Radio frequency signal processing method and system based on deep learning

Abstract

The present application provides a radio frequency signal processing method and system based on deep learning. When performing radio frequency signal processing, the target radio frequency signal to be transmitted is obtained, and the target radio frequency signal is decomposed into radio frequency signal segments according to the frequency band to obtain x radio frequency signal segments of the target radio frequency signal; the spectrum distribution of each radio frequency signal segment in the target radio frequency signal is obtained, and the x radio frequency signal segments are mined for implicit representation to obtain the signal implicit representation of each radio frequency signal segment in the x radio frequency signal segments; for each radio frequency signal segment in the x radio frequency signal segments, the radio frequency signal segment is used as a graph element of the signal relationship graph to obtain the signal relationship graph; the target radio frequency signal is processed according to the signal relationship graph. The present application can improve the processing quality of the target radio frequency signal, complete the accurate reasoning of the quality evaluation coefficient of the target radio frequency signal, and provide an accurate quality judgment basis for the optimization of the target radio frequency signal.

Description

Technical Field

[0001] The present application relates to the field of deep learning, and specifically to a radio frequency signal processing method and system based on deep learning. Background Art

[0002] With the rapid development of wireless communication technology, radio frequency signals play a vital role in communication systems. However, with the continuous increase in communication frequency and the increasingly complex communication environment, the processing of radio frequency signals has become increasingly challenging. The quality assessment of radio frequency signals before transmission is a prerequisite for ensuring communication stability and reliability. Traditional evaluation methods place too much emphasis on the quality assessment of local signals, ignoring the overall composition characteristics of the signal and lacking in-depth analysis and modeling of the internal structure of the signal. In particular, the insufficient use of key information such as the spectrum distribution of radio frequency signals and the correlation between signal segments has limited the further improvement of signal processing quality. Summary of the invention

[0003] The present application provides a radio frequency signal processing method and system based on deep learning.

[0004] According to one aspect of the present application, a deep learning-based RF signal processing method is provided, including: obtaining a target RF signal to be transmitted, decomposing the target RF signal into RF signal segments according to the frequency band included in the target RF signal, and obtaining x RF signal segments of the target RF signal, where x≥2; obtaining the spectrum distribution of each of the x RF signal segments in the target RF signal, and performing implicit representation mining on the x RF signal segments to obtain an implicit signal representation of each of the x RF signal segments; for each of the x RF signal segments, using the RF signal segment as an element of a signal relationship graph, and determining the element information of the element based on the implicit signal representation and spectrum distribution of the RF signal segment, and obtaining the signal relationship graph of the target RF signal; processing the target RF signal based on the signal relationship graph of the target RF signal.

[0005] According to another aspect of the present application, a computer system is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method described above.

[0006] The present application includes at least the following beneficial effects: in the process of performing RF signal processing, the present application obtains a target RF signal to be transmitted, and decomposes the target RF signal into RF signal segments according to the frequency band included in the target RF signal to obtain x RF signal segments of the target RF signal; obtains the spectrum distribution of each of the x RF signal segments in the target RF signal, and performs implicit representation mining on the x RF signal segments to obtain the implicit signal representation of each of the x RF signal segments; for each of the x RF signal segments, the RF signal segment is used as a graph element of a signal relationship graph, and the graph element information of the graph element is determined according to the implicit signal representation and spectrum distribution of the RF signal segment to obtain the signal relationship graph of the target RF signal; based on the signal relationship graph of the target RF signal, the target RF signal is processed. Based on this, the present application divides the target RF signal into x RF signal segments based on the frequency band included in the target RF signal, and takes each RF signal segment as a graph element, and determines the graph element information of the graph element based on the implicit signal representation and spectrum distribution of the RF signal segment, to generate a signal relationship graph of the target RF signal. This signal relationship graph includes the spectrum distribution of the RF signal segment in the target RF signal. In this way, when the target RF signal is processed according to the signal relationship graph, the component structure of the target RF signal can be extracted, the processing quality of the target RF signal can be improved, and the quality evaluation coefficient of the target RF signal can be accurately inferred, providing an accurate quality judgment basis for the optimization of the target RF signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 A schematic diagram of an application scenario of a radio frequency signal processing method based on deep learning according to an embodiment of the present application is shown.

[0008] Figure 2 A flowchart of a radio frequency signal processing method based on deep learning according to an embodiment of the present application is shown.

[0009] Figure 3 A schematic diagram of the composition of a computer system according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0010] Figure 1 A schematic diagram of an application scenario provided according to an embodiment of the present application is shown, which includes one or more receiving devices 101, a computer system 1000, and one or more communication networks 110 coupling the one or more receiving devices 101 to the computer system 1000. In an embodiment of the present application, the computer system 1000 may run one or more services or software applications that enable the execution of a deep learning-based radio frequency signal processing method.

[0011] In some embodiments, the computer system 1000 may also provide other services or software applications, which may include non-virtual environments and virtual environments. In some embodiments, these services may be provided as web-based services or cloud services, such as provided to users of the receiving device 101 under a software as a service (SaaS) model.

[0012] exist Figure 1 In the configuration shown, the computer system 1000 may include one or more components that implement the functions performed by the computer system 1000. These components may include software components, hardware components, or a combination thereof that can be executed by one or more processors. A user operating the receiving device 101 may in turn utilize one or more applications to interact with the computer system 1000 to utilize the services provided by these components. Figure 1 is an example of a scenario for implementing the various methods described herein and is not intended to be limiting.

[0013] The computer system 1000 is used to send a radio frequency signal to the receiving device 101 through the network 110. The receiving device 101 may include various types of computer devices, such as portable handheld devices, general-purpose computers (such as personal computers and laptop computers), workstation computers, etc. These computer devices can run various types and versions of software applications and operating systems. The receiving device 101 can execute a variety of different applications, such as various Internet-related applications, communication applications (such as email applications), short message service (SMS) applications, and can use various communication protocols. The network 110 can be any type of network well known to those skilled in the art, which can use any of a variety of available protocols (including but not limited to TCP / IP, SNA, IPX, etc.) to support data communication. The computer system 1000 may include one or more general-purpose computers, special-purpose computer system computers (such as PC (personal computer) computer systems, UNIX computer systems, mid-range computer systems), blade computer systems, mainframe computers, computer system clusters, or any other appropriate arrangement and / or combination. The computing unit in the computer system 1000 can run one or more operating systems including any of the above-mentioned operating systems and any commercially available computer system operating system.

[0014] In some embodiments, the computer system 1000 may include one or more applications to analyze and consolidate data feeds and / or event updates received from users of the receiving device 101. The computer system 1000 may also include one or more applications to display data feeds and / or real-time events via one or more display devices of the receiving device 101. The computer system 1000 may also include one or more databases 130. In some embodiments, these databases may be used to store data and other information. For example, one or more of the databases 130 may be used to store radio frequency signals. The databases 130 may reside in various locations. For example, the database used by the computer system 1000 may be local to the computer system 1000, or may be remote from the computer system 1000 and may communicate with the computer system 1000 via a network-based or dedicated connection. The databases 130 may be of different types. In some embodiments, the databases used by the computer system 1000 may be, for example, relational databases. One or more of these databases may store, update, and retrieve data to and from the databases in response to commands.

[0015] Please refer to Figure 2 , is a flowchart of a radio frequency signal processing method based on deep learning provided in an embodiment of the present application, including:

[0016] Step S100: acquiring a target radio frequency signal to be transmitted, and performing radio frequency signal segment decomposition on the target radio frequency signal according to a frequency band included in the target radio frequency signal to obtain x radio frequency signal segments of the target radio frequency signal, where x≥2;

[0017] In a real-time scenario, assume that a wireless communication base station is receiving RF signals from multiple mobile devices, which may overlap in the frequency domain and carry different information. The computer system on the base station identifies and separates specific target RF signals (for example, call signals from a specific user) for processing. Specifically, the computer system captures RF signals in the wireless environment through the receiving antenna of the base station. These signals contain mixed information from different sources (such as multiple mobile devices). In order to focus on a specific user, the system uses a known synchronization sequence or a specific identifier to identify and lock the target RF signal. The target RF signal usually covers a certain frequency range, which can be subdivided into multiple sub-bands, each of which may carry different information or be used for different communication purposes. The computer system performs frequency domain analysis techniques such as fast Fourier transform (FFT) on the target signal to reveal its frequency band structure. ; where X[k] is the frequency domain representation, x[n] is the time domain signal sample, N is the number of samples, and i is the imaginary unit.

[0018] The computer system decomposes the target RF signal into multiple RF signal segments based on its frequency band structure. For example, if the target signal covers the frequency range from 800 MHz to 1 GHz, and the system detects three distinct frequency bands within this range (800-850 MHz, 850-900 MHz, 900-1000 MHz), the computer will treat each of these frequency bands as a separate RF signal segment.

[0019] Signal segment 1: 800 MHz~850 MHz;

[0020] Signal segment 2: 850 MHz~900 MHz;

[0021] Signal segment 3: 900 MHz~1000 MHz;

[0022] Each signal segment contains all the information of the target signal within that frequency band, but is relatively isolated from the information in other frequency bands.

[0023] After the RF signal segment decomposition, the computer system obtains x RF signal segments (x=3 in this case), each of which is an independent processing unit that can be further analyzed or processed in subsequent steps. In this way, the computer system can effectively decompose complex RF signals into more manageable components, laying the foundation for subsequent tasks such as signal identification, modulation method identification, and channel estimation. This decomposition not only improves the flexibility of signal processing, but also helps reduce interference between different signal segments and improve the overall performance of the communication system.

[0024] Step S200: Obtain the spectrum distribution of each of the x RF signal segments in the target RF signal, and perform implicit representation mining on the x RF signal segments to obtain an implicit signal representation of each of the x RF signal segments.

[0025] In the embodiment of the present application, step S200 is a step of further analyzing the RF signal segments decomposed by step S100, aiming to deeply understand the spectrum characteristics of each RF signal segment in the target RF signal and capture the high-order features of these signal segments through implicit representation mining.

[0026] For example, assuming that the target radio frequency signal has been decomposed into x radio frequency signal segments through step S100, each signal segment represents information of the target signal in a specific frequency band. Now, the computer system performs a more detailed analysis on these signal segments so that the information can be used more effectively in subsequent processing.

[0027] First, the computer system determines the specific location information of each RF signal segment in the target RF signal spectrum. This includes the center frequency of the band, the bandwidth, and the relative position of the band in the entire spectrum. This information is critical to understanding the physical characteristics of the signal segments and their role in the communication system.

[0028] Regarding center frequency: For example, the center frequency of the first RF signal segment may be 850 MHz, while the center frequency of the second RF signal segment may be 950 MHz.

[0029] Regarding bandwidth: The bandwidth of each signal segment may be different, reflecting the width of the signal occupied within the frequency band. For example, the bandwidth of the first signal segment may be 50 MHz, while the bandwidth of the second signal segment may be 70 MHz.

[0030] Regarding relative position: This describes the order and spacing of the signal segments in the entire spectrum. For example, it can be expressed as the ratio of the difference between the start frequency of the signal segment and the lowest frequency of the spectrum to the width of the entire spectrum.

[0031] Next, the computer system mines the implicit representation of each RF signal segment. This process usually involves using machine learning algorithms (such as autoencoders, variational autoencoders, etc.) to extract high-level features of the signal segment that can better represent the intrinsic properties and structure of the signal segment. The resulting implicit representation can be a feature vector.

[0032] As an example of an autoencoder, a computer system can train an autoencoder to learn a low-dimensional implicit representation of each RF signal segment. The autoencoder compresses the high-dimensional RF signal segment data into a low-dimensional implicit vector through the encoder part, and then attempts to reconstruct the original signal segment through the decoder part. During the training process, the parameters of the encoder and decoder are optimized by minimizing the reconstruction error.

[0033] Assume that the data of the RF signal segment is represented by a matrix X∈R m×n , where m is the number of samples (in this case it may be only 1, because one signal segment is processed at a time), and n is the number of features (such as the number of sampling points of the signal). The encoder function of the autoencoder can be expressed as f enc (X)=z, where z∈R d is a low-dimensional implicit vector (d≪n). The decoder function attempts to reconstruct the original signal segment from z, i.e., g dec (z)=X'.

[0034] The resulting implicit representation vector z captures the key features of the RF signal segment, which can be used for subsequent tasks such as signal classification, recognition, compression or denoising. For example, in a signal classification task, z can be passed as input to a classifier (such as a support vector machine, neural network, etc.) to predict the category of the signal segment.

[0035] In summary, step S200 provides rich feature information for subsequent RF signal processing by acquiring the spectrum distribution of the RF signal segment and performing implicit representation mining, so that the computer system can understand and process complex RF signals more accurately.

[0036] Step S300: For each of the x RF signal segments, use the RF signal segment as a graph element of the signal relationship graph, and determine the graph element information of the graph element based on the implicit signal representation and spectrum distribution of the RF signal segment to obtain the signal relationship graph of the target RF signal.

[0037] In an embodiment of the present application, step S300 integrates the RF signal segments and their characteristics extracted in the previous steps into a unified structure to facilitate subsequent analysis and processing. This step is automatically performed by the computer system, which intuitively displays the relationship and properties between the RF signal segments by creating a signal relationship graph. For example, assume that after processing steps S100 and S200, the computer system has obtained x RF signal segments of the target RF signal, and each signal segment has its corresponding spectrum distribution and implicit representation. Now, the system needs to integrate this information into a signal relationship graph to better understand the overall structure and characteristics of the signal. In the signal relationship graph, each RF signal segment is regarded as an independent graph element (that is, a vertex). These graph elements represent the performance of the signal within a specific frequency band. For example, if x=3, the graph will contain three graph elements, corresponding to three RF signal segments. For each graph element, the computer system determines the detailed information of the graph element based on the implicit signal representation and spectrum distribution of the corresponding RF signal segment. This information may include:

[0038] Center frequency: For example, 850 MHz, which indicates the center position of the signal segment in the spectrum.

[0039] Bandwidth: For example, 50 MHz, which describes the coverage of the signal segment in the frequency domain.

[0040] Implicit representation vector: A low-dimensional vector that captures the high-level features of a signal segment, such as z=[z1,z2,…,zd], where d is the dimension of the implicit representation.

[0041] Other attributes: such as signal strength, modulation mode, etc. (if this information has been extracted in the previous step).

[0042] In some cases, the signal relationship graph may also contain the connection relationships between the graph elements. These connections can be based on similarity, adjacency or other logical associations between signal segments. However, in basic RF signal processing scenarios, this step may be optional because the main goal is to analyze the characteristics of a single signal segment.

[0043] Finally, the computer system generates a signal relationship map based on all the graph elements and their information. This map shows the structure of the RF signal in a graphical way, allowing researchers or algorithms to intuitively understand the composition and characteristics of the signal.

[0044] For example, suppose there are three RF signal segments, corresponding to GSM, LTE, and 5G signals. Each signal segment has its own unique spectrum distribution and implicit representation. In the signal relationship graph, these three signal segments will be represented as three independent graph elements, each of which is annotated with its center frequency, bandwidth, and a simplified form of the implicit representation vector (such as only showing a few key dimensions in the vector).

[0045] GSM signal segment: element 1, center frequency 850 MHz, bandwidth 25 MHz, implicit representation zGSM=[0.2,−0.1,…];

[0046] LTE signal segment: Element 2, center frequency 1800 MHz, bandwidth 100 MHz, implicit representation zLTE=[0.5,0.3,…];

[0047] 5G signal segment: element 3, center frequency 3.5 GHz, bandwidth 400 MHz, implicit representation z5G = [−0.1, 0.8, …];

[0048] The signal relationship graph integrates this information to form a comprehensive view of the target RF signal, providing strong support for subsequent tasks such as signal identification, classification, and interference analysis.

[0049] Step S400: Processing the target radio frequency signal according to the signal relationship graph of the target radio frequency signal.

[0050] In the embodiment of the present application, step S400 aims to optimize the signal processing effect through the information contained in the signal relationship graph, and improve the performance of possible derivative tasks such as signal quality identification, classification, and denoising. For example, it is assumed that the computer system has successfully obtained the signal relationship graph of the target RF signal through steps S100 to S300, and the graph describes in detail the spectral distribution, implicit representation, and relationship between each RF signal segment in the signal. Now, the computer system further processes the target RF signal according to this graph.

[0051] For example, in the application scenario of signal recognition and classification, the computer system can use the implicit representation vector in the signal relationship graph as a feature input into the classifier. For example, machine learning models such as support vector machine (SVM), random forest or deep neural network can be used for classification. These models have been trained with a large amount of labeled data and can learn the characteristic differences of signals of different categories. The specific process can refer to the existing technology. For example, as a simple introduction, the implicit representation vector of each RF signal segment can be extracted from the signal relationship graph as a feature. Then the feature is input into the trained classifier to obtain the category prediction result of the target RF signal.

[0052] In the signal denoising scenario, the spectrum distribution information in the signal relationship graph can help the computer system identify and remove the noise components in the signal. By comparing the difference between the performance of the signal segment in the spectrum and the known noise characteristics, the system can locate and suppress the noise. Let's take a simple example. First, use the spectrum distribution information in the signal relationship graph to analyze the distribution characteristics of the noise in the frequency domain. According to the noise characteristics, design the corresponding filter or denoising algorithm to denoise the target RF signal.

[0053] In the embodiment of the present application, as mentioned above, signal quality assessment is an important task before transmission. In the application scenario of signal quality assessment, the computer system can comprehensively evaluate the quality of the signal based on multiple indicators in the signal relationship graph (such as signal strength, bandwidth, specific dimensions of the implicit representation vector, etc.). This can be achieved by defining a comprehensive scoring function, which weights and sums multiple indicators to obtain a quantitative score of the signal quality. The computer system can extract multiple indicators related to signal quality from the signal relationship graph. Then define a scoring function, such as , where I i is the value of the ith indicator, w i is the corresponding weight, and n is the total number of indicators. By adjusting the weight, the impact of certain indicators on signal quality can be emphasized or ignored.

[0054] For example, suppose the target RF signal is a mixed signal containing multiple communication protocols, and the computer system identifies the GSM, LTE, and 5G signal segments. In the signal relationship graph, each signal segment has its own unique implicit representation vector. The computer system first extracts the implicit representation vector of each signal segment as a feature. Then, these features are input into a pre-trained deep neural network classifier. The classifier outputs the probability that each signal segment belongs to GSM, LTE, or 5G based on the pattern learned from the features. Finally, the computer system further processes the target RF signal based on the classification results, such as separating signal segments of different protocols, or adjusting the parameters of the receiver based on the signal quality evaluation results to optimize the reception effect.

[0055] Through step S400, the computer system can make full use of the rich information in the signal relationship graph to efficiently and accurately process the target radio frequency signal to meet the needs of various practical application scenarios.

[0056] As a possible implementation manner, in step S300, determining the image element information of the image element according to the implicit signal representation and spectrum distribution of the radio frequency signal segment may specifically include:

[0057] Step S310: obtaining the time-frequency signal strength distribution of the radio frequency signal segment;

[0058] Step S320: Determine the graphic element information of the graphic element according to the implicit signal representation, spectrum distribution and time-frequency signal strength distribution of the radio frequency signal segment.

[0059] In step S310, the computer system expands the RF signal segment from a single time domain or frequency domain representation to the time-frequency domain, thereby revealing the variation of signal strength over time and frequency. This representation method has a significant effect on understanding the complex dynamic characteristics of the signal, especially when processing non-stationary signals.

[0060] Specifically, the computer system performs necessary preprocessing on the RF signal segment, such as denoising and filtering, to improve signal quality and reduce the impact of interference factors on subsequent analysis. Next, the system uses time-frequency analysis technology to convert the RF signal segment from the time domain to the time-frequency domain. Commonly used time-frequency transformation methods include short-time Fourier transform (STFT), wavelet transform, Wigner-Ville Distribution, etc. Here we take STFT as an example: ;in, is the representation of the RF signal segment in the time domain, w(τ−t) is the window function used to intercept the signal in the local time range, and f is the frequency variable. The result of STFT is a complex matrix whose elements represent the signal strength or amplitude at a specific time t and frequency f.

[0061] The computer system visualizes the results of the STFT as a time-frequency diagram. In this two-dimensional image, the horizontal axis represents time and the vertical axis represents frequency. The grayscale or color depth of each pixel represents the signal strength at the corresponding time point and frequency. For example, high-intensity signals may be displayed in bright colors, while low-intensity signals are represented in dark colors. Through image processing technology, the system can extract the spatial distribution of signal intensity from the time-frequency diagram. This usually involves statistical analysis of pixel values, such as calculating the mean, variance, histogram, etc., to quantify the overall distribution characteristics of signal strength.

[0062] Assume that a wireless communication system is receiving radio frequency signals from multiple users. These signals may be affected by factors such as multipath effects and Doppler frequency shift during transmission, causing the signals to exhibit complex dynamic characteristics in the time-frequency domain. In order to accurately identify each user's signal and estimate its channel parameters, the computer system performs time-frequency analysis on the received radio frequency signal segments.

[0063] Through step S310, the system can generate a time-frequency diagram for each RF signal segment and observe the change of signal strength over time and frequency. For example, for a mobile user, its signal may show a slant line that changes over time on the time-frequency diagram, which is due to the Doppler frequency shift effect caused by the user's movement. For a stationary user, its signal may be concentrated on a fixed frequency point, and the signal strength distribution on the time-frequency diagram is relatively stable.

[0064] After obtaining the time-frequency signal strength distribution of the RF signal segment, in step S320, the computer system combines it with the implicit representation of the signal and the spectrum distribution to jointly determine the element information of the element in the signal relationship graph. This process involves the fusion and comprehensive processing of multi-source information, aiming to fully reflect the intrinsic characteristics and external performance of the RF signal segment.

[0065] Specifically, first, the computer system designs an effective information fusion strategy to integrate three different types of information: implicit signal representation, spectrum distribution, and time-frequency signal strength distribution. This can be achieved through feature concatenation, weighted summation, deep learning model fusion, etc. In order to construct the primitive information of the primitive, the system needs to extract the most representative features from the fused information. These features should be able to fully reflect the essential properties of the RF signal segment, such as center frequency, bandwidth, modulation mode, signal strength distribution pattern, etc. Based on the extracted features, the system can construct the primitive information of each primitive in the signal relationship map. This information may include the unique identifier of the primitive, the description of the signal segment to which it belongs, spectrum parameters, time-frequency characteristic description, implicit representation vector, etc. Finally, the system integrates the primitive information of all primitives into the signal relationship map and updates and maintains the map as needed. With the addition of new RF signal segments or changes in the characteristics of existing signal segments, the map also needs to be adjusted and expanded accordingly.

[0066] Continuing to take the wireless communication system as an example, assuming that the system has obtained the implicit representation and spectrum distribution information of each radio frequency signal segment through step S200, the system further combines the time-frequency signal strength distribution to improve the construction of the signal relationship map.

[0067] In step S320, the computer system first uses the time-frequency graph generated in step S310 to analyze the time-frequency characteristics of each signal segment. For example, by observing the distribution pattern of signal strength on the time-frequency graph, the system can identify possible interference components or changes in modulation methods in the signal. Then, the system combines the time-frequency characteristics with the implicit representation and the spectrum distribution to jointly determine the primitive information of the primitive.

[0068] Taking image element 1 as an example (assuming it represents the first RF signal segment), its image element information may include:

[0069] Unique identifier: ID_Segment1;

[0070] Spectrum parameters: center frequency 850 MHz, bandwidth 50 MHz;

[0071] Description of time-frequency characteristics: The signal strength changes relatively steadily over time, but there is a brief increase in a specific time period, which may correspond to the burst transmission of user data.

[0072] Implicitly represents the vector: z Segment1 =[0.2,−0.1,0.3,…], this vector captures the high-order features of the signal segment and helps with subsequent classification, recognition and other tasks.

[0073] By constructing such detailed primitive information, the computer system can more comprehensively understand the structure and characteristics of the RF signal, providing strong support for subsequent tasks such as signal processing, interference suppression, and user identification. At the same time, as new signal segments are added or the characteristics of existing signal segments change, the system can update the signal relationship map in real time to maintain its timeliness and accuracy.

[0074] As a possible implementation manner, the step S400, processing the target radio frequency signal according to the signal relationship graph of the target radio frequency signal, may specifically include:

[0075] Step S410: determining a reference RF signal set, wherein the reference RF signal set includes a quality evaluation coefficient of each of t first RF signals, and a signal implicit representation or a signal relationship graph of each first RF signal, wherein the signal implicit representation of the first RF signal is determined based on the signal relationship graph of the first RF signal, and t≥2;

[0076] Step S420: determining a commonality measurement result between the target RF signal and each of the t first RF signals according to the signal relationship graph of the target RF signal and the signal implicit representation or signal relationship graph of each first RF signal;

[0077] Step S430: selecting s first radio frequency signals from the t first radio frequency signals according to the commonality measurement result, wherein 1<s≤t;

[0078] Step S440: Determine the quality evaluation coefficient of the target RF signal according to the quality evaluation coefficients of the s first RF signals.

[0079] In step S410, the computer system constructs a set of multiple reference RF signals. These reference signals should not only reach a certain scale in number (at least t, where t≥2), but also be as diverse as possible in type and characteristics, so as to fully reflect the diversity of RF signals. Each reference RF signal is accompanied by its quality evaluation coefficient and implicit signal representation or signal relationship map.

[0080] Specifically, the computer system can collect reference RF signals from multiple channels, including but not limited to historical records of communication base stations, public RF signal data sets, simulated signals generated in experimental environments, etc. The collected signals should cover different communication protocols, modulation methods, channel conditions, etc. to ensure the breadth and representativeness of the reference set. For each reference RF signal, the computer system first performs preprocessing, such as denoising, filtering, synchronization, etc., to improve the signal quality. Subsequently, a deep learning model (such as an autoencoder, a convolutional neural network, etc.) is used to extract features from the signal to obtain an implicit representation vector for each signal. These vectors can capture the high-order features of the signal and provide a basis for subsequent analysis. Based on the extracted implicit representation vector and the spectral distribution information of the signal, the computer system further constructs a signal relationship map for each reference RF signal. This map graphically displays the intrinsic structure and characteristics of the signal, providing an intuitive basis for commonality measurement. Each reference RF signal is accompanied by a quality evaluation coefficient, which reflects the quality level of the signal. This coefficient can be determined in a variety of ways, such as based on performance indicators such as the signal-to-noise ratio, bit error rate, and demodulation success rate of the signal; or obtained through subjective evaluation methods such as expert scoring and user feedback. For example, suppose a wireless communication system needs to perform quality assessment on the received target RF signal. To accomplish this task, the computer system collects reference RF signals from multiple sources, including GSM, LTE, 5G and other communication protocols. For each reference signal, the system first performs preprocessing and feature extraction to obtain its implicit representation vector. Then, based on these vectors and the signal's spectrum distribution information, the system constructs a signal relationship graph for each reference signal. Finally, based on performance indicators such as the signal-to-noise ratio and bit error rate of the signal, the system assigns a quality evaluation coefficient to each reference signal.

[0081] After determining the reference RF signal set, in step S420, the computer system compares the similarity or commonality between the target RF signal and each reference signal. This process is achieved by calculating the commonality measurement result, which provides a basis for subsequent signal selection and quality evaluation.

[0082] Specifically, the computer system first obtains the signal relationship graph of the target RF signal. Then, the graph is matched and compared with the signal relationship graph or implicit representation vector of the reference RF signal. The matching process may involve a variety of methods such as similarity measurement of graph structure and distance calculation of feature vectors. In order to accurately calculate the commonality measurement results, the computer system selects a suitable measurement algorithm. These algorithms may include similarity measurement based on graph edit distance, vector distance calculation based on cosine similarity, feature matching based on deep learning, etc. The choice of algorithm should be determined according to the specific application scenario and requirements. After the calculation is completed, the computer system will output the commonality measurement results between the target RF signal and each reference RF signal, and store them in the corresponding data structure for subsequent use.

[0083] For example, let's continue with the quality assessment of wireless communication systems. In this step, the computer system first obtains the signal relationship graph of the target RF signal. Then, the system uses a graph-based similarity measurement algorithm to compare the similarity between the target signal and the graph of each reference signal. The algorithm may consider multiple dimensions such as the number of nodes in the graph, the connection pattern of the edges, and the characteristic vectors of the nodes to calculate the commonality measurement results. Through this process, the system obtains the commonality measurement values ​​between the target signal and each reference signal, which reflect the degree of similarity between them.

[0084] After obtaining the commonality measurement results, in step S430, the computer system selects s ​​signals (where 1<s≤t) that are most similar to the target RF signal from the reference RF signal set according to these results. This process aims to approximate the characteristics of the target signal through a few representative reference signals.

[0085] Specifically, the computer system first sorts the reference RF signals according to the commonality measurement results. The basis for sorting can be the size of the commonality measurement value, the degree of similarity with the target signal, etc. Then, the system selects the first s signals from the sorted list as the final selection result. In some cases, in order to ensure that the selected signals have a certain representativeness and quality level, the computer system may set a commonality measurement threshold. Only those signals whose commonality measurement values ​​exceed the threshold are eligible to be selected. In addition, the system can also perform further screening based on other conditions (such as the quality evaluation coefficient of the signal, source reliability, etc.).

[0086] For example, in the quality assessment of wireless communication systems, the computer system sorts the reference RF signals according to the commonality metric results. In order to ensure that the selected signals have a certain representativeness and quality level, the system sets a commonality metric threshold. Then, the system selects the top s signals with the highest commonality metric values ​​from the sorted list as the final selection results. These signals have a high similarity with the target RF signal in terms of communication protocol, modulation mode, channel conditions, etc., and can therefore serve as a reliable reference for evaluating the quality of the target signal.

[0087] After selecting the s reference signals most similar to the target RF signal, in step S440, the computer system determines the quality evaluation coefficient of the target signal based on the quality evaluation coefficients of these signals. This process usually involves multiple methods such as weighted average, voting mechanism, fuzzy comprehensive evaluation, etc.

[0088] Specifically, the computer system can use a weighted average method to calculate the quality evaluation coefficient of the target radio frequency signal. Specifically, the system first assigns corresponding weights to the selected s reference signals (the weights can be determined based on factors such as the commonality measurement results and the reliability of the signal source). Then, the system multiplies the quality evaluation coefficients of these signals by the corresponding weights and sums them to obtain the quality evaluation coefficient of the target signal. This method is simple, intuitive and easy to implement. Alternatively, in some cases, in order to more fairly reflect the opinions and contribution of each reference signal, the computer system can use a voting mechanism to determine the quality evaluation coefficient of the target radio frequency signal. Specifically, the system first sets a voting weight for each of the selected s reference signals (the weights can be determined based on the commonality measurement results). Then, each reference signal votes for the target signal according to its own quality evaluation coefficient (for example, using the quality evaluation coefficient as a voting value). Finally, the system calculates the quality evaluation coefficient of the target signal based on the voting results (for example, by summing or averaging).

[0089] Alternatively, for some quality characteristics that are difficult to describe with precise numerical values ​​(such as signal stability, anti-interference ability, etc.), the computer system can use a fuzzy comprehensive evaluation method to determine the quality evaluation coefficient of the target RF signal. Specifically, the system first constructs a fuzzy comprehensive evaluation model and determines the evaluation index system and weight distribution; then the selected s reference signals are input into the model as evaluation samples for calculation; finally, the quality evaluation coefficient of the target signal and its membership distribution are obtained based on the calculation results. For example, in the quality assessment of wireless communication systems, the computer system uses a weighted average method to calculate the quality evaluation coefficient of the target RF signal. First, the system assigns corresponding weights to the selected s reference signals based on the commonality measurement results (the weight size is proportional to the commonality measurement value). Then the system multiplies the quality evaluation coefficients of these signals with the corresponding weights and sums them up to obtain the quality evaluation coefficient Q of the target signal: ; where w i is the weight of the i-th reference signal; q i is the quality evaluation coefficient of the i-th reference signal; s is the number of selected reference signals. Through this process, the system obtains an evaluation coefficient Q that can reflect the overall quality level of the target RF signal. This coefficient can be used in subsequent decision support, performance optimization and other tasks.

[0090] In summary, step S400 achieves in-depth analysis and processing of the target RF signal through a series of operations such as constructing a reference RF signal set, calculating commonality measurement results, selecting representative signals, and determining quality evaluation coefficients. This process not only relies on advanced features such as signal relationship graphs and implicit representations, but also combines the powerful capabilities of machine learning models and algorithms to provide a more accurate and efficient solution for the field of RF signal processing.

[0091] As a possible implementation manner, the step S420, determining the commonality measurement result of the target RF signal and each of the t first RF signals according to the signal relationship graph of the target RF signal and the signal implicit representation or signal relationship graph of each first RF signal, may specifically include:

[0092] Step S421: if the reference RF signal set includes the signal relationship graphs of the t first RF signals, determining a commonality measurement result between the target RF signal and each of the t first RF signals according to the signal relationship graph of the target RF signal and the signal relationship graphs of the t first RF signals;

[0093] Step S422: If the reference RF signal set includes the implicit signal representation of the t first RF signals, then determine the commonality measurement result between the target RF signal and each of the t first RF signals based on the implicit signal representation of the target RF signal and the implicit signal representation of the t first RF signals.

[0094] In the embodiment of the present application, step S420 evaluates the similarity or commonality between the target RF signal and each signal in the reference set by calculating the commonality measurement result. This process not only helps the subsequent signal selection and quality evaluation, but also provides an important decision-making basis for the processing of RF signals.

[0095] In step S421, when the reference RF signal set includes signal relationship graphs of t first RF signals, the computer system determines the commonality measurement result between them by comparing the signal relationship graph of the target RF signal with the signal relationship graph of each first RF signal.

[0096] Specifically, first, the computer system performs structural alignment on the signal relationship graphs of the target RF signal and each first RF signal. Since the graphs of different signals may contain different numbers of nodes and edges, the alignment process may involve operations such as scaling, cropping, or interpolation of the graphs to ensure that the graphs are structurally comparable. After the graph structures are aligned, the system needs to extract the feature vectors of the nodes in each graph. These feature vectors may be obtained by extracting features from the nodes in the graph using a deep learning model (such as a convolutional network GCN), which captures high-order features of the nodes and their neighborhoods. Then, the system determines the similarity between the target graph and the reference graph by comparing the feature vectors of the corresponding nodes.

[0097] Based on the matching results of node feature vectors, the system can further calculate the similarity between the entire graphs. There are many methods for measuring graph similarity, including but not limited to Graph Edit Distance (GED), Graph Kernels, and learning-based graph matching algorithms. For deep learning scenarios, a common practice is to use Graph Neural Networks (GNNs) to directly learn similarity functions between graphs.

[0098] After the calculation is completed, the system outputs the commonality measurement results between the target RF signal and each first RF signal. These results are usually expressed in numerical form, and the larger the numerical value, the higher the commonality, that is, the more similar the two signals are in structure and characteristics. For example, suppose a wireless communication base station is performing a quality assessment on the received target RF signal. The base station has a reference set containing signal relationship graphs of various communication protocols (such as GSM, LTE, 5G, etc.). In order to evaluate the quality of the target signal, the base station first generates a signal relationship graph of the target signal and performs structural alignment and feature vector matching with each graph in the reference set. Then, the base station uses a graph similarity measurement method based on a graph convolutional network (GCN) to calculate the commonality measurement results between the target signal and the reference signal. By comparing these results, the base station can identify the reference signal that is most similar to the target signal and infer the quality level of the target signal based on this.

[0099] In step S422, when the reference RF signal set includes signal implicit representations of t first RF signals, the computer system determines the commonality measurement result between them by comparing the signal implicit representation of the target RF signal with the signal implicit representation of each first RF signal.

[0100] Specifically, first, the computer system ensures that implicit representations of the target RF signal and each first RF signal are available. These implicit representations are usually obtained by extracting features from the original signal through a deep learning model (such as an autoencoder, a variational autoencoder, etc.), and they capture the high-order features of the signal in the form of a low-dimensional vector. After extracting the implicit representation, the system can evaluate the similarity between the target signal and the reference signal by calculating the distance in the vector space. Common distance measurement methods include Euclidean distance, Manhattan distance, cosine similarity, etc. For deep learning scenarios, the system may also use more complex metric learning methods to optimize the distance metric function. In addition to directly calculating the distance, the system can also design more complex commonality metric functions to comprehensively consider multiple factors. For example, the system can adjust the weight parameters of the metric function according to the specific application scenario of the signal; or use an integrated learning method to combine the output results of multiple base metrics to obtain a more accurate commonality metric value.

[0101] After the calculation is completed, the system outputs the commonality measurement results between the target RF signal and each first RF signal, and may provide additional explanatory information to help users understand the meaning and source of the measurement results. For example, consider a satellite communication system that needs to identify and classify uplink signals from different ground stations. Due to the large number of ground stations and different signal characteristics, the satellite communication system requires an efficient method to distinguish these signals. To this end, the system first extracts implicit representation vectors for each ground station signal and stores them in a reference set. When a new uplink signal is received, the system first extracts its implicit representation vector and performs cosine similarity calculation with the vectors in the reference set. By comparing the size and distribution characteristics of the calculation results, the system can identify which ground station's signal the new signal is most similar to, thereby completing the signal identification and classification tasks. In addition, the system can also evaluate the quality level or abnormality of the new signal based on the size of the commonality measurement results, providing an important reference for subsequent signal processing.

[0102] In summary, step S420 implements the similarity evaluation between the target RF signal and the reference RF signal set through a commonality measurement method based on a signal relationship graph or implicit signal representation. These two methods have their own advantages and disadvantages and are suitable for different application scenarios and requirements. In practical applications, the computer system can select an appropriate method to calculate the commonality measurement result according to the specific situation, and combine other information to comprehensively judge the quality and characteristics of the target RF signal.

[0103] As a possible implementation manner, in the step S421, determining the commonality measurement result between the target RF signal and each of the t first RF signals according to the signal relationship graph of the target RF signal and the signal relationship graph of the t first RF signals may specifically include:

[0104] Step S4211: determining a signal implicit representation of the target radio frequency signal according to a signal relationship graph of the target radio frequency signal;

[0105] Step S4212: determining implicit signal representations of the t first radio frequency signals according to the signal relationship graph of the t first radio frequency signals;

[0106] Step S4213: determining a commonality measurement result between the target RF signal and each of the t first RF signals according to the implicit signal representation of the target RF signal and the implicit signal representation of the t first RF signals.

[0107] In step S4211, the computer system extracts an implicit representation that can characterize the intrinsic characteristics of the signal from the signal relationship graph of the target RF signal. This implicit representation is usually a low-dimensional vector that captures information in the signal that is difficult to observe intuitively but is critical for subsequent processing.

[0108] Specifically, the computer system first parses the signal relationship graph of the target RF signal. This graph is a complex graph structure consisting of multiple nodes and edges, where each node represents an RF signal segment in the signal, and the edges reflect the relationship between nodes (such as frequency adjacency, time order, etc.). In order to extract useful features from the graph, the computer system can use deep learning models such as graph neural networks (GNNs). GNNs are able to process graph-structured data and update the representation of nodes through information transfer and aggregation between nodes. In this process, the representation of each node gradually fuses the information of its neighborhood to form a richer feature vector. For example, a graph convolutional network (GCN) can be used to update the representation of a node. GCN iteratively updates node features in the following way: ; where Ã=A+I is the adjacency matrix A of the undirected graph G plus the matrix after self-connection, I is the identity matrix, is the degree matrix of Ã, is the node feature matrix of the lth layer, W(l) is the learnable weight matrix, and σ is the nonlinear activation function (such as ReLU). After multiple rounds of graph neural network iterations, the representation of each node has integrated a large amount of information in the graph. The computer system can select the node features output by the last layer of GNN as the implicit representation of the signal. For the spectrum of the target RF signal, the representations of all the nodes in it can be aggregated (such as summation, averaging, etc.) to obtain the implicit representation vector of the entire signal.

[0109] For example, consider a wireless communication system that needs to evaluate the similarity of a received target RF signal with multiple known signals. The graph of the target signal contains multiple RF signal segments and the relationships between them. The computer system first uses a graph neural network to extract features from each node in the graph and obtains an implicit representation vector of the entire signal through aggregation. This vector may contain information about multiple dimensions such as signal modulation, bandwidth, frequency distribution, etc., providing a basis for subsequent commonality measurements.

[0110] Step S4212 is similar to step S4211, and the computer system also performs the same processing on each first radio frequency signal in the reference set to extract their signal implicit representations. This step is processed in parallel, that is, the spectra of t signals are analyzed and feature extracted at the same time.

[0111] Specifically, in order to improve efficiency, the computer system can use parallel computing technology to process multiple signals in the reference set. This means that multiple GNN model instances can be started at the same time, and each instance is responsible for processing the spectrum of a signal. For the spectrum of each first RF signal, the computer system performs the same feature extraction and representation generation process as in step S4211. That is, the nodes in the spectrum are iteratively updated using a graph neural network, and finally the implicit representation vector of the signal is aggregated. The extracted implicit representation vector of each first RF signal is stored in the memory of the computer system for subsequent comparison with the implicit representation of the target signal.

[0112] For example, let's continue with the wireless communication system. In this system, the reference set contains multiple known first radio frequency signals and their corresponding signal relationship graphs. In order to evaluate the similarity between the target signal and these signals, the computer system analyzes the graph of each reference signal in parallel. The implicit representation vector of each signal is calculated and stored. These vectors contain the key feature information of the reference signal, providing a data basis for subsequent commonality measurement.

[0113] In step S4213, the computer system uses the previously extracted implicit representation of the signal to calculate the commonality measurement result between the target RF signal and each first RF signal. This process usually involves distance calculation or similarity evaluation in a vector space.

[0114] Specifically, a simple and intuitive method is to calculate the Euclidean distance between the target signal and each reference signal implicit representation vector. The Euclidean distance measures the straight-line distance between two vectors in space. The smaller the distance, the more similar the two vectors are. Where v1 and v2 are implicit representation vectors of the target signal and a reference signal, respectively, and n is the dimension of the vector. In addition to distance calculation, cosine similarity can also be used to evaluate the similarity between two vectors. Cosine similarity measures the closeness of two vectors in direction, regardless of their size. in, represents the dot product operation of vectors, The Euclidean norm of the representation vector. Based on the calculated distance or similarity value, the computer system can generate a commonality measurement result between the target signal and each reference signal. These results are usually expressed in numerical form and used for subsequent sorting, screening or decision support. For example, in a wireless communication system, in order to evaluate the commonality between the target RF signal and each first RF signal in the reference set, the computer system first calculates the Euclidean distance or cosine similarity between the implicit representation vector of the target signal and each reference signal. Then, a commonality measurement result is generated based on these calculation results. These results help the system identify one or more signals in the reference signal set that are most similar to the target signal. For example, in a signal recognition task, the system may select the first few reference signals with the highest commonality measurement results as candidate signals for further analysis; in a signal quality assessment task, the receiving parameters of the target signal may be adjusted or error correction operations may be performed based on the commonality measurement results. In summary, step S421 implements the similarity assessment between the target RF signal and the reference RF signal set through a series of operations such as parsing the signal relationship graph, extracting the implicit representation of the signal, and calculating the commonality measurement results. This process not only relies on advanced deep learning models (such as neural networks) to extract high-order features of signals, but also combines effective distance calculation and similarity assessment methods to quantify the degree of commonality between signals.

[0115] As a possible implementation method, the step S430, based on the commonality measurement result, selects s ​​first RF signals from the t first RF signals, which may specifically include: among the t first RF signals, determining s first RF signals whose commonality measurement results with the target RF signal are not less than the first reference commonality value.

[0116] In the application scenario of RF signal processing, step S430 is a key step in selecting reference signals that are most similar to or have commonalities with the target RF signal. By comparing the commonality measurement results, the computer system can select several signals that are closest to the target signal from a large number of reference signals, providing strong support for subsequent signal processing or quality assessment.

[0117] In step S430, the computer system selects s ​​signals from t first RF signals according to the commonality measurement results. To achieve this goal, the system first sets a first reference commonality value, which is used as a threshold for screening signals to determine which signals' commonality measurement results have reached a sufficient similarity or commonality level. Specifically, the system traverses the commonality measurement results of t first RF signals and compares them with the first reference commonality value one by one. For the commonality measurement result of each signal, if it is not less than the first reference commonality value, then the signal is deemed to have sufficient commonality with the target RF signal and is selected as one of the candidate signals. This process continues until the system finds s signals that meet the conditions or traverses all reference signals. For example, assume that a wireless communication base station is receiving RF signals from different user devices and needs to perform quality assessment on these signals. The base station has constructed a reference RF signal set containing multiple communication protocols and modulation modes through the previous steps, and each signal has a corresponding commonality measurement result, indicating their similarity to the target RF signal.

[0118] The base station sets a first reference commonality value, which is determined based on historical data analysis and expert experience, and is intended to screen out reference signals that are sufficiently similar to the target signal. For example, if the commonality measurement result is expressed in cosine similarity, then the first reference commonality value may be set to 0.8, which means that only those reference signals whose cosine similarity with the target signal is greater than or equal to 0.8 will be considered.

[0119] Now, the base station starts to execute step S430 to select signals from the reference signal set. The system first checks the commonality metric result of the first reference signal. If the result is not less than 0.8, the signal is added to the candidate signal list. Then, the system continues to check the next signal and repeats this process until s signals that meet the conditions are found or all signals are traversed.

[0120] Assume that the base station sets s to 3, that is, it needs to select three reference signals that are most similar to the target RF signal. During the screening process, the system may first find two reference signals with commonality measurement results of 0.85 and 0.9, and then find a signal with a result of 0.82. Since the commonality measurement results of these three signals all meet the condition of not less than 0.8, they are selected as the final candidate signal set.

[0121] The calculation of commonality measurement results often depends on the machine learning model (such as graph neural network) and feature extraction method used in the previous step. For example, in step S420, the system may use a graph neural network (GNN) to extract implicit representations of the signal from the signal relationship graph and calculate the commonality measurement results based on these representations. These implicit representations capture the high-order features of the signal, such as frequency distribution, modulation mode, signal strength changes, etc., providing a rich information basis for commonality measurement.

[0122] As a possible implementation method, the step S440 determines the quality evaluation coefficient of the target RF signal based on the quality evaluation coefficients of the s first RF signals, and specifically may include: if s is not less than 2, then taking the average result of the quality evaluation coefficients of the s first RF signals as the quality evaluation coefficient of the target RF signal.

[0123] In the embodiment of the present application, step S440 infers the quality level of the target signal through a reasonable calculation method based on the quality evaluation coefficients of the s first RF signals most similar to the target RF signal screened in the previous step.

[0124] In step S440, the computer system estimates the quality evaluation coefficient of the target RF signal using the quality evaluation coefficients of the s first RF signals. When s is not less than 2, that is, when at least two reference signals similar to the target signal are screened out, the system usually uses an averaging method to calculate the quality evaluation coefficient of the target signal. Specifically, the system performs arithmetic averaging on the quality evaluation coefficients of the s reference signals, and the result is used as the quality evaluation coefficient of the target signal. This method is based on an assumption that reference signals similar to the target signal should also be similar in quality. Therefore, by averaging the quality evaluation coefficients of these similar signals, a relatively accurate target signal quality assessment can be obtained. For example, assume that a wireless communication system is monitoring the quality of multiple received RF signals. The system has screened out the s first RF signals that are most similar to the target RF signal through previous steps (such as signal decomposition, feature extraction, commonality measurement, etc.). These signals may come from different user devices or communication links, but they have a high similarity with the target signal in terms of spectrum characteristics, modulation mode, etc.

[0125] Each first radio frequency signal has a corresponding quality evaluation coefficient, which may be calculated based on multiple indicators such as the signal-to-noise ratio (SNR), bit error rate (BER), demodulation success rate, etc. For example, the quality evaluation coefficient of a first radio frequency signal may be calculated using the following formula: Quality Score i =α·SNR i +β·(1−BER i)+γ·Demodulation Success Rate i ; Among them, α, β, γ are weight coefficients, which are used to adjust the influence of different indicators on the quality evaluation coefficient; SNR i , BER i 、Demodulation Success Rate i are the signal-to-noise ratio, bit error rate and demodulation success rate of the i-th first RF signal respectively. Now, assuming that the system screens out s=3 first RF signals, and their quality evaluation coefficients are 0.85, 0.88 and 0.82 respectively. According to the implementation method of step S440, the system uses the average of these three coefficients as the quality evaluation coefficient of the target RF signal:

[0126] Target Quality Score=30.85+0.88+0.82=0.85;

[0127] Therefore, the quality evaluation coefficient of the target RF signal is determined to be 0.85, which reflects that the target signal is close to the average level of the screened reference signals in quality.

[0128] In summary, step S440 estimates the quality level of the target RF signal by averaging the quality evaluation coefficients of the screened similar signals. This method is simple, effective and easy to implement, and has important application value in quality monitoring of wireless communication systems. At the same time, the machine learning and feature extraction technologies widely used in the entire RF signal processing process also provide strong support for this process. As a possible implementation method, the step S410, determining the reference RF signal set, may specifically include:

[0129] Step S411: Acquire the t first radio frequency signals;

[0130] Step S412: determining a signal relationship graph and a quality evaluation coefficient of each of the t first radio frequency signals;

[0131] Step S413: Determine the reference RF signal set according to the signal relationship graphs and the quality evaluation coefficients of the t first RF signals.

[0132] In the embodiment of the present application, step S410 is a key step in determining the reference RF signal set, which is crucial for subsequent signal processing, classification, identification and quality assessment tasks. This process not only requires the computer system to effectively collect and organize a large amount of RF signal data, but also requires the ability to accurately construct a signal relationship map for each signal and assign it a reasonable quality evaluation coefficient.

[0133] In step S411, the computer system obtains t first radio frequency signals for constructing a reference radio frequency signal set. These signals may come from a variety of channels, including but not limited to wireless communication base stations, satellite communication systems, radar detection systems, etc. In order to ensure the diversity and representativeness of the reference set, the system needs to collect radio frequency signals covering different frequency bands, modulation modes, channel conditions, and noise levels. Specifically, the computer system collects radio frequency signals in real time or at regular intervals through receivers (such as antennas, sensors, etc.) deployed at specific locations. These receivers have sufficient sensitivity and dynamic range to capture weak radio frequency signals and avoid signal distortion. The collected original radio frequency signals often contain unfavorable factors such as noise and interference, so preprocessing is required to improve signal quality. The preprocessing step may include operations such as filtering, denoising, amplification, and frequency conversion to eliminate or reduce the impact of these unfavorable factors. The preprocessed radio frequency signals are stored in the database of the computer system for subsequent analysis and processing. These signals are usually represented in digital form, including time domain sampling values, frequency domain spectra, etc.

[0134] For example, suppose a wireless communication system needs to evaluate the quality of the received target RF signal. In order to construct a reference RF signal set, the system first retrieves RF signals covering multiple communication protocols (such as GSM, LTE, 5G, etc.), different modulation methods (such as QPSK, QAM, etc.) and different channel conditions (such as multipath fading, shadowing, etc.) from its historical database as candidate signals. These signals come from communication sessions at different times and locations and are widely representative.

[0135] In step S412, the computer system constructs a signal relationship graph for each first radio frequency signal and assigns it a reasonable quality evaluation coefficient. This process involves the application of complex signal processing technology and machine learning algorithms.

[0136] First, the system decomposes each first RF signal and divides it into multiple RF signal segments. This step is similar to spectrum analysis, but it focuses more on the local characteristics of the signal in the time and frequency domains. Decomposition methods may include short-time Fourier transform (STFT), wavelet transform, etc. For each RF signal segment, the system extracts its feature vector through a deep learning model (such as graph convolution network GCN, graph attention network GAT, etc.). These feature vectors capture the characteristics of the signal segment in the time domain, frequency domain, and higher dimensions, providing a basis for subsequent graph construction. Based on the relationship between the extracted feature vectors and signal segments (such as frequency adjacency, time order, etc.), the system constructs a signal relationship graph. This graph is a complex graph structure consisting of multiple nodes and edges. Each node represents an RF signal segment, and the edges reflect the relationship between nodes.

[0137] The system calculates multiple performance indicators of each first RF signal, such as signal-to-noise ratio (SNR), bit error rate (BER), demodulation success rate, etc. These indicators directly reflect the quality level of the signal. Based on the calculation results of multiple performance indicators, the system uses a comprehensive evaluation method to determine the quality evaluation coefficient of each signal. The comprehensive evaluation method may include weighted average method, fuzzy comprehensive evaluation method, prediction method based on machine learning model, etc. For example, the system can use algorithms such as support vector machine (SVM) or random forest to classify or regress signal quality.

[0138] For example, let's continue with the quality assessment of wireless communication systems. When constructing the signal relationship graph, the system first decomposes each candidate RF signal and divides it into multiple RF signal segments. Then, the system uses a graph convolutional network (GCN) to extract features from each signal segment to obtain its implicit representation vector. Next, the system constructs a signal relationship graph based on the frequency adjacency relationship between the signal segments. When determining the quality evaluation coefficient, the system calculates performance indicators such as the signal-to-noise ratio (SNR) and bit error rate (BER) of each candidate RF signal. Then, the system uses a weighted average method to combine these performance indicators into a single quality evaluation coefficient. Specifically, the system may give a higher weight to the signal-to-noise ratio (such as 0.7) because the signal-to-noise ratio is an important indicator for measuring signal purity; and a lower weight to the bit error rate (such as 0.3) because the bit error rate directly reflects the error in signal transmission, but is affected by many factors (such as channel conditions, modulation methods, etc.).

[0139] In step S413, the computer system determines the final reference RF signal set according to the signal relationship graph and the quality evaluation coefficient of each first RF signal. This process may involve operations such as signal screening, sorting and combination.

[0140] Specifically, the system screens the t first radio frequency signals according to the quality evaluation coefficients, and excludes those signals with poor quality or unclear features. The screening criteria may be based on the threshold setting of the quality evaluation coefficient, such as only retaining signals with quality evaluation coefficients higher than a certain value.

[0141] For the signals retained after screening, the system can sort them according to the quality evaluation coefficient so that the signals with higher quality can be given priority in subsequent processing. The sorting operation helps to improve the efficiency and accuracy of signal processing. Finally, the system combines the sorted signals into a reference RF signal set. This set contains multiple representative and high-quality RF signals, providing rich data support for subsequent tasks such as signal processing, classification, identification and quality assessment.

[0142] For example, let's go back to the quality assessment scenario of wireless communication systems. After determining the signal relationship graph and quality evaluation coefficient of each candidate RF signal, the system screens and sorts the signals according to the quality evaluation coefficient. Specifically, the system sets a threshold for the quality evaluation coefficient (such as 0.6) and excludes all signals below the threshold. Then, the system sorts the remaining signals in order of quality evaluation coefficient from high to low.

[0143] Finally, the system selects the first N signals after sorting (N is the preset reference set size) as the reference RF signal set. This set contains multiple RF signals from different communication protocols, modulation modes and channel conditions, and each signal has high quality and representativeness. By using this reference set, the system can more accurately evaluate the quality level of the target RF signal and provide strong support for its subsequent processing. As a possible implementation method, the step S411, obtaining the t first RF signals, can specifically include:

[0144] Step S4111: obtaining u control radio frequency signals and v candidate radio frequency signals, where u and v are both natural numbers greater than 0;

[0145] Step S4112: according to the u control RF signals, determine one or more candidate RF signals from the v candidate RF signals as the first RF signal.

[0146] In an embodiment of the present application, step S411 is a key link in obtaining the basic data of the reference RF signal set, and its goal is to screen out a first RF signal that is representative and diverse from a large number of signal samples. In step S4111, the primary task of the computer system is to collect a sufficient number of RF signal samples, which are divided into two categories: control RF signals and alternative RF signals. The control RF signals usually have known and stable characteristics, such as known signal-to-noise ratios, bit error rates, etc., which are used as benchmarks for subsequent screening and evaluation. The alternative RF signals contain a series of unknown or to-be-evaluated signals, which may come from different communication protocols, modulation methods, channel conditions, etc., and have greater diversity.

[0147] Specifically, the computer system can obtain RF signal samples through various channels, including wireless communication base stations, satellite communication systems, radar detection systems, etc. These systems will generate a large amount of RF signal data in daily operation, which can be transmitted to the computer system in real time or regularly for processing. The acquired RF signal samples are stored in the database of the computer system for subsequent analysis and processing. The database needs to support efficient query and retrieval functions to quickly locate specific signal samples. Before storage, the RF signal samples may need to undergo a series of preprocessing operations, such as denoising, filtering, synchronization, etc., to improve the quality and stability of the signal. For example, suppose a wireless communication system is preparing a signal quality assessment plan for an upcoming large-scale event. In order to ensure the accuracy and comprehensiveness of the assessment, the system needs to build a reference set containing multiple types of RF signals. To this end, the system first retrieves u control RF signals (for example, u=100) from the historical database. These signals cover common communication protocols and modulation methods, and the performance indicators of each signal, such as signal-to-noise ratio and bit error rate, are strictly calibrated and verified. Then, the system selects v candidate RF signals (for example, v=500) from the RF signals collected in real time. These signals may come from different user devices, different geographical locations, and different communication time periods.

[0148] In step S4112, the computer system selects high-quality samples from the candidate RF signals as the first RF signal according to the characteristics of the reference RF signal. The screening process may involve a variety of methods such as signal similarity analysis, performance index comparison, and machine learning model prediction.

[0149] Specifically, the computer system first calculates the similarity metric values ​​(such as cosine similarity, Euclidean distance, etc.) between each candidate RF signal and the reference RF signal. These metric values ​​reflect the similarity between the candidate signal and the reference signal in terms of spectral characteristics, modulation mode, etc.

[0150] In addition to similarity metrics, the system can also directly compare the performance indicators of the candidate signal and the control signal (such as signal-to-noise ratio, bit error rate, etc.). These indicators provide an intuitive basis for evaluating signal quality.

[0151] In order to further improve the accuracy and efficiency of screening, the computer system can also use machine learning models (such as classifiers, regressors, etc.) to predict and evaluate candidate signals. These models have learned the mapping relationship between the control signal and the high-quality signal through a large amount of training data, and can automatically identify candidate signals with similar characteristics. Based on the above analysis results, the system can screen and sort the candidate RF signals. The screening process may involve operations such as setting thresholds and selecting top-ranked signals. The sorted signals are arranged from high to low in terms of quality, which facilitates the subsequent construction of reference sets. For example, let's continue to take the signal quality evaluation of the wireless communication system as an example. In step S4112, the system first calculates the cosine similarity between each candidate RF signal and the control RF signal. Then, the system combines the signal-to-noise ratio and bit error rate and other performance indicators for comprehensive evaluation. In order to improve the accuracy of screening, the system also introduces a classifier based on support vector machine (SVM) to classify and predict the candidate signals. The input features of the classifier include the spectral characteristics, modulation mode, performance indicators, etc. of the signal; the output is the prediction result of whether the signal belongs to the high-quality category. Finally, the system screens and sorts the candidate signals according to information in multiple dimensions such as similarity metrics, performance index prediction results, and classifier outputs, and selects t optimal candidate RF signals as the first RF signal (for example, t=100). These signals will be used in subsequent reference set construction and quality assessment tasks. As a possible implementation method, step S4112, based on the u control RF signals, determines one or more candidate RF signals from the v candidate RF signals as the first RF signal, which may specifically include:

[0152] Step S41121: for the ath candidate RF signal among the v candidate RF signals, determine a commonality measurement result between the ath candidate RF signal and each of the u reference RF signals, 1≤a≤v;

[0153] Step S41122: If a commonality measurement result between the ath candidate RF signal and one or more of the u reference RF signals is greater than a second reference commonality value, the ath candidate RF signal is determined as the first RF signal.

[0154] In the embodiment of the present application, step S4112 selects a signal having sufficient commonality with the target or reference RF signal from a large number of candidate RF signals as the first RF signal. Specifically, in step S41121, the task of the computer system is to evaluate the degree of commonality between each candidate RF signal and all reference RF signals one by one. The commonality measurement result reflects the similarity or consistency of the two signals in terms of spectral characteristics, modulation mode, signal strength, etc.

[0155] Specifically, for each candidate RF signal and control RF signal, the computer system first performs feature extraction, which usually involves converting the signal from the time domain or frequency domain to a feature space that is easier to analyze, such as using short-time Fourier transform (STFT) to extract time-frequency features, or using deep learning models (such as convolutional neural networks (CNNs) and graph neural networks (GNNs)) to extract high-order features.

[0156] For the ath candidate RF signal and each control RF signal, the system calculates the commonality measurement result between them. The commonality measurement method may include calculating the cosine similarity, Euclidean distance, Manhattan distance, etc. between the feature vectors, or using a more complex machine learning model to predict the degree of commonality between the two signals. The calculated commonality measurement results are stored in the memory of the computer system for subsequent comparison and screening operations. For example, suppose a wireless communication system is preparing a signal quality monitoring plan for an upcoming large-scale event. The system has collected u control RF signals (such as high-quality signals in historical records) and v candidate RF signals (such as live signals collected in real time). In order to screen out the candidate signal that has sufficient commonality with the control signal as the first RF signal, the system performs a commonality measurement calculation on each candidate signal and all control signals. For example, the system may use cosine similarity as a commonality measurement indicator to calculate the similarity between each candidate signal and the control signal feature vector.

[0157] In step S41122, the computer system determines whether the candidate RF signal meets the screening condition according to the commonality measurement result and a preset second reference commonality value.

[0158] Specifically, for the ath candidate RF signal, the system compares the commonality measurement results between it and each reference RF signal with a preset second reference commonality value. If at least one commonality measurement result is greater than the threshold, it is considered that the candidate signal has sufficient commonality with the reference signal.

[0159] According to the comparison results, the system selects all candidate RF signals that meet the conditions as the first RF signal. These signals will be used for subsequent reference set construction and quality assessment tasks.

[0160] The screened first RF signal is output to the reference set and may be stored together with other relevant information (such as commonality measurement results, signal source, etc.). For example, let's continue to take the signal quality monitoring of the wireless communication system as an example. In step S41122, the system sets a second reference commonality value (such as 0.8) to determine whether the degree of commonality between the candidate RF signal and the control RF signal is high enough. For each candidate signal, if its commonality measurement result with at least one control signal is greater than 0.8, the system determines it as the first RF signal and adds it to the reference set. This process ensures that the signals in the reference set are not only sufficient in number, but also reliable in quality, providing strong support for subsequent signal quality evaluation.

[0161] As a possible implementation manner, the step S41121, determining the commonality measurement result between the ath candidate RF signal and each of the u reference RF signals, may specifically include:

[0162] Step S411211: determining the implicit signal representation of the ath candidate RF signal and the implicit signal representation of each of the u reference RF signals;

[0163] Step S411212: Determine the commonality measurement result between the ath candidate RF signal and each of the u control RF signals based on the implicit signal representation of the ath candidate RF signal and the implicit signal representation of each of the u control RF signals. In an embodiment of the present application, step S41121 is a key step in evaluating the commonality between the candidate RF signal and the control RF signal. This process is achieved by comparing the implicit representation of the signal, which can capture the high-order features of the signal, thereby more accurately reflecting the similarity between the signals.

[0164] In step S411211, the first task of the computer system is to extract the implicit representation of each RF signal. The implicit representation is a low-dimensional vector that captures information in the signal that is difficult to observe intuitively but is critical for subsequent processing. For the candidate RF signal and the control RF signal, the system needs to extract their implicit representations respectively.

[0165] Specifically, in order to extract the implicit representation of the signal, the computer system first needs to select a suitable feature extraction model. This model can be based on deep learning, such as an autoencoder, a convolutional neural network (CNN), or a graph neural network (GNN). The choice of model depends on the characteristics of the signal and the processing requirements. For example, if the signal has obvious time-frequency characteristics, then using CNN may be a good choice; if the signal can be represented as a graph structure (such as a signal relationship graph), GNN may be more suitable. Model training usually requires the use of a large amount of labeled data. In an embodiment of the present application, these data may include radio frequency signals of known characteristics and their corresponding implicit representations. Through supervised learning, the system can train a model that can accurately extract the implicit representation of the signal.

[0166] For each of the a-th candidate RF signal and the u-th control RF signal, the system uses the trained feature extraction model to extract their implicit representations. This process usually involves inputting the signal into the model and obtaining the low-dimensional vector output by the model as the implicit representation.

[0167] For example, suppose a wireless communication system is preparing a signal quality monitoring solution for an upcoming large-scale event. The system has collected u control RF signals (such as high-quality signals in historical records) and v candidate RF signals (such as live signals collected in real time). In order to evaluate the commonalities between the candidate signals and the control signals, the system needs to extract their implicit representations.

[0168] The system selects a feature extraction model based on graph convolutional networks (GCNs) because RF signals can be represented as signal relationship graphs. Through training with a large amount of labeled data, the system obtains a GCN model that can accurately extract the implicit representation of the signal. Then, the system uses this model to extract the implicit representation of the a-th candidate RF signal and u-th control RF signals. The implicit representation of each signal is a low-dimensional vector, such as z 备选 =[z1,z2,…,zd] and z 对照i =[z1′,z2′,…,zd′] (where i represents the index of the control signal and d is the dimension of the implicit representation).

[0169] After extracting the implicit representation of the signal, the computer system next needs to calculate the commonality measure between the candidate RF signal and each reference RF signal. This process usually involves calculating the distance or similarity between the implicit representation vectors.

[0170] Specifically, the system can select a variety of methods to calculate the commonality measurement results between implicit representation vectors, such as cosine similarity, Euclidean distance, Manhattan distance, etc. Which method to choose depends on the specific application scenario and requirements. For example, if the system is more concerned with the similarity of the signal in direction, then cosine similarity may be a better choice; if the system is more concerned with the difference in amplitude of the signal, then Euclidean distance may be more appropriate. For the ath candidate RF signal and each control RF signal, the system uses the selected commonality measurement method to calculate the distance or similarity between their implicit representation vectors. This process usually involves vector operations and the application of mathematical formulas. For example, let's continue to take the signal quality monitoring of the wireless communication system as an example. In step S411212, the system selects cosine similarity as the commonality measurement indicator. For the ath candidate RF signal and each control RF signal, the system calculates the cosine similarity between their implicit representation vectors.

[0171] Through calculation, the system obtains the cosine similarity value between the ath candidate RF signal and each control RF signal. These values ​​reflect the degree of commonality between the candidate signal and the control signal, providing a basis for subsequent signal screening.

[0172] As a possible implementation manner, the step S411211, determining the implicit signal representation of the ath candidate RF signal and the implicit signal representation of each of the u reference RF signals, may specifically include:

[0173] Step S4112111: determining a signal relationship graph of the ath candidate radio frequency signal and a signal relationship graph of each of the u reference radio frequency signals;

[0174] Step S4112112: determining a signal implicit representation of the ath candidate RF signal according to the signal relationship graph of the ath candidate RF signal;

[0175] Step S4112113: for each of the u control RF signals, determine the signal implicit representation of the control RF signal according to the signal relationship graph of the control RF signal. In the embodiment of the present application, step S411211 is one of the key steps to extract the implicit representation of the RF signal, which relies on the construction and analysis of the signal relationship graph.

[0176] In step S4112111, the computer system constructs a signal relationship graph for each radio frequency signal. A signal relationship graph is a graphical representation method for describing the relationship between components within a signal. In the embodiment of the present application, these components may include different frequency components, modulation methods, channel characteristics, etc.

[0177] Specifically, first, the system decomposes each RF signal into multiple sub-signals or signal segments. This step is similar to spectrum analysis, but focuses more on the local characteristics of the signal in the time and frequency domain. Decomposition methods may include short-time Fourier transform (STFT), wavelet transform, or deep learning-based signal segmentation algorithms. For each sub-signal or signal segment obtained by decomposition, the system extracts its feature vector. These feature vectors may include spectral characteristics, time domain waveforms, modulation parameters, etc., which are used for subsequent spectrum construction.

[0178] Based on the relationship between the extracted feature vectors and sub-signals (such as frequency adjacency, time order, modulation similarity, etc.), the system constructs a signal relationship graph. The nodes in the graph represent sub-signals or signal segments, and the edges represent the relationship between them.

[0179] For example, suppose a wireless communication system is analyzing received RF signals. The system first decomposes each RF signal (including the ath candidate RF signal and uth control RF signals) to obtain multiple sub-signal segments. Then, the system extracts the feature vectors of each sub-signal segment, such as center frequency, bandwidth, modulation mode, etc. Finally, the system constructs a signal relationship graph based on the relationship between these feature vectors and sub-signal segments. For example, if two sub-signal segments are adjacent in frequency and have the same modulation mode, they may be connected by an edge in the graph.

[0180] In step S4112112, the computer system uses the signal relationship graph of the ath candidate radio frequency signal to extract its implicit signal representation. The implicit signal representation is a low-dimensional vector that captures the high-order features of the signal and is helpful for subsequent signal processing and analysis.

[0181] Specifically, the system uses graph neural networks (GNNs) to process signal relationship graphs. GNNs can process graph-structured data and update node representations through information transfer and aggregation between nodes. In this process, the representation of each node will gradually integrate the information of its neighborhood to form a richer feature vector.

[0182] After multiple rounds of GNN iterations, the system selects the representation of a specific node in the graph (such as the root node representing the entire signal) as the implicit representation of the signal. Alternatively, the system can aggregate the representations of all nodes in the graph (such as summing, averaging, etc.) to obtain the implicit representation of the entire signal.

[0183] For example, let's continue with the wireless communication system. When extracting the implicit representation of the ath candidate RF signal, the system uses a graph convolutional network (GCN) to process the signal relationship graph. GCN integrates the information in the graph by iteratively updating the node representation. After multiple rounds of iterations, the system selects the root node representation of the graph as the implicit representation vector of the candidate signal. This vector contains high-order feature information of the signal, such as frequency distribution pattern, modulation characteristics, etc.

[0184] Step S4112113 is similar to step S4112112, but is for each of the u reference RF signals. The system repeats the operations in steps S4112111 and S4112112 for each reference RF signal to extract their signal implicit representations.

[0185] Specifically, for each control RF signal, the system first builds its signal relationship graph (if it has not been built yet). Then, the system processes the graph using the same graph neural network model to extract the implicit representation of the signal. This process is parallel, that is, the system can process the graphs of multiple control RF signals at the same time. For example, in a wireless communication system, the system repeats the above process for each control RF signal. For each control signal, the system builds its signal relationship graph and uses GCN to extract implicit representations. These implicit representation vectors will be used in the subsequent calculation of commonality metrics with alternative signals.

[0186] In summary, step S411211 provides an effective feature extraction method for RF signal processing by constructing a signal relationship graph and using a graph neural network to extract implicit representations. This method can not only capture the high-order feature information of the signal, but also provide strong support for subsequent signal analysis, classification, recognition and other tasks.

[0187] As a possible implementation manner, determining the quality evaluation coefficient of each of the t first RF signals includes: step S4121: if the first RF signal is the a-th candidate RF signal, then determining, based on the commonality measurement result between the a-th candidate RF signal and each of the u-th reference RF signals, w reference RF signals whose commonality measurement results with the a-th candidate RF signal are not less than a third reference commonality value among the u-th reference RF signals;

[0188] Step S4122: Determine the quality evaluation coefficient of the ath candidate RF signal according to the quality evaluation coefficients of the w reference RF signals.

[0189] In the embodiment of the present application, determining the quality evaluation coefficient of each first RF signal (in this scenario, specifically the signal selected from the candidate RF signals) is a key step, which helps with subsequent tasks such as signal processing, classification or quality monitoring. In step S4121, the goal of the computer system is to find a control RF signal with a higher commonality measurement result with the ath candidate RF signal. The commonality measurement result reflects the similarity or consistency of the two signals in terms of spectral characteristics, modulation mode, signal strength, etc.

[0190] Specifically, for the ath candidate RF signal, the computer system has calculated the commonality measurement results between it and each of the u control RF signals according to the previous steps. Now, the system needs to compare these results with a preset third reference commonality value. The third reference commonality value is a threshold value used to determine whether the commonality between two signals is significant enough.

[0191] The system traverses all the control RF signals and finds those signals whose commonality measurement results with the ath candidate RF signal are not less than the third reference commonality value. Assuming there are w such signals, they are considered to be control signals that are closer in characteristics to the candidate signals.

[0192] For example, suppose that a wireless communication system is performing a quality assessment on RF signals collected in real time. The system has screened out t first RF signals (including the ath candidate RF signal) from v candidate RF signals through the previous steps, and calculated the commonality measurement results between them and u historical control RF signals. Now, the system sets the third reference commonality value to 0.8, which means that only when the commonality measurement results of two signals are greater than or equal to 0.8, they are considered to have significant commonality. By comparison, the system finds w control RF signals whose commonality measurement results with the ath candidate RF signal are not less than 0.8.

[0193] In step S4122, the computer system uses the quality evaluation coefficients of the w reference radio frequency signals screened in the previous step to determine the quality evaluation coefficient of the ath candidate radio frequency signal.

[0194] Specifically, for the w selected control RF signals, the system already knows their respective quality evaluation coefficients. Now, the system can use a variety of methods to aggregate these coefficients to obtain the quality evaluation coefficient of the ath candidate RF signal. Common aggregation methods include weighted average, median, mode, etc. Assuming that the system selects the weighted average method as the aggregation method, the quality evaluation coefficient of the ath candidate RF signal can be calculated by the following formula: QS 备选a = Among them, QS 备选a is the quality evaluation coefficient of the ath candidate RF signal, QS 对照iis the quality evaluation coefficient of the ith reference RF signal, w i is the weight of the ith control signal (which can be dynamically adjusted based on the commonality measurement results or other factors).

[0195] For example, let's continue with the quality evaluation of the wireless communication system. After determining w control RF signals with higher commonality measurement results with the ath candidate RF signal, the system uses the weighted average method to calculate the quality evaluation coefficient of the candidate signal. The system assigns the same weight (i.e., equal weight) to each control signal, and then sums and averages the quality evaluation coefficients of these signals. Assuming w=3, and the quality evaluation coefficients of the three control signals are 0.9, 0.85, and 0.92, respectively, the quality evaluation coefficient of the ath candidate RF signal is: QS 备选a =(0.9+0.85+0.92) / 3=0.89.

[0196] This quality evaluation coefficient reflects that the quality of the ath candidate RF signal is similar to the average level of the screened control signal. In this way, the system can assign a reasonable quality evaluation coefficient to each first RF signal (including the signal screened from the candidate signals), providing strong support for subsequent signal processing and quality monitoring.

[0197] As a possible implementation method, when determining the implicit signal representation of the RF signal based on the signal relationship graph of the RF signal, it can specifically include: based on a graph convolutional network, feature extraction of the graph element information of each graph element in the signal relationship graph of the second RF signal, and obtaining the implicit signal representation of the second RF signal, wherein the second RF signal is the above-mentioned target RF signal, the ath alternative RF signal or the control RF signal.

[0198] In the above embodiments, when processing RF signals, a signal relationship graph is used to extract deep features of the signal, thereby obtaining an implicit representation of the signal. This process typically involves using a graph convolutional network (GCN) to extract features from each graph element in the signal relationship graph. Specifically, when a computer system processes an RF signal, it first constructs a signal relationship graph. In this graph, each RF signal segment is regarded as a graph element, and each graph element carries graph element information about the RF signal segment, including but not limited to the spectrum distribution of the signal segment, signal strength distribution, and other related attributes.

[0199] In order to extract meaningful features from the signal relationship graph, the system uses a graph convolutional network (GCN) to process the graph element information. GCN is a special deep learning model that can operate on graph data, capture the relationship between nodes, and extract features from them. In this case, GCN will process each graph element in the signal relationship graph and update the representation of the graph element based on the graph element information.

[0200] For example, suppose that a computer system is processing a target RF signal, which is decomposed into three RF signal segments. The primitive information of each RF signal segment includes the time-frequency signal strength distribution, which can be represented by a two-dimensional matrix, where each element in the matrix represents the signal strength at a certain time point and frequency. For example, for the first RF signal segment, the time-frequency signal strength distribution can be represented as: ;here, represents the signal strength at the i-th time point and the j-th frequency.

[0201] Next, GCN will process this signal strength matrix and update the representation of each primitive. GCN works by aggregating the information of neighboring nodes to update the representation of each node. Specifically, for each RF signal segment (i.e., primitive), GCN will consider its directly adjacent signal segments (i.e., neighboring primitives) and update the representation of the current signal segment based on their primitive information. This process can be expressed as the following formula: ;here, is the representation of node i at the l+1th layer, is the representation of node j at layer l, is the set of neighbors of node i, is the normalization constant, is the weight matrix, and is the activation function. In this example, the computer system uses GCN to process the signal relationship graph of the target RF signal, as well as the signal relationship graph of other RF signals (such as candidate RF signals or control RF signals). Through this process, GCN will generate a new representation for each RF signal segment, and this new representation contains the implicit representation of the signal segment. Ultimately, the computer system will obtain the implicit signal representation of the entire RF signal, which will be an important basis for evaluating the quality of the RF signal in subsequent steps.

[0202] Through such a processing flow, the computer system can effectively extract key features from complex RF signals, laying the foundation for further signal processing and quality assessment.

[0203] As a possible implementation method, when determining the signal relationship graph of the radio frequency signal, it includes:

[0204] Step S1: Decomposing the third RF signal into RF signal segments according to the frequency band included in the third RF signal to obtain b RF signal segments of the third RF signal, wherein the third RF signal is the ath candidate RF signal or the reference RF signal or the first RF signal, and b≥2;

[0205] Step S2: obtaining the spectrum distribution of each of the b RF signal segments in the third RF signal, and performing implicit representation mining on the b RF signal segments to obtain an implicit signal representation of each of the b RF signal segments;

[0206] Step S3: For each of the b RF signal segments, use the RF signal segment as a graph element of the signal relationship graph, and determine the graph element information of the graph element based on the implicit signal representation and spectrum distribution of the RF signal segment to obtain the signal relationship graph of the third RF signal.

[0207] When processing a radio frequency signal, the computer system will construct a signal relationship map according to a series of steps. Specifically, the computer system first obtains a radio frequency signal (referred to as the third radio frequency signal herein), which may be one of the ath candidate radio frequency signal, the reference radio frequency signal or the first radio frequency signal. Next, the system will decompose the signal into multiple radio frequency signal segments according to the different frequency bands contained in the signal. For example, if the third radio frequency signal covers a frequency band range from 1 GHz to 4 GHz, the system may decompose it into three radio frequency signal segments: the first radio frequency signal segment covers 1 GHz to 2 GHz, the second radio frequency signal segment covers 2 GHz to 3 GHz, and the third radio frequency signal segment covers 3 GHz to 4 GHz. In this example, b is equal to 3, that is, the third radio frequency signal is decomposed into 3 radio frequency signal segments. For each radio frequency signal segment, the system needs to obtain its spectrum distribution, which includes information such as the center frequency of the frequency band, the bandwidth, and the relative position of the frequency band in the entire signal spectrum. For example, for the first radio frequency signal segment (1 GHz to 2 GHz), its center frequency may be 1.5 GHz, the bandwidth is 1 GHz, and it is the lowest frequency part in the entire signal spectrum.

[0208] Next, the system performs implicit representation mining on these RF signal segments to obtain the implicit signal representation of each RF signal segment. This usually involves using a deep learning model, such as a graph convolutional network (GCN), to extract the key features of the RF signal segment. For example, assume that the time-frequency signal strength distribution of the first RF signal segment can be represented as a two-dimensional matrix, each row of the matrix represents a time point, each column represents a frequency, and each element in the matrix represents the signal strength at that time point and frequency. Once the implicit signal representation and spectrum distribution of each RF signal segment are obtained, the system can build a signal relationship graph. In this graph, each RF signal segment is regarded as a graph element (node), and each graph element carries graph element information about the RF signal segment. This information includes the implicit signal representation and the spectrum distribution. For example, for the first RF signal segment, the graph element information may include the implicit signal representation (feature vector extracted by the graph convolutional network) and the spectrum distribution information (center frequency, bandwidth, etc.). In this way, the computer system is able to build a signal relationship graph that not only contains the relationship between the RF signal segments, but also retains the key characteristics of each RF signal segment. This helps the system gain a deeper understanding of RF signals and lays a solid foundation for further processing and analysis.

[0209] According to the embodiments of this application, please refer to Figure 3 , is a block diagram of a computer system 1000 provided in the present application, and the computer system 1000 includes a computing unit 1001, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 1002 or a computer program loaded from a storage unit 1008 into a random access memory (RAM) 1003. In the RAM 1003, various programs and data required for the operation of the computer system 1000 can also be stored. The computing unit 1001, the ROM 1002, and the RAM 1003 are connected to each other via a bus 1004. An input / output (I / O) interface 1005 is also connected to the bus 1004.

[0210] Multiple components in the computer system 1000 are connected to the I / O interface 1005, including: an input unit 1006, an output unit 1007, a storage unit 1008, and a communication unit 1009. The input unit 1006 can be any type of device that can input information to the computer system 1000. The input unit 1006 can receive input digital or character information, and generate key signal input related to user settings and / or function control of the electronic device, and can include but is not limited to a mouse, a keyboard, a touch screen, a track pad, a track ball, a joystick, a microphone, and / or a remote control. The output unit 1007 can be any type of device that can present information, and can include but is not limited to a display, a speaker, a video / audio output terminal, a vibrator, and / or a printer. The storage unit 1008 can include but is not limited to a disk and an optical disk. The communication unit 1009 allows the computer system 1000 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks, and may include but is not limited to a modem, a network card, an infrared communication device, a wireless communication transceiver and / or a chipset, such as a Bluetooth™ device, an 802.11 device, a WiFi device, a WiMax device, a cellular communication device and / or the like.

[0211] The computing unit 1001 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 1001 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 1001 performs the various methods and processes described above. For example, in some embodiments, the method may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit 1008. In some embodiments, part or all of the computer program may be loaded and / or installed on the computer system 1000 via ROM 1002 and / or communication unit 1009. When the computer program is loaded into RAM 1003 and executed by the computing unit 1001, one or more steps of the method described above may be performed. Alternatively, in other embodiments, the computing unit 1001 may be configured to execute the method in any other appropriate manner (e.g., by means of firmware).

Claims

1. A radio frequency signal processing method based on deep learning, characterized in that: include: Acquire a target radio frequency signal to be transmitted, and decompose the target radio frequency signal into radio frequency signal segments according to a frequency band included in the target radio frequency signal to obtain x radio frequency signal segments of the target radio frequency signal, where x≥2; Obtaining the spectrum distribution of each of the x radio frequency signal segments in the target radio frequency signal, and performing implicit representation mining on the x radio frequency signal segments to obtain an implicit signal representation of each of the x radio frequency signal segments; For each of the x radio frequency signal segments, use the radio frequency signal segment as a graph element of the signal relationship graph, obtain the time-frequency signal strength distribution of the radio frequency signal segment, determine the graph element information of the graph element according to the signal implicit representation, spectrum distribution and time-frequency signal strength distribution of the radio frequency signal segment, and obtain the signal relationship graph of the target radio frequency signal; Determine a reference RF signal set, the reference RF signal set including a quality evaluation coefficient of each of t first RF signals, and a signal implicit representation or a signal relationship graph of each first RF signal, the signal implicit representation of the first RF signal is determined based on the signal relationship graph of the first RF signal, and t≥2; Determine a commonality measurement result between the target RF signal and each of the t first RF signals according to the signal relationship graph of the target RF signal and the signal implicit representation or signal relationship graph of each first RF signal; According to the commonality measurement result, s first radio frequency signals are selected from the t first radio frequency signals, wherein 1<s≤t; The quality evaluation coefficient of the target radio frequency signal is determined according to the quality evaluation coefficients of the s first radio frequency signals.

2. The method according to claim 1, characterized in that The determining, based on the signal relationship graph of the target radio frequency signal and the signal implicit representation or signal relationship graph of each first radio frequency signal, a commonality measurement result between the target radio frequency signal and each of the t first radio frequency signals includes: If the reference RF signal set includes the signal relationship graphs of the t first RF signals, determining a commonality measurement result between the target RF signal and each of the t first RF signals according to the signal relationship graph of the target RF signal and the signal relationship graphs of the t first RF signals; If the reference RF signal set includes implicit signal representations of the t first RF signals, then based on the implicit signal representation of the target RF signal and the implicit signal representations of the t first RF signals, determine the commonality measurement result between the target RF signal and each of the t first RF signals.

3. The method according to claim 2, characterized in that The determining, based on the signal relationship graph of the target radio frequency signal and the signal relationship graph of the t first radio frequency signals, a commonality measurement result between the target radio frequency signal and each of the t first radio frequency signals includes: Determining a signal implicit representation of the target radio frequency signal according to a signal relationship graph of the target radio frequency signal; Determining implicit signal representations of the t first radio frequency signals according to the signal relationship graphs of the t first radio frequency signals; A commonality measurement result between the target RF signal and each of the t first RF signals is determined according to the implicit signal representation of the target RF signal and the implicit signal representation of the t first RF signals.

4. The method according to claim 1, characterized in that The selecting s first radio frequency signals from the t first radio frequency signals according to the commonality measurement result includes: Determine, among the t first radio frequency signals, s first radio frequency signals whose commonality measurement results with the target radio frequency signal are not less than a first reference commonality value; The determining the quality evaluation coefficient of the target radio frequency signal according to the quality evaluation coefficients of the s first radio frequency signals includes: If s is not less than 2, taking the average result of the quality evaluation coefficients of the s first radio frequency signals as the quality evaluation coefficient of the target radio frequency signal; The determining of the reference radio frequency signal set includes: Acquire the t first radio frequency signals; Determine a signal relationship graph and a quality evaluation coefficient of each of the t first radio frequency signals; Determining the reference radio frequency signal set according to the signal relationship graphs and the quality evaluation coefficients of the t first radio frequency signals; The step of obtaining the t first radio frequency signals includes: Obtain u control radio frequency signals and v candidate radio frequency signals, where u and v are both natural numbers greater than 0; According to the u control radio frequency signals, one or more candidate radio frequency signals are determined from the v candidate radio frequency signals as the first radio frequency signal.

5. The method according to claim 4, characterized in that The step of determining, based on the u control radio frequency signals, one or more candidate radio frequency signals from the v candidate radio frequency signals as the first radio frequency signal comprises: For an ath candidate RF signal among the v candidate RF signals, determine a commonality measurement result between the ath candidate RF signal and each of the u reference RF signals, 1≤a≤v; If a commonality measurement result between the ath candidate RF signal and one or more of the u reference RF signals is greater than a second reference commonality value, determining the ath candidate RF signal as the first RF signal; The determining of the commonality measurement result between the ath candidate radio frequency signal and each of the u reference radio frequency signals includes: Determining an implicit signal representation of the ath candidate radio frequency signal and an implicit signal representation of each of the u reference radio frequency signals; Determining a commonality measurement result between the ath candidate RF signal and each of the u reference RF signals according to the implicit signal representation of the ath candidate RF signal and the implicit signal representation of each of the u reference RF signals; The step of determining the implicit signal representation of the ath candidate radio frequency signal and the implicit signal representation of each of the u reference radio frequency signals comprises: Determine a signal relationship graph of the ath candidate radio frequency signal and a signal relationship graph of each of the u reference radio frequency signals; Determining a signal implicit representation of the ath candidate radio frequency signal according to a signal relationship graph of the ath candidate radio frequency signal; For each of the u reference radio frequency signals, determining a signal implicit representation of the reference radio frequency signal according to a signal relationship graph of the reference radio frequency signal; Determining a quality evaluation coefficient of each of the t first radio frequency signals includes: If the first RF signal is the a-th candidate RF signal, then according to the commonality measurement result between the a-th candidate RF signal and each of the u reference RF signals, w reference RF signals whose commonality measurement results with the a-th candidate RF signal are not less than a third reference commonality value are determined among the u reference RF signals; The quality evaluation coefficient of the ath candidate radio frequency signal is determined according to the quality evaluation coefficients of the w reference radio frequency signals.

6. The method according to claim 3 or 5, characterized in that: Determining a signal implicit representation of the second radio frequency signal according to a signal relationship graph of the second radio frequency signal includes: Based on a graph convolutional network, feature extraction is performed on the graph element information of each graph element in the signal relationship graph of the second RF signal to obtain an implicit signal representation of the second RF signal, where the second RF signal is the target RF signal, the ath candidate RF signal or the control RF signal.

7. The method according to claim 4 or 5, characterized in that: Determining a signal relationship graph of a third radio frequency signal includes: Decomposing the third radio frequency signal into radio frequency signal segments according to the frequency band included in the third radio frequency signal to obtain b radio frequency signal segments of the third radio frequency signal, where the third radio frequency signal is the ath candidate radio frequency signal or the reference radio frequency signal or the first radio frequency signal, and b≥2; Obtaining the spectrum distribution of each of the b radio frequency signal segments in the third radio frequency signal, and performing implicit representation mining on the b radio frequency signal segments to obtain an implicit signal representation of each of the b radio frequency signal segments; For each of the b RF signal segments, the RF signal segment is used as a graph element of the signal relationship graph, and based on the implicit signal representation and spectrum distribution of the RF signal segment, the graph element information of the graph element is determined to obtain the signal relationship graph of the third RF signal.

8. A computer system, characterized in that: include: at least one processor; and a memory communicatively coupled to the at least one processor; The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method described in any one of claims 1 to 7.

Images