Wireless signal action recognition method based on fragment enhancement

Abstract

The invention belongs to the technical field of wireless signal behavior perception, and discloses a wireless signal action recognition method based on fragment enhancement, which comprises the following steps of: acquiring a wireless radio frequency signal, preprocessing the wireless radio frequency signal, and converting the wireless radio frequency signal into time sequence matrix sequence data; constructing a wireless signal action recognition model based on fragment enhancement, extracting global context features and local context features based on attention guidance from time sequence matrix sequence data, and fusing the global context features and the local context features; and training the wireless signal action recognition model based on fragment enhancement, recognizing to-be-recognized sample data by using the trained wireless signal action recognition model, and outputting a prediction result of an action category. According to the method, feature expression of key action segments can be adaptively concerned and enhanced, and effective modeling is carried out on internal dynamic and external contexts of the segments, so that the capability of distinguishing similar complex actions is improved.

Description

Technical Field

[0001] This invention belongs to the field of wireless signal behavior perception technology, specifically relating to a method for wireless signal action recognition based on segment enhancement. Background Technology

[0002] With the rapid development of the Internet of Things (IoT) and ubiquitous computing technologies, non-contact human behavior sensing using wireless radio frequency signals such as Wi-Fi channel state information and millimeter-wave radar signals has become an important research direction. This technology analyzes the disturbances caused by human movements to the propagation path of wireless signals, extracts features that can characterize specific actions, and then completes the recognition task, showing broad application prospects in fields such as smart homes, health monitoring, and human-computer interaction.

[0003] However, existing wireless signal action recognition technologies still have significant shortcomings in practical applications. On the one hand, when dealing with different actions with similar local motion patterns, existing technologies face the core challenge of insufficient recognition accuracy. Traditional feature extraction methods typically process the entire signal time series uniformly, failing to accurately capture key time segments that determine the action category and possess strong discriminative power, and also failing to deeply model the micro-dynamic features within key segments and the temporal contextual relationships between key segments. This technical limitation of lacking a hierarchical and refined understanding of the action temporal structure directly leads to insufficient recognition of the extracted action features. When faced with similar actions, the model is prone to misclassifying common sub-patterns of one type of action as another complete action, significantly reducing the accuracy and robustness of action recognition. Even though some methods based on convolutional neural networks and recurrent neural networks can achieve temporal feature extraction, they still suffer from limited receptive fields and insufficient long-distance dependency modeling, making it difficult to achieve comprehensive and hierarchical feature learning of the global temporal structure and local key details of actions, and thus failing to effectively solve the problem of recognizing similar actions.

[0004] On the other hand, wireless signal data action recognition scenarios commonly suffer from data imbalance and small sample sizes. Samples of abnormal actions such as falls and specific gestures are often difficult to collect, resulting in a severe class imbalance in the action recognition dataset, which seriously restricts the generalization ability of the recognition model. Existing solutions to this problem mostly rely on generative adversarial networks (GANs) to generate virtual samples at the original signal level to expand the dataset. However, this approach also has significant limitations: First, signal generation distortion: wireless radio frequency signals have complex physical characteristics, and virtual data generated directly in the original signal space often lacks real physical semantic features, making it difficult to effectively complement real samples and thus having a very limited effect on improving model training performance. Second, high computational cost: the training process of GANs consumes a lot of computing resources, resulting in low model inference efficiency. This makes it unsuitable for the real-time and lightweight deployment requirements of edge computing devices, hindering its practical application. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a wireless signal action recognition method based on fragment enhancement. This method adaptively focuses on and enhances the feature representation of key action fragments through an attention-guided mechanism, and effectively models the internal dynamics and external context of these fragments. This improves the ability to distinguish similar complex actions and significantly enhances the recognition accuracy and robustness of the model in data-sparse scenarios.

[0006] The present invention provides a method for wireless signal action recognition based on fragment enhancement, comprising the following steps: Step 1: Acquire the wireless radio frequency signal and preprocess it to convert the wireless radio frequency signal into time-series matrix sequence data; Step 2: Construct a wireless signal action recognition model based on fragment enhancement, process the time-series matrix sequence data, extract global context features and attention-guided local context features, and fuse them to obtain deep fusion features; Step 3: Train the wireless signal action recognition model based on fragment enhancement, calculate the loss between the prediction result and the real label based on the deep fusion features, update the parameters of the wireless signal action recognition model, and obtain the trained wireless signal action recognition model. Step 4: Use the trained wireless signal action recognition model to identify the sample data to be identified and output the prediction results of the action category.

[0007] Furthermore, the fragment-enhanced wireless signal action recognition model includes a feature extraction network and a classifier; the feature extraction network is used to extract global and local context features and fuse them into deep fusion features to obtain the final sample feature representation; the classifier is used to map the features into action category probabilities. The feature extraction network includes a convolutional neural network encoder, a salient sequence extraction module, and a hierarchical local context modeling module; The convolutional neural network editor extracts frame-level features from the temporal matrix sequence data in chronological order. The salient sequence extraction module extracts global contextual features from frame-level features using a global Transformer model and generates attention guidance information; based on the attention guidance information, key action segments are adaptively selected. The hierarchical local context modeling module is used to perform hierarchical local context modeling on key action segments and extract local context features. By fusing global and local contextual features, the final sample feature representation is obtained.

[0008] Furthermore, based on the salient sequence extraction module, global contextual features are extracted from frame-level features using a global Transformer model, and attention-guided information is generated; based on the attention-guided information, key action segments are adaptively selected; including: 1) Frame-level features Projected onto the hidden layer dimension d of the Transformer model and with standard sinusoidal position encoding Combination: , in, and Projection parameters; sequence go through The global Transformer encoder layer is used for processing; each layer applies multi-head self-attention and position-wise feedforward networks, and uses residual connections and layer normalization; the output feature sequence matrix of the last Transformer layer is... T represents the total number of time steps of the input timing data; By using average pooling over the time dimension and a subsequent final projection layer, Aggregate into global embedding : , , in, Embed dimensions for the target; This represents the aggregated global feature vector obtained after average pooling along the time dimension; t represents the time step index. This represents a vector consisting of all the feature values ​​of the global Transformer encoder output sequence at time step t. Indicates projection layer operation; 2) Calculate the global attention score using the following formula: , , in, For attention weights, It quantifies how much attention is given to key position j in the query position i in the header h; 3) Based on From frame features Dynamic recognition A significant time segment; the significant time segment refers to a time segment with a global attention score. In the example, the feature sequence is extracted centered on the K time steps with the highest numerical ranking.

[0009] Furthermore, based on the hierarchical local context modeling module, local context features are extracted from key action segments, including: 1) For each segment composed of frame feature sequences, an intra-segment Transformer encoder is used. This is used to capture the microscopic dynamics within the fragment; then, the output is aggregated through time-averaged pooling to produce fragment-level embeddings. : , in, and Let each represent the nth segment and its corresponding embedding vector; this results in a set of unordered segment representations. ; 2) Convert the unordered set (i.e., the unordered fragment representation) into a time-ordered sequence matrix Z: ; 3) Input Z into the inter-fragment Transformer encoder In this process, the final local context embedding is generated. : .

[0010] Furthermore, by fusing global and local contextual features, the final sample feature representation is obtained, including: Complementary global representation and detailed local representation splicing together to form a comprehensive, integrated, and embedded system. : , Here, Concat is the concatenation operation; the fused embedding is input into the final linear classification layer to produce category logits: , in and These are classification parameters.

[0011] The beneficial effects of this invention are as follows: The method described in this invention, through its designed attention-guided mechanism, can adaptively locate the most discriminative key time segments from wireless signal sequences; and through a hierarchical context modeling network, it deeply mines and understands the micro-dynamic features within these key segments and the temporal correlation information between segments; compared with the prior art, this mechanism allows this invention to no longer rely on a fixed sliding window, but to adaptively select the most discriminative salient segments through the attention-guided mechanism, generating highly condensed and discriminative feature representations, which is particularly effective in distinguishing complex actions with similar local patterns, thereby greatly improving the recognition accuracy and robustness of the model in data-sparse scenarios, and solving the class imbalance problem in action recognition in a low-cost and high-efficiency manner. Attached Figure Description

[0012] Figure 1 This is a flowchart of the method described in this invention; Figure 2 This is a schematic diagram of the structure of a wireless signal feature extraction model based on fragment enhancement; Figure 3 This is a comparison of recall rates for WIFI datasets; Figure 4 This is a schematic diagram illustrating the performance of a millimeter-wave radar dataset. Detailed Implementation

[0013] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0014] like Figure 1 As shown, the wireless signal action recognition method based on fragment enhancement according to the present invention includes the following steps: Step 1: Acquire the wireless radio frequency signal and preprocess it to convert the wireless radio frequency signal into time-series matrix sequence data; Step 2: Construct a wireless signal action recognition model based on fragment enhancement, process the time-series matrix sequence data, extract global context features and attention-guided local context features, and fuse them to obtain deep fusion features; Step 3: Train the segment-enhanced wireless signal action recognition model using training sample data, calculate the loss between the prediction result and the real label based on the deep fusion features, update the parameters of the wireless signal action recognition model, and obtain the trained wireless signal action recognition model. Step 4: Use the trained wireless signal action recognition model to identify the sample data to be identified and output the prediction results of the action category.

[0015] For Wi-Fi signals, when the radio frequency signal is Wi-Fi channel state information, the preprocessing aims to eliminate environmental noise and extract the motion-related body-coordinate velocity profile (BVP). In this embodiment, step 1, the preprocessing of the acquired radio frequency signal, includes: 1) Signal cleaning and denoising: First, the received original signal is subjected to conjugate multiplication to eliminate phase shift, and principal component analysis is used to remove the static component caused by static environmental reflection, while retaining the dynamic component caused by human movement. 2) Time-frequency analysis: Perform short-time Fourier transform on the denoised signal to obtain the Doppler frequency shift spectrum, which reflects the radial velocity of different parts of the human body at different times; 3) Coordinate transformation and BVP generation: Since the Doppler frequency shift only reflects radial velocity, in order to more intuitively represent the action, the Doppler spectrum is mapped to the body coordinate system through a coordinate transformation algorithm to generate the BVP matrix; 4) Data normalization: The generated BVP matrix sequence is used as the input X of the model of this invention, and its dimension corresponds to T×H×W of the time series matrix sequence data, where T, H and W represent the sequence length, height and width, respectively.

[0016] For millimeter-wave radar signals, when the radio frequency signal is the raw data from the millimeter-wave radar, the preprocessing aims to generate a dynamic range-angle image (DRAI) through multi-dimensional frequency domain transformation and noise filtering; the specific steps include: 1) 3D-FFT processing: First, the original signal acquired by the radar is subjected to three-dimensional fast Fourier transform in the range dimension, Doppler dimension and angle dimension in sequence to extract the target's range, velocity and angle of arrival information respectively, thereby constructing a three-dimensional feature matrix containing range-Doppler-angle information; 2) Noise cancellation: In order to preserve motion-related dynamic signals and remove environmental interference, Doppler characteristics are used for filtering. First, components with Doppler frequencies of zero or below a certain velocity threshold are set to zero to remove reflections from stationary objects such as walls (static clutter). Second, by calculating the signal intensity distribution in the Doppler dimension and setting an adaptive energy threshold, low-intensity multipath reflection noise is filtered out, and only high-intensity dynamic components generated by human motion are retained. 3) DRAI generation: The distance-Doppler-angle 3D matrix after the above noise reduction processing is summed along the Doppler dimension, thereby projecting the dynamic signal energy onto the two-dimensional distance-angle plane to generate a single frame of dynamic distance-angle image; 4) Sequence construction and normalization: Repeat the above steps for consecutive time frames to form a continuous DRAI image sequence matrix data, and obtain a uniform size T×H×W.

[0017] like Figure 2 As shown, the fragment-enhanced wireless signal feature extraction model includes a convolutional neural network encoder, a salient sequence extraction module, and a hierarchical local context modeling module. The convolutional neural network editor extracts frame-level features from temporal matrix sequence data in chronological order; The salient sequence extraction module extracts global contextual features from frame-level features using a global Transformer model and generates attention guidance information; based on the attention guidance information, key action segments are adaptively selected. The hierarchical local context modeling module includes an intra-segment feature encoder (intra-segment Transformer encoder) and an inter-segment relationship modeling unit (inter-segment Transformer encoder), which is used to perform a processing flow that includes intra-segment feature encoding and inter-segment relationship modeling arranged in chronological order, and extract local context features hierarchically for key action segments. By fusing global and local contextual features, the final sample feature representation is obtained.

[0018] In the embodiments of the present invention, the global Transformer model, the intra-segment Transformer encoder, and the inter-segment Transformer encoder all adopt the standard Transformer structure; the only difference is that the global Transformer model and the intra-segment Transformer encoder are both 2-layer and have 2 heads; while the inter-segment Transformer encoder has 1 layer and 1 head.

[0019] In this embodiment, each wireless signal is converted into a 3D matrix sequence; for example, channel state information and millimeter-wave radar data for each frame.

[0020] Given an input sequence Where T, H, and W represent the sequence length, height, and width, respectively, a convolutional neural network encoder processes each frame independently using shared weights. This encoder typically consists of a standard convolutional block followed by a projection layer to extract spatial features. This produces a sequence of frame-level feature vectors: , Where D represents the dimension of the frame-level feature.

[0021] The salient sequence extraction module consists of two parts: attention score calculation and attention-guided segment selection. Specifically, it first utilizes a global Transformer to analyze the overall time series structure. Modeling is performed to generate an overall feature representation. The self-attention weights are aggregated into attention scores; then, sequence segments with adaptive lengths are selected based on these attention scores.

[0022] 1) Global feature generation; Frame-level features Projected onto the hidden layer dimension d of the Transformer model and with standard sinusoidal position encoding Combination: , in, and For projection parameters; subsequently, the sequence go through The layer-wise global Transformer encoder layer processing is used; each layer applies a multi-head self-attention and position-wise feedforward network, and uses residual connections and layer normalization; the output of the last layer is... ; By using average pooling over the time dimension and a subsequent final projection layer, Aggregate into global embedding : , , in, Embed dimensions for the target.

[0023] 2) Attention score calculation; To guide local analysis towards significant time regions, attention information derived from the global context modeling process is extracted. Specifically, attention weights from the final multi-head self-attention module are utilized. This weight reflects the most fully processed and context-rich representation of the sequence. Within multi-head self-attention, This quantifies how much "attention" is given to key position j at query position i in the header h. Aggregating the attention from all queries to a specific key can filter out noise and more consistently identify the truly critical time steps.

[0024] To determine the overall importance of time step j, the total attention from all other time steps across all heads is calculated. This reflects the collective attention the model pays to that step when constructing the global context. It is achieved by summing over query dimension i and averaging over attention head h, as described in the formula: ; Higher weights indicate that the corresponding time step contributes more significantly to the entire sequence; in other words, time steps with high attention scores have structural importance to the overall characteristics of the sequence. The global attention score is formulated as follows: , This data-driven guidance mechanism enables the model to intelligently focus on subsequent local analysis.

[0025] 3) Selection of segments for attention guidance; based on This module extracts frame features Dynamic recognition The selection process focuses on discriminative regions while ensuring pattern diversity. The process, as shown in the figure, includes candidate point selection, category-adaptive segment generation, and overlap constraint elimination.

[0026] Specifically, in order to smooth the original attention score, firstly... Apply the Exponential Moving Average (EMA) function to obtain the smoothed score vector. Then, candidate time steps are identified by selecting the index corresponding to the highest Top-K score, where, The aim is to provide a sufficiently large candidate pool; the formula is as follows: , in, yes A set of indices for the Top-K values ​​is used; fragments can be further generated centered around these points. To accommodate different action durations, shorter time windows are adaptively assigned to transient gestures, and longer time windows to sustained movements. Specifically, the average sequence length is calculated based on the training set for each category. For test samples during inference, the fragment length for a particular category depends on the pseudo-label.

[0027] Suppose a certain sample belongs to the first... Class, candidate point fragments It can be represented as: , in, It is the first The fragment length of the class. To avoid information redundancy and improve the diversity of local patterns, an overlap constraint elimination strategy is further adopted to remove fragments with high overlap rates. Formally, the overlap rate (IoU) is defined as follows: , in and There are two candidate segments. If If the threshold is exceeded, the next part is removed. This process continues until a threshold is found. The fragment pool or candidate pool is exhausted. If fewer than [number] fragments are found through this bootstrapping process... If there are 100 patches, then a uniform sampling strategy is adopted to generate more patches.

[0028] Furthermore, hierarchical local context modeling is performed on key action segments to extract local context features.

[0029] While independent feature extraction of salient segments can characterize action sub-patterns, it neglects temporal context information; therefore, different actions sharing common segments are easily confused. To address this issue, a hierarchical local context modeling module is used to perform a processing flow that includes intra-segment feature encoding and inter-segment relationship modeling arranged in temporal order.

[0030] In the first stage, the aim is to extract each variable-length salient segment into a compact, fixed-dimensional representation. For each segment composed of a sequence of frame features, an intra-segment Transformer encoder is employed. This is used to capture the microscopic dynamics within the fragment; then, the output is aggregated through time-averaged pooling to produce fragment-level embeddings. : , in, and Let each represent the nth segment and its corresponding embedding vector; this results in a set of unordered segment representations. .

[0031] Processing fragment embeddings solely according to selection order (e.g., based on attention scores) ignores their inherent temporal structure. To preserve this crucial contextual flow, a key temporal ordering step is introduced, transforming the unordered set into a temporally ordered sequence matrix Z that reflects the true temporal progression of salient events: ; Then, Z is input into the inter-fragment Transformer encoder. In this model, high-level dependencies between these key local patterns are modeled. The encoder operates on the sequence of fragment representations to produce the final local context embeddings. : .

[0032] By independently processing salient local patterns and then modeling their temporal relationships, this hierarchical approach effectively distinguishes sequences that share similar sub-patterns based on contextual information. This directly enhances the model's ability to disambiguate and identify complex and easily confused minority-sample categorical actions.

[0033] Furthermore, by fusing global and local contextual features, the final sample feature representation is obtained, including: Global branches capture the overall essence of a sequence, while local context modeling can explore subtle and discriminative patterns crucial for distinguishing fine-grained actions. Therefore, complementary global representations... and detailed local representation They are pieced together to form a comprehensive, integrated embedding. : , Here, Concat is the concatenation operation; the fused embedding is input into the final linear classification layer to produce category logits: , in and These are classification parameters.

[0034] The method of this invention is implemented in the deep learning framework PyTorch, and the model is trained on a server equipped with an NVIDIA RTX GPU. Inference latency tests were conducted on a Raspberry Pi 4B edge computing platform. Experimental data comes from two publicly available wireless radio frequency datasets: the WiFi CSI dataset Widar3.0 and the millimeter-wave radar dataset MCD-Gesture, covering various gesture and action categories. To verify the effectiveness of this invention under different data scarcity levels, various imbalanced test scenarios were designed, and multi-dimensional metrics such as accuracy, recall, F1 score, and geometric mean were used for evaluation, comprehensively verifying the performance of this invention in feature extraction and key fragment enhancement classification tasks.

[0035] like Figure 3As shown in the figure, this diagram visually compares the recall rates of the proposed method with those of a traditional baseline model (CNN+Transformer) across 22 different action categories in the WIFI dataset. Overall, the proposed model outperforms the baseline model in the vast majority of action categories, demonstrating the robustness of the proposed method. In contrast, the proposed method adaptively extracts key segments through an attention-guided mechanism and performs hierarchical modeling of the local context, successfully capturing subtle differences among similar actions and improving the recognition rate of these difficult-to-classify actions to over 60%. This experimental result confirms that the segment enhancement-based method proposed in this invention can effectively solve the problem of confusion between similar actions and improve the model's ability to recognize complex and subtle action features.

[0036] like Figure 4 As shown in the figure, this figure illustrates the changing trends of various key performance indicators of this method under different data scarcity ratios (gradually increasing from 40% to 95%) on a millimeter-wave radar dataset. Figure 4 The four curves correspond to precision, recall, F1 score, and geometric mean, respectively. From... Figure 4 The overall trend of the data shows that as data scarcity increases, i.e., the number of training samples decreases significantly, various performance indicators, as expected, exhibit a certain downward trend, but the decline is relatively gradual. It is worth noting that even under the extremely stringent condition of a data scarcity of up to 95%, this method still maintains a high performance level, with all core indicators remaining above 85%. This experimental result fully demonstrates the robustness of this method in handling small sample sizes and class imbalance problems. Furthermore, combined with the experimental conclusions on the aforementioned WiFi dataset, this result further verifies that this method possesses good cross-modal generalization ability, applicable not only to WiFi signal data but also effectively to millimeter-wave radar wireless sensing data, confirming the versatility of this method in practical multimodal wireless sensing scenarios.

[0037] Table 1 Method Deployment Evaluation

[0038] Table 1 shows the evaluation data of this method in terms of model complexity and performance on actual edge devices. Experiments were conducted on a typical resource-constrained edge computing platform, the Raspberry Pi 4B. The data shows that the model has only 2.00M parameters and the number of floating-point operations is controlled at 0.16G, reflecting a lightweight network design. In terms of inference speed, processing a complete action sample containing 20 frames takes only 131.50 milliseconds, equivalent to approximately 6.58 milliseconds per frame, far lower than the frame interval of typical video streams. This meets the low-latency requirements of real-time action recognition, indicating that this method can be deployed in low-computing-power, low-power embedded devices and has certain engineering application value.

[0039] The above description is merely a preferred embodiment of the present invention and is not intended to further limit the present invention. All equivalent changes made based on the description and drawings of the present invention are within the protection scope of the present invention.

Claims

1. A method for action recognition of wireless signals based on fragment enhancement, characterized in that, Includes the following steps: Step 1: Acquire the wireless radio frequency signal and preprocess it to convert the wireless radio frequency signal into time-series matrix sequence data; Step 2: Construct a wireless signal action recognition model based on fragment enhancement, process the time-series matrix sequence data, extract global context features and attention-guided local context features, and fuse them to obtain deep fusion features; Step 3: Train the wireless signal action recognition model based on fragment enhancement, calculate the loss between the prediction result and the real label based on the deep fusion features, update the parameters of the wireless signal action recognition model, and obtain the trained wireless signal action recognition model. Step 4: Use the trained wireless signal action recognition model to identify the sample data to be identified and output the prediction results of the action category.

2. The method for wireless signal action recognition based on fragment enhancement according to claim 1, characterized in that, The fragment-enhanced wireless signal action recognition model includes a feature extraction network and a classifier; the feature extraction network is used to extract global and local context features and fuse them into deep fusion features to obtain the final sample feature representation; the classifier is used to map the features into action category probabilities. The feature extraction network includes a convolutional neural network encoder, a salient sequence extraction module, and a hierarchical local context modeling module; The convolutional neural network editor extracts frame-level features from the temporal matrix sequence data in chronological order. The salient sequence extraction module extracts global contextual features from frame-level features using a global Transformer model and generates attention guidance information. Based on the attention guidance information, key action segments are adaptively selected; The hierarchical local context modeling module is used to perform hierarchical local context modeling on key action segments and extract local context features. By fusing global and local contextual features, the final sample feature representation is obtained.

3. The method for wireless signal action recognition based on fragment enhancement according to claim 2, characterized in that, The salient sequence extraction module extracts global contextual features from frame-level features using a global Transformer model and generates attention guidance information. Based on the attention-guided information, key action segments are adaptively selected; including: 1) Frame-level features Projected onto the hidden layer dimension d of the Transformer model and with standard sinusoidal position encoding Combination: ， in, and Projection parameters; sequence go through The global Transformer encoder layer is used for processing; each layer applies multi-head self-attention and position-wise feedforward networks, and uses residual connections and layer normalization; the output feature sequence matrix of the last Transformer layer is... T represents the total number of time steps of the input timing data; By using average pooling over the time dimension and a subsequent final projection layer, Aggregate into global embedding : ， ， in, Embed dimensions for the target; This represents the aggregated global feature vector obtained after average pooling along the time dimension; t represents the time step index. This represents a vector consisting of all the feature values ​​of the global Transformer encoder output sequence at time step t. Indicates projection layer operation; 2) Calculate the global attention score using the following formula: ， ， in, For attention weights, It quantifies how much attention is given to key position j in the query position i in the header h; 3) Based on From frame features Dynamic recognition A significant time segment; the significant time segment refers to a time segment with a global attention score. In the example, the feature sequence is extracted centered on the K time steps with the highest numerical ranking.

4. The method for wireless signal action recognition based on fragment enhancement according to claim 3, characterized in that, The hierarchical local context modeling module extracts local context features from key action segments, including: 1) For each segment composed of frame feature sequences, an intra-segment Transformer encoder is used. This is used to capture the microscopic dynamics within the fragment; then, the output is aggregated through time-averaged pooling to produce fragment-level embeddings. : ， in, and Let each represent the nth segment and its corresponding embedding vector; this results in a set of unordered segment representations. ; 2) Convert the unordered fragment representation into a time-ordered sequence matrix Z: ； 3) Input Z into the inter-fragment Transformer encoder In this process, the final local context embedding is generated. : 。 5. The method for wireless signal action recognition based on fragment enhancement according to claim 4, characterized in that, By fusing global and local contextual features, the final sample feature representation is obtained, including: Complementary global representation and detailed local representation splicing together to form a comprehensive, integrated, and embedded system. : ， Here, Concat is the concatenation operation; the fused embedding is input into the final linear classification layer to produce category logits: ， in and These are classification parameters.

Images