={u i},u∈R d The source and target datasets contain n data points. S x and u are learnable d-dimensional deep activations φ(I) of input I. It is the j-th dimension of the i-th source (target) data, C S (C T ) represents the covariance matrix, Loss C The loss is the distance between the second-order statistics of the source features and the target features, as shown below:

[0131]

[0132] In the above formula, Let Frobenius norm represent the square matrix. The covariance matrices of the source and target data are shown below:

[0133]

[0134] Where l is a column vector with all elements equal to 1. The gradient of the input feature is calculated using the chain rule:

[0135]

[0136] The feature vectors from the two domains are then input into the feature aligner to calculate the loss. C This is used to calculate the degree of difference between two domains. The output is optimized through gradient inversion and network parameter adjustments to reduce the loss. M and Loss C The parameters that minimize the loss are then shared between the training and testing networks to obtain a domain-adaptive network model. This model is then trained to ultimately obtain the fall recognition result.

[0137] This invention addresses the CSI (Continuous Sensor) fall detection scenario in shower environments with interference. It employs a linear transformation to correct the CSI phase, resulting in accurate CSI phase. Furthermore, it proposes a shower interference detection method to identify whether shower interference exists in the CSI data. A neighborhood-adaptive neural network model is constructed to learn the differences between the source and target domains, effectively reducing the impact of environmental changes on the recognition results. In summary, this method significantly improves the accuracy of fall detection in shower environments and can be extended to other areas of CSI, contributing to improved operational efficiency.

[0138] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A CSI (Continuous Sensory Injury) method for human fall detection in a bathroom shower environment with interference, characterized in that, The method specifically includes the following steps: S1: Obtain fall data in the bathroom area under shower interference environment, calculate the spatial channel constant R by CSI amplitude and CSI phase, calculate the shower signal energy ratio SSER by power spectral density PSD, and calculate the shower interference index SII by combining spatial channel parameter R and shower signal energy ratio SSER for interference discrimination; Step S1 specifically includes the following steps: S11: Data was collected in the bathroom area under shower interference environment according to the types of activities: no activity, falling, and normal bathing. The amplitude, phase, and power spectral density of the CSI data were extracted according to the activity type and time sequence. S12: Calculate the spatial equivalent dielectric constant R using the amplitude and phase of the CSI, perform frequency domain analysis on the CSI to obtain the power spectral density PSD of the CSI, and further calculate the shower signal energy ratio SSER based on the characteristics of the shower interference frequency energy distribution. Finally, calculate the shower interference index SII using the spatial equivalent dielectric constant R and the shower signal energy ratio SSER. This process includes the following steps: S121: CSI phase calibration is performed using a linear transformation scheme. After calibration, the CSI amplitude and CSI phase are extracted to calculate the spatial equivalent dielectric constant. : in, and These are the CSI amplitudes of the k-th subcarrier received by adjacent antennas on the same network card. and These are the CSI phases of the k-th subcarriers received by adjacent antennas on the same network card; Calculated for different subcarriers on the same antenna pair Different values ​​are used to fully describe the shower water information in the space, so the SEP representation method is adopted. It is assumed that the receiver has... m If there are 1 antenna, then the number of antenna pairs formed is 1. When the collected CSI data packets contain k subcarriers, calculate... A dimensional R vector, defining SEP as: Calculate the mean of the SEP matrix. ; The degree of shower interference is characterized by fusing the CSI amplitude and phase information of k subcarriers on different receiving antennas in the bathroom space; S122: Calculate the Shower Signal Energy Ratio (SSER) index. Use Fast Fourier Transform (FFT) to convert the acquired time-domain CSI signal to obtain the power spectral density in the frequency domain. , representing the power distribution of the CSI signal in the frequency domain, and statistical analysis determines the energy distribution of the shower signal; S123: Combination The Shower Interference Index (SII) is calculated using the SSER parameter to quantitatively represent the impact of shower interference on wireless signals. in, It is the vacuum permittivity; S2: Based on step S1, the collected CSI data of the same action are divided into two datasets: data without shower interference is divided into the source domain dataset, and data with shower interference is divided into the target domain dataset. A neighborhood-adaptive network model is constructed and trained. Specifically, features are extracted from the source and target datasets using a feature extractor and then shared. Next, a feature weighting algorithm is combined with a convolutional layer to select the CSI data features that best represent the human fall action. The feature vectors of the two domains are input into a feature aligner, and the distance between the second-order statistics of the source and target features is used as the aligner. The difference between the two domains is calculated; finally, the output is obtained by gradient inversion and network parameter optimization, which identifies the fall action.

2. The CSI human fall detection method according to claim 1, characterized in that, Step S122 specifically includes the following steps: (1) First, assume the experimental data collection environment; (2) Data collection process and fitting ratio definition: 10 data points were collected in a shower-free environment and 20 data points were collected with the shower on, each lasting 1 minute; the CSI transmitting device sent 1000 data packets per second, dividing the 1-minute CSI file into 6 10-second CSI files, each file generating one spectrogram; the shower area contained in each spectrogram was used for the KS distribution test; the Kolmogorov-Smirnov test was used to test the distribution of the size of each sample dataset. If the test did not show the data following the specified distribution at a certain significance level, the result was 1, otherwise 0. Seven distributions were selected, and the KS test was performed on each sample dataset. The following fitting ratio was used to evaluate the fitting effect; the fitting ratio of the distribution was defined as FP: in, This indicates the number of times the KS test result is 0. This indicates the total number of KS tests performed; Based on the fitting ratios of the seven distributions, it is concluded that the Rayleigh distribution is no longer suitable for this shower environment, while the Weibull distribution is the expected distribution and is suitable for most shower scenarios. The probability density function of the Weibull random variable x is given by the following equation: in, It is a probability density function that follows a Weibull distribution. These are shape parameters. It is the scaling factor of the distribution; The energy information of N shower background areas is represented as follows: ; The reciprocal of is given by the following formula: Therefore, the estimate of k is: Therefore, the estimated values ​​of the shape parameters of the Weibull distribution were obtained, and these were treated as known values. The maximum likelihood estimation method is used to estimate the... for: Assuming frequency range , representing the frequency range of the shower interference signal, where the signal follows a Weibull distribution, and the total energy of the shower signal. Calculate using the following formula: in, It is a frequency point. and These are the Weibull distribution parameters obtained by fitting data; make The power spectral density (PSD) represents the energy at each window frequency point, and the shower signal energy ratio (SSER) is expressed as: in, This represents the estimated frequency energy caused by shower interference in the power spectral density spectrum (PSD). Represents each frequency point in the power density spectrum (PSD) i The energy value.

3. The CSI human fall detection method according to claim 1, characterized in that, Step S2 specifically includes the following steps: S21: The segmented CSI data is fed into a deep transfer network model based on domain-invariant features. First, motion-related features in the CSI data are extracted and selected. The CSI sample data includes time-domain features and frequency-domain features. Time-domain features include mean, maximum, minimum, skewness, kurtosis, variance, and average line throughput to capture CSI waveform patterns. Frequency-domain features include normalized entropy, normalized energy, and FFT peak value. The ReliefF algorithm is used to calculate the weights of different feature categories, and a voting mechanism is used to select the most effective features to identify human motion. The feature projection neural network layer uses the weights obtained by ReliefF as prior knowledge and automatically focuses on effective features as continuously trained recognition models. S22: Based on the shower interference detection model, the source domain dataset and target domain dataset are divided. Features of the source and target datasets are extracted through step S21, and weights are shared to classify human actions and obtain prediction results. Calculate the predicted results and the actual results Losses between : in, C The number of action data. Input data; Assume the source dataset is labeled as , L The total number of labeled samples in the source domain dataset. It is the first i The labels of each sample; , The feature set of the source domain dataset. It is the first i The feature vector of each sample These feature vectors represent points in a d-dimensional real vector space, and the unlabeled target dataset is... , The target dataset is the set of feature vectors. The data size in the source dataset and the target dataset is... , and It is a learnable d-dimensional deep activation of input I. , It is the j-th dimension of the i-th source. It is the j-th dimension of the i-th target. These represent the covariance matrices of the source data and the target data, respectively. The loss is the distance between the second-order statistics of the source features and the target features, as shown below: In the above formula, The Frobenius norm of the square matrix is ​​shown below; the covariance matrices of the source and target data are as follows: in, Let the input features be column vectors, where all elements are 1; use the chain rule to calculate the gradient of the input features: The feature vectors from the two domains are then input into the feature aligner for calculation. This is used to calculate the degree of difference between two domains; the output is optimized through gradient inversion and network parameters to achieve... and The parameters that minimize the loss are then shared between the training and testing networks to obtain a domain-adaptive network model. This model is then trained to ultimately obtain the fall recognition result.

4. The CSI human fall detection method according to claim 3, characterized in that, In step S21, the specific steps of the ReliefF algorithm are as follows: (1) Initialize the weight of each feature to 0, and randomly select a sample R from the dataset; (2) Find the nearest neighbor sample: find the nearest neighbor sample H that is in the same class as sample R, and for each class, find the nearest neighbor sample that is in a different class than sample R; (3) Update feature weights: For each feature A Update weights : in, Representation of features A exist R and H The differences between them Representation of features A exist R and M The differences between them; m This is the total number of samples, used for normalization; (4) Feature selection: If a feature is greater than a preset threshold ,Right now Then the features A Add to the result set; (5) Repeat steps (1)-(4) m Next, select the feature set T.

Images