Optical signal identification and model training method and device, electronic equipment and storage medium

Abstract

The invention discloses an optical signal identification and model training method and device, electronic equipment and a storage medium. The method comprises the steps that a to-be-recognized optical signal is acquired, and the to-be-recognized optical signal is optical signal data in a complex number form; the target complex numerical neural network model is adopted to recognize the optical signal to be recognized, signal parameter information corresponding to the optical signal to be recognized is obtained, and the signal parameter information comprises a modulation format and an optical signal-to-noise ratio; wherein the target complex numerical neural network model is obtained by training based on a finite memory quasi-Newton method, the finite memory quasi-Newton method is used for accelerating convergence of the model in a mode of updating an approximate matrix according to m pieces of historical iteration information, m is a memory scale, and m is an integer greater than or equal to 1. The memory scale is determined according to an error between a model prediction value and an actual value corresponding to different memory scales in a historical iteration process. The technical problem that an optical signal performance monitoring method in the prior art is too high in complexity and poor in precision is solved.

Description

Technical Field

[0001] The present application relates to the field of optical transmission technology, and more specifically, to an optical signal recognition and model training method, device, electronic device and storage medium. Background Art

[0002] Optical signal performance monitoring is an essential part of optical transmission links. It will directly affect the performance and reliability of the system. By identifying the signal modulation format and monitoring the optical signal-to-noise ratio, the selection of signal processing algorithms can be dynamically adjusted to ensure signal transmission efficiency, distance and quality. However, most of the optical signal performance monitoring methods in related technologies have technical problems such as high application complexity and poor accuracy.

[0003] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0004] The embodiments of the present application provide a method, device, electronic device and storage medium for optical signal recognition and model training, so as to at least solve the technical problems of high complexity and poor accuracy of optical signal performance monitoring methods in related technologies.

[0005] According to one aspect of an embodiment of the present application, a method for optical signal recognition is provided, including: obtaining an optical signal to be recognized, wherein the optical signal to be recognized is optical signal data in complex form; using a target complex-valued neural network model to recognize the optical signal to be recognized, and obtain signal parameter information corresponding to the optical signal to be recognized, wherein the signal parameter information includes: modulation format, optical signal-to-noise ratio; wherein the target complex-valued neural network model is trained based on a finite-memory quasi-Newton method, and the finite-memory quasi-Newton method is used to accelerate the convergence of the model by updating an approximate matrix based on m historical iteration information, where m is a memory scale, and during the model training process, each iteration corresponds to a memory scale, and the memory scale is determined based on the error between the model prediction value and the actual value corresponding to different memory scales in the historical iteration process.

[0006] Optionally, the target complex-valued neural network model is trained based on a target training data set, and the target training data set includes a first data set and a second data set, wherein the first data set includes: a plurality of data groups, and a first label corresponding to the data group, each data group includes a plurality of orthogonal amplitude modulation signal data in complex form corresponding to optical signals of the same modulation format, and the first label is used to characterize the modulation format corresponding to the orthogonal amplitude modulation signal data in the data group, and the modulation format includes at least one of the following: quadrature phase shift keying QPSK, sixteen-state orthogonal amplitude modulation 16QAM, sixty-four-state orthogonal amplitude modulation 64QAM, two hundred and fifty-six-state orthogonal amplitude modulation 256QAM and giga-state orthogonal amplitude modulation 1024QAM; the second data set includes: a plurality of data groups, and a second label corresponding to the data group, each data group includes a plurality of orthogonal amplitude modulation signal data in complex form corresponding to optical signals of the same optical signal-to-noise ratio, and the second label is used to characterize the optical signal-to-noise ratio corresponding to the orthogonal amplitude modulation signal data in the data group.

[0007] Optionally, the training steps of the target complex-valued neural network model include: obtaining an initial complex-valued neural network model; using the initial complex-valued neural network model to make predictions based on the data of the data group in the target training data set to obtain predicted values ​​of signal parameter information; determining a loss function value based on the predicted value of the signal parameter information and the true value of the signal parameter information represented by the first label and / or the second label in the target training data set, wherein the loss function value is used to characterize the error between the predicted value and the true value; using a finite memory quasi-Newton method to adjust the model parameters of the complex-valued neural network model based on the loss function value, and repeating the above training steps using the adjusted complex-valued neural network model until the loss function value is less than a preset error threshold, thereby obtaining the target complex-valued neural network model, wherein the model parameters include: an approximate matrix.

[0008] Optionally, a finite memory quasi-Newton method is used to adjust the model parameters of the complex-valued neural network model according to the loss function value, including: obtaining a preset set of candidate memory scales, wherein the candidate memory scale set includes multiple different memory scales; updating the candidate memory scale set according to the loss function value corresponding to the memory scale in the candidate memory scale set to obtain a target memory scale set; when the model parameters need to be adjusted, selecting a memory scale in the target memory scale set, and updating the approximate matrix corresponding to the complex-valued neural network model according to m historical iteration information of the memory scale.

[0009] Optionally, the candidate memory scale set is updated according to the loss function value corresponding to the memory scale in the candidate memory scale set to obtain the target memory scale set, including: in each iteration of model training, determining the loss function value corresponding to the complex-valued neural network model when making predictions according to the model parameters corresponding to each memory scale in the candidate memory scale set, wherein the model parameters corresponding to the memory scale are the approximate matrices corresponding to the complex-valued neural network model determined according to m historical iteration information of the memory scale; sorting the memory scales in the candidate memory scale set in order of the loss function value from small to large to obtain a first sequence; after a first preset number of iterations, respectively determining the average values ​​of the first second preset number of memory scales in the first preset number of first sequences to obtain a first preset number of average values; and combining the first preset number of average values ​​into a target memory scale set.

[0010] Optionally, the complex-valued neural network model consists of an input layer, a multi-connected layer and an output layer, wherein the input layer is used to receive the original data input into the complex-valued neural network model, and the number of neurons contained in the input layer is determined by the number of data contained in the data group in the target training data set; the multi-connected layer contains multiple network layers, which are used to extract features of the data input into the complex-valued neural network model and learn the intrinsic representation of the data through nonlinear transformation; the output layer is used to determine the final prediction result based on the output result of the multi-connected layer, and the number of neurons contained in the output layer is determined by the number of types of the first label or the second label in the target training data set.

[0011] Optionally, the optical signal to be identified is optical signal data in a complex form transmitted in a coherent optical communication system, wherein the coherent optical communication system includes: a transmitting end and a receiving end, wherein the transmitting end is used to map a pseudo-random binary sequence into signals with different amplitude levels through orthogonal amplitude modulation encoding, and upsample and filter the signal; import the signal after upsampling and filtering into a waveform generator to obtain an analog level signal; load the analog level signal onto an orthogonal amplitude IQ modulator to obtain a modulation format signal; load the modulation format signal onto a narrow linewidth laser to obtain an optical signal, and send the optical signal to a receiving end through a transmission link; the receiving end is used to perform analog-to-digital conversion on the received optical signal, and perform digital signal processing on the optical signal after the analog-to-digital conversion to obtain the optical signal to be identified, wherein the digital signal processing includes: resampling, matched filtering, IQ orthogonalization, dispersion compensation, and clock recovery.

[0012] According to another aspect of an embodiment of the present application, a model training method is also provided, including: training a complex-valued neural network model based on a finite memory quasi-Newton method to obtain a target complex-valued neural network model, wherein the target complex-valued neural network model is used to identify signal parameter information corresponding to an optical signal to be identified, the optical signal to be identified is optical signal data in complex form, and the signal parameter information includes: modulation format, optical signal-to-noise ratio; wherein the finite memory quasi-Newton method is used to accelerate the convergence of the model by updating an approximate matrix based on m historical iteration information, where m is a memory scale, and during the model training process, each iteration corresponds to a memory scale, and the memory scale is determined based on the error between the model prediction value and the actual value corresponding to different memory scales in the historical iteration process.

[0013] According to another aspect of an embodiment of the present application, there is also provided an optical signal recognition device, including: a signal acquisition module, used to acquire an optical signal to be recognized, wherein the optical signal to be recognized is optical signal data in complex form; a signal recognition module, used to use a target complex-valued neural network model to recognize the optical signal to be recognized, and obtain signal parameter information corresponding to the optical signal to be recognized, wherein the signal parameter information includes: modulation format, optical signal-to-noise ratio; wherein the target complex-valued neural network model is trained based on a finite memory quasi-Newton method, and the finite memory quasi-Newton method is used to accelerate the convergence of the model by updating an approximate matrix based on m historical iteration information, where m is a memory scale, and during the model training process, each iteration corresponds to a memory scale, and the memory scale is determined based on the error between the model prediction value and the actual value corresponding to different memory scales in the historical iteration process.

[0014] According to another aspect of an embodiment of the present application, there is also provided an electronic device, including: a memory and a processor, the processor being used to run a program stored in the memory, wherein the program executes an optical signal recognition method or a model training method when running.

[0015] According to another aspect of an embodiment of the present application, a non-volatile storage medium is further provided, the non-volatile storage medium including a stored computer program, wherein the device where the non-volatile storage medium is located executes the optical signal recognition method or the model training method by running the computer program.

[0016] According to another aspect of the embodiments of the present application, a computer program product is provided, including a computer program, which implements the steps of the optical signal recognition method or the model training method when the computer program is executed by a processor.

[0017] In an embodiment of the present application, an optical signal to be identified is obtained, wherein the optical signal to be identified is optical signal data in complex form; a target complex-valued neural network model is used to identify the optical signal to be identified, and signal parameter information corresponding to the optical signal to be identified is obtained, wherein the signal parameter information includes: modulation format, optical signal-to-noise ratio; wherein the target complex-valued neural network model is trained based on a finite memory quasi-Newton method, and the finite memory quasi-Newton method is used to accelerate the convergence of the model by updating an approximate matrix based on m historical iteration information, m is a memory scale, and during the model training process, each iteration corresponds to a memory scale, and the memory scale is determined according to the error between the model prediction value and the actual value corresponding to different memory scales in the historical iteration process. Based on the finite memory quasi-Newton algorithm, the memory scale is adaptively adjusted, thereby achieving the purpose of reducing the convergence speed and complexity of the complex-valued neural network model, thereby solving the technical problems of high complexity and poor accuracy of the optical signal performance monitoring method in the related art. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0019] Figure 1 It is a hardware structure block diagram of a computer terminal (or electronic device) for implementing a method for optical signal recognition provided in an embodiment of the present application;

[0020] Figure 2 is a schematic diagram of a method flow for optical signal recognition provided according to an embodiment of the present application;

[0021] Figure 3 is a schematic diagram of the architecture of a coherent optical communication system provided according to an embodiment of the present application;

[0022] Figure 4 is a schematic diagram of a format of a training data set provided according to an embodiment of the present application;

[0023] Figure 5 is a schematic diagram of a complex-valued neural network architecture provided according to an embodiment of the present application;

[0024] Figure 6 is a schematic diagram of a neuron structure provided according to an embodiment of the present application;

[0025] Figure 7 is a schematic diagram of a modulation format recognition accuracy result provided according to an embodiment of the present application;

[0026] Figure 8It is a schematic diagram of a comparison between an estimated optical signal-to-noise ratio value output by a neural network and a true value according to an embodiment of the present application;

[0027] Fig. 9 It is a structural schematic diagram of an optical signal recognition device provided according to an embodiment of the present application. DETAILED DESCRIPTION

[0028] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0029] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0030] In order to facilitate those skilled in the art to better understand the embodiments of the present application, some technical terms or nouns involved in the embodiments of the present application are explained as follows:

[0031] CV-NNs (Complex-Valued Neural Networks): A neural network model that uses complex weights and activation functions. Compared with traditional real-valued neural networks, CV-NNs can more effectively process data with phase information, such as signal processing, image recognition, and speech recognition.

[0032] Modulation format recognition: Modulation format recognition is a key task in the field of communication, especially in wireless communication and radar systems. It involves identifying the modulation type of the signal, such as ASK (Amplitude Shift Keying), FSK (Frequency Shift Keying), PSK (Phase Shift Keying) and its variants QPSK, 8PSK, etc.

[0033] OSNR (Optical Signal-to-Noise Ratio): An important parameter in optical fiber communication systems that measures signal quality. It is defined as the ratio of signal power to noise power, usually expressed in decibels (dB).

[0034] In coherent optical communications, it is crucial to find a superior modulation format recognition and optical signal-to-noise ratio estimation method. In related technologies, there are endless solutions based on machine learning and deep neural networks. For example, support vector machines and artificial neural networks are used to process the cumulative distribution function of signals for modulation format recognition and OSNR estimation, but they have limitations in signal noise resistance. The method of using the amplitude histogram of the signal as a data set for deep learning training also shows weaknesses in complexity and signal noise resistance.

[0035] With the improvement of signal modulation formats, convolutional neural networks are unable to distinguish the information characteristics of high-order signal constellation diagrams. At this time, complex-valued neural networks have shown their advantages in processing such complex-valued signals, and have advantages in dealing with phase-sensitive problems, reducing model complexity, and converging more smoothly and stably. At present, the schemes using complex-valued neural networks for identification in related technologies, such as the traditional complex-valued neural network mentioned in CN113517926A, do not actually use the characteristics of the complex-valued signal itself, but use the grayscale image as the network input. In essence, it is still based on the idea of ​​convolutional neural networks, and the complex-valued neural network is based on the traditional L-BFGS algorithm, which has high convergence complexity and is not conducive to the development of actual optical modules.

[0036] In order to solve the above problems, relevant solutions are provided in the embodiments of the present application, which are described in detail below.

[0037] According to an embodiment of the present application, a method embodiment of optical signal recognition is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0038] The method embodiments provided in the embodiments of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Figure 1 FIG. 1 shows a hardware structure block diagram of a computer terminal (or electronic device) for implementing an optical signal recognition method. Figure 1As shown, the computer terminal 10 (or electronic device) may include one or more (102a, 102b, ..., 102n are used to illustrate) processors 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, and a transmission device 106 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the BUS bus), a network interface, a power supply and / or a camera. It can be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations are shown.

[0039] It should be noted that the one or more processors 102 and / or other data processing circuits described above may generally be referred to herein as "data processing circuits". The data processing circuits may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuit may be a single independent processing module, or may be incorporated in whole or in part into any of the other components in the computer terminal 10 (or electronic device). As described in the embodiments of the present application, the data processing circuit acts as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0040] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the optical signal recognition method in the embodiment of the present application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, the above-mentioned optical signal recognition method is realized. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories may be connected to the computer terminal 10 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0041] The transmission device 106 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the computer terminal 10. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0042] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the computer terminal 10 (or electronic device).

[0043] In the above operating environment, the embodiment of the present application provides a method for optical signal recognition. Figure 2 is a schematic diagram of a method flow for optical signal recognition provided according to an embodiment of the present application, such as Figure 2 As shown, the method comprises the following steps:

[0044] Step S202, obtaining a light signal to be identified, wherein the light signal to be identified is light signal data in a plural form;

[0045] Step S204, using the target complex-valued neural network model to identify the optical signal to be identified, and obtaining signal parameter information corresponding to the optical signal to be identified, wherein the signal parameter information includes: modulation format and optical signal-to-noise ratio;

[0046] Among them, the target complex-valued neural network model is trained based on the finite memory quasi-Newton method. The finite memory quasi-Newton method is used to accelerate the convergence of the model by updating the approximate matrix based on m historical iteration information. m is the memory scale. During the model training process, each iteration corresponds to a memory scale. The memory scale is determined based on the error between the model prediction value and the actual value corresponding to different memory scales in the historical iteration process.

[0047] Through the above steps, based on the adaptive finite memory quasi-Newton algorithm, on the basis of the limited memory quasi-Newton algorithm, the memory scale is adaptively adjusted, thereby achieving the purpose of reducing the convergence speed and complexity of the complex-valued neural network model, thereby solving the technical problems of high complexity and poor accuracy of the optical signal performance monitoring method in the related technology.

[0048] The optical signal recognition method in steps S202 to S204 of the embodiment of the present application is further introduced below.

[0049] In an embodiment of the present application, a complex-valued neural network (AL-CV-NNs) based on an adaptive finite-memory quasi-Newton learning algorithm is proposed to realize a scheme for modulation format recognition and OSNR monitoring of a signal. In this embodiment, the adaptive finite-memory quasi-Newton learning algorithm is illustrated by taking an adaptive finite-memory BFGS algorithm (AL-BFGS, hereinafter referred to as the AL algorithm) as an example, and the complex-valued neural network is illustrated by taking AL-CV-NNs as an example. Compared with traditional complex-valued neural networks, AL-CV-NNs has efficient learning and a simple structure, can adapt to many types of modulation signals from low-order to high-order, and exhibits excellent performance in modulation format recognition and OSNR estimation, which is described in detail below.

[0050] The optical signal identified by the solution of the present application is an optical signal to be identified after being transmitted via a coherent optical communication system, wherein the structure of the coherent optical communication system is as follows.

[0051] In some embodiments of the present application, the optical signal to be identified is optical signal data in a complex form transmitted in a coherent optical communication system, wherein the coherent optical communication system includes: a transmitting end and a receiving end, wherein the transmitting end is used to map a pseudo-random binary sequence into signals with different amplitude levels through orthogonal amplitude modulation encoding, and upsample and filter the signal; the signal after upsampling and filtering is introduced into a waveform generator to obtain an analog level signal; the analog level signal is loaded onto an orthogonal amplitude IQ modulator to obtain a modulation format signal; the modulation format signal is loaded onto a narrow linewidth laser to obtain an optical signal, and the optical signal is sent to the receiving end through a transmission link; the receiving end is used to perform analog-to-digital conversion on the received optical signal, and perform digital signal processing on the optical signal after analog-to-digital conversion to obtain the optical signal to be identified, wherein the digital signal processing includes: resampling, matched filtering, IQ orthogonalization, dispersion compensation, and clock recovery.

[0052] Specifically, Figure 3 is a schematic diagram of the architecture of a coherent optical communication system provided according to an embodiment of the present application, such as Figure 3As shown, at the transmitting end, the pseudo-random binary sequence is mapped into amplitude levels through QAM (Quadrature Amplitude Modulation) coding, and then imported into the arbitrary waveform generator by upsampling and root raised cosine filtering to generate analog signal levels. Different amplitude levels are loaded into the IQ modulator to generate the transmitted modulation format signal, and then loaded into the narrow linewidth laser to obtain the optical signal; then the optical signal is amplified by the erbium-doped fiber amplifier in the transmission link, and the attenuator adjusts the signal quality, and then input into the standard single-mode optical fiber for transmission to reach the receiving end; after the receiver receives the optical signal, it performs analog-to-digital conversion, and the data after clock recovery is obtained through the digital signal processing preprocessing at the receiving end, that is, the optical signal to be identified that needs to be identified for modulation format identification and OSNR estimation. This system architecture can generate and transmit signals of various modulation formats such as QPSK to 1024QAM.

[0053] The following introduces the training process of the target complex-valued neural network model used for modulation format recognition and OSNR estimation in the embodiment of the present application, including: training data set construction, model structure construction, model training method, etc. The following first introduces the preparation of the training data set.

[0054] In some embodiments of the present application, a target complex-valued neural network model is obtained by training based on a target training data set, and the target training data set includes a first data set and a second data set, wherein the first data set includes: a plurality of data groups, and a first label corresponding to the data group, each data group includes a plurality of orthogonal amplitude modulation signal data in complex form corresponding to optical signals of the same modulation format, and the first label is used to characterize the modulation format corresponding to the orthogonal amplitude modulation signal data in the data group, and the modulation format includes at least one of the following: quadrature phase shift keying QPSK, sixteen-state orthogonal amplitude modulation 16QAM, sixty-four-state orthogonal amplitude modulation 64QAM, two hundred and fifty-six-state orthogonal amplitude modulation 256QAM and giga-state orthogonal amplitude modulation 1024QAM; the second data set includes: a plurality of data groups, and a second label corresponding to the data group, each data group includes a plurality of orthogonal amplitude modulation signal data in complex form corresponding to optical signals of the same optical signal-to-noise ratio, and the second label is used to characterize the optical signal-to-noise ratio corresponding to the orthogonal amplitude modulation signal data in the data group.

[0055] Specifically, Figure 4 As shown, signal data a n-m +jb n-m represents the transmitted QAM signal, which is a complex value, where n represents the nth signal and m represents the mth modulation format, for example a 2-2 +jb 2-2That is, it represents the second signal of the second modulation format. Since the highest order modulation format to be identified is 1024QAM, which is displayed as 1024 constellation points on the constellation diagram, in order to ensure that the constellation points of each signal can cover it and extract sufficient information features, in this embodiment, each data group selects 2048 original signals.

[0056] In addition, in this embodiment, a 5-bit binary number can be used to represent the five modulation formats respectively, which is used as a label (ie, the first label mentioned above) to facilitate subsequent network training. Figure 4 The data shown in the figure are data groups corresponding to the five modulation formats. In practical applications, the amount of data transmitted in real time by the communication system is relatively large. For convenience, in this implementation, the amount of data is 2048000 as an example. The signal of each modulation format can be cut into 1000 groups. Figure 4 The signal datasets shown are used to ensure sufficient data training and test sets, and each dataset is marked Figure 4 After the above process, the data set of the signal modulation format recognition part, that is, the above-mentioned first data set, can be obtained. The preparation of the second data set of the OSNR monitoring part is similar. It is only necessary to replace the first label representing the modulation format type with the second label representing different OSNR values. It will not be repeated here.

[0057] The following is an introduction to the model structure of the initial complex-valued neural network model constructed in the embodiments of the present application (taking a low-complexity complex-valued neural network based on the AL-BFGS algorithm as an example), as follows.

[0058] In some embodiments of the present application, a complex-valued neural network model is composed of an input layer, a multi-connected layer and an output layer, wherein the input layer is used to receive the original data input into the complex-valued neural network model, and the number of neurons contained in the input layer is determined by the number of data contained in the data group in the target training data set; the multi-connected layer includes multiple network layers, which are used to extract features of the data input into the complex-valued neural network model and learn the intrinsic representation of the data through nonlinear transformation; the output layer is used to determine the final prediction result based on the output result of the multi-connected layer, and the number of neurons contained in the output layer is determined by the number of types of the first label or the second label in the target training data set.

[0059] like Figure 5 As shown, the complex-valued neural network model in the embodiment of the present application is composed of an input layer, a multi-connection layer and an output layer, wherein the middle multi-connection layer can adjust the number of network layers and the number of neurons according to actual needs; the number of neurons included in the input layer is determined by the number of data included in the data group in the target training data set. For example, due to Figure 4One data group in the training data set shown contains 2048 data. Therefore, in this embodiment, the input layer can select 2048 neurons; the output layer is determined according to the number of modulation format types and the number of OSNR monitoring values. Taking QPSK-1024QAM as an example, there are 5 modulation format types in this embodiment, and the number of output neurons can be set to 5.

[0060] In the embodiments of the present application, Figure 5 The neuron shown in c i 、c k The detailed structure of Figure 6 The corresponding output formula is as follows:

[0061]

[0062] Among them, x n Represents the output of the previous layer as the input value to the next layer of neurons. i Represents the weight matrix, ∑ represents the sum, b represents the bias of the neurons in this layer. In this embodiment, the loss function of the network can adopt the minimum mean square error function, f represents the activation function of the neural network. In this embodiment, the activation function can use the complex hyperbolic tangent function to ensure the boundedness and resolvability of the activation function. The formula is as follows: f(x)=tanh(Re(x))+jtanh(Im(x)).

[0063] After obtaining the target training data set and the initial complex-valued neural network model, model training can begin. The specific steps are as follows.

[0064] In some embodiments of the present application, the training steps of the target complex-valued neural network model include: obtaining an initial complex-valued neural network model; using the initial complex-valued neural network model to make predictions based on the data of the data group in the target training data set to obtain predicted values ​​of signal parameter information; determining a loss function value based on the predicted value of the signal parameter information and the true value of the signal parameter information represented by the first label and / or the second label in the target training data set, wherein the loss function value is used to characterize the error between the predicted value and the true value; using a finite memory quasi-Newton method to adjust the model parameters of the complex-valued neural network model based on the loss function value, and repeating the above training steps using the adjusted complex-valued neural network model until the loss function value is less than a preset error threshold, thereby obtaining the target complex-valued neural network model, wherein the model parameters include: an approximate matrix.

[0065] In this embodiment, the limited memory quasi-Newton method is explained by taking L-BFGS as an example. L-BFGS is a second-order learning algorithm in a neural network. It is extended to the complex domain and has a high convergence efficiency when solving unconstrained optimization problems. It is a solution that trades time for space. Specifically, m vectors containing previous and subsequent iteration information (historical iteration information) are used to assist in calculating the approximate matrix of the neural network, where m is called the memory scale of the algorithm, and the selection of m also affects the performance of the algorithm. Usually, the algorithm updates the memory scale m in each iteration process, and m is selected from [1,2…,M] (i.e., the candidate memory scale set), M is the upper limit value, and the value selected by M is usually larger.

[0066] The complex-valued neural network model is trained in the above manner. The network needs to perform forward propagation and backward propagation to calculate the gradient. At this time, the network performance also depends on the memory scale m, and the m parameter value needs to be continuously adjusted. This undoubtedly increases the complexity of the entire network training. For example, in order to reduce the complexity of network training, in the embodiment of the present application, the M value can be updated again by taking the average value, that is, the candidate memory scale set is updated. The specific steps are as follows.

[0067] In some embodiments of the present application, a finite memory quasi-Newton method is used to adjust the model parameters of a complex-valued neural network model according to a loss function value, including the following steps: obtaining a preset set of candidate memory scales, wherein the candidate memory scale set includes a plurality of different memory scales; updating the candidate memory scale set according to the loss function values ​​corresponding to the memory scales in the candidate memory scale set to obtain a target memory scale set; when the model parameters need to be adjusted, selecting a memory scale from the target memory scale set, and updating the approximate matrix corresponding to the complex-valued neural network model according to m historical iteration information of the memory scale.

[0068] In some embodiments of the present application, updating the candidate memory scale set according to the loss function value corresponding to the memory scale in the candidate memory scale set to obtain the target memory scale set includes the following steps: in each iteration of model training, determining the loss function value corresponding to the complex-valued neural network model when making predictions based on the model parameters corresponding to each memory scale in the candidate memory scale set, wherein the model parameters corresponding to the memory scale are the approximate matrices corresponding to the complex-valued neural network model determined based on m historical iteration information of the memory scale; sorting the memory scales in the candidate memory scale set in order of the loss function value from small to large to obtain a first sequence; after a first preset number of iterations, respectively determining the average values ​​of the first second preset number of memory scales in the first preset number of first sequences to obtain a first preset number of average values; and combining the first preset number of average values ​​into a target memory scale set.

[0069] Specifically, in each iteration of model training, the error value e corresponding to each memory scale m in the candidate memory scale set [1, 2…, M] is first calculated. m (loss function value), then, according to the error value e m Rearrange multiple m values ​​in the candidate memory scale set in ascending order to obtain a new sequence m about m p (i.e., the first sequence mentioned above); each replacement process can obtain a first sequence after the candidate memory scale set is reordered. After multiple (e.g., a first preset number of) iterations, multiple new sequences (first sequences) about m can be obtained. After that, each first sequence m can be intercepted. p The average value of the first second preset number (for example, 3) m values ​​with smaller error values ​​(i.e., the first second preset number of memory scales in the first sequence) is used to form a new sequence m again. q , and use this as a new set of m selected values, that is, the above target memory scale set. Through the above steps, the complexity of subsequent network training can be greatly reduced, making it more in line with the actual optical module development needs.

[0070] According to the training process of the above training data set construction, model structure construction, model training method, etc., the embodiment of the present application can realize the recognition of five commonly used rectangular constellation signals from low-order to ultra-high-order coherent optical communication systems, including QPSK, 16QAM, 64QAM, 256QAM and 1024QAM, under low OSNR conditions, with extremely strong anti-noise performance and without the need for too much intermediate link processing; in addition, for the modulation format types commonly used in existing coherent optical communication commercial environments, such as QPSK, 16QAM, 64QAM and 256QAM, the embodiment of the present application can also be further optimized, so that the four modulation format signals can achieve 100% recognition accuracy under the condition of 10dB ultra-low optical signal-to-noise ratio. For OSNR estimation, the embodiment of the present application selected three commonly used modulation formats, QPSK, 16QAM and 64QAM, and tested them within the OSNR range of the commercial environment, and the deviation was controlled at about 0.01dB, as follows.

[0071] For the modulation format recognition part, this embodiment produces training data sets for the five modulation formats of QPSK-1024QAM, 1000 groups for each modulation format, and conducts experiments according to the training test ratio of 7:3. Since the data set directly uses complex-valued signals, the signals carry phase amplitude information, and the complex-valued network is easier to extract signal features. In the signal modulation format recognition stage, the network only needs to set a single intermediate connection layer. The input neurons are set to 2048 according to the data set, and the intermediate layer neurons are set to 15-20. Since five types of modulation formats are identified and classified, five output neurons are finally selected. The network learning rate is set to 0.001, and iterates 150 times. The memory scale of the complex-valued neural network is adjusted automatically according to the adaptive rules, which greatly reduces the complexity of the network.

[0072] For the OSNR estimation part, this solution takes QPSK-64QAM as an example. The three modulation formats are selected from the lower limit range of OSNR required for commercial use, which are 11-18dB, 17-24dB, and 23-30dB respectively. Each OSNR value corresponds to 300 sets of data sets, and the training test is also carried out in a 7:3 ratio. For the OSNR test data, this embodiment repeats each set of test data for more than 100 times and takes the average value to obtain the final estimate. Since the OSNR estimation stage involves more data sets and OSNR values, this network sets two layers of intermediate connection layers. The input neurons are still selected as 2048 according to the data set format. The two layers of intermediate layer neurons are preferably selected as 20-25, and the last 8 are output neurons. The network learning rate is set to 0.001, and it can converge to a stable value after 200 iterations.

[0073] The output of the modulation format recognition of the complex-valued neural network model is as follows: Figure 7 As shown, the five characteristic curves in the figure correspond to the recognition accuracy of five modulation formats, namely QPSK, 16QAM, 64QAM, 256QAM and 1024QAM. The horizontal axis represents the optical signal-to-noise ratio of the signal, and the vertical axis represents the recognition accuracy of the complex-valued neural network. From the information in the figure, it can be seen that the three signals of QPSK, 16QAM and 64QAM can achieve 100% recognition accuracy when the OSNR value is less than 10dB. For ultra-high-order signals 256QAM and 1024QAM, the recognition accuracy can reach 100% at 12dB and 15dB respectively, and as the optical signal-to-noise ratio of the transmission signal increases, the accuracy is also constantly improving. From the recognition effect, it reaches requirements far higher than those of actual commercial links. The present application scheme is applied to the recognition of ultra-high-order signal modulation formats for the first time, and can achieve 100% recognition accuracy under ultra-low optical signal-to-noise ratio conditions.

[0074] The output of the OSNR estimation of the complex-valued neural network is as follows: Figure 8As shown in the figure, the straight line in the figure represents the reference line of OSNR, and the completely accurate value should fall on this line. The horizontal axis represents the optical signal-to-noise ratio of the signal, and the vertical axis on the left represents the estimated value of the complex-valued neural network. It can be seen from the points on the straight line that the estimated value of the network output is basically accurate. For details, please refer to the vertical axis on the right, which represents the deviation of the estimated OSNR value. Here, only the QPSK-64QAM signal is estimated, and the error is basically controlled within 0.01dB, achieving a good effect.

[0075] This application scheme can not only realize the monitoring of low-order signals, but also can be applied to ultra-high-order optical signals. Five signal types of QPSK-1024QAM are identified, ensuring that 1024QAM can achieve 100% recognition accuracy under 15dB. At the same time, the OSNR of QPSK-64QAM signal is estimated, and the error value is controlled within 0.01dB, which ensures the effectiveness of the method and provides a good supplement for future large-capacity optical transmission.

[0076] The present application proposes a complex-valued neural network model based on an adaptive limited memory BFGS algorithm to identify multiple modulation formats from low to high order and monitor the optical signal-to-noise ratio. It optimizes the complexity and accuracy of the complex-valued neural network in the convergence process in the related technology, reduces the computational complexity by adaptively selecting the memory scale in the BFGS algorithm, greatly improves the network training efficiency, and is more conducive to the development of commercial optical modules. It can be applied to the digital signal processing demodulation algorithm of the traditional coherent optical communication system. By adding this module in the preprocessing stage, modulation format recognition and OSNR estimation can be performed to realize optical performance monitoring, realize flexible selection of subsequent equalization algorithms, ensure the transmission efficiency and stability of the system, and realize the correct identification of ultra-high-order signals for the first time. In addition, the data set processed by the present application is a complex-valued signal with phase amplitude, which is more conducive to the feature extraction of the complex-valued neural network, and shows excellent performance in the test results.

[0077] According to an embodiment of the present application, a model training method is also provided, the method comprising the following steps:

[0078] Based on the limited memory quasi-Newton method, the complex-valued neural network model is trained to obtain a target complex-valued neural network model, wherein the target complex-valued neural network model is used to identify signal parameter information corresponding to the optical signal to be identified, the optical signal to be identified is optical signal data in complex form, and the signal parameter information includes: modulation format and optical signal-to-noise ratio;

[0079] Among them, the limited memory quasi-Newton method is used to accelerate the convergence of the model by updating the approximate matrix based on m historical iteration information. m is the memory scale. During the model training process, each iteration corresponds to a memory scale. The memory scale is determined based on the error between the model prediction value and the actual value corresponding to different memory scales in the historical iteration process.

[0080] It should be noted that the model training method provided in this embodiment is Figure 2 The optical signal recognition method shown corresponds to a method embodiment, therefore, the relevant explanations and descriptions of the above optical signal recognition method are also applicable to the embodiments of the present application and will not be repeated here.

[0081] According to an embodiment of the present application, an embodiment of an optical signal recognition device is also provided. Fig. 9 is a schematic diagram of the structure of an optical signal recognition device provided according to an embodiment of the present application. Fig. 9 As shown, the device comprises:

[0082] A signal acquisition module 90, used to acquire the optical signal to be identified, wherein the optical signal to be identified is optical signal data in plural form;

[0083] The signal recognition module 92 is used to use the target complex value neural network model to recognize the optical signal to be recognized, and obtain the signal parameter information corresponding to the optical signal to be recognized, wherein the signal parameter information includes: modulation format and optical signal-to-noise ratio;

[0084] Among them, the target complex-valued neural network model is trained based on the finite memory quasi-Newton method. The finite memory quasi-Newton method is used to accelerate the convergence of the model by updating the approximate matrix based on m historical iteration information. m is the memory scale. During the model training process, each iteration corresponds to a memory scale. The memory scale is determined based on the error between the model prediction value and the actual value corresponding to different memory scales in the historical iteration process.

[0085] Optionally, the target complex-valued neural network model is trained based on a target training data set, and the target training data set includes a first data set and a second data set, wherein the first data set includes: a plurality of data groups, and a first label corresponding to the data group, each data group includes a plurality of orthogonal amplitude modulation signal data in complex form corresponding to optical signals of the same modulation format, and the first label is used to characterize the modulation format corresponding to the orthogonal amplitude modulation signal data in the data group, and the modulation format includes at least one of the following: quadrature phase shift keying QPSK, sixteen-state orthogonal amplitude modulation 16QAM, sixty-four-state orthogonal amplitude modulation 64QAM, two hundred and fifty-six-state orthogonal amplitude modulation 256QAM and giga-state orthogonal amplitude modulation 1024QAM; the second data set includes: a plurality of data groups, and a second label corresponding to the data group, each data group includes a plurality of orthogonal amplitude modulation signal data in complex form corresponding to optical signals of the same optical signal-to-noise ratio, and the second label is used to characterize the optical signal-to-noise ratio corresponding to the orthogonal amplitude modulation signal data in the data group.

[0086] Optionally, the training steps of the target complex-valued neural network model include: obtaining an initial complex-valued neural network model; using the initial complex-valued neural network model to make predictions based on the data of the data group in the target training data set to obtain predicted values ​​of signal parameter information; determining a loss function value based on the predicted value of the signal parameter information and the true value of the signal parameter information represented by the first label and / or the second label in the target training data set, wherein the loss function value is used to characterize the error between the predicted value and the true value; using a finite memory quasi-Newton method to adjust the model parameters of the complex-valued neural network model based on the loss function value, and repeating the above training steps using the adjusted complex-valued neural network model until the loss function value is less than a preset error threshold, thereby obtaining the target complex-valued neural network model, wherein the model parameters include: an approximate matrix.

[0087] Optionally, a finite memory quasi-Newton method is used to adjust the model parameters of the complex-valued neural network model according to the loss function value, including: obtaining a preset set of candidate memory scales, wherein the candidate memory scale set includes multiple different memory scales; updating the candidate memory scale set according to the loss function value corresponding to the memory scale in the candidate memory scale set to obtain a target memory scale set; when the model parameters need to be adjusted, selecting a memory scale in the target memory scale set, and updating the approximate matrix corresponding to the complex-valued neural network model according to m historical iteration information of the memory scale.

[0088] Optionally, the candidate memory scale set is updated according to the loss function value corresponding to the memory scale in the candidate memory scale set to obtain the target memory scale set, including: in each iteration of model training, determining the loss function value corresponding to the complex-valued neural network model when making predictions according to the model parameters corresponding to each memory scale in the candidate memory scale set, wherein the model parameters corresponding to the memory scale are the approximate matrices corresponding to the complex-valued neural network model determined according to m historical iteration information of the memory scale; sorting the memory scales in the candidate memory scale set in order of the loss function value from small to large to obtain a first sequence; after a first preset number of iterations, respectively determining the average values ​​of the first second preset number of memory scales in the first preset number of first sequences to obtain a first preset number of average values; and combining the first preset number of average values ​​into a target memory scale set.

[0089] Optionally, the complex-valued neural network model consists of an input layer, a multi-connected layer and an output layer, wherein the input layer is used to receive the original data input into the complex-valued neural network model, and the number of neurons contained in the input layer is determined by the number of data contained in the data group in the target training data set; the multi-connected layer contains multiple network layers, which are used to extract features of the data input into the complex-valued neural network model and learn the intrinsic representation of the data through nonlinear transformation; the output layer is used to determine the final prediction result based on the output result of the multi-connected layer, and the number of neurons contained in the output layer is determined by the number of types of the first label or the second label in the target training data set.

[0090] Optionally, the optical signal to be identified is optical signal data in a complex form transmitted in a coherent optical communication system, wherein the coherent optical communication system includes: a transmitting end and a receiving end, wherein the transmitting end is used to map a pseudo-random binary sequence into signals with different amplitude levels through orthogonal amplitude modulation encoding, and upsample and filter the signal; import the signal after upsampling and filtering into a waveform generator to obtain an analog level signal; load the analog level signal onto an orthogonal amplitude IQ modulator to obtain a modulation format signal; load the modulation format signal onto a narrow linewidth laser to obtain an optical signal, and send the optical signal to a receiving end through a transmission link; the receiving end is used to perform analog-to-digital conversion on the received optical signal, and perform digital signal processing on the optical signal after the analog-to-digital conversion to obtain the optical signal to be identified, wherein the digital signal processing includes: resampling, matched filtering, IQ orthogonalization, dispersion compensation, and clock recovery.

[0091] It should be noted that the various modules in the above-mentioned optical signal recognition device can be program modules (for example, a set of program instructions that implement a certain specific function) or hardware modules. For the latter, it can be expressed in the following forms, but is not limited to this: the expression form of each of the above-mentioned modules is a processor, or the functions of each of the above-mentioned modules are implemented by a processor.

[0092] It should be noted that the optical signal recognition device provided in this embodiment can be used to perform Figure 2 The optical signal recognition method shown, therefore, the relevant explanations and descriptions of the above optical signal recognition method are also applicable to the embodiments of the present application and will not be repeated here.

[0093] The embodiment of the present application also provides a non-volatile storage medium, the non-volatile storage medium includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the following optical signal recognition method or model training method by running the computer program: obtaining an optical signal to be recognized, wherein the optical signal to be recognized is optical signal data in complex form; using a target complex-valued neural network model to recognize the optical signal to be recognized, and obtain signal parameter information corresponding to the optical signal to be recognized, wherein the signal parameter information includes: modulation format, optical signal-to-noise ratio; wherein the target complex-valued neural network model is obtained by training based on a finite memory quasi-Newton method, and the finite memory quasi-Newton method is used to accelerate the convergence of the model by updating an approximate matrix based on m historical iteration information, where m is a memory scale, and during the model training process, each iteration corresponds to a memory scale, and the memory scale is determined based on the error between the model prediction value and the actual value corresponding to different memory scales in the historical iteration process.

[0094] Based on the finite memory quasi-Newton method, the complex-valued neural network model is trained to obtain a target complex-valued neural network model, wherein the target complex-valued neural network model is used to identify signal parameter information corresponding to an optical signal to be identified, the optical signal to be identified is optical signal data in complex form, and the signal parameter information includes: modulation format, optical signal-to-noise ratio; wherein the finite memory quasi-Newton method is used to accelerate the convergence of the model by updating the approximate matrix based on m historical iteration information, m is the memory scale, and during the model training process, each iteration corresponds to a memory scale, and the memory scale is determined based on the error between the model prediction value and the actual value corresponding to different memory scales in the historical iteration process.

[0095] The embodiment of the present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the optical signal recognition method or model training method described in each embodiment of the present application: obtaining an optical signal to be recognized, wherein the optical signal to be recognized is optical signal data in complex form; using a target complex-valued neural network model to recognize the optical signal to be recognized, and obtain signal parameter information corresponding to the optical signal to be recognized, wherein the signal parameter information includes: modulation format, optical signal-to-noise ratio; wherein the target complex-valued neural network model is trained based on a finite memory quasi-Newton method, and the finite memory quasi-Newton method is used to accelerate the convergence of the model by updating an approximate matrix based on m historical iteration information, where m is a memory scale, and during the model training process, each iteration corresponds to a memory scale, and the memory scale is determined based on the error between the model prediction value and the actual value corresponding to different memory scales in the historical iteration process.

[0096] Based on the finite memory quasi-Newton method, the complex-valued neural network model is trained to obtain a target complex-valued neural network model, wherein the target complex-valued neural network model is used to identify signal parameter information corresponding to an optical signal to be identified, the optical signal to be identified is optical signal data in complex form, and the signal parameter information includes: modulation format, optical signal-to-noise ratio; wherein the finite memory quasi-Newton method is used to accelerate the convergence of the model by updating the approximate matrix based on m historical iteration information, m is the memory scale, and during the model training process, each iteration corresponds to a memory scale, and the memory scale is determined based on the error between the model prediction value and the actual value corresponding to different memory scales in the historical iteration process.

[0097] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0098] In the above embodiments of the present application, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0099] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0100] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0101] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0102] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, disk or optical disk and other media that can store program codes.

[0103] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for optical signal recognition, characterized in that: include: Acquire a light signal to be identified, wherein the light signal to be identified is light signal data in a plural form; Using a target complex-valued neural network model to identify the optical signal to be identified, and obtaining signal parameter information corresponding to the optical signal to be identified, wherein the signal parameter information includes: modulation format and optical signal-to-noise ratio; Among them, the target complex-valued neural network model is trained based on the finite memory quasi-Newton method, and the finite memory quasi-Newton method is used to accelerate the convergence of the model by updating the approximate matrix based on m historical iteration information. m is the memory scale. During the model training process, each iteration corresponds to one memory scale. The memory scale is determined based on the error between the model prediction value and the actual value corresponding to different memory scales in the historical iteration process.

2. The optical signal recognition method according to claim 1, characterized in that: The target complex-valued neural network model is obtained by training based on a target training data set, wherein the target training data set includes a first data set and a second data set, wherein: The first data set includes: a plurality of data groups and first labels corresponding to the data groups, each of the data groups includes a plurality of orthogonal amplitude modulation signal data in complex form corresponding to optical signals of the same modulation format, the first label is used to characterize the modulation format corresponding to the orthogonal amplitude modulation signal data in the data group, and the modulation format includes at least one of the following: quadrature phase shift keying QPSK, sixteen-state orthogonal amplitude modulation 16QAM, sixty-four-state orthogonal amplitude modulation 64QAM, two hundred and fifty-six-state orthogonal amplitude modulation 256QAM and giga-state orthogonal amplitude modulation 1024QAM; The second data set includes: multiple data groups and second labels corresponding to the data groups, each of the data groups includes multiple orthogonal amplitude modulation signal data in complex form of optical signals corresponding to the same optical signal-to-noise ratio, and the second label is used to characterize the optical signal-to-noise ratio corresponding to the orthogonal amplitude modulation signal data in the data group.

3. The optical signal recognition method according to claim 2, characterized in that: The training steps of the target complex-valued neural network model include: Obtain an initial complex-valued neural network model; Using the initial complex-valued neural network model, prediction is performed based on the data of the data group in the target training data set to obtain a predicted value of the signal parameter information; Determine a loss function value according to the predicted value of the signal parameter information and the true value of the signal parameter information represented by the first label and / or the second label in the target training data set, wherein the loss function value is used to represent the error between the predicted value and the true value; The finite memory quasi-Newton method is used to adjust the model parameters of the complex-valued neural network model according to the loss function value, and the above training steps are repeated using the adjusted complex-valued neural network model until the loss function value is less than a preset error threshold, thereby obtaining the target complex-valued neural network model, wherein the model parameters include: the approximate matrix.

4. The optical signal recognition method according to claim 3, characterized in that: Using the limited memory quasi-Newton method and adjusting the model parameters of the complex-valued neural network model according to the loss function value includes: Acquire a preset candidate memory scale set, wherein the candidate memory scale set includes a plurality of different memory scales; According to the loss function value corresponding to the memory scale in the candidate memory scale set, the candidate memory scale set is updated to obtain a target memory scale set; When the model parameters need to be adjusted, a memory scale is selected from the target memory scale set, and the approximate matrix corresponding to the complex-valued neural network model is updated based on m historical iteration information of the memory scale.

5. The optical signal recognition method according to claim 4, characterized in that: According to the loss function value corresponding to the memory scale in the candidate memory scale set, the candidate memory scale set is updated to obtain a target memory scale set including: In each iteration of model training, determining the loss function value corresponding to the complex-valued neural network model when making predictions based on the model parameters corresponding to each memory scale in the candidate memory scale set, wherein the model parameters corresponding to the memory scale are the approximate matrices corresponding to the complex-valued neural network model determined based on the m historical iteration information of the memory scale; Sorting the memory scales in the candidate memory scale set in order of the loss function values ​​from small to large to obtain a first sequence; After a first preset number of the iterative processes, respectively determining the average values ​​of the first second preset number of the memory scales in the first preset number of the first sequences, to obtain the first preset number of the average values; The first preset number of the average values ​​are combined into the target memory scale set.

6. The optical signal recognition method according to claim 3, characterized in that: The complex-valued neural network model consists of an input layer, a multi-connection layer and an output layer, wherein: The input layer is used to receive raw data input into the complex-valued neural network model, and the number of neurons included in the input layer is determined by the number of data included in the data group in the target training data set; The multi-connected layer includes a plurality of network layers for extracting features of data input into the complex-valued neural network model and learning the intrinsic representation of the data through nonlinear transformation; The output layer is used to determine the final prediction result based on the output result of the multi-connected layer, and the number of neurons included in the output layer is determined by the number of types of the first label or the second label in the target training data set.

7. The optical signal recognition method according to claim 1, characterized in that: The optical signal to be identified is a complex optical signal data transmitted in a coherent optical communication system, wherein the coherent optical communication system includes: a transmitting end and a receiving end, wherein: The transmitting end is used to map the pseudo-random binary sequence into signals with different amplitude levels through orthogonal amplitude modulation coding, and perform up-sampling and filtering processing on the signals; import the signals after the up-sampling and filtering processing into a waveform generator to obtain an analog level signal; load the analog level signal onto an orthogonal amplitude IQ modulator to obtain a modulation format signal; load the modulation format signal onto a narrow linewidth laser to obtain an optical signal, and send the optical signal to the receiving end through a transmission link; The receiving end is used to perform analog-to-digital conversion on the received optical signal, and perform digital signal processing on the optical signal after analog-to-digital conversion to obtain the optical signal to be identified, wherein the digital signal processing includes: resampling, matched filtering, IQ orthogonalization, dispersion compensation, and clock recovery.

8. A model training method, characterized in that: include: Based on the limited memory quasi-Newton method, the complex-valued neural network model is trained to obtain a target complex-valued neural network model, wherein the target complex-valued neural network model is used to identify signal parameter information corresponding to the optical signal to be identified, the optical signal to be identified is optical signal data in complex form, and the signal parameter information includes: modulation format and optical signal-to-noise ratio; Among them, the limited memory quasi-Newton method is used to accelerate the convergence of the model by updating the approximate matrix based on m historical iteration information, where m is the memory scale. During the model training process, each iteration corresponds to one memory scale, and the memory scale is determined based on the error between the model prediction value and the actual value corresponding to different memory scales in the historical iteration process.

9. An optical signal recognition device, characterized in that: include: A signal acquisition module, used for acquiring an optical signal to be identified, wherein the optical signal to be identified is optical signal data in a plural form; A signal recognition module, used to use a target complex-valued neural network model to recognize the optical signal to be recognized, and obtain signal parameter information corresponding to the optical signal to be recognized, wherein the signal parameter information includes: modulation format and optical signal-to-noise ratio; Among them, the target complex-valued neural network model is trained based on the finite memory quasi-Newton method, and the finite memory quasi-Newton method is used to accelerate the convergence of the model by updating the approximate matrix based on m historical iteration information. m is the memory scale. During the model training process, each iteration corresponds to one memory scale. The memory scale is determined based on the error between the model prediction value and the actual value corresponding to different memory scales in the historical iteration process.

10. An electronic device, characterized in that: include: A memory and a processor, wherein the processor is used to run a program stored in the memory, wherein the program, when running, executes the optical signal recognition method described in any one of claims 1 to 7 or the model training method described in claim 8.

11. A non-volatile storage medium, characterized in that: The non-volatile storage medium includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the optical signal recognition method described in any one of claims 1 to 7 or the model training method described in claim 8 by running the computer program.

12. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the optical signal recognition method described in any one of claims 1 to 7 or the model training method described in claim 8 are implemented.

Images