Short wave signal collection, detection and identification and model updating method, system and medium

Abstract

This invention discloses a method, system, and medium for shortwave signal acquisition, detection, identification, and model updating, belonging to the field of communication and signal processing technology. It includes the following steps: shortwave signal acquisition and adaptive preprocessing; signal detection and dual-layer identification; logical consistency verification and comprehensive credibility calculation; difficult sample screening, processing, and training dataset construction; incremental model updating and online hot deployment. By utilizing the design and coordination of corresponding processing strategies in each step, a dual closed-loop process of data and strategy—"acquisition-detection-identification-evaluation-screening-updating-deployment"—is ultimately formed. The method, system, and medium of this invention can solve the problems of performance degradation under complex channel conditions, difficulty in identifying unknown protocols, inability to quantify the credibility of identification results, reliance on manual model updates, and lack of data and engineering closed loops in existing shortwave signal identification systems, achieving continuous adaptive evolution of the shortwave communication signal identification model.

Description

Technical Field

[0001] This invention belongs to the field of communication and signal processing technology, specifically relating to shortwave signal acquisition, detection and identification, and model updating methods, systems, and media. Background Technology

[0002] In the field of radio communication technology, shortwave communication has significant application value in military, emergency, maritime, and aviation communications due to its advantages such as long propagation distance, strong diffraction capability, and independence from satellite links. Under ionospheric reflection conditions, shortwave signals can achieve intercontinental communication. However, ionospheric propagation is affected by factors such as solar activity, geomagnetic disturbances, and terrain obstruction, resulting in channel characteristics that are highly time-varying, exhibiting multipath effects and frequency-selective fading, increasing the difficulty of shortwave signal detection and identification.

[0003] Traditional shortwave signal detection and identification methods are mostly based on techniques such as energy detection, cyclostationary feature analysis, or template matching. These methods typically rely on fixed rules or feature extraction. While they can meet the requirements for shortwave signal detection and identification to a certain extent, they are prone to problems such as detection errors, decreased recognition rates, or protocol parsing failures in complex and changing electromagnetic environments. With the increasing number of shortwave communication protocol types and the evolution of modulation methods, identification methods based on fixed rules are becoming increasingly unable to cover protocol and standard changes.

[0004] In recent years, deep learning and pattern recognition technologies have been introduced into shortwave signal recognition, extracting signal features through convolutional networks, recurrent networks, or time-frequency graph recognition to improve recognition performance. However, research and application have revealed that these methods still have many shortcomings; for example, model training is mostly conducted offline, making it difficult to adapt to channel changes after deployment; uncovered protocols and variant protocols are difficult to identify; recognition results lack a reliability evaluation mechanism; and model updates lack data feedback and system triggering mechanisms.

[0005] Furthermore, existing shortwave recognition systems often process detection, recognition, and model training modules independently, lacking data management and a closed-loop strategy. When faced with complex electromagnetic environment changes or the emergence of new protocols, the system's recognition performance is prone to decline, requiring manual intervention in sample collection, annotation, and model updates. This results in long update cycles and high maintenance costs, posing significant limitations. Summary of the Invention

[0006] In response to one or more of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a method, system and medium for shortwave signal acquisition, detection and identification and model updating, which can construct a closed-loop adaptive mechanism for shortwave communication, realize the continuous adaptive evolution of the shortwave communication signal identification model and improve the reliability of shortwave signal communication.

[0007] To achieve the above objectives, one aspect of the present invention provides a method for shortwave signal acquisition, detection, identification, and model updating, comprising the following steps: (1) Acquire shortwave signals and perform preprocessing to obtain feature data for identification and channel quality indicators for subsequent confidence calculation; (2) Perform signal presence detection and dual-layer parallel identification on the feature data, and output the identification results of physical layer modulation mode and protocol layer standard type respectively; (3) Based on the preset protocol specification library, the logical consistency of the two identification results is checked, and the comprehensive credibility is calculated by combining the channel quality index and the identification probability distribution; (4) Select samples that fail the logical consistency check or whose overall credibility is lower than the preset threshold as difficult samples, and perform clustering and semi-supervised pre-labeling on the difficult samples to construct an incremental training dataset; (5) Use the incremental training dataset to perform incremental updates on the recognition model, and then deploy the updated model online using the shadow verification mode; In the incremental update process, knowledge distillation constraints and consistency constraint loss functions based on protocol constraint relationships are introduced. The constraint relationship between modulation method and protocol type in the protocol specification library is encoded as training constraints to constrain the matching relationship between physical layer identification results and protocol layer identification results.

[0008] As a further improvement of the present invention, in step (1), the preprocessing includes at least one of radio frequency downconversion, analog-to-digital conversion, channel compensation, multi-domain feature extraction, feature fusion and dimensionality reduction.

[0009] As a further improvement of the present invention, the multi-domain feature extraction includes time domain features, frequency domain features, spatial domain features, and code domain features; and / or The feature fusion and dimensionality reduction process includes: normalizing the features of each domain, performing feature dimensionality reduction using principal component analysis or a deep learning encoder, and retaining the main feature components.

[0010] As a further improvement of the present invention, in step (2): The signal presence detection employs at least one of energy detection, cyclostationary feature detection, or deep learning detection methods, and outputs a signal presence flag and detection confidence level. and / or The dual-layer parallel recognition is based on two parallel branches: physical layer modulation recognition and protocol layer standard recognition. The two branches share the underlying feature extraction network and output the physical layer modulation method recognition result and the protocol layer standard type recognition result respectively through their respective classification heads.

[0011] As a further improvement of the present invention, in step (3), the preset protocol specification library includes: The correspondence between modulation and protocol is constrained, defining the set of allowed modulation methods for each protocol type; Parameter range constraints define the allowed range of parameters for each protocol type. Frame structure constraints define the frame structure characteristics of each protocol type; and The logical consistency verification process includes: querying the protocol specification library and obtaining constraints; checking whether the physical layer modulation type conforms to the set of modulation methods allowed by the protocol layer; checking whether the parameters are within the range defined in the protocol specification library; and calculating the overall consistency score.

[0012] As a further improvement of the present invention, the calculation of the overall confidence level is output through a constructed multidimensional confidence assessment matrix, which includes the following dimensions: The channel quality dimension calculates a channel quality score based on channel metrics, which include at least signal-to-noise ratio, multipath parameters, and bandwidth consistency. The identification probability dimension is determined by calculating the identification deterministic score based on the maximum probability distribution and probability entropy of the physical layer and protocol layer identification. In terms of consistency, a consistency score is calculated based on a comprehensive consistency score derived from logical consistency verification. The overall credibility score is obtained by weighted fusion of scores from each dimension, with the weights either set to a fixed preset weight or dynamically adjusted based on the reliability of each dimension.

[0013] As a further improvement of the present invention, in step (4): The clustering process employs density-based clustering, hierarchical clustering, or deep feature-based clustering methods; difficult samples are divided into several similarity clusters based on signal characteristic similarity, which facilitates batch pre-labeling and discovery of new protocol categories; and the clustering feature space uses dimensionality-reduced features or intermediate layer features of the recognition network. and / or The semi-supervised pre-labeling process includes: for samples in a cluster whose similarity to historical knowledge base samples is higher than a threshold, inheriting the labels of historical samples; for newly emerging clusters, marking them as potential new categories, triggering manual assisted labeling or automatically assigning temporary labels; for sample groups with high consistency within a cluster, generating pseudo-labels using majority voting or confidence weighting.

[0014] As a further improvement of the present invention, in step (5): The knowledge distillation mechanism includes: using the old model as the teacher model to generate soft labels for training samples, where the soft labels reflect the probability distribution of the old model for each category; the training loss function of the new model includes a weighted combination of hard label loss and soft label distillation loss, where the hard label loss uses cross-entropy loss and the soft label distillation loss uses KL divergence or mean squared error; for samples of new categories, only hard label loss is used; for samples of known categories, both hard label loss and soft label distillation loss are used. The consistency constraint loss function, together with the classification loss and distillation loss, constitutes a joint optimization objective function, which is used to achieve synergistic optimization of recognition accuracy and cross-layer consistency. The shadow verification mode includes: performing parallel inference between the updated model and the current online model for the same real-time signal, and comparing their performance metrics; The online hot deployment process of the updated model includes: deploying the updated model in shadow mode and performing parallel inference on real-time signals with the current online model; collecting and comparing the key performance indicators of the two models within a preset verification period; and performing an online hot switch when the updated model meets the preset conditions in terms of hard sample recognition performance and overall stability, making the updated model the master model and the original master model downgraded to a backup model or retired.

[0015] Another aspect of the present invention provides a shortwave signal acquisition, detection, identification, and model update system, comprising: The acquisition and preprocessing module is used to acquire shortwave signals and perform preprocessing to obtain feature data for identification and channel quality indicators for subsequent confidence calculation. The detection module is used to perform signal presence detection and time-frequency localization on the feature data, and output the detection results and detection confidence level; The dual-layer recognition module is used to perform dual-layer parallel recognition on the feature data to output the physical layer modulation mode recognition result and the protocol layer standard type recognition result. The consistency verification module stores a predefined protocol specification library, which is used to perform logical consistency verification on the dual-layer identification results and output a consistency score. The confidence assessment module is used to construct a multi-dimensional confidence assessment matrix, integrate channel quality indicators, identification probability distribution and consistency verification results, and output a comprehensive confidence score. The sample management module is used to manage difficult samples, perform clustering and semi-supervised pre-labeling, and generate training datasets. The model update module is used to perform incremental learning and knowledge distillation training, initiate the model update process according to the trigger conditions, and generate the updated recognition model. The shadow verification and deployment module is used to manage parallel inference of shadow models, compare performance metrics, and perform online hot switching or rollback operations.

[0016] In another aspect, the present invention also provides a storage medium storing a processor-executable program, which, when executed by the processor, is used to perform the aforementioned shortwave signal acquisition, detection and identification, and model update method.

[0017] The aforementioned improved technical features can be combined with each other as long as they do not conflict with each other.

[0018] In summary, the beneficial effects of the above-described technical solutions conceived by this invention compared with the prior art include: The shortwave signal acquisition, detection, identification, and model update method of the present invention includes the following steps: shortwave signal acquisition and adaptive preprocessing; signal detection and dual-layer identification; logical consistency verification and comprehensive credibility calculation; difficult sample screening, processing, and training dataset construction; incremental model update and online hot deployment. By utilizing the design and coordination of corresponding processing strategies in each step, a dual closed-loop process of data and strategy, namely "acquisition-detection-identification-evaluation-screening-update-deployment", is finally formed. This solves the problems of recognition performance degradation under complex channel conditions, difficulty in identifying unknown protocols, inability to quantify the credibility of recognition results, reliance on manual model updates, and lack of data and engineering closed loops in existing shortwave signal identification systems, thereby realizing the continuous adaptive evolution of the shortwave communication signal identification model.

[0019] By utilizing the shortwave signal acquisition, detection and identification and model update method and system settings in this invention, the shortwave signal identification system has the following capabilities: (1) maintaining stable identification performance with changes in the shortwave electromagnetic environment; (2) having adaptive processing capabilities when facing unknown protocols or variant protocols; (3) improving the reliability of the identification link through trusted quantification; and (4) having an update mechanism with low manual dependence to meet engineering deployment requirements. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the shortwave signal acquisition, detection and identification, and model update method in an embodiment of the present invention; Figure 2 This is an architecture diagram of the shortwave signal acquisition, detection, identification, and model update system in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the model update and deployment module of the system in this embodiment of the invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0023] In the description of this invention, it should be understood that, unless otherwise expressly specified and limited, the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," "circumferential," etc., indicating the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, are only for the convenience of describing this invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0024] Furthermore, unless otherwise expressly defined, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise expressly and specifically defined.

[0025] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0026] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0027] Below, for reference Figures 1-3 The present invention describes a method, system, and medium for shortwave signal acquisition, detection, identification, and model updating according to a preferred embodiment, and elaborates on the following three embodiments.

[0028] Example 1: As one aspect of the present invention, a method for shortwave signal acquisition, detection and identification and model updating is provided, which aims to realize intelligent identification of shortwave signals that supports adaptive model evolution. The whole process includes signal acquisition from the front end, detection, identification and evaluation, and then sample management and model updating. The verified model is deployed online through the deployment module to realize long-term adaptive operation of the model.

[0029] Specifically, such as Figure 1 As shown in the preferred embodiment, the shortwave signal acquisition, detection, identification, and model update method includes the following steps: (1) Shortwave signal acquisition and adaptive preprocessing Shortwave signals are acquired and preprocessed to obtain feature data for identification and channel quality indicators for subsequent confidence calculation.

[0030] Specifically, radio frequency signals are acquired through a broadband shortwave receiving front-end, and then the acquired shortwave signals are preprocessed. In a preferred embodiment, the preprocessing process includes at least one of the following sequentially executed steps: radio frequency down-conversion, analog-to-digital conversion, channel compensation, multi-domain feature extraction, feature fusion, and dimensionality reduction. After the preprocessing steps are performed, a multi-source feature stream and channel quality indicators are output for subsequent detection and identification.

[0031] More specifically, the aforementioned RF downconversion includes shifting the target frequency band to zero intermediate frequency or low intermediate frequency; channel compensation includes Doppler frequency offset compensation and multipath interference suppression, wherein multipath interference suppression preferably employs an adaptive equalizer, and the number of equalizer taps is adaptively adjusted according to the multipath delay spread.

[0032] Meanwhile, the preferred methods for multi-domain feature extraction in preprocessing include: Time-domain feature extraction: Extract instantaneous amplitude, instantaneous phase, and instantaneous frequency sequences, and calculate higher-order cumulants and cyclic autocorrelation functions; Frequency domain feature extraction: Perform fast Fourier transform on the signal to extract features such as power spectral density, peak distribution, occupied bandwidth, and spectral entropy; Spatial domain feature extraction: When using a multi-antenna system, the corresponding signal angle of arrival is extracted (estimated); Code domain feature extraction: For spread spectrum signals, capture pseudo-random codes and estimate chip rate and spreading gain.

[0033] Furthermore, for feature fusion and dimensionality reduction, the preferred processing steps include: normalizing the features of each domain, using principal component analysis or a deep learning encoder to reduce the dimensionality of the features, and retaining the main feature components.

[0034] Furthermore, the preprocessed data, namely the feature data (multi-source feature stream) and channel quality indicators used for identification, are preferably passed to the corresponding detection and identification modules in the form of a data structure for subsequent detection and identification. Among them, the channel quality indicators, as input data for subsequent confidence assessment, preferably include parameters such as signal-to-noise ratio, signal-to-interference-plus-noise ratio, and multipath delay spread.

[0035] (2) Signal detection and two-layer recognition The feature data is subjected to signal presence detection and dual-layer parallel identification, and the physical layer modulation mode identification result and the protocol layer standard type identification result are output respectively.

[0036] Specifically, by invoking the detection module, signal presence detection and time-frequency localization can be performed on the multi-dimensional feature stream (i.e., feature data). Signal presence detection preferably employs at least one of energy detection, cyclostationary feature detection, or deep learning detection methods, and outputs a signal presence flag and detection confidence level accordingly. Time-frequency localization preferably determines the start and end times of the signal through the energy envelope and determines the center frequency and occupied bandwidth through spectral peak search.

[0037] Meanwhile, utilizing the dual-layer recognition function design, a recognition module is configured with two parallel branches: physical layer modulation recognition and protocol layer standard recognition. The two branches share the underlying feature extraction network and output the physical layer modulation method recognition result and the protocol layer standard type recognition result respectively through their respective classification heads.

[0038] More specifically, the physical layer modulation identification results preferably include the output modulation type (including analog modulation and digital modulation), modulation identification probability distribution, and probability entropy; the protocol layer standard type identification results preferably include the output protocol type, protocol identification probability distribution, and probability entropy.

[0039] (3) Logical consistency verification and comprehensive credibility calculation Based on a pre-defined protocol specification library, the logical consistency of the two identification results is verified, and the overall credibility is calculated by combining channel quality indicators and identification probability distribution.

[0040] Specifically, in practice, logical consistency verification is preferably implemented through the aforementioned identification module. Correspondingly, the two-layer identification module is preferably trained using a multi-task learning framework, with the loss function including physical layer identification loss, protocol layer identification loss, and consistency constraint loss. The two-layer identification module can output a structured identification report, which further includes detection results, classification probability distribution, physical layer identification results, protocol layer identification results, consistency verification results, and timestamps.

[0041] Of course, it is understandable that a separate verification module can be set up to perform logical consistency checks, which will not be elaborated here.

[0042] More specifically, in the preferred embodiment, the logical consistency check is performed based on a predefined protocol specification library, which further includes: The correspondence between modulation and protocol is constrained, defining the set of allowed modulation methods for each protocol type; Parameter range constraints define the allowed range of parameters for each protocol type, such as the allowed range of parameters like symbol rate and bandwidth usage. Frame structure constraints define the frame structure characteristics of each protocol type, such as preamble length, synchronization word mode, and other frame structure characteristics.

[0043] Based on the aforementioned protocol specification library, the logical consistency verification process in the preferred embodiment specifically includes the following: Query the protocol specification library and obtain the constraints; check whether the physical layer modulation type conforms to the set of modulation methods allowed by the protocol layer; check whether the estimated parameters are within the range defined in the protocol specification library; calculate the overall consistency score.

[0044] More specifically, based on the obtained logical consistency verification results, classification probability distribution, and channel quality indicators, a multi-dimensional confidence evaluation matrix can be constructed, thereby outputting the comprehensive confidence score corresponding to the current recognition result.

[0045] More specifically, the multidimensional confidence assessment matrix in the preferred embodiment preferably includes the following dimensions: The channel quality dimension calculates the channel quality score based on channel metrics, which include at least the signal-to-noise ratio, multipath parameters, and bandwidth consistency. The identification probability dimension is determined by calculating the identification deterministic score based on the maximum probability distribution and probability entropy of the physical layer and protocol layer identification. In terms of consistency, a consistency score is calculated based on a comprehensive consistency score derived from logical consistency verification. Accordingly, the overall credibility score is obtained by weighted fusion of scores from each dimension, with the weights either set to a fixed preset weight or dynamically adjusted based on the reliability of each dimension.

[0046] (4) Selection, processing and construction of training dataset for difficult samples Samples that fail the logical consistency check or whose overall credibility is below a preset threshold are selected as difficult samples. These difficult samples are then clustered and semi-supervised pre-labeled to construct an incremental training dataset.

[0047] Specifically, through the comprehensive credibility calculation in the aforementioned process (3), the samples can be classified and labeled. At this time, a corresponding comparison threshold can be preset. For example: Set the first judgment threshold Second determination threshold ,in, When the overall credibility score is lower than the first judgment threshold When the overall credibility score is below the second judgment threshold, the sample is marked as a low-confidence sample; when the overall credibility score is below the second judgment threshold, the sample is marked as a low-confidence sample. When this happens, the sample is marked as a difficult sample.

[0048] In actual screening, the screening conditions for difficult samples are further optimized as follows: the overall credibility score is lower than the preset threshold; the maximum recognition probability value of the physical layer or protocol layer (obtained from the recognition probability distribution) is lower than the uncertainty threshold; the logical consistency verification score is lower than the consistency threshold; and the recognition result of the physical layer or protocol layer is an unknown category.

[0049] Meanwhile, the processing of difficult samples mainly includes clustering and semi-supervised pre-labeling. Among them, clustering is based on clustering algorithms, which use density-based clustering, hierarchical clustering, or deep feature-based clustering methods to divide difficult samples into several similarity clusters according to the similarity of signal characteristics, thereby facilitating batch pre-labeling and the discovery of new protocol categories.

[0050] More specifically, the clustering feature space uses the dimensionality-reduced features extracted in step (1) or the intermediate layer features of the recognition network.

[0051] Furthermore, the preferred semi-supervised pre-labeling process for difficult samples includes: for samples in a cluster whose similarity to historical knowledge base samples is higher than a threshold, inheriting the labels of historical samples; for newly emerging clusters, marking them as potential new categories, triggering manual assisted labeling or automatically assigning temporary labels; and for sample groups with high consistency within a cluster, generating pseudo-labels using majority voting or confidence weighting.

[0052] Through the aforementioned clustering and semi-supervised pre-labeling processes, an incremental training dataset (to be trained) can be output. At this point, the training dataset specifically includes hard sample features, pre-labeled labels, confidence scores, and other content.

[0053] (5) Incremental model update and online hot deployment Incremental updates to the recognition model are performed using an incremental training dataset, and the updated model is then hot-deployed online using a shadow verification mode. During the incremental update process, knowledge distillation constraints and a consistency constraint mechanism based on protocol constraints are introduced. The constraint relationship between modulation methods and protocol types in the protocol specification library is encoded as training constraints to constrain the matching relationship between the physical layer recognition results and the protocol layer recognition results.

[0054] Specifically, this step mainly involves two processes: "incremental update of the model" and "online deployment of the model," among which: When the number of difficult samples reaches the update trigger threshold or the recognition performance degrades beyond a preset range, the incremental update process of the model is initiated. An incremental learning strategy is used to update and train the recognition module. Knowledge distillation constraints are introduced during the training process, and the updated model is obtained accordingly.

[0055] After completing the incremental update of the model, the online hot deployment process of the model is then performed, that is: The updated model is deployed in shadow mode and used for parallel inference of real-time signals with the current online model. Within a preset verification period, key performance indicators of the two models are collected and compared. When the updated model meets the preset conditions in terms of hard sample recognition performance and overall stability, an online hot switch is performed, making the updated model the master model and the original master model downgraded to a backup model or retired.

[0056] In the actual design, further details were designed for incremental updates and online deployment of the model, as detailed below: First, the preferred triggering conditions for the aforementioned incremental updates include: the number of samples in the hard sample repository exceeds a preset threshold; the recent recognition accuracy has decreased by more than a preset percentage compared to the baseline; a new cluster has emerged and the number of samples is sufficient; and periodic updates are triggered regularly.

[0057] Meanwhile, the preferred incremental learning strategies include: Construct the training dataset: Mix the pre-labeled samples in the hard sample repository with the historical training samples in a proportional manner, and make the weight of hard samples higher than the weight of historical samples; Initialize a new model: Initialize based on the current model parameters, or add new category output nodes by expanding the network structure.

[0058] Furthermore, the preferred knowledge distillation mechanism in the incremental update process of the model includes the following: The old model is used as the teacher model, and soft labels are generated for the training samples. The soft labels reflect the probability distribution of the old model for each category. The training loss function of the new model consists of a weighted combination of hard label loss and soft label distillation loss, where the hard label loss uses cross-entropy loss and the soft label distillation loss uses KL divergence or mean squared error. For samples of new categories, only hard label loss is used; for samples of known categories, both hard label loss and soft label distillation loss are used.

[0059] Meanwhile, during model training (incremental update), a consistency constraint loss function is preferably introduced to encode the modulation-protocol constraint relationship in the protocol specification library into a computable constraint loss term, which is used to constrain the matching relationship between the physical layer modulation identification result and the protocol layer identification result.

[0060] More specifically, the consistency constraint loss function is used to penalize recognition results that do not meet the protocol constraints. It, together with the classification loss and distillation loss, constitutes a joint optimization objective function, which is used to guide the model to form cross-layer consistent representations during feature learning, thereby improving the logical consistency and robustness of the recognition results.

[0061] By introducing a knowledge distillation mechanism during training, the old model provides soft label constraints for the new model, thereby enhancing the performance of new category recognition while avoiding catastrophic forgetting of the original category recognition ability, thus achieving forgetting-resistant training of the model.

[0062] More specifically, the anti-forgetting training strategy for the model preferably further includes: Constrain the recognition probability distribution of historical samples to prevent the performance of the new model from degrading on the original category; at the same time, use methods such as elastic weight consolidation or experience replay to protect important network parameters.

[0063] Furthermore, the preferred termination condition for model training is that the overall performance index on the validation set converges, or the preset maximum number of training rounds is reached.

[0064] Furthermore, for the online deployment of the incrementally updated model, it is preferable to deploy it in shadow mode. In this case, the updated model and the current model run separately and execute the following configuration: The updated model and the current model simultaneously receive real-time signal data streams; the two models independently execute the detection, recognition, and confidence assessment processes; the system only outputs the recognition results of the current main model to the outside world, and the results of the shadow model are only used for internal verification.

[0065] More specifically, for the updated model, the preferred online hot deployment process includes: (a) Deploy the updated model in shadow mode and infer in parallel with the current online model on real-time signals; (b) Within a preset verification period, collect and compare the key performance indicators of the two models; wherein, the key performance indicators collected and compared preferably include: Overall recognition accuracy: The accuracy rate of recognition on the entire sample set; Hard sample recognition performance: recognition accuracy or F1 score on hard sample sets; Confidence calibration: The consistency between the predicted confidence level and the actual accuracy; Inference latency: the average processing time for a single sample; Stability metric: The magnitude of performance fluctuation within a continuous time window.

[0066] (c) When the updated model meets the preset conditions in both hard sample recognition performance and overall stability, an online hot-swap is performed, making the updated model the primary model, and the original primary model is downgraded to a backup model or decommissioned. The hot-swap conditions and procedures are as follows: Hot-switch conditions include: The updated model's recognition performance on difficult sample sets is at least a predetermined improvement over the current model; The updated model's recognition accuracy on the full sample set is no lower than that of the current model; The inference latency of the updated model meets real-time requirements; The updated model maintained stable performance during the validation period, with fluctuations within an acceptable range.

[0067] The hot-swapping process includes: Stop updating the weights of the current main model and freeze its state as a backup; Switch the shadow model to the main model and start outputting recognition results. Retain the backup model for a period of time, and perform a rollback operation if an anomaly is found. Record the switchover log, including switchover time, performance comparison data, triggering conditions, and other information.

[0068] After completing the online hot deployment of the updated model, the complete closed-loop process from shortwave signal acquisition to model adaptive update is completed. Afterwards, step (1) can be returned to continue processing new signal data streams, forming a continuously evolving adaptive recognition process.

[0069] Of course, if the updated model does not meet the hot switching conditions, the update is abandoned and the reason for failure is recorded. The incremental training strategy for the next round is then adjusted based on the reason for failure.

[0070] Example 2: As another aspect of the present invention, a shortwave signal acquisition, detection, identification, and model updating system is also provided, the system architecture of which is as follows: Figure 2 As shown, and whose design logic corresponds to the aforementioned methods, it mainly includes: The acquisition and preprocessing module is used to acquire shortwave signals and perform preprocessing to obtain feature data for identification and channel quality indicators for subsequent confidence calculation. The detection module is used to perform signal presence detection and time-frequency localization on the feature data, and output the detection results and detection confidence level; The dual-layer recognition module is used to perform dual-layer parallel recognition on the feature data to output the physical layer modulation mode recognition result and the protocol layer standard type recognition result. The consistency verification module stores a predefined protocol specification library, which is used to perform logical consistency verification on the dual-layer identification results and output a consistency score. The confidence assessment module is used to construct a multi-dimensional confidence assessment matrix, integrate channel quality indicators, identification probability distribution and consistency verification results, and output a comprehensive confidence score. The sample management module is used to manage difficult samples, perform clustering and semi-supervised pre-labeling, and generate training datasets. The model update module is used to perform incremental learning and knowledge distillation training, initiate the model update process according to the trigger conditions, and generate the updated recognition model. The shadow verification and deployment module is used to manage parallel inference of shadow models, compare performance metrics, and perform online hot switching or rollback operations.

[0071] Through the collaborative setup of the above modules, the settings of each module correspond to the corresponding processing procedures in each step of the aforementioned method. In actual setup, the modules are preferably connected through standardized data interfaces, and data is transferred using message queues or shared memory, forming a closed-loop, self-evolving process of "detection-identification-evaluation-screening-update-deployment" to achieve continuous engineering-level adaptation of the identification model.

[0072] In practical setups, the acquisition and preprocessing module, detection module, and dual-layer recognition module can be deployed on a digital signal processor, field-programmable gate array, or graphics processor. The sample management module, model update module, and shadow verification and deployment module can be deployed on a general-purpose server or cloud computing platform.

[0073] Furthermore, in order to better illustrate the design functions and implementation details of each module, the details of each module are described in detail below.

[0074] The acquisition and preprocessing module is mainly used to receive, frequency convert, sample, extract features, and fuse features of shortwave radio frequency signals. This module can be deployed in a digital signal processor, field-programmable gate array, graphics processor, or general-purpose processor. Its input is the shortwave radio frequency signal, and its output is a multi-source feature stream (i.e., feature data) and channel quality parameters (i.e., channel quality indicators).

[0075] More specifically, the acquisition and preprocessing module preferably includes a radio frequency front-end unit, an intermediate frequency conversion unit, an analog-to-digital conversion unit, a channel compensation unit, and a feature extraction unit, wherein: The radio frequency front-end unit is used to receive shortwave radio frequency signals and perform pre-selection filtering, gain control and preliminary amplification to improve signal quality and suppress out-of-band interference; The intermediate frequency conversion unit is used to downconvert shortwave radio frequency signals to zero intermediate frequency or the intermediate frequency bandwidth range for easier subsequent digital processing. The analog-to-digital conversion unit is used to sample and digitally store the intermediate frequency signal. The sampling rate is set according to the target signal bandwidth and the Nyquist criterion. The channel compensation unit is used to perform Doppler frequency offset compensation, multipath suppression and normalization processing on the digital signal to reduce the impact of the channel on subsequent identification performance; The feature extraction unit is used to perform feature extraction on a digital signal in at least one domain, including the time domain, frequency domain, time-frequency domain, spatial domain, or code domain.

[0076] In an optional embodiment of the present invention, the extraction method for different domains is preferably as follows: (a) Time domain feature extraction, the extracted features preferably include instantaneous amplitude, instantaneous phase, instantaneous frequency, symbol energy sequence, higher-order cumulants, etc.; (b) Frequency domain feature extraction, the extracted features preferably include the spectral envelope, power spectral density, bandwidth, spectral peak distribution, spectral entropy, etc. obtained by Fast Fourier Transform (FFT); (c) Time-frequency domain feature extraction, wherein the extracted features preferably include a two-dimensional time-frequency matrix obtained by short-time Fourier transform, Wigner-Ville distribution, Chirp Z transform or continuous wavelet transform; (d) Spatial domain feature extraction, mainly applied in multi-antenna systems, preferably obtained by obtaining the spatial spectrum of the signal through direction of arrival estimation; (e) Code domain feature extraction, mainly used in spread spectrum or frequency hopping communication, obtains code domain features through chip rate detection, pseudo-random sequence matching or frequency hopping sequence tracking.

[0077] In addition, the above-mentioned features are preferably scaled using a normalization algorithm, and feature dimensionality reduction is achieved through principal component analysis, linear discriminant analysis or deep autoencoder, thereby forming a data structure suitable for the input of the subsequent recognition module.

[0078] Furthermore, the acquisition and preprocessing module simultaneously calculates channel quality indicators, including but not limited to parameters such as signal-to-noise ratio, signal-to-interference-to-noise ratio, multipath delay spread, and phase offset, which are used by the subsequent confidence assessment module.

[0079] Correspondingly, the multi-source feature stream and channel quality parameters output by the acquisition and preprocessing module are transmitted to the detection and recognition modules through the internal bus or cache to complete the acquisition and preprocessing stage.

[0080] The detection module is mainly used to detect the presence of signals and locate the time-frequency signals from the multi-source feature streams and channel quality parameters output by the acquisition and preprocessing module, so as to determine the effective range and spectrum range of the shortwave communication signal.

[0081] In actual setup, the detection module can be deployed on a DSP, FPGA, or CPU. The input is a digital feature stream, and the output includes information such as signal presence flag, detection confidence level, start and end time, center frequency, and occupied bandwidth.

[0082] More specifically, the detection module in the preferred embodiment includes a signal detection unit, a time-domain positioning unit, and a frequency-domain positioning unit, wherein: The signal detection unit is used to determine whether the target signal exists, and the deployed detection method includes at least one of the following detection methods: (a) Energy detection method: The energy of the sampled sequence is accumulated and compared with the detection threshold obtained based on noise estimation to determine the presence of the signal and the detection confidence. (b) Cyclic stationary feature detection method: Perform cyclic autocorrelation or spectral correlation operations on the input signal sequence to detect whether there are cyclic stationary features consistent with the communication signal, so as to improve the detection reliability under low signal-to-noise ratio conditions; (c) Deep learning detection method: Input the time domain or time-frequency domain features into the detection neural network model, and output the signal existence probability or binary classification label. This model can be implemented by convolutional neural network, time-frequency graph detection network or time series network.

[0083] In actual operation, the signal presence indicator and detection confidence level output by the signal detection unit are transmitted to the time-domain positioning unit and the frequency-domain positioning unit.

[0084] For the time-domain localization unit, which is used to determine the start and end points of the signal on the time axis, the preferred methods are the time envelope method, the energy gradient method, or the model-based sequence slicing method. Wherein: The temporal envelope method calculates the sequence of the sliding energy window, detects energy abrupt change points, and uses them as candidate start and end points; the energy gradient method calculates the first difference of the energy sequence and detects its significant abrupt change positions; the model-based sequence slicing method identifies effective signal segments by training a sequence segmentation network.

[0085] In actual operation, the time-domain positioning unit outputs the start and end time information of the signal, which facilitates the identification module to extract the features of the corresponding time period.

[0086] For the frequency domain positioning unit, which is used to determine the center frequency and occupied bandwidth of the signal, the following method is preferred: (a) Spectrum peak search: Detect the main energy concentration area based on the power spectral density distribution, estimate the center frequency through the peak frequency point, and estimate the bandwidth through the energy proportion threshold method; (b) Spectral energy envelope analysis: The boundary positions of the energy distribution are determined by spectral envelope scanning, and the bandwidth range and bandwidth boundary points are obtained; (c) Time-frequency density matrix analysis: When the time-frequency matrix is ​​input, the center frequency and bandwidth information can be obtained through two-dimensional energy peak cluster detection.

[0087] In actual operation, the frequency domain positioning unit outputs the center frequency and occupied bandwidth to constrain the search range of subsequent identification modules.

[0088] By utilizing the design of the detection module, a structured output is ultimately generated. The output information includes signal presence flag, detection confidence level, signal start and end time, center frequency, occupied bandwidth, and timestamp information. This output is transmitted to the recognition module and confidence evaluation module in the form of messages or data structures for subsequent recognition and confidence calculation.

[0089] Furthermore, in the preferred embodiment, the identification module is used to perform physical layer modulation identification and protocol layer standard identification of the shortwave communication signal on the signal to be identified output by the detection module. Its output includes physical layer modulation identification results, protocol layer standard identification results and their probability distribution information. The output is transmitted to the consistency verification module and the confidence evaluation module in the form of structured data for subsequent verification and reliability calculation.

[0090] In actual setup, the recognition module is preferably deployed on a CPU, GPU or hardware platform with tensor operation acceleration capability. Its input is the multi-source feature stream (from the acquisition and preprocessing module) and the time-frequency positioning information output by the detection module. The output is the physical layer modulation type, protocol layer standard type and their probability distribution.

[0091] More specifically, in practical design, the recognition module is further optimized to include a feature normalization unit, a shared feature extraction unit, a physical layer recognition branch, and a protocol layer recognition branch, wherein: The feature normalization unit is used to standardize the multi-source feature stream output by the acquisition and preprocessing module, including amplitude normalization, mean-variance normalization, and feature scale unification, in order to adapt to the input requirements of the deep recognition model.

[0092] The shared feature extraction unit is used to extract deep features based on normalized features and output a general representation vector to support the shared feature space for physical layer and protocol layer recognition tasks.

[0093] In practical construction, the shared feature extraction unit preferably adopts at least one structure from convolutional neural networks, recurrent neural networks, Transformer networks, or multi-branch fusion networks, and different network structures can be selected according to the deployment hardware capabilities and signal characteristics. Meanwhile, the input method of the shared feature extraction unit preferably adopts any one of the following: a two-dimensional convolutional structure for processing time-frequency matrix input, a one-dimensional convolutional structure for processing time series input, or a self-attention network for processing feature sequence input.

[0094] The physical layer identification branch is used to identify the physical layer modulation scheme based on shared feature vectors. This branch preferably employs a multi-layer fully connected network, a convolutional classification network, or an attention-based classification network to output the modulation scheme category label and its probability distribution. Simultaneously, the probability distribution is preferably obtained through a SoftMax activation function, and the probability entropy is calculated to represent the identification uncertainty.

[0095] For example, the physical layer modulation scheme preferably includes common modulation types in the art, such as AM, FM, ASK, FSK, PSK, QAM, OFDM, etc., and other modulation types can also be extended according to the characteristics of the shortwave channel.

[0096] The protocol layer identification branch is used to identify the protocol layer type based on shared feature vectors. This branch preferably employs a classification network structure similar to the physical layer branch to output protocol category labels and their probability distributions. Simultaneously, the probability distribution output by the protocol layer identification branch can also be used to calculate probability entropy for subsequent confidence assessment modules.

[0097] For example, the protocol layer standard type preferably includes standards commonly used in shortwave communication in this field, such as ALE, STANAG, MIL-STD series protocols, FSK link protocol, data link protocol, etc., and it is also preferable to extend other protocol types according to the actual system configuration.

[0098] More specifically, for the recognition module in the preferred embodiment, a multi-task learning framework is further preferably adopted during the training process, with the shared feature extraction unit as a shared part, and task losses are calculated separately for the physical layer recognition branch and the protocol layer recognition branch. The loss function preferably includes at least one of classification cross-entropy loss, weighted classification loss, and consistency constraint loss.

[0099] Furthermore, in the preferred embodiment, the consistency verification module is used to perform a logical consistency check on the physical layer modulation type and protocol layer standard type output by the identification module to determine whether the identification result meets the constraints of the protocol specification library. In this case, the input to the consistency verification module is the dual-layer identification result from the identification module and channel parameter information (i.e., channel quality indicators), and the output is a consistency score and a verification flag.

[0100] In practical design, the consistency verification module preferably includes a rule base management unit, a parameter extraction unit, a model verification unit, and a scoring output unit, wherein: The rule base management unit is used to store the specification constraint information of the shortwave communication protocol, which includes, but is not limited to: (a) The set of allowed modulation schemes corresponding to the protocol; (b) The symbol rate range corresponding to the protocol; (c) The bandwidth range corresponding to the protocol; (d) Frame structure information (including preamble length, synchronization word mode, frame header format, etc.); (e) Spreading parameters (including chip rate, spreading gain, etc.); (f) Frequency hopping parameters (including frequency hopping rate, frequency hopping sequence mode, etc.).

[0101] In the specific construction process, the aforementioned rule base is preferably established through manual input, table import, or program initialization, and can be expanded and updated as needed.

[0102] The parameter extraction unit is used to extract parameter information for consistency verification from the output of the detection module and the recognition module, and then pass the parameter information to the model verification unit after formatting the parameter information.

[0103] More specifically, the aforementioned parameter information includes, but is not limited to: (a) Modulation mode identification results; (b) Protocol type identification results; (c) Center frequency; (d) Bandwidth usage; (e) Symbol rate or chip rate; (f) Frame header structure features.

[0104] The model validation unit is used to perform consistency checks on the recognition results based on the protocol specification library. The specific validation rules must include at least one of the following: (a) Modulation-protocol correspondence verification: Check whether the modulation method belongs to the set of modulation methods allowed by the protocol. If it does not belong, mark it as inconsistent; (b) Spectrum parameter verification: Check whether the occupied bandwidth is within the bandwidth allowed by the protocol. If it exceeds the range, it is marked as inconsistent. (c) Symbol rate verification: Check whether the symbol rate or chip rate meets the protocol requirements; (d) Frame structure verification: If the protocol includes frame structure requirements, check whether the frame structure features such as synchronization word mode and preamble length meet the constraints. (e) Channel scenario verification: In the case of special protocol specifications, scenario consistency checks are performed based on parameters such as channel attenuation type and frequency hopping mode.

[0105] It is understandable that, in actual setup, different protocol types can be configured with different verification items, and the model verification unit preferably selects the rules to participate in the verification dynamically according to the protocol number.

[0106] More specifically, the scoring output unit quantifies the above verification results into a consistency score, and ultimately outputs the consistency score and a binary verification flag (pass / fail). This consistency score is preferably generated using one of the following methods: (a) Score or take a weighted average of all verification items; (b) Deduct points for inconsistencies; (c) Set mandatory constraint flags for key items (i.e., if a key inconsistency occurs, the item will be judged as failing).

[0107] In actual output, the consistency verification results are preferably passed to the confidence assessment module and the sample management module in a structured format for subsequent use. When the consistency verification result is unsuccessful or the consistency score is lower than the preset threshold, the identified sample is regarded as a potentially difficult sample and enters the subsequent sample screening process.

[0108] Furthermore, the confidence assessment module calculates a comprehensive confidence score for the identification results based on the outputs of the detection module, identification module, and consistency verification module. Simultaneously, the output of the confidence assessment module is passed in a structured format to the sample management module, model update module, and deployment module for subsequent decision-making.

[0109] Specifically, the confidence assessment module takes into account the physical layer identification probability distribution, the protocol layer identification probability distribution, the channel quality index, and the consistency score as input signals. The output is a comprehensive confidence score and a confidence level, which are used to guide the screening of difficult samples and the model update process.

[0110] In practical configuration, the confidence assessment module preferably includes a channel quality assessment unit, an identification probability assessment unit, a consistency score assessment unit, and a confidence fusion unit. Among these: The channel quality assessment unit is used to quantify and score channel conditions. This unit preferably constructs a channel quality score based on parameters such as signal-to-noise ratio (SNR), signal-to-interference-to-noise ratio (SINR), multipath delay spread, and frequency offset estimate calculated by the detection module, and uses this score to reflect whether the signal itself is suitable for the identification task.

[0111] More specifically, the channel quality score is preferably calculated using the following method: (a) Convert the signal-to-noise ratio to a weight in the [0, 1] interval using a normalized mapping; (b) Calculate and normalize the stability index based on multipath delay spread or Doppler frequency offset; (c) The channel quality score is obtained by weighted fusion of each channel index.

[0112] The probability assessment unit is used to quantitatively analyze the probability distribution output by the identification module. It can calculate probability scores for both the physical layer and protocol layer identification results, and perform weighted fusion based on configuration. This includes: (a) Maximum probability value: that is, the maximum classification probability output by the SoftMax activation function; (b) Probability entropy: Used to measure the uncertainty of a probability distribution, defined as... ;in, For probability entropy, To indicate the first i The probability of each possible outcome. At the same time, the higher the entropy value, the more uncertain the identification; (c) Distribution difference: the difference between the highest probability and the second highest probability, used to measure the clarity of the classification boundary.

[0113] The consistency score evaluation unit is used to normalize the consistency score output by the consistency verification module. Its output consistency score value is used to enhance the credibility of correct identification and suppress the credibility of incorrect combinations.

[0114] In actual evaluation, if the consistency check result is unsuccessful, the consistency score can be directly set to zero or a low value to reflect the logical conflict in the identification conclusion.

[0115] The credibility fusion unit is used to weight and fuse the channel quality score, the identification probability score, and the logical consistency score to obtain the final comprehensive credibility score.

[0116] In practical design, the credibility fusion unit preferably supports at least one of the following strategies: (a) Linear weighted fusion: characterized as ,in, A comprehensive credibility score is given. For configurable weighting coefficients, it is preferable to set them based on historical data statistics, and ; Score the channel quality; To identify probability scores; Scoring is given based on logical consistency.

[0117] (b) Nonlinear fusion functions: including the Sigmoid function, the SoftMax function, or decision tree-based fusion strategies; (c) Neural network-based fusion model: adaptive fusion of multidimensional scores is achieved by training a small network.

[0118] Meanwhile, the output of the credibility fusion unit includes a comprehensive credibility score and a credibility level. The credibility level can be divided into three levels: high credibility, medium credibility, and low credibility, or into two categories: credibility and untrustworthiness.

[0119] More specifically, the credibility level is preferably divided according to a preset threshold, for example, setting a threshold. and : (a) when At that time, it was judged as highly reliable; (b) When At that time, it was judged as low confidence; (c) When At that time, it was judged to be credible.

[0120] When the overall credibility score is lower than the first threshold, the identified sample can be identified as a low-credibility sample; when the overall credibility score is lower than the second threshold or the consistency check fails, the sample can be identified as a difficult sample and enter the sample management module for further processing.

[0121] Furthermore, the sample management module is used to collect, store, cluster, and semi-supervised pre-label low-confidence and difficult samples to form a training dataset for incremental model updates. Simultaneously, the difficult sample set, cluster information, and pre-labeled sample set output by the sample management module are passed to the model update module for incremental training and model updates.

[0122] Specifically, the inputs of the sample management module include the comprehensive confidence score and confidence level output by the confidence assessment module, the verification results output by the consistency verification module, and the feature representation output by the recognition module. The outputs include the set of difficult samples, cluster information, and the set of pre-labeled samples.

[0123] More specifically, the sample management module preferably includes a sample acquisition unit, a sample storage unit, a cluster analysis unit, and a pre-labeling unit, wherein: The sample acquisition unit is used to receive and filter sample data from the confidence assessment module and the consistency verification module. The filtering strategy preferably includes at least one of the following conditions: (a) The overall credibility score is lower than the preset first threshold; (b) Consistency check failed; (c) The entropy value of the probability distribution exceeds the uncertainty threshold; (d) The physical layer or protocol layer identification result is an unknown category; (e) The system detected abnormal channel conditions, causing the identification to be interrupted.

[0124] Samples that meet any of the above conditions are judged as low-confidence samples or difficult samples and are entered into the sample storage unit.

[0125] Secondly, the sample storage unit is used to manage the storage space of difficult samples. It preferably implements rolling storage, hierarchical storage or partitioned storage according to the set caching strategy to support long-term accumulation and efficient access in online scenarios.

[0126] For the actual stored content, the preferred options include, but are not limited to: (a) Original signal segment or sampled sequence; (b) Multi-domain eigenvectors or dimensionality-reduced eigenvectors; (c) Channel quality metrics; (d) Detection and identification results and probability distribution; (e) Consistency verification flags and scores; (f) Timestamp and signal source information.

[0127] The clustering analysis unit performs similarity clustering on stored hard samples to distinguish between low-confidence samples of known categories and samples of potentially unknown categories. The clustering feature vectors used for clustering analysis can come from the dimensionality-reduced features output by the acquisition and preprocessing module, or from the intermediate layer representation of the shared feature extraction network of the recognition module, to enhance the separability of clusters in the feature space. Simultaneously, the output of the clustering analysis unit is cluster information, including cluster number, cluster center, number of samples within each cluster, and feature density distribution.

[0128] More specifically, the cluster analysis unit preferably employs at least one of the following clustering methods: (a) Density-based clustering methods (DBSCAN, OPTICS, etc.); (b) Hierarchical clustering methods based on distance or similarity; (c) Embedded clustering method based on deep feature space; (d) Spectral clustering methods based on graph structures or spectral methods.

[0129] Furthermore, the pre-labeling unit is used to perform semi-supervised pre-labeling on the clustered hard samples in order to form a dataset for incremental training.

[0130] Specifically, the pre-annotation strategy in the preferred embodiment further preferably includes: (a) Knowledge base-based similarity matching: If the similarity between a feature in a cluster and a sample of a certain category in the historical knowledge base exceeds the similarity threshold, then the sample of that cluster is assigned a pre-label of the corresponding category; (b) Based on majority voting strategy: For samples that already have some credible labels within a cluster, cluster-level pseudo-labels can be generated through majority voting or weighted voting; (c) Uncertainty-based filtering strategy: Samples with excessively high probability entropy are not labeled, but their features are retained for subsequent model training or manual labeling; (d) New category detection strategy: For clusters whose features have low similarity to historical categories, they can be marked as potential new categories or temporary categories, and then upgraded to formal categories after confirmation by manual auxiliary annotation units or accumulation of more samples.

[0131] Meanwhile, the final output of the pre-labeling unit is a set of pre-labeled samples, which includes sample features, pre-labels, and confidence values.

[0132] In this embodiment, the model update module is used to combine the set of difficult samples and the set of pre-labeled samples output by the sample management module to perform incremental training and anti-forgetting optimization on the recognition model, so as to achieve adaptation to new protocols, new modulation methods or special communication scenarios.

[0133] Specifically, the inputs to the model update module are the pre-labeled sample set, historical training samples or sample summaries, old model weights and teacher model outputs, and the outputs are the updated recognition model parameters and version information.

[0134] In practical configuration, the model update module preferably includes a trigger management unit, a data construction unit, an incremental training unit, and a knowledge distillation unit, wherein: The trigger management unit is used to determine whether to initiate the model evolution process. When the trigger management unit detects that the evolution conditions are met, the incremental training process is initiated; otherwise, the model state remains unchanged.

[0135] More specifically, the triggering conditions for the model evolution process preferably include at least one of the following conditions: (a) The number of samples in the hard sample set exceeds a set threshold; (b) New clusters emerge and accumulate to a certain sample size; (c) Recent recognition accuracy has decreased by more than a preset percentage compared to the baseline; (d) An unknown protocol category or variant appears; (e) Triggered on a timed or periodic basis (e.g., daily, weekly, or monthly); (f) Triggered manually or by maintenance instructions.

[0136] Meanwhile, the data construction unit is used to prepare the incremental training dataset and output the training set, validation set and test set, providing single-class sample count information for subsequent distillation strategy allocation.

[0137] More specifically, the preferred data construction process includes: (a) Select the pre-labeled samples output by the sample management module as new category or new scene samples; (b) Select historical training samples or their summary samples (e.g., feature vectors and label pairs) to maintain the old class decision boundary; (c) Balance the sampling according to the category sample ratio to avoid excessive proportion of new or old categories; (d) Employ a weight amplification strategy for pre-labeled samples to give them a larger share in the training loss; (e) Perform uncertainty filtering or data augmentation on low-confidence samples to improve training quality.

[0138] Furthermore, the incremental training unit is used to incrementally train the recognition model based on the data output by the data construction unit, and the training method adopted by the incremental training unit preferably includes at least one of the following methods: (a) Incremental training based on parameter fine-tuning: Using the current model parameters as the initial weights, only the back-end classification layer is fine-tuned, which is suitable for cases with few category expansions; (b) Incremental training based on full network fine-tuning: Fine-tuning the entire network structure, suitable for large-scale scene adaptation; (c) Incremental training based on structural expansion: When there are many new categories or the model structure needs to be adapted to new tasks, expand the classification head or add new modules for training.

[0139] The preferred incremental training loss includes classification cross-entropy loss, sample weighting loss, center loss, or other losses that preserve feature separability.

[0140] Furthermore, the knowledge distillation unit is used to avoid the decline in the ability to identify old categories due to incremental training (catastrophic forgetting), and its configuration mechanism preferably includes: (a) Using the old model as the teacher model, generate soft label probability distributions for the training set and historical samples; (b) Calculate the distillation loss for the predicted output of the new model and the soft label of the teacher model. The distillation loss is preferably KL divergence, mean square error or cross-entropy loss after temperature scaling. (c) For new category samples, only calculate the hard label loss; (d) For old category samples, calculate both hard label loss and distillation loss; (e) Assign higher distillation weights to key or low-frequency categories to maintain their identification boundaries.

[0141] Simultaneously, the knowledge distillation process is optimized and incremental training is performed concurrently, ultimately forming a fusion loss function, which can be represented as follows:

[0142] In the formula, This represents the total amount of loss. This is the hard label loss; For distillation loss, For distillation weight.

[0143] In addition, the incremental training termination condition preferably includes at least one of the following: (a) The accuracy on the validation set converges or shows no significant improvement; (b) The training loss converges or reaches the upper limit of the number of iterations; (c) Distillation loss decreases below a set threshold; (d) The training time has reached the set limit.

[0144] After incremental training is completed, the model update module outputs the updated model weights, class label mapping table and version number information, and transmits them to the shadow validation and deployment module.

[0145] In this embodiment, the shadow verification and deployment module is used to perform shadow mode verification and online hot-swap deployment on the updated model output by the model update module, to ensure that the identification model completes the update process without interrupting business operations. Simultaneously, the shadow verification and deployment module ultimately outputs the current online master model and version information, and stores the deployment logs, performance metrics, and model status in the system log management module for auditing and traceability.

[0146] In actual operation, the inputs of the shadow validation and deployment module include updated model weights, version number information, old model weights, and evaluation metrics, and the outputs are the currently running online model and the switching log.

[0147] More specifically, the shadow verification and deployment module preferably includes a shadow verification unit, a performance evaluation unit, a hot-switching unit, and a rollback management unit, wherein: The shadow verification unit is used to run the updated model in shadow mode after the model is updated. The preferred mode of operation is: the updated model and the current main model receive real-time input data streams in parallel and independently complete the detection, recognition and confidence evaluation process. However, the recognition results output by the shadow model are not released to the public and are only used for internal testing and comparative analysis.

[0148] As an example, the shadow verification unit is preferably configured to perform the following tasks: (a) Synchronous data input: Using a data distribution mechanism, the same signal feature stream is synchronously input to the old and new models; (b) Independent inference execution: The updated model independently performs inference, generating recognition results and confidence scores; (c) Result Cache Record: Save the shadow model output to the cache area for subsequent performance analysis.

[0149] The shadow verification cycle is preferably set to the hour, day, or week level according to the system configuration.

[0150] Secondly, the performance evaluation unit is used to statistically analyze the performance of the updated model in shadow mode and compare it with the current main model. Preferred evaluation metrics include, but are not limited to: (a) Full sample recognition accuracy; (b) Recognition accuracy or F1 score on difficult sample sets; (c) The degree of model confidence calibration; (d) Single-sample inference delay or average processing delay; (e) Performance fluctuation range or stability indicators; (f) Resource consumption metrics (GPU utilization, memory usage, etc.).

[0151] The performance evaluation unit preferably scores the indicators comprehensively according to preset weights and generates a model evaluation report. If the updated model's performance is stable and meets or exceeds the preset performance conditions of the current main model during the evaluation period, a verification pass signal will be sent to the hot-switching unit.

[0152] More specifically, in the preferred embodiment, the hot-switching unit is used to perform an online model replacement operation in order to upgrade the shadow model to the main model.

[0153] In actual configuration, the hot-switching unit preferably supports at least one of the following switching modes: (a) Instant switching mode: The updated model is immediately switched to the main model when the verification conditions are met; (b) Delayed handover mode: Handover is performed during periods of low service load; (c) Partition switching mode: First, test the update model on some channels or data sources, and then gradually expand it to the full business.

[0154] The preferred execution steps of the hot-switching unit are as follows: (a) Freeze the current main model state and record the version number; (b) Switch the shadow model to the new main model; (c) Update the model mapping table and the category label table; (d) Mark the available time and deployment time of the model version.

[0155] In addition, the hot-swap process is transparent to external business operations and does not interrupt the system's data processing flow.

[0156] Furthermore, the rollback management unit is used to perform rollback operations when an abnormal situation occurs after the updated model is deployed.

[0157] More specifically, the preferred exception detection conditions during the operation of the rollback management unit include: (a) Recognition performance degrades beyond a set threshold; (b) Inference latency exceeds real-time requirements; (c) Stability indicators are abnormal or fluctuate excessively; (d) The system malfunctions and issues an error alarm or resource overload.

[0158] Furthermore, the preferred rollback operations for the model include: (a) Switch the old model to the current main model; (b) Record the reason for the update failure and the context information; (c) Mark the updated model status as "to be improved" or "to be taken offline"; (d) Trigger the model update module to adjust the training strategy or data construction strategy.

[0159] After the rollback is completed, the shadow verification and deployment module will prioritize the notification sample management module to continue accumulating hard samples to support the next round of model evolution.

[0160] By combining the aforementioned modules, a shortwave signal acquisition, detection, identification, and model update system can be accurately obtained. This system accurately realizes a dual closed loop of data and strategies for shortwave signal acquisition, detection, identification, evaluation, screening, updating, and deployment, effectively improving the link stability and adaptive processing capabilities of the shortwave signal identification system, thereby meeting the actual needs of engineering deployment.

[0161] Example 3: It is understood that the method steps in Embodiment 1 can be implemented by an electronic device executing program instructions, or by software, hardware, or a combination of both. The program instructions can be stored in a computer-readable storage medium, and the electronic device involved includes a processor and a memory. The processor executes the program instructions in the memory to implement the method steps in Embodiment 1. Conversely, the software is preferably deployed on a server, embedded device, or other processing device with computing capabilities.

[0162] Therefore, as another aspect of the present invention, a storage medium is also provided, wherein a processor-executable program is stored, which, when executed by a processor, is used to perform the shortwave signal acquisition, detection and identification and model update method described in Embodiment 1.

[0163] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A short wave signal acquisition, detection and identification and model updating method, characterized in that, Includes the following steps: (1) Acquire shortwave signals and perform preprocessing to obtain feature data for identification and channel quality indicators for subsequent confidence calculation; (2) Perform signal presence detection and dual-layer parallel identification on the feature data, and output the identification results of physical layer modulation mode and protocol layer standard type respectively; (3) Based on the preset protocol specification library, the logical consistency of the two identification results is checked, and the comprehensive credibility is calculated by combining the channel quality index and the identification probability distribution; (4) Select samples that fail the logical consistency check or whose overall credibility is lower than the preset threshold as difficult samples, and perform clustering and semi-supervised pre-labeling on the difficult samples to construct an incremental training dataset; (5) Use the incremental training dataset to perform incremental updates on the recognition model, and then deploy the updated model online using the shadow verification mode; In the incremental update process, knowledge distillation constraints and consistency constraint loss functions based on protocol constraint relationships are introduced. The constraint relationship between modulation method and protocol type in the protocol specification library is encoded as training constraints to constrain the matching relationship between physical layer identification results and protocol layer identification results.

2. The short wave signal acquisition, detection and identification and model updating method according to claim 1, characterized in that, In step (1), the preprocessing includes at least one of radio frequency downconversion, analog-to-digital conversion, channel compensation, multi-domain feature extraction, feature fusion and dimensionality reduction.

3. The short wave signal acquisition, detection and identification and model updating method according to claim 2, characterized in that, The multi-domain feature extraction includes time-domain features, frequency-domain features, spatial-domain features, and code-domain features; and / or The feature fusion and dimensionality reduction process includes: normalizing the features of each domain, performing feature dimensionality reduction using principal component analysis or a deep learning encoder, and retaining the main feature components.

4. The short wave signal acquisition, detection and identification and model updating method according to any one of claims 1 to 3, characterized in that, In step (2): The signal presence detection employs at least one of energy detection, cyclostationary feature detection, or deep learning detection methods, and outputs a signal presence flag and detection confidence level. and / or The dual-layer parallel recognition is based on two parallel branches: physical layer modulation recognition and protocol layer standard recognition. The two branches share the underlying feature extraction network and output the physical layer modulation method recognition result and the protocol layer standard type recognition result respectively through their respective classification heads.

5. The short wave signal acquisition, detection and identification and model updating method according to any one of claims 1 to 3, characterized in that, In step (3), the preset protocol specification library includes: The correspondence between modulation and protocol is constrained, defining the set of allowed modulation methods for each protocol type; Parameter range constraints define the allowed range of parameters for each protocol type. Frame structure constraints define the frame structure characteristics of each protocol type; and The logical consistency verification process includes: querying the protocol specification library and obtaining constraints; checking whether the physical layer modulation type conforms to the set of modulation methods allowed by the protocol layer; checking whether the parameters are within the range defined in the protocol specification library; and calculating the overall consistency score.

6. The method for shortwave signal acquisition, detection, identification, and model updating according to claim 5, characterized in that, The overall confidence level is calculated and output through a constructed multidimensional confidence assessment matrix, which includes the following dimensions: The channel quality dimension calculates a channel quality score based on channel metrics, which include at least signal-to-noise ratio, multipath parameters, and bandwidth consistency. The identification probability dimension is determined by calculating the identification deterministic score based on the maximum probability distribution and probability entropy of the physical layer and protocol layer identification. In terms of consistency, a consistency score is calculated based on a comprehensive consistency score derived from logical consistency verification. The overall credibility score is obtained by weighted fusion of scores from each dimension, with the weights either set to a fixed preset weight or dynamically adjusted based on the reliability of each dimension.

7. The method for shortwave signal acquisition, detection, identification, and model updating according to any one of claims 1 to 3 or claim 6, characterized in that, In step (4): The clustering process employs density-based clustering, hierarchical clustering, or deep feature-based clustering methods; difficult samples are divided into several similarity clusters based on signal characteristic similarity, which facilitates batch pre-labeling and discovery of new protocol categories; and the clustering feature space uses dimensionality-reduced features or intermediate layer features of the recognition network. and / or The semi-supervised pre-labeling process includes: for samples in a cluster whose similarity to historical knowledge base samples is higher than a threshold, inheriting the labels of historical samples; for newly emerging clusters, marking them as potential new categories, triggering manual assisted labeling or automatically assigning temporary labels; for sample groups with high consistency within a cluster, generating pseudo-labels using majority voting or confidence weighting.

8. The method for shortwave signal acquisition, detection, identification, and model updating according to any one of claims 1 to 3 or claim 6, characterized in that, In step (5): The knowledge distillation mechanism includes: using the old model as the teacher model to generate soft labels for training samples, where the soft labels reflect the probability distribution of the old model for each category; the training loss function of the new model includes a weighted combination of hard label loss and soft label distillation loss, where the hard label loss uses cross-entropy loss and the soft label distillation loss uses KL divergence or mean squared error; for samples of new categories, only hard label loss is used; for samples of known categories, both hard label loss and soft label distillation loss are used. The consistency constraint loss function, together with the classification loss and distillation loss, constitutes a joint optimization objective function, which is used to achieve synergistic optimization of recognition accuracy and cross-layer consistency. The shadow verification mode includes: performing parallel inference between the updated model and the current online model for the same real-time signal, and comparing their performance metrics; The online hot deployment process of the updated model includes: deploying the updated model in shadow mode and performing parallel inference on real-time signals with the current online model; collecting and comparing the key performance indicators of the two models within a preset verification period; and performing an online hot switch when the updated model meets the preset conditions in terms of hard sample recognition performance and overall stability, making the updated model the master model and the original master model downgraded to a backup model or retired.

9. A shortwave signal acquisition, detection, identification, and model updating system, characterized in that, include: The acquisition and preprocessing module is used to acquire shortwave signals and perform preprocessing to obtain feature data for identification and channel quality indicators for subsequent confidence calculation. The detection module is used to perform signal presence detection and time-frequency localization on the feature data, and output the detection results and detection confidence level; The dual-layer recognition module is used to perform dual-layer parallel recognition on the feature data to output the physical layer modulation mode recognition result and the protocol layer standard type recognition result. The consistency verification module stores a predefined protocol specification library, which is used to perform logical consistency verification on the dual-layer identification results and output a consistency score. The confidence assessment module is used to construct a multi-dimensional confidence assessment matrix, integrate channel quality indicators, identification probability distribution and consistency verification results, and output a comprehensive confidence score. The sample management module is used to manage difficult samples, perform clustering and semi-supervised pre-labeling, and generate training datasets. The model update module is used to perform incremental learning and knowledge distillation training, initiate the model update process according to the trigger conditions, and generate the updated recognition model. The shadow verification and deployment module is used to manage parallel inference of shadow models, compare performance metrics, and perform online hot switching or rollback operations.

10. A storage medium storing a processor-executable program, characterized in that, When executed by a processor, the program is used to perform the shortwave signal acquisition, detection and identification, and model update method as described in any one of claims 1 to 8.

Images