An intelligent elderly care method based on WiFi

Abstract

The present invention discloses an intelligent elderly care method based on WiFi, and the steps are as follows: collect CSI data of the monitoring area through WiFi devices; preprocess the CSI data to determine whether it belongs to a static environment or a dynamic environment; when it is determined to be a static environment, train the preprocessed CSI data through GA-SVM to obtain a trained target model after machine learning, and use the trained target model to identify "staring blankly, sleeping, no one" in the CSI data, and then calculate the heart rate and breathing conditions during sleep; when it is determined to be a dynamic environment, train the preprocessed CSI data through GA-SVM to obtain a trained target model after machine learning, and use the pre-trained target to identify "elderly falling", "indoor intrusion", and other situations in the preprocessed CSI data; the static data processor and the dynamic data processor store the results in the memory, and train and update the pre-trained target model.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent monitoring, and particularly relates to an intelligent monitoring method for the elderly based on WiFi. Background Art

[0002] With the acceleration of the aging process, the guardianship problem of the elderly at home has gradually become a hot topic in society. Currently, the more popular guardianship means on the market are divided into two categories: Device-based and Device-free. Device-based is based on wearable devices, and the most common one is the sensor-based method. Device-based collects relevant data through sensors attached to people, and has advantages such as being decorative and low cost. However, wearable sensors have problems such as short battery life and inconvenience in carrying. The elderly are prone to forgetting to wear them or not knowing how to use them, which will make it difficult for wearable sensors to achieve all-weather guardianship. The non-invasive guardianship method uses general computer vision technology in the market as the monitoring means, but the implementation carrier of this technology - the surveillance camera, is costly for ordinary families, and has low detection accuracy in dark environments, complex azimuth layouts, and does not support non-line-of-sight guardianship. In addition, video guardianship may violate the privacy of the elderly.

[0003] In order to solve the guardianship problem of the elderly, and at the same time to solve the problems of privacy protection of traditional monitoring devices and network cameras, and the implementation of machine vision algorithms for dangerous situation recognition requires complex software and processing requirements; and to overcome the deficiencies of traditional invasive guardianship devices such as being easy to forget and having low battery life.

[0004] At the current stage, there are already patents for elderly guardianship on the market, such as an intelligent elderly care service system based on Internet of Things technology (Patent No.: CN113180947A), and an intelligent monitoring method, system and medium for the elderly based on WiFi (Patent No.: CN108614989B). The former implants various sensor modules with specific functions in the medical care bed, and uses a variety of touchless sensors to monitor the breathing rate, heart rate, arterial pulsation, etc. of users, and adopts an intelligent control method to help guardians better monitor and manage the elderly; the latter performs weighted operations on the JS divergence of the CSI amplitude value and the CSI phase after random forest, and compares the calculation result with the model data to achieve the purpose of danger recognition. Compared with the present system, the systems on the market are often relatively single in function and lack the idea of analysis from the perspective of machine learning and intelligent algorithms. Summary of the Invention

[0005] The purpose of the present invention is to solve the above-mentioned defects in the prior art and provide an intelligent monitoring method for the elderly based on WiFi.

[0006] The object of the present invention can be achieved by adopting the following technical solutions:

[0007] An elderly intelligent monitoring method based on WiFi, the elderly intelligent monitoring method comprising the following steps:

[0008] S1. Collect CSI data of the monitored home area through WiFi devices;

[0009] Further, the process of collecting CSI data in step S1 is as follows: Install a WiFi signal transmitter and receiver in the monitored home area, and install a CSI data real-time collection and processing tool. The working frequency of WiFi is selected as 5 GHz.

[0010] S2. Preprocess the collected CSI data to obtain preprocessed CSI data;

[0011] S3. Determine whether it belongs to a static environment or a dynamic environment according to the preprocessed CSI data;

[0012] S4. When it is determined to be a static environment, input the preprocessed CSI data into a static data processor. The static data processor obtains a first pre-trained target model after machine learning, and uses the first pre-trained target model to identify the three situations of "staring blankly, sleeping, and no one" for the preprocessed CSI information, and calculates the heart rate and breathing conditions during sleep time;

[0013] S5. When it is determined to be a dynamic environment, input the preprocessed CSI data into a dynamic data processor. The dynamic data processor obtains a second pre-trained target model after machine learning, and uses the second pre-trained target model to identify dangerous behaviors for the preprocessed CSI information.

[0014] Further, the process of step S2 is as follows:

[0015] S2.1. Separate the collected CSI raw data to obtain amplitude data and phase data;

[0016] S2.2. Perform linear interpolation and completion on the amplitude data and phase data. Use the mean value of the two adjacent data on the left and right of the point to be interpolated in the data sequence as the estimated value. Using this estimation method, compared with high-order interpolation, the operation complexity is low and the speed is fast;

[0017] S2.3. Obtain phase difference data by calculating the difference of the phase data of different subcarriers, and the value of the phase difference needs to be limited between [-π, π];

[0018] S2.4. Remove noise from the amplitude data and phase difference data by means of Hampel filtering and principal component analysis. The window size selected for filtering is 50. Locate each sample data in the selected window interval that differs from the median μ of the interval by more than three standard deviations σ as an outlier and remove it, thereby eliminating the influence of outliers on the result and improving the accuracy of the result.

[0019] S2.5. Calculate the variances of the amplitude data and phase difference data, and select the amplitude data and phase difference data with the largest variances as the optimal subcarrier amplitude data and optimal phase difference data. The selected optimal subcarrier amplitude data and phase difference data generally contain the most information. Using these two sets of data as the objects for subsequent operations can greatly reduce the amount of computation and improve the computation speed while ensuring the accuracy of the result.

[0020] Further, the process of step S3 is as follows:

[0021] S3.1. Intercept the amplitude data and phase data at every 2 - second interval through a sliding window, and calculate the data mean. To ensure sensitivity, the intercepted window should not be too long, as it may lead to a large amount of single - time computation; nor should it be too short, as it may result in too little information contained in the amplitude data and phase data, leading to misjudgment. A 2 - second window size is more appropriate.

[0022] S3.2. Compare the calculation results with the static data and dynamic data in the memory to determine whether the current data acquisition period is in a static state or a dynamic state. The comparison step does not involve model training and the operation is relatively simple, enabling rapid classification of the data.

[0023] Further, the process in step S4 is as follows:

[0024] S4.1. Initialize the first target model before pre - training.

[0025] S4.2. Collect CSI data in an experimental scenario without any moving objects, and repeatedly collect CSI data in static environments under different scenarios as training data. After pre - processing the training data, generate the first pre - trained target model and save it to the memory.

[0026] S4.3. Perform band - pass filtering on the optimal amplitude data using a Butterworth filter. The pass - band interval is selected as 0.1 - 0.5 Hz. By filtering out frequency components other than respiration, remove the interference of other frequencies, then obtain the spectral data within the respiration range, and finally calculate the respiration rate through fast Fourier transform.

[0027] S4.4. Extract the DWT detail coefficients from the optimal phase data, obtain the spectral data within the heart rate range through band-pass filtering. Use a Butterworth filter with a passband range of 1 - 1.67 Hz to filter out frequency components outside the heart rate range, and finally calculate the heart rate through fast Fourier transform.

[0028] Furthermore, the scenarios required for collecting training data in the static environment in step S4.2 include: "someone is in a daze in the environment", "someone is sleeping in the environment", and "no one is in the environment". In the scenario of "someone is sleeping in the environment", the sleeping position should be the position where people often sleep. By replacing different people and the "daze" positions, multiple sets of data are collected to achieve the collection of high-quality training data, ensure the richness of the training data, and make the model more robust.

[0029] Furthermore, the process in step S5 is as follows:

[0030] S5.1. Initialize the second target model before pre-training;

[0031] S5.2. In the experimental environment, simulate falling and indoor intrusion behaviors. For falling, data under different conditions need to be collected at different positions in the experimental environment, with different postures, and by different experimental personnel; for indoor intrusion, different people are required to conduct experiments to collect more data. Repeat the collection of multiple sets of data as training data, generate the second pre-training target model with the training data, and save it to the memory.

[0032] Furthermore, the scenarios required for collecting training data in the dynamic environment in step S5.2 include: "the elderly fall", "indoor intrusion", and other situations. For falling, data on falls under different conditions need to be collected by replacing different people, with different postures, and by different experimental personnel; for indoor intrusion, different people are required to conduct intrusion experiments to collect more data. Multiple sets of data are collected. When collecting data, the number of people in the environment needs to be strictly controlled, preferably one person. In addition, the falling and intrusion times need to be strictly controlled. The falling time is controlled within 1 second, and the intrusion time is controlled within 1.5 seconds to ensure that the training data is more in line with the real situation and achieve the collection of high-quality training data. The data on falls and intrusions in different situations make the training data more abundant, consider more situations, and make the model more robust.

[0033] Furthermore, the first pre-training target model and the second pre-training target model are trained by GA-SVM. Through global optimization by the GA algorithm, the kernel function of SVM is reconstructed, and the penalty factor C and the regularization coefficient γ are evolved 500 times. At this time, the models obtained by training with the penalty factor and the regularization coefficient γ will be more stable, reducing the redundant information between eigenvalues, improving the classification accuracy of the pre-training target model, and enhancing the classification efficiency and the stability of the model.

[0034] Furthermore, the intelligent elderly care method further includes the following steps:

[0035] S6. The static data processor and the dynamic data processor store the pre-training target model, amplitude data, and phase difference data in the memory. When the model performs classification discrimination each time, the corresponding amplitude data and phase difference data can be stored, the data is calibrated, and the stored data is used as training data at regular intervals to retrain the first pre-training target model and the second pre-training target model, so as to realize the training and updating of the first pre-training target model and the second pre-training target model. Each update of the model is based on more training data, making the model more stable, more accurate, and more adaptable.

[0036] The present invention has the following advantages and effects compared with the prior art:

[0037] (1) The present invention classifies the living environment state of the elderly at home into a static environment and a dynamic environment. For the static environment, its monitoring algorithm pays more attention to the calculation and analysis of physiological data. By analyzing the static environment, the daily routine and physiological data (breathing, heart rate) of the elderly can be monitored, which helps to monitor the physical health of the elderly. For the dynamic environment, its monitoring algorithm pays more attention to events such as accidental falls and indoor intrusions. By detecting the occurrence of events in a timely manner and issuing warnings in a timely manner, it helps the elderly to obtain timely assistance when an accident occurs.

[0038] (2) Starting from the perspective of intelligent classification algorithms and the idea of multi-faceted care for the elderly, the present invention discriminates data types through the GA-SVM model, reducing the training time and improving the recognition accuracy. In the GA-SVM model provided by the present invention, for classification and recognition, the preprocessed CSI data will be classified and discriminated. The preprocessing target model is trained on the artificially labeled data set. The parameters of the support vector machine kernel function established by SVM and genetic algorithm are optimized, and the processor stores the recognition result information in the memory to train and update the training target model, which can further improve the classification and recognition accuracy, enhance the adaptability of the data to the current environment, improve the utilization rate of the data, and enhance the stability of the model.

[0039] (3) The elderly care method provided by the present invention starts from both static and dynamic scenarios. By classifying the application scenarios, the modularity and stability of the system can be effectively ensured. At the same time, corresponding training models are adopted for different scenarios, which can ensure the stability and accuracy of model training. And by continuously updating the target model in actual applications, the accuracy of the monitoring model can be effectively improved. Description of the Drawings

[0040] The drawings described herein are used to provide a further understanding of the present invention and form a part of this application. The illustrative embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention. In the drawings:

[0041] Figure 1 is a flowchart of the WiFi-based intelligent elderly care method provided in this embodiment;

[0042] Figure 2 is a specific working principle diagram of the WiFi-based intelligent elderly care method provided in this embodiment;

[0043] Figure 3 is a flowchart of the respiratory rate calculation provided in this embodiment;

[0044] Figure 4 is a flowchart of the heart rate calculation provided in this embodiment;

[0045] Figure 5 is a flowchart of the dynamic and static classification discrimination provided in this embodiment;

[0046] Figure 6 is a flowchart of the genetic - support vector machine training model provided in this embodiment. Detailed Embodiments

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0048] Embodiment 1

[0049] The following describes the WiFi-based multifunctional elderly care method of this embodiment with reference to the drawings. Refer to Figure 1 As shown, the WiFi-based multifunctional home elderly care method provided in this embodiment includes the following steps:

[0050] S1. Collect CSI data of the monitored home area through WiFi devices. Specifically, select suitable locations in the homes of the elderly in need of guardianship to install CSI collection and transmission devices to collect CSI data of the guardianship area. To increase the robustness of CSI, for the same area, multiple different locations may be selected to install transmission devices to ensure that the CSI signal is sensitive to changes at various locations within the area.

[0051] S2. Preprocess the collected CSI data to obtain preprocessed CSI data.

[0052] In this embodiment, step S2 is specifically as follows:

[0053] S2.1. Separate the collected raw CSI data to obtain amplitude data and phase data. The calculation method is as follows:

[0054] Where the k-th subcarrier h k is expressed as: real k and imag k are the real and imaginary parts of the k-th subcarrier, |h k | and ∠h k are the amplitude and phase of the k-th subcarrier;

[0055] S2.2. Interpolate and complete the amplitude data and phase data.

[0056] S2.3. Calculate the phase difference of the interpolated and completed phase data. The calculation method is as follows:

[0057]

[0058] Where is the phase of the n 1 , n 2 -th subcarriers, is the phase difference of the n 1 , n 2 -th subcarriers;

[0059] For the phase difference 1 , n 2 -th subcarriers , limit its value to the range of [-π, π]. The calculation method is as follows:

[0060] S2.4. Remove noise from the amplitude data and phase difference data through hampel filtering and principal component analysis.

[0061] S2.5. Select the optimal subcarrier amplitude data and optimal phase difference data from the amplitude data and phase difference data.

[0062] Specific reference Figure 1 As shown, the collected data information is transmitted to the preprocessing module, and the corresponding amplitude and phase angle are extracted from the collected CSI data. The amplitude and phase angle data contain the main information of the CSI. Next, the amplitude and phase data are complemented by linear interpolation to avoid data loss caused by packet loss and affect the results. Then, by calculating the phase difference of the phase data, a large amount of noise in the phase data can be eliminated. Then, by means of hampel filtering, each data in the measurement window that differs from the window median by more than the specified standard deviation is eliminated, so as to eliminate the influence of outliers in the sample. By means of principal component analysis, the useless signal components are filtered out, most of the useful information is retained, the data dimension is reduced, the complexity of subsequent operations is reduced, and the operation rate is improved. Then, by obtaining the subcarrier with the largest coefficient of variation or variance, the subcarrier with the most information is obtained, and the data suitable for using the training model for classification and discrimination in the amplitude data and phase difference data in the data is selected. The data contains more information, while reducing the burden of subsequent data calculation and improving the operation speed.

[0063] As an example, the original CSI data is transmitted to the preprocessing module for preprocessing of the collected original data. For a sampling frequency of f, the number of transmitting and receiving antennas is m and n, then the number of data packets collected in 1 s is about f×m×n. After preprocessing, only f amplitude data and f phase data are retained. The preprocessed data still retains the main information, while reducing the burden of subsequent operations and improving the calculation rate.

[0064] S3. Judge whether it belongs to a static environment or a dynamic environment according to the preprocessed CSI data;

[0065] In this embodiment, step S3 is specifically as follows:

[0066] S3.1. Intercept the amplitude data and phase data of each specified length period through a sliding window, and calculate the data mean. The calculation formula is as follows:

[0067] where |h k (t x ) and ∠h k (t x ) are the amplitude and phase of the kth subcarrier at time t x , and z is the number of subcarriers.

[0068] S3.2. Compare the calculation results with the static data and dynamic data in the memory to judge whether the current data acquisition period is in a static state or a dynamic state.

[0069] Specific reference Figure 1As shown, after the CSI raw data is processed by the preprocessing module, the corresponding environment will be identified, and the data characteristics of the amplitude and phase difference information of the preprocessed CSI will be used to determine whether the scene is a dynamic scene or a static scene.

[0070] The CSI amplitude and phase difference data obtained after preprocessing in the current scenario are normalized according to the following formula:

[0071]

[0072]

[0073] Where A(t x ), t x The values ​​of the amplitude and phase difference at the moment, minA and maxA are the minimum and maximum amplitude in the window respectively. They are the minimum and maximum values ​​of the phase difference in the window respectively. The current environment is identified by the following formula:

[0074]

[0075] Among them, f(t x )、f(t x+1 ) is t in static environment x ,t x+1 The amplitude at the moment, g(t x )、g(t x+1 ) is t in static environment x ,t x+1 The phase difference at time, f p (t x ), g p (t x ) is t in static environment x The normalized amplitude and phase difference at time, f p (t x+1 ), g p (t x+1 ) is t in static environment x+1 The amplitude and phase difference after normalization at each moment, z is the number of subcarriers. If the left equation is greater than or equal to the right equation, the current environment is identified as a dynamic environment, otherwise the current environment is identified as a static environment.

[0076] S4. When it is determined to be a static environment, the pre-processed CSI data is input into a static data processor, and the static data processor obtains a first pre-trained target model after machine learning, and identifies the three conditions of "daze, sleep, and no one" for the pre-processed CSI information through the first pre-trained target model, and calculates the heart rate and breathing conditions during sleep time;

[0077] Specifically, when the judgment scenario is a static environment, the preprocessed image information obtained by the preprocessing module is transmitted to the processor for data processing. The recognition and classification tasks for three specific situations are mainly achieved through data transmission and processing with the established first pre-trained target model. The pre-classified data in the real world is recognized and classified by the trained first pre-trained target model.

[0078] In this embodiment, the process of step S4 is as follows:

[0079] S4.1. Initialize the first target model before pre-training;

[0080] S4.2. Repeatedly collect training data for static environments in different scenarios. After preprocessing the training data, generate the first pre-trained target model and save it in the memory;

[0081] S4.3. Perform band-pass filtering on the optimal amplitude data to obtain spectral data within the breathing range, and calculate the breathing rate through fast Fourier transform;

[0082] Reference Figure 3 , the calculation method of the breathing rate of a person in a static environment in step S4.3 is as follows: Specifically, after filtering out most of the noise through band-pass filtering, the corresponding spectrum can be obtained through fast Fourier transform. The frequency corresponding to the maximum value of the spectrum amplitude is the breathing frequency. Through the formula:

[0083] v 呼 = f 呼 × 60

[0084] where v 呼 is the breathing rate and f 呼 is the breathing frequency.

[0085] S4.4. Extract the DWT detail coefficients from the optimal phase data, obtain spectral data within the heart rate range through band-pass filtering, and calculate the heart rate through fast Fourier transform.

[0086] Reference Figure 4 , the calculation method of the heart rate of a person in a static environment in step S4.4 is as follows:

[0087] For the relationship between the phase difference Δ∠h and the approximation coefficient a L and the detail coefficient d l is as follows:

[0088]

[0089] where L is the decomposition times, n s is the number of data packets, is the wavelet basis.

[0090] The corresponding spectrum is obtained through fast Fourier transform, and the frequency corresponding to the maximum value of the spectrum amplitude is the heart rate. The calculation formula is as follows: v 心率 = f 心跳 × 60

[0091] where v 心率 is the heart rate, and f 心跳 is the heart beat frequency.

[0092] In this embodiment, in step S4.2, the training data of the static environment in different scenarios is collected as follows:

[0093] S4.2.1. Set up a home environment; S4.2.2. The required scenarios for the experiment are "someone is in a daze in the environment", "someone is sleeping in the environment", and "no one is in the environment". Multiple groups of data are collected by replacing different people to achieve the collection of excellent training data.

[0094] Specifically for the training model GA-SVM, the selected hyperplane equation is:

[0095] w T x + b = 0

[0096] where w is the normal vector of the hyperplane, b represents the distance between the hyperplane and the origin, x is the data vector, and y is the label.

[0097] When y is the label of the data point and its value is -1 or 1, the distance from the label to the hyperplane is represented by the distance formula y(w T x + b), then the objective function is:

[0098]

[0099] where w is the normal vector of the hyperplane, b represents the distance between the hyperplane and the origin, x is the data vector, and y is the label.

[0100] Correspondingly, the dual problem of the above formula is:

[0101] where,

[0102] The KKT conditions of this dual problem are:

[0103] where α is the normal vector of the hyperplane, α = (α 1 , α 2 , α 3 ,..., α ns ), x i and x j are the decision variables of the original problem, y i and y jis the decision variable of the dual problem.

[0104] The training data is passed to the training model GA - SVM for training. The selected kernel function is the RBF function, and the expression of the RBF function is:

[0105] where \(x\) is the data vector, \(y\) is the label, and \(\gamma\) is the regularization coefficient.

[0106] Combined with Figure 6 As shown, after constructing the kernel function and the objective function, the training model GA - SVM autonomously optimizes and updates the regularization coefficient \(\gamma\) and the penalty coefficient \(C\), calculates the accuracy through non - training set data, and then finds the optimal parameter solution.

[0107] After initializing the parameters, the binary encoding method is selected. The regularization coefficient \(\gamma\) and the penalty coefficient \(C\) are encoded. The fitness function is used to evaluate the current data, calculate the accuracy of the current model, and record the optimal parameters. Then, through the crossover algorithm and mutation, new parameters are generated for the currently selected parameters, and a new model is generated to calculate the accuracy of the new model. Finally, the optimal SVM model is obtained.

[0108] During the static environmental monitoring process, when a person is recognized in the environment, the computing module will calculate the physiological conditions (respiration rate and heart rate) of the person in the environment. When the respiration and heart rate in the environment are lower than the set thresholds, an alarm will be issued.

[0109] S5. When it is determined to be a dynamic environment, the pre - processed CSI data is input into the dynamic data processor. The dynamic data processor obtains the second pre - trained target model after machine learning and uses the second pre - trained target model to identify dangerous behaviors in the pre - processed CSI information.

[0110] For the training model GA - SVM, the selected hyperplane equation is: \(w\) T \(x + b=0\)

[0111] When \(y\) is the label of the data point and its value is - 1 or 1, the distance from the label to the hyperplane can be expressed by the distance formula \(y(w\) T \(x + b)\), then the objective function is:

[0112]

[0113] where \(w\) is the normal vector of the hyperplane, \(b\) represents the distance between the hyperplane and the origin, \(x\) is the data vector, and \(y\) is the label.

[0114] Correspondingly, the dual problem of the above formula is:

[0115] where,

[0116] The KKT conditions of the dual problem are as follows:

[0117] where α is the normal vector of the hyperplane, α = (α 1 , α 2 , α 3 ,..., α ns ), x i and x j are the decision variables of the original problem, and y i and y j are the decision variables of the dual problem.

[0118] The training data is passed to the SVM for training, and the selected kernel function is the RBF function. The expression of the RBF function is:

[0119] where x is the data vector, y is the label, and γ is the regularization coefficient.

[0120] As shown in combination with Figure 6 , after constructing the kernel function and the objective function, the training model GA - SVM autonomously optimizes and updates the regularization coefficient γ and the penalty coefficient C, calculates the accuracy through the non - training set data, and then finds the optimal parameter solution.

[0121] After initializing the parameters, the binary encoding method is selected. The regularization coefficient γ and the penalty coefficient C are encoded, the current data is evaluated through the fitness function, the accuracy of the current model is calculated, and the optimal parameters are recorded. Then, through the crossover algorithm and mutation, new parameters are generated for the currently selected parameters, a new model is generated to calculate the accuracy of the new model, and finally the optimal SVM model is obtained.

[0122] When detecting the dynamic environment, when behaviors such as "the elderly falling" and "indoor intrusion" are detected, the system will issue an alarm.

[0123] S5.1. Initialize the second target model before pre - training;

[0124] S5.2. Repeatedly collect the training data of the dynamic environment, pre - process the training data, generate the second pre - training target model, and save it to the memory.

[0125] The process of collecting the training data of the dynamic environment in step S5.2 of this embodiment is as follows:

[0126] S5.2.1. Build a home environment;

[0127] S5.2.2. In the experimental scenario, simulate indoor intrusion and elderly people falling, and collect multiple sets of data by replacing different people to achieve the collection of high-quality training data.

[0128] S6. The static data processor and the dynamic data processor store the result information in the memory, and train and update the pre-trained target model.

[0129] It should be noted that in the static environment recognition and dynamic environment recognition provided by the present invention, after the model is updated every Q times, the latest data is transmitted to the target model to realize the training and updating of the target model, thereby ensuring the stability of the model training. By using more accurate data, the model parameters are updated, which can effectively improve the accuracy of model classification and recognition.

[0130] Example 2

[0131] like Figure 2 As shown, this embodiment provides a WiFi-based intelligent monitoring system for the elderly. The system mainly includes a CSI information collection module, a preprocessing module, a static and dynamic discrimination module, a CPU processor, and a storage module. The specific functions of each module are as follows:

[0132] The CSI information collection module is used to collect CSI data in the environment in real time and transmit the collected data to the preprocessing module;

[0133] The preprocessing module is used to preprocess the collected CSI data, obtain amplitude data and phase difference data, and transmit the data to the static and dynamic recognition module and the CPU processor;

[0134] The static and dynamic discrimination module is used to classify the environment into a static environment and a dynamic environment, and feed back the discrimination result to the CPU processor;

[0135] The CPU processor will select the corresponding pre-trained model according to the environmental conditions. If the environment is a static environment, the first pre-trained target model will be called to classify and distinguish the three conditions of "daze", "sleep" and "no one" in the environment. When the environment is detected as "daze" and "sleep", the breathing rate and heart rate of the elderly in the environment will be calculated. When the breathing rate is lower than the threshold, the result will be sent to the server. After receiving the result, the cloud server will alarm. If the environment is a dynamic environment, the second pre-trained target model will be called to classify and distinguish "elderly fall", "indoor intrusion" and other conditions in the environment. When the two conditions of "elderly fall" and "indoor intrusion" are identified, the results will be sent to the cloud server. After receiving the result, the cloud server will alarm and regularly update the first pre-trained target model and the second pre-trained target model.

[0136] A storage module, used to store the first pre-trained target model, the second pre-trained target model, and the results of classification and discrimination of the two models and the corresponding amplitude data and phase difference data, wherein the stored data is used for classification and discrimination and model updating;

[0137] In this embodiment, the first pre-trained target model and the second pre-trained target model are both trained by GA-SVM, and the specific implementation can refer to the above embodiment. It should be noted that the static and dynamic discrimination module enables the entire CPU processor to call only one pre-trained target model each time it operates, and whether to calculate the heart rate and respiratory rate will be determined based on the results of the static and dynamic discrimination module. Compared with simply classifying and discriminating "dazed", "sleeping", "no one", "elderly falling", "indoor invasion", and other situations at the same time, the CPU processor operation amount is greatly reduced, the complexity of model training is reduced, and the CPU processor's computing power requirements are relatively low. Since each model has fewer classification situations, the classification accuracy is improved, and the time required for each classification is less. The application of GA-SVM optimizes the penalty factor C and the regularization coefficient γ through the GA algorithm, and the GA algorithm is an efficient and global search and optimization method with strong solution ability and wide adaptability. Compared with the traditional SVM parameter value, it is more reasonable, and the classification performance and the running speed of the system are improved. Each update of the model will improve the accuracy of the system, make the model more suitable for the current environment, and greatly improve the robustness and accuracy of the system.

[0138] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. An intelligent elderly care method based on WiFi, characterized in that, the intelligent elderly care method includes the following steps: S1. Collect CSI data of the monitored home area through WiFi devices; S2. Preprocess the collected CSI data to obtain preprocessed CSI data, and obtain amplitude data and phase difference data; S3. Judge whether it belongs to a static environment or a dynamic environment according to the preprocessed CSI data. The process is as follows: S3.

1. Intercept the amplitude data and phase data of each specified-length time period through a sliding window, and calculate the data mean. The calculation formula is as follows: where |h k (t x )| and ∠h k (t x ) are the amplitude and phase of the k-th subcarrier at time t x , and z is the number of subcarriers; S3.

2. Compare the calculation results with the static data and dynamic data in the memory to judge whether the current data acquisition period is in a static state or a dynamic data state. The process is as follows: For the CSI amplitude and phase difference data obtained after preprocessing in the current scenario, normalize the amplitude and phase difference data according to the following formula: where A(t x ), is the value of the amplitude and phase difference at time t, minA and maxA are respectively the minimum and maximum values of the amplitude within the window, x and are respectively the minimum and maximum values of the phase difference within the window. The current environment is identified through the following formula: Among them, f(t x ) and f(t x+1 ) are the amplitudes at times t x and t x+1 in a static environment, g(t x ) and g(t x+1 ) are the phase differences at times t x and t x+1 in a static environment, f p (t x ) and g p (t x ) are the normalized amplitude and phase difference at time tx in a static environment, f p (t x+1 ) and g p (t x+1 ) are the normalized amplitude and phase difference at time t x+1 in a static environment, z is the number of subcarriers. If the left side is greater than or equal to the right side, the current environment is identified as a dynamic environment; otherwise, the current environment is identified as a static environment. S4. When it is judged to be a static environment, input the preprocessed CSI data into a static data processor. The static data processor obtains a first pre-trained target model after machine learning, and uses the first pre-trained target model to identify the three situations of "daze", "sleep" and "no one" for the preprocessed CSI information, and calculate the heart rate and breathing conditions during sleep time; S5. When it is judged to be a dynamic environment, input the preprocessed CSI data into a dynamic data processor. The dynamic data processor obtains a second pre-trained target model after machine learning, and uses the second pre-trained target model to identify dangerous behaviors for the preprocessed CSI information.

2. An intelligent elderly care method based on WiFi according to claim 1, characterized in that, the process of step S2 is as follows: S2.

1. Separate the collected original CSI data to obtain amplitude data and phase data; S2.

2. Interpolate and complete the amplitude data and phase data; S2.

3. Calculate the phase difference for the interpolated and completed phase data; S2.

4. Remove noise from the amplitude data and phase difference data by means of hampel filtering and principal component analysis; S2.

5. Select the optimal subcarrier amplitude data and optimal phase difference data from the amplitude data and phase difference data.

3. An intelligent elderly care method based on WiFi according to claim 1, characterized in that, the process in step S4 is as follows: S4.

1. Initialize the first target model before pre-training; S4.

2. Repeatedly collect training data in static environments under different scenarios, generate the first pre-trained target model after preprocessing the training data, and save it to the memory; S4.

3. Perform band-pass filtering on the optimal amplitude data to obtain spectral data within the breathing range, and calculate the breathing rate through fast Fourier transform; S4.

4. Extract the DWT detail coefficients from the optimal phase data, obtain spectral data within the heart rate range through band-pass filtering, and calculate the heart rate through fast Fourier transform.

4. An intelligent elderly care method based on WiFi according to claim 3, characterized in that, The scenarios required for collecting training data in the static environment in step S4.2 include: "Someone is in a daze in the environment", "Someone is sleeping in the environment", and "No one is in the environment". Multiple groups of data are collected by replacing different people to achieve the collection of excellent training data.

5. A method for intelligent elderly care based on WiFi according to claim 1, characterized in that the process in step S5 is as follows: S5.1, initialize the second target model before pre-training; S5.2, repeatedly collect training data in the dynamic environment, generate the second pre-training target model through the training data, and save it to the memory.

6. A method for intelligent elderly care based on WiFi according to claim 5, characterized in that the scenarios required for collecting training data in the dynamic environment in step S5.2 include: "The elderly fall" and "Indoor intrusion". Multiple groups of data are collected by replacing different people to achieve the collection of excellent training data.

7. A method for intelligent elderly care based on WiFi according to claim 1, characterized in that the first pre-training target model and the second pre-training target model are trained by GA-SVM, and the optimal penalty factor C and regularization coefficient γ are obtained through the GA algorithm.

8. A method for intelligent elderly care based on WiFi according to claim 1, characterized in that the method for intelligent elderly care further includes the following steps: S6, the static data processor and the dynamic data processor store the result information in the memory, and train and update the first pre-training target model and the second pre-training target model.

Images