Fast adaptive radiation source individual open set identification method and system based on edge cloud collaborative reasoning

Abstract

The invention discloses a rapid adaptive radiation source individual open set identification method and system based on edge cloud collaborative reasoning, and belongs to the field of radiation source individual identification. According to the rapid self-adaptive radiation source individual open set identification method based on edge cloud cooperative reasoning, through an edge end rapid reasoning mode, newly appearing unknown class signals can be rapidly found, and channel environment changes can be perceived. Meanwhile, the reliability of side end reasoning can be ensured in a cloud asynchronous verification mode. The recognition system comprises an edge end node and a cloud end node, the edge end node is used for processing a lightweight real-time radiation source individual recognition task, and when edge end resources are insufficient, the cloud end node can be called through an edge cloud cooperation mechanism to assist reasoning; and the cloud node is used for providing an asynchronous verification function for the side end node to ensure the reliability of a reasoning result.

Description

Technical Field

[0001] This invention relates to the field of radiation source individual identification, and in particular to a fast adaptive open set identification method and system for radiation source individuals based on edge-cloud collaborative reasoning. Background Technology

[0002] Individual radiation source identification (IRI) is a technology that uses the differences in radio frequency fingerprints (RF fingerprints) of wireless signals (such as transient response, phase noise, and harmonic distortion) to distinguish different individual radiation sources. It has significant applications in both military and civilian fields. In the military, since friendly and enemy forces may use the same type of electronic equipment or communication signals, traditional signal analysis methods struggle to differentiate them, easily leading to misjudgments. The uniqueness of RF fingerprints allows this technology to effectively distinguish between friendly and enemy equipment, reducing the risk of accidental damage. In the civilian sector, the surge in the number of IoT devices poses security risks to traditional cryptographic authentication. IRI technology, based on physical layer signal recognition, provides secure physical layer identity authentication and has been widely applied in scenarios such as fake base station detection, smart home device authentication, and industrial IoT security, effectively solving the identity recognition problem and ensuring system security.

[0003] However, most existing research on individual radiation source identification focuses on designing better radio frequency fingerprint feature extraction and classification methods to achieve accurate identification of individual radiation sources. These studies assume that the individual radiation source identification model is always deployed in an ideal closed environment, that is, the individual radiation sources that the model needs to identify in the training and testing phases come from the same category set, and the data used by the model in the training and testing phases come from the same channel environment. However, the above assumptions do not hold true in an open environment. Specifically, individual radiation source identification faces two challenges in an open environment: (1) the existence of unknown categories in an open environment reduces the credibility of the identification results of the individual radiation source identification model; (2) changes in the channel environment in an open environment cause a distribution shift in the radio frequency fingerprint features of the radiation source, resulting in a decrease in the identification accuracy of the individual radiation source identification model.

[0004] To address these issues, existing research primarily relies on open set identification techniques, which distinguish between known and unknown radiation sources by constructing classification boundaries for known and unknown categories of data. While these methods possess the ability to identify unknown categories of radiation sources and overcome the impact of channel environment changes, the number of unknown categories that the algorithm can identify and its adaptability to the channel environment depend on the unlabeled data of the target domain collected during the training phase. This dependency is determined during model training and remains unchanged after training. However, after the radiation source identification model is deployed, new unknown categories of radiation sources will still emerge in the open environment, and the channel environment will also change. At this point, it is necessary to collect unlabeled data from the current environment and use this data to adjust the radiation source identification model in order to identify new unknown categories of radiation sources and adapt to changes in the current channel environment. Since data collection and model training are time-consuming processes, radiation source identification models deployed in open environments cannot quickly identify newly emerging unknown radiation sources or adapt to changes in the channel environment, making it difficult to meet the application requirements of real-world open environments. Therefore, it is of great significance to design a fast adaptive radiation source identification system for open environments, which can quickly adapt to newly emerging radiation source individuals and channel changes in open environments. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a fast adaptive open set identification method and system for individual radiation sources based on edge-cloud collaborative reasoning.

[0006] The first aspect of this invention provides a fast adaptive open set identification method for individual radiation sources based on edge-cloud collaborative reasoning, which performs the following steps in a radiation source identification system including edge devices and cloud devices:

[0007] Step 1: The edge device acquires the original radiation source signal input and extracts depth and manifold features based on the local feature extractor. The manifold features are input into the manifold nearest neighbor classifier, and the depth features are input into the Euclidean nearest neighbor classifier to obtain the original manifold nearest neighbor classification output and the original Euclidean nearest neighbor classification output.

[0008] Step 2: The edge device performs conflict resolution on the initial manifold nearest neighbor classification output and the original Euclidean nearest neighbor classification output, that is, conflict resolution on the classification results output by the dual classifier; and adjusts the weights of the manifold nearest neighbor classifier and the Euclidean nearest neighbor classifier respectively based on the resolution results.

[0009] Step 3: The edge device makes a classification decision based on the output of the dual classifier after conflict resolution to obtain confused samples; and submits the current confused samples and their dual classifier classification results to the cloud device for asynchronous verification in real time.

[0010] Step 4: After receiving the asynchronous verification request and the obfuscated sample, the cloud device performs the same feature extraction, classification, and conflict handling as the edge device based on its local feature extractor, manifold nearest neighbor classifier, and Euclidean nearest neighbor classifier to obtain the cloud's dual classifier classification result. Based on the cloud's dual classifier classification result and the classification result of the received obfuscated sample, it performs edge-cloud collaborative inference processing (based on the configured edge-cloud collaborative inference algorithm), outputs the edge-cloud collaborative inference result, and feeds it back to the edge device. The edge device corrects the classification result of the corresponding asynchronous verification obfuscated sample based on the real-time feedback result.

[0011] Furthermore, the edge devices and cloud devices are configured with the same feature extractor, manifold nearest neighbor classifier, and Euclidean nearest neighbor classifier. Additionally, the cloud devices can also distribute the feature extractor, manifold nearest neighbor classifier, and Euclidean nearest neighbor classifier to the edge devices.

[0012] Furthermore, deep feature and manifold feature extraction based on feature extractors specifically includes:

[0013] The temporal features of each time-domain signal sample in each batch are extracted based on the feature extractor.

[0014] The time-domain signal is converted into the frequency-domain signal using the Discrete Fourier Transform, and then converted into polar coordinates. The resulting amplitude vector and phase vector are used as the frequency-domain features of each signal sample.

[0015] The time-domain features and frequency-domain features are fused by splicing to obtain the time-frequency fusion features of each signal sample, i.e., the depth features;

[0016] Calculate the autocorrelation matrix of the depth features of the signal sample, and then perform singular value decomposition on it;

[0017] Take the first left singular matrix obtained from singular value decomposition The column yields the shape features, among which, This is a preset value. Preferably, The rule for selecting the value is: before selection A singular value whose energy percentage exceeds a specified proportion.

[0018] Furthermore, the feature extractor can employ a backbone network based on a convolutional neural network recognition network.

[0019] Furthermore, the classification results of the Euclidean nearest neighbor classifier and the manifold nearest neighbor classifier include: classification category and classification confidence; wherein, the classification confidence is set as follows: , The nearest neighbor distance output by the classifier. This is the distance threshold for unknown categories in the classifier.

[0020] Furthermore, conflict handling of classification results output by the dual classifier includes:

[0021] definition , These represent the classification categories output by the Euclidean nearest neighbor classifier and the manifold nearest neighbor classifier, respectively.

[0022] definition , These represent the classification confidence scores output by the Euclidean nearest neighbor classifier and the manifold nearest neighbor classifier, respectively.

[0023] definition , Let represent the edge Euclidean classification confidence weight and the edge popular classification confidence weight, respectively, and initialize them. , ;

[0024] Conflict handling determination: If the two classifiers output the same classification category, it means there is no conflict, and the conflict is resolved. or As the current edge category ; and will As edge popularity confidence , As a marginal European confidence level ;

[0025] If the dual classifier outputs a classification category, conflict handling is performed, specifically including:

[0026] judge Is it greater than or equal to? If so, then As the current edge category And based on the current , Perform the assignment operation:

[0027] ;

[0028] Then reduce the downward adjustment based on the settings. Increase based on the setting's upward adjustment range ;

[0029] Otherwise, As the current edge category And based on the current , Perform the assignment operation:

[0030] , ;

[0031] Then reduce the downward adjustment based on the settings. Increase based on the setting's upward adjustment range .

[0032] Furthermore, step 3 includes:

[0033] Edge classification decision:

[0034] If the current edge popular confidence level and edge European confidence All are greater than the preset high-confidence threshold for known categories. If the current signal sample is of a known category, then it is based on the current edge category. Obtain the final category label;

[0035] like and All are less than the preset high-confidence threshold for unknown categories. If the current signal sample is of an unknown category, it is assigned a label in ascending order starting from the maximum value of the known category label. Based on the assigned label, the manifold feature and depth feature of the signal sample are marked. The labeled manifold feature and depth feature are then stored as new known class sample features in the edge sample feature database. At the same time, the signal sample and its labeled manifold feature and depth feature are uploaded to the cloud device so that it can be updated with new unknown class sample data in the cloud device in a timely manner.

[0036] And will not be satisfied The signal sample, or does not meet the requirements The signal samples are treated as scrambled samples; and the scrambled samples and their current signals are treated as scrambled samples. , , and Submit to the cloud device for asynchronous verification; where the symbol " " indicates logical AND;

[0037] Edge decision processing: If the classification confidence scores of the two classifiers output by the signal samples satisfy... and or satisfy and If so, the current signal sample is considered as a domain offset sample;

[0038] The momentum update method is used to update the manifold and depth features of signal samples with the same category label as the current domain offset sample.

[0039] Furthermore, the manifold and depth features of signal samples with the same class label as the domain offset samples are updated using the momentum update method, specifically as follows:

[0040]

[0041]

[0042] in, , These represent the updated popular features and deep features, respectively, with weight coefficients. , The sum of the two is 1, and Greater than , , These represent the popularity features and depth features before the update, respectively. E represents the depth features of the domain offset sample. and Log These are the exponential and logarithmic mapping functions of the manifold space, respectively.

[0043] Furthermore, step 4 specifically includes:

[0044] Step 4-1: For the currently received confused samples, the cloud device performs the same feature extraction, classification, and conflict handling as the edge device based on its local feature extractor, manifold nearest neighbor classifier, and Euclidean nearest neighbor classifier to obtain the cloud category. And the corresponding cloud-based European confidence level after conflict resolution. Cloud popularity confidence Cloud-based European classification confidence weights Cloud-based popular classification confidence weights ;

[0045] Step 4-2: Perform cloud collaborative reasoning processing, output the edge-cloud collaborative reasoning results, and feed them back to the edge device:

[0046] Calculate confidence in the cloud Cloud-based trust weight ;

[0047] Calculate edge confidence Edge confidence weight ;

[0048] like Then the following assignment operation is performed:

[0049]

[0050]

[0051] Otherwise, perform the following assignment operation:

[0052]

[0053]

[0054] in, This indicates the category determined by cloud collaborative reasoning. This indicates the downward adjustment range set. Indicates the upward adjustment range set;

[0055] Calculate the confidence score for joint classification of edge clouds:

[0056]

[0057] in, , These are the confidence scores for the combined popular classification of edge-cloud and the combined Euclidean classification of edge-cloud, respectively.

[0058] based on , and The classification process is performed to obtain the edge-cloud collaborative inference results; the classification rule is as follows: if and ,but For a known category; if and ,but Unknown category; otherwise This refers to the domain offset category, i.e., the known category affected by the domain offset.

[0059] Step 4-3: The edge device corrects the classification results of the corresponding asynchronously verified confused samples based on the feedback results received in real time.

[0060] A second aspect of the present invention provides a fast adaptive open-set identification system for individual radiation sources based on edge-cloud collaborative reasoning, comprising:

[0061] Edge nodes are used to handle lightweight real-time individual radiation source identification tasks. When the computing resources of edge nodes are insufficient, cloud nodes are called through the edge-cloud collaboration mechanism to assist in reasoning. Cloud nodes are used to provide asynchronous verification and assisted reasoning functions for edge nodes.

[0062] The functional layers of the edge node include an edge signal processing layer, an edge data management layer, an edge model algorithm layer, and an edge human-computer interaction layer.

[0063] in,

[0064] The edge signal processing layer is used to continuously provide signal sample data in a specified format to the edge model algorithm layer. It includes a signal acquisition module and a signal preprocessing module. The signal acquisition module is used to continuously acquire signal file data as raw data. The signal preprocessing module is used to process the raw data acquired by the signal acquisition module into the data format required by the edge model algorithm layer.

[0065] The edge data management layer stores the prototype features of known categories of radiation sources required for inference calculations at the edge model algorithm layer, as well as log data generated at the edge, and temporarily stores signal samples for a period of time; the prototype features include depth features and manifold features.

[0066] The edge model algorithm layer includes a classification model module, a conflict resolution algorithm module, an edge inference algorithm module, and a data update module. The classification model module extracts the manifold and depth features of the signal samples and uses a dual classifier to classify the two types of features, obtaining a preliminary classification result for the signal samples. The dual classifiers are a manifold nearest neighbor classifier and an Euclidean nearest neighbor classifier. The conflict resolution algorithm module maintains consistency between the two classification results output by the classification model module and dynamically adjusts the classification confidence weights of the dual classifiers. The edge inference algorithm module generates the final classification result for the signal samples and uploads confused samples to the cloud node for asynchronous verification. The data update module updates the prototype features of known-category radiation source individuals based on the final classification result of the edge inference algorithm module, or adds the features of identified unknown-category radiation source individuals to the prototype feature library. It also receives new model parameters from the cloud node to update the feature extraction model of the edge node. This feature extraction model extracts the temporal features of the signal samples and fuses the frequency domain features to obtain the depth features of the signal samples.

[0067] The edge-side human-computer interaction layer provides users with a visual interface, including a signal display module, a recognition result display module, and a recognition log display module. The signal display module provides time-domain waveforms and spectrum diagrams of signal samples. The recognition result display shows the final classification results. The recognition log display model provides recognition statistics during the operation of the edge node, including the total number of recognized signal samples, the total number of signal samples of known categories, and the total number of signal samples of unknown categories.

[0068] The functional layers of the cloud node include a cloud data management layer, a cloud model algorithm layer, a cloud system management layer, and a cloud human-computer interaction layer;

[0069] in,

[0070] The cloud data management layer is used to store sample features (manifold features and depth features) of known categories of radiation sources required for inference calculations at the cloud model algorithm layer, log data generated in the cloud, all collected signal sample data uploaded from the edge, and all historical model parameters of the edge-cloud hybrid system.

[0071] The cloud-based model algorithm layer includes a classification model module, a conflict resolution algorithm module, an edge-cloud collaborative inference algorithm module, and a data update module. The classification model module and conflict resolution algorithm module have the same functions as those on the edge. The edge-cloud collaborative inference algorithm module generates the final classification result for the signal samples and sends this result to the edge nodes. The data update module updates the local sample feature library and, based on received model automatic training notifications, updates the cloud-based feature extraction model with the newly trained model parameters and sends the new model parameters to the edge nodes.

[0072] The cloud system management layer includes an edge-cloud conflict monitoring module and a model automatic training module. The edge-cloud conflict monitoring module records the number of edge-cloud conflicts that occur during the execution of the edge-cloud collaborative inference algorithm. When the number of conflicts exceeds the system setting, it sends a model training command to the model automatic training module. The model automatic training module automatically starts the model training process for the feature extraction model when it receives the model training command from the edge-cloud conflict monitoring module or when the system running time exceeds a specified threshold. After the model training is completed, it notifies the data update module to update.

[0073] The cloud-based human-computer interaction layer includes a signal display module, a model file list display module, and a model training display module. The signal display module displays a list of all signal samples currently stored in the system; the model file list display module displays a list of all model files currently stored in the system; and the model training display module displays information about each automatic training iteration of the model.

[0074] The technical solution provided by this invention brings at least the following beneficial effects:

[0075] This invention utilizes edge-based fast inference technology to quickly discover newly emerging unknown signal categories and perceive changes in the channel environment. Simultaneously, combined with an asynchronous verification mechanism in the cloud, it can effectively ensure the accuracy and reliability of edge-based inference results, forming a closed-loop optimization system for edge-cloud collaboration. Attached Figure Description

[0076] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0077] Figure 1 This is a flowchart illustrating the fast adaptive open set identification method for individual radiation sources based on edge-cloud collaborative reasoning, according to an embodiment of the present invention.

[0078] Figure 2This is a schematic diagram of the structure of the fast adaptive open set identification system for individual radiation sources based on edge-cloud collaborative reasoning, according to an embodiment of the present invention. Detailed Implementation

[0079] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be described in detail and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Generally, the components of the embodiments of the present invention described and shown in the accompanying drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the present invention.

[0080] In one embodiment, this invention provides a fast adaptive open-set identification method for individual radiation sources based on edge-cloud collaborative inference. This method, through fast edge-side inference, can quickly discover newly emerging unknown signal categories and perceive changes in the channel environment. Simultaneously, it ensures the reliability of edge-side inference through asynchronous verification in the cloud. See also... Figure 1 The fast adaptive open set identification method for individual radiation sources based on edge-cloud collaborative reasoning provided in this embodiment of the invention includes the following steps:

[0081] Step 1: The edge device acquires the original radiation source signal (test signal sample) input. After extracting manifold features and depth features from the test signal sample, the two types of features are input into the manifold nearest neighbor classifier (corresponding to manifold features) and the Euclidean nearest neighbor classifier (corresponding to depth features) respectively to obtain the original manifold nearest neighbor classification output and the original Euclidean nearest neighbor classification output.

[0082] Step 2: Input the original classification output of the dual classifier into the edge conflict handling module, and adjust the weights of the dual classifiers based on the processing results;

[0083] Step 3: The edge performs a fast classification decision based on the output of the dual classifier after conflict resolution and displays the classification results in real time. For signal samples marked as easily confused, the edge submits the sample data and the dual classifier classification results to the cloud for asynchronous verification.

[0084] Step 4: After receiving the verification request information and obfuscated samples sent by the edge, the cloud uses a similar feature extraction, dual-classifier classification, and classifier conflict handling process as the edge to process the obfuscated samples. After that, the dual-classifier classification results of the edge and the cloud are used as input, and the edge-cloud collaborative inference algorithm is used to output the edge-cloud collaborative inference results and feed them back to the edge. The edge continuously receives the feedback results sent by the cloud and uses the feedback results to correct the previous incorrect inference.

[0085] Step 1 includes the following steps:

[0086] Step 1-1: Deep feature extraction, for each batch of sample data A one-dimensional residual network, ResNet18, was used as the temporal feature extraction network to extract temporal features. The discrete Fourier transform is used to convert the time-domain signal into frequency-domain information, and then converts it into polar coordinates to obtain the amplitude vector. and phase vector As a frequency domain feature, the specific process is as follows: First, the sample data... Performing a discrete Fourier transform yields the frequency domain representation. :

[0087]

[0088] Frequency domain representation Low-pass filtering is performed to retain only the low-frequency information represented in the frequency domain.

[0089] Define the filter response H :

[0090]

[0091] The frequency domain representation after low-pass filtering is as follows:

[0092]

[0093] Frequency domain after low-pass filtering Perform amplitude-phase decomposition to obtain the amplitude vector. and phase vector :

[0094]

[0095]

[0096] The amplitude vector and phase vector are concatenated to obtain the frequency domain features. Finally, the time-domain and frequency-domain features are fused by concatenation to obtain the time-frequency fused features:

[0097]

[0098] Among the formulas involved in the above process, It refers to the number of sample data in each batch. It is the length of the sample data. It is the size of the dimension of the time-domain feature. The dimension of the frequency domain feature. The cutoff frequency of a low-pass filter is represented in the frequency domain. Index in This represents a discrete time index, i.e., a sampling point. express The real part, express The imaginary part, Return to origin to point azimuth angle, This represents the feature splicing function.

[0099] Step 1-2: Manifold feature extraction, the features extracted by the time-frequency feature extractor are: , That is, the time-frequency fusion feature of the j-th sampling point, for each feature in each batch of sample data. ,in To index the samples, first calculate the autocorrelation matrix of the sample data. :

[0100]

[0101] Next, the autocorrelation matrix Perform singular value decomposition:

[0102]

[0103] Take the left singular matrix The former The column yields the subspace representation of the data on the Grassmann manifold. ,in The choice of value makes the previous The principle is used to determine the energy percentage of singular values, which must exceed 90%. In the formula, It is a left singular matrix, and its column vectors are an orthonormal basis ( ), It is a singular value matrix.

[0104] Steps 1-3: Collaborative decision-making by the edge node dual classifiers. The deep features obtained in Step 1-1 are input into the Euclidean nearest neighbor classifier in the edge node dual classifier structure to obtain the classification category. and classification confidence The manifold features obtained in steps 1-2 are input into the manifold nearest neighbor classifier in the dual classifier structure to obtain the classification category. and classification confidence ,in Defined as:

[0105]

[0106] In the formula, The nearest neighbor distance output by the classifier. The threshold value for unknown class distances in the classifier is initialized as follows: For a manifold nearest neighbor classifier, calculate the distances from all samples in the popular feature database to their nearest neighbors, and then take the 99th percentile of these distances as the threshold value. This refers to the unknown class distance threshold for the manifold nearest neighbor classifier. For the Euclidean nearest neighbor classifier, the distances from all samples in the deep feature database to their nearest neighbors are statistically analyzed, and then the 99th percentile of these distances is taken as the threshold. , which is the unknown class distance threshold of the Euclidean nearest neighbor classifier.

[0107] Step 2 includes the following steps:

[0108] Step 2-1: Conflict handling and determination, obtaining the output results of the dual classifier in Step 1-3 and the classification confidence weights of the manifold nearest neighbor classifier. The classification confidence weights of the Euclidean nearest neighbor classifier ,like If so, proceed to step 2-2; otherwise, proceed to step 2-3.

[0109] Step 2-2: If the category outputs are the same, no processing is performed, that is:

[0110]

[0111]

[0112] in, , Let represent the classification confidence scores of the manifold nearest neighbor classifier and the Euclidean nearest neighbor classifier at the edge, respectively. Indicates the edge classification category.

[0113] Steps 2-3: Calculation and And compare them, if If this happens, the classification confidence of the Euclidean nearest neighbor classifier is reduced and its classification confidence weight is lowered, while the classification confidence weight of the manifold nearest neighbor classifier is increased. Finally, the following assignment operation is performed:

[0114]

[0115]

[0116] in, , The downward and upward parameters can be set based on the actual application scenario and are not limited to these two specific values.

[0117] like If this is the case, then the classification confidence of the manifold nearest neighbor classifier is reduced and its classification confidence weight is lowered, while the classification confidence weight of the Euclidean nearest neighbor classifier is increased. Finally, the following assignment operation is performed:

[0118]

[0119]

[0120] In this embodiment, the classification confidence weight layer is initialized with a preset value. The classification confidence weight of the manifold nearest neighbor classifier is... Initialized to 1.5, the classification confidence weights of the Euclidean nearest neighbor classifier. The initial value is set to 1.2.

[0121] Step 3 includes the following steps:

[0122] Step 3-1: Fast classification decision for edge nodes. Edge nodes classify their categories based on the bi-classifier after conflict resolution in Step 2. Classification confidence of manifold nearest neighbor classifier Euclidean nearest neighbor classifier classification confidence Given a sample, classify it according to the following rules: If If the sample is a highly reliable sample with a known class, the class label will be returned directly.

[0123]

[0124] like This means that the samples are highly reliable samples of unknown categories, and labels are assigned to them in ascending order, starting from the highest known category label value:

[0125]

[0126] Then, the manifold and depth features of the signal sample are labeled using this label, and the labeled features are stored as new known class sample features in the edge sample feature database. At the same time, the sample is uploaded to the remote end so that the cloud can update new unknown class sample data in a timely manner.

[0127] The appeal process For known categories, a high-confidence threshold is set. These are high-confidence thresholds for unknown categories, used to control the degree of cloud intervention in edge inference. In practical applications, users can independently weigh the system's recognition accuracy against the current system workload, finding the optimal balance between system recognition accuracy and operational efficiency by adjusting the thresholds. In this embodiment, they are respectively set to... If none of the conditions described during the appeal process are met, the sample is determined to be a confusing sample, and step 3-2 is executed.

[0128] Step 3-2: After determining that a sample is a confused sample, give the decision result of the edge node according to the following rules, upload the result to the cloud and execute step 4:

[0129]

[0130] The existence of known categories affected by domain offset stems from the fact that continuous channel variations in open environments cause a slow shift in the signal's feature distribution. When this domain offset is large enough, even a test sample of a known category may have a distance exceeding a threshold T from its nearest neighbor in the reference set, leading the system to misclassify it as an unknown category. The approach using a dual classifier is based on the assumption of inconsistency in the impact of domain offset on low-dimensional manifold features and high-dimensional depth features. That is, for the same known category, either its manifold nearest neighbor distance or its depth nearest neighbor distance will always exceed its corresponding threshold first. Therefore, a dual classifier structure can be used to capture this difference, thereby determining whether the sample is a known category sample affected by domain offset. If the category is known and affected by the domain offset, then proceed to step 3-3.

[0131] Step 3-3: When a known category sample with domain offset is found at the edge (i.e., a known category affected by the domain offset), the momentum update method is used to update the original prototype features of the corresponding category, and the old prototype features are recorded as follows. The characteristics of the domain offset samples are Then the new prototype of the manifold feature is:

[0132]

[0133] Among them, E and Log For the exponential and logarithmic mapping functions of the manifold space, the new prototype of deep features is:

[0134]

[0135] Step 4 includes the following steps:

[0136] Step 4-1: After receiving the verification request information and confused samples from the edge, the cloud uses the same feature extraction, dual-classifier classification, and classifier conflict handling process as the edge to obtain the predicted category. Furthermore, the classification confidence score and classification confidence weights are adjusted for the output of the binary classifier. , , , .

[0137] Step 4-2: Execute the edge-cloud collaborative inference algorithm, output the edge-cloud collaborative inference results, and feed them back to the edge:

[0138] First, perform the following assignment operation:

[0139]

[0140]

[0141] Next comparison and ,like Then the following assignment operation is performed:

[0142]

[0143]

[0144] Otherwise, perform the following assignment operation:

[0145]

[0146]

[0147] Finally, calculate the confidence score for the joint classification of edge and cloud:

[0148]

[0149] And based on , and Classify according to the rules in step 3-2.

[0150] Step 4-3: The edge continuously receives feedback results from the cloud (based on...) , and (The classification results) are used to correct previous incorrect inferences using the feedback results.

[0151] Figure 2 This is a schematic diagram of the structure of the fast adaptive open set identification system for individual radiation sources based on edge-cloud collaborative reasoning, as described in an embodiment of the present invention. Figure 2 As shown, the system includes:

[0152] Edge nodes are deployed on resource-constrained terminal systems to handle lightweight real-time individual radiation source identification tasks. When edge resources are insufficient, cloud nodes can be invoked through the edge-cloud collaboration mechanism to assist in inference.

[0153] The architecture of the edge nodes is based on a hierarchical design principle, consisting of a signal processing layer, a data management layer, a model algorithm layer, and a human-computer interaction layer. The functions of each layer are as follows:

[0154] Signal Processing Layer: This layer consists of two functional modules: signal acquisition and signal preprocessing. It is primarily used to continuously provide signal sample data in a specified format to the edge model algorithm. The signal acquisition module continuously reads signal file data from the RF signal folder, simulating the process of acquiring signal data using an RF receiver in practical applications. The signal preprocessing module processes the raw data acquired by the signal acquisition module into the data format required by the model algorithm.

[0155] Data Management Layer: This layer stores the prototype features of known class radiation sources required for edge model algorithm inference calculations, as well as log data generated by the edge (including edge inference result logs and prototype feature update logs), and temporarily stores sample data for a period of time.

[0156] Model Algorithm Layer: This layer consists of four functional modules: a classification model, a conflict resolution algorithm, an edge-based fast inference algorithm, and a data update module. The classification model extracts the manifold and depth features of the signal and uses two classifiers (manifold nearest neighbor classifier and Euclidean nearest neighbor classifier) ​​to classify the features, obtaining a preliminary classification result. The conflict resolution algorithm (corresponding to step 2 above) maintains the consistency of the two classification results output by the classification model and dynamically adjusts the classification confidence weights of the two classifiers. The edge-based fast inference algorithm generates the final classification result for the signal samples (corresponding to step 3 above) and uploads easily confused samples to the cloud for verification. The data update module updates the prototype features of known-category radiation source individuals based on the classification results of the edge-based fast inference algorithm, or adds the features of identified unknown-category radiation source individuals to the prototype feature library. It also receives new model parameters from the cloud to update the edge feature extraction model, i.e., updates the temporal feature extraction network.

[0157] Human-Computer Interaction Layer: This layer primarily provides users with a visual interface, consisting of three modules: signal display, recognition result display, and recognition log display. The signal display module provides the time-domain waveform and spectrum of the signal. The recognition result display module displays the classification results and classification confidence levels. The recognition log display model provides recognition statistics during the operation of the edge system, including the total number of recognized signal samples, the total number of signal samples of known categories, and the total number of signal samples of unknown categories.

[0158] Cloud nodes, deployed on resource-rich cloud server platforms, provide asynchronous verification functionality for edge nodes, ensuring the reliability of inference results.

[0159] The structure of the cloud node is based on a hierarchical design principle, consisting of a data management layer, a model algorithm layer, a system management layer, and a human-computer interaction layer. The functions of each layer are as follows:

[0160] Data Management Layer: This layer stores the full sample features of known categories of radiation sources required for cloud model algorithm inference calculations, cloud-generated log data (including cloud inference result logs, sample feature update logs, and model training logs), all collected signal sample data uploaded from the edge, and all historical model parameters from the edge-cloud hybrid system.

[0161] Model Algorithm Layer: This layer consists of four functional modules: classification model, conflict resolution algorithm, edge-cloud collaborative inference algorithm, and data update. The edge-cloud collaborative inference algorithm generates the final classification result for the signal samples and sends the result to the edge. The data update module, in addition to updating the sample feature library, also receives notifications of automatic model training, updates the feature extraction model in the cloud using the newly trained model parameters, and sends the new model parameters to the edge. The remaining modules function essentially the same as the edge modules and will not be described further.

[0162] System Management Layer: This layer consists of two functional modules: edge-cloud conflict monitoring and automatic model training. The edge-cloud conflict monitoring module records the number of edge-cloud conflicts occurring during the execution of the edge-cloud collaborative inference algorithm. When the number of conflicts exceeds the system setting, it sends a model training command to the automatic model training module. The automatic model training module automatically starts the model training process upon receiving the command from the edge-cloud conflict monitoring module or when the system runtime exceeds a specified threshold. After model training is complete, it notifies the data update module to update the system's model.

[0163] Human-Computer Interaction Layer: This layer consists of three modules: signal sample list display, model file list display, and model training display. The signal sample display module shows a list of all signal samples currently stored in the system. The model file list display module shows a list of all model files currently stored in the system. The model training display module shows information about each automatic training iteration of the model.

[0164] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0165] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A fast adaptive open set identification method for individual radiation sources based on edge-cloud collaborative reasoning, characterized in that, Perform the following steps in a radiation source identification system that includes both edge devices and cloud devices: Step 1: The edge device acquires the original radiation source signal input and extracts depth and manifold features based on the local feature extractor. The manifold features are input into the manifold nearest neighbor classifier, and the depth features are input into the Euclidean nearest neighbor classifier to obtain the original manifold nearest neighbor classification output and the original Euclidean nearest neighbor classification output. Step 2: The edge device performs conflict resolution on the initial manifold nearest neighbor classification output and the original Euclidean nearest neighbor classification output, and adjusts the weights of the manifold nearest neighbor classifier and the Euclidean nearest neighbor classifier respectively based on the resolution result. Step 3: The edge device makes a classification decision based on the output of the dual classifier after conflict resolution to obtain confused samples; and submits the current confused samples and their dual classifier classification results to the cloud device for asynchronous verification in real time. Step 4: After receiving the asynchronous verification request and the obfuscated sample, the cloud device performs the same feature extraction, classification, and conflict handling as the edge device based on its local feature extractor, manifold nearest neighbor classifier, and Euclidean nearest neighbor classifier to obtain the cloud's dual classifier classification result. Based on the cloud's dual classifier classification result and the classification result of the received obfuscated sample, it performs edge-cloud collaborative inference processing, outputs the edge-cloud collaborative inference result, and feeds it back to the edge device. The edge device corrects the classification result of the corresponding asynchronous verification obfuscated sample based on the real-time feedback result.

2. The method as described in claim 1, characterized in that, Edge devices and cloud devices are configured with the same feature extractor, manifold nearest neighbor classifier, and Euclidean nearest neighbor classifier.

3. The method as described in claim 1, characterized in that, Deep feature and manifold feature extraction based on feature extractors specifically includes: The temporal features of each time-domain signal sample in each batch are extracted based on the feature extractor. The time-domain signal is converted into the frequency-domain signal using the Discrete Fourier Transform, and then converted into polar coordinates. The resulting amplitude vector and phase vector are used as the frequency-domain features of each signal sample. The time-domain features and frequency-domain features are fused by splicing to obtain the time-frequency fusion features of each signal sample, i.e., the depth features; Calculate the autocorrelation matrix of the depth features of the signal sample, and then perform singular value decomposition on it; Take the first left singular matrix obtained from singular value decomposition The column yields the shape features, among which, This is a preset value.

4. The method as described in claim 1, characterized in that, The feature extractor can employ the backbone network of a recognition network based on a convolutional neural network.

5. The method as described in claim 1, characterized in that, The classification results of the Euclidean nearest neighbor classifier and the manifold nearest neighbor classifier include: classification category and classification confidence; where the classification confidence is set as follows: , The nearest neighbor distance output by the classifier. This is the distance threshold for unknown categories in the classifier.

6. The method as described in claim 1, characterized in that, Conflict resolution includes: definition , These represent the classification categories output by the Euclidean nearest neighbor classifier and the manifold nearest neighbor classifier, respectively. definition , These represent the classification confidence scores output by the Euclidean nearest neighbor classifier and the manifold nearest neighbor classifier, respectively. definition , Let represent the edge Euclidean classification confidence weight and the edge popular classification confidence weight, respectively, and initialize them. , ; Conflict handling determination: If the two classifiers output the same classification category, it means there is no conflict, and the conflict is resolved. or As the current edge category ; and will As edge popularity confidence , As a marginal European confidence level ; If the dual classifier outputs a classification category, conflict handling is performed, specifically including: judge Is it greater than or equal to? If so, then As the current edge category And based on the current , Perform the assignment operation: ； ； Then reduce the downward adjustment based on the settings. Increase based on the setting's upward adjustment range ; Otherwise, As the current edge category And based on the current , Perform the assignment operation: ， ； Then reduce the downward adjustment based on the settings. Increase based on the setting's upward adjustment range .

7. The method as described in claim 1, characterized in that, Step 3 includes: Edge classification decision: If the current edge popular confidence level and edge European confidence All are greater than the preset high-confidence threshold for known categories. If the current signal sample is of a known category, then it is based on the current edge category. Obtain the final category label; like and All are less than the preset high-confidence threshold for unknown categories. If the current signal sample is of an unknown category, it is assigned a label in ascending order starting from the maximum value of the known category label. Based on the assigned label, the manifold feature and depth feature of the signal sample are labeled, and the labeled manifold feature and depth feature are stored as new known class sample features in the edge sample feature database. At the same time, the signal sample and its labeled manifold feature and depth feature are uploaded to the cloud device so that it can be updated with new unknown class sample data in the cloud device in a timely manner. And will not be satisfied The signal sample, or does not meet the requirements The signal samples are treated as scrambled samples; and the scrambled samples and their current signals are treated as scrambled samples. , , and Submit to the cloud device for asynchronous verification; where the symbol " " indicates logical AND; Edge decision processing: If the classification confidence scores of the two classifiers output by the signal samples satisfy... and or satisfy and If so, the current signal sample is considered as a domain offset sample; The momentum update method is used to update the manifold and depth features of signal samples with the same category label as the current domain offset sample.

8. The method as described in claim 7, characterized in that, The manifold and depth features of signal samples with the same class label as the domain offset samples are updated using the momentum update method as follows: ； ； in, , These represent the updated popular features and deep features, respectively, with weight coefficients. , The sum of the two is 1, and Greater than , , These represent the popularity features and depth features before the update, respectively. E represents the depth features of the domain offset sample. and Log These are the exponential and logarithmic mapping functions of the manifold space, respectively.

9. The method as described in claim 1, characterized in that, Step 4 specifically includes: Step 4-1: For the currently received confused samples, the cloud device performs the same feature extraction, classification, and conflict handling as the edge device based on its local feature extractor, manifold nearest neighbor classifier, and Euclidean nearest neighbor classifier to obtain the cloud category. And the corresponding cloud-based European confidence level after conflict resolution. Cloud popularity confidence Cloud-based European classification confidence weights Cloud-based popular classification confidence weights ; Step 4-2: Perform cloud collaborative reasoning processing, output the edge-cloud collaborative reasoning results, and feed them back to the edge device: Calculate confidence in the cloud Cloud-based trust weight ; Calculate edge confidence Edge confidence weight ; like Then the following assignment operation is performed: ； ； ； ； Otherwise, perform the following assignment operation: ； ； ； ； in, This indicates the category determined by cloud collaborative reasoning. This indicates the downward adjustment range set. Indicates the upward adjustment range set; Calculate the confidence score for joint classification of edge clouds: ； in, , These are the confidence scores for the combined popular classification of edge-cloud and the combined Euclidean classification of edge-cloud, respectively. based on , and The classification process is performed to obtain the edge-cloud collaborative inference results; the classification rule is as follows: if and ,but For a known category; if and ,but Unknown category; otherwise For domain offset category; Step 4-3: The edge device corrects the classification results of the corresponding asynchronously verified confused samples based on the feedback results received in real time.

10. A fast adaptive open-set identification system for individual radiation sources based on edge-cloud collaborative reasoning, characterized in that, include Edge nodes are used to handle lightweight real-time individual radiation source identification tasks. When the computing resources of edge nodes are insufficient, cloud nodes are called through the edge-cloud collaboration mechanism to assist in reasoning. Cloud nodes are used to provide asynchronous verification and assisted reasoning functions for edge nodes. The functional layers of the edge node include an edge signal processing layer, an edge data management layer, an edge model algorithm layer, and an edge human-computer interaction layer. in, The edge signal processing layer is used to continuously provide signal sample data in a specified format to the edge model algorithm layer. It includes a signal acquisition module and a signal preprocessing module. The signal acquisition module is used to continuously acquire signal file data as raw data. The signal preprocessing module is used to process the raw data acquired by the signal acquisition module into the data format required by the edge model algorithm layer. The edge data management layer stores the prototype features of known categories of radiation sources required for inference calculations at the edge model algorithm layer, as well as log data generated at the edge, and temporarily stores signal samples for a period of time; the prototype features include depth features and manifold features. The edge model algorithm layer includes a classification model module, a conflict resolution algorithm module, an edge inference algorithm module, and a data update module. The classification model module extracts the manifold and depth features of the signal samples and uses a dual classifier to classify the two types of features, obtaining a preliminary classification result for the signal samples. The dual classifiers are a manifold nearest neighbor classifier and an Euclidean nearest neighbor classifier. The conflict resolution algorithm module maintains consistency between the two classification results output by the classification model module and dynamically adjusts the classification confidence weights of the dual classifiers. The edge inference algorithm module generates the final classification result for the signal samples and uploads confused samples to the cloud node for asynchronous verification. The data update module updates the prototype features of known-category radiation source individuals based on the final classification result of the edge inference algorithm module, or adds the features of identified unknown-category radiation source individuals to the prototype feature library. It also receives new model parameters from the cloud node to update the feature extraction model of the edge node. This feature extraction model extracts the temporal features of the signal samples and fuses the frequency domain features to obtain the depth features of the signal samples. The edge-side human-computer interaction layer provides users with a visual interface, including a signal display module, a recognition result display module, and a recognition log display module. The signal display module provides time-domain waveforms and spectrum diagrams of signal samples. The recognition result display shows the final classification results. The recognition log display model provides recognition statistics during the operation of the edge node, including the total number of recognized signal samples, the total number of signal samples of known categories, and the total number of signal samples of unknown categories. The functional layers of the cloud node include a cloud data management layer, a cloud model algorithm layer, a cloud system management layer, and a cloud human-computer interaction layer; in, The cloud data management layer is used to store sample features of known categories of radiation sources required for inference calculations at the cloud model algorithm layer, log data generated in the cloud, all collected signal sample data uploaded from the edge, and all historical model parameters of the edge-cloud hybrid system. The cloud-based model algorithm layer includes a classification model module, a conflict resolution algorithm module, an edge-cloud collaborative inference algorithm module, and a data update module. The classification model module and conflict resolution algorithm module have the same functions as those on the edge. The edge-cloud collaborative inference algorithm module generates the final classification result for the signal samples and sends this result to the edge nodes. The data update module updates the local sample feature library and, based on received model automatic training notifications, updates the cloud-based feature extraction model with the newly trained model parameters and sends the new model parameters to the edge nodes. The cloud system management layer includes an edge-cloud conflict monitoring module and a model automatic training module. The edge-cloud conflict monitoring module records the number of edge-cloud conflicts that occur during the execution of the edge-cloud collaborative inference algorithm. When the number of conflicts exceeds the system setting, it sends a model training command to the model automatic training module. The model automatic training module automatically starts the model training process for the feature extraction model when it receives the model training command from the edge-cloud conflict monitoring module or when the system running time exceeds a specified threshold. After the model training is completed, it notifies the data update module to update. The cloud-based human-computer interaction layer includes a signal display module, a model file list display module, and a model training display module. The signal display module displays a list of all signal samples currently stored in the system; the model file list display module displays a list of all model files currently stored in the system; and the model training display module displays information about each automatic training iteration of the model.

Images