An all-optical convolution computing system based on multi-core fiber

Abstract

The application belongs to the technical field of optical information computing, and particularly relates to a kind of all-optical convolution computing systems based on multi-core optical fiber, including optical pulse source modulation module, image input module, multi-core optical fiber convolution kernel module and optical accumulation module, optical pulse source modulation module, image input module, multi-core optical fiber convolution kernel module and optical accumulation module are communicatively connected;Optical pulse source modulation module sends optical pulse to the programmed convolution kernel in multi-core optical fiber convolution kernel module, and image pixels are sequentially input into image input module and then enter multi-core optical fiber convolution kernel module, and convolution kernel carries out weighted multiplication on each pixel in image input module, and then carries out addition through optical accumulation module.The base in the application adopts multi-core optical fiber, the signals transmitted in the fiber core do not interfere with each other, and can be used in harsh environmental conditions such as electromagnetic interference and extreme temperature;In addition, it can be directly coupled with the existing optical communication network without alignment problem.

Description

Technical Field

[0001] The present invention belongs to the technical field of optical information computing, and in particular relates to an all-optical convolution computing system based on multi-core optical fibers. Background Art

[0002] The all-optical convolutional neural network (CPNN) architecture based on optical convolution kernels typically has fixed weights that do not require real-time updates, and can easily and passively perform weighted multiplication and accumulation through phase modulation or wavelength division multiplexing. Once the weights of the all-optical convolutional neural network (CPNN) are set, the network's latency is only limited by the speed of light and the final optoelectronic conversion output. Compared with traditional electrical convolutional neural networks, all-optical convolutional neural networks (CPNN) have advantages such as high modulation rate, low power consumption, and large channel parallelism. In the all-optical convolutional neural network (CPNN), the input signal is individually weighted by nodes with different weights. Multiple nodes can form a convolution kernel, and the convolution kernel can be designed with various weights to calculate the input image, such as extracting edge features, filtering, and sharpening.

[0003] The present invention provides an all-optical convolution computing system based on multi-core optical fiber. Non-volatile phase-change material film layers and anti-oxidation film layers are sequentially deposited on the multi-core optical fiber to construct a convolution kernel. Laser pulses of different powers are injected into each fiber core to change the state of the phase-change material film layer, i.e., programming the convolution kernel. The different refractive indices and absorption coefficients of the phase-change materials in different phases result in different degrees of weighted multiplication being performed on the light passing through. The light is then accumulated through an optical fiber coupler to realize all-optical convolution computing. Summary of the Invention

[0004] The purpose of the present invention is to provide an all-optical convolution computing system based on multi-core optical fiber, in which a non-volatile phase change material film layer and an anti-oxidation film layer are sequentially deposited on the multi-core optical fiber to construct a convolution kernel; laser pulses of different powers are injected into each fiber core to change the state of the phase change material film layer, that is, to program the convolution kernel; the different refractive indices and absorption coefficients of the phase change materials in different phases cause the passing light to be subjected to different degrees of weighted multiplication, and then accumulated through a fiber coupler to realize all-optical convolution computing.

[0005] The technical solutions adopted by the present invention are as follows:

[0006] An all-optical convolution computing system based on multi-core optical fiber includes an optical pulse source modulation module, an image input module, a multi-core optical fiber convolution kernel module and an optical accumulation module. The optical pulse source modulation module, the image input module, the multi-core optical fiber convolution kernel module and the optical accumulation module are communicatively connected.

[0007] Preferably, the optical pulse source modulation module includes a pulse laser; the three pulse lasers program the convolution kernel in the multi-core optical fiber convolution kernel module by emitting pulse sequences of different powers.

[0008] Preferably, the power of the light pulses emitted by the three pulse lasers is adjustable, and the power corresponds one-to-one to the weight of the convolution kernel in the multi-core optical fiber convolution kernel module.

[0009] Preferably, the image input module includes a broadband light source, an n-channel demultiplexer, an adjustable attenuator, a 2×2 fiber coupler, a fiber isolator, an n-channel optical wavelength division multiplexer, and a multi-core fiber fan-in unit.

[0010] Preferably, the n-channel demultiplexer is capable of decomposing the mixed optical signal output by the broadband light source into multiple optical wavelength signals, and its n output ports correspondingly output n different wavelengths, and the number of ports n is the same as the number of the three pulse lasers, the three adjustable attenuators, and the three 2×2 fiber couplers;

[0011] The two input ports of the three 2×2 fiber couplers are connected to the three pulse lasers and the three adjustable attenuators respectively;

[0012] One output port of the three 2×2 optical fiber couplers is directly connected to the multi-core optical fiber fan-in unit, and the other output port is respectively connected to the three first optical fiber isolators, coupled through the first n-channel optical wavelength division multiplexer, and then input into the multi-core optical fiber fan-in unit as a reference path;

[0013] The first n-channel optical wavelength division multiplexer is capable of multiplexing n-channel optical wavelength signals into one channel for propagation, and the wavelength and number n correspond to those of the n-channel demultiplexer;

[0014] The input channels of the multi-core fiber fan-in unit correspond one-to-one to the cores of the multi-core fiber convolution core module;

[0015] The three first optical fiber isolators only allow light to pass in one direction.

[0016] Preferably, the multi-core optical fiber convolution core module includes four multi-core optical fiber convolution cores of a multi-core optical fiber, and each of the multi-core optical fiber convolution cores includes a fiber core, a phase change material film layer and an anti-oxidation film layer;

[0017] The phase change material film layer is a sulfur compound;

[0018] The phase change material film layer has at least two phases that can be adjusted by light pulses;

[0019] The material of the anti-oxidation film is indium tin oxide.

[0020] Preferably, the optical accumulation module includes a multi-core optical fiber fan-out unit, three second optical fiber isolators, a second n-channel optical wavelength division multiplexer, a photoelectric balance detector, a data acquisition card, and a host computer;

[0021] The output channels of the multi-core fiber fan-out unit correspond one-to-one to the cores of the multi-core fiber convolution core module.

[0022] Preferably, the second optical fiber isolator only allows light to pass in one direction;

[0023] The number of ports n of the second n-channel optical wavelength division multiplexer is one less than the number of output channels of the multi-core fiber fan-out unit;

[0024] The photoelectric balance detector has a positive input port and a negative input port, and the output value of the photoelectric balance detector depends on the difference between the light intensity input at the positive input port and the light intensity at the negative input port;

[0025] The positive input port of the photoelectric balanced detector is connected to the output end of the second n-channel optical wavelength division multiplexer;

[0026] The negative input port of the photoelectric balanced detector is connected to the output end of the multi-core fiber fan-out unit corresponding to the core of the multi-core fiber convolution core module input by the first n-channel optical wavelength division multiplexer in the image input module;

[0027] The photoelectric balance detector converts the input light signal into an electrical signal and transmits it to the data acquisition card, which then transmits the data to the host computer for monitoring and display.

[0028] The host computer interacts with the data acquisition card.

[0029] The technical effects achieved by the present invention are:

[0030] 1. The convolution calculation performed in this invention is entirely optical, thus having the transmission speed of light waves, which significantly improves the calculation speed compared to electrical convolution calculation schemes. The wide bandwidth of lasers allows light waves to carry more information, which is used for parallel convolution calculations.

[0031] 2. The convolution kernel in the present invention uses non-volatile phase-change materials, so no additional energy is required to maintain the state when performing the convolution operation, resulting in zero power loss;

[0032] 3. The substrate of the present invention uses a multi-core optical fiber, and the signals transmitted in the fiber cores will not interfere with each other, and can be used in harsh environmental conditions such as electromagnetic interference and extreme temperatures; in addition, it can be directly coupled with the existing optical communication network without alignment issues.

[0033] 4. A convolution kernel is constructed by sequentially depositing non-volatile phase-change material layers and anti-oxidation film layers on a multi-core optical fiber. Laser pulses of varying powers are injected into each fiber core to alter the state of the phase-change material layer, thereby programming the convolution kernel. The different refractive indices and absorption coefficients of the phase-change materials in different phases result in varying degrees of weighted multiplication of the light passing through them. The light is then accumulated through a fiber coupler, achieving all-optical convolution calculations. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A schematic diagram of an all-optical convolution calculation solution based on multi-core optical fiber provided by the present invention;

[0035] Figure 2 Schematic diagram of the all-optical convolution calculation example based on multi-core optical fiber provided by the present invention.

[0036] In the accompanying drawings, the components represented by the reference numerals are as follows:

[0037] Optical pulse source modulation module; 2. Image input module; 3. Multi-core fiber convolution kernel module; 4. Optical accumulation module; 101. Pulsed laser; 102. Pulsed laser; 103. Pulsed laser; 201. Broad spectrum light source; 202. N-channel demultiplexer; 203. Adjustable attenuator; 204. Adjustable attenuator; 205. Adjustable attenuator; 206. 2×2 fiber coupler; 207. 2×2 fiber coupler; 208. 2×2 fiber coupler; 209. Fiber isolator; 210. Fiber isolator; 211. Fiber isolator; 212. First n-channel optical wavelength division multiplexer; 213, multi-core fiber fan-in unit; 301, multi-core fiber convolution core; 302, multi-core fiber convolution core; 303, multi-core fiber convolution core; 304, multi-core fiber convolution core; 3011, fiber core; 3012, phase change material film layer; 3013, anti-oxidation film layer; 401, multi-core fiber fan-out unit; 402, second optical fiber isolator; 403, second optical fiber isolator; 404, second optical fiber isolator; 405, second n-channel optical wavelength division multiplexer; 406, photoelectric balance detector; 407, data acquisition card; 408, host computer. DETAILED DESCRIPTION

[0038] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is described in detail below with reference to the following examples. It should be understood that the following text is only used to describe one or more specific embodiments of the present invention and does not strictly limit the scope of protection of the present invention.

[0039] like Figure 1As shown, an all-optical convolution computing system based on multi-core optical fiber includes an optical pulse source modulation module 1, an image input module 2, a multi-core optical fiber convolution kernel module 3 and an optical accumulation module 4. The optical pulse source modulation module 1, the image input module 2, the multi-core optical fiber convolution kernel module 3 and the optical accumulation module 4 are communicatively connected.

[0040] The optical pulse source modulation module 1 sends an optical pulse to the multi-core optical fiber convolution kernel module 3 to program the convolution kernel. The image pixels are input into the image input module 2 in turn and then enter the multi-core optical fiber convolution kernel module 3. The convolution kernel performs weighted multiplication on each pixel in the image input module 2, and then adds them through the optical accumulation module 4.

[0041] The optical pulse source modulation module 1 includes pulse lasers 101, 102, and 103; the three pulse lasers 101, 102, and 103 program the convolution kernel in the multi-core optical fiber convolution kernel module 3 by emitting pulse sequences of different powers.

[0042] Among them, the power of the light pulses emitted by the three pulse lasers 101, 102, and 103 is adjustable, and the power corresponds to the weight of the convolution kernel in the multi-core optical fiber convolution kernel module 3 one by one.

[0043] The image input module 2 includes a broadband light source 201, an n-channel demultiplexer 202, adjustable attenuators 203, 204, 205, 2×2 fiber couplers 206, 207, 208, fiber isolators 209, 210, 211, a first n-channel optical wavelength division multiplexer 212, and a multi-core fiber fan-in unit 213.

[0044] A broadband light source 201, an n-channel demultiplexer 202, and adjustable attenuators 203, 204, and 205 are used to convert image pixel information into light intensity information. The image pixel size corresponds to the light intensity.

[0045] Specifically, the broadband light output by the broadband light source 201 is split into n signal lights of different wavelength bands by the n-channel demultiplexer 202, and the light intensities are then attenuated to different degrees by the adjustable attenuators 203, 204, and 205.

[0046] The attenuation strength of the adjustable attenuators 203 , 204 , and 205 is controlled by the host computer 408 , and the degree of light attenuation depends on the pixel size of the image being calculated.

[0047] The n-channel demultiplexer 202 is capable of decomposing the mixed optical signal output by the broadband light source 201 into multiple optical wavelength signals. Its n output ports correspond to outputting n different wavelengths. The number of ports n is the same as the number of the three pulse lasers 101, 102, 103, the three adjustable attenuators 203, 204, 205, and the three 2×2 fiber couplers 206, 207, 208.

[0048] The two input ports of the three 2×2 fiber couplers 206 , 207 , and 208 are connected to the three pulse lasers 101 , 102 , and 103 and the three adjustable attenuators 203 , 204 , and 205 , respectively;

[0049] One output port of the three 2×2 fiber couplers 206, 207, and 208 is directly connected to the multi-core fiber fan-in unit 213, and the other output ports are respectively connected to the three first fiber isolators 209, 210, and 211, and then coupled through the first n-channel optical wavelength division multiplexer 212 before being input into the multi-core fiber fan-in unit 213 for use as a reference path;

[0050] The first n-channel optical wavelength division multiplexer 212 is capable of multiplexing n-channel optical wavelength signals into one channel for propagation, and the wavelength and number n correspond to the n-channel demultiplexer 202;

[0051] The input channels of the multi-core fiber fan-in unit 213 correspond one-to-one to the cores of the multi-core fiber convolution core module 3, which is described in detail as one of the input channels of the multi-core fiber fan-in unit 213 passing through only one core of the multi-core fiber convolution core module 3;

[0052] The three first optical fiber isolators 209 , 210 , and 211 only allow light to pass in one direction.

[0053] Among them, the multi-core optical fiber convolution core module 3 includes four multi-core optical fiber convolution cores 301, 302, 303, 304 of the multi-core optical fiber, and each multi-core optical fiber convolution core 301, 302, 303, 304 includes a fiber core, a phase change material film layer and an anti-oxidation film layer;

[0054] For example, the multi-core optical fiber convolution core 301 includes a fiber core 3011, a phase change material film layer 3012, and an anti-oxidation film layer 3013;

[0055] The number of multi-core fiber convolution kernels 301, 302, 303, 304 is one more than the number of ports n of the first n-channel optical wavelength division multiplexer 212;

[0056] n of the multi-core fiber convolution kernels 301, 302, 303, and 304 are used as convolution kernels, and the weighted state of the remaining convolution kernel is fixed to an intermediate state as a reference optical path;

[0057] The phase change material film layer 3012 is a chalcogenide compound containing at least two or more elements of Ge, Sb, and Te, such as germanium antimony tellurium alloy Ge2Sb2Te5, germanium antimony selenium tellurium alloy Gs2Sb2Se4Te1, etc.;

[0058] The phase change material film layer 3012 has at least two phases that can be modulated by light pulses: a crystalline state and an amorphous state. There is also an intermediate state between the crystalline state and the amorphous state. Different phases have different refractive indices and light absorption rates. As a result, when the multi-core optical fiber convolution kernels 301, 302, 303, and 304 are in the crystalline state, the weighting coefficient is small, that is, the transmittance to light is low. When the multi-core optical fiber convolution kernels 301, 302, 303, and 304 are in the amorphous state, the weighting coefficient is large, that is, the transmittance to light is high.

[0059] The phase change material film layer 3012 is combined with the optical fiber by radio frequency magnetron sputtering. The thickness of the sputtered phase change material film layer 3012 is controlled by setting the voltage, argon concentration, and sputtering time during sputtering.

[0060] When the phase change material film layer 3012 is located within 0.5µm of the fiber core 3011, the modulation effect of the optical pulse in the fiber core is optimal, that is, the energy of the optical pulse regulates the state of the phase change material film layer 3012 through evanescent field coupling.

[0061] The anti-oxidation film 3013 is made of indium tin oxide (ITO) to prevent the optical phase change material film layer 3012 from being oxidized by air.

[0062] The anti-oxidation film 3013 is deposited on the phase change material film layer 3012 by radio frequency magnetron sputtering;

[0063] The weight states of the multi-core optical fiber convolution kernels 301, 302, 303, and 304 are adjusted by light pulses. When the phase change material film layer 3012 is in a completely crystalline state, the weighting coefficient is the smallest and the transmittance to light is the lowest. When the phase change material film layer 3012 is in a completely amorphous state, the weighting coefficient is the largest and the transmittance to light is the highest.

[0064] The optical accumulation module 4 includes a multi-core optical fiber fan-out unit 401, three second optical fiber isolators 402, 403, 404, a second n-channel optical wavelength division multiplexer 405, a photoelectric balance detector 406, a data acquisition card 407, and a host computer 408;

[0065] The output channels of the multi-core fiber fan-out unit 401 correspond one-to-one to the cores of the multi-core fiber convolution core module 3. It is described in detail as one of the output channels of the multi-core fiber fan-out unit 401 only outputs the optical signal in one core of the multi-core fiber convolution core module 3.

[0066] The second optical fiber isolators 402, 403, and 404 only allow light to pass in one direction.

[0067] The number of ports n of the second n-channel optical wavelength division multiplexer 405 is one less than the number of output channels of the multi-core fiber fan-out unit 401;

[0068] The second n-channel optical wavelength division multiplexer 405 is used for accumulating the weighted detection light intensity of the n convolution kernels in the multi-core fiber convolution kernel module 3;

[0069] The photoelectric balance detector 406 has a positive input port and a negative input port, and the output value of the photoelectric balance detector 406 depends on the difference between the light intensity input by the positive input port and the light intensity of the negative input port;

[0070] The positive input port of the photoelectric balance detector 406 is connected to the total light intensity output by the second n-channel optical wavelength division multiplexer 405;

[0071] The negative input port of the photoelectric balance detector 406 is connected to the output end of the multi-core fiber fan-out unit 401 corresponding to the fiber core of the multi-core fiber convolution kernel module 3 input by the multi-core fiber fan-in unit 213 in the image input module 2;

[0072] The photoelectric balance detector 406 converts the input optical signal into an electrical signal and transmits it to the data acquisition card 407, and then transmits the data to the host computer 408 through the data acquisition card 407 for monitoring and display;

[0073] The host computer 408 interacts with the data acquisition card 407, programs the function of the data acquisition card 407, and obtains the data transmitted by the data acquisition card 407;

[0074] The host computer 408 can display the pixels of the image, and control the attenuation degree of the adjustable attenuators 203, 204 and 205 of the image input module 2 by the pixel value, and the number of pixels at each time step is consistent with the number of adjustable attenuators 203, 204 and 205.

[0075] The working principle of the application is that, for example, the constructed multi-core fiber-based all-optical convolution calculation scheme is used to realize convolution calculation on a handwritten character "7" to extract edge features.

[0076] As Figure 2As shown in the figure, it is a schematic diagram of the example of extracting the "bottom", "top", "right" and "left" edge features of the handwritten character "7" by convolution calculation; a five-core optical fiber is selected for demonstration, in which one core is used as a reference path, the phase change material near it is pre-modulated to an intermediate state, and the remaining four cores are used as convolution kernels to form a 2×2 matrix; the pulse laser of the optical pulse source modulation module 1 is controlled to emit pulse sequences of different powers to program the 2×2 convolution kernel in the five-core optical fiber convolution kernel module 3; the high-energy pulse emitted by the pulse laser can modulate the convolution kernel to a high-weight state 1, and the low-energy pulse energy It is enough to modulate the convolution kernel to a low weight state (0); for extracting the "lower" edge feature of the handwritten character "7", the state of the 2×2 convolution kernel needs to be modulated to "00;11"; for extracting the "upper" edge feature of the handwritten character "7", the state of the 2×2 convolution kernel needs to be modulated to "11;00"; for extracting the "right" edge feature of the handwritten character "7", the state of the 2×2 convolution kernel needs to be modulated to "01;01"; for extracting the "left" edge feature of the handwritten character "7", the state of the 2×2 convolution kernel needs to be modulated to "10;10".

[0077] The foregoing is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained herein shall, unless otherwise specified or limited, be implemented in accordance with conventional means in the art.

Claims

1. An all-optical convolution computing system based on multi-core optical fiber, characterized by: It comprises an optical pulse source modulation module (1), an image input module (2), a multi-core optical fiber convolution kernel module (3) and an optical accumulation module (4), wherein the optical pulse source modulation module (1), the image input module (2), the multi-core optical fiber convolution kernel module (3) and the optical accumulation module (4) are communicatively connected; The optical pulse source modulation module (1) comprises pulse lasers (101, 102, 103); the three pulse lasers (101, 102, 103) program the convolution kernel in the multi-core optical fiber convolution kernel module (3) by emitting pulse sequences of different powers; The power of the light pulses emitted by the three pulse lasers (101, 102, 103) is adjustable, and the power corresponds to the weight of the convolution kernel in the multi-core optical fiber convolution kernel module (3) in a one-to-one manner; The image input module (2) includes a broadband light source (201), an n-channel demultiplexer (202), an adjustable attenuator (203, 204, 205), a 2×2 optical fiber coupler (206, 207, 208), a first optical fiber isolator (209, 210, 211), a first n-channel optical wavelength division multiplexer (212), and a multi-core optical fiber fan-in unit (213); The number n of ports of the n-channel demultiplexer (202) is the same as the number of the three pulse lasers (101, 102, 103), the three adjustable attenuators (203, 204, 205), and the three 2×2 fiber couplers (206, 207, 208); The n-channel demultiplexer (202) is capable of decomposing the mixed optical signal output by the wide-spectrum light source (201) into multi-path optical wavelength signals, and its n output ports correspondingly output n different wavelengths; The two input ports of the three 2×2 fiber couplers (206, 207, 208) are respectively connected to the three pulse lasers (101, 102, 103) and the three adjustable attenuators (203, 204, 205); One output port of the three 2×2 optical fiber couplers (206, 207, 208) is directly connected to the multi-core optical fiber fan-in unit (213), and the other output port is respectively connected to the three first optical fiber isolators (209, 210, 211), coupled through the first n-channel optical wavelength division multiplexer (212), and then input into the multi-core optical fiber fan-in unit (213) for use as a reference path; The first n-channel optical wavelength division multiplexer (212) is capable of multiplexing n-channel optical wavelength signals into one channel for propagation, and the wavelength and number n correspond to the n-channel demultiplexer (202); The input channels of the multi-core optical fiber fan-in unit (213) correspond one-to-one to the cores of the multi-core optical fiber convolution kernel module (3); The three first optical fiber isolators (209, 210, 211) only allow light to pass in one direction; The multi-core optical fiber convolution core module (3) comprises four multi-core optical fiber convolution cores (301, 302, 303, 304) of a multi-core optical fiber, and each of the multi-core optical fiber convolution cores (301, 302, 303, 304) comprises a fiber core, a phase change material film layer and an anti-oxidation film layer; The phase change material film layer is a sulfur compound; The phase change material film layer has at least two phases that can be adjusted by light pulses; The material of the anti-oxidation film is indium tin oxide; The optical accumulation module (4) includes a multi-core optical fiber fan-out unit (401), three second optical fiber isolators (402, 403, 404), a second n-channel optical wavelength division multiplexer (405), a photoelectric balance detector (406), a data acquisition card (407), and a host computer (408); The output channels of the multi-core optical fiber fan-out unit (401) correspond one-to-one to the cores of the multi-core optical fiber convolution core module (3); The second optical fiber isolator (402, 403, 404) only allows light to pass in one direction; The number of ports n of the second n-channel optical wavelength division multiplexer (405) is one less than the number of output channels of the multi-core optical fiber fan-out unit (401); the second n-channel optical wavelength division multiplexer (405) is used to accumulate the weighted detection light intensities of the n convolution kernels in the multi-core optical fiber convolution kernel module (3); The photoelectric balance detector (406) has a positive input port and a negative input port, and the output value of the photoelectric balance detector (406) depends on the difference between the light intensity input at the positive input port and the light intensity at the negative input port; The positive input port of the photoelectric balance detector (406) is connected to the output end of the second n-channel optical wavelength division multiplexer (405); The negative input port of the photoelectric balance detector (406) is connected to the output end of the multi-core fiber fan-out unit (401) corresponding to the fiber core of the multi-core fiber convolution kernel module (3) inputted from the multi-core fiber fan-in unit (213) in the image input module (2); The photoelectric balance detector (406) converts the input optical signal into an electrical signal and transmits it to the data acquisition card (407), which then transmits the data to the host computer (408) for monitoring and display. The host computer (408) interacts with the data acquisition card (407).

Images