CN116796826A - Pulsar search network training method, pulsar search device and pulsar search equipment - Google Patents

Abstract

The invention provides a pulsar search network training method, which includes: based on a training sample set and the pulse profile image of the candidate body, performing first training on the backbone network of the pulsar search network, and determining the initial network parameters of the backbone network; Based on the training sample set and the above-mentioned and pulsar label values, the backbone network of the pulsar search network is secondly trained to obtain the target parameters of the backbone network of the pulsar search network. This can reduce the total amount of training data and eliminate the need for Under the premise of re-implementing supervised training, the accuracy of pulsar search network training can be steadily improved, the over-fitting of the pulsar search network can be reduced, and the generalization ability of the pulsar search network model can be enhanced. The present invention also provides a pulsar search method, device, electronic equipment, computer program product and storage medium. Embodiments of the present invention can be applied to various scenarios such as cloud technology, artificial intelligence, smart transportation, and assisted driving.

Description

Pulsar search network training method, pulsar search method, device and equipment

Technical field

The present invention relates to image processing technology in the field of artificial intelligence, and in particular to pulsar search network training methods, pulsar search methods, devices, electronic equipment, computer program products and storage media.

Background technique

Category recognition based on deep learning has always been an important tool for solving large amounts of data in various application scenarios. For example, in application scenarios such as images and natural language processing, large-scale classification and recognition of large amounts of data can be achieved to quickly and accurately obtain relevant classification prediction results and accelerate the implementation of functions in the application scenarios.

In the classification and prediction of images, different tasks can be performed according to the deployed application scenarios. When performing pulsar searches, a large number of candidate files are usually generated in pulsar search tasks. For example, FAST will There will be annual increments of the order of PB, and it is expected that tens of millions of candidates will be generated. Therefore, in the search for pulsar candidates, the accuracy of the search model and the inference speed of the search model will be crucial.

Contents of the invention

In view of this, embodiments of the present invention provide a pulsar search network training method, pulsar search method device, electronic equipment, computer program products and storage media, which can achieve training of the pulsar search network and reduce training data. Under the premise of total amount and no need to re-supervised training, it can steadily improve the accuracy of pulsar search network training, reduce the over-fitting of pulsar search network, enhance the generalization ability of pulsar search network model, and facilitate the timely training of pulsar search network. The pulsar search network model is deployed in network terminals to realize large-scale application of the pulsar search network model.

The technical solution of the embodiment of the present invention is implemented as follows:

Embodiments of the present invention provide a pulsar search network training method, which includes:

Obtain the time phase distribution image and frequency phase distribution image corresponding to the pulsar search candidate signal;

The time phase distribution image and the frequency phase distribution image are combined to form a training sample set for the pulsar search network;

Obtain the pulse profile image and pulsar label value of the candidate body corresponding to the obtained pulsar search candidate body signal;

Based on the training sample set and the pulse profile image of the candidate body, perform a first training on the backbone network of the pulsar search network, and determine the initial network parameters of the backbone network;

Based on the training sample set and the sum pulsar label value, perform a second training on the backbone network of the pulsar search network to adjust the initial parameters of the backbone network of the pulsar search network to obtain the pulsar search network. Target parameters of the backbone network of the star search network.

Embodiments of the present invention also provide a pulsar search method, including:

Receive unlabeled pulsar search candidate signals;

Obtaining a time phase distribution image and a frequency phase distribution image corresponding to the pulsar search candidate signal;

Process the time phase distribution image and frequency phase distribution image through the backbone network of the pulsar search network to determine the image processing results;

Through the classification task fully connected layer network of the pulsar search network, the image processing results are processed to obtain the prediction results of the pulsar search candidate signals.

An embodiment of the present invention also provides a pulsar search network training device, which includes:

The information transmission module is used to obtain the time phase distribution image and frequency phase distribution image corresponding to the pulsar search candidate signal;

An information processing module, configured to combine the time phase distribution image and the frequency phase distribution image to form a training sample set for the pulsar search network;

The information processing module is used to obtain the pulse profile image and pulsar label value of the candidate corresponding to the obtained pulsar search candidate signal;

The information processing module is configured to perform first training on the backbone network of the pulsar search network based on the training sample set and the pulse profile image of the candidate body, and determine the initial network parameters of the backbone network;

The information processing module is configured to perform a second training on the backbone network of the pulsar search network based on the training sample set and the sum pulsar label value, so as to adjust the backbone network of the pulsar search network. The initial parameters are used to obtain the target parameters of the backbone network of the pulsar search network.

In the above scheme,

The information processing module is used to determine the noise value of the time phase distribution image and the frequency phase distribution image, and perform denoising processing on the determined time phase distribution image and frequency phase distribution image respectively;

The information processing module is used to standardize the time phase distribution image and frequency phase distribution image that have been denoised, respectively, to obtain a time phase distribution image and a frequency phase distribution image that comply with single-channel grayscale image standards;

The information processing module is used to combine time phase distribution images and frequency phase distribution images that comply with single-channel grayscale image standards to form a training sample set for the pulsar search network.

In the above scheme,

The information processing module is configured to determine a dynamic noise threshold that matches the usage environment of the pulsar search network based on the position of the target area corresponding to the pulsar search candidate signal;

The information processing module is configured to perform noise reduction processing on the time phase distribution image and the frequency phase distribution image according to the dynamic noise threshold to form a time phase distribution image and frequency that match the dynamic noise threshold. Phase distribution image.

In the above scheme,

The information processing module is configured to determine a fixed noise threshold that matches the usage environment of the pulsar search network according to the radio telescope type corresponding to the pulsar search candidate signal;

The information processing module is configured to perform noise reduction processing on the time phase distribution image and the frequency phase distribution image according to the fixed noise threshold to form a time phase distribution image and frequency that match the fixed noise threshold. Phase distribution image.

In the above scheme,

The information processing module is used to obtain the original pulse profile image of the candidate body corresponding to the pulsar search candidate body signal;

The information processing module is used to standardize the sequence values of the original pulse profile image to obtain the standard sequence values of the original pulse profile image;

The information processing module is used to adjust the standard sequence value of the original pulse profile image to obtain the pulse profile image corresponding to the pulsar search candidate signal, so as to match the accuracy of the pulsar search network. .

In the above scheme,

The information processing module is configured to determine the label value of the pulse profile image of the candidate body based on the pulse profile image of the candidate body;

The information processing module is configured to determine the first loss function corresponding to the first training based on the label value of the pulse profile image through the pre-trained fully connected layer network corresponding to the backbone network;

The information processing module is used to bring the training samples in the training sample set into the first loss function;

The information processing module is configured to determine initial network parameters of the backbone network when the first loss function satisfies the first convergence condition.

In the above scheme,

The information processing module is configured to determine the second loss function corresponding to the second training based on the pulsar label value through the fully connected layer network of the classification task corresponding to the backbone network;

The information processing module is used to bring the training samples in the training sample set into the second loss function;

The information processing module is configured to determine the update parameters of the backbone network when the second loss function satisfies the second convergence condition;

The information processing module is configured to adjust the initial parameters of the backbone network of the pulsar search network based on the updated parameters of the backbone network to obtain target parameters of the backbone network of the pulsar search network.

An embodiment of the present invention also provides a pulsar search device, including:

The signal transmission module is used to receive unlabeled pulsar search candidate signals;

A signal processing module, used to obtain the time phase distribution image and frequency phase distribution image corresponding to the pulsar search candidate signal;

The signal processing module is used to process the time phase distribution image and frequency phase distribution image through the backbone network of the pulsar search network, and determine the image processing results;

The signal processing module is used to process the image processing results through the fully connected layer network of the classification task of the pulsar search network to obtain the prediction result of the pulsar search candidate signal.

An embodiment of the present invention also provides an electronic device. The electronic device includes:

Memory, used to store executable instructions;

The processor is configured to implement the aforementioned pulsar search network training method when running executable instructions stored in the memory.

An embodiment of the present invention also provides a computer-readable storage medium storing executable instructions, which is characterized in that when the executable instructions are executed by a processor, the aforementioned pulsar search network training method is implemented.

The embodiments of the present invention have the following beneficial effects:

The present invention obtains the time phase distribution image and frequency phase distribution image corresponding to the pulsar search candidate signal; combines the time phase distribution image and the frequency phase distribution image to form a training sample set of the pulsar search network; obtains Obtaining the pulse profile image and pulsar label value of the candidate body corresponding to the pulsar search candidate body signal; based on the training sample set and the pulse profile image of the candidate body, conducting the backbone network of the pulsar search network The first training is to determine the initial network parameters of the backbone network; based on the training sample set and the sum pulsar label value, the second training is performed on the backbone network of the pulsar search network to adjust the pulse The initial parameters of the backbone network of the star search network are obtained to obtain the target parameters of the backbone network of the pulsar search network. Therefore, it is possible to reduce the total amount of training data and eliminate the need for re-supervision by training the pulsar search network. Under the premise of training, it can steadily improve the accuracy of pulsar search network training, reduce over-fitting of pulsar search network, enhance the generalization ability of pulsar search network model, and facilitate the deployment of timely trained pulsar search network model. In network terminals, large-scale application of the pulsar search network model is realized.

Description of the drawings

Figure 1 is a schematic diagram of the usage environment of the pulsar search network training method provided by the embodiment of the present invention;

Figure 2 is a schematic structural diagram of an electronic device provided by an embodiment of the present invention;

Figure 3 is an optional flow diagram of a pulsar search network training method provided by an embodiment of the present invention;

Figure 4A is a schematic diagram of the sample set of the pulsar search network in the embodiment of the present invention;

Figure 4B is a schematic diagram of the sample set of the pulsar search network in the embodiment of the present invention;

Figure 5 is an optional flow diagram of the pulsar search network training method provided by the embodiment of the present invention;

Figure 6 is a schematic network structure diagram of the pulsar search network in the embodiment of the present invention;

Figure 7 is an optional flow diagram of the pulsar search method provided by the embodiment of the present invention;

Figure 8 is a schematic diagram of pulsar search candidate signals in an embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be described in further detail below in conjunction with the accompanying drawings. The described embodiments should not be regarded as limiting the present invention. Those of ordinary skill in the art will not make any All other embodiments obtained under the premise of creative work belong to the scope of protection of the present invention.

In the following description, references to "some embodiments" describe a subset of all possible embodiments, but it is understood that "some embodiments" may be the same subset or a different subset of all possible embodiments, and Can be combined with each other without conflict.

Before further describing the embodiments of the present invention in detail, nouns and terms involved in the embodiments of the present invention will be described. The nouns and terms involved in the embodiments of the present invention are suitable for the following explanations.

1) Response is used to represent the conditions or states on which the performed operations depend. When the dependent conditions or states are met, the one or more operations performed may be in real time or may have a set delay; Unless otherwise specified, there is no restriction on the execution order of the multiple operations performed.

2) Downsampling processing, a sample sequence is sampled at intervals of several samples, so that the new sequence obtained is the downsampling of the original sequence. For example: for an image whose I size is M*N, downsample it by s times. , that is, a resolution image of (M/s)*(N/s) size is obtained, where s should be the common divisor of M and N

3) Model training, performing multi-classification learning on image data sets. The model can be constructed using deep learning frameworks such as TensorFlow and torch, and uses multi-layer combinations of neural network layers such as CNN to form a multi-classification model. The input of the model is the three-channel or original channel matrix formed by reading the image through tools such as openCV. The output of the model is multi-classification probability, and the image category is finally output through algorithms such as softmax. During training, the model approaches the correct trend through objective functions such as cross-entropy.

4) Noisy recognition: Training of image recognition tasks based on noise samples. Noisy samples include samples with wrong category labels and samples with inaccurate category labels. For example, images do not completely correspond to category labels, and the concept of two category labels With partial overlap, an image has properties of both of the above class labels, but has only one class label.

5) Contrastive loss: Contrastive loss function, which can learn a mapping relationship, which can make points of the same category but far apart in a high-dimensional space become closer after being mapped to a low-dimensional space through a function , points of different categories but close together become further apart in the low-dimensional space after mapping. The result is that in a low-dimensional space, points of the same type will produce a clustering effect, and means of different types will be separated. Similar to fisher dimensionality reduction, but fisher dimensionality reduction does not have the effect of out-of-sample extension and cannot affect new samples.

6) Client, a carrier that implements specific functions in the terminal. For example, a mobile client (APP) is a carrier of specific functions in the mobile terminal, such as a program that performs the function of user gesture recognition.

7)Soft max: Normalized exponential function, which is a generalization of the logical function. It can "compress" a K-dimensional vector containing any real number into another K-dimensional real vector, so that each element ranges between [0, 1], and the sum of all elements is 1.

Figure 1 is a schematic diagram of the usage scenario of the pulsar search network training method provided by the embodiment of the present invention. Referring to Figure 1, terminals (including terminal 10-1 and terminal 10-2) are provided with corresponding clients capable of performing different functions. The server 200 stores the pulse signals received by the radio telescope 400. The client is a terminal (including terminal 10-1 and terminal 10-2) that obtains different pulsar candidate signals from the corresponding server 200 through the network 300 for browsing, or obtains the corresponding For the identification result of the pulsar candidate signal, the terminal connects to the server 200 through the network 300. The network 300 can be a wide area network or a local area network, or a combination of the two, using wireless links to realize data transmission, in which the terminal (including the terminal 10-1 The corresponding pulsar candidate signals obtained from the corresponding server 200 through the network 300 and the terminal 10-2) may be the same or different. For example, the terminals (including the terminal 10-1 and the terminal 10-2) may either The network 300 obtains a pulsar candidate signal with annotation information related to a certain known pulsar from the corresponding server 200. It can also obtain a pulsar candidate signal without annotation information from the corresponding server 200 through the network 300. It can also obtain The final identification result of the pulsar candidate signal without annotation information (for unknown pulsars). The server 200 may store pulsar candidate signals without annotation information corresponding to different target objects and corresponding recognition results. In some embodiments of the present invention, the pulsar candidate signals without annotation information stored in the server 200 are collected by the radio telescope 400 . At the same time, the server 200 may also store pulse signals of known pulsars, for example: (1) HTRU: High Time Resolution Universe Survey, the data set contains 1196 known pulsars and There are 89996 non-pulsars, and most of the positive samples are relatively strong pulsar signals. The number of negative samples is much more than the positive samples, which belongs to the extremely imbalanced classification task. (2) PMPS, PMPS-26k, P309: Parkes Multi-beam PulsarSurvey. PMPS has a total of 4.25T, most of which are preliminary screening samples. PMPS-26k contains 2,000 pulsar samples and 20,000 radio interference samples. There are 2000 other signals and 2000 unlabeled samples, totaling 26k. P309 contains 2698 pulsars and 1656 radio interferences. The data set generally contains a wide variety of samples, and the positive samples include both samples with strong signals and samples with weak signals. Taking the new generation 500m aperture spherical radio telescope (FAST Five Hundred Meter Aperture Spherical Telescope) as an example, the data stored in the server contains 1163 pulsars (837 available training sets) and 14319 non-pulse signals (998 available training sets).

The pulsar search network trained by the pulsar search network training method provided in this application can realize the search for pulsars using unlabeled pulsar search candidate signals. Of course, before use, the pulsar search network needs to be first Training, specifically including:

Obtain the time phase distribution image and frequency phase distribution image corresponding to the pulsar search candidate signal; combine the time phase distribution image and the frequency phase distribution image to form a training sample set of the pulsar search network; obtain the acquisition Pulsar search candidate body's pulse profile image and pulsar label value corresponding to the candidate body signal; based on the training sample set and the pulse profile image of the candidate body, perform first training on the backbone network of the pulsar search network , determine the initial network parameters of the backbone network; based on the training sample set and the sum pulsar label value, perform a second training on the backbone network of the pulsar search network to adjust the pulsar search network The initial parameters of the backbone network are used to obtain the target parameters of the backbone network of the pulsar search network.

In order to better understand the methods provided by the embodiments of the present application, artificial intelligence, various branches of artificial intelligence, and the application fields, cloud technologies and artificial intelligence cloud services involved in the methods provided by the embodiments of the present application are first described.

Artificial Intelligence (AI) is a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is the study of the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making.

Artificial intelligence technology is a comprehensive subject that covers a wide range of fields, including both hardware-level technology and software-level technology. Basic artificial intelligence technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation/interaction systems, mechatronics and other technologies. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning/deep learning. Each direction is explained below.

Natural language processing (NLP, Nature Language processing) is an important direction in the fields of computer science and artificial intelligence. It studies various theories and methods that enable effective communication between humans and computers using natural language. Natural language processing is a science that integrates linguistics, computer science, and mathematics. Therefore, research in this field will involve natural language, that is, the language that people use every day, so it is closely related to the study of linguistics. Natural language processing technology usually includes text processing, semantic understanding, machine translation, robot question answering, knowledge graph and other technologies.

Machine Learning (ML) is a multi-field interdisciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying 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 make computers intelligent. Its applications cover all fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning and other technologies.

Cloud technology refers to a hosting technology that unifies a series of resources such as hardware, software, and networks within a wide area network or local area network to realize data calculation, storage, processing, and sharing. Cloud technology is a general term for network technology, information technology, integration technology, management platform technology, application technology, etc. based on the cloud computing business model. It can form a resource pool and use it on demand, which is flexible and convenient. Cloud computing technology will become an important support. The background services of technical network systems require a large amount of computing and storage resources, such as video websites, picture websites and more portal websites. With the rapid development and application of the Internet industry, in the future each item may have its own identification mark, which needs to be transmitted to the backend system for logical processing. Data at different levels will be processed separately, and all types of industry data need to be powerful. System backing support can only be achieved through cloud computing.

The so-called artificial intelligence cloud service, also generally called AIaaS (AI as a Service), is currently a mainstream artificial intelligence platform service method. Specifically, the AIaaS platform will provide several common AI services. Split it up and provide standalone or packaged services in the cloud. This service model is similar to opening an AI theme mall: all developers can access one or more artificial intelligence services provided by the platform through API interfaces, and some senior developers can also use the platform. Provide AI framework and AI infrastructure to deploy and operate your own exclusive cloud artificial intelligence services.

The solutions provided by the embodiments of this application involve artificial intelligence natural language processing, machine learning, artificial intelligence cloud services and other technologies, which are specifically explained through the following embodiments.

The pulsar search network training method provided by the embodiment of the present application will be described with reference to the exemplary application and implementation of the terminal provided by the embodiment of the present application.

The structure of the electronic device according to the embodiment of the present invention is described in detail below. The electronic device can be implemented in various forms, such as a dedicated terminal with an endoscopic image processing function, or an electronic device with an endoscopic image processing function. Or a cloud server, such as the server 200 in Figure 1 mentioned above. Figure 2 is a schematic structural diagram of an electronic device provided by an embodiment of the present invention. It can be understood that Figure 2 only shows an exemplary structure of the electronic device rather than the entire structure. Part or all of the structure shown in Figure 2 can be implemented as needed. .

The electronic device provided by the embodiment of the present invention includes: at least one processor 201, a memory 202, a user interface 203, and at least one network interface 204. The various components in the electronic device are coupled together through bus system 205 . It can be understood that the bus system 205 is used to implement connection communication between these components. In addition to the data bus, the bus system 205 also includes a power bus, a control bus and a status signal bus. However, for the sake of clarity, the various buses are labeled as bus system 205 in FIG. 2 .

The user interface 203 may include a display, keyboard, mouse, trackball, click wheel, keys, buttons, touch pad or touch screen, etc. Among them, terminals in embodiments of the present invention include but are not limited to mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle-mounted terminals, etc. Embodiments of the present invention can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, smart transportation, assisted driving, etc. When the pulsar search network training method provided by the present invention is executed through different terminals, the specific usage scenarios are as follows: Invention without limits

It can be understood that the memory 202 can be a volatile memory or a non-volatile memory, and can also include both volatile and non-volatile memories. The memory 202 in the embodiment of the present invention can store data to support the operation of the terminal (such as 10-1). Examples of such data include: any computer programs, such as operating systems and applications, used to operate on a terminal such as 10-1. Among them, the operating system includes various system programs, such as framework layer, core library layer, driver layer, etc., which are used to implement various basic services and process hardware-based tasks. Applications can contain various applications.

In some embodiments, the pulsar search network training device provided by the embodiment of the present invention can be implemented by combining software and hardware. As an example, the pulsar search network training device provided by the embodiment of the present invention can be implemented by using a hardware decoding processor. A processor in the form of a processor, which is programmed to execute the pulsar search network training method provided by the embodiment of the present invention. For example, a processor in the form of a hardware decoding processor can use one or more Application Specific Integrated Circuits (ASICs), DSPs, Programmable Logic Devices (PLDs), Complex Programmable Logic Devices (CPLDs) , Complex Programmable Logic Device), Field-Programmable Gate Array (FPGA, Field-Programmable Gate Array) or other electronic components.

As an example of the pulsar search network training device provided by the embodiment of the present invention using a combination of software and hardware, the pulsar search network training device provided by the embodiment of the present invention can be directly embodied as a combination of software modules executed by the processor 201. The software modules It can be located in a storage medium, and the storage medium is located in the memory 202. The processor 201 reads the executable instructions included in the software module in the memory 202, and completes the present invention in combination with the necessary hardware (for example, including the processor 201 and other components connected to the bus 205). The pulsar search network training method provided by the embodiment of the invention.

As an example, the processor 201 may be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gates, or transistor logic devices. , discrete hardware components, etc., wherein the general processor can be a microprocessor or any conventional processor, etc.

As an example of the pulsar search network training device provided by the embodiment of the present invention being implemented in hardware, the device provided by the embodiment of the present invention can be directly executed by a processor 201 in the form of a hardware decoding processor, for example, by one or more Application Specific Integrated Circuit (ASIC, Application Specific Integrated Circuit), DSP, Programmable Logic Device (PLD, Programmable Logic Device), Complex Programmable Logic Device (CPLD, Complex Programmable Logic Device), Field Programmable Gate Array (FPGA, Field -Programmable GateArray) or other electronic components to implement the pulsar search network training method provided by the embodiment of the present invention.

The memory 202 in the embodiment of the present invention is used to store various types of data to support the operation of the electronic device. Examples of these data include: any executable instructions for operating on an electronic device, such as executable instructions, and a program that implements the pulsar search network training method of an embodiment of the present invention may be included in the executable instructions.

In other embodiments, the pulsar search network training device provided by the embodiment of the present invention can be implemented in software. Figure 2 shows the pulsar search network training device 2020 stored in the memory 202, which can be a program and a plug-in. and other forms of software, and includes a series of modules. As an example of a program stored in the memory 202, it may include a pulsar search network training device 2020. The pulsar search network training device 2020 includes the following software modules:

The information transmission module 2081 is used to obtain the time phase distribution image and frequency phase distribution image corresponding to the pulsar search candidate signal.

The information processing module 2082 is used to combine the time phase distribution image and the frequency phase distribution image to form a training sample set for the pulsar search network.

The information processing module 2082 is used to obtain the pulse profile image and pulsar label value of the candidate corresponding to the obtained pulsar search candidate signal.

The information processing module 2082 is configured to perform first training on the backbone network of the pulsar search network based on the training sample set and the pulse profile image of the candidate body, and determine the initial network parameters of the backbone network.

The information processing module 2082 is configured to perform a second training on the backbone network of the pulsar search network based on the training sample set and the sum pulsar label value, so as to adjust the backbone network of the pulsar search network The initial parameters of the network are used to obtain the target parameters of the backbone network of the pulsar search network.

It should be noted that after the pulsar search network training is completed, it can also be deployed in electronic equipment (such as a star finding server or a data processing server of a radio telescope) to process the received pulsar search candidate signals. Therefore, electronic equipment can also include:

The signal transmission module is used to receive unlabeled pulsar search candidate signals.

A signal processing module, configured to obtain a time phase distribution image and a frequency phase distribution image corresponding to the pulsar search candidate signal.

The signal processing module is used to process the time phase distribution image and frequency phase distribution image through the backbone network of the pulsar search network, and determine the image processing results.

The signal processing module is used to process the image processing results through the fully connected layer network of the classification task of the pulsar search network to obtain the prediction result of the pulsar search candidate signal.

According to the electronic device shown in Figure 2, in one aspect of the present application, the present application also provides a computer program product or computer program. The computer program product or computer program includes computer instructions, and the computer instructions are stored in a computer-readable file. in the storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the various optional implementations of the pulsar search network training method provided by this application. Methods.

The pulsar search network training method provided by the embodiment of the present invention will be described with reference to the pulsar search network training device shown in Figure 2. Refer to Figure 3, which is an optional process of the pulsar search network training method provided by the embodiment of the present invention. Schematic diagram, it can be understood that the steps shown in Figure 3 can be executed by various electronic devices running the pulsar search network training device, for example, it can be a dedicated terminal, server or server cluster with a pulsar search function. The steps shown in Figure 3 will be described below.

Step 301: The pulsar search network training device obtains the time phase distribution image and frequency phase distribution image corresponding to the pulsar search candidate signal.

Step 302: The pulsar search network training device combines the time phase distribution image and the frequency phase distribution image to form a training sample set of the pulsar search network.

Referring to Figure 4A, Figure 4A is a schematic diagram of a sample set of the pulsar search network in an embodiment of the present invention, in which the time phase distribution image and the frequency phase distribution image are combined to form a training sample of the pulsar search network Collections can be implemented in the following ways:

Determine the noise values of the time phase distribution image and frequency phase distribution image, perform denoising processing on the determined time phase distribution image and frequency phase distribution image respectively; perform denoising processing on the time phase distribution image and The frequency phase distribution images are standardized separately to obtain a time phase distribution image and a frequency phase distribution image that meet the single-channel grayscale image standard; the time phase distribution image and frequency phase distribution image that meet the single-channel grayscale image standard are combined to form A collection of training samples for the pulsar search network. Taking the radio telescope shown in Figure 1 as an example, after feature reading of the fid format file obtained by receiving pulsar signals, the time-phase distribution image resolution is 256*64, and the frequency-phase distribution image resolution is 96*64; After feature extraction, the phcx format files obtained by other radio telescopes generally have a time-phase distribution image resolution of (16~24)*64 and a frequency-phase distribution image resolution of 16*64. The sizes of the above images are not equal. , in addition, the characteristic background distinction between the phase distribution image and the frequency phase distribution image is also different. For example, in the phcx format file image, the pulse characteristics are black and the background is white, while in the fid format file image, the pulse characteristics are white and the background is black, so it needs to be Denoising and normalizing to facilitate training and use of pulsar search networks.

In some embodiments of the present invention, referring to Figure 4B, which is a schematic diagram of the sample set of the pulsar search network in the embodiment of the present invention, the time-phase distribution diagram and the frequency-phase distribution diagram can be subjected to data standardization operations respectively, to obtain The corresponding 64x64 single-channel grayscale image is then stacked in the channel dimension into a 2-channel 2-dimensional signal as the input feature vector of the backbone network. In pre-training tasks and classification tasks, the output results of the backbone network are used to adjust the output of the backbone network to the output dimension that can calculate the task loss through different fully connected layers.

In some embodiments of the present invention, the noise values of the time phase distribution image and the frequency phase distribution image are determined, and the determined time phase distribution image and the frequency phase distribution image are separately denoised in the following manner. accomplish:

According to the position of the target area corresponding to the pulsar search candidate signal, a dynamic noise threshold matching the usage environment of the pulsar search network is determined; according to the dynamic noise threshold, the time phase distribution image and the The frequency phase distribution image is subjected to noise reduction processing to form a time phase distribution image and a frequency phase distribution image that match the dynamic noise threshold. Among them, when detecting pulsars in the low-frequency radio band (for example, below 350MHz), since the radiation spectrum appears as a power law spectrum above a very low frequency lower limit, it is expected that there should be a higher flux density relative to high frequency. Therefore, it is easier to detect, but because the galaxies in which pulsars appear are different, and because the interstellar medium has different refractive index (i.e., dispersion) for radio waves of different frequencies, pulsar pulses of different frequencies aligned at the signal emission site, After long-distance propagation, different frequency components will be staggered from each other, and the very obvious periodic pulse profile in the time domain will become almost impossible to directly identify and detect through the dispersion effect. Therefore, detecting pulsars in the low-frequency radio band requires a dynamic noise threshold to denoise the time phase distribution image and the frequency phase distribution image to form a time phase distribution image and frequency phase that match the dynamic noise threshold. Distribute images to ensure the accuracy of pulsar search networks.

In some embodiments of the present invention, the noise values of the time phase distribution image and the frequency phase distribution image are determined, and the determined time phase distribution image and the frequency phase distribution image are separately denoised in the following manner. accomplish:

According to the type of radio telescope corresponding to the pulsar search candidate signal, a fixed noise threshold matching the usage environment of the pulsar search network is determined; according to the fixed noise threshold, the time phase distribution image and the The frequency phase distribution image undergoes noise reduction processing to form a time phase distribution image and a frequency phase distribution image that match the fixed noise threshold. Among them, because the number of pulsars that have been detected and can be used to adjust the parameters of the detection algorithm is very limited, and unknown pulsars cannot be used as a correction training sample set before they are actually searched. Therefore, for a brand new radio telescope or an existing radio telescope that has never detected a pulsar, the time phase distribution image and the frequency phase distribution image can be reduced according to the dynamic noise threshold. Noise processing is performed to form a time phase distribution image and a frequency phase distribution image that match the dynamic noise threshold, and a noise threshold that matches the radio telescope is determined.

Further, when the pulsar search network is fixed in the terminal equipment of the radio telescope, the time phase distribution image and the frequency phase distribution image are denoised according to the fixed noise threshold to form a pixel that matches the fixed noise threshold. The time phase distribution image and frequency phase distribution image can continue to reduce the training time of the pulsar search network and improve the training efficiency of the pulsar search network.

Step 303: The pulsar search network training device obtains the pulse profile image and pulsar label value of the candidate corresponding to the acquired pulsar search candidate signal.

In some embodiments of the present invention, obtaining the pulse profile image of the candidate body corresponding to the pulsar search candidate body signal can be achieved in the following manner:

Obtain the original pulse profile image of the candidate body corresponding to the pulsar search candidate body signal; perform standardization processing on the sequence value of the original pulse profile image to obtain the standard sequence value of the original pulse profile image; The standard sequence value of the profile image is adjusted to obtain a pulse profile image corresponding to the pulsar search candidate signal, so as to match the accuracy of the pulsar search network. Among them, in order to increase the difficulty of the pre-training task, the original pulse profile sequence can be data standardized, and then the 64-dimensional sequence is linearly interpolated to a 100-dimensional sequence as the label value of the pre-training task. When the training accuracy increases, different parameters can be flexibly adjusted. Dimension sequences are linearly interpolated to adapt to different needs of pulsar search networks.

Step 304: The pulsar search network training device performs first training on the backbone network of the pulsar search network based on the training sample set and the pulse profile image of the candidate body, and determines the initial network parameters of the backbone network.

Step 305: The pulsar search network training device performs a second training on the backbone network of the pulsar search network based on the training sample set and the sum pulsar label value to adjust the backbone network of the pulsar search network The initial parameters of the network are used to obtain the target parameters of the backbone network of the pulsar search network.

Continuing to describe the pulsar search network training method provided by the embodiment of the present invention with reference to the pulsar search network training device shown in Figure 2, see Figure 5. Figure 5 is an optional version of the pulsar search network training method provided by the embodiment of the present invention. It is understandable that the steps shown in Figure 5 can be executed by various electronic devices running the pulsar search network training device, for example, a dedicated terminal, a server or a server cluster with a pulsar search function. The steps shown in Figure 5 will be described below.

Step 501: The pulsar search network training device determines the label value of the pulse profile image of the candidate body based on the pulse profile image of the candidate body.

Step 502: The pulsar search network training device determines the first loss function corresponding to the first training based on the label value of the pulse profile image through the pre-trained fully connected layer network corresponding to the backbone network.

Step 503: The pulsar search network training device brings the training samples in the training sample set into the first loss function.

Step 504: When the first loss function satisfies the first convergence condition, the pulsar search network training device determines the initial network parameters of the backbone network.

Referring to Figure 6, Figure 6 is a schematic network structure diagram of a pulsar search network in an embodiment of the present invention, in which, through the processing of steps 501 to 504, the time-phase distribution diagram and frequency-phase distribution diagram of the candidate body signal are used as the pulse For the input features of the star search network, the pulse profile of the candidate body is used as the label value, and the root mean square error is used as the loss function to calculate the initial network parameters of the pulsar search network. Through the backbone network, the time-phase distribution diagram and the frequency-phase distribution diagram can be summed in column dimensions, and then the two summed sequences are summed again to obtain a 64-dimensional original pulse profile sequence, where Figure 6 shows The backbone network in the network structure shown can be a ResNet18 network, and the network connected to the backbone network can include multiple fully connected layers. It is worth noting that in order to balance the space occupied by image features and image quantized coding, the dimension of the image feature can be set smaller than the dimension of the image quantized coding. Otherwise, excessive dimensions of the image feature will occupy too much storage space.

In the process of training the pulsar search network, the following steps are mainly involved.

1) Parameter initialization. For the initial feature extraction network, it can be trained through open source image sets (such as ImageNet image set, Open Image image set, etc.) to complete the initialization of weight parameters. For other networks, such as max pooling networks, embedded representation networks and quantized coding networks, the weight parameters can be randomly initialized. For example, a Gaussian distribution with a variance of 0.01 and a mean of 0 can be used for initialization.

2) Determine the weight parameters that need to be learned (that is, need to be updated). This can be divided into two training stages. In the first training stage, the weight parameters of the initial feature extraction network, maximum pooling network, and embedding representation network are updated; in the second training stage, the weight parameters of the quantized encoding network are updated. The weight parameters are updated.

3) Determine the learning rate. In this embodiment of the present application, each network in the pulsar search network can adopt the same learning rate, for example, the learning rate is 0.005. The learning rate can be updated to 0.1 times the original value after every 10 iterations.

4) Learning process (training process). Here, the pulsar search network can be iterated in epoch rounds. epoch is an integer greater than 1 and can be set according to the actual application scenario. In each round of iteration, a full amount of samples are processed once. For example, the full amount of similar image pairs can be divided according to the Batch Size (which can be set according to the actual application scenario) to obtain Nb batches. For each batch (Batch), execute The following processing:

①Forward propagation processing. Here, each training image in the batch is forward propagated through the pulsar search network to obtain image features and image quantization coding.

②Loss value calculation. Here, the feature extraction network and the quantization encoding network correspond to different loss values.

③Update weight parameters. For example, the stochastic gradient descent algorithm or other gradient descent algorithms can be used to backpropagate the loss value, and during the backpropagation process, the weight parameters of the network to be trained are updated along the gradient descent direction.

It should be noted that the specific structures in the above embodiments are only exemplary, and other model structures can be selected as needed. For example, the backbone network can also use resnet50, inceptionv4, resnet18, etc., which is not limited in this application.

After completing the pre-training task through steps 501 to 504, continue to step 505.

Step 505: The pulsar search network training device determines the second loss function corresponding to the second training based on the pulsar label value through the fully connected layer network of the classification task corresponding to the backbone network.

Step 506: The pulsar search network training device brings the training samples in the training sample set into the second loss function.

Step 507: When the second loss function satisfies the second convergence condition, the pulsar search network training device determines the update parameters of the backbone network.

Step 508: The pulsar search network training device adjusts the initial parameters of the backbone network of the pulsar search network based on the updated parameters of the backbone network, and obtains the target parameters of the backbone network of the pulsar search network.

Among them, through steps 505 to 508, the feature dimension is mapped to the classification dimension through the fully connected layer, and then the cross-entropy loss is calculated between the predicted value and the pulsar label value. In order to prevent the pre-trained backbone network from over-fitting during the optimization process of the classification task, the backbone network parameter learning rate can be set to 0.25 times the learning rate of the fully connected parameters of the classification task.

After the pulsar search network is trained through the previous embodiment, it can be deployed in the corresponding data processing equipment to predict the obtained pulsar search candidate signals. See Figure 7, which is provided by the embodiment of the present invention. An optional flow diagram of the pulsar search method. Understandably, the steps shown in Figure 7 can be executed by various electronic equipment running the pulsar search device, such as a dedicated terminal with a pulsar search function, a radio Telescope data server or server cluster. The steps shown in Figure 7 will be described below.

Step 701: The pulsar search device receives unlabeled pulsar search candidate signals.

Step 702: The pulsar search device acquires the time phase distribution image and frequency phase distribution image corresponding to the pulsar search candidate signal.

Referring to Figure 8, Figure 8 is a schematic diagram of the pulsar search candidate signal in an embodiment of the present invention. The time phase distribution image corresponding to the pulsar search candidate signal is as shown in (1), and the frequency phase distribution image is as (2). ), when non-numeric values (NaN Not a Number) appear in the time phase distribution image and frequency phase distribution image, all NaN values can be replaced, for example, the linear interpolation method is used to complete and replace the time phase distribution image and frequency phase distribution image. If non-numeric values appear in the data, if there is missing data, data filtering can also be performed to ensure the integrity of the non-numeric values in the time phase distribution image and frequency phase distribution image.

Step 703: The pulsar search device processes the time phase distribution image and frequency phase distribution image through the backbone network of the pulsar search network, and determines the image processing result.

Referring to Figure 8, taking the new generation 500m aperture spherical radio telescope (FAST Five Hundred Meter Aperture Spherical Telescope) as an example, as the pulsar search proceeds, a large number of pulsar candidate signals without annotation information will be generated. However, there is an inherent connection between the different modal signals of the pulsar. For example, the integrated profile of the pulsar shown in (3) in Figure 8 is the time-phase distribution diagram integrated over time or the frequency-phase distribution diagram. It is obtained by integrating the frequency. The inherent relationships between different signals in the pulsar search candidate signals can be determined through a trained pulsar search network. Among them, the backbone network of the pulsar search network can adopt network structures such as resnet50, inceptionv4, resnet18, etc., which is not limited in this application.

Step 704: The pulsar search device processes the image processing result through the fully connected layer network of the classification task of the pulsar search network to obtain the prediction result of the pulsar search candidate signal.

Among them, as shown in (4) in Figure 8 China, through the fully connected layer network of the classification task of the pulsar search network, the image processing results are processed to obtain the image processing results of the time phase distribution image and the frequency phase distribution image, which can be The discovery of a new pulsar was confirmed.

Beneficial technical effects:

The present invention obtains the time phase distribution image and frequency phase distribution image corresponding to the pulsar search candidate signal; combines the time phase distribution image and the frequency phase distribution image to form a training sample set of the pulsar search network; obtains Obtaining the pulse profile image and pulsar label value of the candidate body corresponding to the pulsar search candidate body signal; based on the training sample set and the pulse profile image of the candidate body, conducting the backbone network of the pulsar search network The first training is to determine the initial network parameters of the backbone network; based on the training sample set and the sum pulsar label value, the second training is performed on the backbone network of the pulsar search network to adjust the pulse The initial parameters of the backbone network of the star search network are obtained to obtain the target parameters of the backbone network of the pulsar search network. Therefore, it is possible to reduce the total amount of training data and eliminate the need for re-supervision by training the pulsar search network. Under the premise of training, it can steadily improve the accuracy of pulsar search network training, reduce over-fitting of pulsar search network, enhance the generalization ability of pulsar search network model, and facilitate the deployment of timely trained pulsar search network model. In network terminals, large-scale application of the pulsar search network model is realized.

The above are only examples of the present invention and are not used to limit the scope of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention shall be included in the present invention. within the scope of protection.

Claims (13)

1. A pulsar search network training method, the method comprising:

acquiring a time phase distribution image and a frequency phase distribution image corresponding to the pulsar search candidate signals;

combining the time phase distribution image and the frequency phase distribution image to form a training sample set of the pulsar search network;

acquiring a pulse contour image and a pulsar label value of a candidate corresponding to the acquired pulsar search candidate signal;

performing first training on a backbone network of the pulsar search network based on the training sample set and the pulse contour image of the candidate, and determining initial network parameters of the backbone network;

and performing second training on the backbone network of the pulsar search network based on the training sample set and the sum pulsar label value to realize adjustment of initial parameters of the backbone network of the pulsar search network and obtain target parameters of the backbone network of the pulsar search network.

2. The method of claim 1, wherein combining the time phase distribution image and frequency phase distribution image to form a training sample set for the pulsar search network comprises:

determining noise values of the time phase distribution image and the frequency phase distribution image, and respectively carrying out noise removal processing on the determined time phase distribution image and the determined frequency phase distribution image;

respectively carrying out standardization processing on the time phase distribution image and the frequency phase distribution image subjected to the denoising processing to obtain a time phase distribution image and a frequency phase distribution image which accord with the single-channel gray level image standard;

and combining the time phase distribution image and the frequency phase distribution image which accord with the single-channel gray level image standard to form a training sample set of the pulsar search network.

3. The method of claim 2, wherein said determining noise values of said time phase distribution image and said frequency phase distribution image, respectively, comprises:

determining a dynamic noise threshold matched with the use environment of the pulsar search network according to the position of the target area corresponding to the pulsar search candidate signal;

And carrying out noise reduction processing on the time phase distribution image and the frequency phase distribution image according to the dynamic noise threshold value so as to form a time phase distribution image and a frequency phase distribution image matched with the dynamic noise threshold value.

4. The method of claim 2, wherein said determining noise values of said time phase distribution image and said frequency phase distribution image, respectively, comprises:

determining a fixed noise threshold matched with the use environment of the pulsar search network according to the type of the radio telescope corresponding to the pulsar search candidate signal;

and carrying out noise reduction processing on the time phase distribution image and the frequency phase distribution image according to the fixed noise threshold value so as to form a time phase distribution image and a frequency phase distribution image matched with the fixed noise threshold value.

5. The method of claim 1, wherein the acquiring the pulse profile image of the candidate corresponding to the pulsar search candidate signal comprises:

acquiring an original pulse contour image of a candidate corresponding to the pulsar search candidate signal;

Performing standardization processing on the sequence value of the original pulse contour image to obtain the standard sequence value of the original pulse contour image;

and adjusting the standard sequence value of the original pulse contour image to obtain the pulse contour image corresponding to the pulsar search candidate signal so as to realize matching with the accuracy of the pulsar search network.

6. The method of claim 5, wherein the first training of the backbone network of the pulsar search network based on the set of training samples and the pulse profile image of the candidate comprises:

determining a label value of the pulse contour image of the candidate based on the pulse contour image of the candidate;

determining a first loss function corresponding to the first training based on the label value of the pulse contour image through a pre-training full-connection layer network corresponding to the backbone network;

bringing training samples in the training sample set into the first loss function;

and when the first loss function meets a first convergence condition, determining initial network parameters of the backbone network.

7. The method of claim 1, wherein the performing a second training on the backbone network of the pulsar search network based on the set of training samples and the sum pulsar label values to achieve adjusting initial parameters of the backbone network of the pulsar search network to obtain target parameters of the backbone network of the pulsar search network comprises:

determining a second loss function corresponding to the second training based on the pulsar label value through a classification task full-connection layer network corresponding to the backbone network;

bringing training samples in the training sample set into the second loss function;

when the second loss function meets a second convergence condition, determining updating parameters of the backbone network;

and based on the updated parameters of the backbone network, adjusting initial parameters of the backbone network of the pulsar search network to obtain target parameters of the backbone network of the pulsar search network.

8. A pulsar search method, said method comprising:

receiving a pulsar search candidate signal without labels;

acquiring a time phase distribution image and a frequency phase distribution image corresponding to the pulsar search candidate signals;

Processing the time phase distribution image and the frequency phase distribution image through a backbone network of a pulsar search network, and determining an image processing result;

and processing the image processing result through the classification task full-connection layer network of the pulsar search network to obtain a prediction result of the pulsar search candidate signal.

9. A pulsar search network training apparatus, said apparatus comprising:

the information transmission module is used for acquiring a time phase distribution image and a frequency phase distribution image corresponding to the pulsar search candidate signals;

the information processing module is used for combining the time phase distribution image and the frequency phase distribution image to form a training sample set of the pulsar search network;

the information processing module is used for acquiring the pulse contour image and the pulsar label value of the candidate corresponding to the acquired pulsar search candidate signal;

the information processing module is used for performing first training on a backbone network of the pulsar search network based on the training sample set and the pulse contour image of the candidate body, and determining initial network parameters of the backbone network;

The information processing module is used for performing second training on the main network of the pulsar search network based on the training sample set and the sum pulsar label value so as to adjust initial parameters of the main network of the pulsar search network and obtain target parameters of the main network of the pulsar search network.

10. A pulsar search device, said device comprising:

the signal transmission module is used for receiving the pulsar search candidate signals without labels;

the signal processing module is used for acquiring a time phase distribution image and a frequency phase distribution image corresponding to the pulsar search candidate signals;

the signal processing module is used for processing the time phase distribution image and the frequency phase distribution image through a backbone network of a pulsar search network and determining an image processing result;

and the signal processing module is used for processing the image processing result through the classification task full-connection layer network of the pulsar search network to obtain the prediction result of the pulsar search candidate signal.

11. An electronic device, the electronic device comprising:

A memory for storing executable instructions;

a processor configured to implement the pulsar search network training method of any one of claims 1 to 7, or implement the pulsar search method of claim 8, when executing the executable instructions stored in the memory.

12. A computer program product comprising a computer program or instructions which, when executed by a processor, implements the pulsar search network training method of any one of claims 1 to 7, or implements the pulsar search method of claim 8.

13. A computer readable storage medium storing executable instructions which when executed by a processor implement the pulsar search network training method of any one of claims 1 to 7, or implement the pulsar search method of claim 8.