A time-domain optical signal processing method and system based on optical neural network

By building a multi-layer optical neural network, using optical phase modulators and nonlinear activation modules to process time domain optical signals, the problem of insufficient time domain signal processing capability of optical neural networks in the prior art is solved, and efficient and flexible optical Signal processing, suitable for optical communication systems.

CN115514411BActive Publication Date: 2024-11-08TONGJI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202211192297.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-09-28
Publication Date
2024-11-08
Estimated Expiration
2042-09-28

AI Technical Summary

Technical Problem

The existing optical neural networks mainly process optical signals in the spatial domain, lack effective processing capabilities for optical signals in the time domain, and the devices are complex, the process is difficult and flexibility are poor, making it difficult to seamlessly connect with optical communication systems.

Method used

Using the time domain optical signal processing method based on optical neural network, the optical neural network with multi-layer structure, including optical phase modulator, dispersion module and nonlinear activation module, uses the back propagation algorithm to train the network to modulate the phase of the optical signal. and dispersion to realize the identification and classification of time domain signals.

Benefits of technology

It realizes the rapid and effective processing of time-domain optical signals, improves data capacity, simplifies device structure, reduces process complexity, enhances flexibility and scalability, and achieves seamless connection with optical communication systems.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115514411B_ABST
    Figure CN115514411B_ABST
Patent Text Reader

Abstract

The present invention relates to a time-domain optical signal processing method and system based on an optical neural network, wherein the method comprises: obtaining a confirmation signal that an optical path has been accurately connected and an encoded mixed time-domain optical signal; constructing an optical neural network and training it, wherein the optical neural network comprises a plurality of intermediate layers for extracting the optical characteristics of the input mixed time-domain optical signal and classifying it, wherein the structure of the intermediate layer comprises an optical phase modulator, a first dispersion module, a nonlinear activation module, and a second dispersion module connected in sequence; training the optical neural network to determine the adjustment parameters of the optical phase modulator; aligning the mixed time-domain optical signal with the optical phase modulator, adjusting the optical phase modulator so that its modulation phase is consistent with the training result, completing the recognition of the time-domain signal, and obtaining a classified time-domain optical signal. Compared with the prior art, the present invention has the advantages of being applicable to time-domain optical signal processing and having a fast recognition speed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of optoelectronic technology, and in particular to a time-domain optical signal processing method and system based on optical neural network. Background Art

[0002] There are many different device implementation methods for optical neural networks, such as optical fiber, Mach-Zehnder interferometer, integrated circuit optical chip, D2NN structure, etc. The optical neural networks basically convert the spatial domain image data into light intensity or optical power input network for processing. There is no solution in the existing solution to process the time domain optical signal commonly used in optical communication systems.

[0003] Comparative document 1 "All-Optical Machine Learning Using Diffractive Deep Neural Networks" (Lin X, Rivenson Y, Yardimci NT, et al. Science (361) 6406 (2018) 1004–1008.) pioneered an all-optical neural network, using the diffraction principle as the network's forward transmission equation and using 3D printing technology to obtain a diffraction mask layer. Each mask constitutes a single layer, and the neurons are small grids with different transmittance, refractive index or mask thickness. Each neuron in the upper layer is connected to all neurons in the next layer through diffraction. The weights are changed by adjusting different phases or amplitudes, and the classification results of the network are obtained by measuring the light intensity with a detector array in the output layer. The input signal processed by the network is an image in the spatial domain, and once the network is trained, it is solidified through 3D printing, is not reconstructible, is not flexible enough, and the network does not have a nonlinear activation function.

[0004] Comparative document 2 "Class-specific differential detection in diffractive optical neural networks improves inference accuracy" (Li J, Mengu D, Luo Y, et al. Advanced Photonics [J], 2019, 1 (4).) proposed a differential task-specific D2NN structure, which increased the original 10 detectors to 10 positive and negative detectors, and replaced the original mean square error loss function with a softmax cross entropy loss function, effectively improving the recognition accuracy of the D2NN structure, which is highly close to the recognition accuracy of electronic neural networks. However, the input data processed by this network is grayscale images, the number of network parameters is too large, training takes a lot of time, and it is still not flexible enough and has no scalability.

[0005] Comparative document 3 "Neural networks within multi-core optic fibers" (Cohen E, Malka D, Shemer A, et al. Scientific Reports, 2016, 6 (1): 29080.) proposes an optical neural network based on multi-core optical fibers, which is simulated using numerical simulation and demonstrates the ability to distinguish different input modes. The network consists of multi-core optical fibers, EDFA, and Y-couplers, and the optical signal is transmitted through the coupling of the light-guiding silicon core. The network inputs the amplified signal power, and the output optical signal is measured by the detector. The neuron is implemented with a single silicon core, and the synaptic weight is implemented by the erbium-doped amplifier in the core. Through computer learning, the learning process is to pass the output signal through a comparator, update the weight of the erbium-doped amplifier light emitter in the control core stored in the memory array, and feed it back to the input. The input signal processed by the network is the optical power in the spatial domain, which requires the preparation of a multi-core optical fiber with a long length and a large core spacing. The manufacturing process is difficult, and the network does not have a nonlinear activation function, and it is also difficult to couple the light source to the multi-core optical fiber.

[0006] Comparative document 4 "Optical neural networks based on optical fiber-communication system" (Cheng Tian-You, Chou Da-Ya, Liu Ching-Chuan, et al. Science Direct [J] Neurocomputing, 2019, 364: 239-244.) proposes an optical neural network based on directional couplers, optical fibers, and optical amplifiers. The network structure is a reservoir calculation, which is a cyclic system composed of time finite element internal states and is perturbed by temporary external inputs. The system consists of an input layer, a storage layer, and an output layer. The input data is a time-correlated optical signal of a laser passing through a Mach-Zehnder modulator loaded with an electrical signal. The input weight matrix of the input layer consists of directional couplers. The neuron consists of two optical couplers, an EDFA, and a 53-meter-long optical fiber. The nonlinear activation function is generated using a full optical ring formed by an optical fiber and an erbium-doped fiber amplifier (EDFA). The readout matrix of the output layer is generated using a Mach-Zehnder phase modulator and a directional coupler. Simulation results show that the system can distinguish between rectangular waves and triangular waves of the input optical signal. The network uses 53-meter-long optical fiber, which requires too much space and is not suitable for miniaturization and integration, so its feasibility is low.

[0007] The above-mentioned optical fiber-based optical neural network devices are relatively complex, with high process difficulty, poor flexibility, lack of scalability, and basically process input signals in the spatial domain. Summary of the invention

[0008] The purpose of the present invention is to provide a time-domain optical signal processing method and system based on optical neural network, which can effectively and quickly process time-domain optical signals, improve the capacity of optical neural network to process data, and achieve seamless connection with optical communication system.

[0009] The purpose of the present invention can be achieved by the following technical solutions:

[0010] A time-domain optical signal processing method based on an optical neural network comprises the following steps:

[0011] Obtain a confirmation signal that the optical path is accurately connected;

[0012] obtaining an encoded hybrid time-domain optical signal;

[0013] Constructing and training an optical neural network, wherein the optical neural network includes a plurality of intermediate layers for extracting optical features of an input mixed time-domain optical signal and classifying the signals, wherein the structure of the intermediate layers includes an optical phase modulator, a first dispersion module, a nonlinear activation module, and a second dispersion module connected in sequence;

[0014] Train an optical neural network and use a back-propagation algorithm to determine the tuning parameters of an optical phase modulator;

[0015] Align the mixed time-domain optical signal with the optical phase modulator in the middle layer to input the signal into the trained optical neural network, adjust the optical phase modulator so that its modulation phase is consistent with the training result, complete the recognition of the time-domain signal, and obtain the classified time-domain optical signal.

[0016] The optical phase modulator loads weights by modulating the phase of the optical signal, encoding the network neuron weights onto the phase of the input optical signal:

[0017]

[0018] in, is the modulation phase, U in (t) is the input mixed time-domain optical signal, is the output signal of the optical phase modulator.

[0019] The optical signal undergoes dispersion transmission through the dispersion module;

[0020]

[0021] in, is the Fourier transform of the time-domain optical signal, is the input signal of the dispersion module, is the output signal of the dispersion module, D is the dispersion coefficient, in ps2 .

[0022] The optical signal with dispersion and phase weight is transformed nonlinearly through the nonlinear activation module to realize the nonlinear activation function:

[0023]

[0024] Among them, γ is the nonlinear coefficient, is the output signal of the nonlinear activation module, is the input signal of the nonlinear activation module.

[0025] The nonlinear activation module is composed of a nonlinear optical fiber or a semiconductor optical amplifier.

[0026] A time-domain optical signal processing system based on an optical neural network, comprising:

[0027] A transmitting module, used for acquiring a confirmation signal sent by the optical path accurate connectivity confirmation module and transmitting an encoded mixed time-domain optical signal;

[0028] An optical signal processing module based on an optical neural network is used to perform the following steps: obtaining an encoded mixed time-domain optical signal; constructing an optical neural network and training it, wherein the optical neural network includes a plurality of intermediate layers, which are used to extract the optical characteristics of the input mixed time-domain optical signal and classify it, wherein the structure of the intermediate layer includes an optical phase modulator, a first dispersion module, a nonlinear activation module, and a second dispersion module connected in sequence; training the optical neural network, and using a back-propagation algorithm to determine the adjustment parameters of the optical phase modulator; aligning the mixed time-domain optical signal with the optical phase modulator of the intermediate layer, realizing inputting the signal into the trained optical neural network, and adjusting the optical phase modulator so that its modulation phase is consistent with the training result;

[0029] A receiving module, used to receive the output signal of the optical signal processing module, complete the recognition of the time domain signal, and obtain the classified time domain optical signal;

[0030] The optical path accurate connectivity confirmation module is used to confirm the positions of the transmitting module, the optical neural network and the receiving module to ensure accurate connectivity of the optical path.

[0031] The optical phase modulator loads weights by modulating the phase of the optical signal, encoding the network neuron weights onto the phase of the input optical signal:

[0032]

[0033] in, is the modulation phase, U in (t) is the input mixed time-domain optical signal, is the output signal of the optical phase modulator.

[0034] The optical signal undergoes dispersion transmission through the dispersion module;

[0035]

[0036] in, is the Fourier transform of the time-domain optical signal, is the input signal of the dispersion module, is the output signal of the dispersion module, D is the dispersion coefficient, in ps 2 .

[0037] The optical signal with dispersion and phase weight is transformed nonlinearly through the nonlinear activation module to realize the nonlinear activation function:

[0038]

[0039] Among them, γ is the nonlinear coefficient, is the output signal of the nonlinear activation module, is the input signal of the nonlinear activation module.

[0040] The nonlinear activation module is composed of a nonlinear optical fiber or a semiconductor optical amplifier.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] (1) Compared with the prior art which can only process optical signals in the spatial domain, the input signal of the present invention is a time domain signal, which can process time domain optical signals, thereby achieving seamless connection with the optical communication system and processing optical communication tasks.

[0043] (2) The present invention uses an optical neural network to process time domain signals, which effectively expands the system's data processing capacity, requires fewer training parameters, and has a fast training speed, which can improve signal recognition speed.

[0044] (3) The optical neural network device of the present invention has a simple structure, low process complexity, and is easy to implement and integrate.

[0045] (4) The present invention can flexibly adjust the parameters of the optical neural network, has good flexibility, can be applied to most optical signal processing scenarios, and has strong scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a schematic diagram of the model structure of the optical neural network;

[0047] Figure 2 It is a schematic diagram of the system structure of the present invention;

[0048] Figure 3 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0049] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0050] Example 1

[0051] A time-domain optical signal processing system based on optical neural network, such as Figure 2 As shown, including:

[0052] 1) A transmitting module, used to obtain a confirmation signal sent by the optical path accurate connectivity confirmation module and transmit an encoded mixed time-domain optical signal;

[0053] The transmitting module transmits an encoded time domain signal, and compared with the conventional optical neural network which usually inputs image signals, the data complexity and capacity of the network are greatly increased.

[0054] 2) An optical signal processing module based on an optical neural network, configured to perform the following steps:

[0055] obtaining an encoded hybrid time-domain optical signal;

[0056] Build and train fiber-based optical neural networks;

[0057] The optical neural network includes a plurality of intermediate layers for extracting the optical characteristics of the input mixed time-domain optical signal and classifying it, wherein the structure of the intermediate layer includes an optical phase modulator, a first dispersion module, a nonlinear activation module, and a second dispersion module connected in sequence, such as Figure 1 As shown,

[0058] ①Optical phase modulator

[0059] The optical phase modulator loads weights by modulating the phase of the optical signal, encoding the network neuron weights onto the phase of the input optical signal:

[0060]

[0061] in, is the modulation phase, U in (t) is the input mixed time-domain optical signal, is the output signal of the optical phase modulator.

[0062] ②Dispersion module

[0063] The dispersion module is composed of common optical fiber or other dispersion devices, such as Bragg grating, chirped grating, etc. In this embodiment, common optical fiber is used.

[0064] The optical signal undergoes dispersion transmission through the dispersion module;

[0065]

[0066] in, is the Fourier transform of the time-domain optical signal, is the input signal of the dispersion module, is the output signal of the dispersion module, D is the dispersion coefficient, in ps 2 .

[0067] ③Non-linear activation module

[0068] The nonlinear activation module is composed of nonlinear media such as nonlinear optical fiber or semiconductor optical amplifier. In this embodiment, a high nonlinear optical fiber HNLF is used, and the high nonlinear optical fiber HNLF produces a strong Kerr effect, causing the light intensity of the optical signal to undergo nonlinear transformation.

[0069] The optical signal with dispersion and phase weight is transformed nonlinearly through the nonlinear activation module to realize the nonlinear activation function:

[0070]

[0071] Among them, γ is the nonlinear coefficient, is the output signal of the nonlinear activation module, is the input signal of the nonlinear activation module.

[0072] Train an optical neural network and use a back-propagation algorithm to determine the tuning parameters of an optical phase modulator;

[0073] The mixed time-domain optical signal is aligned with the optical phase modulator in the middle layer to input the signal into the trained optical neural network, and the optical phase modulator is adjusted so that its modulation phase is consistent with the training result.

[0074] In this embodiment, the number of neurons in the input layer of the optical neural network is 128, the number of neurons in the middle layer is 5 layers, the number of neurons in each layer is 128, and the number of neurons in the output layer is 128. The phase adjustment parameters of the optical phase modulator correspond to the training results one by one.

[0075] 3) A receiving module is used to receive the output signal of the optical signal processing module, complete the recognition of the time domain signal, and obtain the classified time domain optical signal. In this embodiment, the receiving module is a receiver of the optical communication system, and the photodiode converts the optical signal into an electrical signal and outputs it to obtain the recognition result.

[0076] 4) An optical path accurate connection confirmation module is used to confirm the positions of the transmitting module, the optical neural network and the receiving module to ensure accurate connection of the optical path. In this embodiment, the relative positions of the transmitting module, the optical fiber, the optical phase modulator, the highly nonlinear optical fiber and the receiving module need to be adjusted.

[0077] Example 2

[0078] A time-domain optical signal processing method based on optical neural network, such as Figure 3 As shown, the system for implementing the embodiment 1 includes the following steps:

[0079] S1, obtaining a confirmation signal that the optical path is accurately connected; after obtaining the signal, executing subsequent steps;

[0080] S2, obtaining an encoded mixed time-domain optical signal;

[0081] S3, constructing an optical neural network based on optical fiber and training it, wherein the optical neural network includes multiple intermediate layers for extracting optical features of the input mixed time-domain optical signal and classifying it, wherein the structure of the intermediate layer includes an optical phase modulator, a first dispersion module, a nonlinear activation module, and a second dispersion module connected in sequence;

[0082] S4, training the optical neural network and using a back-propagation algorithm to determine the adjustment parameters of the optical phase modulator;

[0083] S5. Align the mixed time-domain optical signal with the optical phase modulator in the middle layer to input the signal into the trained optical neural network, adjust the optical phase modulator so that its modulation phase is consistent with the training result, complete the recognition of the time-domain signal, and obtain the classified time-domain optical signal.

[0084] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning, or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A time-domain optical signal processing method based on optical neural network, It is characterized in that The following steps are involved: Obtain a confirmation signal that the optical path is accurately connected; obtaining an encoded hybrid time-domain optical signal; Constructing and training an optical neural network, wherein the optical neural network includes a plurality of intermediate layers for extracting optical features of an input mixed time-domain optical signal and classifying the signals, wherein the structure of the intermediate layers includes an optical phase modulator, a first dispersion module, a nonlinear activation module, and a second dispersion module connected in sequence; Train an optical neural network and use a back-propagation algorithm to determine the tuning parameters of an optical phase modulator; Align the mixed time-domain optical signal with the optical phase modulator in the middle layer, input the signal into the trained optical neural network, adjust the optical phase modulator so that its modulation phase is consistent with the training result, complete the recognition of the time-domain signal, and obtain the classified time-domain optical signal; The optical phase modulator loads weights by modulating the phase of the optical signal, encoding the network neuron weights onto the phase of the input optical signal: in, is the modulation phase, U in (t) is the input mixed time-domain optical signal, is the output signal of the optical phase modulator; The optical signal is transmitted through the dispersion module; in, is the Fourier transform of the time-domain optical signal, is the input signal of the dispersion module, is the output signal of the dispersion module, D is the dispersion coefficient, in ps 2 ; The optical signal with dispersion and phase weight is transformed nonlinearly through the nonlinear activation module to realize the nonlinear activation function: Among them, γ is the nonlinear coefficient, is the output signal of the nonlinear activation module, is the input signal of the nonlinear activation module.

2. According to claim 1, a time-domain optical signal processing method based on optical neural network, It is characterized in that The nonlinear activation module is composed of a nonlinear optical fiber or a semiconductor optical amplifier.

3. A time-domain optical signal processing system based on optical neural network, It is characterized in that include: A transmitting module, used for acquiring a confirmation signal sent by the optical path accurate connectivity confirmation module and transmitting an encoded mixed time-domain optical signal; An optical signal processing module based on an optical neural network is used to perform the following steps: obtaining an encoded mixed time-domain optical signal; constructing an optical neural network and training it, wherein the optical neural network includes a plurality of intermediate layers, which are used to extract the optical characteristics of the input mixed time-domain optical signal and classify it, wherein the structure of the intermediate layer includes an optical phase modulator, a first dispersion module, a nonlinear activation module, and a second dispersion module connected in sequence; training the optical neural network, and using a back-propagation algorithm to determine the adjustment parameters of the optical phase modulator; aligning the mixed time-domain optical signal with the optical phase modulator of the intermediate layer, realizing inputting the signal into the trained optical neural network, and adjusting the optical phase modulator so that its modulation phase is consistent with the training result; A receiving module, used to receive the output signal of the optical signal processing module, complete the recognition of the time domain signal, and obtain the classified time domain optical signal; The optical path accurate connection confirmation module is used to confirm the position of the transmitting module, the optical neural network and the receiving module to ensure the accurate connection of the optical path; The optical phase modulator loads weights by modulating the phase of the optical signal, encoding the network neuron weights onto the phase of the input optical signal: in, is the modulation phase, U in (t) is the input mixed time-domain optical signal, is the output signal of the optical phase modulator; The optical signal is transmitted through the dispersion module; in, is the Fourier transform of the time-domain optical signal, is the input signal of the dispersion module, is the output signal of the dispersion module, D is the dispersion coefficient, in ps 2 ; The optical signal with dispersion and phase weight is transformed nonlinearly through the nonlinear activation module to realize the nonlinear activation function: Among them, γ is the nonlinear coefficient, is the output signal of the nonlinear activation module, is the input signal of the nonlinear activation module.

4. The time-domain optical signal processing system based on optical neural network according to claim 3, It is characterized in that The nonlinear activation module is composed of a nonlinear optical fiber or a semiconductor optical amplifier.