WIFI CSI-based behavior prediction model system and execution method therefor

Abstract

According to the described present invention, fine movement and presence can be detected in an indoor environment through changes in the amplitude and phase of a signal by using WiFi CSI, and abnormal behavior can be predicted on the basis of the detection. In addition, according to the present invention, AI learns and classifies reflection, distortion, and the like of packet data contained in a WiFi signal waveform, and thus can detect and monitor related phenomena. In addition, according to the present invention, the location of an object located in a wide space and the movement path of the object can be tracked merely through WiFi sensing, flexible application to environmental changes is possible due to AI-based automatic learning, and privacy protection for individuals is possible through a non-camera method.

Description

WIFI CSI-based behavior prediction model system and its execution method

[0001] The present invention relates to a WiFi CSI-based behavior prediction model system and an execution method thereof, and more specifically, to a WiFi CSI-based behavior prediction model system and an execution method thereof that can detect minute movements and presence in an indoor environment through amplitude and phase changes of signals using WiFi CSI, predict abnormal behavior based on the same, recognize a user's location without blind spots, correct accuracy through an AI model, and control IoT-linked control.

[0002]

[0003] With the recent advancement of Internet of Things (IoT) technology, interest in building smart environments is rapidly increasing. In particular, active research is being conducted to maximize user convenience and energy efficiency in smart home and smart building systems. A critical element of these systems is accurately identifying the number of people in a room and providing services based on this information.

[0004] Conventional methods for indoor occupancy counting have primarily utilized sensors such as video cameras, infrared sensors, ultrasonic sensors, and mmWave, as well as BLE and UWB (Ultra-wideband)-based location tracking systems. However, these methods are significantly expensive to purchase and install due to the physical sensors and related equipment. Covering large spaces requires a large number of sensors, further increasing costs.

[0005] Additionally, installing these sensors requires selecting the optimal location, can be complex, and can lead to privacy violations, blind spots, increased installation costs, and reliance on additional equipment, along with significant maintenance costs.

[0006] The detection range of these physical sensors is limited, and their performance is heavily dependent on the installation environment. For example, obstacles or complex indoor structures can negatively impact detection accuracy. As an alternative, approaches utilizing WiFi signals are currently being studied.

[0007] WiFi Channel State Information (CSI) provides detailed information about the signal propagation path at the physical layer of a WiFi network. This information includes changes in signal amplitude and phase, enabling the detection of subtle motion and presence in indoor environments.

[0008] Korean Patent Publication No. 2024-0030003 describes an indoor occupancy detection system using physical sensors, focusing on sensor placement and data processing methods. However, cost and installation challenges remain.

[0009] Previously, Korean Patent No. 10-2659202 relates to a CNN-based human activity recognition system and method using a Wi-Fi signal, and the CNN-based human activity recognition system using a Wi-Fi signal discloses a configuration of an input unit that is connected to a terminal that receives a Wi-Fi signal transmitted from a Wi-Fi AP and inputs raw CSI data of the Wi-Fi signal; an extraction unit that extracts a spectrum amplitude value of the raw CSI data; and a CNN classification unit that extracts feature points for defining a human activity between the Wi-Fi AP and the terminal based on the spectrum amplitude value and classifies and learns the activity based on the extracted feature points. However, it does not suggest an additional configuration for AI to learn and classify data changes such as reflection and distortion of packet data to determine whether an abnormality has occurred.

[0010] In addition, when acquiring data in spaces with different indoor structures or interference environments, there may be inconveniences in having to retrain the AI ​​model from scratch to fit the indoor structure or interference environment, and there is a need for a meta-learning-based structure that can normalize pattern differences in CSI data according to various conditions.

[0011] Additionally, there is a need to classify the behavioral status of people located in indoor spaces in real time and to detect abnormal conditions such as falls, trips, and cardiac arrests early and respond quickly.

[0012] Therefore, there is a need for a WiFi CSI-based behavior prediction model system and its execution method that can detect subtle movements and presence in an indoor environment through amplitude and phase changes of signals using WiFi CSI and predict abnormal behavior based on the same.

[0013]

[0014] The present invention aims to provide a WiFi CSI-based behavior prediction model system and an execution method thereof that can detect minute movements and presence in an indoor environment through amplitude and phase changes of signals using WiFi CSI and predict abnormal behavior based on the same.

[0015] In addition, the present invention aims to provide a WiFi CSI-based behavior prediction model system and an execution method thereof that enable AI to learn and classify reflections, distortions, etc. of packet data contained in a WiFi signal waveform, thereby detecting and monitoring related phenomena.

[0016] The objectives of the present invention are not limited to those mentioned above. Other objectives and advantages of the present invention not mentioned above can be understood through the following description and will be more clearly understood through the embodiments of the present invention. Furthermore, it will be readily apparent that the objectives and advantages of the present invention can be realized by the means and combinations thereof set forth in the claims.

[0017]

[0018] To achieve these objectives, a WiFi CSI-based behavior prediction model system includes: a noise detection AI model learning unit that generates a noise detection AI model using normal signal data and noise data; an object classification and pattern classification AI model learning unit that generates an object classification and pattern classification AI model using CSI data collected through a WiFi transmitter and receiver in various environments and situations; an abnormality occurrence prediction AI model learning unit that generates an abnormality occurrence prediction AI model using time domain features, frequency domain features, and spatial domain features; a WiFi CSI signal collection and preprocessing unit that deletes noise if noise is included in a WiFi CSI signal using the noise detection AI model, and if the noise is deleted, executes labeling for each object using the object classification and pattern classification AI model and then stores object pattern data; and an abnormality occurrence prediction unit that collects and analyzes spatial data and then analyzes the behavior pattern type for each object and then checks whether there is an abnormality sign using the abnormality occurrence prediction AI model and provides the result to an administrator terminal.

[0019] In one embodiment, the noise detection AI model learning unit may collect normal signal data and noise data, extract features from the data, label each data with a normal signal label or a noise label, select a machine learning or deep learning model suitable for noise detection, and then generate a noise detection AI model.

[0020] In one embodiment, the object classification and pattern classification AI model learning unit collects CSI data through a WiFi transmitter and receiver in various environments and situations, labels each data sample with a label that can identify the corresponding object, selects a machine learning or deep learning model suitable for object classification and pattern classification, and trains the model to create an object classification and pattern classification AI model.

[0021] In one embodiment, the above-described abnormal situation occurrence prediction AI model learning unit collects time domain data and spatial domain data of an object, labels the data with object behavior patterns and situation occurrence signs, extracts meaningful features to analyze the object's behavior patterns, evaluates the importance of the features, and then selects features according to the importance to configure an abnormal situation occurrence sign prediction AI model.

[0022] In addition, a WiFi CSI-based behavior prediction model method executed in a WiFi CSI-based behavior prediction model system to achieve these purposes includes a step of creating a noise detection AI model using normal signal data and noise data, a step of creating an object segmentation and pattern classification AI model using CSI data collected through a WiFi transmitter and receiver in various environments and situations, a step of creating an abnormal situation occurrence prediction AI model using time domain features, frequency domain features, and spatial domain features, a step of deleting noise if noise is included in a WiFi CSI signal using the noise detection AI model, and if the noise is deleted, a step of executing labeling for each object using the object segmentation and pattern classification AI model and then storing object pattern data, and a step of analyzing the behavior pattern type for each object after collecting and analyzing spatial data, and then checking whether there is an abnormal situation sign using the abnormal situation occurrence prediction AI model and providing the result to an administrator terminal.

[0023]

[0024] According to the present invention as described above, there is an advantage in that it is possible to detect minute movements and presence in an indoor environment through amplitude and phase changes of signals using WiFi CSI and predict abnormal behavior based on this.

[0025] In addition, according to the present invention, there is an advantage in that AI can learn and classify reflections, distortions, etc. of packet data contained in a WiFi signal waveform, and detect and monitor related phenomena.

[0026] In addition, according to the present invention, the location of an object located in a wide space and the movement path of the object can be tracked using only WiFi sensing, and there are advantages in that flexible application to environmental changes is possible due to AI-based automatic learning, and privacy protection for individuals is possible through a non-camera method.

[0027] In addition, according to the present invention, there is an advantage in that learning time can be shortened through fine tuning based on a pre-learned model through automatic environmental exploration and feature extraction, and accuracy can be improved by adapting with minimal data after field installation.

[0028] In addition, according to the present invention, it is possible to analyze behavioral states while ensuring privacy by applying it to users who require care, such as in silver towns, hospitals, restrooms, and single-person households, and thus has the advantage of being able to quickly and automatically detect dangerous states, such as prolonged inactivity or collapse.

[0029]

[0030] FIG. 1 is a network configuration diagram for explaining a WiFi CSI-based behavior prediction model system according to one embodiment of the present invention.

[0031] FIG. 3 is a flowchart illustrating one embodiment of a behavior prediction model method based on WiFi CSI according to the present invention.

[0032] FIG. 3 is a flowchart illustrating another embodiment of a behavior prediction model method based on WiFi CSI according to the present invention.

[0033] Figure 4 is an example diagram for explaining the execution process of Figure 3.

[0034] FIG. 5 and FIG. 6 are flowcharts illustrating another embodiment of a behavior prediction model method based on WiFi CSI according to the present invention.

[0035]

[0036] A noise detection AI model learning unit that collects normal signal data and noise data for WiFi CSI data and then uses them to create a noise detection AI model;

[0037] A learning unit for object classification and pattern classification AI models that collects CSI data through WiFi transmitters and receivers in various environments and situations, labels objects based on the movement and stationary status of each object detected from the collected data, and creates an object classification and pattern classification AI model;

[0038] An abnormal situation occurrence prediction AI model learning unit that collects time domain data including sensor signals, log data, and traffic data for objects using multiple sensors, and spatial domain data including image and video 3D coordinate data;

[0039] A WiFi CSI signal collection and preprocessing unit that removes noise contained in a WiFi CSI signal using a noise detection AI model, performs object-specific labeling using an object segmentation and pattern classification AI model, and then stores object pattern data;

[0040] It includes an abnormal situation occurrence prediction unit that collects and analyzes spatial data, analyzes the type of behavior pattern for each object, and then uses an abnormal situation occurrence prediction AI model to check whether there are signs of an abnormal situation and provides the result to the administrator terminal.

[0041] The above noise detection AI model learning unit

[0042] After extracting features from the collected normal signal data and noise data, label the normal signal label or noise label for each extracted data, select a machine learning or deep learning model suitable for noise detection to build a noise detection AI model, separate the data into training data and validation data, train the model to distinguish between normal signals and noise using the training data, and evaluate the performance of the model using the validation data to input the feature vector of the training data sample and the corresponding label into the noise detection AI model.

[0043] The AI ​​model learning department that predicts the occurrence of the above abnormal situation

[0044] A WiFi CSI-based behavior prediction model system characterized by including an AI model for predicting situational signs by collecting time domain data and space domain data of an object, labeling the data with object behavior patterns and situational signs, extracting features for analyzing the object's behavior patterns, evaluating the importance of the extracted features, and selecting features based on the evaluated importance.

[0045]

[0046] The above-described objects, features, and advantages will be described in detail below with reference to the accompanying drawings, so that those skilled in the art can easily practice the technical idea of ​​the present invention. In describing the present invention, if it is determined that a detailed description of known technologies related to the present invention may unnecessarily obscure the gist of the present invention, a detailed description thereof will be omitted. Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the drawings, the same reference numerals are used to indicate the same or similar components.

[0047] Among the terms used in this specification, "WiFi CSI (Channel State Information) data" refers to information that represents the physical characteristics of a WiFi signal, providing detailed information about how the signal changed along the path it took to reach the receiver. This information primarily includes changes in the signal's amplitude and phase, and can be useful for detecting subtle changes in the indoor environment.

[0048] FIG. 1 is a network configuration diagram for explaining a WiFi CSI-based behavior prediction model system according to one embodiment of the present invention.

[0049] Referring to FIG. 1, the WiFi CSI-based behavior prediction model system includes a WiFi CSI signal collection and preprocessing unit (110), a noise detection AI model learning unit (120), an object classification and pattern classification AI model learning unit (130), an abnormal situation occurrence prediction AI model learning unit (140), and an abnormal situation occurrence prediction unit (150).

[0050] The WiFi CSI signal collection and preprocessing unit (110) uses the noise detection AI model (121) generated by the noise detection AI model learning unit (120) to delete noise if noise is included in the WiFi CSI signal.

[0051] The WiFi CSI signal collection and preprocessing unit (110) executes object-specific labeling using the object classification and pattern classification AI model (131) generated by the object classification and pattern classification AI model learning unit (130) after noise has been removed, and then stores object pattern data.

[0052] The abnormal situation occurrence prediction unit (150) collects and analyzes spatial data, analyzes the behavior pattern type for each object, and then uses the abnormal situation occurrence prediction AI model generated by the abnormal situation occurrence prediction AI model learning unit (140) to check whether there are signs of an abnormal situation and provides the result to the administrator terminal.

[0053] The noise detection AI model learning unit (120) collects normal signal data and noise data and then uses them to create a noise detection AI model (121). At this time, the data is presented in matrix form, with each row representing an antenna and each column representing a subcarrier.

[0054] In one embodiment, the noise detection AI model training unit (120) collects normal WiFi CSI data in various environments. For example, the noise detection AI model training unit (120) collects data under various conditions, such as in an unoccupied room, a room with moving people, and a room with furniture.

[0055] In another embodiment, the noise detection AI model training unit (120) collects WiFi CSI data in various noise situations. For example, the noise detection AI model training unit (120) collects data in various situations, such as electromagnetic interference, interference from other WiFi networks, and interference from specific devices.

[0056] After that, the noise detection AI model learning unit (120) removes noise from the collected data to improve the quality of the signal, and normalizes the signal intensity and phase values ​​to make the range of the model input values ​​constant. At this time, noise is removed and time-series features are extracted through a moving average filter, DWT (Discrete Wavelet Transform), FFT (Fast Fourier Transform), etc.

[0057] Then, the noise detection AI model learning unit (120) can extract features from the data and then label each data with a normal signal label if it is normal signal data, and with a noise label if it is noise data.

[0058] The noise detection AI model learning unit (120) selects a machine learning or deep learning model suitable for noise detection and then builds a noise detection AI model (121).

[0059] To this end, the noise detection AI model learning unit (120) separates the data into training data and validation data, trains the model using the training data, and evaluates the model's performance using the validation data. The training data is used to train the model to distinguish between normal signals and noise. Therefore, the noise detection AI model learning unit (120) inputs the feature vectors and corresponding labels of the training data samples into the noise detection AI model.

[0060] In one embodiment, the noise detection AI model learning unit (120) calculates the difference between the label predicted by the model and the actual label using a loss function (e.g., cross entropy loss function, etc.), and updates the weights of the noise detection AI model (121) using a backpropagation algorithm to minimize the loss function.

[0061] Next, the noise detection AI model training unit (120) evaluates the performance of the noise detection AI model using the validation data during training. This data, separate from the training data, is used to determine how well the noise detection AI model (121) generalizes.

[0062] As described above, the noise detection AI model learning unit (120) uses test data to evaluate the final model performance. The noise detection AI model receives the feature vector of the test data, predicts the label, and compares this prediction result with the actual label to evaluate performance.

[0063] The object classification and pattern classification AI model learning unit (130) collects CSI data through a WiFi transmitter and receiver in various environments and situations, and then performs labeling for each data sample to identify the corresponding object.

[0064] Additionally, data collected from multiple sensors for the same situation are integrated and analyzed by applying one of the fusion algorithms, either the channel-to-channel shared method or the confidence score-based method, to the CSI data collected through the WiFi transmitter and receiver.

[0065] In one embodiment, the object classification and pattern classification AI model learning unit (130) collects data including moving people, fixed objects, furniture, etc. in various locations, and labels the data of moving people as 'people', the data of fixed objects as 'objects', etc.

[0066] After that, the object classification and pattern classification AI model learning unit (130) removes noise from the CSI data, normalizes the signal intensity and phase values ​​to a certain range, matches the range of the model input values ​​to a certain range, and extracts useful features in the time and frequency domains of the signal. For example, the features may include signal amplitude, phase change, FFT transformation results, etc. At this time, the object classification and pattern classification AI model learning unit (130) stores the feature vector and label in pairs.

[0067] The object classification and pattern classification AI model learning unit (130) selects a machine learning or deep learning model suitable for object classification and pattern classification. For example, machine learning or deep learning models suitable for object classification and pattern classification may include, but are not limited to, a convolutional neural network (CNN), a recurrent neural network (RNN), a transformer model, and a meta learning (MAML) model.

[0068] Then, the object classification and pattern classification AI model learning unit (130) configures the selected model. For example, in the case of CNN, it is configured with an input layer, multiple convolutional layers, a pooling layer, and a fully connected layer, in the case of RNN, LSTM (Long Short-Term Memory) or GRU (Gated Recurrent Unit) can be used, in the case of Transformer, an attention mechanism, etc. can be utilized, and in the case of MAML, a meta-learning structure can be merged to configure an environment-adaptive location prediction module, and other machine learning or deep learning models suitable for object classification and pattern classification other than these models can be configured as another component. In addition, it can be utilized as a structure that quickly fine-tunes even with a small amount of data based on learned model parameters and pre-learning weights, etc.

[0069] The object classification and pattern classification AI model learning unit (130) divides the collected data into training data and test data, inputs the data's feature vector and label into the object classification and pattern classification AI model, and calculates the difference between the label predicted by the object classification and pattern classification AI model and the actual label using a loss function (e.g., cross entropy loss function).

[0070] After that, the object classification and pattern classification AI model learning unit (130) updates the weights of the object classification and pattern classification AI model using a backpropagation algorithm to minimize the loss function, and adjusts hyperparameters such as the learning rate, batch size, and number of epochs to optimize the performance of the object classification and pattern classification AI model.

[0071] The abnormal situation occurrence prediction AI model learning unit (140) collects time domain data and spatial domain data of objects using various sensors.

[0072] In one embodiment, the abnormal situation occurrence prediction AI model learning unit (140) may collect object movement data using sensors based on lidar, radar, cameras, or wireless signals. For example, time domain data may include sensor signals, log data, traffic data, etc., while spatial domain data may include images, videos, and 3D coordinate data.

[0073] The abnormal situation occurrence prediction AI model learning unit (140) labels object behavior patterns and situation occurrence signs in the data. For example, the abnormal situation occurrence prediction AI model learning unit (140) labels object behavior patterns such as walking, running, and stopping, and labels situation occurrence signs such as abnormal behavior, intrusion, and abnormal movement.

[0074] At this time, the abnormal situation occurrence prediction AI model learning unit (140) includes a posture classification AI model learning unit (141) to determine the status of the current behavior of the object, such as sitting, standing, lying down, falling, or not responding.

[0075] Then, the abnormal situation occurrence prediction AI model learning unit (140) extracts meaningful features to analyze the object's behavioral pattern. Here, the features may include changes in position, changes in speed, acceleration, path patterns, etc.

[0076] The detailed classification AI model learning unit (141) determines that an abnormal state of an object, i.e., a state including lying down, falling down, and unresponsiveness, continues based on significant features extracted from the behavioral pattern of the extracted object, and provides this to the administrator terminal.

[0077] Meanwhile, an abnormal state is defined as one or more of the following: maintaining the same posture for a certain period of time, detecting a pattern of rapid falls from a specific object, lack of movement toward a specific object, and suspected cardiac arrest.

[0078] In one embodiment, the abnormal situation occurrence prediction AI model learning unit (140) can extract time domain features of statistical features (e.g., mean, variance, standard deviation, maximum value, minimum value, etc.) and peak features (number of peaks, interval between peaks, peak height, etc.).

[0079] In another embodiment, the abnormal situation occurrence prediction AI model learning unit (140) can convert a signal into a frequency domain and then analyze the frequency components to extract frequency domain features such as main frequency components, spectrum peaks, and energy of a frequency band.

[0080] In another embodiment, the abnormal situation occurrence prediction AI model learning unit (140) can extract edge detection, corner detection, histogram-based features, etc. from image data, and extract spatial domain features such as object tracking, motion vectors, and frame-to-frame changes from video data.

[0081] The abnormal situation occurrence prediction AI model learning unit (140) generates composite features using time domain features, frequency domain features, and spatial domain features.

[0082] In one embodiment, the abnormal situation occurrence prediction AI model learning unit (140) analyzes temporal change patterns by analyzing time-frequency composite features. For example, the abnormal situation occurrence prediction AI model learning unit (140) calculates the temporal average, variance, etc. of major frequency components after Fourier transform, and calculates the number of peaks, inter-peak intervals, etc. in a specific frequency band after wavelet transform.

[0083] In another embodiment, the abnormal situation occurrence prediction AI model learning unit (140) analyzes spatial-temporal composite features to combine the temporal positional changes of a specific object with spatial features in the image. For example, the abnormal situation occurrence prediction AI model learning unit (140) may combine the object's movement path with features such as edges and corners in each frame. In another example, the abnormal situation occurrence prediction AI model learning unit (140) may combine the object's motion vector with histogram-based features of the image.

[0084] In another embodiment, the abnormal situation occurrence prediction AI model learning unit (140) combines frequency domain features and spatial features of image data. For example, the abnormal situation occurrence prediction AI model learning unit (140) may combine major frequency components and spatial features such as edges and corners after Fourier transforming the image. In another example, the abnormal situation occurrence prediction AI model learning unit (140) may combine energy distribution in the frequency domain with texture features of the image.

[0085] After that, the abnormal situation occurrence prediction AI model learning unit (140) uses a random forest model to evaluate the importance of each feature.

[0086] In one embodiment, the abnormal situation occurrence prediction AI model learning unit (140) uses a random forest model to evaluate the importance of each feature.

[0087] In the above embodiment, the abnormal situation occurrence prediction AI model learning unit (140) trains the model, calculates the degree to which each feature contributes to the prediction performance, and selects features with high importance to use in the final model.

[0088] In another embodiment, the abnormal situation occurrence prediction AI model learning unit (140) uses an XGBoost model to evaluate the importance of each feature.

[0089] In the above embodiment, the abnormal situation occurrence prediction AI model learning unit (140) trains the model, calculates the degree to which each feature contributes to the prediction performance, and then selects features with high importance to use in the final model.

[0090] After that, the abnormal situation occurrence prediction AI model learning unit (140) analyzes the correlation between features.

[0091] In one embodiment, the abnormal situation occurrence prediction AI model learning unit (140) calculates the correlation between features, removes duplicate features, and then finds pairs of features with high correlation coefficients to remove or combine one of them to create a new feature.

[0092] In another embodiment, the abnormal situation occurrence prediction AI model learning unit (140) finds pairs of features with a nonlinear relationship and removes or combines one to create a new feature.

[0093] The abnormal situation occurrence prediction AI model learning unit (140) selects features with high importance and constructs an abnormal situation occurrence prediction AI model.

[0094] In one embodiment, the abnormal situation occurrence prediction AI model learning unit (140) trains a machine learning model using features with high importance, evaluates the performance of the model through cross-validation, and optimizes it through hyperparameter tuning to construct an abnormal situation occurrence prediction AI model.

[0095] In another embodiment, a WiFi CSI-based behavior prediction model system configures a system that performs adaptive learning for a new environment through a machine learning or deep learning model based on CSI data collected via a WiFi transmitter and receiver. The system extracts environmental features from training data based on a pre-trained meta AI model. The extracted features are fine-tuned using MAML to perform predictions optimized for the environment.

[0096] Afterwards, the initial model parameters are automatically identified from the CSI data, including key environmental factors such as location information, structural obstacles, and reflector conditions for the new environment, and the initial model parameters are adjusted, and the learned initial model parameters θ are configured to be quickly relearned for feature values ​​for a small amount of environments.

[0097] Machine learning or deep learning models applied as meta AI models may include CNN (Convolutional Neural Network), RNN (Recurrent Neural Network), Transformer model, MAML (Meta learning), etc.

[0098] FIG. 2 is a flowchart illustrating one embodiment of a behavior prediction model method based on WiFi CSI according to the present invention.

[0099] Referring to FIG. 2, the transmission module of the WiFi CSI-based behavior prediction model system transmits a signal at 100 ms intervals for each antenna by applying a TDMA method that enables data transmission and reception to multiple users (step S201).

[0100] The receiving module of the WiFi CSI-based behavior prediction model system receives and stores packet data transmitted from the transmitting module, and collects environmental data including temperature gradient, etc. (step S202).

[0101] The WiFi CSI-based behavior prediction model system classifies a signal (step S203), and then, if noise is included in the WiFi CSI signal (step S204), the noise is deleted (step S205) using a noise detection AI model.

[0102] The WiFi CSI-based behavior prediction model system executes object-specific labeling (step S208) using an object segmentation and pattern classification AI model (step S207) after noise has been removed (step S206), and then stores object pattern data (step S209).

[0103] FIG. 3 is a flowchart illustrating another embodiment of a behavior prediction model method based on WiFi CSI according to the present invention. FIG. 4 is an exemplary diagram illustrating the execution process of FIG. 3.

[0104] Referring to FIG. 3, the WiFi CSI-based behavior prediction model system collects CSI data through a WiFi transmitter and receiver in various environments and situations, and then labels each data sample with a label that can identify the corresponding object (step S301, step S302).

[0105] In one embodiment, the WiFi CSI-based action prediction model system collects data including moving people, stationary objects, furniture, etc. at various locations, and labels the data of moving people as 'people', the data of stationary objects as 'objects', etc.

[0106] Afterwards, the WiFi CSI-based behavior prediction model system removes noise from the CSI data, normalizes the signal intensity and phase values ​​to a certain range, thereby ensuring a consistent range of model input values, and extracts useful features from the time and frequency domains of the signal (step S303). For example, the features may include signal amplitude, phase shift, FFT transform results, etc. At this time, the WiFi CSI-based behavior prediction model system stores feature vectors and labels in pairs.

[0107] The WiFi CSI-based behavior prediction model system selects a machine learning or deep learning model suitable for object segmentation and pattern classification (step S304). Examples include a Convolutional Neural Network (CNN), a Recurrent Neural Network (RNN), a Transformer model, and Meta Learning (MAML), but any model capable of achieving the same goal is also acceptable.

[0108] Next, the WiFi CSI-based behavior prediction model system configures the selected model (step S305). For example, a CNN consists of an input layer, multiple convolutional layers, a pooling layer, and a fully connected layer. An RNN uses a Long Short-Term Memory (LSTM) or a Gated Recurrent Unit (GRU). A Transformer uses an attention mechanism. For MAML, the initial weights of the model are structured to meta-learn, enabling rapid adaptation even with a small amount of training data.

[0109] The WiFi CSI-based behavior prediction model system divides the collected data into training data and test data, inputs the data's feature vectors and labels into the object classification and pattern classification AI model, and calculates the difference between the label predicted by the object classification and pattern classification AI model and the actual label using a loss function (e.g., cross-entropy loss function).

[0110] After that, the WiFi CSI-based action prediction model system uses the backpropagation algorithm to update the weights of the object classification and pattern classification AI model to minimize the loss function, and adjusts hyperparameters such as learning rate, batch size, and number of epochs to optimize the performance of the object classification and pattern classification AI model.

[0111] Meanwhile, in another embodiment, a behavior prediction model system based on WiFi CSI collects CSI data through a WiFi transmitter and receiver in various environments and situations (step S401), removes noise from the collected CSI, normalizes amplitude and phase values ​​to a certain range, and adjusts the range of model input values ​​to a certain range (step S402), and extracts useful features from the normalized CSI data (step S403).

[0112] Then, a machine learning or lip learning model is selected and the extracted features are input to estimate the location of the object (step S404), and the estimated location of the object is transmitted to an external system including a manager terminal, platform, or API (step S405), and an IoT control command is transmitted using the MQTT (Message Queuing Telemetry Transport) or Matter protocol based on the estimated location of the transmitted object (step S406).

[0113] FIG. 5 is a flowchart illustrating another embodiment of a behavior prediction model method based on WiFi CSI according to the present invention.

[0114] Referring to FIG. 5, the WiFi CSI-based behavior prediction model system collects time domain data and spatial domain data of an object by utilizing various sensors (step S501).

[0115] The WiFi CSI-based behavior prediction model system analyzes the behavior pattern type for each object and then uses an abnormal situation occurrence prediction AI model to check whether there are signs of an abnormal situation (step S502).

[0116] The WiFi CSI-based behavior prediction model system performs classification of the abnormal situation (step S503) when there is an abnormal situation sign (step S502).

[0117] If the abnormality indication is an emergency situation (step S504), the WiFi CSI-based behavior prediction model system provides the information to the administrator terminal (step S505). Meanwhile, if the abnormality indication is not an emergency situation (step S504), the WiFi CSI-based behavior prediction model system provides the information to the administrator terminal (step S506).

[0118] FIG. 6 is a flowchart illustrating another embodiment of a behavior prediction model method based on WiFi CSI according to the present invention.

[0119] Referring to FIG. 6, the WiFi CSI-based behavior prediction model system collects time-domain and spatial-domain data of objects using various sensors (step S601). The WiFi CSI-based behavior prediction model system performs object-specific labeling using an object segmentation and pattern classification AI model, and then stores object pattern data (step S602).

[0120] The WiFi CSI-based behavior prediction model system uses an AI model for predicting abnormal situations to predict whether there are any abnormal situation signs (step S603). If the WiFi CSI-based behavior prediction model predicts abnormal situation signs using the AI ​​model for predicting abnormal situations (step S604), the system provides the predicted abnormal situation signs to the administrator terminal (step S605).

[0121] Meanwhile, in another embodiment, in a behavior prediction model system based on WiFi CSI, an environment adaptive learning system based on MAML sets pre-learned model initial values ​​(step S601) and configures CSI data collected in a new environment as a small amount of learning data (step S602).

[0122] Afterwards, the loss for the environment sample is calculated through the inner loop and the initial model parameters are fine-tuned (step S603), and the initial model parameters are updated based on the meta-loss calculated through the outer loop (step S604).

[0123] Meanwhile, the description of the technology disclosed in this specification is merely an example for structural and functional explanation, and therefore, the scope of the rights of the disclosed technology should not be construed as being limited by the embodiments described in the text. That is, since the embodiments can be modified in various ways and can take various forms, the scope of the rights of the disclosed technology should be understood to include equivalents that can realize the technical idea. Furthermore, the purpose or effect presented in the disclosed technology does not mean that a specific embodiment must include all of them or only such effects, and therefore, the scope of the rights of the disclosed technology should not be construed as being limited thereby.

[0124] Furthermore, when a component is said to be "connected" to another component, it should be understood that while it may be directly connected to that other component, there may also be other components intervening. Conversely, when a component is said to be "directly connected" to another component, it should be understood that there are no other intervening components. Similarly, other expressions describing relationships between components, such as "between" and "between," or "adjacent to" and "directly adjacent to," should be interpreted similarly.

[0125] Singular expressions should be understood to include plural expressions unless the context clearly indicates otherwise, and terms such as "comprises" or "has" should be understood to specify the presence of a stated feature, number, step, operation, component, part, or combination thereof, but not to exclude the possibility of the presence or addition of one or more other features, numbers, steps, operations, components, parts, or combinations thereof.

[0126]

[0127] The present invention is not limited to the specific preferred embodiments described above, and anyone with ordinary skill in the art to which the invention pertains can make various modifications without departing from the gist of the present invention claimed in the claims, and as long as it relates to technical ideas forming such modifications, it is within the scope of the claims.

Claims

1. A noise detection AI model learning unit that collects normal signal data and noise data for WiFi CSI data and then uses them to create a noise detection AI model; An object classification and pattern classification AI model learning unit that collects CSI data through WiFi transmitters and receivers in various environments and situations, labels objects based on the movement and stationary status of each object detected from the collected data, and creates an object classification and pattern classification AI model; An abnormal situation occurrence prediction AI model learning unit that collects time domain data including sensor signals, log data, and traffic data for objects using multiple sensors, and spatial domain data including image and video 3D coordinate data; A WiFi CSI signal collection and preprocessing unit that removes noise contained in a WiFi CSI signal using a noise detection AI model, performs object-specific labeling using an object segmentation and pattern classification AI model, and then stores object pattern data; It includes an abnormal situation occurrence prediction unit that collects and analyzes spatial data, analyzes the type of behavioral pattern for each object, and then uses an abnormal situation occurrence prediction AI model to check whether there are signs of an abnormal situation and provides the information to the administrator terminal. The above noise detection AI model learning unit After extracting features from the collected normal signal data and noise data, label the normal signal label or noise label for each extracted data, select a machine learning or deep learning model suitable for noise detection to build a noise detection AI model, separate the data into training data and validation data, train the model to distinguish between normal signals and noise using the training data, and evaluate the performance of the model using the validation data to input the feature vector of the training data sample and the corresponding label into the noise detection AI model. The AI ​​model learning department that predicts the occurrence of the above abnormal situation A WiFi CSI-based behavior prediction model system characterized by including an AI model for predicting situational signs by collecting time domain data and space domain data of an object, labeling the data with object behavior patterns and situational signs, extracting features for analyzing the object's behavior patterns, evaluating the importance of the extracted features, and selecting features based on the evaluated importance.

2. In paragraph 1, The above noise detection AI model learning unit The difference between the label predicted by the noise detection AI model and the actual label is calculated using a loss function, and the weights of the noise detection AI model are updated using the backpropagation algorithm to minimize the loss function. A WiFi CSI-based behavior prediction model system characterized by evaluating the performance of a noise detection AI model using validation data during training.

3. In paragraph 1, The above object classification and pattern classification AI model learning unit A WiFi CSI-based behavior prediction model system characterized in that after removing noise from CSI data, signal intensity and phase values ​​are normalized to a certain range to make the range of model input values ​​constant, features including signal amplitude, phase change, and FFT change results in the time and frequency domains of the signal are extracted, feature vectors and labels are stored in pairs, and a machine learning or deep learning model suitable for object classification and pattern classification is selected and trained to create an object classification and pattern classification AI model.

4. In a method for executing a WiFi CSI-based behavior prediction model system according to any one of the first to third clauses, the method for executing the behavior prediction model system comprises: A step of creating a interference signal detection AI model using normal signal data and interference signal data; A step of creating an object classification and pattern classification AI model using CSI data collected through a WiFi transmitter and receiver in various environments and situations; A step of creating an AI model for predicting the occurrence of an abnormal situation using time domain features, frequency domain features, and spatial domain features; A step of using the above interference signal detection AI model to delete an interference signal if the WiFi CIS signal contains an interference signal, and if the interference signal is deleted, executing labeling for each object using the object classification and pattern classification AI model and then storing object pattern data; A step of collecting and analyzing spatial data, and then using the AI ​​model predicting the occurrence of abnormal situations that analyzes the type of behavioral pattern for each object, to check whether there are signs of abnormal situations and providing the result to the administrator terminal; The steps for creating an AI model to predict the occurrence of the above abnormal situation are as follows: A method for executing a behavior prediction model system based on WiFi CSI, comprising the steps of collecting time domain data and space domain data of an object, labeling the data with object behavior patterns and situation occurrence signs, extracting meaningful features to analyze the object behavior patterns, evaluating the importance of the features, and then selecting features according to the importance to construct an AI model for predicting situation occurrence signs.

Images