Environment perception detection method based on wi-fi channel state information

Abstract

This invention discloses an environmental perception and detection method based on Wi-Fi channel state information, belonging to the field of mobile target detection technology. This method is based on a contactless perception scheme using Wi-Fi signals, completely avoiding image acquisition and fundamentally eliminating the risk of privacy leakage. It is also unaffected by environmental interference such as light, obstruction, dust, and severe weather, and is safe for use in sensitive scenarios such as warehouses, industrial workshops, and confidential parks. At the same time, the solution directly reuses existing commercial Wi-Fi equipment without the need to deploy dedicated hardware, significantly reducing system construction and maintenance costs and possessing strong promotion and popularization value. Furthermore, through an original four-level preprocessing process, it effectively suppresses Gaussian noise, phase shift, and static multipath interference. Combined with multi-scale time-frequency domain-spatial joint feature extraction, it can accurately capture changes in channel state information caused by mobile targets of different speeds and types.

Description

Technical Field

[0001] This invention relates to the field of mobile target detection technology, specifically to an environmental perception and detection method based on Wi-Fi channel state information. Background Technology

[0002] In modern society, with the rapid development of the Internet of Things, industrial automation and intelligent security technologies, indoor and outdoor moving object detection has become a core technology requirement for scenarios such as warehousing and logistics, industrial production and park security. Existing detection technologies are mainly divided into two categories: contact and non-contact. Contact solutions can only achieve fixed-point passing detection and cannot obtain complete information such as motion status and speed. They also have poor deployment flexibility, are susceptible to environmental corrosion, and have high maintenance costs. Traditional visual non-contact solutions have comprehensive recognition capabilities, but are greatly affected by lighting, occlusion, dust, and severe weather. Their detection accuracy drops significantly in low light and complex environments, and there is a risk of privacy leakage. They cannot be applied to sensitive scenarios such as confidential and closed warehouses. Millimeter-wave radar and lidar have high accuracy, but the equipment cost is high. They are prone to misjudgment and missed detection in indoor environments with severe multipath effects and industrial workshops with dense metal equipment. Wi-Fi signal-based sensing solutions have become popular due to their advantages such as widespread deployment of commercial equipment, no visual acquisition required, and low cost. However, solutions based on received signal strength indication suffer from coarse signal granularity and poor stability, only achieving rough presence detection. CSI-based sensing solutions, on the other hand, can acquire amplitude and phase information of each subcarrier, possessing the potential for high-precision sensing. However, existing CSI sensing technologies still have significant shortcomings: raw data is susceptible to noise and frequency offset interference; current preprocessing cannot effectively eliminate phase shifts and static multipath, leading to decreased detection accuracy in complex environments; complex deep learning models have high computational overhead, making it difficult to achieve real-time inference on commercial equipment; fixed thresholds and models cannot adapt to dynamic environmental changes, resulting in poor generalization ability; furthermore, high accuracy and low latency are difficult to achieve simultaneously, failing to meet the detection needs of diverse moving objects such as low-speed AGVs and objects with minute displacements. Therefore, developing a Wi-Fi channel state information sensing and detection method that combines high precision, strong anti-interference, real-time performance, and environmental adaptability is of great practical significance and application value. Summary of the Invention

[0003] The purpose of this invention is to provide an environmental perception and detection method based on Wi-Fi channel state information to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an environment perception and detection method based on Wi-Fi channel state information, comprising the following steps: S1. Collect raw CSI data corresponding to each OFDM subcarrier in the target environment using commercial Wi-Fi devices; S2. Preprocess the acquired raw CSI data to eliminate noise interference and phase shift in the data, and obtain standardized effective CSI data. S3. The standardized CSI effective data are grouped by subcarrier, and multi-scale time-frequency domain feature extraction is performed to obtain the CSI feature sequence characterizing the motion characteristics of the moving object. S4. Input the CSI feature sequence into the pre-trained lightweight moving object perception and recognition model, and combine it with adaptive environment threshold judgment to output the moving object detection results in the target environment in real time. S5. Simultaneously display the CSI time-frequency domain feature map and moving object status corresponding to the detection results, and save the feature data and detection log.

[0005] Preferably, the commercial Wi-Fi device collects raw Channel State Information (CSI) data within the target environment. Specifically, this includes using a commercial Wi-Fi transceiver that supports the 802.11a / n / ac / ax protocol. The transmitting end sends standard OFDM data frames, and the receiving end collects raw CSI data for each OFDM subcarrier in the target environment through a Wi-Fi network card. The raw CSI data includes amplitude and phase information for each subcarrier. The collection frequency band covers the 2.4GHz and 5GHz Wi-Fi bands, and the number of subcarriers collected is no less than 30.

[0006] Preferably, step S2 specifically includes the following steps: S201. For the CSI amplitude timing sequence of each subcarrier, use... The criteria include outlier handling and calculation of the statistical mean of the sequence. with standard deviation The amplitude exceeds Data points in the interval are identified as outliers and removed. For missing data points resulting from the removal, linear interpolation between two adjacent valid data points is used to fill in the missing data points. S202. For the amplitude sequence after outlier removal, perform 4-level wavelet decomposition using the db4 wavelet basis to obtain one set of low-frequency approximation coefficients and four sets of high-frequency detail coefficients. Perform threshold quantization on the high-frequency detail coefficients using a soft thresholding function, setting the threshold... ,in , The noise standard deviation representing the high-frequency coefficient. Indicates the length of the sequence to be processed. Represents the natural logarithm; S203. After quantization, the signal is reconstructed by inverse wavelet transform. The reconstructed signal is then smoothed by a moving average filter with a window length of 7 to further suppress random burst noise.

[0007] Preferably, step S2 further includes the following steps: S204, CSI phase sequence The linear slope was obtained by fitting using the least squares method. With fixed offset and determine the center frequency of the subcarrier. ,use and Calculate the calibrated phase information ,in ; S205. Phase calibration is performed sequence by sequence for each subcarrier phase sequence of the receiving antenna; S206. Using the min-max normalization method, the amplitude sequence and calibrated phase information of each subcarrier are normalized and mapped to the [0,1] interval. The specific normalization formula is as follows: , in, Represents the original data value. This represents the minimum value in the current sequence. This represents the maximum value in the current sequence. This represents standardized data after normalization. S207. Through normalization processing, the amplitude differences between different subcarriers and the hardware gain differences between antennas are eliminated to obtain standardized CSI effective data.

[0008] Preferably, step S3 specifically includes the following steps: S301. Divide the standardized CSI effective data into several continuous subcarrier groups according to the subcarrier frequency correlation, and extract time-domain features, frequency-domain features and spatial features for each subcarrier group respectively. S302. Concatenate all features according to the time series to form a fixed-dimensional CSI feature sequence with a length of 256 dimensions. Use a sliding time window to extract continuous features. Set the sliding window length to 100ms and the window sliding step size to 20ms.

[0009] Preferably, step S4 specifically includes the following steps: S401. A lightweight moving object perception and recognition model receives a CSI feature sequence with a dimension of 256 as input. This input sequence is then fed into a first depthwise separable convolutional layer to decouple spatial feature extraction from channel feature fusion. The feature map output by the convolution is downsampled by a first max pooling layer to obtain pre-processed features. S402. The pre-processed features are sequentially passed through the second deep separable convolutional layer and the second max pooling layer. The multi-dimensional feature map output by the pooling layer is fed into the Flatten layer. All features are flattened into a one-dimensional vector and input into the GRU layer. After the GRU layer, the Dropout layer is connected. The regularized features are fed into the fully connected layer for linear transformation and dimension mapping, and finally output through the Softmax classification layer to output the probability distribution of each category. S403. The model finally outputs the probability distribution of 6 types of detection results, namely no moving object, object stationary, low-speed movement, medium-speed movement, high-speed movement, and area intrusion. The category corresponding to the maximum probability is taken as the final detection result. Low-speed movement is defined as movement speed less than or equal to 0.5 m / s, medium-speed movement is defined as movement speed greater than 0.5 m / s and less than or equal to 2 m / s, and high-speed movement is defined as movement speed greater than 2 m / s.

[0010] Preferably, step S4 further includes the following steps: S404. Collect 30 minutes of environmental CSI data with no moving objects and calculate the baseline mean of the corresponding feature sequences. and baseline standard deviation and the initial sensitivity coefficient Set to 2.5, utilize and Calculate the threshold ,in ; S405. Real-time calculation of the current CSI feature sequence With environmental baseline Feature Difference ,in , The current real-time 256-dimensional CSI feature sequence is the first... 3D eigenvalues The first environmental baseline characteristic sequence 3D eigenvalues The fixed total dimension representing the feature sequence; S406, when When a moving object is detected in the environment, the subsequent fine-grained model recognition process is immediately triggered. When the object is determined to be stationary, model inference is not triggered. S407. Set up an automatic environmental baseline update mechanism, updating the environmental baseline automatically every 12 hours. The update period is set to the non-operational period of the warehouse. Environmental data without moving objects during this period is collected to update the baseline mean and standard deviation. The sensitivity coefficient is adjusted according to scenario requirements. .

[0011] Preferably, step S5 specifically includes the following steps: S501: The local touch interface of the industrial gateway displays the current moving object detection results, CSI amplitude time-domain waveform and frequency-domain spectrum in real time. It also displays the timeline of moving events in the last 24 hours, supports real-time viewing and historical data backtracking within 30 days, and can retrieve the corresponding CSI raw data, feature data and detection results by time node. S502. The local device automatically saves feature data, detection results and corresponding timestamps in a preset format. The main storage format is CSV, and a backup JSON format is also provided. The data fields include timestamp, detection result category, recognition confidence, feature sequence mean, corresponding subcarrier frequency band information and moving object motion parameters. S503: All stored data is encrypted using AES-256 symmetric encryption. The encryption key is set locally by the administrator and can only be decrypted and exported offline after authorization. S504. In response to an abnormal intrusion event in the area, after the detection result is triggered, the local device immediately saves the full CSI time-frequency domain data for 30 seconds before and after the event, and pushes the alarm information to the warehouse management system via industrial Ethernet.

[0012] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention is a contactless sensing solution based on Wi-Fi signals that completely avoids image acquisition, fundamentally eliminating the risk of privacy leakage. It is also unaffected by environmental interference such as light, obstruction, dust, and severe weather, and is safe for use in sensitive scenarios such as warehouses, industrial workshops, and confidential parks. At the same time, the solution directly reuses existing commercial Wi-Fi devices without the need to deploy dedicated hardware, which significantly reduces the system construction and maintenance costs and has great potential for widespread adoption. 2. This invention effectively suppresses Gaussian noise, phase shift and static multipath interference through an original four-level preprocessing process. Combined with multi-scale time-frequency domain-space joint feature extraction, it can accurately capture channel state information changes caused by moving targets of different speeds and types. 3. This invention can automatically adapt to static layout changes through dynamic environmental baseline updates and adaptive threshold judgment mechanisms, avoiding repeated retraining of the model for different environments. Combined with federated learning, it achieves multi-scenario data aggregation, greatly improving the model's generalization ability to different environments and different types of moving objects, and ensuring long-term stable operation in multiple scenarios such as warehousing and logistics, industrial production, and park security. Attached Figure Description

[0013] Figure 1 Detailed flowcharts are provided for embodiments of the present invention; Figure 2 A flowchart of the CSI preprocessing and feature extraction algorithm is provided for embodiments of the present invention; Figure 3 A flowchart of the model reasoning and environment adaptation mechanism is provided for embodiments of the present invention; Figure 4 An overall method flowchart is provided for embodiments of the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Example 1: Please refer to Figure 4 This invention provides a technical solution: an environmental perception and detection method based on Wi-Fi channel state information, comprising the following steps: S1. Collect raw CSI data corresponding to each OFDM subcarrier in the target environment using commercial Wi-Fi devices; S2. Preprocess the acquired raw CSI data to eliminate noise interference and phase shift in the data, and obtain standardized effective CSI data. S3. The standardized CSI effective data are grouped by subcarrier, and multi-scale time-frequency domain feature extraction is performed to obtain the CSI feature sequence characterizing the motion characteristics of the moving object. S4. Input the CSI feature sequence into the pre-trained lightweight moving object perception and recognition model, and combine it with adaptive environment threshold judgment to output the moving object detection results in the target environment in real time. S5. Simultaneously display the CSI time-frequency domain feature map and moving object status corresponding to the detection results, and save the feature data and detection log.

[0016] Commercial Wi-Fi devices collect raw Channel State Information (CSI) data within the target environment. Specifically, this involves using commercial Wi-Fi transceivers that support the 802.11a / n / ac / ax protocols. The transmitting end sends standard OFDM data frames, and the receiving end collects raw CSI data for each OFDM subcarrier in the target environment via a Wi-Fi network card. The raw CSI data includes amplitude and phase information for each subcarrier. The collection frequency bands cover the 2.4GHz and 5GHz Wi-Fi bands, and the number of subcarriers collected is no less than 30. S2 specifically includes the following steps: S201. For the CSI amplitude timing sequence of each subcarrier, use... The criteria include outlier handling and calculation of the statistical mean of the sequence. with standard deviation The amplitude exceeds Data points in an interval are identified as outliers and removed. For missing data points resulting from the removal, linear interpolation between two adjacent valid data points is used to fill in the missing data points, ensuring the continuity of the time series. S202. For the amplitude sequence after outlier removal, perform 4-level wavelet decomposition using the db4 wavelet basis to obtain one set of low-frequency approximation coefficients and four sets of high-frequency detail coefficients. Perform threshold quantization on the high-frequency detail coefficients using a soft thresholding function, setting the threshold... ,in , The noise standard deviation representing the high-frequency coefficient. Indicates the length of the sequence to be processed. Represents the natural logarithm; S203. After quantization, the signal is reconstructed by inverse wavelet transform, which effectively eliminates Gaussian white noise in the industrial environment. For the reconstructed signal, a moving average filter with a window length of 7 is used for smoothing to further suppress random burst noise, while retaining the effective signal change characteristics brought about by the motion of moving objects. S2 also includes the following steps: S204, CSI phase sequence The linear slope was obtained by fitting using the least squares method. With fixed offset and determine the center frequency of the subcarrier. ,use and Calculate the calibrated phase information ,in ; S205. Phase calibration is performed sequence by sequence for each subcarrier phase sequence of the receiving antenna; S206. Using the min-max normalization method, the amplitude sequence and calibrated phase information of each subcarrier are normalized and mapped to the [0,1] interval. The specific normalization formula is as follows: , in, Represents the original data value. This represents the minimum value in the current sequence. This represents the maximum value in the current sequence. This represents standardized data after normalization. S207. Through normalization processing, the amplitude differences between different subcarriers and the hardware gain differences between antennas are eliminated to obtain standardized CSI effective data.

[0017] S3 specifically includes the following steps: S301. The standardized CSI effective data is divided into several continuous subcarrier groups according to the subcarrier frequency correlation. Time domain features, frequency domain features and spatial features are extracted for each subcarrier group to achieve multi-scale and multi-dimensional feature capture. S302. All features are concatenated according to the time series to form a fixed-dimensional CSI feature sequence with a length of 256 dimensions. A sliding time window is used for continuous feature extraction. The sliding window length is set to 100ms and the window sliding step size is 20ms to ensure the continuity and real-time performance of the detection and achieve millisecond-level feature updates. S4 specifically includes the following steps: S401. A lightweight moving object perception and recognition model receives a CSI feature sequence with a dimension of 256 as input. This input sequence is then fed into a first depthwise separable convolutional layer to decouple spatial feature extraction from channel feature fusion. The feature map output by the convolution is downsampled by a first max pooling layer to obtain pre-processed features. S402. The pre-processed features are sequentially passed through the second deep separable convolutional layer and the second max pooling layer. The multi-dimensional feature map output by the pooling layer is fed into the Flatten layer. All features are flattened into a one-dimensional vector and input into the GRU layer. After the GRU layer, the Dropout layer is connected. The regularized features are fed into the fully connected layer for linear transformation and dimension mapping, and finally output through the Softmax classification layer to output the probability distribution of each category. S403. The model finally outputs the probability distribution of 6 types of detection results, namely no moving object, object stationary, low-speed movement, medium-speed movement, high-speed movement and area intrusion. The category corresponding to the maximum probability is taken as the final detection result. Low-speed movement is the movement speed less than or equal to 0.5m / s, medium-speed movement is the movement speed greater than 0.5m / s and less than or equal to 2m / s, and high-speed movement is the movement speed greater than 2m / s. S4 also includes the following steps: S404. Collect 30 minutes of environmental CSI data with no moving objects and calculate the baseline mean of the corresponding feature sequences. and baseline standard deviation and the initial sensitivity coefficient Set to 2.5, utilize and Calculate the threshold ,in ; S405. Real-time calculation of the current CSI feature sequence With environmental baseline Feature Difference ,in , The current real-time 256-dimensional CSI feature sequence is the first... 3D eigenvalues The first environmental baseline characteristic sequence 3D eigenvalues The fixed total dimension representing the feature sequence; S406, when When a moving object is detected in the environment, the subsequent fine-grained model recognition process is immediately triggered. When the object is determined to be stationary, model inference is not triggered. S407. Set up an automatic environmental baseline update mechanism, updating the environmental baseline automatically every 12 hours. The update period is set to the non-operational period of the warehouse. Environmental data without moving objects during this period is collected to update the baseline mean and standard deviation. The sensitivity coefficient is adjusted according to scenario requirements. The value ranges from 1.5 to 3.0. The smaller the value, the higher the sensitivity for detecting moving objects with low speed and small displacement, making it suitable for monitoring the displacement of stored goods. The larger the value, the stronger the anti-interference ability, which is suitable for high dynamic scenarios where forklifts operate frequently; S5 specifically includes the following steps: S501: The local touch interface of the industrial gateway displays the current moving object detection results, CSI amplitude time-domain waveform and frequency-domain spectrum in real time. It also displays the timeline of moving events in the last 24 hours, supports real-time viewing and historical data backtracking within 30 days, and can retrieve the corresponding CSI raw data, feature data and detection results by time node. S502. The local device automatically saves feature data, detection results and corresponding timestamps in a preset format. The main storage format is CSV, and a backup JSON format is also provided. The data fields include timestamp, detection result category, recognition confidence, feature sequence mean, corresponding subcarrier frequency band information and moving object motion parameters. S503: All stored data is encrypted using AES-256 symmetric encryption. The encryption key is set locally by the administrator and can only be decrypted and exported offline after authorization. S504. In response to an abnormal intrusion event in the area, after the detection result is triggered, the local device immediately saves the full CSI time-frequency domain data for 30 seconds before and after the event, and pushes the alarm information to the warehouse management system via industrial Ethernet.

[0018] Example 2: Please refer to Figures 1-3 This invention targets core scenarios in industrial warehousing and logistics, enabling functions such as moving object presence detection, static / moving state recognition, motion speed level classification, intrusion behavior determination, and movement trajectory tracking. The method includes the following steps: S1. Collect raw Channel State Information (CSI) data within the target environment using commercial Wi-Fi devices; S2. Preprocess the acquired raw CSI data to eliminate noise interference and phase shift in the data, and obtain standardized effective CSI data. S3. The standardized CSI effective data are grouped by subcarrier, and multi-scale time-frequency domain feature extraction is performed to obtain the CSI feature sequence characterizing the motion characteristics of the moving object. S4. Input the CSI feature sequence into the pre-trained lightweight moving object perception and recognition model, and combine it with adaptive environment threshold judgment to output the moving object detection results in the target environment in real time. S5. Simultaneously display the CSI time-frequency domain feature map and moving object status corresponding to the detection results, and save the feature data and detection log.

[0019] S1 collects raw Channel State Information (CSI) data within the target environment using commercial Wi-Fi devices. The specific implementation is as follows: This embodiment uses four industrial gateways equipped with Intel Wi-Fi 6 AX210 commercial wireless network cards as Wi-Fi transceiver devices. Both transceivers natively support the IEEE 802.11ax protocol. Each device is configured with two 5GHz omnidirectional Wi-Fi antennas with a gain of 5dBi and a horizontal beamwidth of 360°, meeting the omnidirectional coverage requirements of the warehouse space. The transmitter operates at 5GHz. The Wi-Fi band, with a center frequency of 5.2GHz and a channel bandwidth of 40MHz, corresponds to 114 effective OFDM subcarriers (excluding DC subcarriers and guard interval subcarriers). The transmitter continuously sends standard OFDM data frames at a packet rate of 2000Hz using the open-source CSI-Tool tool under the Linux system. The frame length is 1500 bytes. The receiver and transmitter maintain microsecond-level time synchronization and synchronously collect raw CSI data corresponding to each OFDM subcarrier in the target warehouse environment. The raw data includes amplitude and phase information of each subcarrier. At the same time, spatial diversity CSI data from two receiving antennas are acquired. The collected raw data is temporarily stored in the local memory of the receiver device in real time. The buffering time for a single acquisition is no less than 60 seconds, which meets the requirements of continuous detection and historical backtracking. The target monitoring environment in this embodiment is a standard industrial warehouse space of 800㎡, which includes warehouse equipment such as high-bay racks, pallets, AGV carts, and forklifts. Four Wi-Fi transceivers are deployed at the four opposite corners of the warehouse space, with a deployment height of 2.5m. There are no strong metal obstructions, which can completely cover the entire area of ​​the warehouse where goods are handled, AGVs are driven, and personnel and vehicles are intruded. It can effectively capture changes in the Wi-Fi signal propagation path caused by the movement of moving objects.

[0020] S2 preprocesses the acquired raw CSI data, specifically by performing a four-stage processing flow: outlier removal, noise suppression, phase calibration, and data normalization, as detailed below: The first step is outlier removal: For the CSI amplitude time series of each subcarrier, outlier processing is performed using the 3σ criterion. For the amplitude sequence of a single subcarrier, its statistical mean μ and standard deviation σ are calculated. Data points with amplitudes exceeding the range [μ-3σ, μ+3σ] are identified as outliers and removed. For missing data points resulting from removal, linear interpolation between two adjacent valid data points is used to fill in the missing data, ensuring the continuity of the time series. In this embodiment, outlier removal is performed on each of the 114 subcarriers of a single antenna, eliminating abnormal data caused by electromagnetic interference in the industrial environment and sampling jitter in the equipment hardware.

[0021] The second step is noise suppression: a two-stage noise reduction scheme combining wavelet denoising and moving average filtering is adopted. First, the amplitude sequence after outlier removal is decomposed into four levels using the db4 wavelet basis, yielding one set of low-frequency approximation coefficients and four sets of high-frequency detail coefficients. The high-frequency detail coefficients are then thresholded using a soft threshold function, with the threshold set to [value missing]. ,in Let be the noise standard deviation of the high-frequency coefficients, N be the length of the sequence to be processed, and log be the natural logarithm (base e, also written as ln). After quantization, the signal is reconstructed using inverse wavelet transform, effectively eliminating Gaussian white noise in the industrial environment. The reconstructed signal is then smoothed using a moving average filter with a window length of 7 to further suppress random burst noise while preserving the effective signal variation characteristics caused by the motion of moving objects.

[0022] The third step, phase calibration: A linear phase correction algorithm is used to eliminate random phase shifts and phase entanglement caused by carrier frequency offset and sampling frequency offset. This is applied to the original phase sequence. The correction formula is: , In the formula, f is the center frequency of the subcarrier, a is the linear slope of the phase offset, and b is the fixed phase offset. The linear slope a and the fixed offset b are obtained by least squares fitting, eliminating the linear error of the phase and obtaining continuous, untangled calibrated phase information. In this embodiment, phase calibration is performed sequence by sequence for each subcarrier phase sequence of the two receiving antennas to ensure the validity of the phase information.

[0023] The fourth step is data normalization: Using the min-max normalization method, the amplitude data and calibrated phase data of each subcarrier are normalized and mapped to the [0,1] interval. The normalization formula is as follows: , In the formula, x is the original data value. This is the minimum value of the sequence. The maximum value of the sequence. This is the normalized data. Through normalization processing, the amplitude differences between different subcarriers and the hardware gain differences between different antennas are eliminated, ultimately yielding standardized CSI valid data.

[0024] The S3 multi-scale time-frequency domain feature extraction is implemented as follows: In this embodiment, the 114 standardized effective subcarriers are divided into 19 consecutive subcarrier groups according to the frequency correlation of adjacent subcarriers. Each group contains 6 consecutive subcarriers. Time-domain features, frequency-domain features, and spatial features are extracted for each subcarrier group to achieve multi-scale and multi-dimensional feature capture. The specific feature extraction dimensions are as follows: 1. Time-domain characteristics: For the amplitude time-series sequence of each subcarrier group, five core time-domain characteristics are calculated, namely: sequence mean, sequence variance, peak factor (the ratio of the sequence peak value to the effective value), zero-crossing rate, and short-time energy within a 50ms sliding time window, which are used to characterize the time-domain fluctuation characteristics of the CSI signal caused by the motion of moving objects.

[0025] 2. Frequency Domain Features: For the time-domain sequence of each subcarrier group, a 1024-point Fast Fourier Transform (FFT) is performed. After conversion to the frequency domain, four core frequency domain features are extracted: spectral centroid, spectral entropy, frequency band energy ratio corresponding to the effective Doppler frequency shift of the moving object in the 0~100Hz range, and Doppler frequency shift peak value. These features are used to distinguish the frequency domain characteristics of different motion speeds and motion types. The Doppler frequency shift corresponding to the low-speed movement of the AGV is concentrated in the 1~20Hz range, while the Doppler frequency shift corresponding to the high-speed movement of the forklift is concentrated in the 10~80Hz range. These features can be accurately captured by the frequency band energy ratio.

[0026] 3. Spatial Features: For the same-frequency subcarriers of two receiving antennas, the phase difference between the two antennas is calculated, and two types of spatial features, the mean and variance of the phase difference, are extracted. These features are used to characterize the changes in the spatial propagation path of Wi-Fi signals caused by the movement of moving objects, improving the detection capability for objects with small displacements and low speeds. In this embodiment, all extracted features are concatenated in time sequence to form a fixed-dimensional CSI feature sequence of length 256. A sliding time window is used for continuous feature extraction, with a sliding window length of 100ms and a window sliding step of 20ms, ensuring the continuity and real-time performance of detection and achieving millisecond-level feature updates.

[0027] The lightweight moving object perception and recognition model in S4, which performs inference and adaptive environment threshold judgment, is implemented as follows: In this embodiment, the lightweight moving object perception and recognition model is an improved CNN-GRU fusion network. The input is a 256-dimensional CSI feature sequence. The model network structure is as follows: Input layer → First depthwise separable convolutional layer (3×3 kernel, stride 1, 32 output channels, Padding=Same) → First max pooling layer (2×2 pooling kernel, stride 2) → Second depthwise separable convolutional layer (3×3 kernel, stride 1, 64 output channels, Padding=Same) → Second max pooling layer (2×2 pooling kernel, stride 2) → Flatten layer → GRU layer (128 hidden layer dimensions, dropout=0.1) → Dropout layer (0.2% inactivation rate) → Fully connected layer (6 output dimensions) → Softmax classification layer.

[0028] The model ultimately outputs probability distributions for six detection categories: no moving object, stationary object, low-speed motion (≤0.5m / s), medium-speed motion (0.5~2m / s), high-speed motion (>2m / s), and area intrusion. The category corresponding to the highest probability is taken as the final detection result. During model training, a federated learning framework is used, uniting eight terminal devices from different warehousing scenarios. During training, each terminal device completes data training and gradient calculation locally, only uploading the model gradient parameters to the central server for model aggregation, without uploading any original scene data, ensuring data security for each warehousing scenario. After model training, it is converted to the TensorFlow Lite lightweight format and deployed to a local industrial gateway. Inference is completed entirely locally. Actual testing shows that the inference latency for a single frame of 256-dimensional feature sequences is ≤30ms, meeting the real-time detection requirements of industrial scenarios.

[0029] In this embodiment, the adaptive environmental threshold judgment is implemented as follows: after the device is deployed, environmental CSI data with no moving objects is collected for 30 minutes during the first startup, and the baseline mean of the corresponding feature sequence is calculated. and baseline standard deviation The initial sensitivity coefficient α is set to 2.5, and the threshold calculation formula is as follows: , Real-time calculation of the feature difference between the current CSI feature sequence and the environmental baseline. The difference is quantified using Euclidean distance, and the calculation formula is as follows: , In the formula, For feature difference degree, The current real-time 256-dimensional CSI feature sequence is the first... 3D eigenvalues The first environmental baseline characteristic sequence The feature values ​​are 256, where D=256 is the fixed total dimension of the feature sequence; when When a moving object is detected in the environment, the subsequent fine-grained model recognition process is immediately triggered. When there is no moving object, the model inference is not triggered, thus reducing the device's computational overhead and power consumption.

[0030] To address the long-term static changes in the warehousing environment, this embodiment establishes an automatic environmental baseline update mechanism. The environmental baseline is automatically updated every 12 hours during non-operational periods. Environmental data without moving objects during this period is collected to update the baseline mean and standard deviation. This adapts to static environmental changes such as warehouse shelf position adjustments, goods stacking changes, environmental temperature and humidity variations, and Wi-Fi channel interference, preventing an increase in false alarm rates over long-term operation and improving the method's environmental adaptability and long-term stability. Users can manually adjust the sensitivity coefficient α according to scenario requirements, with a value ranging from 1.5 to 3.0. A smaller α value results in higher sensitivity for detecting slow-moving, minutely displaced objects, suitable for warehousing goods displacement monitoring scenarios; a larger α value provides stronger anti-interference capabilities, suitable for high-dynamic scenarios with frequent forklift operations.

[0031] The S5 synchronously displays detection results and saves data, implemented as follows: The local touch interface of the industrial gateway displays the current moving object detection results, CSI amplitude time-domain waveform, and frequency-domain spectrum in real time. It also displays a timeline of movement events over the past 24 hours, supporting real-time viewing and historical data backtracking within 30 days. Raw CSI data, feature data, and detection results for the corresponding time period can be retrieved by time node. Simultaneously, the local device automatically saves feature data, detection results, and corresponding timestamps in a preset format. The main storage format is CSV, with backups in JSON format. Data fields include: timestamp, detection result category, recognition confidence, feature sequence mean, corresponding subcarrier frequency band information, and moving object motion parameters. All stored data is encrypted using AES-256 symmetric encryption. The encryption key is set locally by administrators and can only be decrypted and exported offline after authorization, ensuring data security in the warehouse scenario. For abnormal events such as area intrusion, after the detection result is triggered, the local device immediately saves the full CSI time-frequency domain data for 30 seconds before and after the event, and simultaneously pushes alarm information to the warehouse management system via industrial Ethernet, meeting the emergency alarm needs of industrial security.

[0032] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0033] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for environment perception detection based on Wi-Fi channel state information, characterized in that, The method includes the following steps: S1. Collect raw CSI data corresponding to each OFDM subcarrier in the target environment using commercial Wi-Fi devices; S2. Preprocess the acquired raw CSI data to eliminate noise interference and phase shift in the data, and obtain standardized effective CSI data. S3. The standardized CSI effective data are grouped by subcarrier, and multi-scale time-frequency domain feature extraction is performed to obtain the CSI feature sequence characterizing the motion characteristics of the moving object. S4. Input the CSI feature sequence into the pre-trained lightweight moving object perception and recognition model, and combine it with adaptive environment threshold judgment to output the moving object detection results in the target environment in real time. S5. Simultaneously display the CSI time-frequency domain feature map and moving object status corresponding to the detection results, and save the feature data and detection log.

2. The environmental perception and detection method based on Wi-Fi channel state information according to claim 1, characterized in that, The commercial Wi-Fi device collects raw Channel State Information (CSI) data within the target environment. Specifically, it uses a commercial Wi-Fi transceiver that supports the 802.11a / n / ac / ax protocol. The transmitting end sends standard OFDM data frames, and the receiving end collects raw CSI data for each OFDM subcarrier in the target environment through a Wi-Fi network card. The raw CSI data includes amplitude and phase information for each subcarrier. The collection frequency band covers the 2.4GHz and 5GHz Wi-Fi bands, and the number of subcarriers collected is no less than 30. 3.The method of claim 1, wherein, S2 specifically includes the following steps: S201. For the CSI amplitude timing sequence of each subcarrier, use... The criteria include outlier handling and calculation of the statistical mean of the sequence. with standard deviation The amplitude exceeds Data points in the interval are identified as outliers and removed. For missing data points resulting from the removal, linear interpolation between two adjacent valid data points is used to fill in the missing data points. S202. For the amplitude sequence after outlier removal, perform 4-level wavelet decomposition using the db4 wavelet basis to obtain one set of low-frequency approximation coefficients and four sets of high-frequency detail coefficients. Perform threshold quantization on the high-frequency detail coefficients using a soft thresholding function, setting the threshold... ,in , The noise standard deviation representing the high-frequency coefficient. Indicates the length of the sequence to be processed. Represents the natural logarithm; S203. After quantization, the signal is reconstructed by inverse wavelet transform. The reconstructed signal is then smoothed by a moving average filter with a window length of 7 to further suppress random burst noise.

4. The method of claim 3, wherein the method further comprises: S2 further includes the following steps: S204, fitting the CSI phase sequence , the linear slope is obtained by least square fitting with a fixed offset , and the center frequency of the subcarrier is determined , the calibrated phase information is calculated using and where ;​ S205. Phase calibration is performed sequence by sequence for each subcarrier phase sequence of the receiving antenna; S206. Using the min-max normalization method, the amplitude sequence and calibrated phase information of each subcarrier are normalized and mapped to the [0,1] interval. The specific normalization formula is as follows: ， wherein, denotes the original data value, denotes the minimum value of the current sequence, denotes the maximum value of the current sequence, denotes the normalized standardized data; S207. Through normalization processing, the amplitude differences between different subcarriers and the hardware gain differences between antennas are eliminated to obtain standardized CSI effective data. 5.The method of claim 1, wherein, S3 specifically includes the following steps: S301. Divide the standardized CSI effective data into several continuous subcarrier groups according to the subcarrier frequency correlation, and extract time-domain features, frequency-domain features and spatial features for each subcarrier group respectively. S302. Concatenate all features according to the time series to form a fixed-dimensional CSI feature sequence with a length of 256 dimensions. Use a sliding time window to extract continuous features. Set the sliding window length to 100ms and the window sliding step size to 20ms. 6.The method of claim 1, wherein, S4 specifically includes the following steps: S401. A lightweight moving object perception and recognition model receives a CSI feature sequence with a dimension of 256 as input. This input sequence is then fed into a first depthwise separable convolutional layer to decouple spatial feature extraction from channel feature fusion. The feature map output by the convolution is downsampled by a first max pooling layer to obtain pre-processed features. S402. The pre-processed features are sequentially passed through the second deep separable convolutional layer and the second max pooling layer. The multi-dimensional feature map output by the pooling layer is fed into the Flatten layer. All features are flattened into a one-dimensional vector and input into the GRU layer. After the GRU layer, the Dropout layer is connected. The regularized features are fed into the fully connected layer for linear transformation and dimension mapping, and finally output through the Softmax classification layer to output the probability distribution of each category. S403. The model finally outputs the probability distribution of 6 types of detection results, namely no moving object, object stationary, low-speed movement, medium-speed movement, high-speed movement, and area intrusion. The category corresponding to the maximum probability is taken as the final detection result. Low-speed movement is defined as movement speed less than or equal to 0.5 m / s, medium-speed movement is defined as movement speed greater than 0.5 m / s and less than or equal to 2 m / s, and high-speed movement is defined as movement speed greater than 2 m / s.

7. The environmental perception and detection method based on Wi-Fi channel state information according to claim 6, characterized in that, S4 specifically includes the following steps: S404. Collect 30 minutes of environmental CSI data with no moving objects and calculate the baseline mean of the corresponding feature sequences. and baseline standard deviation and the initial sensitivity coefficient Set to 2.5, utilize and Calculate the threshold ,in ; S405. Real-time calculation of the current CSI feature sequence With environmental baseline Feature Difference ,in , The current real-time 256-dimensional CSI feature sequence is the first... 3D eigenvalues The first environmental baseline characteristic sequence 3D eigenvalues The fixed total dimension representing the feature sequence; S406、When a moving object is determined to exist in the environment, immediately trigger the subsequent model fine recognition process, and when no moving object is determined to exist, do not trigger the model inference. S407, set the environmental baseline automatic updating mechanism, update the environmental baseline every 12 hours, the update period is set to the non-operation period of the warehouse, collect the environmental data without moving objects in the period to update the baseline mean and standard deviation, and adjust the sensitivity coefficient according to the scene requirements . 8.The method of claim 1, wherein, S5 specifically includes the following steps: S501: The local touch interface of the industrial gateway displays the current moving object detection results, CSI amplitude time-domain waveform and frequency-domain spectrum in real time. It also displays the timeline of moving events in the last 24 hours, supports real-time viewing and historical data backtracking within 30 days, and can retrieve the corresponding CSI raw data, feature data and detection results by time node. S502. The local device automatically saves feature data, detection results and corresponding timestamps in a preset format. The main storage format is CSV, and a backup JSON format is also provided. The data fields include timestamp, detection result category, recognition confidence, feature sequence mean, corresponding subcarrier frequency band information and moving object motion parameters. S503: All stored data is encrypted using AES-256 symmetric encryption. The encryption key is set locally by the administrator and can only be decrypted and exported offline after authorization. S504. In response to an abnormal intrusion event in the area, after the detection result is triggered, the local device immediately saves the full CSI time-frequency domain data for 30 seconds before and after the event, and pushes the alarm information to the warehouse management system via industrial Ethernet.

Images