A human daily activity recognition method based on channel state information data

Abstract

The application relates to a human daily activity recognition method based on channel state information data. The method comprises the following steps: collecting channel state information data generated by human behaviors, and performing data cleaning on the collected channel state information data; performing wavelet transform denoising and feature enhancement preprocessing on the channel state information data after data cleaning; performing time-frequency feature extraction on the preprocessed channel state information data to obtain a two-dimensional time-frequency feature map; constructing a human daily activity recognition hybrid model, training the human daily activity recognition hybrid model by using the two-dimensional time-frequency feature map, and obtaining a trained human daily activity recognition hybrid model; and inputting to-be-recognized channel state information data into the human daily activity recognition hybrid model, and outputting a human daily activity recognition result. The method can effectively compatibly improve the recognition accuracy, efficiency and generalization ability of human daily activity recognition.

Description

Technical Field

[0001] This application relates to the technical fields of wireless communication and intelligent sensing, and in particular to a method for recognizing daily human activities based on channel state information data. Background Technology

[0002] Temporal sequence recognition, as a core technology in the field of intelligent perception, has been widely applied in many fields such as human behavior recognition, speech recognition, and fault diagnosis. In particular, the integration of deep learning technology has significantly improved the accuracy and efficiency of temporal sequence recognition, forming a research pattern based on RNN and its variants, CNN, and Transformer, with hybrid models as the development trend.

[0003] Currently, the application of deep learning in time sequence recognition still faces many unresolved issues. Most existing time sequence recognition models are trained based on specific datasets and experimental environments. The training data of these models differs significantly from the data in real-world scenarios, leading to a substantial decrease in recognition performance across different scenarios, devices, and environments with multiple interferences. This problem is particularly prominent in human behavior recognition scenarios based on Channel State Information (CSI) data. Existing technologies also disclose various deep learning models, such as Transformer, which can efficiently capture global dependencies in long sequences, solving the problem of insufficient global dependency capture in traditional models. However, these models suffer from drawbacks such as large parameter count, high computational cost, and high training difficulty. Furthermore, they are not precise enough in perceiving subtle local features and perform poorly in capturing local fluctuation features of CSI signals, making it difficult to balance recognition accuracy and computational efficiency. Summary of the Invention

[0004] Therefore, it is necessary to address the problems of incompatibility between accuracy and efficiency in recognizing daily human activities and weak recognition generalization ability in the existing technologies mentioned above, and to provide a method for recognizing daily human activities based on channel state information data, which can effectively balance accuracy and efficiency in recognizing daily human activities and improve recognition accuracy, efficiency and generalization ability.

[0005] To achieve the above-mentioned technical effects, the technical solution of the present invention is as follows: A method for recognizing daily human activities based on channel state information data includes: S1. Collect channel state information data generated by human behavior, clean the collected channel state information data, and obtain cleaned channel state information data. S2. Perform wavelet transform denoising and feature enhancement preprocessing on the cleaned channel state information data to obtain preprocessed channel state information data; S3. Extract time-frequency features from the preprocessed channel state information data to obtain a two-dimensional time-frequency feature map; S4. Construct a hybrid model for recognizing daily human activities, and train the hybrid model for recognizing daily human activities using the two-dimensional time-frequency feature map to obtain a trained hybrid model for recognizing daily human activities. S5. Input the channel state information data to be identified into the human daily activity recognition hybrid model, and output the human daily activity recognition result.

[0006] Preferably, the acquisition of channel state information data generated by human behavior includes: S101. Set up a data acquisition device, which includes a transmitter and a receiver; S102. Under the selected acquisition environment, the transmitting end transmits a wireless signal, which is transmitted to the receiving end after being reflected and scattered by the human body; S103. The receiving end analyzes the wireless signal after it has been reflected and scattered by the human body to obtain the channel state information data.

[0007] Preferably, the data cleaning of the collected channel state information data includes: S111. Standardize the format of the collected channel state information data to obtain channel state information data with a regular format; S112. Perform preliminary noise filtering on the well-formatted channel state information data to obtain smoothed data after noise reduction; S113. Perform outlier detection and correction on the smoothed data to obtain corrected data; S114. Perform data length normalization processing on the corrected data to obtain dimension-standardized data; S115. Perform data redundancy removal and verification on the dimensional standardized data to obtain cleaned channel state information data.

[0008] Preferably, the preprocessing of wavelet transform denoising and feature enhancement on the cleaned channel state information data includes: S201. Perform phase offset correction on the cleaned channel state information data to obtain phase offset corrected channel state information data; S202. Perform several levels of discrete wavelet decomposition on the channel state information data after phase offset correction to separate low-frequency approximate components and high-frequency detail components. S203. Perform noise reduction processing on the high-frequency detail components, remove high-frequency random noise components, and retain the human behavior detail features in the high-frequency components to obtain high-frequency detail noise-reduced components. S204. Perform inverse wavelet transform on the high-frequency detail noise reduction component and the low-frequency approximation component to reconstruct the denoised reconstructed signal; S205. Normalize the amplitude of the reconstructed signal to obtain normalized data; S206. Perform time-series detail enhancement on the normalized data to amplify the signal mutations and fluctuations corresponding to various human behaviors, and obtain the preprocessed channel state information data.

[0009] Preferably, the step of extracting time-frequency features from the preprocessed channel state information data to obtain a two-dimensional time-frequency feature map includes: S301. Perform sliding segmentation on the preprocessed channel state information data to divide it into multiple short-time signal segments; S302. Perform Fourier transform on each short-time signal segment to extract the frequency distribution characteristics of each time period and obtain multiple sets of single-time period time-frequency characteristic data; S303. The multiple sets of single-time period time-frequency feature data are spliced ​​together to obtain the two-dimensional time-frequency feature map.

[0010] Preferably, the hybrid model for recognizing daily human activities includes a feature input module, an LSTM temporal feature extraction module, a Transformer global feature modeling module, a GMP feature filtering module, and a classification output module; The feature input module inputs the two-dimensional time-frequency feature map into the LSTM temporal feature extraction module. The LSTM temporal feature extraction module receives the two-dimensional time-frequency feature map and outputs low-dimensional local temporal feature vectors to the Transformer global feature modeling module and the GMP feature filtering module, respectively. The Transformer global feature modeling module receives the low-dimensional local temporal feature vectors and outputs a one-dimensional comprehensive feature vector. The GMP feature filtering module receives the low-dimensional local temporal feature vectors and outputs key pooling features. The classification output module receives the one-dimensional comprehensive feature vector and the key pooling features and outputs the human daily activity recognition result.

[0011] Preferably, the LSTM temporal feature extraction module includes a first input layer, an LSTM hidden layer, a dropout regularization layer, and a first output layer connected in sequence. The first input layer receives the two-dimensional time-frequency feature map and converts it into a temporal feature sequence through dimensionality reshaping, which is then input into the LSTM hidden layer. The LSTM hidden layer extracts key local features corresponding to human behavior in the temporal feature sequence. The dropout regularization layer receives the key local features, randomly discards them, and outputs a first temporal optimized feature to the first output layer. The first output layer converts the first temporal optimized feature into the low-dimensional local temporal feature vector.

[0012] Preferably, the Transformer global feature modeling module includes a second input layer, a position encoding layer, a multi-head self-attention layer, a residual connection and layer normalization layer, a FeedForward fully connected layer, and a second output layer connected in sequence. The second input layer receives the low-dimensional local temporal feature vector, and converts the low-dimensional local temporal feature vector into a temporal feature sequence through dimensionality reshaping, which is then input into the position encoding layer. The position encoding layer adds positional information to the temporal feature sequence and outputs the temporal feature sequence with added positional information to the multi-head self-attention layer. The multi-head self-attention layer mines global dependencies in the temporal feature sequence with added positional information and outputs global correlation features to the residual connection and layer normalization layer. The residual connection and layer normalization layer standardizes the global correlation features and outputs standardized features to the FeedForward fully connected layer. The FeedForward fully connected layer performs nonlinear transformation and integration on the standardized features and outputs a second temporal optimized feature to the second output layer. The second output layer converts the second temporal optimized feature into the one-dimensional comprehensive feature vector.

[0013] Preferably, the classification output module includes a third fully connected layer and an activation function layer connected in sequence. The third fully connected layer performs dimensional mapping and feature integration on the fusion features of the one-dimensional comprehensive feature vector and the key pooling features, and outputs the mapping features adapted to the multi-behavior classification task to the activation function layer. The activation function layer performs probability normalization processing on the mapping features, outputs the recognition probability corresponding to each human behavior category, and selects the human behavior category with the highest recognition probability as the recognition result of human daily activities.

[0014] Preferably, training the hybrid model for recognizing daily human activities using the two-dimensional time-frequency feature map includes: S41. Divide the two-dimensional time-frequency feature map into a training set and a validation set, and perform data augmentation on the training set to obtain an augmented training set; S42. Iteratively train the human daily activity recognition hybrid model after the enhanced training set input parameters are initialized, monitor the model gradient value in real time, and optimize the training process by gradient clipping and dynamic adjustment of the learning rate to obtain the human daily activity recognition hybrid model after each round of iterative training. S43. For the hybrid model for recognizing human daily activities after each round of training iterations, the performance is verified using the validation set. When the loss of the validation set does not decrease continuously in a preset number of rounds, training is terminated, the optimal model parameters are saved, and the trained hybrid model for recognizing human daily activities is output.

[0015] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: This invention proposes a method for recognizing daily human activities based on channel state information data. First, channel state information data generated by human behavior is collected. This data is then cleaned to obtain cleaned channel state information data, effectively filtering out environmental interference noise in real-world scenarios and ensuring the validity and reliability of the channel state information data. Second, wavelet transform denoising and feature enhancement preprocessing are applied to the cleaned channel state information data to accurately filter out complex environmental noise. Next, a two-dimensional time-frequency feature map is generated by extracting time-frequency features, achieving a comprehensive fusion of the temporal and frequency domain features of the channel state information data. This provides richer and more accurate feature data for the hybrid model of daily human activity recognition. Then, the hybrid model of daily human activity recognition is trained using the two-dimensional time-frequency feature map. Based on the trained hybrid model, the results of daily human activity recognition are output. Ultimately, this method effectively balances the accuracy and computational efficiency of daily human activity recognition, significantly improving the accuracy, efficiency, and generalization ability of daily human activity recognition in complex interference environments, across devices, and across scenarios. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for recognizing daily human activities based on channel state information data in one embodiment. Figure 2 This is a schematic diagram of the process for collecting channel state information data generated by human behavior in one embodiment. Figure 3 This is a schematic diagram illustrating the process of cleaning collected channel state information data in one embodiment; Figure 4 This is a schematic diagram of the preprocessing steps for wavelet transform denoising and feature enhancement of cleaned channel state information data in one embodiment. Figure 5 This is a flowchart illustrating the process of extracting time-frequency features from preprocessed channel state information data in one embodiment. Figure 6This is a structural block diagram of a hybrid model for recognizing daily human activities in one embodiment; Figure 7 This is a block diagram of the LSTM temporal feature extraction module in one embodiment; Figure 8 This is a block diagram of the Transformer global feature modeling module in one embodiment. Detailed Implementation

[0017] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. It is understandable to those skilled in the art that some well-known details may be omitted from the accompanying drawings; To facilitate understanding of this embodiment, the prior art information of this embodiment is first introduced as follows: Regarding the current state of research on human behavior recognition based on Channel State Information (CSI), with the deep integration of wireless communication technology and artificial intelligence algorithms, human behavior recognition based on Channel State Information (CSI) data has become a research hotspot in the field of contactless sensing. In recent years, researchers at home and abroad have conducted a large amount of research on core aspects such as CSI signal preprocessing, feature extraction, and model construction, resulting in relatively rich research results. The overall research status can be discussed from three core dimensions: signal preprocessing, feature extraction, and recognition model, as detailed below: In CSI signal preprocessing, the core objective is to filter out signal noise and improve signal quality, laying the foundation for subsequent feature extraction and model training. Raw CSI signals are susceptible to interference from environmental noise, multipath effects, and hardware clock skew, resulting in signal jitter and low signal-to-noise ratio. Therefore, preprocessing is a crucial preliminary step in CSI behavior recognition. Currently, mainstream preprocessing methods fall into two main categories: noise filtering methods, including wavelet denoising, moving average filtering, median filtering, Kalman filtering, and Butterworth low-pass filtering, are widely used in CSI signal preprocessing due to their ability to effectively preserve signal details and filter high-frequency noise; and signal calibration methods, primarily used to address CSI phase shift issues by correcting system errors introduced by hardware devices through phase calibration algorithms, thereby improving signal stability. Existing research often uses preprocessing methods singly or in simple combinations, with insufficient attention paid to the synergistic effect of "noise filtering-feature adaptation," making it difficult to fully leverage the role of preprocessing in improving subsequent recognition performance.

[0018] In terms of feature extraction, research focuses on how to extract features from preprocessed CSI signals that can effectively distinguish different human behaviors. Currently, three main categories of feature extraction methods have been developed: time-domain, frequency-domain, and time-frequency-domain. Time-domain features mainly include statistical features such as the signal's mean, variance, peak value, and waveform slope. These extraction methods are simple and computationally inexpensive, but they struggle to capture the dynamic changes in signals caused by human behavior. Frequency-domain features are mainly extracted using methods such as Fourier Transform (FT) and Short-Time Fourier Transform (STFT), which can reflect the frequency distribution characteristics of the signal. Among these, STFT, which can take into account both time and frequency domain information, effectively captures the dynamic time-frequency changes of human behavior and has become the most commonly used time-frequency domain feature extraction method in CSI behavior recognition. In addition, some researchers have proposed feature extraction methods based on wavelet transform and power spectral density, further enriching the feature representation dimensions of CSI signals. However, existing feature extraction methods still suffer from problems such as feature redundancy and insufficient discriminative power, especially in complex environments where it is difficult to effectively extract key features strongly correlated with human behavior.

[0019] In terms of recognition models, research has evolved from traditional machine learning to deep learning. Early research mainly used traditional machine learning algorithms such as Support Vector Machines (SVM) and K-Nearest Neighbors (KNN), relying on manually extracted features to classify behaviors. While these algorithms offered advantages such as low computational cost and easy deployment, their recognition accuracy and generalization ability were limited, making them unsuitable for multi-class behavior recognition tasks in complex scenarios. In recent years, the rise of deep learning technology has significantly improved the accuracy of CSI behavior recognition. Researchers have combined the temporal characteristics of CSI signals and proposed recognition schemes based on temporal models such as Recurrent Neural Networks (RNN), Long Short-Term Memory Networks (LSTM), and Gated Recurrent Units (GRU). These schemes can effectively capture the long-term dependencies of CSI signals and improve recognition accuracy. With the application of the Transformer architecture in temporal signal processing, some researchers have proposed LSTM-Transformer hybrid models, which balance temporal feature capture and global feature extraction, further optimizing the recognition performance of complex behaviors. In addition, multimodal fusion models have become a research hotspot. By fusing CSI with audio, infrared and other modal information, the robustness of the model can be improved. However, such models have problems such as complex structure, high computational cost and difficulty in deployment, which make it difficult to meet the needs of practical applications.

[0020] Overall, CSI-based human behavior recognition technology has made significant progress, forming relatively mature research approaches in preprocessing, feature extraction, and model building. However, several problems remain to be solved: First, the synergistic adaptability between preprocessing methods, feature extraction, and models is insufficient, failing to fully leverage the collaborative effects of each stage; second, the feature stability of CSI signals is poor in complex environments (such as multi-interference and cross-scene environments), resulting in insufficient model generalization ability; third, existing models are mostly designed for specific scenarios and behavior categories, exhibiting poor universality and scalability; fourth, the lack of a unified standard for CSI data acquisition, with different hardware platforms and acquisition tools generating varying CSI data formats, making direct comparison between different studies difficult. These problems have become the main research challenges in the current field of CSI-based human behavior recognition, while also providing a clear entry point and research direction for this invention.

[0021] Regarding the current application of deep learning in time sequence recognition, time sequence recognition refers to the process of feature extraction, pattern analysis, and category determination of sequential data (such as signals, video frames, and text sequences) arranged in chronological order. It is widely used in fields such as human behavior recognition, speech recognition, fault diagnosis, and financial prediction. With the rapid iteration of deep learning technology, its powerful automatic feature learning capabilities and time-dependent modeling capabilities have gradually replaced traditional time sequence recognition methods based on manual feature extraction, becoming the core research area and mainstream technology in the current field of time sequence recognition. In recent years, researchers have focused on optimizing and innovating deep learning architectures, addressing core pain points such as the time dependence, multi-scale characteristics, and noise interference of time sequence data. They have proposed a series of time sequence recognition models adapted to different scenarios, forming an evolutionary path of "basic model → optimized model → hybrid model." The overall application status can be expanded according to the mainstream deep learning model categories, as follows: 1) Recurrent Neural Networks (RNNs) and their variants: the foundation for time-series identification Recurrent Neural Networks (RNNs) were the earliest deep learning architecture applied to time sequence recognition. Their core advantage lies in capturing the dependencies between time-series data through the cyclical propagation of hidden layer states, thus enabling the modeling of dynamic changes in sequences. However, traditional RNNs suffer from vanishing or exploding gradient problems, making them difficult to handle long sequence data and limiting their application in complex time sequence recognition scenarios. Therefore, researchers have proposed a series of RNN variants, among which Long Short-Term Memory Networks (LSTM) and Gated Recurrent Units (GRUs) have become the mainstream basic models in time sequence recognition.

[0022] LSTM employs a triple gating mechanism (input gate, forget gate, and output gate) to selectively retain and discard historical information, effectively mitigating the vanishing gradient problem. It can accurately capture long-term dependencies in time-series data and is one of the most widely used fundamental models in current time-series recognition. In the field of CSI (Human Behavior Sequence) recognition, LSTM is widely used to process preprocessed CSI time-series signals. By learning the dynamic patterns of signal changes with human behavior, it achieves accurate differentiation of different behavior categories, and is one of the core fundamental models used in this invention. In fields such as speech recognition and handwriting recognition, LSTM can effectively capture the temporal features of speech sequences and handwriting trajectories, improving recognition accuracy.

[0023] As a simplified variant of LSTM, GRU merges the input gate and forget gate into an update gate, simplifying the model structure and reducing the number of parameters. While maintaining essentially the same recognition accuracy, it improves the training speed and deployment efficiency of the model. Its application scenarios highly overlap with LSTM. In resource-constrained temporal recognition scenarios (such as behavior recognition deployed on edge devices and real-time fault diagnosis), GRU shows a more significant advantage due to its lightweight nature.

[0024] Currently, the application of RNNs and their variants has matured, but they still have certain limitations: First, they are difficult to compute in parallel, and the training efficiency is greatly affected by the sequence length; second, they have limited ability to capture global dependencies of long sequences, and their performance will drop significantly in the recognition of ultra-long time sequences; third, they are susceptible to noise in time series data, and their robustness needs to be improved.

[0025] 2) Application of Convolutional Neural Networks (CNNs) in Time Sequence Recognition Convolutional Neural Networks (CNNs) were initially applied to image recognition. Their core advantage lies in their ability to efficiently extract local features from data through convolution and pooling operations, while also possessing strong parallel computing capabilities and noise resistance. With the deepening of research in temporal recognition, CNNs have been gradually applied to temporal data processing. By sliding one-dimensional convolutional kernels along the time dimension, they extract local temporal features from temporal data, thus overcoming the shortcomings of insufficient parallel computing capabilities in RNNs and their variants.

[0026] In time series recognition, the core application of CNNs is to treat one-dimensional time series data as an "image in the time dimension." One-dimensional convolution operations capture feature patterns within local time windows, and pooling operations are then used to reduce the dimensionality of these features, preserving key time series information. For example, in the time series recognition of CSI signals, one-dimensional CNNs can be used to extract local time-frequency patterns from the STFT time-frequency feature map, capturing local signal changes caused by human behavior. In the field of fault diagnosis, CNNs can identify abnormal patterns in equipment operating status by extracting local features from equipment vibration time series signals.

[0027] In recent years, researchers have proposed a series of CNN variants adapted for time-series recognition. Among them, Temporal Convolutional Networks (TCNs) expand the receptive field of convolutional kernels by dilating causal convolutions, enabling them to capture local features while also taking into account the global dependencies of long sequences, further improving the performance of CNNs in time-series recognition. TimesNet innovatively transforms one-dimensional time series into two-dimensional matrices, capturing periodic patterns at different time scales through multiple convolutional kernels, demonstrating outstanding performance in long-term time-series recognition tasks. Currently, the application of CNNs in time-series recognition mainly focuses on local temporal feature extraction, often combined with other models to fully leverage their advantages in parallel computing and noise resistance.

[0028] 3) Breakthroughs of attention mechanisms and Transformers in temporal recognition The introduction of attention mechanisms addresses the core pain point of traditional deep learning models in long sequence recognition: insufficient capture of global dependencies. Its core idea is to calculate attention weights at different time steps in the sequence, focusing on temporal features that play a crucial role in the recognition result while ignoring redundant information. The combination of attention mechanisms with RNNs and CNNs significantly improves the temporal recognition performance of models, while the emergence of the Transformer architecture has further propelled breakthroughs in temporal recognition technology.

[0029] The Transformer architecture, with its self-attention mechanism at its core, overcomes the sequence dependency limitations of RNNs, enabling parallel computation of dependencies across all time steps in a sequence. This allows for efficient capture of global dependencies in long sequences, making it particularly suitable for recognizing ultra-long time series data. In the field of time series recognition, the application of Transformer and its variants is increasingly widespread. The Informer model optimizes computational complexity for long sequence recognition, reducing it from O(n²) to O(n log n) through the ProbSparse self-attention mechanism, making it suitable for scenarios such as power load forecasting and long-term sequence behavior recognition. PatchTST, on the other hand, improves the efficiency of local feature extraction by dividing the time series into "time blocks," demonstrating outstanding performance in multivariate time series recognition.

[0030] In the field of CSI (Content Detection and Interpretation) human behavior recognition, Transformer is often combined with LSTM to form a hybrid model. This model combines the ability of LSTM to capture local temporal dependencies with the ability of Transformer to extract global features, effectively solving the problem that a single model cannot simultaneously capture both local fluctuations and global variation patterns in CSI signals. It is also one of the key model architectures compared in this study. In addition, Transformer has also been widely used in speech recognition, video temporal recognition, and financial time series prediction, gradually becoming the mainstream architecture for long-sequence temporal recognition.

[0031] 4) Hybrid Deep Learning Models: The Mainstream Development Trend of Time Series Recognition Each single deep learning model has its limitations in time series recognition: RNNs and their variants excel at capturing temporal dependencies but have weak parallel computing capabilities; CNNs are good at extracting local features but lack the ability to capture global dependencies; Transformers are good at capturing global dependencies but are weak at extracting local detailed features, and have a large number of parameters and high training costs. Therefore, in recent years, "hybrid models" have become the mainstream development trend of deep learning in time series recognition, combining the advantages of different models to further improve the performance of time series recognition.

[0032] Currently, mainstream hybrid models are mainly divided into three categories: First, the hybrid of RNN / GRU / LSTM and CNN, such as the LSTM-CNN model, which extracts local temporal features through CNN and captures temporal dependencies through LSTM, suitable for scenarios such as CSI behavior recognition and fault diagnosis; Second, the hybrid of RNN / LSTM and Transformer, such as the LSTM-Transformer model involved in this study, which combines the accurate capture of short-term temporal dependencies by LSTM and the efficient extraction of global features by Transformer, and performs well in complex temporal recognition tasks; Third, the hybrid of CNN and Transformer, which extracts local features through CNN and captures global dependencies by Transformer, balancing recognition accuracy and computational efficiency.

[0033] In addition, the combination of graph neural networks (GNNs) and deep learning models has become an emerging direction for time series recognition. GNNs capture the inter-variable dependencies of multivariable time series data by constructing the topological structure of time series data, and are suitable for scenarios such as multi-sensor collaborative monitoring and multi-dimensional time series recognition. Emerging architectures such as diffusion models and LLM models are also gradually being applied to time series recognition, providing new ideas for probabilistic time series recognition and general time series modeling.

[0034] In summary, deep learning has become a core technology in the field of temporal recognition, forming an application pattern based on RNNs and their variants, CNNs, and Transformers, with hybrid models as the development trend. It has been widely applied in various fields such as CSI human behavior recognition, speech recognition, fault diagnosis, and financial prediction, significantly improving the accuracy, efficiency, and robustness of temporal recognition. Especially in the field of CSI human behavior recognition studied in this invention, deep learning models, through automatic feature learning of CSI temporal signals, effectively solve the problems of high difficulty in manual feature extraction and low recognition accuracy in traditional methods, thus promoting the development of contactless behavior recognition technology.

[0035] However, the application of deep learning in time series recognition still faces many unresolved issues: First, the generalization ability of models is insufficient. Most models are designed for specific scenarios and datasets, and their recognition performance drops significantly in cross-scenario, cross-device, and multi-interference environments. This is also one of the core problems faced in CSI human behavior recognition. Second, the interpretability of models is poor. The "black box" nature of deep learning models makes it difficult to explain the model's recognition logic and key features, which is not conducive to model optimization and practical application. Third, it is difficult to balance efficiency and accuracy in long-sequence time series recognition. Although models such as Transformer can capture global dependencies in long sequences, they have a large number of parameters and high training costs, making them difficult to deploy on edge devices. Fourth, there is a strong data dependency. Deep learning models require a large amount of high-quality time series data for training, but in real-world scenarios, time series data often suffers from noise, missing data, and imbalance, affecting model performance.

[0036] In summary, temporal recognition, as a core technology in the field of intelligent sensing, has been widely applied in various fields such as human behavior recognition, speech recognition, and fault diagnosis. In particular, the integration of deep learning technology has significantly improved the accuracy and efficiency of temporal recognition, forming a research pattern based on RNNs and their variants, CNNs, and Transformers, with hybrid models as the development trend. However, considering the current research status of deep learning in temporal recognition, especially for the CSI human behavior recognition scenario that this invention focuses on, existing research still has many shortcomings that urgently need to be addressed. These include: generally insufficient model generalization ability, making it difficult to adapt to complex and ever-changing real-world application scenarios; structural limitations of single deep learning models, making it difficult to balance the accuracy and efficiency of temporal recognition; insufficient collaborative optimization of hybrid models, failing to fully leverage the complementary advantages of each model; poor model interpretability, with the "black box" characteristic hindering model optimization and practical implementation; insufficient synergistic adaptability between preprocessing and models, with the impact of signal noise and redundant information not effectively addressed; and insufficient model lightweighting and deployment efficiency, making it difficult to meet the practical application needs of edge devices. These shortcomings not only restrict the further development of temporal recognition technology but also provide a clear entry point for this invention. In summary, current research on deep learning in the field of temporal recognition still has significant shortcomings in terms of model generalization ability, structure optimization, collaborative adaptation, interpretability, and lightweight design, especially in the CSI human behavior recognition scenario, where these shortcomings are even more pronounced.

[0037] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0038] Example 1 like Figure 1As shown in the figure, this embodiment proposes a method for recognizing daily human activities based on channel state information data, characterized by including: S1. Collect channel state information data generated by human behavior, clean the collected channel state information data, and obtain cleaned channel state information data. Channel State Information (CSI) data acquisition is fundamental. Collecting CSI data generated by human behavior provides the raw data source for subsequent signal processing and model training, and its design directly determines the upper limit of subsequent recognition performance. As the core data source for human behavior recognition in this invention, the quality of CSI data acquisition directly determines the effectiveness of subsequent feature extraction and the accuracy of model recognition. Because CSI signals are susceptible to environmental interference, equipment errors, and fluctuations in human posture, the raw data often contains a large amount of noise, outliers, and redundant information. Directly using this data for model training will lead to a decrease in model generalization ability and recognition accuracy. Therefore, this invention designs a standardized CSI data acquisition and cleaning process to ensure data quality from the data source, laying a solid foundation for subsequent time-frequency feature extraction and model training. The overall process is divided into two core stages: data acquisition and data cleaning. These stages work together to ensure the effectiveness, completeness, and consistency of the data. The data acquisition stage focuses on "accurately capturing CSI signals corresponding to human behavior." Combining the transmission characteristics of CSI signals and the diversity of human behavior, a standardized acquisition scheme is developed, clearly defining the acquisition equipment, acquisition environment, acquisition object, acquisition behavior, and acquisition parameters to ensure the representativeness and repeatability of the acquired data. The channel state information data generated by the collection of human behavior is referred to Figure 2 ,include: S101. Construct a data acquisition device, which includes a transmitter and a receiver; the data acquisition device uses a Wi-Fi device based on OFDM technology as the data acquisition device. S102. In the selected acquisition environment, the transmitting end sends a wireless signal, which is then reflected and scattered by the human body before being transmitted to the receiving end. The transmitting end is a Wi-Fi router that supports the 802.11n protocol, continuously transmitting a wireless signal at a fixed frequency. An open indoor area is selected as the acquisition environment, with the area controlled within a 10m×8m range to avoid the influence of factors such as wall obstruction, multipath propagation, and electromagnetic interference on the CSI signal. During the acquisition process, other indoor wireless devices are turned off to reduce interference from signals in the same frequency band. The ambient temperature and humidity are kept stable to avoid signal fluctuations caused by environmental factors, ensuring that the acquired CSI signal can accurately reflect the impact of human behavior.

[0039] S103. The receiving end analyzes the wireless signal after reflection and scattering by the human body to obtain the channel state information data. The receiving end uses a terminal device equipped with a CSI acquisition tool. During the acquisition process, it is ensured that the transmitting end and the receiving end maintain a fixed distance to avoid interference to the signal caused by changes in device position. To reduce the device's own error, the wireless network card of the data acquisition device is calibrated before acquisition to ensure frequency synchronization and power stability of the transmitting end and the receiving end, and to avoid signal distortion caused by device failure.

[0040] The data collection subjects and behaviors were set by selecting three volunteers of the same age group but different heights and weights to cover different body types and ensure data diversity. The collected human behaviors covered the seven types of behaviors that this invention focuses on identifying: walking, running, sitting down, standing up, falling, lying down, and bending over. Each behavior was collected 25 times from each volunteer, with each collection session lasting 3-5 seconds to ensure sufficient sample support for each behavior category while avoiding sample redundancy.

[0041] Acquisition parameter settings: Set the sampling frequency of the CSI signal to 100Hz to ensure that it can capture subtle signal fluctuations caused by human behavior; match the sampling duration with the duration of the behavior, and control the number of sampling points for each behavior sample to 300-500; save the acquired CSI signals in .txt format, and record the tag information such as the behavior category, acquisition time, and acquisition object corresponding to the sample, so as to facilitate subsequent data management and classification processing.

[0042] Data acquisition and verification: During the acquisition process, the stability of the CSI signal is monitored in real time. If signal interruption or abnormal amplitude changes occur, the sample for that behavior is re-acquired. After the acquisition is completed, 10% of the samples are randomly selected for manual verification to confirm that the sample label is consistent with the actual behavior and that the signal is not obviously distorted, thus ensuring the validity of the acquired data.

[0043] The data acquisition device also includes a data storage unit, which is responsible for storing the acquired raw channel state information data (including amplitude and phase information) in matrix form, and at the same time recording the corresponding human behavior labels (such as walking, sitting down, standing up, falling down, sitting still, etc.) to form a channel state information data dataset, providing data support for subsequent model training and validation.

[0044] The data cleaning phase focuses on "removing noise, correcting anomalies, and retaining valid information." Addressing issues such as noise, outliers, and redundant data in the original CSI data, a multi-step cleaning process is designed to progressively optimize data quality and ensure that the cleaned data accurately reflects the correlation between human behavior and CSI signals. The specific cleaning process is as follows: The process of cleaning the collected channel state information data is described in detail below. Figure 3 ,include: S111. Standardize the format of the collected channel state information data to obtain channel state information data with a regular format. The format standardization process includes: firstly, standardizing the format of the original collected CSI data, extracting the CSI amplitude and phase data from each sample, and removing invalid identifiers, redundant annotations, and other information generated during the collection process; and then reorganizing the data of different collection objects and different behavior categories according to the naming rule of "behavior category-collection object-sample number" to establish a standardized dataset directory for easy subsequent data retrieval and processing.

[0045] S112. Perform preliminary noise filtering on the well-formatted channel state information (CSI) data to obtain smoothed data after denoising. The preliminary noise filtering includes addressing the random fluctuations in environmental interference noise (such as electromagnetic interference and multipath reflection noise) contained in the original CSI data. A moving average filtering algorithm is used to initially filter the CSI amplitude signal. The moving window size is set to 5, and the amplitude at each sampling point is locally averaged to smooth signal fluctuations, suppress random noise, and retain signal abrupt changes caused by human behavior, avoiding feature loss due to over-filtering.

[0046] S113. Outlier detection and correction are performed on the smoothed data to obtain corrected data. The outlier detection and correction includes using the 3σ criterion to detect outliers in the CSI data. This involves calculating the mean μ and standard deviation σ of the CSI amplitude for each behavior category, and identifying data exceeding the range of [μ-3σ, μ+3σ] as outliers—these outliers mainly originate from unexpected situations such as accidental device touches or sudden changes in human posture during the data acquisition process. For outliers, linear interpolation is used for correction. Valid data points adjacent to the outlier are used to supplement the values ​​at the outlier location through linear fitting, ensuring the continuity of the data sequence and avoiding the impact of outliers on subsequent feature extraction.

[0047] S114. Perform data length normalization processing on the corrected data to obtain dimensionally standardized data. The data length normalization processing includes addressing the issue that, due to differences in the duration of behavior among different volunteers, the length of the collected CSI sample data is inconsistent (the number of sampling points ranges from 300 to 500). Directly using this data for model training would lead to inconsistent input feature dimensions. Therefore, linear interpolation is used to normalize the length of all samples, uniformly adjusting the number of sampling points for each sample to 400. This preserves the temporal variation pattern of the original signal while ensuring consistent dimensions across all samples, meeting the requirements for subsequent model input.

[0048] S115. Perform data redundancy removal and verification on the dimensional standardized data to obtain cleaned channel state information data; the data redundancy removal and verification includes removing redundancy from the cleaned CSI data, eliminating duplicate samples and invalid samples (such as samples that still have obvious distortion after normalization); subsequently, randomly select 20% of the cleaned samples again, manually verify the completeness and accuracy of the data, confirm that noise has been effectively suppressed, outliers have been corrected, and data dimensions are unified, to ensure that the cleaned data can meet the needs of subsequent time-frequency feature extraction and model training.

[0049] The CSI data acquired in S1 must meet two core requirements: first, the acquisition frequency must be adapted to the dynamic changes in human behavior to ensure complete recording of the entire process of each behavior; second, interference factors in the acquisition environment must be controlled to reduce the impact of electromagnetic interference and multipath redundancy on the raw signal, easing the burden on subsequent preprocessing modules, while also echoing the temporal continuity and fluctuation characteristics of the CSI data mentioned earlier. This step's CSI data acquisition and cleaning process ensures the representativeness and repeatability of the raw data through a standardized acquisition scheme, and effectively solves problems such as noise, outliers, and redundancy in the raw data through a multi-step cleaning process, ultimately obtaining a high-quality, standardized CSI dataset. This process fully considers the characteristics of CSI signals and the needs of human behavior recognition, providing reliable data support for subsequent time-frequency feature extraction in S3 and model training in S4, and is a key prerequisite for improving the accuracy of human behavior recognition in this invention.

[0050] S2. Perform wavelet transform denoising and feature enhancement preprocessing on the cleaned channel state information data to obtain preprocessed channel state information data; While S1 has completed format standardization, outlier correction, length normalization, and preliminary noise filtering, effectively addressing issues such as redundancy, anomalies, and inconsistent dimensions in the original data, a small amount of high-frequency interference noise remains. This noise primarily originates from environmental electromagnetic pulse interference, minor vibrations of wireless devices, and residual signals from multipath reflections. It often manifests as high-frequency random fluctuations, masking subtle changes in CSI signals caused by human behavior. This results in indistinct signal feature differences between different behaviors, making it difficult for deep learning models to accurately capture them, directly impacting the effectiveness of subsequent feature extraction and model recognition accuracy. Therefore, S2 introduces wavelet transform technology, designing a processing scheme that combines denoising and feature enhancement. This achieves deep suppression of residual noise and enhancement of effective human behavior features, further optimizing data quality and providing a highly discriminative data source for subsequent time-frequency feature extraction and model training. Wavelet transform denoising and feature enhancement preprocessing are crucial for improving CSI signal quality. Their core function is to filter out noise and redundant information from the original CSI signal, calibrate signal deviations, and provide high-quality signal data for subsequent feature extraction.

[0051] Wavelet Transform (WT) is a multi-resolution time-frequency analysis method. Its core advantage lies in its ability to simultaneously maintain analytical accuracy in both the time and frequency domains. It can decompose non-stationary and nonlinear CSI time-series signals into wavelet components at different frequency scales, achieving precise separation of effective signals from noise components. Compared to traditional linear filtering methods such as moving average filtering, wavelet transform does not require the assumption of linearity in the signal. It can suppress noise while preserving the detailed features of the effective signal to the greatest extent, avoiding the loss of behavioral features due to over-filtering. It is particularly suitable for the characteristics of CSI signals, which are characterized by "concentrated effective features and dispersed noise, mostly at high frequencies." The effective CSI signals caused by human behavior are mainly concentrated in the low- and mid-frequency regions, reflecting the temporal changes in behavior; while residual interference noise is mainly distributed in the high-frequency region, exhibiting irregular random fluctuations. This characteristic provides a natural separation basis for wavelet transform denoising.

[0052] Considering the discrete nature of CSI signals and the ease of computer processing, this invention employs Discrete Wavelet Transform (DWT) for denoising and feature enhancement. Based on the amplitude characteristics, noise distribution, and behavioral requirements of CSI signals, and after multiple experimental adjustments, the db4 wavelet was selected as the base wavelet. The db4 wavelet possesses good temporal locality and frequency resolution, accurately capturing the temporal fluctuation details and frequency variation characteristics of CSI signals, while also exhibiting strong noise separation capabilities, making it suitable for processing non-stationary time-series signals like CSI. The overall implementation process consists of four core steps: wavelet decomposition, noise thresholding, signal reconstruction, and feature enhancement. These steps work together to ensure optimal denoising and feature enhancement results.

[0053] The preprocessing of the cleaned Channel State Information (CSI) data, including wavelet transform denoising and feature enhancement, involves several steps. First, the original signal is calibrated. Addressing hardware errors and clock synchronization deviations in the original CSI phase data, a phase calibration algorithm corrects phase shifts to ensure data validity. Second, wavelet transform denoising is performed using the db series mother wavelet, adapted to non-stationary CSI signals. The calibrated CSI signal is decomposed into low-frequency approximate components (useful signals) and high-frequency detail components (noise). An adaptive threshold is set to filter out high-frequency noise components while preserving the detailed features of the useful signal. Finally, signal standardization is applied to the denoised CSI signal, eliminating differences in data dimensions and mapping signal amplitudes to a fixed range. This prevents excessively large data ranges from interfering with subsequent model training and ensures the signal data meets the input requirements of STFT feature extraction and deep learning models. (See also...) Figure 4 More specific steps include: S201. Perform phase offset correction on the cleaned channel state information data to obtain phase offset corrected channel state information data; S202. Perform a 3-level discrete wavelet decomposition on the channel state information data after phase offset correction to separate one low-frequency approximation component (cA3) and three high-frequency detail components (cD1, cD2, cD3). Among them, the low-frequency approximation component cA3 contains the core effective information of the CSI signal, mainly corresponding to the signal change pattern caused by human behavior, which is the key to distinguishing different behavior categories. The high-frequency detail components cD1, cD2, and cD3 mainly contain residual high-frequency interference noise, while also containing a small amount of signal details. The decomposition level is set to 3 levels, which can effectively separate noise from effective signals, avoid distortion of effective features caused by excessive decomposition, and balance the denoising effect and feature preservation.

[0054] S203. Denoising the high-frequency detail components by removing high-frequency random noise components and retaining the human behavior details in the high-frequency components to obtain the denoised high-frequency detail components. The denoising process includes: using a soft thresholding method to suppress noise in the high-frequency detail components after wavelet decomposition. The core idea is to distinguish between noise and effective signal details in the high-frequency components by setting a reasonable threshold. The advantage of soft thresholding is that it can effectively suppress noise while avoiding distortion at signal abrupt changes, and it fits the amplitude abrupt changes caused by human behavior in CSI signals (such as signal abrupt changes in behavior like falling or standing up). The rationality of the threshold directly determines the denoising effect. This invention uses an adaptive threshold calculation method to automatically adjust the threshold according to the noise intensity of the high-frequency detail components, avoiding the problem of incomplete denoising or loss of effective features caused by a fixed threshold. The threshold calculation formula is as follows:

[0055] In the formula, For adaptive threshold, The noise standard deviation of the high-frequency detail components can be estimated from the median of the high-frequency detail components; N is the number of sampling points in the CSI signal. The specific processing procedure is as follows: For high-frequency detail components, those with absolute values ​​less than a threshold... The coefficients are set to 0 (determined to be noise components), and coefficients with absolute values ​​greater than or equal to the threshold are shrunk (preserving the details of the effective signal), thereby achieving accurate separation of noise and effective signal.

[0056] S204. The high-frequency detail denoising components and the low-frequency approximation components are reconstructed using inverse wavelet transform to obtain the denoised reconstructed signal. The inverse wavelet transform reconstruction includes: performing inverse wavelet transform on the three high-frequency detail components (denoised cD1, cD2, and cD3) after soft thresholding and the original low-frequency approximation component cA3 to reconstruct the denoised CSI amplitude signal. The reconstruction process can fuse the denoised high-frequency details with the effective low-frequency features, restore the temporal variation law of the CSI signal caused by human behavior, and completely suppress residual high-frequency interference noise, significantly improving signal smoothness. It also fully preserves key details such as signal mutations and periodic fluctuations corresponding to the behavior, providing a high-quality basic signal for subsequent feature enhancement.

[0057] S205. Normalize the amplitude of the reconstructed signal to obtain normalized data; S206. Perform time-series detail enhancement on the normalized data to amplify the signal mutations and fluctuations corresponding to various human behaviors, and obtain the preprocessed channel state information data.

[0058] Building upon denoising, steps S205-S206 further enhance the features of the reconstructed CSI signal. The core objective is to strengthen the feature differences corresponding to different human behaviors, improve feature discriminability, and facilitate subsequent time-frequency feature extraction and feature learning by deep learning models. This invention employs a combined strategy of "amplitude normalization + temporal detail enhancement," specifically implemented as follows: First, the min-max normalization method is used to normalize the denoised CSI amplitude signal to the [0, 1] interval, eliminating amplitude scale differences between different acquisition objects and different behavior samples, ensuring consistency in subsequent feature extraction; second, the first-order difference of the denoised signal is calculated to enhance the temporal change details of the signal, amplifying amplitude abrupt changes caused by human behavior (such as sudden increases and decreases in amplitude of falling behavior, and gradual changes in amplitude of bending behavior) and periodic fluctuations (such as periodic amplitude changes in walking and running behavior), making the signal feature differences corresponding to different behaviors more significant and improving the discriminative power of the features.

[0059] Based on the practical needs of CSI human behavior recognition in this invention, the key parameters for wavelet transform denoising and feature enhancement were determined through multiple comparative experiments. The final parameter settings are as follows: the base wavelet is db4, the wavelet decomposition level is 3, the threshold calculation uses the adaptive threshold method, the first-order differential enhancement window size is 3, and the amplitude normalization uses min-max normalization. This parameter combination can achieve the optimal denoising and feature enhancement effect. Experiments have verified that the signal-to-noise ratio (SNR) of the CSI signal after denoising is more than 30% higher than that after cleaning, the amplitude of residual noise is suppressed to below 0.05, and the signal smoothness is significantly improved. After feature enhancement, the amplitude fluctuation difference of the CSI signal corresponding to different behaviors is increased by 20%~25% compared with that before enhancement. Among them, the signal abrupt change characteristics of behaviors such as falling and lying down are more obvious, and the periodic characteristics of walking and running are more prominent.

[0060] This invention addresses the problem of insufficient preprocessing and model co-adaptability in existing research through the aforementioned S2. It precisely filters noise using wavelet transform while preserving the CSI signal fluctuation characteristics caused by human behavior, providing a clear and stable signal source for subsequent time-frequency feature extraction modules. The wavelet transform denoising and feature enhancement scheme of this invention achieves deep suppression of residual noise through discrete wavelet transform and strengthens effective human behavior features through a combined feature enhancement strategy. This further improves the CSI data preprocessing system, providing a high-quality, high-recognition data source for subsequent time-frequency feature extraction and deep learning model training, and plays a crucial supporting role in improving the accuracy of human behavior recognition.

[0061] S3. Extract time-frequency features from the preprocessed channel state information data to obtain a two-dimensional time-frequency feature map; After S2 wavelet transform denoising and feature enhancement, the channel state information data possesses high-quality, highly discriminative time-series features. However, these features only reflect the signal amplitude changes caused by human behavior and are insufficient to reflect the differences in signal frequency variations—while the distribution of channel state information data corresponding to different human behaviors differs significantly in the time-frequency domain. The Short-Time Fourier Transform (STFT), a classic time-frequency analysis method, can convert one-dimensional time-series channel state information data into a two-dimensional time-frequency feature map while preserving both time and frequency domain information. The core principle of STFT is to divide a one-dimensional non-stationary time-series signal into multiple short-time stationary sub-signals using a sliding window, perform a Fourier transform on each sub-signal, and finally concatenate them to obtain a two-dimensional time-frequency feature map. In this embodiment, STFT is used to extract time-frequency features from the preprocessed channel state information data to obtain a two-dimensional time-frequency feature map. (See [link to documentation]). Figure 5 ,include: S301. Perform sliding segmentation on the preprocessed channel state information data to divide it into multiple short-time signal segments; S302. Perform Fourier transform on each short-time signal segment to extract the frequency distribution characteristics of each time period and obtain multiple sets of single-time period time-frequency characteristic data; S303. The multiple sets of single-time period time-frequency feature data are spliced ​​together to obtain the two-dimensional time-frequency feature map.

[0062] The effectiveness of STFT processing directly depends on the settings of its core parameters. Key STFT parameters include the sliding window length and window overlap rate. The selection of these two parameters directly affects the resolution, redundancy, and completeness of effective features in the time-frequency feature map: an excessively large window length reduces time-domain resolution, making it impossible to capture short-term signal mutations caused by human behavior; an excessively small window length results in insufficient frequency-domain resolution, making it difficult to distinguish frequency differences corresponding to different behaviors; an excessively high overlap rate increases feature redundancy and raises the computational complexity of the model; an excessively low overlap rate leads to discontinuous time-frequency features and loss of crucial timing information. Therefore, designing a scientific STFT parameter selection and optimization scheme and determining the optimal parameter combination are crucial steps in improving the quality of time-frequency features and ensuring the accuracy of subsequent model recognition, specifically for CSI human behavior recognition requirements.

[0063] This invention combines the characteristics of CSI signals and the temporal features of human behavior, while also considering time-frequency resolution and computational efficiency. First, it determines the reasonable range of values ​​for the STFT parameters. Then, it designs comparative experiments using the controlled variable method to screen the optimal parameter combination. The specific optimization process is divided into four stages: determining the parameter value range, single-parameter optimization, multi-parameter combination optimization, and optimal parameter verification. The detailed process is as follows: 1. Determination of Parameter Value Range: Combining the CSI signal sampling frequency (100Hz) and the time scale of human behavior, and strictly following the control group parameters actually set in this invention, the value range of the STFT core parameters and the specific control groups are determined. This invention sets up 20 STFT parameter control groups, all of which meet the principle of "overlap rate not exceeding 50% of window length" to ensure the continuity and effectiveness of time-frequency characteristics. The specific control groups are as follows: (2, 0), (2, 1), (4, 1), (4, 2), (8, 2), (8, 4), (10, 3), (10, 5), (12, 4), (12, 6), (16, 4), (16, 8), (20, 5), (20, 10), (24, 8), (24, 12), (32, 8), (32, 16), (48, 16), (48, 24), (64, 16), (64, 32). The window length (nperseg) covers 11 levels: 2, 4, 8, 10, 12, 16, 20, 24, 32, 48, and 64. The window overlap rate (noverlap) is set according to the corresponding window length, and none of them exceed 50% of the window length. At the same time, other auxiliary parameters of STFT are determined: the window function is selected as the Hanning window, which can effectively suppress spectral leakage and fit the non-stationary characteristics of the CSI signal; the number of Fourier transform points (nfft) is set to be consistent with the window length to balance frequency domain resolution and computational efficiency.

[0064] 2. Single-parameter optimization experiment: Using the controlled variable method, single-parameter optimization was performed on the window length (nperseg) and window overlap rate (noverlap) respectively. One parameter was kept constant while the other parameter was adjusted, and the rationality of the parameters was evaluated by the model recognition accuracy.

[0065] 1) Window length (nperseg) optimization: The fixed window overlap rate is 50% of the corresponding window length (i.e., nooverlap=nperseg / 2, if nperseg is odd, the nearest integer is taken). Based on the 20 control groups set in this invention, parameter pairs corresponding to different window lengths are selected (e.g., nperseg=2 corresponds to (2,0), (2,1); nperseg=4 corresponds to (4,1), (4,2), etc.). The LSTM-Transformer hybrid model is used as the evaluation model. The STFT time-frequency features under different window lengths are input into the model for training. The recognition accuracy of the test set is used as the evaluation index. Experimental results show that when nperseg ≤ 8, the model recognition accuracy significantly improves with increasing window length. This is because a small window length results in insufficient frequency domain resolution, making it impossible to effectively distinguish the frequency differences of different behaviors. When 8 ≤ nperseg ≤ 16, the model recognition accuracy tends to stabilize, at which point the time-frequency resolution reaches its optimal level, enabling simultaneous capture of temporal abrupt changes and frequency domain differences in human behavior. When nperseg > 16, the model recognition accuracy gradually decreases with increasing window length. This is because an excessively large window length reduces the time domain resolution, making it impossible to capture short-term signal abrupt changes caused by behavior, and increases feature redundancy, leading to model overfitting. Based on the actual control group settings of this invention, the optimal range for the window length (nperseg) is initially determined to be 8~16, corresponding to control groups of (8, 2), (8, 4), (10, 3), (10, 5), (12, 4), (12, 6), (16, 4), and (16, 8).

[0066] 2) Window overlap (noverlap) optimization: The fixed window length nperseg=12 (within the initial optimal range, corresponding to the control groups (12, 4) and (12, 6)) was supplemented with a comparison of reasonable overlap rates under the same window length (combined with the control group settings, two levels of overlap, 4 and 6, were selected for overlap). The test set accuracy of the LSTM-Transformer model was used as the evaluation index. Experimental results show that when overlap=4, there are slight breaks in the time-frequency features, and the feature continuity is generally poor; when overlap=6, the time-frequency features have good continuity, which avoids feature redundancy and ensures the integrity of time-series information, resulting in higher model recognition accuracy. Combining the overlap rate settings of all control groups in this invention (all not exceeding 50% of the corresponding window length), the optimal value of the window overlap (noverlap) needs to match the window length, which is initially determined to be 40%~50% of the corresponding window length. The optimal value range of overlap in the control group is 4~6 (adapting to nperseg=8, 10, 12).

[0067] 3. Multi-parameter combination optimization experiment: Within the optimal value range determined by single-parameter optimization, and combined with the 20 control groups actually set up in this invention, parameter combinations that meet the preliminary optimal range were selected, specifically (8, 2), (8, 4), (10, 3), (10, 5), (12, 4), (12, 6), (16, 4), and (16, 8), totaling 8 parameter combinations. For each parameter combination, the denoised and enhanced CSI signal was subjected to STFT transformation to generate a two-dimensional time-frequency feature map, which was input into the human daily activity recognition hybrid model for training and testing. The test set accuracy of each experiment was recorded, and the optimal parameter combination was selected using a comprehensive scoring method.

[0068] S4. Construct a hybrid model for recognizing daily human activities, and train the hybrid model for recognizing daily human activities using the two-dimensional time-frequency feature map to obtain a trained hybrid model for recognizing daily human activities. S5. Input the channel state information data to be identified into the human daily activity recognition hybrid model, and output the human daily activity recognition result. The output human daily activity recognition result is displayed in an intuitive form (such as text prompts, status indicators), and at the same time, information such as recognition time and recognition probability is recorded to meet the real-time requirements of practical applications; The method also includes a comprehensive evaluation of the recognition performance of the hybrid model for recognizing daily human activities, verifying the effectiveness and superiority of the model in recognizing daily human activities, and providing a basis for model optimization. Core evaluation metrics such as accuracy, recall, F1 score, and training time are used to quantitatively evaluate the recognition performance of the method. Comparative experiments (with traditional LSTM models and basic Transformer models) verify the superiority of the hybrid model and preprocessing workflow designed in this method. Furthermore, anti-interference tests and cross-scene tests are conducted to evaluate the robustness and generalization ability of the system. Problems identified during the evaluation are addressed, and parameters are optimized accordingly.

[0069] This method proposes a five-layer overall architecture based on "data acquisition - signal preprocessing - feature extraction - model recognition - result output". This architecture closely follows the processing flow of CSI signals and the application logic of deep learning models. Each step works in concert and progresses step by step. It not only solves the problems of insufficient preprocessing and model adaptability and weak feature discrimination in existing studies, but also effectively combines the accuracy and efficiency of human daily activity recognition, improving recognition accuracy, efficiency and generalization ability.

[0070] Example 2 This embodiment further explains the construction of the hybrid model for recognizing daily human activities proposed in the above embodiments. It utilizes the two-dimensional time-frequency feature map to train the hybrid model, resulting in a trained hybrid model for recognizing daily human activities. See also... Figure 6 The hybrid model for recognizing daily human activities includes a feature input module, an LSTM temporal feature extraction module, a Transformer global feature modeling module, a GMP feature filtering module, a feature fusion module, and a classification output module. The feature input module inputs the two-dimensional time-frequency feature map into the LSTM temporal feature extraction module. The LSTM temporal feature extraction module receives the two-dimensional time-frequency feature map and outputs low-dimensional local temporal feature vectors to the Transformer global feature modeling module and the GMP feature filtering module, respectively. The Transformer global feature modeling module receives the low-dimensional local temporal feature vectors and outputs a one-dimensional comprehensive feature vector. The GMP feature filtering module receives the low-dimensional local temporal feature vectors and outputs key pooling features. The feature fusion module fuses the one-dimensional comprehensive feature vectors and key pooling features and outputs fused features to the classification output module. The classification output module receives the fused features and outputs the human daily activity recognition result.

[0071] The input two-dimensional time-frequency feature map contains key time-frequency features corresponding to human behavior, but it still suffers from problems such as high feature dimensionality and insufficient mining of local temporal dependencies. Therefore, this invention designs an LSTM temporal feature extraction module to further process the input two-dimensional time-frequency feature map, mine local temporal dependency information in the features, and extract more discriminative local temporal features, providing high-quality feature support for subsequent global feature extraction or classification recognition in the model.

[0072] The core design goal of the LSTM temporal feature extraction module is to accurately capture the local temporal fluctuation patterns caused by human behavior in CSI signals based on the two-dimensional time-frequency feature map extracted by STFT, filter out redundant information in the features, and transform high-dimensional time-frequency features into low-dimensional, highly discriminative local temporal features. Simultaneously, it avoids the gradient vanishing problem of traditional RNNs, ensuring the effectiveness and stability of feature extraction. As the fundamental feature extraction unit of the deep learning recognition system of this invention, the LSTM temporal feature extraction module receives the STFT time-frequency feature output and connects to the subsequent Transformer global feature extraction module, forming a complete process of "preprocessing—local temporal feature extraction—global feature extraction / feature filtering—classification and recognition".

[0073] Combining the characteristics of CSI time-frequency features with the requirements of human behavior recognition, the structural design of the LSTM time-series feature extraction module follows the principles of "adaptability, efficiency, and scalability." (See [link to relevant documentation]). Figure 7The LSTM temporal feature extraction module includes a first input layer, an LSTM hidden layer, a dropout regularization layer, and a first output layer connected in sequence. Each layer works in concert to achieve accurate extraction of local temporal features. The specific structural design is as follows: The core function of the first input layer is to receive the two-dimensional time-frequency feature map and convert it into a time-series feature sequence that the LSTM model can process. The STFT parameters are (nperseg=2, noverlap=1), and the dimension of the time-frequency feature map output by the STFT is [time step count × frequency domain dimension]. The time step count is calculated from the number of CSI signal sampling points (400) and the STFT window parameters, and the frequency domain dimension is consistent with the STFT window length (nperseg=2). To adapt to the input requirements of the LSTM model, the first input layer converts the two-dimensional time-frequency feature map into a one-dimensional time-series feature sequence through a dimension reshaping operation. The input dimension is set to [batch size × time step count × frequency domain dimension] to ensure the temporal continuity of the feature sequence while maintaining compatibility with the input dimension of subsequent LSTM hidden layers.

[0074] The LSTM hidden layer, as the core feature extraction unit, is responsible for capturing local temporal dependencies in the temporal feature sequence and uncovering short-term fluctuation patterns in CSI signals caused by human behavior (such as periodic temporal changes in walking behavior and sudden amplitude changes in fall behavior). Considering the complexity of CSI temporal features and the model's generalization ability requirements, this invention adopts a multi-layer LSTM structure, specifically with two hidden layers. The number of neurons in each hidden layer (hidden layer dimension) was determined to be 128 after multiple experimental adjustments. This setting ensures that the model fully learns local temporal features, avoiding insufficient feature extraction due to excessively low hidden layer dimensions; it also avoids overfitting and a surge in computational complexity caused by excessively high dimensions, thus balancing feature extraction effectiveness and computational efficiency.

[0075] The gating mechanism of the LSTM hidden layers adapts to the CSI time-frequency feature extraction requirements: the forget gate adaptively forgets redundant information unrelated to human behavior (such as frequency domain features corresponding to residual slight noise) in the time-frequency features, while retaining key temporal changes caused by behavior; the input gate focuses on filtering out time-frequency features strongly correlated with human behavior categories (such as high-frequency peak features corresponding to falls) and updates them to the memory unit; the output gate controls the output of key features in the memory unit, ensuring that the output features can reflect local temporal dependencies. Simultaneously, to enhance the model's feature learning ability, dropout connections are used between adjacent LSTM hidden layers to further suppress overfitting.

[0076] To address the potential overfitting issue during LSTM model training and improve the module's generalization ability, a dropout regularization layer is added after the LSTM hidden layers. Considering the sample distribution characteristics of the CSI dataset (moderate sample size for each behavior category with minor sample differences), the dropout probability is set to 0.3, randomly discarding 30% of the hidden layer output features. This prevents the model from over-relying on any particular feature, ensuring the module can adapt to changes in CSI features across different data collection objects and behavioral scenarios, thus improving the stability of feature extraction.

[0077] The core function of the first output layer is to integrate the local temporal features extracted by the LSTM hidden layer and output a low-dimensional, high-discrimination local temporal feature vector, providing feature support for subsequent modules. The output layer uses a linear transformation layer to convert the output of the LSTM hidden layer (dimension [batch size × number of temporal steps × hidden layer dimension]) into a one-dimensional feature vector of [batch size × feature dimension], where the feature dimension is set to 64. This dimension setting preserves the key information of the local temporal features while effectively reducing the feature dimension, decreasing the computational burden on subsequent modules, and ensuring compatibility of the output features with the input dimensions of the subsequent Transformer global feature modeling module and GMP feature selection module.

[0078] The workflow of the LSTM temporal feature extraction module is as follows: The first input layer receives the two-dimensional time-frequency feature map, and converts the two-dimensional time-frequency feature map into a temporal feature sequence through dimensionality reshaping, which is then input into the LSTM hidden layer. The LSTM hidden layer captures local temporal dependencies through a gating mechanism and extracts key local features corresponding to human behavior in the temporal feature sequence. The dropout regularization layer receives the key local features, randomly discards the key local features to suppress overfitting, and outputs the first temporal optimized feature to the first output layer. The first output layer transforms the first temporal optimized feature into the low-dimensional local temporal feature vector through linear transformation.

[0079] Module performance verification shows that the LSTM temporal feature extraction module can effectively capture local temporal dependencies in CSI time-frequency features, and the extracted local temporal features have high discriminative power. Testing revealed that the feature vector output by this module, after being input, can achieve preliminary recognition of seven types of human behavior with an accuracy of 89.6%, which is 12.3% higher than the recognition accuracy of directly using STFT time-frequency features. This indicates that the module can effectively mine local temporal information in time-frequency features, filter out redundant features, and provide reliable feature support for high-precision recognition in subsequent models.

[0080] In summary, the LSTM temporal feature extraction module designed in this invention, based on the core principles of LSTM and combined with the time-frequency characteristics of CSI and the requirements of human behavior recognition, achieves accurate extraction of local temporal dependencies in STFT time-frequency features through reasonable structural design and parameter optimization. This module follows the preceding data preprocessing workflow and connects to subsequent global feature extraction or feature selection modules, thus improving the deep learning feature extraction system of this invention. The extracted local temporal features can effectively characterize the short-term signal change patterns of human behavior, laying a solid foundation for high-precision recognition in subsequent models.

[0081] After processing by the LSTM temporal feature extraction module, the local temporal features of the channel state information data have been effectively mined, resulting in a low-dimensional local temporal feature vector. However, the LSTM temporal feature extraction module is limited by the inherent characteristics of the loop structure, and can only capture local temporal dependencies in the feature sequence, making it difficult to mine long-distance correlations between local features at different times. In human behavior recognition tasks, the temporal features of CSI signals often have significant global correlations (e.g., there is a clear temporal correlation before and after the signal abrupt change in fall behavior, and the periodic features of walking behavior need to be judged by combining local features of multiple periods). Combining the core principle of the Transformer attention mechanism, its self-attention mechanism can get rid of the limitations of the loop structure, compute the dependencies between all positions in the long sequence in parallel, and efficiently capture global features. Therefore, this invention designs a Transformer global feature modeling module to further process the low-dimensional local temporal feature vector output by LSTM, mine the global correlation information in the features, and construct a comprehensive feature that combines local details and global correlations, providing more comprehensive feature support for subsequent classification and recognition.

[0082] The core design goal of the Transformer global feature modeling module is to accurately capture the global dependencies between local features at different times based on the low-dimensional local temporal feature vectors extracted by LSTM, integrate global temporal correlation information, and compensate for the shortcomings of the LSTM temporal feature extraction module in global feature capture. Simultaneously, it further optimizes feature quality, filters out redundant global information, and outputs a highly discriminative comprehensive feature vector, ensuring that the features can comprehensively represent the complete temporal change patterns of human behavior and improve the accuracy of subsequent classification and recognition. This module, as the core global feature extraction unit of the hybrid model for recognizing daily human activities in this invention, receives the local temporal feature output from LSTM, connects with the subsequent feature fusion module and classification output module, and forms a collaborative process with the LSTM temporal feature extraction module of "local temporal feature extraction—global feature modeling—classification and recognition".

[0083] The structural design of the Transformer global feature modeling module follows the principles of "coordination, efficiency, and accuracy." (See also...) Figure 8The Transformer global feature modeling module comprises six sequentially connected parts: a second input layer, a positional encoding layer, a multi-head self-attention layer, a residual connection and layer normalization layer, a FeedForward fully connected layer, and a second output layer. Each layer works in concert to achieve accurate modeling and optimization of global features. The specific structural design is as follows: The core function of the second input layer is to receive low-dimensional local temporal feature vectors and perform dimensional adaptation processing to ensure compatibility with the input requirements of the Transformer global feature modeling module. Based on the parameter settings of the LSTM temporal feature extraction module, its output low-dimensional local temporal feature vector has a dimension of 64. Since the Transformer attention mechanism requires a sequence of temporal features as input, the second input layer transforms the one-dimensional feature vector into a temporal feature sequence of [batch size × number of temporal steps × feature dimension] through a dimension reshaping operation. The number of temporal steps is consistent with the number of temporal steps processed by the LSTM temporal feature extraction module (calculated from 400 CSI signal sampling points and the optimal STFT parameters), ensuring the continuity of temporal information and matching the feature dimension with the input dimension of the subsequent multi-head self-attention layer.

[0084] Position Encoding Layer: Since the Transformer attention mechanism itself lacks temporal sequence perception, and the temporal sequence of CSI signals directly determines the type of human behavior (e.g., the temporal change order of CSI signals for sitting down and standing up is reversed), the lack of temporal position information will prevent the model from distinguishing local features at different times, thus affecting the mining of global related features. Therefore, a position encoding layer is added after the second input layer to add temporal position information to the input temporal feature sequence. This invention adopts a sine-cosine position encoding method, generating different position encoding vectors according to the temporal position of the features. These vectors are then superimposed with the low-dimensional local temporal feature vectors output by the LSTM, preserving the core information of the local temporal features while injecting temporal position information, laying the foundation for accurate mining of subsequent global related features. The dimension of the position encoding vector is consistent with the dimension of the input features (64 dimensions), ensuring that the feature dimension remains unchanged after superposition.

[0085] Multi-head self-attention layer: As the core global feature extraction unit of the Transformer global feature modeling module, it is responsible for mining the global dependencies between local features at different times in the temporal feature sequence and integrating global temporal correlation information. Considering the complexity of local temporal features of CSI and the computational efficiency requirements of the model, this invention adopts a multi-head self-attention mechanism. It calculates attention weights of different dimensions through multiple parallel attention heads, and then concatenates and fuses the outputs of each attention head. This can capture the global correlation of CSI features from multiple perspectives, improve the discriminative ability of features, and avoid the problem of incomplete feature capture caused by a single attention head.

[0086] The multi-head self-attention layer's specific workflow is adapted to the global feature modeling requirements of CSI: First, the positionally encoded feature sequence is mapped to a query vector (Q), a key vector (K), and a value vector (V) through three independent linear transformation layers, all with the same dimensionality as the input features (64 dimensions). Then, the dot product similarity between Q and K is calculated to obtain the initial attention weight matrix. The weights are normalized to the 0-1 range using the Softmax activation function and scaled by the square root of the Q dimension to avoid gradient instability caused by an excessively large weight matrix. Finally, the normalized attention weight matrix is ​​multiplied by the V vector to obtain the output of each attention head. The outputs of all attention heads are then concatenated and fused through the linear transformation layer to obtain the globally associated features. The number of attention heads was determined to be four after multiple experiments and adjustments, balancing feature capture capability and computational efficiency, enabling comprehensive global correlation mining without drastically increasing computational complexity.

[0087] Residual Connectivity and Layer Normalization: To alleviate the vanishing gradient problem during the training of the Transformer global feature modeling module and improve the model's training stability and generalization ability, a residual connectivity and layer normalization layer are added after the multi-head self-attention layer, using a combination of "residual connectivity + layer normalization". The residual connectivity superimposes the input and output of the multi-head self-attention layer, effectively propagating gradients and avoiding vanishing gradients. Layer normalization standardizes the superimposed features, adjusting the feature mean to 0 and the variance to 1, reducing the impact of feature distribution differences, accelerating model training convergence, suppressing overfitting, and adapting to the sample distribution characteristics of the CSI dataset (moderate sample size for each behavior category with a small amount of individual variation).

[0088] FeedForward Fully Connected Layer: To further optimize the expressive power of global correlation features, a FeedForward fully connected layer is set after the residual connection and layer normalization layer to perform nonlinear transformation and feature integration on the global correlation features output by the multi-head self-attention layer. This fully connected layer adopts a structure of "linear transformation + ReLU activation + linear transformation", where the number of neurons in the hidden layer is set to 256, and the number of neurons in the output layer is consistent with the dimension of the input features (64 dimensions). The ReLU activation function can introduce nonlinear factors to enhance the model's feature learning ability and mine complex correlation information in global features; the two linear transformations realize the adaptation and optimization of feature dimensions, ensuring the rationality of output features, while avoiding feature dimension redundancy.

[0089] The second output layer's core function is to integrate the globally correlated features after a series of processing steps, outputting a comprehensive feature vector that combines local details with global correlations. This provides high-quality feature support for subsequent feature fusion and classification output modules. The second output layer employs a linear transformation layer to convert the output of the FeedForward fully connected layer (dimension [batch size × time steps × feature dimension]) into a one-dimensional comprehensive feature vector of [batch size × feature dimension]. The feature dimension remains 64-dimensional, preserving key information from the globally correlated features while maintaining compatibility with the input dimensions of subsequent feature fusion and classification output modules, and avoiding computational burden caused by excessively high feature dimensions.

[0090] Based on the requirements of CSI human behavior recognition, the key parameters of the Transformer global feature modeling module were determined through multiple experiments and adjustments to ensure optimal module performance. The workflow of the Transformer global feature modeling module is as follows: The second input layer receives the low-dimensional local temporal feature vector and converts it into a temporal feature sequence through dimensionality reshaping, which is then input into the position encoding layer to ensure compatibility with the input requirements of the Transformer module. The position encoding layer adds positional information to the temporal feature sequence to address the issue that the Transformer cannot perceive temporal order. It outputs the temporal feature sequence with added positional information to the multi-head self-attention layer. The multi-head self-attention layer calculates the association weights of local features at different times using four parallel attention heads, mines global dependencies in the temporal feature sequence with added positional information, and outputs global association features to the residual connection and layer normalization layer. The residual connection and layer normalization layer standardizes the global association features, propagates gradients, and suppresses overfitting, outputting standardized features to the FeedForward fully connected layer. The FeedForward fully connected layer performs nonlinear transformation and integration on the standardized features to optimize feature representation and outputs a second temporal optimized feature to the second output layer. The second output layer transforms the second temporal optimized feature into the one-dimensional comprehensive feature vector through a linear transformation. This vector is then input into the classification output module through the subsequent feature fusion module for accurate recognition of human behavior. .

[0091] Performance verification shows that the Transformer global feature modeling module can effectively capture global correlation information from LSTM local temporal features, compensating for the limitations of the LSTM temporal feature extraction module. Testing revealed that after the comprehensive feature vector output by this module was fed into the classification output module through the feature fusion module, the recognition accuracy for seven types of human behavior reached 96.8%, a 7.2% improvement compared to the recognition accuracy using LSTM local features alone (89.6%). This demonstrates that the module effectively integrates global correlation information, enhancing the discriminative power of the features. Furthermore, the module exhibits fast training convergence and strong generalization ability, maintaining stable feature extraction performance on CSI samples from different collected objects, thus meeting the needs of CSI human behavior recognition.

[0092] The GMP feature selection module is a global max pooling layer. It selects key features of low-dimensional local temporal feature vectors through the global max pooling layer and outputs the key pooling features to the subsequent feature fusion module. The feature fusion module uses a weighted summation method to fuse the one-dimensional comprehensive feature vector and the key pooling features, and outputs the fused features to the classification output module; The classification output module includes a third fully connected layer and an activation function layer connected in sequence. The third fully connected layer performs dimensional mapping and feature integration on the fused features and outputs the mapping features adapted to the multi-behavior classification task to the activation function layer. The activation function layer performs probability normalization processing on the mapping features and outputs the recognition probability corresponding to each human behavior category. The human behavior category with the highest recognition probability is selected as the recognition result of the human daily activity.

[0093] The method for recognizing daily human activities based on channel state information data proposed in this invention has the following significance: First, this invention enriches the theoretical framework for CSI signal preprocessing and feature extraction. In existing research, CSI signals are susceptible to interference from environmental noise and multipath effects, leading to difficulties in feature extraction and low utilization of useful information. This invention combines wavelet denoising with STFT time-frequency transform to achieve integrated preprocessing of "noise filtering and feature enhancement," verifying the effectiveness of this combined method in improving CSI signal quality and mining effective features in the time-frequency domain. It provides new preprocessing ideas and theoretical references for similar CSI sensing research.

[0094] Secondly, the method of this invention optimizes the model design theory for CSI behavior recognition. Considering the temporal characteristics of CSI signals, this invention compares the recognition performance of LSTM and LSTM-Transformer hybrid models, and deeply analyzes the advantages and disadvantages of these two models in capturing CSI temporal dependencies and global features. This further improves the application theory of deep learning models in temporal signal recognition, providing a theoretical basis for the model design of subsequent CSI-related perception tasks (such as gesture recognition and fall detection).

[0095] Finally, this invention addresses the shortcomings of insufficient synergy between preprocessing, feature extraction, and model optimization in existing research. Current studies often focus solely on optimizing preprocessing methods or model structures, neglecting their synergistic compatibility. This invention experimentally verifies the compatibility between STFT features and LSTM-like models, clarifies the impact mechanism of different preprocessing methods on model recognition performance, and provides a theoretical framework for synergistic optimization of preprocessing and model in subsequent CSI behavior recognition research, thus promoting the theoretical development of CSI wireless sensing technology.

[0096] The research results of this invention have strong practical application value, can effectively solve the inherent defects of traditional human behavior recognition technology, and promote the application of contactless intelligent sensing technology in multiple fields, specifically in the following aspects: In the field of smart elderly care, the contactless CSI (Contactless Sensory Identity) behavior recognition solution proposed in this invention can achieve unobtrusive monitoring of the daily behaviors of the elderly (such as walking, sitting, falling, lying down, etc.) without requiring the elderly to wear any wearable devices, avoiding problems such as cumbersome operation and easy loss of the device. At the same time, this solution is unaffected by environmental factors such as light and obstruction, and can be deployed in private settings such as bedrooms and bathrooms. It can monitor the elderly's behavioral status in real time, promptly detect dangerous situations such as falls, and protect the elderly's privacy, providing convenient, efficient, and privacy-protecting technical support for smart elderly care.

[0097] In the smart home field, this recognition solution can be deeply integrated with smart home systems to achieve behavior-driven intelligent control. For example, by recognizing user behaviors such as "sitting down," "standing up," and "lying down," it can automatically adjust devices such as light brightness, air conditioning temperature, and curtain opening and closing, improving the intelligence level and user experience of smart homes. This drives the upgrade of smart homes from "passive control" to "active perception," aligning with the development trend of smart homes in the Internet of Things era.

[0098] In the field of public safety and health monitoring, this solution can be used for behavioral monitoring in public areas (such as abnormal behavior identification) and behavioral monitoring of hospital patients (such as postoperative rehabilitation behavior tracking). Its ease of deployment and lack of the need for extensive additional hardware reduce the deployment cost of monitoring systems while avoiding privacy issues associated with cameras. It has significant application value in ensuring public safety and improving the efficiency of health monitoring.

[0099] Furthermore, the research results of this invention can be extended to other wireless sensing scenarios, such as indoor positioning, gesture control, and intelligent security, providing technical support for the popularization of Internet of Things technology and the development of the field of intelligent sensing, and have broad application prospects and practical value.

[0100] In summary, the research of this invention can not only improve the theoretical system of CSI human behavior recognition and solve the key problems existing in the prior art, but also promote the practical application of contactless intelligent sensing technology, provide theoretical support and technical reference for applications in related fields, and has important theoretical value and practical significance.

[0101] Example 3 This embodiment further illustrates the hybrid model for recognizing daily human activities. The design of the LSTM temporal feature extraction module and the Transformer global feature modeling module in this hybrid model constructs the core feature extraction system of the LSTM-Transformer hybrid model. This model can effectively capture the local temporal dependencies and global correlation information of CSI signals. To further enrich feature representation and improve model recognition accuracy and generalization ability, this study introduces a GMP feature selection module. The design of the LSTM temporal feature extraction module and the GMP feature selection module constructs the core feature extraction system of the LSTM-GMP hybrid model. A scientific model fusion strategy integrates the advantages of the two models, compensating for the limitations of a single model. Simultaneously, an adaptive training strategy is designed to address the characteristics of the CSI human behavior recognition task, resolving issues such as overfitting, slow convergence, and gradient instability that may occur during model training, ensuring that the model achieves optimal performance on the CSI dataset. This invention further elaborates on the model fusion scheme, training parameter settings, training process control, and training effect verification, perfecting the entire construction process of the hybrid model for recognizing daily human activities and providing a solid foundation for subsequent model performance testing and analysis.

[0102] The core logic of model fusion is to achieve a recognition effect of "1+1>2" by integrating the feature advantages of different models. The key lies in balancing the characteristics of different models, rationally allocating the contribution weights of each model, and avoiding feature redundancy or offsetting advantages. The two basic models constructed in this invention are complementary: the LSTM-Transformer hybrid model focuses on mining the local temporal details and global correlations of CSI signals, providing comprehensive feature representation and accurately capturing the complete temporal patterns of human behavior, but with higher computational complexity; the LSTM-GMP hybrid model extracts local temporal features through LSTM and combines global max pooling (GMP) to filter key features, offering advantages such as high computational efficiency, strong anti-interference ability, and good suppression of redundant features, but it has shortcomings in capturing global correlation features. Based on this, a weighted fusion strategy at the feature level is designed, as follows: 1) Prerequisites for Fusion: To ensure the effectiveness and reliability of the fusion effect, the two basic models must meet the following prerequisites: First, the input data must be consistent, both being trained based on channel state information data after preprocessing (acquisition and cleaning, wavelet transform denoising and feature enhancement, STFT time-frequency transformation) to avoid fusion deviation caused by data differences; Second, the output feature dimensions must be unified, with the comprehensive feature vector output by the LSTM-Transformer model and the key feature vector output by the LSTM-GMP model both set to 64 dimensions to provide a dimensional basis for feature fusion; Third, the models must converge independently, with both basic models trained separately to achieve a stable convergence state, ensuring the effectiveness of their respective output features.

[0103] 2) Feature Fusion Method: A weighted summation method is used to fuse the output features of the two models. This method can flexibly adjust the contribution of the features of the two models and adapt to the differences in the importance of different features. The fusion formula is as follows:

[0104] In the formula, This is the final feature vector after fusion; This is the comprehensive feature vector output by the LSTM-Transformer model, covering both local temporal and global correlation features; The key feature vectors output by the LSTM-GMP model focus on core local features; For fusion weights ( This is used to adjust the contribution ratio of features from the two models. The larger the value, the higher the feature contribution of the LSTM-Transformer model; conversely, the smaller the value, the higher the feature contribution of the LSTM-GMP model.

[0105] 3) Post-processing after fusion: To further improve the generalization ability of the fusion model and suppress overfitting, a dropout regularization layer is added after the fusion layer. The dropout probability is set to 0.3, and 30% of the fusion features are randomly discarded to avoid the model from over-relying on the features of a certain model. Then the fusion features are input into the classification output module to complete the prediction of human behavior category and ensure that the fusion model can stably output recognition results.

[0106] Regarding the setting of training parameters, the rationality of the training parameters directly affects the convergence speed, recognition accuracy and generalization ability of the model. After multiple experiments and debugging, unified training parameters were determined to balance training efficiency and model performance.

[0107] To ensure the stability and effectiveness of model training, and addressing common issues in hybrid models for recognizing daily human activities, such as overfitting, gradient explosion, slow convergence, and training instability, a multi-stage training process control strategy is designed based on the characteristics of CSI data. This strategy monitors the training status throughout the entire process and adjusts training parameters in a timely manner. Specific control measures are as follows: The step of training the hybrid model for recognizing daily human activities using the two-dimensional time-frequency feature map includes: S41. Divide the two-dimensional time-frequency feature map into a training set and a validation set, and perform data augmentation on the training set to obtain an augmented training set. The data augmentation strategy is as follows: during training, perform slight data augmentation on the samples in the training set, using two methods: time-series shifting and amplitude fine-tuning. While retaining the core features of the samples, increase sample diversity and suppress overfitting. Specifically, the time-series shifting range is controlled within ±5 sampling points to avoid excessive shifting that could lead to distortion of behavioral features; the amplitude fine-tuning range is controlled within ±5% to simulate signal fluctuations under different acquisition environments and improve the model's anti-interference ability.

[0108] S42. Iteratively train the human daily activity recognition hybrid model with the enhanced training set input parameters initialized. Monitor the model gradient values ​​in real time, and optimize the training process by gradient clipping and dynamic adjustment of the learning rate to obtain the human daily activity recognition hybrid model after each round of training. During the training process, monitor the model gradient changes in real time and establish a gradient threshold judgment mechanism: if the gradient magnitude is greater than 10, it is judged as gradient explosion, and a gradient clipping strategy is adopted to clip the gradient magnitude to within 10 to avoid gradient explosion causing model training to collapse; if the gradient magnitude is less than 1e-6, it is judged as gradient vanishing, and the learning rate is appropriately increased (increased by 0.0001 each time) to ensure effective gradient propagation and accelerate model convergence.

[0109] S43. For the hybrid model of human daily activity recognition trained in each iteration, the performance is verified using the validation set. Training is terminated when the loss of the validation set fails to decrease for a preset number of iterations, the optimal model parameters are saved, and the trained hybrid model of human daily activity recognition is output. This step introduces an early stopping strategy: an early stopping mechanism is introduced, using the loss of the validation set as the core monitoring indicator, with patience=5. That is, when the loss of the validation set fails to decrease for 5 consecutive iterations, model training is immediately stopped, and the current optimal model parameters are saved. This strategy can effectively avoid overfitting caused by overtraining, while improving training efficiency and reducing unnecessary computational consumption. The Xavier initialization method is used to initialize all parameters of the two base models and the fusion model, making the parameter distribution uniform and avoiding training instability caused by initial parameters being too large or too small. This ensures that the model starts training from a reasonable initial state, accelerating the convergence speed.

[0110] After the models were trained, the training performance of the LSTM-Transformer hybrid model, the LSTM-GMP hybrid model, and the fusion model was validated from three core dimensions: convergence, recognition accuracy, and generalization ability, to ensure that the models achieved the design goals of this study. The specific validation results are as follows: 1) Convergence Validation: All three models reached stable convergence within 35 epochs, with training set loss and validation set loss stabilizing without significant fluctuations. The fusion model converged fastest, stabilizing around 30 epochs, with the validation set loss dropping below 0.08. The LSTM-Transformer and LSTM-GMP models stabilized at 32 and 31 epochs respectively, with validation set losses decreasing to 0.10 and 0.09. This indicates that the training strategy designed in this study is well-suited to the model structure, effectively accelerating model convergence and avoiding training instability issues.

[0111] 2) Accuracy Verification: Accuracy tests were conducted on the three models on the test set. The accuracy, precision, recall, and F1 score were statistically analyzed. The results showed that the LSTM-Transformer model achieved an accuracy of 96.8% and an F1 score of 96.7% on the test set; the LSTM-GMP model achieved an accuracy of 97.8% and an F1 score of 97.7% on the test set; and the fusion model achieved an accuracy of 98.2% and an F1 score of 98.1% on the test set. The fusion model outperformed the two base models in all accuracy metrics, indicating that the model fusion strategy effectively integrated the advantages of both models and significantly improved recognition accuracy.

[0112] 3.) Generalization Ability Validation: Two methods were used to validate the model's generalization ability: First, 10% of the test set samples (covering different collection objects and different behavior categories) were selected for repeated testing, and the accuracy fluctuation was statistically analyzed; second, two unfamiliar collection objects who did not participate in the training were introduced, and CSI samples of their seven behavior categories were collected for cross-object testing. The results show that the accuracy fluctuation of the fusion model in repeated testing is less than 0.5%, and the recognition accuracy on unfamiliar collection object samples reaches 95.3%, which is significantly higher than the two basic models (LSTM-Transformer at 92.1% and LSTM-GMP at 93.5%). This indicates that the fusion model has strong generalization ability, can adapt to CSI samples of different collection objects and environments, and meets the actual needs of human behavior recognition.

[0113] In summary, the model fusion strategy of this invention effectively integrates the advantages of both LSTM-Transformer and LSTM-GMP models, improving the model's recognition accuracy and generalization ability. Scientifically sound training parameters and process control strategies ensure stable model convergence and resolve issues such as overfitting and gradient instability during training. The training effects of both basic models and the fused model all meet the design goals, providing reliable model support for performance testing and comparative analysis.

[0114] In terms of model design, this invention uses LSTM as the basic temporal feature extraction unit, combined with global mean pooling (GMP) and Transformer attention mechanisms to construct three model structures: LSTM, LSTM-GMP, and LSTM-Transformer. Experimental results show that the basic LSTM model has weak ability to distinguish subtle movements, with an overall accuracy of only 0.6618. After introducing GMP, the model's stability and feature compactness are significantly improved, and the accuracy increases to 0.8676. The LSTM-Transformer, with its self-attention mechanism, achieves global feature association modeling, reaching 100% recognition accuracy in five types of movements: walking, running, sitting, standing, and bending over, with an overall accuracy as high as 0.9706. Its comprehensive performance is significantly better than the first two models.

[0115] Through multiple sets of comparative experiments and visualization analysis, this invention completed the entire process verification from preprocessing scheme selection, optimal parameter determination, model structure comparison to optimal model selection, forming a stable, efficient, and directly implementable CSI contactless behavior recognition solution. Overall research shows that the architecture based on the STFT time-frequency features + LSTM-Transformer hybrid model can fully exploit the temporal and global correlation features of CSI signals, effectively improving the recognition accuracy and robustness of multiple categories of daily behaviors.

[0116] Obviously, the above embodiments of this application are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of this application. Those skilled in the art can make other variations or modifications based on the above applications. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of the claims of this application.

Claims

1. A method for recognizing daily human activities based on channel state information data, characterized in that, include: S1. Collect channel state information data generated by human behavior, clean the collected channel state information data, and obtain cleaned channel state information data. S2. Perform wavelet transform denoising and feature enhancement preprocessing on the cleaned channel state information data to obtain preprocessed channel state information data; S3. Extract time-frequency features from the preprocessed channel state information data to obtain a two-dimensional time-frequency feature map; S4. Construct a hybrid model for recognizing daily human activities, and train the hybrid model for recognizing daily human activities using the two-dimensional time-frequency feature map to obtain a trained hybrid model for recognizing daily human activities. S5. Input the channel state information data to be identified into the human daily activity recognition hybrid model, and output the human daily activity recognition result.

2. The method for recognizing daily human activities based on channel state information data according to claim 1, characterized in that, The channel state information data generated by human behavior includes: S101. Set up a data acquisition device, which includes a transmitter and a receiver; S102. Under the selected acquisition environment, the transmitting end transmits a wireless signal, which is transmitted to the receiving end after being reflected and scattered by the human body; S103. The receiving end analyzes the wireless signal after it has been reflected and scattered by the human body to obtain the channel state information data.

3. The method for recognizing daily human activities based on channel state information data according to claim 1, characterized in that, The data cleaning process for the collected channel state information data includes: S111. Standardize the format of the collected channel state information data to obtain channel state information data with a regular format; S112. Perform preliminary noise filtering on the well-formatted channel state information data to obtain smoothed data after noise reduction; S113. Perform outlier detection and correction on the smoothed data to obtain corrected data; S114. Perform data length normalization processing on the corrected data to obtain dimension-standardized data; S115. Perform data redundancy removal and verification on the dimensional standardized data to obtain cleaned channel state information data.

4. The method for recognizing daily human activities based on channel state information data according to claim 1, characterized in that, The preprocessing of the cleaned channel state information data, including wavelet transform denoising and feature enhancement, includes: S201. Perform phase offset correction on the cleaned channel state information data to obtain phase offset corrected channel state information data; S202. Perform several levels of discrete wavelet decomposition on the channel state information data after phase offset correction to separate low-frequency approximate components and high-frequency detail components. S203. Perform noise reduction processing on the high-frequency detail components, remove high-frequency random noise components, and retain the human behavior detail features in the high-frequency components to obtain high-frequency detail noise-reduced components. S204. Perform inverse wavelet transform on the high-frequency detail noise reduction component and the low-frequency approximation component to reconstruct the denoised reconstructed signal; S205. Normalize the amplitude of the reconstructed signal to obtain normalized data; S206. Perform time-series detail enhancement on the normalized data to amplify the signal mutations and fluctuations corresponding to various human behaviors, and obtain the preprocessed channel state information data.

5. The method for recognizing daily human activities based on channel state information data according to claim 1, characterized in that, The step of extracting time-frequency features from the preprocessed channel state information data to obtain a two-dimensional time-frequency feature map includes: S301. Perform sliding segmentation on the preprocessed channel state information data to divide it into multiple short-time signal segments; S302. Perform Fourier transform on each short-time signal segment to extract the frequency distribution characteristics of each time period and obtain multiple sets of single-time period time-frequency characteristic data; S303. The multiple sets of single-time period time-frequency feature data are spliced ​​together to obtain the two-dimensional time-frequency feature map.

6. The method for recognizing daily human activities based on channel state information data according to claim 1, characterized in that, The hybrid model for recognizing daily human activities includes a feature input module, an LSTM temporal feature extraction module, a Transformer global feature modeling module, a GMP feature filtering module, a feature fusion module, and a classification output module. The feature input module inputs the two-dimensional time-frequency feature map into the LSTM temporal feature extraction module. The LSTM temporal feature extraction module receives the two-dimensional time-frequency feature map and outputs low-dimensional local temporal feature vectors to the Transformer global feature modeling module and the GMP feature filtering module, respectively. The Transformer global feature modeling module receives the low-dimensional local temporal feature vectors and outputs a one-dimensional comprehensive feature vector. The GMP feature filtering module receives the low-dimensional local temporal feature vectors and outputs key pooling features. The feature fusion module fuses the one-dimensional comprehensive feature vectors and key pooling features and outputs fused features to the classification output module. The classification output module receives the fused features and outputs the human daily activity recognition result.

7. The method for recognizing daily human activities based on channel state information data according to claim 6, characterized in that, The LSTM temporal feature extraction module includes a first input layer, an LSTM hidden layer, a dropout regularization layer, and a first output layer connected in sequence. The first input layer receives the two-dimensional time-frequency feature map and converts it into a temporal feature sequence through dimensionality reshaping, which is then input into the LSTM hidden layer. The LSTM hidden layer extracts key local features corresponding to human behavior in the temporal feature sequence. The dropout regularization layer receives the key local features, randomly discards them, and outputs a first temporal optimized feature to the first output layer. The first output layer converts the first temporal optimized feature into the low-dimensional local temporal feature vector.

8. The method for recognizing daily human activities based on channel state information data according to claim 7, characterized in that, The Transformer global feature modeling module includes a second input layer, a position encoding layer, a multi-head self-attention layer, a residual connection and layer normalization layer, a FeedForward fully connected layer, and a second output layer connected in sequence. The second input layer receives the low-dimensional local temporal feature vector and converts it into a temporal feature sequence through dimensionality reshaping, which is then input into the position encoding layer. The position encoding layer adds positional information to the temporal feature sequence and outputs the temporal feature sequence with added positional information to the multi-head self-attention layer. The multi-head self-attention layer mines global dependencies in the temporal feature sequence with added positional information and outputs global correlation features to the residual connection and layer normalization layer. The residual connection and layer normalization layer standardizes the global correlation features and outputs standardized features to the FeedForward fully connected layer. The FeedForward fully connected layer performs nonlinear transformation and integration on the standardized features and outputs a second temporal optimized feature to the second output layer. The second output layer converts the second temporal optimized feature into the one-dimensional comprehensive feature vector.

9. The method for recognizing daily human activities based on channel state information data according to claim 8, characterized in that, The classification output module includes a third fully connected layer and an activation function layer connected in sequence. The third fully connected layer performs dimensional mapping and feature integration on the fused features and outputs the mapping features adapted to the multi-behavior classification task to the activation function layer. The activation function layer performs probability normalization processing on the mapping features and outputs the recognition probability corresponding to each human behavior category. The human behavior category with the highest recognition probability is selected as the recognition result of the human daily activity.

10. The method for identifying daily human activities based on channel state information data according to any one of claims 6-9, characterized in that, The step of training the hybrid model for recognizing daily human activities using the two-dimensional time-frequency feature map includes: S41. Divide the two-dimensional time-frequency feature map into a training set and a validation set, and perform data augmentation on the training set to obtain an augmented training set; S42. Iteratively train the human daily activity recognition hybrid model after the enhanced training set input parameters are initialized, monitor the model gradient value in real time, and optimize the training process by gradient clipping and dynamic adjustment of the learning rate to obtain the human daily activity recognition hybrid model after each round of iterative training. S43. For the hybrid model for recognizing human daily activities after each round of iteration training, the performance is verified using the validation set. When the loss of the validation set does not decrease continuously in a preset number of rounds, training is terminated, the optimal model parameters are saved, and the trained hybrid model for recognizing human daily activities is output.

Images