A Blind Parsing Method for Communication Protocol Signal Frame Structure Based on Contrastive Learning

Abstract

A blind parsing method for communication protocol signal frame structure based on contrastive learning is proposed. First, a cross-scale waveform feature representation model is constructed, employing a dual-branch contrastive learning strategy at both the instance and segment levels, combined with phase shift and waveform reconstruction enhancement, to extract noise-resistant and semantically robust deep features. Second, a sliding window scanning feature extraction mechanism is used to detect waveform abrupt changes, achieving initial signal segmentation. Subsequently, a variational Dirichlet process Gaussian mixture model is introduced to perform unsupervised clustering of segment features, automatically identifying functional and data segment categories. Finally, redundant boundaries are eliminated through structural inverse optimization, and similar segments are merged to output an accurate frame structure. This invention requires no prior information such as synchronization sequences, modulation schemes, and frame lengths, and can still achieve high-precision, adaptive blind parsing of frame structures under low signal-to-noise ratio and non-ideal channels with hardware impairments, making it suitable for spectrum monitoring and non-cooperative communication signal processing.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, and specifically relates to a blind parsing method for communication protocol signal frame structure based on contrastive learning. Background Technology

[0002] With the development of modern communication technology, communication systems typically exchange information according to pre-designed protocols and frame structures. The frame structure is an important organizational form of physical layer communication signals, usually containing different functional areas such as synchronization fields, control fields, training fields, payload fields, and check fields. The waveform characteristics, modulation schemes, statistical properties, and time-domain lengths of each functional area may differ. Accurately identifying the location and category of each functional area within a frame is a crucial prerequisite for subsequent parameter estimation, modulation identification, synchronization processing, demodulation recovery, and protocol reverse engineering.

[0003] Under cooperative communication conditions, the receiving end, knowing the communication protocol and frame structure, can directly utilize the synchronization word, pilot sequence, and field positions specified in the standard to complete frame synchronization and data parsing. However, in non-cooperative scenarios such as electronic reconnaissance, spectrum monitoring, and communication countermeasures, there is no information sharing relationship between the third-party receiving end and the communicating parties, making it impossible to directly obtain protocol format and frame structure information. Therefore, the receiving end must infer the frame structure from the intercepted waveform without prior knowledge of the protocol, pilot, or synchronization sequence.

[0004] Communication signals are also affected by complex channels and receiver non-ideals during actual propagation. For example, multipath propagation can cause inter-symbol interference, additive noise can reduce the saliency of signal features, carrier frequency offset and phase offset can change the constellation rotation state, and sampling deviation and oversampling factor changes can affect the sampling structure of time-domain waveforms. These factors, superimposed on the intra-frame multi-modulation complex structure, make the analysis of non-cooperative frame structures a key challenge in blind processing of communication signals.

[0005] The existing implementations most similar to this invention mainly include the following categories.

[0006] The first category is blind recognition methods based on statistical features and correlations. Early frame synchronization recognition research mainly relied on the inherent statistical properties of synchronization words or pilot sequences in the time domain, such as periodicity, autocorrelation, and higher-order statistics. For example, Kil et al., in their paper "Analysis of Blind Frame Recognition and Synchronization Based on Sync Word Periodicity," used cross-correlation functions to estimate the frame length multiple for scenarios with unknown frame lengths, and reconstructed the synchronization word in a low signal-to-noise ratio environment by calculating the mean of periodic samples, verifying the asymptotic convergence characteristics and robustness of the correlation detection method under sample accumulation. Bai Yu et al., in their paper "m-sequence frame synchronization code recognition based on higher-order statistical processing techniques," introduced higher-order statistical theory and effectively solved the fuzzy aliasing problem between distributed m-sequence synchronization codes and random information sequences by utilizing the peak significance of the third-order correlation function.

[0007] The second category is analytical methods based on channel coding characteristics and sequence mining. With the widespread application of error-correcting coding technology, methods relying solely on synchronization words have limited performance at low signal-to-noise ratios. Researchers have begun to utilize the algebraic structure of channel coding or data mining techniques for bit-level structural analysis. Imad et al., in their paper "Blind framesynchronization for error correcting codes having a sparse parity checkmatrix," proposed a blind synchronization framework based on maximum a posteriori probability. By calculating the log-likelihood ratio of the checksums in the sparse parity check matrix, they achieved frame synchronization without relying on hard-decision synchronization words, improving the algorithm's noise robustness. Choqueuse et al., in their paper "Blind Recognition of Linear Space–Time Block Codes: A Likelihood-Based Approach," implemented blind recognition of linear space-time block codes for multiple-input multiple-output (MIMO) systems using second-order statistics and a maximum likelihood classifier. In the field of protocol reverse engineering, Song Jiang introduced the multi-sequence alignment technique from bioinformatics into the field of communication in his paper "Exploration and Research on the Discovery of Unknown Protocols in Wireless Network Environment". He inferred the format of unknown protocols by analyzing the association rules of feature sequences in bit streams.

[0008] The third category is the emerging trend of feature mining and intelligent recognition based on deep learning. In recent years, deep learning has received widespread attention due to its powerful feature extraction capabilities, gradually expanding from sequence learning to visual classification of multi-dimensional feature matrices. On the one hand, researchers have attempted to directly utilize neural networks to process data sequences. For example, Li Yonghui, in his paper "Research on Non-cooperative Signal Link Layer Analysis Technology," constructed a non-cooperative signal analysis framework, using deep neural networks to achieve equal-length frame synchronization and combining convolutional networks to identify channel coding types. However, its input features still mainly rely on the demodulated bit sequence. On the other hand, mapping one-dimensional signals to two-dimensional images and using visual models for classification has become the mainstream approach. Jin Mingchao, in his paper "Research on Blind Recognition Technology of Frame Synchronization in Non-cooperative Communication," rearranged multi-frame bitstreams containing synchronization codes into an analysis matrix and converted them into grayscale images, using ResNet to extract texture features to estimate frame length. Although this method improves the recognition effect through image processing, it essentially still falls within the scope of visual modeling of the correlation of demodulated bitstreams.

[0009] In summary, while existing research on frame structure recognition encompasses various technical approaches such as statistical inference, correlation detection, and deep learning, most methods still primarily rely on the analysis of demodulated bit-level sequences. This presents a dual bottleneck in non-cooperative reception scenarios: firstly, unknown frame structures limit the accuracy of correlation demodulation, leading to reduced reliability of the recovered bitstream and severely impacting the performance of bit-feature-based recognition algorithms; secondly, the time-varying characteristics of signal parameters combined with complex channel distortions easily cause synchronization template adaptation, resulting in performance degradation of traditional correlation detection methods. To address the error propagation introduced during demodulation and the template matching failure in dynamic and complex environments, this study proposes a physical layer IQ data frame structure analysis method based on contrastive learning. By mining discriminative robust features from the original waveform, it achieves highly robust analysis of frame structures in complex electromagnetic environments. Summary of the Invention

[0010] To overcome the problems of the existing technologies, this invention aims to propose a blind parsing method for communication protocol signal frame structure based on contrastive learning. This method does not rely on a demodulation pre-processing stage, directly extracting discriminative representations from the original baseband IQ waveform. It achieves cross-scale feature alignment between local details and global dependencies by constructing a Deep Residual Convolutional Transform (DRCT) backbone network. Utilizing a supervised contrastive learning framework and a specially designed physical layer data augmentation mechanism, this invention can accurately locate the boundaries of functional and data segments within a frame under conditions of no protocol prior and the presence of frequency offset, multipath, and other impairments. It also establishes a description of the organizational relationships between waveform segments, providing a highly reliable structural prior for subsequent modulation identification, parameter estimation, and communication system identification tasks.

[0011] To achieve the above objectives, the technical solution of the present invention is as follows: A blind parsing method for communication protocol signal frame structure based on contrastive learning includes the following steps: Step 1: A supervised contrastive learning framework is adopted to train two independent branches: an instance-level feature extraction network and a segment-level feature extraction network. The instance-level and segment-level feature extraction networks share the topology of the DRCT unified feature extraction backbone network, and their parameter sets are independent and not shared during training. The instance-level feature extraction network takes a complete signal instance as input and learns an instance-level global modulation semantic representation. The segment-level feature extraction network takes a signal segment captured by a sliding window as input and learns a segment-level local waveform representation. The contrastive learning framework introduces a waveform reconstruction enhancement strategy, which cuts the same signal instance into multiple segments, randomly shuffles the order, and reassembles them to construct semantically equivalent positive sample pairs. Through cross-scale alignment loss constraints, semantic consistency between the instance-level global modulation semantic representation and the segment-level local waveform representation is established, resulting in a DRCT unified feature extraction backbone network with noise and phase perturbation resistance capabilities. After training, only the weights of the segment-level feature extraction network are retained for subsequent inference. Step 2: Input the signal to be tested into the trained segment-level feature extraction network, use sliding window scanning to extract feature sequences, calculate the cosine similarity between adjacent features and align and fuse them along the time axis into a global similarity sequence, and obtain the global difference trajectory based on this. Candidate boundaries are detected by adaptive thresholding, and the signal is initially divided into multiple quasi-stationary short segments. Step 3: Calculate the feature mean of each quasi-stationary short segment as the central representation of the segment, perform unsupervised clustering on the set of central representations composed of all central representations, assign semantic labels to each short segment, and identify the functional segment and data segment categories. Step 4: Traverse the semantic tags of adjacent short segments. If adjacent tags are the same, they are identified as redundant boundaries and removed. Similar segments are merged into a complete physical region. If adjacent tags are different, they are identified as real frame structure state switching points and retained. Output the final set of frame boundaries and segment group categories to complete the signal frame structure parsing.

[0012] The unified feature extraction backbone network for DRCT described in step 1 adopts a deep residual convolutional transform model. The instance-level feature extraction network and the segment-level feature extraction network are two instantiations of this unified feature extraction backbone network: they use the same network topology and module configuration, differing only in the observation scale of the input signal. The instance-level feature extraction network takes a complete signal instance as input, while the segment-level feature extraction network takes a local signal segment captured by a sliding window as input. During the training phase, their parameter sets are independent and not shared; during the inference phase, only the segment-level feature extraction network is called. The unified feature extraction backbone network for DRCT consists of three cascaded modules: Waveform encoding module: A 9-layer fully convolutional network is used to extract hierarchical features from the input complex baseband IQ signal. The terminal global average pooling layer is removed to preserve the temporal resolution. High-dimensional local feature sequences are output through residual connections and group normalization. Global temporal modeling module: Sine-cosine position codes are superimposed on the high-dimensional local feature sequence and input to a module consisting of L layers of Transformer encoders. The module models long-range statistical dependence and global modulation law through a multi-head self-attention mechanism. Feature mapping module: The global description output by the global temporal modeling module is mapped to a low-dimensional embedding space through the projection head, and then constrained to the unit hypersphere by L2 normalization, so that the similarity between samples is uniquely determined by the angle between the feature vectors.

[0013] In step 1, the supervised contrastive learning framework is optimized using a joint loss function that includes the following three terms: Global scale contrast loss: Align the original signal instance, phase shift enhancement instance, and waveform reconstruction enhancement instance under the same supervision to reduce the impact of phase uncertainty and segment rearrangement on the instance-level global modulation semantic representation; Describe the instance-level features of the original signal samples Two types of enhancements are constructed for each sample: phase shift enhancement produces... Group features Waveform reconstruction enhancement Group features The global scale contrast loss is defined as: Local scale contrast loss: It brings the features of segments at different time locations within the same signal instance closer together under the supervision of the same type, and constrains their consistency with the corresponding segments of the enhanced samples, thereby improving the stability of segment-level representations to positional changes and phase perturbations. Record the original sample number The characteristics of each segment are: Phase shift enhancement sample The enhancement in the first The characteristics of each segment are: Accordingly, define the set of fragments within a batch. Local scale contrast loss, using the first segment of each sample as the anchor reference, is defined as: Cross-scale alignment loss: Under supervised constraints, semantic consistency is established between instance-level global modulation semantic representation and fragment-level local features, so that local fragment discrimination information is corrected under instance-level global semantic reference; Record No. The global features of each phase shift enhancement instance are: and with Indicates the first A set of fragments, then .

[0014] The phase offset enhancement is achieved by applying an additional global phase rotation to the received signal to simulate phase and frequency offset disturbances in the actual channel; the waveform reconstruction enhancement is specifically achieved by cutting the original waveform sequence into multiple segments, randomly shuffling their order, and then reassembling them; the original signal instance, the phase offset enhancement instance, and the waveform reconstruction enhancement instance all participate in supervised comparative optimization.

[0015] The input signals of the instance-level feature extraction network include original signal instances, phase-shift enhancement instances, waveform reconstruction enhancement instances, similar signal instances, and dissimilar signal instances. The input signals of the segment-level feature extraction network include signal segments extracted from the above instances using a sliding window. The signals are defined as follows: Example of raw signal: The raw complex baseband IQ signal without enhancement processing, used as a benchmark sample for contrastive learning; Phase offset enhancement example: A sample generated by applying additional global phase rotation to the original signal example, used to simulate phase and frequency offset disturbances in a real channel; Waveform reconstruction enhancement example: The original signal instance is cut into multiple segments, the order is randomly shuffled and reassembled to generate samples, which are used to construct semantically equivalent positive sample pairs; Similar signal instances: Signal samples with the same modulation scheme as the original signal instances are used as positive samples for comparative learning within the batch; Outlier signal instances: Signal samples with different modulation schemes from the original signal instances are used as negative samples in the batch for comparative learning. The global scale contrast loss, local scale contrast loss, and cross-scale alignment loss are all constructed based on the similarity calculation between the above signal instances.

[0016] The multi-starting-point parallel scanning mechanism described in step 2 specifically includes: pre-setting a set of different starting offset indices and performing sliding window feature extraction in parallel; for each starting offset, calculating the cosine similarity between adjacent features in the feature sequence to obtain a similarity sequence under each scanning path; aligning the similarity sequences of each path along the time axis and taking the average to merge them into a global similarity sequence; and calculating the global dissimilarity. , For the global difference trajectory, For global similarity sequences; set a threshold. If the position is a discrete point with a value greater than the threshold, it is directly extracted as the initial boundary; if it is a continuous high-discrepancy region, the local maximum value of the region is extracted as the initial boundary, and finally the initial segmentation boundary set is determined (here the initial segmentation boundary set may result in over-segmentation). It is the maximum value of the global difference trajectory G over the entire sequence of signals to be tested.

[0017] Step 3 specifically includes: for each sub-segment interval divided by the initial boundary, calculate the statistical mean of all feature vectors in the interval as the central representation of the segment; use the variational Dirichlet process Gaussian mixture model to cluster the central representation set and output the semantic label sequence of each segment, where different labels correspond to functional segments or data segments.

[0018] The structural reverse optimization described in step 4 specifically involves traversing adjacent short segments in chronological order. and and its semantic tags and If adjacent labels are the same If two segments are determined to belong to the same physical region, the dividing point is removed. And merge the two segments; if adjacent labels are different, ≠ The points that are identified as real frame structure state transition points are retained; the final output is a set of precise frame boundaries. The corresponding segment group category; after step 4, a more accurate boundary set can be obtained.

[0019] The signal under test is a single-carrier communication signal. The frame contains functional segments and data segments with different modulation methods, different symbol set constraints, or different statistical characteristics. The parsing process does not require the demodulated bit-level sequence, protocol prior information, or synchronization sequence prior.

[0020] The dataset used in the training phase of step 1 includes hardware impairment simulations of IQ amplitude-phase imbalance and IQ phase quadrature error. The IQ amplitude-phase imbalance factor ranges from [-10, 10] dB, and the IQ phase quadrature error ranges from [-10, 10] degrees. Combined with additive white Gaussian noise, multipath fading, carrier frequency offset, and initial phase offset, a training environment under non-ideal reception conditions is constructed.

[0021] Compared with the prior art, the present invention has the following advantages: 1. Existing technologies typically rely on manually designed features (such as cyclic spectra and higher-order moments) or standard deep learning models, which are highly susceptible to phase rotation, frequency offset, and hardware impairments (IQ imbalance), leading to a sharp performance drop at low signal-to-noise ratios. This invention constructs a cross-scale waveform feature representation model (DRCT architecture) and utilizes a dual-branch contrastive learning mechanism at both the instance and segment levels to force the network to ignore irrelevant perturbations such as phase and frequency offset, focusing on mining the inherent modulation semantic structure of the signal. In particular, a waveform recombination enhancement strategy is introduced, enabling the model to resist data rearrangement changes under unknown protocols. Experiments demonstrate that this method maintains high segmentation accuracy even under low signal-to-noise ratio conditions, solving the problem of feature mismatch in complex electromagnetic environments.

[0022] 2. Existing blind parsing methods often require assumptions about known synchronization headers, specific pilot sequences, or frame length ranges, making it difficult to handle unknown or dynamically changing communication protocols. This invention adopts a purely data-driven unsupervised clustering approach. The front end uses multi-starting-point scanning to capture waveform abrupt changes, while the back end uses a variational Dirichlet process Gaussian mixture model (VDPGMM) to automatically infer the number of functional segment categories in the frame structure, eliminating the need to manually set the number of cluster centers or frame length thresholds.

[0023] 3. Compared to traditional methods that require full-sequence decoding or large-scale matrix operations, this invention optimizes the computational process. The front-end feature extraction of this invention employs a lightweight convolutional and Transformer hybrid architecture, avoiding parameter explosion in fully connected layers; the back-end clustering is performed directly on the segment mean features, rather than through point-by-point computation. This two-stage processing mode of coarse segmentation and fine clustering significantly reduces computational power consumption while maintaining high accuracy, enabling near real-time signal frame structure parsing on general-purpose GPUs or high-performance embedded platforms, meeting the deployment requirements of practical spectrum surveillance and electronic reconnaissance systems.

[0024] In summary, this invention systematically solves the failure problem of existing communication protocol frame structure parsing technology in complex environments from four aspects: feature characterization (interference resistance), parsing logic (blind identification), and engineering deployment (high efficiency), filling the technical gap in unsupervised deep structure mining of unknown signals. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of waveform representation based on contrastive learning. The diagram includes five types of input signals: original signal instance, phase shift enhancement instance, waveform reconstruction enhancement instance, similar signal instance, and dissimilar signal instance. The arrows indicate the feature similarity calculation and constraint relationship between each signal instance under the contrastive learning framework, and the dashed box indicates the construction range of positive sample pairs. Figure 2This diagram shows the network structure of instance-level / fragment-level feature extraction. Different colored blocks represent the three-level cascaded modules of the unified feature extraction backbone network for DRCT: blue blocks represent waveform encoding modules, orange blocks represent global temporal modeling modules, and green blocks represent feature mapping modules. The instance-level and fragment-level feature extraction networks are two instances of the same backbone network, both using the same blue-orange-green serial structure, differing only in the observation scale of the input signal. The arrows indicate the transmission path of the feature data stream. Figure 3 Flowchart of waveform structure analysis strategy; Figure 4 The feature space distribution diagram of the DRCT segment-level waveform representation model after extracting features from some modulation types; Figure 5 Visualization of the waveform structure analysis of a 4dB Sig-I signal-to-noise ratio signal; Figure 6 Visualize the waveform structure analysis of a 10dB Sig-I signal-to-noise ratio signal; Figure 7 The feature distribution of each segment of a single-frame Sig-I signal after passing through a feature extraction network; Figure 8 The graph shows the analytical performance of each algorithm on the waveform structure of the Sig-I signal under different signal-to-noise ratios. Figure 9 The graph shows the analytical performance of each algorithm on the Sig-II signal waveform structure under different signal-to-noise ratios. Figure 10 The graphs show the analytical performance of each algorithm on the Sig-III signal waveform structure under different signal-to-noise ratios. Figure 11 The graph shows the performance of each algorithm in analyzing the waveform structure of the Sig-IV signal under different signal-to-noise ratios. Figure 12 The graph shows the analytical performance of each algorithm on the Sig-I signal waveform structure under different frequency deviations. Figure 13 The graph shows the analytical performance of each algorithm on the Sig-II signal waveform structure under different frequency deviations. Figure 14 The graph shows the analytical performance of each algorithm on the Sig-III signal waveform structure under different frequency deviations. Figure 15 The graph shows the analytical performance of each algorithm on the Sig-IV signal waveform structure under different frequency deviations. Detailed Implementation

[0026] The embodiments and effects of the present invention will be further described in detail below with reference to the accompanying drawings.

[0027] To address the problem of frame structure parsing in communication protocol signals, this section proposes a contrastive learning-based processing architecture. This architecture consists of two cascaded stages: waveform feature representation and structure mining. First, a cross-scale feature alignment strategy is employed to construct and train a segment-level feature extraction network. Second, a sliding window mechanism is used to extract high-dimensional sequence features of local waveforms. Finally, a combined segmentation and clustering algorithm is used to locate segment boundaries and aggregate homogeneous segments in single-carrier communication frame signals, thereby achieving effective parsing of the signal frame structure.

[0028] 1. Waveform Feature Representation Based on Contrastive Learning To address the non-stationary nature of single-carrier communication signal frames, different segments (such as functional segments and data segments) exhibit variations in temporal morphology and statistical distribution. Directly modeling the entire frame globally can easily lead to feature aliasing across multiple segments, thereby weakening the discriminative power for subtle structural changes at boundaries. To solve this problem, this invention employs a sliding window mechanism for local feature extraction. This mechanism effectively isolates neighboring waveform interference through short-time window processing, thereby improving the accuracy of boundary localization during frame structure segmentation. Figure 1 As shown, the segment-level network is responsible for encoding and representing local signal segments within the sliding window. Here, "segment-level" specifically refers to the local signal observation unit extracted from the short-term sliding window.

[0029] However, due to the physical scale of the sliding window, fragment-level networks face certain challenges in processing high-order modulation signals. Because short-time observations cannot meet the statistical sufficiency required for higher-order modulations and cannot cover the complete symbol sequence or statistical period, the intra-class distribution in the feature space exhibits a discretization trend (i.e., insufficient intra-class compactness), which in turn affects the stability of local discrimination.

[0030] To address the feature divergence problem caused by limited observation scale, this paper employs a cross-scale feature alignment strategy at both the fragment-level and instance-level. In contrast to the "fragment-level," which is limited to a local scale, the "instance-level" here refers to a global signal sequence with a longer observation period, capable of reflecting complete macroscopic statistical characteristics. This strategy introduces an instance-level model to model long-term signals, capturing richer modulation semantics. By constructing consistency constraints between fragment-level local representations and instance-level global representations, this strategy utilizes global semantic priors to guide and correct local embeddings. This method effectively suppresses semantic shifts in local representations while retaining the advantages of the sliding window inference paradigm, improving the aggregation and separability of higher-order modulation signal fragments in the feature space. Furthermore, to enhance the distinguishability and representation robustness between samples of different modulation types, this paper introduces two types of data augmentation strategies within a supervised contrastive learning framework. The first method is phase offset-based enhancement. By applying an additional global phase rotation to the received signal, it simulates the unavoidable phase / frequency offset disturbances in actual channels, thereby prompting the model to learn discriminative features invariant to phase changes. The second method is waveform segment recombination enhancement. Specifically, we cut the original waveform sequence into multiple segments, randomly shuffle their order, and then reassemble the shuffled sequences into new waveform samples. Although the temporal information is disturbed, the statistical characteristics of the modulated signal are preserved, and the sample remains semantically equivalent to the original sample, thus constructing effective positive sample pairs. The positive sample pairs constructed through the above enhancements, together with cross-class negative sample pairs, participate in supervised contrastive optimization, which helps to further improve the intra-class compactness and inter-class separability of the feature space. See [link to relevant documentation]. Figure 1 .

[0031] (1) Waveform representation model To support the proposed cross-scale feature alignment strategy and achieve comparability and alignment of waveform representations across different observation scales in a homogeneous semantic space, this paper constructs a unified temporal feature extraction backbone network, which is instantiated into instance-level and segment-level feature extraction networks. Both maintain a consistent network topology and module configuration, but their parameter sets are independent and not shared during the comparative learning optimization process. It should be noted that the training phase does not involve learning on frame-structured signals containing concatenated synchronization and data segments, but rather on data area signals containing only a single waveform and modulation scheme: the instance-level network learns instance-level global modulation semantic representations using complete signal instances as input; the segment-level network learns local waveform representations using local segments obtained by sliding window segmentation of the same signal instance as input. This dual-scale modeling enables the model to obtain complementary representations of both local and global types during training. By using representation association constraints, the fragment-level representations are semantically aligned with the instance-level global modulation semantic representations. During the inference phase, when the model faces multi-segment signals with frame structures, the fragment-level network can perform local encoding and discrimination within the frame using a sliding window approach. Guided by instance-level global semantic priors, it alleviates representation divergence caused by insufficient statistical sufficiency of short windows, thereby improving the robustness of frame structure parsing.

[0032] The feature extraction network structure proposed in this paper is as follows: Figure 2 As shown, this model is named the Deep Residual Convolutional Transformer (DRCT). It should be noted that... Figure 2 The different colored blocks in the diagram represent the three-level cascaded modules of the unified feature extraction backbone network of DRCT. The differences and relationships between the modules are as follows: Blue block (waveform coding module): As the front-end input layer, it receives complex baseband IQ signals and outputs high-dimensional local feature sequences that retain temporal resolution, providing a basic representation for subsequent modules; Orange block (global temporal modeling module): As an intermediate layer, it is superimposed on the output features of the blue block. It models long-range statistical dependencies and global modulation patterns through the Transformer encoder, thereby enhancing the semantic consistency of the features. Green block (feature mapping module): As the final output layer, it maps the output of the orange block to a low-dimensional embedding space and constrains it to the unit hypersphere by L2 normalization, providing a representation basis for the calculation of cross-scale alignment loss.

[0033] During the training phase, both the instance-level feature extraction network and the fragment-level feature extraction network adopt the aforementioned blue-orange-green serial structure. The difference lies in the input scale of the blue blocks: the instance-level network inputs complete signal instances, and the orange blocks can model long-range dependencies across the entire frame; the fragment-level network inputs local fragments captured by a sliding window, and the orange blocks only fuse local discriminative features within the window. The cross-scale alignment loss constrains the feature vectors output by the green blocks of both networks to maintain semantic consistency, thereby correcting the local representations for global semantics.

[0034] This backbone network sampling hybrid architecture design integrates the convolutional neural network's ability to extract fine local structures with the Transformer's ability to model long-term temporal dependencies globally. The overall architecture consists of three cascaded modules: waveform encoding, global temporal modeling, and feature mapping.

[0035] ① Waveform encoding module The waveform encoding module is used to map the input complex baseband IQ temporal sequence into a high-dimensional local representation. This module uses a 9-layer fully convolutional network as the front-end backbone. Its design removes the global average pooling and fully connected layers at the end of conventional CNNs, avoiding premature compression of temporal information in the early stages of feature extraction, thereby preserving feature sequences rich in temporal details for the downstream Transformer encoder.

[0036] Given input (in Indicates IQ channel, (where the sequence length is 1), firstly, preliminary feature extraction and downsampling are performed through convolutional layers, and then stable nonlinear mapping is obtained by combining Group Normalization and ReLU: in, Represents ReLU. This indicates Group Normalization. For convolution kernel weights, This corresponds to the convolution stride. Subsequently, the features... Hierarchical representation learning is performed sequentially through four residual stages. Stage 1 maintains temporal resolution, Stage 2 and Stage 3 use convolution with a stride of 2 to achieve downsampling, and Stage 4 uses a stride of 1 to preserve the end time length. At the same time, the number of channels is gradually increased from 64 to 128, 256 and 512, forming a multi-level representation of waveform structures at different time scales.

[0037] Let the first The residual input is Its main branch can be represented as: in, and These represent the weight parameters of the two convolutional layers in the main branch. The sampling step size specified for the current Stage. When the main branch introduces downsampling or changes the number of channels, the residual branch uses... One-dimensional convolution is used for linear projection alignment; otherwise, identity mapping is applied. Let this alignment operator be denoted as . Then the first The final output of each residual block is defined as: in for The weights of the projected convolution kernels. Through the residual convolutional encoding described above, the network provides a high-dimensional temporal feature representation rich in discriminative information for the subsequent Transformer module.

[0038] ② Global Temporal Modeling Module The global temporal modeling module is used to model long-range statistical dependencies and global modulation patterns on the feature sequences after convolutional coding, in order to further improve the comprehensive robustness of the representation to channel distortions under complex electromagnetic environments such as frequency offset, multipath, and IQ imbalance, while enhancing the semantic consistency of the features. Let the waveform coding module output sequence features be... in This represents the encoded time steps. To explicitly inject position information, sine-cosine position coding is introduced. And then overlay them: The position code of the first The position, the and The dimensions are respectively Subsequently, Input by A module composed of stacked Transformer encoders. For the first... Layers are first obtained through linear mapping. : And establish full sequence interactions by scaling dot product attention: Multi-head attention concatenates the outputs of each head and linearly maps them to obtain... The nonlinear transformation is then completed via a feedforward network, and the final output is obtained through residual connections. After being processed by stacked L layers of Transformer encoders, the final output is the encoded sequence representation. H(L) is the final output of the Transformer in the global temporal modeling module.

[0039] Finally, the encoded sequence representation is obtained. .

[0040] This module is used to model long-range dependencies in both fragment-level and instance-level networks. In instance-level networks, because the input is a complete signal, this module can aggregate sufficient global statistical information based on a longer observation sequence. In fragment-level networks, its role is to fuse discriminative features within a limited observation window. In subsequent representation association constraints, the representations generated by instance-level networks are more stable because they are based on complete signals and can serve as global semantic priors, guiding fragment-level networks to improve the intra-class consistency and feature compactness of their representations.

[0041] ③ Feature mapping module The feature mapping module maps the global description output from the upstream to a low-dimensional embedding space through a projection head, and then... Normalization constrains feature representations to a unit hypersphere. Since the similarity between samples is entirely determined by the angle between feature vectors, it is mathematically consistent with the cosine similarity metric. This provides a representational basis for subsequent alignment of instance-level and fragment-level representations, as well as loss calculation in supervised contrastive learning.

[0042] (2) Loss function design To effectively utilize label and supervisory information to drive the model to learn highly discriminative signal representations, this paper adopts the Supervised Contrastive Learning (SupCon) paradigm to construct the loss function. Unlike traditional self-supervised contrastive methods that rely solely on data augmentation views to construct positive samples, SupCon expands the definition of positive samples, considering samples belonging to the same modulation category within the same batch as a set of positive samples. This aims to maximize the mutual information of similar samples in the feature embedding space, thereby achieving a feature distribution geometry that is "compact within classes and separable between classes."

[0043] Specifically, to capture the semantic features of signals at different scales, this paper employs a multi-level joint optimization objective, comprising three core components: a global scale contrastive loss for aligning the overall representation of the signal, a local scale contrastive loss for mining the temporal invariance of local waveforms, and a cross-scale alignment loss for establishing a cross-level semantic mapping between global and local features. To uniformly describe the above optimization process, we first define a general supervised contrastive loss operator, and then elaborate on the specific forms of the loss functions at each level.

[0044] ① Basic definitions and symbols Suppose the current batch contains N samples. For any two feature sets... and We define the general supervised contrastive loss operator as follows: For anchor point samples Its positive sample set Defined as the set of sample indices with the same label in a batch. The general loss term is defined as follows: in, Represents cosine similarity. This is a temperature scalar parameter used to control the model's attention to difficult negative samples. The summation term in the denominator... Indicates the calculation anchor point The sum of similarities with all samples in the batch except itself.

[0045] ② Global scale contrast loss Describe the instance-level features of the original signal samples Two types of enhancements are constructed for each sample: phase shift enhancement produces... Group features Waveform reconstruction enhancement Group features The global scale contrast loss is defined as: The first term aligns the original instance with the phase-shifted enhanced instance under the same supervision to reduce the impact of phase uncertainty on the global semantic representation. The second term constrains the consistency between different phase-shifted enhancements, suppressing intra-class distribution expansion caused by enhancement perturbations. The third term aligns the original instance with the waveform reconstruction enhanced instance, ensuring that the global representation remains stable for the rearrangement of equivalent segments under the same modulation semantics. These three terms work together to improve the intra-class compactness and inter-class separability of instance-level embeddings.

[0046] ③Local scale contrast loss Each instance sample is divided at the fragment level. The original sample contains 128 segments, each with a length of 128. The characteristics of each segment are: Phase shift enhancement sample The enhancement in the first The characteristics of each segment are: Accordingly, define the set of fragments within a batch. Local scale contrast loss, using the first segment of each sample as the anchor reference, is defined as: The first term brings different time segments of the same semantic category closer together under the same supervision, thereby strengthening the stability of segment-level representation to changes in segment position and improving intra-class consistency. The second term further aligns the representations of the original anchor segment and the augmented sample at different segment positions, so that the segment-level embedding is consistent with changes in segment position and phase perturbation, thereby reducing the drift of segment representation caused by augmentation perturbation and improving the stability of segment-level discriminative representation.

[0047] ④ Cross-scale alignment loss To explicitly establish consistent alignment constraints between instance-level global semantics and fragment-level local semantics, a cross-scale alignment loss is introduced, ensuring that instance-level embeddings and fragment-level features maintain alignment within the same metric space. Let the... The global features of each phase shift enhancement instance are: and with Indicates the first A set of fragments, then This approach ensures that the semantic direction of "global instance representation - local fragment representation" remains consistent under supervised constraints, thereby enabling the discrimination information of local fragments to be corrected and constrained under global semantic reference, and improving the comparability and consistency of cross-scale representations.

[0048] ⑤ Total Loss Combining the above three types of complementary objectives, the final optimization objective function is formed: In the above formula, , as well as These represent the weights for the global scale loss, local scale loss, and cross-scale alignment loss, respectively. In this paper, all these weights are set to 1.

[0049] (3) Training dataset and parameter settings ① Training dataset settings To verify the effectiveness of the waveform feature representation model based on supervised contrastive learning proposed in this paper, the experiment did not directly use the publicly available automatic modulation recognition benchmark dataset RadioML 2018.01A. Although this dataset is widely used in modulation recognition research, its generation mechanism mainly focuses on the multipath effect and clock synchronization error during the wireless channel propagation process of the model, without fully considering the IQ amplitude and phase imbalance effect commonly found in actual receiver hardware. IQ imbalance can cause image interference in the signal, leading to a shift in the feature space distribution and severely affecting the ability of the representation model to extract robust general features. To comprehensively evaluate the robustness of the model under non-ideal hardware conditions and complex channel environments, this paper constructs a dedicated simulation dataset containing IQ imbalance features based on the Matlab platform. This dataset covers four major types of modulation signals: ASK, PSK, APSK, and QAM. In the data generation process, not only additive white Gaussian noise, Ricean channel multipath fading, carrier frequency offset, and phase offset were introduced, but random IQ amplitude imbalance and phase quadrature error were also specifically introduced to address the non-ideal characteristics of the hardware. In addition, to eliminate the differences in received power caused by channel fading and different transmission distances, and to accelerate the convergence of the neural network, we perform amplitude-based maximum normalization on the original IQ signal.

[0050] The specific simulation parameter settings for the training dataset are shown in Table 1.1 below.

[0051] Simulation training dataset generation parameter configuration table ② Setting parameters related to model training The training parameter settings for the model in this chapter are shown in the table below: Model training related parameter settings A Waveform Structure Analysis Method Driven by Segmentation-Clustering Collaboration The waveform structure analysis strategy proposed in this section analyzes the signal frame structure from two basic dimensions: waveform segments and segment groups. The former focuses on abrupt changes in the signal's time dimension, while the latter focuses on the homogeneity of the signal's features. Through the synergistic analysis and integration of these two dimensions, a concise and structured representation of communication signal frames can be constructed, thus providing effective feature support for subsequent downstream tasks such as waveform parameter estimation and modulation identification. To achieve the automatic extraction of the above information, this paper proposes the following... Figure 3 The analysis strategy shown mainly includes two core processing stages: waveform segmentation and segment composition mining. The processing procedure is described in detail below.

[0052] (1) Waveform segmentation based on sliding window scanning In blind segmentation tasks, since the true physical boundaries of the received sequence T are unknown, this paper uses a fixed-step sliding window scanning mechanism to continuously extract local waveform representations and measure differences, thereby locating the unknown boundaries. The core steps are as follows: ① Sliding window scanning and representation extraction: A fixed-length sliding window is set, and continuous sliding scanning is performed on the received sequence T at specified step sizes. At each sliding position, a local waveform segment is captured and input into the segment-level waveform representation model to extract the corresponding waveform features.

[0053] ② Global Dissimilarity Trajectory Construction: Based on the extracted feature sequences Calculate the cosine similarity between the feature vectors of adjacent sliding windows. Then, use the similarity sequence obtained along the sliding window scanning path. Aligned along the time axis and merged into a global similarity sequence Subsequently, the global dissimilarity trajectory is calculated using the following formula. : ③ Threshold setting and coarse boundary determination: Calculation global maximum value Set adjustment coefficient Define threshold for: Use this threshold to filter global dissimilarity trajectories If a position greater than the threshold is a discrete point, it is directly extracted as the initial boundary; if it is a continuous region of high difference, the local maximum value of that region is extracted as the initial boundary. This ultimately forms the initial set of segmented boundaries. .

[0054] It should be noted that the candidate boundaries detected in this step are not the final true frame boundaries. Due to noise interference or instantaneous waveform fluctuations, a single scan path may produce false peaks. This invention fuses the global difference trajectory G through parallel scanning at multiple starting points and performs threshold filtering on G. The final output initial segmentation boundary set Y is the candidate boundary set after filtering and denoising. This set contains all possible frame structure switching points, and redundant boundaries need to be further eliminated in subsequent steps through clustering and inverse optimization.

[0055] (2) Segment semantic clustering based on VDPGMM After initial segmentation, the signal was divided into several quasi-stationary short segments. To identify the semantic categories of these segments, this paper establishes a clustering structure in the representation space. The specific steps are as follows: ①Targeting the Each sub-slice end interval Calculate the statistical mean of its eigenvectors, and use it as the central representation of the segment. : ② Subsequently, the VDPGMM (Variational Dirichlet Process Gaussian Mixture Model) algorithm was used to process the set. Perform unsupervised clustering. After clustering, output the semantic label sequence for each segment. ,in , This represents the number of segments.

[0056] (3) Structural reverse optimization To address the potential oversegmentation problem in the initial segmentation results—that is, a single physical region (such as a complete data area) is represented as a sequence of multiple continuous sub-segments—this section proposes a structural inverse optimization mechanism. This mechanism aims to remove redundant boundaries and fuse homogeneous segments in the temporal segmentation results based on semantic-level identity constraints.

[0057] Specifically, the algorithm utilizes the semantic label sequence obtained in the previous stage. Execute the decision logic based on homogeneous merging: Traverse two adjacent sub-segments and and their corresponding semantic tags and If the two tags are the same (i.e. If the two segments belong to the same physical region, then the dividing point in between is considered the boundary. This is considered an invalid redundant boundary. At this point, the algorithm performs a fusion operation, removing the boundary and... and Merge into a single fragment .

[0058] Conversely, if adjacent labels are different If the boundary is determined to be a true frame structure state switching point, it will be retained.

[0059] Through this reverse optimization process, the algorithm can effectively eliminate invalid internal boundaries, reassemble fragmented segments of the same type into complete physical layer segments, and finally output a more accurate set of frame boundaries. By matching the corresponding segment group categories, the structure of the frame signal can be analyzed.

[0060] 3 Simulation Experiment 3.1 Experimental Setup (1) Test dataset Based on the typical single-carrier multi-waveform communication signal protocol frame format, this paper generates four simulation test signals, and their specific parameters and structural designs are as follows: Test dataset settings signal number Functional segment Data segment Single-frame structure Single-frame symbol sequence length Sig-I A random sequence of 32 symbols, modulated using QPSK. A 288-symbol random sequence, modulated using 16QAM. [Function segment, Data segment, Function segment] 352 symbols Sig-II A 32-symbol random sequence, modulated using 8PSK. A 288-symbol random sequence, modulated using 32QAM. [Function segment, Data segment, Function segment] 352 symbols Sig-III 32-symbol fixed sequence, using 8PSK modulation A 96-symbol random sequence, modulated using 8PSK. [Function segment, data segment, function segment, data segment, function segment] 288 symbols Sig-IV A 32-symbol random sequence, using QPSK modulation. A 288-symbol random sequence, modulated using 32QAM. [Function segment, Data segment, Function segment] 352 symbols It should be noted that although both the functional segment and data segment of the Sig-III signal in the table use 8PSK modulation, their symbol set constraints are different. The functional segment consists of two sets of 8PSK modulated signals {0,0,0,0,0,2,4,6,0,4,0,4,0,6,4,2}; while the data segment is a random symbol sequence, whose symbols can be found in the complete 8PSK symbol set. Randomly selected values.

[0061] After generating multi-frame signals based on the frame structure in Table 1.3 above, non-ideal factors are then added to the signals, such as frequency offset, multipath, phase shift, and IQ imbalance. The setting range of non-ideal factors is the same as that in Table 1.1, and will not be repeated here.

[0062] (2) Comparison method To systematically evaluate the overall performance of the proposed algorithm, we adopted a comprehensive experimental design combining comparison with classic time segmentation algorithms and feature model replacement verification. First, we selected two widely accepted time segmentation algorithms as benchmark comparison methods to establish a reference system for performance comparison. Second, to further verify the superiority of the proposed feature extraction model in its representational capabilities, in the waveform structure mining process described above, we replaced the waveform feature mining module with two widely used time-series feature extraction models and conducted comparative tests under the same experimental settings. This design aims to comprehensively analyze the algorithm's performance through multi-faceted comparison. For brevity and ease of subsequent reference, this paper refers to the proposed waveform mining algorithm based on DRCT model feature extraction as DRCT-WS (DRCT-based Waveform Segmentation).

[0063] Simulation Test Parameter Setting Table Comparison Methods Method Description TICC TICC is a multivariate time series joint segmentation and clustering method that models the dependencies between variables by estimating the inverse covariance matrix specific to each state, and outputs a state sequence with consistent statistical properties. GGS GGS is an efficient change point detection algorithm based on a greedy strategy. It directly divides a sequence into multiple continuous segments that follow different Gaussian distributions by iteratively inserting split points to maximize the global likelihood. Resnet18-WS Waveform segmentation algorithm that replaces instance-level and fragment-level feature extraction networks with an improved one-dimensional ResNet18 pre-trained model. CNN-WS The waveform segmentation algorithm replaces the instance-level and fragment-level feature extraction networks with an improved multi-layer one-dimensional CNN pre-trained model. (3) Waveform structure analysis evaluation index To evaluate the performance of the proposed temporal segmentation method, this paper constructs a comprehensive evaluation framework and uses the following three types of indicators for quantitative analysis: coverage score as a segmentation evaluation indicator to measure the accuracy of boundary matching between predicted segments and actual segments; and adjusted RAND index and standardized mutual information as clustering evaluation indicators to quantify the consistency between predicted segment groups and actual segment groups in clustering results.

[0064] ①Coverage Score Coverage score primarily measures the degree of overlap between predicted time segments and ground truth segments over a given time span. This metric tends to reward predictions that cover the ground truth segments to the greatest extent possible, regardless of the specific numerical value of the segment labels.

[0065] Assume the true set of segments is The predicted segment set is For each true segment Find the predicted segment with the highest Jaccard similarity index. Perform matching. The coverage score is calculated as the weighted average of the maximum overlap of all true segments: in, This represents the total length of the time series. Indicates the first The time length of each true segment. The Jaccard similarity coefficient is defined as the ratio of the intersection to the union of two time intervals: The range of values ​​for the coverage score is: A higher value indicates a more accurate prediction of the boundary delineation.

[0066] ② Adjust the Rand Index (ARI) The Rand Index evaluates clustering performance by calculating the consistency between the predicted and actual assignments of sample pairs. Adjusted Rand Index introduces a correction for random guessing, addressing the issue of the baseline value not being zero as the number of clusters increases.

[0067] Given a dataset containing N time points, ARI treats the predicted label U and the true label V as two distinct partitions. ARI is calculated based on a contingency table, and the specific formula is as follows: RI stands for RAND Corporation Index. The RI value represents the period under random clustering, where a is the number of logarithmic samples of the same class assigned to the same cluster, and b is the number of logarithmic samples of different classes assigned to different clusters. For the sample logarithm, The total number of samples is denoted as ARI. The ARI value is usually in the range of [-1, 1], where 1 indicates complete agreement and 0 indicates that the result is close to random.

[0068] ③ Standard mutual information Normalized mutual information (NMI) is based on the concept of entropy in information theory and is used to quantify the amount of shared information between two cluster distributions. Compared to pure mutual information, NMI eliminates the bias of the number of clusters on the evaluation results through normalization, making it more suitable for comparing results with different numbers of clusters.

[0069] NMI is defined as the true label distribution. With predictive labels Mutual information between The ratio of the geometric mean (or arithmetic mean) of the entropies of the two: in, Shannon entropy represents mutual information. The calculation formula is: In the above formula, the value of NMI ranges from [0,1]. The closer the value is to 1, the more information about the true result is contained in the prediction result, that is, the better the clustering effect.

[0070] 3.2 Feasibility Analysis This section analyzes the feasibility of the segment-level waveform representation model and the DRCT-WS waveform structure analysis algorithm proposed in this paper.

[0071] Figure 4 This paper presents the two-dimensional spatial distribution of high-dimensional features extracted from signals of certain modulation types using the DRCT segment-level representation model, after dimensionality reduction using the t-SNE algorithm. The experiment was conducted with a sample size of 50 for each modulation type, a sequence length of 128, an oversampling factor of 8, and a signal-to-noise ratio of 20 dB. The feature distribution map shows that feature points of similar modulation signals are tightly clustered in the two-dimensional space, while feature clusters of different modulation types are independent and have clear boundaries, demonstrating good feature separability.

[0072] During the feasibility test of the DRCT-WS waveform structure analysis algorithm, the sliding window length was set to 128. The following analysis uses an 8x oversampled Sig-I signal as an example to illustrate the algorithm's feasibility.

[0073] Figure 5 and Figure 6 The figure shows the waveform structure resolution performance of the Sig-I signal at signal-to-noise ratios of 4dB and 10dB, with 3 test frames for the Sig-I signal. As can be seen from the figure, the proposed algorithm achieves good waveform structure resolution performance at both 4dB and 10dB signal-to-noise ratios.

[0074] Figure 7 This demonstrates the performance of 100 Sig under a signal-to-noise ratio of 10 dB. The features of a single-frame signal after processing by a feature extraction network are used as t The distribution of each signal segment in the two-dimensional feature space is visualized using the SNE dimensionality reduction method. Based on the parameter settings mentioned earlier, the number of segments Q of the Sig-I signal is 3, and the number of segment groups is 2. As can be seen from the figure, the features of signals belonging to the same segment Seg1 and Seg3, after extraction by the segment-level representation model, are relatively clustered in the feature space, while the features corresponding to signals from different segments Seg1 and Seg2 exhibit significant distinguishability in the space.

[0075] 3.3 Analysis of Performance Comparison Experiment Results This section discusses Sig-I and Sig... II, Sig III and Sig IV. Four waveform structures were tested. The impact of signal-to-noise ratio, phase shift, and frequency shift on algorithm performance was analyzed through simulation experiments.

[0076] (1) Waveform structure analysis algorithm is affected by signal-to-noise ratio To evaluate the performance of the proposed waveform structure analysis algorithm under varying noise conditions, this section first conducts performance comparison experiments under different signal-to-noise ratio (SNR) conditions. Specifically, different SNR levels are set on four test signals from Sig-I to Sig-IV, and the algorithm is compared with various other methods. The curves of the three evaluation indicators, CS, ARI, and NMI, as a function of SNR are statistically analyzed, thus providing the overall performance comparison results of the proposed method under different noise levels.

[0077] The performance graphs of each algorithm for analyzing the waveform structure of the Sig-I signal under different signal-to-noise ratios are shown below. Figures 8 to 11 It can be seen that under different signal-to-noise ratio (SNR) conditions for the four sets of test signals, the three evaluation metrics exhibit a consistent performance ranking: the proposed method achieves optimal or near-optimal resolution results across all signal types and the entire SNR range, with particularly significant advantages under low to medium SNR conditions. Specifically, Sig-III, due to the use of the same modulation format for both functional and data segments, introduces statistical differences only through symbol set constraints, resulting in relatively weak inter-segment boundaries and a higher overall resolution difficulty compared to the other three signal sets. The performance metrics of all methods are generally lower, but DRCT-WS still maintains the best performance. With increasing SNR, the CS, ARI, and NMI metrics of DRCT-WS significantly improve and gradually stabilize in the medium-to-high SNR range, indicating that after the noise impact is reduced, this method can more fully recover the statistical differences and boundary structures of intra-frame segments.

[0078] The fundamental reason for this performance advantage lies in the fact that DRCT-WS, under the sliding window inference paradigm, does not rely solely on local transient morphological differences. Instead, it learns discriminative representations and jointly encodes local morphological cues with statistical stability features, thus maintaining high consistency even under strong noise conditions. In contrast, traditional segmentation methods mainly rely on statistical assumptions such as segment independence and covariance structure. When the signal exhibits significant non-stationary characteristics and there are receiver disturbances such as phase shift, frequency shift, and IQ imbalance, the bias of local statistics is significantly amplified, resulting in a lack of stable basis for change point detection and state division. Therefore, it performs poorly on more stringent consistency metrics such as ARI and NMI, and is not sensitive to improvements in signal-to-noise ratio.

[0079] Compared to the learned replacement baseline, ResNet18-WS achieves higher consistency at medium to high signal-to-noise ratios (SNR), but its improvement is limited in the low SNR range, showing a persistent gap with DRCT-WS. This indicates that while local features extracted using only convolutional residual networks possess some morphological representation capabilities, their ability to represent cross-temporal statistical dependencies and segment-level global patterns is insufficient due to limitations in their receptive field and local modeling mechanisms. In situations with strong noise or high modulation orders, the number of observable symbols within the sliding window is limited, making local features more prone to intra-class discrepancies and boundary blurring, thus affecting segmentation consistency. The performance ceiling of CNN-WS is even more pronounced because shallow or purely local convolutional representations rely more on local texture and short-term morphological differences, lacking an effective aggregation mechanism for structural features such as distribution differences under different modulation schemes and symbol sequence statistical periods, resulting in poor waveform structure resolution performance.

[0080] In summary, the test results not only demonstrate that DRCT-WS outperforms traditional segmentation methods and feature replacement baselines in terms of CS, ARI, and NMI, but also validate the source of its advantages at the mechanistic level: under conditions of limited sliding window observation and multiple receiver perturbations, by learning more discriminative and statistically stable waveform representations, DRCT-WS can more reliably distinguish statistical differences between different segments and maintain robustness in boundary discrimination. Especially in low to medium signal-to-noise ratio scenarios, DRCT-WS exhibits stronger noise immunity and superior waveform structure analysis performance, providing a more reliable performance guarantee for communication signal waveform structure analysis tasks.

[0081] (2) Waveform structure analysis algorithm is affected by frequency shift The robustness of the proposed waveform structure analysis algorithm is further evaluated from the perspective of frequency offset perturbation. Specifically, while keeping the signal-to-noise ratio fixed at 10dB, simulations are performed to generate signals with different degrees of frequency offset, with a normalized frequency offset range of [0, 0.01]. The changes in three indices, CS, ARI, and NMI, are statistically analyzed on four types of test signals from Sig-I to Sig-IV to analyze the impact of frequency offset on the consistency and stability of frame structure analysis.

[0082] The performance graphs of each algorithm on the Sig-II signal waveform structure analysis under different frequency deviations are shown below. Figures 12 to 15 .

[0083] Experimental results on frequency deviation robustness show that, at a signal-to-noise ratio of 10dB, the CS, ARI, and NMI curves of DRCT-WS and Resnet18-WS on the four signal classes Sig-I to Sig-IV are approximately horizontal with changes in frequency deviation, exhibiting only minor fluctuations. This indicates that both methods are highly robust to the phase rotation introduced by the normalized frequency deviation over time. In contrast, the overall performance of Cnn-WS is significantly lower and more volatile. Traditional methods consistently maintain low performance across all signals, showing little improvement with frequency deviation, indicating their inability to maintain stable segmentation and clustering consistency under frequency deviation disturbances. From a signal structure perspective, DRCT-WS and Resnet18-WS maintain consistent performance levels in Sig-I, Sig-II, and Sig-IV. However, in the Sig-III scenario, the overall task difficulty increases, and the differences between methods become more pronounced. Nevertheless, DRCT-WS still maintains the highest ARI / NMI with a stable curve, demonstrating its stronger ability to identify differences in weaker regions.

[0084] From a mechanistic perspective, frequency deviation causes linear phase shift of the signal over time, thus weakening the segment discrimination stability based on local morphology or simple statistical assumptions. DRCT-WS maintains high and stable performance under various frequency deviation conditions, indicating that its learned feature representation is more robust to this type of phase evolution, maintaining the compactness of similar samples in the feature space and ensuring segment separability. In contrast, Cnn-WS relies more on local texture and short-term morphology, making it susceptible to phase drift perturbations and resulting in representational fluctuations. Traditional methods such as GGS / TICC exhibit more significant statistical bias under non-stationary waveforms and receiver mismatch conditions, thus limiting their overall analytical performance. In summary, this set of experiments further verifies the stable analytical capability of the proposed method under frequency deviation perturbations, especially maintaining more reliable consistency in the Sig-III scenario where segment differences are relatively weak.

Claims

1. A blind parsing method for the signal frame structure of a communication protocol based on contrastive learning, characterized in that, Includes the following steps: Step 1: A supervised contrastive learning framework is adopted to train two independent branches: an instance-level feature extraction network and a segment-level feature extraction network. The instance-level and segment-level feature extraction networks share the topology of the DRCT unified feature extraction backbone network, and their parameter sets are independent and not shared during the training phase. The instance-level feature extraction network takes a complete signal instance as input and learns the instance-level global modulation semantic representation. The segment-level feature extraction network takes a signal segment truncated by a sliding window as input and learns the segment-level local waveform representation. The contrastive learning framework introduces a waveform reconstruction enhancement strategy, which cuts the same signal instance into multiple segments, randomly shuffles the order, and reassembles them to construct semantically equivalent positive sample pairs. Through cross-scale alignment loss constraints, semantic consistency between the instance-level global modulation semantic representation and the segment-level local waveform representation is established, resulting in a DRCT unified feature extraction backbone network with noise resistance and phase perturbation resistance. After training, only the network weights for fragment-level feature extraction are retained for subsequent inference; Step 2: Input the signal to be tested into the trained segment-level feature extraction network, use sliding window scanning to extract feature sequences, calculate the cosine similarity between adjacent features and align and fuse them along the time axis into a global similarity sequence, and obtain the global difference trajectory based on this. Candidate boundaries are detected by adaptive thresholding, and the signal is initially divided into multiple quasi-stationary short segments. Step 3: Calculate the feature mean of each quasi-stationary short segment as the central representation of the segment, perform unsupervised clustering on the set of central representations composed of all central representations, assign semantic labels to each short segment, and identify the functional segment and data segment categories. Step 4: Traverse the semantic tags of adjacent short segments. If adjacent tags are the same, they are identified as redundant boundaries and removed. Similar segments are merged into a complete physical region. If adjacent tags are different, they are identified as real frame structure state switching points and retained. Output the final set of frame boundaries and segment group categories to complete the signal frame structure parsing.

2. The method according to claim 1, characterized in that, The unified feature extraction backbone network for DRCT described in step 1 adopts a deep residual convolutional transform model. The instance-level feature extraction network and the segment-level feature extraction network are two instantiations of this unified feature extraction backbone network: they use the same network topology and module configuration, differing only in the observation scale of the input signal. The instance-level feature extraction network takes a complete signal instance as input, while the segment-level feature extraction network takes a local signal segment captured by a sliding window as input. During the training phase, their parameter sets are independent and not shared; during the inference phase, only the segment-level feature extraction network is called. The unified feature extraction backbone network for DRCT consists of three cascaded modules: Waveform encoding module: A 9-layer fully convolutional network is used to extract hierarchical features from the input complex baseband IQ signal. The terminal global average pooling layer is removed to preserve the temporal resolution. High-dimensional local feature sequences are output through residual connections and group normalization. Global temporal modeling module: Sine-cosine position codes are superimposed on the high-dimensional local feature sequence and input to a module consisting of L layers of Transformer encoders. The module models long-range statistical dependence and global modulation law through a multi-head self-attention mechanism. Feature mapping module: The global description output by the global temporal modeling module is mapped to a low-dimensional embedding space through the projection head, and then constrained to the unit hypersphere by L2 normalization, so that the similarity between samples is uniquely determined by the angle between the feature vectors.

3. The method according to claim 2, characterized in that, In step 1, the supervised contrastive learning framework is optimized using a joint loss function that includes the following three terms: Global scale contrast loss: Align the original signal instance, phase shift enhancement instance, and waveform reconstruction enhancement instance under the same supervision to reduce the impact of phase uncertainty and segment rearrangement on the instance-level global modulation semantic representation; Describe the instance-level features of the original signal samples Two types of enhancements are constructed for each sample: phase shift enhancement produces... Group features Waveform reconstruction enhancement Group features The global scale contrast loss is defined as: Local scale contrast loss: It brings the features of segments at different time locations within the same signal instance closer together under the supervision of the same type, and constrains their consistency with the corresponding segments of the enhanced samples, thereby improving the stability of segment-level representations to positional changes and phase perturbations. Record the original sample number The characteristics of each segment are: Phase shift enhancement sample The enhancement in the first The characteristics of each segment are: Accordingly, define the set of fragments within a batch. Local scale contrast loss, using the first segment of each sample as the anchor reference, is defined as: Cross-scale alignment loss: Under supervised constraints, semantic consistency is established between instance-level global modulation semantic representation and fragment-level local features, so that local fragment discrimination information is corrected under instance-level global semantic reference; Record No. The global features of each phase shift enhancement instance are: and with Indicates the first A set of fragments, then 。 4. The method according to claim 3, characterized in that, The phase offset enhancement is achieved by applying an additional global phase rotation to the received signal to simulate phase and frequency offset disturbances in the actual channel; the waveform reconstruction enhancement is specifically achieved by cutting the original waveform sequence into multiple segments, randomly shuffling their order, and then reassembling them; the original signal instance, the phase offset enhancement instance, and the waveform reconstruction enhancement instance all participate in supervised comparative optimization.

5. The method according to claim 1, characterized in that, The input signals of the instance-level feature extraction network include original signal instances, phase-shift enhancement instances, waveform reconstruction enhancement instances, similar signal instances, and dissimilar signal instances. The input signals of the segment-level feature extraction network include signal segments extracted from the above instances using a sliding window. The signals are defined as follows: Example of raw signal: The raw complex baseband IQ signal without enhancement processing, used as a benchmark sample for contrastive learning; Phase offset enhancement example: A sample generated by applying additional global phase rotation to the original signal example, used to simulate phase and frequency offset disturbances in a real channel; Waveform reconstruction enhancement example: The original signal instance is cut into multiple segments, the order is randomly shuffled and reassembled to generate samples, which are used to construct semantically equivalent positive sample pairs; Similar signal instances: Signal samples with the same modulation scheme as the original signal instances are used as positive samples for comparative learning within the batch; Outlier signal instances: Signal samples with different modulation schemes from the original signal instances are used as negative samples in the batch for comparative learning. The global scale contrast loss, local scale contrast loss, and cross-scale alignment loss are all constructed based on the similarity calculation between the above signal instances.

6. The method according to claim 1, characterized in that, The multi-starting-point parallel scanning mechanism described in step 2 specifically includes: pre-setting a set of different starting offset indices and performing sliding window feature extraction in parallel; for each starting offset, calculating the cosine similarity between adjacent features in the feature sequence to obtain a similarity sequence under each scanning path; aligning the similarity sequences of each path along the time axis and taking the average to merge them into a global similarity sequence; and calculating the global dissimilarity. , For the global difference trajectory, For global similarity sequences; set a threshold. If the position is a discrete point with a value greater than the threshold, it is directly extracted as the initial boundary; if it is a continuous high-difference region, the local maximum value of the region is extracted as the initial boundary, and the initial segment boundary set is finally determined. It is the maximum value of the global difference trajectory G over the entire sequence of signals to be tested.

7. The method according to claim 1, characterized in that, Step 3 specifically includes: for each sub-segment interval divided by the initial boundary, calculate the statistical mean of all feature vectors in the interval as the central representation of the segment; use the variational Dirichlet process Gaussian mixture model to cluster the central representation set and output the semantic label sequence of each segment, where different labels correspond to functional segments or data segments.

8. The method according to claim 1, characterized in that, The structural reverse optimization described in step 4 specifically involves traversing adjacent short segments in chronological order. and and its semantic tags and If adjacent labels are the same If two segments are determined to belong to the same physical region, the dividing point is removed. And merge the two segments; if adjacent labels are different, ≠ The points that are identified as real frame structure state transition points are retained; the final output is a set of precise frame boundaries. Corresponding segment group category.

9. The method according to claim 1, characterized in that, The signal under test is a single-carrier communication signal. The frame contains functional segments and data segments with different modulation methods, different symbol set constraints, or different statistical characteristics. The parsing process does not require the demodulated bit-level sequence, protocol prior information, or synchronization sequence prior.

10. The method according to claim 1, characterized in that, The dataset used in the training phase of step 1 includes hardware impairment simulations of IQ amplitude-phase imbalance and IQ phase quadrature error. The IQ amplitude-phase imbalance factor ranges from [-10, 10] dB, and the IQ phase quadrature error ranges from [-10, 10] degrees. Combined with additive white Gaussian noise, multipath fading, carrier frequency offset, and initial phase offset, a training environment under non-ideal reception conditions is constructed.

Images