Dynamic spectrum signal identification method and device, electronic equipment and storage medium

Abstract

Embodiments of the application disclose a dynamic spectrum signal identification method and device, electronic equipment and storage medium, and the application relates to the fields of maps, intelligent transportation, artificial intelligence and the like. The method comprises: acquiring a dynamic spectrum signal to be identified, the dynamic spectrum signal having a time dimension and a frequency dimension; dividing the dynamic spectrum signal along the time dimension by a preset size threshold to obtain a plurality of signal blocks of the same frequency dimension; inputting the signal blocks into a signal classification model respectively to obtain a confidence level of the dynamic spectrum signal for a single pulse signal type; and when the confidence level reaches a preset threshold, taking the single pulse signal type as the signal type of the dynamic spectrum signal. Through the division of the dynamic spectrum signal, the model can process signal blocks of the same resolution, thereby improving the accuracy of model classification.

Description

Technical Field

[0001] This application relates to the field of computer technology, specifically to a dynamic spectrum signal recognition method, apparatus, electronic device, and storage medium. Background Technology

[0002] With the development of astrophysics and radio signal technology, more and more cosmic signals can be observed and recorded by radio telescopes deployed in various regions of the world. These signals can help astronomers study the universe and have significant scientific research value.

[0003] As sensitivity and observation range increase, the amount of data that astronomical telescopes can receive also increases explosively. Processing this data often means that the desired signal needs to be sifted out from a large amount of data. Due to the large amount of data, the efficiency and difficulty of screening and judgment are increased. Summary of the Invention

[0004] This application provides a method, apparatus, electronic device, and storage medium for dynamic spectrum signal recognition, which can improve the efficiency and accuracy of dynamic spectrum signal recognition.

[0005] In a first aspect, embodiments of this application provide a dynamic spectrum signal identification method, including: Acquire the dynamic spectrum signal to be identified, wherein the dynamic spectrum signal has a time dimension and a frequency dimension; The dynamic spectrum signal is divided along the time dimension by a preset size threshold to obtain several signal blocks with the same frequency dimension; The signal blocks are input into the signal classification model to obtain the confidence level of the dynamic spectrum signal for the single-pulse signal type. When the confidence level reaches a preset threshold, the single pulse signal type is used as the signal type of the dynamic spectrum signal.

[0006] Secondly, embodiments of this application also provide a dynamic spectrum signal identification device, comprising: The acquisition module is used to acquire the dynamic spectrum signal to be identified, wherein the dynamic spectrum signal has a time dimension and a frequency dimension; The processing module is used to segment the dynamic spectrum signal along the time dimension by a preset size threshold to obtain several signal blocks with the same frequency dimension; The analysis module is used to input the signal blocks into the signal classification model to obtain the confidence level of the dynamic spectrum signal for the single-pulse signal type. The determination module is used to determine the single-pulse signal type as the signal type of the dynamic spectrum signal when the confidence level reaches a preset threshold.

[0007] In some embodiments of this application, the signal classification model includes a discriminant network and a fusion network, and the analysis module includes: A discrimination unit is used to input the signal blocks into the discrimination network respectively to obtain the local confidence of each signal block for the single-pulse signal type; A fusion unit is used to input each of the local confidence scores into the fusion network to obtain a comprehensive confidence score; A determining unit is used to use the overall confidence level as the confidence level of the dynamic spectrum signal for the single-pulse signal type.

[0008] In some embodiments of this application, the discriminant network includes convolutional layers and discriminant layers, and the discriminant unit includes: An extraction subunit is used to input the signal block into the convolutional layer for each signal block to obtain the block signal feature information corresponding to the signal block; The discrimination subunit is used to input the block signal feature information into the discrimination layer to obtain the local confidence level of the signal block for the single-pulse signal type.

[0009] In some embodiments of this application, the device further includes a training module, which includes: The acquisition unit is used to acquire sample training data and sample signal classification model. The sample training data includes sample dynamic spectrum signals and sample signal types corresponding to the sample dynamic spectrum signals. The sample signal classification model is the model corresponding to the signal classification model before training. The segmentation unit is used to segment the sample dynamic spectrum signal along the time dimension according to the preset size threshold to obtain several sample signal blocks with the same frequency dimension. The first input unit is used to input the sample signal block into the discriminant network of the sample signal classification model for each sample signal block, so as to obtain the sample local confidence of the sample signal block for each sample signal type. The second input unit is used to input the local confidence scores of each sample into the fusion network of the sample signal classification model to obtain the comprehensive confidence score of the sample dynamic spectrum signal for each sample signal type. The training unit is used to train the sample signal classification model based on the sample comprehensive confidence level and the sample signal type of the sample dynamic spectrum signal, so as to obtain the trained signal classification model.

[0010] In some embodiments of this application, the training unit includes: The benchmark loss determination subunit is used to determine the benchmark loss based on the sample comprehensive confidence level and the sample signal type of the sample dynamic spectrum signal; The training subunit is used to train the sample signal classification model based on the benchmark loss to obtain the trained signal classification model.

[0011] In some embodiments of this application, the training unit further includes: A positive and negative sample determination subunit is used to determine positive and negative samples from the sample training data based on the sample signal type; A block-level loss determination subunit is used to determine the block-level loss of a sample signal block for each sample signal block of a negative sample, based on the sample local confidence of the sample signal block for each sample signal type and the sample signal type to which the sample signal block belongs. The sample-level loss determination subunit is used to determine the sample-level loss of the negative sample based on the block-level loss of each sample signal block in the negative sample. The training subunit is specifically used for: When the training data is positive, the sample signal classification model is trained according to the baseline loss to obtain the trained signal classification model. When the training data is negative, the sample signal classification model is trained based on the baseline loss and the sample-level loss to obtain the trained signal classification model.

[0012] In some embodiments of this application, the second input unit includes: The first input subunit is used to input the local confidence scores of each sample into the decision layer to obtain the sample fusion weights corresponding to the local confidence scores of each sample. The second input subunit is used to input the local confidence scores of each sample and the fusion weights of each sample into the fusion layer to obtain the comprehensive confidence score of the sample dynamic spectrum signal for each sample signal type.

[0013] In some embodiments of this application, the training unit includes: A fusion loss determination subunit is used to determine the fusion loss based on the entropy of the distribution of the fusion weights of each sample; The benchmark loss determination subunit is used to determine the benchmark loss based on the sample comprehensive confidence level and the sample signal type of the sample dynamic spectrum signal; The training subunit is used to train the sample signal classification model based on the benchmark loss to obtain the trained signal classification model.

[0014] Thirdly, embodiments of this application also provide an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps in the dynamic spectrum signal recognition method described above.

[0015] Fourthly, embodiments of this application also provide a storage medium storing a computer program, which, when executed by a processor, implements the steps in the dynamic spectrum signal recognition method described above.

[0016] Fifthly, embodiments of this application also provide a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations described in embodiments of this application.

[0017] In this embodiment, a dynamic spectrum signal to be identified is acquired. This dynamic spectrum signal has a time dimension and a frequency dimension. The dynamic spectrum signal is segmented along the time dimension by a preset size threshold to obtain several signal blocks with the same frequency dimension. Each signal block is input into a signal classification model to obtain the confidence level of the dynamic spectrum signal for a single-pulse signal type. When the confidence level reaches the preset threshold, the single-pulse signal type is taken as the signal type of the dynamic spectrum signal. Since the dynamic spectrum signal is a signal with a relatively fixed frequency observation interval but a time observation range that varies with the observed sky area, this embodiment segments the dynamic spectrum signal based on the time dimension. This ensures that the width of each segmented signal block in the time dimension is the same, meaning each signal block has the same resolution. When the model analyzes signal blocks with the same resolution, it can improve the accuracy of the model's analysis of each signal block, thereby improving the accuracy of dynamic spectrum signal identification. Analyzing the signal type corresponding to each dynamic spectrum signal through a model improves the efficiency of dynamic spectrum signal identification compared to exhaustive search or manual screening methods in related technologies. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram illustrating that the dynamic spectrum signal provided in this application is a single-pulse signal; Figure 2 This is a schematic diagram illustrating that the dynamic spectrum signal provided in the embodiments of this application is not a single pulse signal; Figure 3 This is a schematic diagram of a scenario for the dynamic spectrum signal recognition method provided in an embodiment of this application; Figure 4 This is a flowchart illustrating the dynamic spectrum signal recognition method provided in an embodiment of this application; Figure 5 This is another flowchart illustrating the dynamic spectrum signal recognition method provided in this application embodiment; Figure 6 This is a schematic diagram of the training process of the sample signal classification model provided in the embodiments of this application; Figure 7 This is a simplified flowchart illustrating the training process of the sample signal classification model provided in the embodiments of this application; Figure 8 This is a schematic diagram of the experimental results obtained by the dynamic spectrum signal recognition method provided in the embodiments of this application; Figure 9 This is a schematic diagram of the dynamic spectrum signal recognition device provided in the embodiments of this application; Figure 10 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0020] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] With the development of astrophysics and radio signal technology, more and more cosmic signals can be observed and recorded by radio telescopes deployed in various regions of the world. Among them, the Five-hundered-meter Aperture Spherical Radio (FAST), the world's largest and most sensitive single-aperture radio telescope, can receive electromagnetic radiation signals from tens of thousands of light-years away. These signals can help astronomers discover celestial objects in the universe that are of great scientific significance.

[0022] At the same time, with the increase in sensitivity and observation range, the amount of data received by the FAST telescope every day has also increased explosively. Processing this data means that electromagnetic signals emitted by extraterrestrial objects need to be selected from the terabyte (TB) of data. For this purpose, large-scale computer data processing and artificial intelligence-assisted celestial object identification systems have been designed to help quickly find extraterrestrial pulsars.

[0023] Currently, some AI-assisted systems have been deployed at astronomical observation sites. However, these systems are mainly used for the identification of periodic pulse signals. For the identification and search of dynamic spectrum signals, there is still no complete AI-based solution. The current approach relies on feature fitting from the dynamic spectrum images of single-pulse signals for screening, leading to numerous false alarms and the time cost of secondary observations. Specifically, the current mainstream single-pulse dynamic spectrum signal discrimination process mainly uses an exhaustive search method for discrimination. First, the astronomical signal processing software processes the signals recorded by the FAST telescope into... Figure 1 , Figure 2 The two-dimensional dynamic spectrum signal array shown is, in which, Figure 1 This is a schematic diagram illustrating that the dynamic spectrum signal provided in this application is a single-pulse signal, wherein... Figure 1 The black box in the middle shows a single-pulse signal pattern. Figure 2 This is a schematic diagram illustrating a dynamic spectral signal that is not a single-pulse signal, as provided in this embodiment. The array is normalized and after a dispersion-reducing process. A sub-region of the array is then acquired in a sliding manner within a fixed time window. Lorentz fitting is performed on the signal points within this sub-region. If the fitting accuracy is higher than a set threshold, and the mean signal intensity within the sub-region is also higher than the set threshold, the signal is considered a potential single-pulse signal candidate sample and is then used for further downstream observation to assist in the judgment. Alternatively, some other schemes directly resample and deform the two-dimensional dynamic spectral signal before feeding it into an artificial intelligence model for training. The trained model is then used to distinguish the same two-dimensional signal.

[0024] However, the Lorentz fitting method involves many fitting parameters that need to be manually adjusted, and its fitting ability varies for different signal morphologies. When the fitting accuracy threshold is too high, it is easy to miss the real positive sample pulse signal. When the threshold is too low, it will cause a large number of false alarms, increasing the time cost and pressure of downstream secondary verification. At the same time, the scheme itself relies on sliding window to fit segment by segment, which has low running efficiency. On the other hand, directly treating the two-dimensional array of dynamic spectrum as an image input to the artificial intelligence model can, to some extent, leverage the strong fitting ability of deep models to cover different signal morphologies and enhance the generalization ability for sample discrimination. Simultaneously, existing graphics processing unit (GPU) frameworks can be used to accelerate the artificial intelligence model. However, since dynamic spectrum signals differ from periodic signals, the frequency observation range of dynamic spectrum signals is relatively fixed, while the time observation range varies with the observed sky area, for example, constantly changing within the range of 0.1s to 10s. When using the same time sampling length (e.g., sampling once every 100us), the temporal resolution of different observed signals will have significant dynamic changes. This large range of resolution changes poses a significant challenge to the fitting ability of deep models. Therefore, when the input two-dimensional array has a large time span, directly resampling the original array will cause severe signal distortion, leading the trained artificial intelligence model to tend to fit distorted signals, resulting in omissions and false alarms.

[0025] Based on the above, embodiments of this application provide a dynamic spectrum signal recognition method, apparatus, electronic device, and storage medium. By processing the dynamic spectrum signal, the model can avoid resampling the dynamic spectrum signal during processing, thus preventing signal distortion and improving the accuracy of dynamic spectrum signal recognition. Specifically, embodiments of this application provide a dynamic spectrum signal recognition method suitable for electronic devices, wherein the electronic device includes a terminal or a server. The terminal includes, but is not limited to, devices such as computers, mobile phones, or astronomical telescopes. The server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The server can be directly or indirectly connected via wired or wireless communication.

[0026] In this embodiment, the dynamic spectrum signal recognition method can be executed by the terminal device alone, by the server alone, or by both the terminal device and the server. Please refer to [link to relevant documentation]. Figure 3 Taking the joint execution of the dynamic spectrum signal recognition method by the terminal device and the server as an example, the specific execution process of the dynamic spectrum signal recognition method is as follows: After acquiring the dynamic spectrum signal, the terminal device 10 sends the dynamic spectrum signal to the server 11, which then identifies and analyzes the dynamic spectrum signal. Specifically: After acquiring the dynamic spectrum signal, server 11 divides the dynamic spectrum signal along the time dimension into a preset size threshold to obtain several signal blocks with the same frequency dimension. Then, the several signal blocks are input into the signal classification model to obtain the confidence level of the dynamic spectrum signal for the pulse signal type. When the confidence level reaches the preset threshold, the single pulse signal type is taken as the signal type of the dynamic spectrum signal.

[0027] In this embodiment of the application, the terminal device can directly observe the dynamic spectrum signal. For example, the terminal device can be an astronomical telescope, or it can receive dynamic spectrum signals sent by other terminal devices, such as receiving dynamic spectrum signals observed by an astronomical telescope.

[0028] The dynamic spectrum signal recognition method provided in this application relates to machine learning in the field of artificial intelligence. This application can improve the efficiency and accuracy of dynamic spectrum signal recognition.

[0029] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to have perception, reasoning, and decision-making capabilities. AI technology is a comprehensive discipline involving a wide range of fields, encompassing both hardware and software technologies. AI software technologies mainly include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0030] Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0031] For example, in this embodiment of the application, the dynamic spectrum signal is identified and processed in the segmented signal block according to the signal classification model to obtain the confidence level of the dynamic spectrum signal for the single pulse signal type; wherein, the signal classification model can be trained by sample data.

[0032] In this embodiment, a dynamic spectrum signal to be identified is acquired. This dynamic spectrum signal has a time dimension and a frequency dimension. The dynamic spectrum signal is segmented along the time dimension by a preset size threshold to obtain several signal blocks with the same frequency dimension. Each signal block is input into a signal classification model to obtain the confidence level of the dynamic spectrum signal for a single-pulse signal type. When the confidence level reaches the preset threshold, the single-pulse signal type is taken as the signal type of the dynamic spectrum signal. Since the dynamic spectrum signal is a signal with a relatively fixed frequency observation interval but a time observation range that varies with the observed sky area, this embodiment segments the dynamic spectrum signal based on the time dimension. This ensures that the width of each segmented signal block in the time dimension is the same, meaning each signal block has the same resolution. When the model analyzes signal blocks with the same resolution, it can improve the accuracy of the model's analysis of each signal block, thereby improving the accuracy of dynamic spectrum signal identification. Analyzing the signal type corresponding to each dynamic spectrum signal through a model improves the efficiency of dynamic spectrum signal identification compared to exhaustive search or manual screening methods in related technologies.

[0033] The following sections provide detailed descriptions of each example. It should be noted that the order in which the embodiments are described is not intended to limit the priority of the embodiments.

[0034] Please see Figure 4 , Figure 4 This is a flowchart illustrating the dynamic spectrum signal recognition method provided in an embodiment of this application. The specific flow of the dynamic spectrum signal recognition method is as follows: 101. Obtain the dynamic spectrum signal to be identified, wherein the dynamic spectrum signal has a time dimension and a frequency dimension.

[0035] In this embodiment of the application, the dynamic spectrum signal is the electromagnetic signal after processing. For example, the electromagnetic signal is processed by dedispersion, interference removal and normalization to obtain a data array, which is the dynamic spectrum signal.

[0036] In this embodiment, the dynamic spectrum signal has a frequency dimension and a time dimension, which are beneficial for the identification of the dynamic spectrum signal.

[0037] In this embodiment of the application, the electromagnetic signal is emitted by celestial bodies in the universe and received by an astronomical telescope. The electromagnetic signal mainly refers to the signal in the radio wave band.

[0038] In this application embodiment, the astronomical telescope is the main tool for observing celestial bodies and capturing celestial information. For example, the Five-hundered-meter Aperture Spherical Radio, which is the world's largest and most sensitive single-aperture radio telescope, is referred to as the FAST astronomical telescope.

[0039] Since dynamic spectrum signals are obtained by converting signals emitted by celestial bodies in the universe, acquiring dynamic spectrum signals facilitates scientific research related to the universe. For example, by analyzing the frequency and time dimensions of dynamic spectrum signals, it is easy to identify the signal type to which the dynamic spectrum signal belongs, and thus identify the type of celestial body that emitted the electromagnetic signal corresponding to the dynamic spectrum signal.

[0040] 102. The dynamic spectrum signal is divided along the time dimension by a preset size threshold to obtain several signal blocks with the same frequency dimension.

[0041] In this embodiment, the segmentation of the dynamic spectrum signal can be understood as dividing the data array corresponding to the dynamic spectrum signal into multiple small sub-data arrays. Each sub-data array corresponds to a signal block, which has the characteristics of having the same frequency dimension and the same width in the time dimension. That is, when the data array corresponding to the dynamic spectrum signal is regarded as an image, each signal block is a sub-image (a small part of the image) corresponding to the image, and each sub-image has the same length and width, that is, each signal block has the same resolution.

[0042] Since dynamic spectrum signals have a relatively fixed frequency observation range, while the time observation range changes with the observation sky area, the width of the time dimension corresponding to different dynamic spectrum signals is different, that is, the resolution of the array corresponding to the dynamic spectrum signal is different. Therefore, if the digital array corresponding to the dynamic spectrum signal is directly resampled, it will cause serious signal distortion, causing the trained artificial intelligence model to tend to fit the distorted signal, resulting in omissions and false alarms.

[0043] Therefore, in this embodiment, by segmenting the dynamic spectrum signal by a preset size threshold, several signal blocks are obtained, so that each signal block has the same resolution, which facilitates the model's identification and analysis of each signal block. Compared with the resampling of the dynamic spectrum signal in related technologies, this embodiment does not destroy the data of the dynamic spectrum signal in the time and frequency dimensions, thus improving the accuracy of model prediction.

[0044] In this embodiment, since the dynamic spectrum signal is a data array obtained by processing signals received by an astronomical telescope, different dynamic spectrum signals have different time dimensions in the processed data array. Furthermore, after segmenting the dynamic spectrum signal, the size selection may not satisfy the time dimensions of all dynamic spectrum signals. Therefore, there is a requirement that the time dimension width of the segmented signal blocks does not meet the preset size threshold. Thus, in this embodiment, to enable the signal classification model to process signal blocks with the same frequency dimension and the same time dimension width (i.e., the same image resolution), the segmented signal blocks whose size does not meet the requirement can be supplemented. Optionally, in some embodiments of this application, the step "segmenting the dynamic spectrum signal along the time dimension according to the preset size threshold to obtain several signal blocks with the same frequency dimension" includes: The dynamic spectrum signal is divided along the time dimension by a preset size threshold to obtain several reference signal blocks with the same frequency dimension; When the size of the reference signal block in the time dimension direction does not reach the preset size threshold, the reference signal block is padded along the time dimension direction to obtain the padded reference signal block. The reference signal block whose size meets the preset size threshold and the padded reference signal block are respectively used as signal blocks.

[0045] The padding of the reference signal blocks can be achieved by padding with zeros in the time dimension. This ensures that each signal block has the same width in the time dimension, guaranteeing that each signal block has the same resolution when processed as an image.

[0046] 103. Input the signal blocks into the signal classification model respectively to obtain the confidence level of the dynamic spectrum signal for the single pulse signal type.

[0047] In this embodiment, the signal classification model is a machine learning model, which can also be understood as an artificial intelligence model. By analyzing each signal block, it is easy to obtain the confidence level of the dynamic spectrum signal relative to the single pulse signal type. By obtaining this confidence level, it is easy to determine whether the dynamic spectrum signal belongs to the single pulse signal.

[0048] In this embodiment, the confidence level of the dynamic spectrum signal for the single-pulse signal type can be obtained by combining the confidence levels of each signal block for the single-pulse signal type. For example, in this embodiment, the confidence level of each signal block for the single-pulse signal type can be obtained separately, and then the confidence level of the dynamic spectrum signal for the single-pulse signal type can be obtained based on the combined result of each confidence level. That is, optionally, in some embodiments of this application, the signal classification model includes a discriminant network and a fusion network, and the step "inputting the signal blocks into the signal classification model respectively to obtain the confidence level of the dynamic spectrum signal for the single-pulse signal type" includes: The signal blocks are respectively input into the discrimination network to obtain the local confidence of each signal block for the single-pulse signal type; Each of the local confidence scores is input into the fusion network to obtain the overall confidence score; The overall confidence level is used as the confidence level of the dynamic spectrum signal for the single-pulse signal type.

[0049] In this embodiment, the signal classification model mainly consists of a discriminant network and a fusion network. The discriminant network is mainly used to obtain the local confidence level of the single-pulse signal type corresponding to each input signal block, while the fusion network is used to fuse the local confidence levels of multiple signal blocks belonging to the same dynamic spectrum signal to obtain the overall confidence level of the dynamic spectrum signal for the single-pulse signal type.

[0050] In this embodiment, when the discriminant network analyzes the input signal block to obtain the local confidence level of the signal block for the single-pulse signal type, the block signal features of each signal block can be obtained through a convolutional network. Then, based on the block signal features, the confidence level of the signal block for the single-pulse signal type is obtained. That is, optionally, in some embodiments of this application, the discriminant network includes a convolutional layer and a discriminant layer. The step "inputting the signal blocks into the discriminant network respectively to obtain the local confidence level of each signal block for the single-pulse signal type" includes: For each signal block, the signal block is input into the convolutional layer to obtain the block signal feature information corresponding to the signal block; The block signal feature information is input into the discrimination layer to obtain the local confidence level of the signal block for the single-pulse signal type.

[0051] In this embodiment, the convolutional layer is used to treat the input signal block as an image, and then extract features from the image to obtain the block signal feature information corresponding to the signal block; the discriminant layer is used to discriminate the block signal feature information to obtain the local confidence level of the block signal feature information for the single-pulse signal type.

[0052] In this embodiment of the application, the specific algorithms for each network layer of the signal network model are not limited.

[0053] In some embodiments of this application, the signal classification model composed of the discriminant network and the fusion network can be obtained by training with sample data. That is, optionally, in some embodiments of this application, before the step of "obtaining the dynamic spectrum signal to be identified", the method further includes: Acquire sample training data and sample signal classification model. The sample training data includes sample dynamic spectrum signals and sample signal types corresponding to the sample dynamic spectrum signals. The sample signal classification model is the model corresponding to the signal classification model before training. The sample dynamic spectrum signal is divided along the time dimension according to the preset size threshold to obtain several sample signal blocks with the same frequency dimension; For each of the sample signal blocks, the sample signal block is input into the discriminant network of the sample signal classification model to obtain the sample local confidence of the sample signal block for each of the sample signal types; The local confidence scores of each sample are input into the fusion network of the sample signal classification model to obtain the comprehensive confidence score of the sample dynamic spectrum signal for each sample signal type. Based on the comprehensive confidence level of the samples and the sample signal type of the dynamic spectrum signal of the samples, the sample signal classification model is trained to obtain the trained signal classification model.

[0054] The process involves training a sample signal classification model using sample training data to obtain the signal classification model after training. By segmenting the dynamic spectrum signal of the sample using a preset size threshold, the model can improve the accuracy of signal block classification for signal blocks of the same time dimension width during training and actual classification.

[0055] Among them, the accuracy of the model classification is improved by training the sample signal classification model through the sample comprehensive confidence level and the sample signal type of the sample dynamic spectrum signal.

[0056] In this embodiment of the application, when training the sample signal classification model based on the sample comprehensive confidence level and the sample signal type of the sample dynamic spectrum signal, the loss can be determined based on the sample comprehensive confidence level and the sample signal type of the sample dynamic spectrum signal, and the model can be trained based on this loss. That is, optionally, in some embodiments of this application, the step "training the sample signal classification model based on the sample comprehensive confidence level and the sample signal type corresponding to the sample dynamic spectrum signal to obtain the trained signal classification model" includes: The baseline loss is determined based on the overall confidence level of the samples and the sample signal type of the dynamic spectrum signal of the samples; The sample signal classification model is trained based on the baseline loss to obtain the trained signal classification model.

[0057] In this embodiment of the application, the corresponding benchmark loss can be obtained based on the difference between the overall confidence level of the sample and the sample signal type.

[0058] Among them, the loss is determined by the classification of the model during the training process and the actual type of the sample data. This makes it easier to train the model based on the loss, so that the classification result of the trained model is closer to the actual classification result of the signal.

[0059] In this embodiment, for the segmented signal blocks, each signal block has a different signal type depending on whether it contains monopulse signal type characteristics (in this embodiment, this mainly refers to monopulse signal type and non-monopulse signal type), as detailed below: When the sample dynamic spectrum signal is a single-pulse signal, the data array corresponding to the dynamic spectrum signal contains characteristics of the single-pulse signal type (such as...). Figure 1 The selected part in the middle box contains a drifting diagonal line in the data array. However, since only a small part of the data array contains this feature, when the dynamic spectrum signal is divided into multiple signal blocks, each signal block may contain this feature or may not. That is, for a dynamic spectrum signal of the single-pulse signal type, each signal block in the divided signal blocks may be of the single-pulse signal type or may not be of the single-pulse signal type. When the sample dynamic spectrum signal is not a single-pulse signal type, the data array corresponding to the dynamic spectrum signal does not contain features of the single-pulse signal type (such as...). Figure 2 In the data array, there are no drifting diagonal lines. Therefore, in each signal block after the dynamic spectrum signal is segmented, since no signal block contains the characteristics of a single pulse signal, correspondingly, for the dynamic spectrum signal of the single pulse signal type, each signal block in the segmented signal block is not a single pulse signal type.

[0060] Therefore, based on the above, a loss for negative samples can be added to increase the accuracy of the prediction results after model training. That is, optionally, in some embodiments of this application, before the step "training the sample signal classification model according to the benchmark loss to obtain the trained signal classification model," the method further includes: Positive and negative samples are determined from the sample training data based on the sample signal type. For each sample signal block of a negative sample, the block-level loss of the sample signal block is determined based on the sample local confidence of the sample signal block for each sample signal type and the sample signal type to which the sample signal block belongs. For the negative sample, the sample-level loss of the negative sample is determined based on the block-level loss of each sample signal block in the negative sample.

[0061] In this embodiment of the application, regarding the loss of negative samples, since it is clear that each signal block in the negative sample is a non-single-pulse signal, the corresponding loss can be calculated based on the classification results of the negative sample through the model and the actual results.

[0062] In this embodiment of the application, the sample-level loss of the corresponding negative sample can be obtained based on the average block-level loss of each signal block.

[0063] In this embodiment of the application, in order to improve the accuracy of model classification, when training the model based on the negative sample, the loss can be added to the training process to improve the accuracy of model prediction. That is, optionally, in some embodiments of this application, the step "training the sample signal classification model according to the benchmark loss to obtain the trained signal classification model" includes: When the training data is positive, the sample signal classification model is trained according to the baseline loss to obtain the trained signal classification model. When the training data is negative, the sample signal classification model is trained based on the baseline loss and the sample-level loss to obtain the trained signal classification model.

[0064] In this case, after segmenting positive samples, analyzing whether each signal block belongs to the single-pulse signal type is time-consuming and prone to inaccurate manual judgment. However, each signal block of negative samples can be directly identified as a non-single-pulse signal. Therefore, when training the model for negative samples, the loss of each signal block is obtained based on the actual type of each signal block and the model classification result. Training the model based on this loss can improve the accuracy of the classification results after training.

[0065] In this embodiment of the application, when fusing the sample local confidence scores of multiple signal blocks, fusion can be performed using a weighted approach. Specifically, in some embodiments of this application, the fusion network includes a decision layer and a fusion layer. The step "inputting each of the sample local confidence scores into the fusion network of the sample signal classification model to obtain the sample dynamic spectrum signal's comprehensive confidence score for each sample signal type" includes: The local confidence scores of each sample are input into the decision layer to obtain the sample fusion weights corresponding to the local confidence scores of each sample. The local confidence scores of each sample and the fusion weights of each sample are input into the fusion layer to obtain the comprehensive confidence score of the sample dynamic spectrum signal for each sample signal type.

[0066] Among them, the accuracy of the overall confidence score of the samples is improved by fusing the local confidence scores of each sample through sample fusion weights.

[0067] In some embodiments of this application, a sparsity constraint can be added to the generated fusion weights to make the fusion selective. That is, optionally, in some embodiments of this application, the step "training the sample signal classification model according to the sample comprehensive confidence level and the sample signal type corresponding to the sample dynamic spectrum signal to obtain the trained signal classification model" includes: The fusion loss is determined based on the entropy of the distribution of the fusion weights of each sample. The baseline loss is determined based on the overall confidence level of the samples and the sample signal type of the dynamic spectrum signal of the samples; The sample signal classification model is trained based on the baseline loss and the fusion loss to obtain the trained signal classification model.

[0068] Entropy reflects the distribution of data (such as the sample fusion weights in this scheme). The smaller the entropy, the sharper the probability distribution, and the more inclined it is to choose a few options. Conversely, the larger the entropy, the smoother the probability distribution. When the probability distribution is uniform, the entropy is the largest, and at this time, the contribution (sample fusion weights) of each data block is the same.

[0069] Please refer to Figure 1 As shown, after segmenting the dynamic spectrum signal, the effective signal often only exists in 1-2 segmented signal blocks. If too many signal blocks without effective signals are considered, it will affect the training of the model, making it difficult for the model to achieve good recognition performance after training.

[0070] The embodiments of this application, by constraining the sparsity of the sample fusion weights, effectively increase the fusion weights corresponding to the local confidence of some samples. In other words, the fusion weights of different signal blocks will not all be very close, but some will be relatively large while most others will be relatively small. This allows the confidence to be mainly concentrated on one or two signal blocks during fusion, while the fusion weights of other signal blocks are as small as possible, thereby improving the accuracy of the calculation of the overall confidence of the samples and further improving the recognition performance of the trained model.

[0071] 104. When the confidence level reaches a preset threshold, the single pulse signal type is used as the signal type of the dynamic spectrum signal.

[0072] In this embodiment, the preset threshold can be set according to actual needs, or the signal type with the highest confidence level can be used as the signal type of the dynamic spectrum signal.

[0073] When the confidence level meets the preset threshold requirement, the single-pulse signal type can be used as the signal type corresponding to the dynamic spectrum signal.

[0074] Accordingly, in the embodiments of this application, when the dynamic spectrum signal is analyzed to be a single-pulse signal type, the celestial body emitting the electromagnetic signal (the signal corresponding to the dynamic spectrum signal) can be used as the pulsar to be searched. That is, optionally, in some embodiments of this application, after the step "when the confidence level reaches a preset threshold, the single-pulse signal type is used as the signal type of the dynamic spectrum signal", the method further includes: When the dynamic spectrum signal is a single pulse signal, the celestial body that emits the dynamic spectrum signal is marked as a pulsar.

[0075] Among these methods, by analyzing the signals emitted by celestial bodies and determining the type of the signal, it is possible to identify the celestial body that emitted the signal. Based on this analysis and identification, it is helpful to assist researchers in identifying and exploring celestial bodies in the universe.

[0076] In this embodiment, a dynamic spectrum signal to be identified is acquired. This dynamic spectrum signal has a time dimension and a frequency dimension. The dynamic spectrum signal is segmented along the time dimension by a preset size threshold to obtain several signal blocks with the same frequency dimension. Each signal block is input into a signal classification model to obtain the confidence level of the dynamic spectrum signal for a single-pulse signal type. When the confidence level reaches the preset threshold, the single-pulse signal type is taken as the signal type of the dynamic spectrum signal. Since the dynamic spectrum signal is a signal with a relatively fixed frequency observation interval but a time observation range that varies with the observed sky area, this embodiment segments the dynamic spectrum signal based on the time dimension. This ensures that the width of each segmented signal block in the time dimension is the same, meaning each signal block has the same resolution. When the model analyzes signal blocks with the same resolution, it can improve the accuracy of the model's analysis of each signal block, thereby improving the accuracy of dynamic spectrum signal identification. Analyzing the signal type corresponding to each dynamic spectrum signal through a model improves the efficiency of dynamic spectrum signal identification compared to exhaustive search or manual screening methods in related technologies.

[0077] Please see Figure 5 , Figure 5This is another flowchart illustrating the dynamic spectrum signal recognition method provided in this application embodiment, wherein the dynamic spectrum signal recognition process specifically includes: 201. Obtain electromagnetic signals recorded by astronomical telescopes; Among them, astronomical telescopes are the main tools for observing celestial bodies and capturing celestial information. By observing and capturing signals emitted by celestial bodies in the universe, astronomical telescopes facilitate the analysis of celestial bodies in the universe.

[0078] 202. Preprocess the electromagnetic signal to obtain the dynamic spectrum signal; In this embodiment of the application, the preprocessing of the electromagnetic signal includes de-dispersion, interference removal, normalization, etc. Through the processing of the electromagnetic signal, a corresponding data array is obtained, which is the dynamic spectrum signal.

[0079] 203. Divide the dynamic spectrum signal into several signal blocks by a preset size threshold; In this embodiment of the application, the dynamic spectrum signal is segmented according to a preset size threshold so that each signal block has the same frequency dimension and the same width in the time dimension. That is, when the dynamic spectrum signal is regarded as an image, each segmented signal block has the same length and width (same resolution).

[0080] For example, for a dynamic spectrum signal represented vertically by frequency and horizontally by time, the corresponding data array... The cutting module in this solution cuts the data at a fixed width. Horizontal segmentation is performed to obtain a bag-of-words (Bag(x)) consisting of multiple data blocks (signal blocks); ), represented as:

[0081] Where k represents the k-th data block (signal block), for example, data block x_k, prediction score s_k, fusion coefficient alpha_k, etc. If the width of the last segmented signal block is insufficient... Then, the signal block is padded with zeros in the horizontal direction to the width. This method of partitioning divides the original data array into... Each data point has the same width. signal block This operation uses horizontal cutting throughout and does not resample the array data, so the original data shape is preserved within each data block, and no deformation problem occurs.

[0082] 204. Input several signal blocks into the trained signal classification model to obtain the confidence level of the dynamic spectrum signal for each sample signal type; Among them, the constructed bag of words Bag(x; The signal blocks (data blocks) in the signal are fed one by one into the discriminant network of a deep neural network (CNN, signal classification model) for feature extraction and discrimination scoring (local confidence for each signal type), obtaining the feature and score information of each signal block. For example, ,in, This represents a feature extractor that transforms an input two-dimensional data block into a feature description vector of length d through convolution and pooling operations. , and The discriminator transforms the input feature description vector into a confidence vector of length 2 through a linear transformation, representing the normalized confidence scores of negative and positive samples, respectively. ; These represent the learnable parameters of the two modules, respectively.

[0083] After obtaining the local confidence scores of each signal block in the dynamic spectrum signal, the local confidence scores can be fused using a fusion network of the signal classification model to obtain the confidence scores of the dynamic spectrum signal for each signal type. For example, the fusion network can be represented as:

[0084]

[0085] Here, k and j represent the k-th or j-th data block, respectively. For example, data x_k, prediction score s_k, fusion coefficient alpha_k, etc. These are the learnable parameters of the module. Through this fusion module, we can leverage the learned baseline vector features. The contribution of different data blocks is evaluated, and the evaluation scores are used. The scores of the data blocks are reweighted to obtain the scores of the original data. .

[0086] In this embodiment, the similarity-based score fusion module can be replaced by other function modules with aggregation and reduction functions, such as mean averaging, maximum value, global momentum update, and other fusion functions.

[0087] In this embodiment, the fusion network is constructed based on an attention mechanism, which fuses the scores of each data block and uses the fused score as the score of the original dynamic spectrum signal (original image) x.

[0088] 205. Based on the confidence level, determine whether the dynamic spectrum signal belongs to the single pulse signal type. If it belongs to the single pulse signal type, proceed to step 206. In this embodiment of the application, the signal type of the dynamic spectrum signal can be determined by whether the confidence level meets a preset threshold.

[0089] 206. The celestial body that emits the electromagnetic signal corresponding to the dynamic spectrum signal is labeled as a pulsar.

[0090] By identifying the type of dynamic spectrum signal, the type of celestial body that emits the corresponding electromagnetic signal can be identified. For example, when the dynamic spectrum signal is a single pulse signal, the celestial body that emits the electromagnetic signal is marked as a pulsar.

[0091] The identification of signals emitted by celestial bodies facilitates the identification of the celestial bodies emitting those signals. Once the corresponding celestial body is identified, it enables astronomers to explore and study the universe. By classifying segmented information blocks of the same resolution using an information classification model, the model can process information blocks of the same resolution for different dynamic spectrum signals, avoiding resampling of the dynamic spectrum signal and improving the accuracy of the model's classification results.

[0092] Please refer to Figure 6 , Figure 6 This is a schematic diagram of the training process for the sample signal classification model provided in an embodiment of this application. The training process includes: 211. Acquire sample electromagnetic signals and the sample signal type corresponding to the sample electromagnetic signals, wherein the sample signal type includes single pulse signal type and non-single pulse signal type; 212. Preprocess the electromagnetic signal to obtain the corresponding sample dynamic spectrum signal; The training dataset with positive and negative labels is represented as a set. ,in, This represents a two-dimensional data array with height H and width W. The physical meaning of height H is the number of frequency channels, and the physical meaning of width W is the temporal resolution. Each sample corresponds to one label. y=0 indicates that the sample is a negative sample, and y=1 indicates that the sample is a positive pulsar signal.

[0093] 213. Divide the sample dynamic spectrum signal along the time dimension according to a preset size threshold to obtain several sample signal blocks with the same frequency dimension; Please refer to Figure 7 , Figure 7 This is a simplified flowchart illustrating the training process of the sample signal classification model provided in this application embodiment, wherein the sample dynamic spectrum signal A is divided into a bag-of-words B, which contains multiple sample signal blocks.

[0094] 214. For each sample signal block, input the sample signal block into the convolutional layer of the sample signal classification model to obtain the sample block signal feature information; Please refer to Figure 7 Each sample signal block in the bag of words is input into the convolutional layer of the discriminant network C of the sample signal classification model to obtain the sample block signal feature information of each sample signal block.

[0095] 215. Input the signal feature information of the sample block into the discriminant layer of the sample signal classification model to obtain the sample local confidence of the sample signal block for each sample signal type; Please refer to Figure 7 The sample block signal feature information of each sample signal block is input into the discriminant layer of the discriminant network C to obtain the sample local confidence of the sample signal block for each sample signal type.

[0096] 216. For each sample dynamic spectrum signal, the sample local confidence of each sample signal block corresponding to the sample dynamic spectrum signal is input into the decision layer and fusion layer of the fusion network D to obtain the sample confidence of the sample dynamic spectrum signal for each sample signal type. In this embodiment of the application, the fusion network can be represented as:

[0097]

[0098] Here, k and j represent the k-th or j-th data block, respectively. For example, data x_k, prediction score s_k, fusion coefficient alpha_k, etc. These are the learnable parameters of the module. Through this fusion module, we can leverage the learned baseline vector features. The contribution of different data blocks is evaluated, and the evaluation scores are used. The scores of the data blocks are reweighted to obtain the scores of the original data.

[0099] Please refer to Figure 7 The local confidence scores of each sample are input into the fusion network to obtain the sample confidence scores of the dynamic spectrum signals for each sample signal type.

[0100] 217. Determine whether the dynamic spectrum signal of the sample is a positive sample based on the sample signal type. If it is, proceed to step 218; otherwise, proceed to step 220. The original input data is: The value of y can only indicate whether the original data x is a positive or negative sample. When x is split into data blocks, there are two cases: 1. When x is a positive sample, i.e., y=1: ​​Since only a small portion of the data array x possesses positive sample characteristics (drift lines), the data blocks divided from x... It could be a positive sample or a negative sample.

[0101] 2. When x is a negative sample, i.e., y=0: Since there are no positive sample features (drift lines) in the original data array x, the data blocks divided from x are... It must be a negative sample.

[0102] Therefore, as Figure 7 As shown, the corresponding losses are calculated for positive and negative samples respectively.

[0103] 218. Determine the baseline loss based on the sample confidence level of the dynamic spectrum signal for each sample signal type and the sample signal type of the dynamic spectrum signal. In this embodiment of the application, the benchmark loss E can be expressed as:

[0104] Where s represents the confidence level (score) of the sample dynamic spectrum signal, y represents the signal type of the sample dynamic spectrum signal. In the embodiments of this application, y can take the values ​​0 and 1, where 0 indicates that the dynamic spectrum signal is a non-single pulse signal type, and 1 indicates that the dynamic spectrum signal is a single pulse signal type. T represents the transpose of a matrix or vector, and the log operation represents taking the logarithm of each element in the vector s.

[0105] 219. Based on this benchmark loss, the sample signal classification model is trained to obtain the signal classification model; 220. Based on the sample confidence of the dynamic spectrum signal for each sample signal type and the sample signal type of the dynamic spectrum signal, determine the baseline loss; and based on the fact that each signal block is a negative sample, determine the sample block-level loss of each signal block, and obtain the sample-level loss of the entire negative sample based on the sample block-level loss. In this embodiment, based on the fact that the blocks of negative samples are still negative samples, this scheme simultaneously calculates the cross-entropy loss at the data block level, which applies to all negative sample data blocks. The sample block-level loss F can be expressed as:

[0106] Where 0 indicates that the dynamic spectrum signal is not a single-pulse signal, 1 indicates that the dynamic spectrum signal is a single-pulse signal, T represents the transpose of a matrix or vector, and the log operation represents the transpose of a vector. Take the logarithm of each element in the array, where k represents the k-th signal block or data block.

[0107] Accordingly, the loss in this embodiment, based on positive and negative samples, is expressed as follows:

[0108] Among them, Bag(x; () represents the bag-of-words containing the dynamic spectrum signal after segmentation. Indicates the width of the signal block. This represents the width of the time dimension of the dynamic spectrum signal. [y=0] means that y=0 is 1, otherwise it is 0.

[0109] 221. The sample signal classification model is trained based on the baseline loss and the sample-level loss to obtain the sample signal classification model.

[0110] Specifically, by introducing a sample-level loss for negative samples on top of the baseline loss, the neural network is guided to accurately identify the more diverse and complex features of negative samples.

[0111] In the embodiments of this application, This represents the local confidence level of the k-th sample signal block (sample data block). The corresponding sample fusion weights, since the effective signal may exist in one or two signal blocks, can be adjusted during training. By imposing sparsity constraints, the overall confidence score of the samples is made to focus more on the fusion weights of individual samples. For example, in this embodiment, the entropy of the distribution of each sample fusion weight can be used as the fusion loss. The sample signal classification model is trained based on this fusion loss to obtain the signal classification model. Accordingly, in this embodiment, the fusion loss... This can be expressed by the following formula, specifically:

[0112] Accordingly, the final loss in the training sample signal classification model of this application embodiment can be expressed as:

[0113] Among them, Bag(x; () represents the bag-of-words containing the dynamic spectrum signal after segmentation. Indicates the width of the signal block. This represents the width of the time dimension of the dynamic spectrum signal. [y=0] means that y=0 is 1, otherwise it is 0.

[0114] In this embodiment, by constraining the sparsity of the sample fusion weights, the fusion weights corresponding to the local confidence of some samples are increased. That is, the fusion weights of different signal blocks are not all very close, but some are relatively large and most of the others are relatively small. This allows the confidence to be mainly concentrated on one or two signal blocks during fusion, while the fusion weights of other signal blocks are as small as possible. This improves the accuracy of the calculation of the overall confidence of the samples and further improves the recognition performance of the trained model.

[0115] Please see Figure 8 , Figure 8 This is a schematic diagram of the experimental results obtained by the dynamic spectrum signal recognition method provided in the embodiments of this application. When the search and recognition of single pulse signals is realized based on the dynamic spectrum signal recognition method (multi-instance learning) in the embodiments of this application, the retrieval recall rate is significantly increased, while still maintaining a lower false alarm rate.

[0116] This application focuses on the dynamic spectrum discrimination and search problem of single-pulse signals during pulsar search. To address the potential array signal distortion problem in the input of artificial intelligence models, it proposes an artificial intelligence model training scheme based on bag-of-words input and multi-instance (signal block) learning. Compared with the scheme that directly uses Lorentz fitting + sliding window, the multi-instance learning method proposed in this scheme helps the search model perform better in the task of pulsar candidate search, significantly improving the algorithm's retrieval ability while further suppressing the false alarm rate and reducing the manual requirements and time cost of subsequent secondary verification.

[0117] To facilitate better implementation of the dynamic spectrum signal recognition method of this application, this application also provides a dynamic spectrum signal recognition device based on the above-described dynamic spectrum signal recognition method. The meaning of the third target term is the same as in the above-described dynamic spectrum signal recognition method; specific implementation details can be found in the description of the method embodiments.

[0118] Please see Figure 9 , Figure 9 This is a schematic diagram of the structure of the dynamic spectrum signal recognition device provided in the embodiments of this application, wherein the dynamic spectrum signal recognition device may include: The acquisition module 301 is used to acquire the dynamic spectrum signal to be identified, wherein the dynamic spectrum signal has a time dimension and a frequency dimension; The processing module 302 is used to divide the dynamic spectrum signal along the time dimension into a preset size threshold to obtain several signal blocks with the same frequency dimension. Analysis module 303 is used to input the signal blocks into the signal classification model respectively to obtain the confidence level of the dynamic spectrum signal for the single pulse signal type; The determination module 304 is used to determine the single pulse signal type as the signal type of the dynamic spectrum signal when the confidence level reaches a preset threshold.

[0119] In some embodiments of this application, the signal classification model includes a discriminant network and a fusion network, and the analysis module 303 includes: A discrimination unit is used to input the signal blocks into the discrimination network respectively to obtain the local confidence of each signal block for the single-pulse signal type; A fusion unit is used to input each of the local confidence scores into the fusion network to obtain a comprehensive confidence score; A determining unit is used to use the overall confidence level as the confidence level of the dynamic spectrum signal for the single-pulse signal type.

[0120] In some embodiments of this application, the discriminant network includes convolutional layers and discriminant layers, and the discriminant unit includes: An extraction subunit is used to input the signal block into the convolutional layer for each signal block to obtain the block signal feature information corresponding to the signal block; The discrimination subunit is used to input the block signal feature information into the discrimination layer to obtain the local confidence level of the signal block for the single-pulse signal type.

[0121] In some embodiments of this application, the device further includes a training module, which includes: The acquisition unit is used to acquire sample training data and sample signal classification model. The sample training data includes sample dynamic spectrum signals and sample signal types corresponding to the sample dynamic spectrum signals. The sample signal classification model is the model corresponding to the signal classification model before training. The segmentation unit is used to segment the sample dynamic spectrum signal along the time dimension according to the preset size threshold to obtain several sample signal blocks with the same frequency dimension. The first input unit is used to input the sample signal block into the discriminant network of the sample signal classification model for each sample signal block, so as to obtain the sample local confidence of the sample signal block for each sample signal type. The second input unit is used to input the local confidence scores of each sample into the fusion network of the sample signal classification model to obtain the comprehensive confidence score of the sample dynamic spectrum signal for each sample signal type. The training unit is used to train the sample signal classification model based on the sample comprehensive confidence level and the sample signal type of the sample dynamic spectrum signal, so as to obtain the trained signal classification model.

[0122] In some embodiments of this application, the training unit includes: The benchmark loss determination subunit is used to determine the benchmark loss based on the sample comprehensive confidence level and the sample signal type of the sample dynamic spectrum signal; The training subunit is used to train the sample signal classification model based on the benchmark loss to obtain the trained signal classification model.

[0123] In some embodiments of this application, the training unit further includes: A positive and negative sample determination subunit is used to determine positive and negative samples from the sample training data based on the sample signal type; A block-level loss determination subunit is used to determine the block-level loss of a sample signal block for each sample signal block of a negative sample, based on the sample local confidence of the sample signal block for each sample signal type and the sample signal type to which the sample signal block belongs. The sample-level loss determination subunit is used to determine the sample-level loss of the negative sample based on the block-level loss of each sample signal block in the negative sample. The training subunit is specifically used for: When the training data is positive, the sample signal classification model is trained according to the baseline loss to obtain the trained signal classification model. When the training data is negative, the sample signal classification model is trained based on the baseline loss and the sample-level loss to obtain the trained signal classification model.

[0124] In some embodiments of this application, the second input unit includes: The first input subunit is used to input the local confidence scores of each sample into the decision layer to obtain the sample fusion weights corresponding to the local confidence scores of each sample. The second input subunit is used to input the local confidence scores of each sample and the fusion weights of each sample into the fusion layer to obtain the comprehensive confidence score of the sample dynamic spectrum signal for each sample signal type.

[0125] In some embodiments of this application, the training unit includes: A fusion loss determination subunit is used to determine the fusion loss based on the entropy of the distribution of the fusion weights of each sample; The benchmark loss determination subunit is used to determine the benchmark loss based on the sample comprehensive confidence level and the sample signal type of the sample dynamic spectrum signal; The training subunit is used to train the sample signal classification model based on the benchmark loss to obtain the trained signal classification model.

[0126] In some embodiments of this application, the processing module 302 includes: A segmentation unit is used to segment the dynamic spectrum signal along the time dimension by a preset size threshold to obtain several reference signal blocks with the same frequency dimension; The padding unit is used to pad the reference signal block along the time dimension direction when the size of the reference signal block in the time dimension direction does not reach the preset size threshold, so as to obtain a padded reference signal block. The processing unit is used to treat the reference signal block whose size meets the preset size threshold and the padded reference signal block as signal blocks respectively.

[0127] In some embodiments of this application, the device further includes an application module, which includes: The application unit is used to mark the celestial body that emits the dynamic spectrum signal as a pulsar when the dynamic spectrum signal is a single pulse signal.

[0128] In this embodiment, the acquisition module 301 acquires the dynamic spectrum signal to be identified. The dynamic spectrum signal has a time dimension and a frequency dimension. Then, the processing module 302 divides the dynamic spectrum signal along the time dimension into a preset size threshold to obtain several signal blocks with the same frequency dimension. Subsequently, the analysis module 303 inputs the signal blocks into a signal classification model to obtain the confidence level of the dynamic spectrum signal for the single-pulse signal type. Then, the determination module 304 determines the single-pulse signal type as the signal type of the dynamic spectrum signal when the confidence level reaches a preset threshold.

[0129] In this embodiment, a dynamic spectrum signal to be identified is acquired. This dynamic spectrum signal has a time dimension and a frequency dimension. The dynamic spectrum signal is segmented along the time dimension by a preset size threshold to obtain several signal blocks with the same frequency dimension. Each signal block is input into a signal classification model to obtain the confidence level of the dynamic spectrum signal for a single-pulse signal type. When the confidence level reaches the preset threshold, the single-pulse signal type is taken as the signal type of the dynamic spectrum signal. Since the dynamic spectrum signal is a signal with a relatively fixed frequency observation interval but a time observation range that varies with the observed sky area, this embodiment segments the dynamic spectrum signal based on the time dimension. This ensures that the width of each segmented signal block in the time dimension is the same, meaning each signal block has the same resolution. When the model analyzes signal blocks with the same resolution, it can improve the accuracy of the model's analysis of each signal block, thereby improving the accuracy of dynamic spectrum signal identification. Analyzing the signal type corresponding to each dynamic spectrum signal through a model improves the efficiency of dynamic spectrum signal identification compared to exhaustive search or manual screening methods in related technologies.

[0130] In addition, this application also provides an electronic device, such as Figure 10 As shown, it illustrates the structural diagram of the electronic device involved in this application, specifically: The electronic device may include components such as a processor 401 with one or more processing cores, a memory 402 with one or more computer-readable storage media, a power supply 403, and an input unit 404. Those skilled in the art will understand that... Figure 10 The electronic device structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein: The processor 401 is the control center of the electronic device, connecting various parts of the device via various interfaces and lines. It executes various functions and processes data by running or executing software programs and / or modules stored in the memory 402, and by calling data stored in the memory 402. Optionally, the processor 401 may include one or more processing cores; preferably, the processor 401 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 401.

[0131] The memory 402 can be used to store software programs and modules. The processor 401 executes various functional applications and data processing by running the software programs and modules stored in the memory 402. The memory 402 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device, etc. In addition, the memory 402 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 402 may also include a memory controller to provide the processor 401 with access to the memory 402.

[0132] The electronic device also includes a power supply 403 that supplies power to the various components. Preferably, the power supply 403 can be logically connected to the processor 401 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 403 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0133] The electronic device may also include an input unit 404, which can be used to receive input digital or character information, and generate keyboard, mouse, joystick, optical or trackball signal inputs related to object settings and function control.

[0134] Although not shown, the electronic device may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 401 in the electronic device loads the executable files corresponding to the processes of one or more application programs into the memory 402 according to the following instructions, and the processor 401 runs the application programs stored in the memory 402, thereby implementing the steps in any of the dynamic spectrum signal recognition methods provided in this application.

[0135] This application embodiment acquires a dynamic spectrum signal to be identified. This dynamic spectrum signal has a time dimension and a frequency dimension. The dynamic spectrum signal is segmented along the time dimension by a preset size threshold, resulting in several signal blocks with the same frequency dimension. These signal blocks are then input into a signal classification model to obtain the confidence level of the dynamic spectrum signal relative to a single-pulse signal type. When the confidence level reaches a preset threshold, the single-pulse signal type is identified as the signal type of the dynamic spectrum signal. Since the dynamic spectrum signal is a signal with a relatively fixed frequency observation interval but a time observation range that varies with the observed sky area, this application embodiment segments the dynamic spectrum signal based on the time dimension. This ensures that the width of each segmented signal block in the time dimension is the same, meaning each signal block has the same resolution. When the model analyzes signal blocks with the same resolution, it can improve the accuracy of the model's analysis of each signal block, thereby improving the accuracy of dynamic spectrum signal identification. Analyzing the signal type corresponding to each dynamic spectrum signal through a model, compared to exhaustive search or manual screening methods in related technologies, can improve the efficiency of dynamic spectrum signal identification.

[0136] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0137] Therefore, this application provides a computer-readable storage medium storing a computer program that can be loaded by a processor to execute the steps in any of the dynamic spectrum signal recognition methods provided in this application.

[0138] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0139] Since the instructions stored in the computer-readable storage medium can execute the steps in any of the dynamic spectrum signal recognition methods provided in this application, the beneficial effects that any of the dynamic spectrum signal recognition methods provided in this application can achieve can be realized, as detailed in the preceding embodiments, and will not be repeated here.

[0140] This application also provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations described in this application.

[0141] The above provides a detailed description of a dynamic spectrum signal identification method, apparatus, electronic device, and computer-readable storage medium provided in this application. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, those skilled in the art will recognize that there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for identifying dynamic spectrum signals, characterized in that, include: Acquire the dynamic spectrum signal to be identified, wherein the dynamic spectrum signal has a time dimension and a frequency dimension; The dynamic spectrum signal is divided along the time dimension by a preset size threshold to obtain several signal blocks with the same frequency dimension; The signal blocks are input into a signal classification model to obtain the confidence level of the dynamic spectrum signal for the single-pulse signal type. The signal classification model includes a discriminant network and a fusion network. The discriminant network includes convolutional layers and a discrimination layer. For each signal block, the signal block is input into the convolutional layer to obtain the block signal feature information corresponding to the signal block. The block signal feature information is input into the discrimination layer to obtain the local confidence level of the signal block for the single-pulse signal type. Each local confidence level is input into the fusion network to obtain the comprehensive confidence level. The comprehensive confidence level is used as the confidence level of the dynamic spectrum signal for the single-pulse signal type. When the confidence level reaches a preset threshold, the single pulse signal type is used as the signal type of the dynamic spectrum signal.

2. The method according to claim 1, characterized in that, Before acquiring the dynamic spectrum signal to be identified, the method further includes: Acquire sample training data and sample signal classification model. The sample training data includes sample dynamic spectrum signals and sample signal types corresponding to the sample dynamic spectrum signals. The sample signal classification model is the model corresponding to the signal classification model before training. The sample dynamic spectrum signal is divided along the time dimension according to the preset size threshold to obtain several sample signal blocks with the same frequency dimension; For each of the sample signal blocks, the sample signal block is input into the discriminant network of the sample signal classification model to obtain the sample local confidence of the sample signal block for each of the sample signal types; The local confidence scores of each sample are input into the fusion network of the sample signal classification model to obtain the comprehensive confidence score of the sample dynamic spectrum signal for each sample signal type. Based on the comprehensive confidence level of the samples and the sample signal type of the dynamic spectrum signal of the samples, the sample signal classification model is trained to obtain the trained signal classification model.

3. The method according to claim 2, characterized in that, The step of training the sample signal classification model based on the sample comprehensive confidence level and the sample signal type corresponding to the sample dynamic spectrum signal to obtain the trained signal classification model includes: The baseline loss is determined based on the overall confidence level of the samples and the sample signal type of the dynamic spectrum signal of the samples; The sample signal classification model is trained based on the baseline loss to obtain the trained signal classification model.

4. The method according to claim 3, characterized in that, Before training the sample signal classification model based on the benchmark loss to obtain the trained signal classification model, the method further includes: Positive and negative samples are determined from the sample training data based on the sample signal type. For each sample signal block of a negative sample, the block-level loss of the sample signal block is determined based on the sample local confidence of the sample signal block for each sample signal type and the sample signal type to which the sample signal block belongs. For the negative sample, the sample-level loss of the negative sample is determined based on the block-level loss of each sample signal block in the negative sample; The step of training the sample signal classification model based on the benchmark loss to obtain the trained signal classification model includes: When the training data is positive, the sample signal classification model is trained according to the baseline loss to obtain the trained signal classification model. When the training data is negative, the sample signal classification model is trained based on the baseline loss and the sample-level loss to obtain the trained signal classification model.

5. The method according to claim 4, characterized in that, The fusion network includes a decision layer and a fusion layer. The step of inputting the local confidence scores of each sample into the fusion network of the sample signal classification model to obtain the comprehensive confidence score of the sample dynamic spectrum signal for each sample signal type includes: The local confidence scores of each sample are input into the decision layer to obtain the sample fusion weights corresponding to the local confidence scores of each sample. The local confidence scores of each sample and the fusion weights of each sample are input into the fusion layer to obtain the comprehensive confidence score of the sample dynamic spectrum signal for each sample signal type.

6. The method according to claim 5, characterized in that, The step of training the sample signal classification model based on the sample comprehensive confidence level and the sample signal type corresponding to the sample dynamic spectrum signal to obtain the trained signal classification model includes: The fusion loss is determined based on the entropy of the distribution of the fusion weights of each sample. The baseline loss is determined based on the overall confidence level of the samples and the sample signal type of the dynamic spectrum signal of the samples; The sample signal classification model is trained based on the baseline loss and the fusion loss to obtain the trained signal classification model.

7. A dynamic spectrum signal identification device, characterized in that, include: The acquisition module is used to acquire the dynamic spectrum signal to be identified, wherein the dynamic spectrum signal has a time dimension and a frequency dimension; The processing module is used to segment the dynamic spectrum signal along the time dimension by a preset size threshold to obtain several signal blocks with the same frequency dimension; An analysis module is used to input the signal blocks into a signal classification model to obtain the confidence level of the dynamic spectrum signal for the single-pulse signal type. The signal classification model includes a discriminant network and a fusion network. The discriminant network includes a convolutional layer and a discrimination layer, comprising: an extraction subunit for inputting the signal block into the convolutional layer to obtain block signal feature information corresponding to the signal block; a discrimination subunit for inputting the block signal feature information into the discrimination layer to obtain the local confidence level of the signal block for the single-pulse signal type; a fusion unit for inputting each local confidence level into the fusion network to obtain a comprehensive confidence level; and a determination unit for using the comprehensive confidence level as the confidence level of the dynamic spectrum signal for the single-pulse signal type. The determination module is used to determine the single-pulse signal type as the signal type of the dynamic spectrum signal when the confidence level reaches a preset threshold.

8. An electronic device, characterized in that, The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the dynamic spectrum signal recognition method as described in any one of claims 1-6.

9. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the dynamic spectrum signal recognition method as described in any one of claims 1-6.

10. A computer program product comprising computer instructions stored in a computer-readable storage medium, wherein a processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to cause the computer device to perform the method for dynamic spectrum signal identification as described in any one of claims 1-6.

Images