Convolutional computing system and method based on fiber stimulated raman scattering and time-frequency mapping

Abstract

The application discloses a convolution calculation system and method based on optical fiber stimulated Raman scattering and time-frequency mapping, and relates to the technical field of photon computing, and comprises the following steps: a wide-spectrum ultrashort pulse is stretched in a first time domain to obtain first seed light; a convolution kernel data to be calculated is modulated onto the spectral intensity of the first seed light to obtain second seed light after data loading; continuous light or a long pulse is used as first pump light, and a matrix data vector to be calculated is modulated onto the time-domain waveform of the first pump light to obtain second pump light; the second seed light and the second pump light are transmitted in opposite directions in an optical computing medium; the parallel multiplication operation of the convolution kernel data to be calculated and the matrix data vector to be calculated is realized by utilizing the stimulated Raman scattering effect of light in the medium, and stimulated Raman scattering light is output; and the convolution calculation result is output after the stimulated Raman scattering light is processed by a second dispersion medium and a photoelectric detector. The application can realize high-robustness, low-cost and high-parallel full-optical convolution calculation.

Description

Technical Field

[0001] This invention relates to the field of photonic computing technology, and in particular to a convolution computing system and method based on fiber-induced Raman scattering and time-frequency mapping. Background Technology

[0002] With the rapid development of artificial intelligence, deep learning, and high-throughput communication technologies, the demands on the speed and energy efficiency of matrix operations such as convolutional neural networks are increasing exponentially. Traditional von Neumann architecture computers, limited by the failure of Moore's Law and the "memory wall" problem, face enormous challenges in power consumption and heat dissipation, as well as transmission latency bottlenecks, when processing convolutional operations on massive amounts of data.

[0003] Optical computing is a key technology for overcoming the bottlenecks of electronic computing. Existing optical computing solutions include spatial optical path solutions, on-chip integrated photonic solutions, and traditional fiber optic solutions. Among them, spatial optical path solutions are mainly based on lenses and diffraction elements, which have problems such as complex structure and large size, and are extremely sensitive to environmental vibrations, making them difficult to apply on mobile platforms or in harsh environments. On-chip integrated photonic solutions, such as MZI interferometer networks, are small in size, but the number of computing units integrated on a single chip is limited by photolithography processes, and they have extremely high requirements for temperature control, resulting in high manufacturing costs. Traditional fiber optic solutions mostly only implement simple logic gates or scalar multiplication, failing to effectively utilize the wavelength dimension of light for large-scale parallel matrix convolution operations. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a convolution calculation system and method based on fiber stimulated Raman scattering and time-frequency mapping, achieving highly robust, low-cost, and highly parallel all-optical convolution calculation.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a convolution computing system based on fiber stimulated Raman scattering and time-frequency mapping, the system comprising a pulsed laser source, a first dispersive medium, a first electro-optic modulator, a monochromatic laser source, a second electro-optic modulator, an optical computing medium, a second dispersive medium, and a photodetector; The broadband ultrashort pulse generated by the pulsed laser source is subjected to a first-stage time-domain stretching in the first dispersive medium to obtain a first seed light after wavelength and time linear mapping; the convolution kernel data to be calculated is modulated onto the spectral intensity of the first seed light by the first electro-optic modulator to obtain a second seed light after data loading. The continuous light or long pulse generated by the monochromatic laser source is used as the first pump light. The matrix data vector to be calculated is modulated onto the time domain waveform of the first pump light by the second electro-optic modulator to obtain the second pump light. The second seed light and the second pump light are transmitted in opposite directions in the optical computing medium. By utilizing the stimulated Raman scattering effect of light in the medium, the parallel multiplication operation of the convolution kernel data to be calculated and the matrix data vector to be calculated is realized, and stimulated Raman scattering light is output. The stimulated Raman scattered light undergoes a second-stage time-domain stretching in the second dispersive medium to obtain a time-domain modulated optical signal. The time-domain modulated optical signal is processed into an electrical signal by the photodetector to obtain the convolution calculation result.

[0006] In one possible implementation, the group delay dispersion introduced by the first dispersive medium is a first segment of dispersion, and the group delay dispersion introduced by the second dispersive medium is a second segment of dispersion. The second segment dispersion and the first segment dispersion meet a preset condition so that the second segment dispersion compensates for or adjusts the wavelength-time delay introduced by the first segment dispersion; the preset condition is that the second segment dispersion is less than the first segment dispersion, or the second segment dispersion and the first segment dispersion are equal in magnitude but opposite in sign.

[0007] In one possible implementation, when the second segment dispersion is less than the first segment dispersion, the second segment dispersion partially compensates for the wavelength-time delay introduced by the first segment dispersion. The signals of different wavelengths in the stimulated Raman scattering light are convolved and summed in a sliding window manner according to their relative delays. The convolution calculation system operates in convolution operation mode.

[0008] In one possible implementation, when the second segment of dispersion is equal in magnitude and opposite in sign to the first segment of dispersion, signals of different wavelengths in the stimulated Raman scattering light are coherently superimposed at the same time to perform vector dot product operation, and the convolution calculation system operates in matrix-vector multiplication and addition mode.

[0009] In one possible implementation, the convolution calculation system further includes a mode switching module; the mode switching module is used to adjust the second segment of dispersion to switch between the convolution operation mode and the matrix-vector multiplication and addition mode.

[0010] In one possible implementation, the pulsed laser source is a laser or light source capable of linear chirping; the optical computing medium is a waveguide medium with nonlinear effects that can generate stimulated Raman scattering light.

[0011] In one possible implementation, the pulsed laser source is a femtosecond pulsed laser; the optical computing medium is an optical fiber with nonlinear effects that can generate stimulated Raman scattering light.

[0012] In one possible implementation, both the first dispersive medium and the second dispersive medium are dispersive optical fibers.

[0013] In one possible implementation, the center wavelength of the first pump light is set at the Raman gain peak shift of the center wavelength of the broadband ultrashort pulse.

[0014] Secondly, the present invention also provides a convolution calculation method based on fiber stimulated Raman scattering and time-frequency mapping, wherein the method is implemented using the convolution calculation system based on fiber stimulated Raman scattering and time-frequency mapping described in any of the above claims, and the method includes: The received broadband ultrashort pulses are subjected to a first-stage time-domain stretching through a first dispersive medium to obtain the first seed light after wavelength-time linear mapping, and the received continuous light or long pulses are used as the first pump light. The convolution kernel data to be calculated is modulated onto the spectral intensity of the first seed light by the first electro-optic modulator to obtain the second seed light after data loading, and the matrix data vector to be calculated is modulated onto the time domain waveform of the first pump light by the second electro-optic modulator to obtain the second pump light. The second seed light and the second pump light are injected into the optical computing medium in a counter-transmission manner to perform parallel multiplication calculations, and stimulated Raman scattering light is output. The stimulated Raman scattered light is sequentially subjected to a second-stage time-domain stretching through a second dispersive medium and then converted from an optical signal to an electrical signal by a photodetector to obtain the convolution calculation result.

[0015] In practical applications, the convolution calculation system based on fiber-optic stimulated Raman scattering and time-frequency mapping of this invention loads the convolution kernel data to be calculated onto different wavelengths of a broadband ultrashort pulse that has undergone a first-stage time-domain stretching to obtain a second seed light; the matrix data vector to be calculated is loaded onto a first pump light to obtain a second pump light; the second seed light and the second pump light are transmitted in opposite directions and meet in the optical computing medium, resulting in stimulated Raman scattering and achieving all-optical multiplication; subsequently, through a second dispersive medium, the calculation results of different wavelengths are translated, aligned, and superimposed in the time domain, and the convolution calculation results are read out through a photodetector; this invention utilizes the optical computing medium, combined with the broadband characteristics of broadband ultrashort pulses and the nonlinear gain mechanism of stimulated Raman scattering, to achieve highly robust, low-cost, and highly parallel all-optical convolution calculation. Attached Figure Description

[0016] Figure 1 A structural block diagram of a convolution calculation system based on fiber stimulated Raman scattering and time-frequency mapping is provided for an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the loading of convolution kernel data to be calculated onto a broadband ultrashort pulse in a convolution calculation system based on fiber stimulated Raman scattering and time-frequency mapping, provided as an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the loading of the matrix data vector to be calculated onto the first pump light in a convolution calculation system based on fiber stimulated Raman scattering and time-frequency mapping, provided as an embodiment of the present invention. Figure 4 A schematic diagram of the vector signal carried by stimulated Raman scattered light in a convolution calculation system based on stimulated Raman scattering and time-frequency mapping in an embodiment of the present invention; Figure 5 A schematic diagram of the light intensity detected by a photodetector in a convolution calculation system based on fiber-induced Raman scattering and time-frequency mapping, provided as an embodiment of the present invention; Figure 6 The flowchart illustrates the steps of a convolution calculation method based on fiber stimulated Raman scattering and time-frequency mapping, as provided in this embodiment of the invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of this disclosure, unless otherwise stated, "a plurality of" means two or more. Furthermore, the use of "based on" or "according to" implies openness and inclusiveness, because processes, steps, calculations, or other actions "based on" or "according to" one or more of the stated conditions or values ​​may in practice be based on additional conditions or beyond the stated values.

[0019] This invention provides a convolution calculation system and method based on fiber-induced Raman scattering and time-frequency mapping. By utilizing optical computing media and combining the broadband characteristics of broadband ultrashort pulses with the nonlinear gain mechanism of stimulated Raman scattering, it achieves highly robust, low-cost, and highly parallel all-optical convolution calculations.

[0020] like Figure 1 As shown, this embodiment of the invention provides a convolution calculation system based on fiber stimulated Raman scattering and time-frequency mapping, including a pulsed laser source, a first dispersive medium, a first electro-optic modulator, a monochromatic laser source, a second electro-optic modulator, an optical calculation medium, a second dispersive medium, and a photodetector.

[0021] In this process, the broadband ultrashort pulse generated by the pulsed laser source undergoes a first-stage time-domain stretching in the first dispersive medium to obtain the first seed light after a linear mapping between wavelength and time.

[0022] The convolution kernel data to be calculated is modulated onto the spectral intensity of the first seed light by the first electro-optic modulator to obtain the second seed light after data loading.

[0023] Among them, broadband ultrashort pulses refer to optical pulses that simultaneously possess the characteristics of "ultrashort duration" and "extremely wide spectral range." "Ultrashort" refers to the extremely short duration of the pulse, typically within femtoseconds (10⁻¹⁰). -15 (seconds) or even attoseconds (10) -18 The pulse has a wavelength of 10 ...

[0024] Furthermore, the pulsed laser source is a laser or light source capable of achieving linear chirp. For example, a femtosecond pulsed laser or other laser or light source capable of achieving linear chirp.

[0025] Temporal stretching refers to the use of dispersive optical fibers or other dispersion effects to distribute light of different wavelengths sequentially along the time axis, thereby allocating an independent time window for each wavelength.

[0026] In the convolution calculation system of the present invention, seed light generation and loading of convolution kernel data to be calculated are achieved by a pulsed laser source, a first dispersive medium, and a first electro-optic modulator.

[0027] Specifically, such as Figure 2 As shown, the broadband ultrashort pulse generated by the pulsed laser source undergoes a first-stage time-domain stretching process through a first dispersive medium, separating the light of different wavelengths in the broadband ultrashort pulse in time to obtain the first seed light. Then, the convolution kernel data c = (c...) to be calculated is modulated by a first electro-optic modulator. 11 , c 12 , …, c mk The data is modulated onto the spectrum of the first seed light obtained after stretching, resulting in a second seed light loaded with convolution kernel data. Here, m represents the total number of convolution kernels, and k represents the total number of data in each convolution kernel.

[0028] After data loading is complete, each specific wavelength in the broadband ultrashort pulse The light intensity represents a weight value in the convolution kernel data.

[0029] The continuous light or long pulse generated by the monochromatic laser source is used as the first pump light. The matrix data vector to be calculated is modulated onto the time domain waveform of the first pump light by the second electro-optic modulator to obtain the second pump light.

[0030] Furthermore, the center wavelength of the first pump light is set at the Raman gain peak shift of the center wavelength of the broadband ultrashort pulse.

[0031] In the convolution calculation system of the present invention, the loading of the matrix data vector to be calculated is achieved by a monochromatic laser source and a second electro-optic modulator.

[0032] Specifically, such as Figure 3 As shown, continuous light or long pulses generated by a monochromatic laser source are used as the first pump light, and the matrix data vector to be calculated is converted by a second electro-optic modulator. The second pump light is obtained by modulating the second pump light onto the time-domain waveform of the first pump light. N This represents the total number of data points in the matrix data vector to be calculated.

[0033] The second seed light and the second pump light are transmitted in opposite directions in the optical computing medium. By utilizing the stimulated Raman scattering effect of light in the medium, the parallel multiplication operation of the convolution kernel data to be calculated and the matrix data vector to be calculated is realized, and the stimulated Raman scattered light is output.

[0034] Furthermore, the optical computing medium is a waveguide medium with nonlinear effects that can generate stimulated Raman scattering light.

[0035] Specifically, the optical computing medium is an optical fiber that has nonlinear effects and can generate stimulated Raman scattering light.

[0036] In the convolution calculation system of the present invention, the second seed light and the second pump light are injected into the optical computing medium from both ends, respectively, so that the two beams of light are transmitted in opposite directions in the optical computing medium, and parallel multiplication operations are realized in the optical computing medium.

[0037] Specifically, when two beams of light meet in the optical computing medium, the stimulated Raman scattering condition is satisfied. The energy of the second pump light is transferred to the second seed light (or a corresponding Stokes beam is generated). The gain intensity is proportional to the product of the intensities of the two beams, thus realizing the multiplication operation of the matrix and the corresponding weights of the convolution kernel, and data copying. Furthermore, the information carried by the stimulated Raman scattered light output from the optical computing medium is as follows: Figure 4 As shown. In other words, by utilizing the stimulated Raman scattering effect of light in the optical computing medium, each wavelength component in the second seed light obtains a gain proportional to the intensity of the second pump light, thereby realizing parallel multiplication operations between the convolution kernel data elements and the matrix vector elements.

[0038] Stimulated Raman scattering light undergoes a second-stage time-domain stretching in the second dispersive medium to obtain a time-domain modulated optical signal.

[0039] After time-domain modulation, the optical signal is processed into an electrical signal by a photodetector, and the convolution calculation result is obtained.

[0040] In the convolution calculation system of this invention, convolution summation and result output are achieved through a second dispersive medium and a photodetector. Stimulated Raman scattering light output from the optical calculation medium is input into the second dispersive medium for time-domain manipulation, so that the wavelength components carrying different product results are superimposed or arranged on the time axis according to a predetermined rule. Then, the photodetector receives the optical signal manipulated by the second dispersive medium, and the output electrical signal is the convolution calculation result.

[0041] Specifically, the stimulated Raman scattering light output by the optical computing medium contains multiple wavelength components. To complete the "summation" or "matrix-vector multiplication and addition" in convolution, a second-stage time-domain stretching of different wavelength components in the stimulated Raman scattering light is performed through a second dispersive medium, such as... Figure 5 As shown, the second delay causes the different wavelength components (representing different product results) that were originally dispersed in time to be superimposed within a specific time window, or to arrive at the photodetector in a predetermined sequence. The light intensity signal detected by the photodetector is the final convolution calculation result.

[0042] Furthermore, in this embodiment, both the first dispersive medium and the second dispersive medium are dispersive optical fibers.

[0043] Pulsed laser sources and monochromatic laser sources can be independent light sources or integrated into a single light source.

[0044] In practical applications, the convolution calculation system based on fiber-optic stimulated Raman scattering and time-frequency mapping of this invention loads the convolution kernel data to be calculated onto different wavelengths of a broadband ultrashort pulse that has undergone a first-stage time-domain stretching to obtain a second seed light; the matrix data vector to be calculated is loaded onto a first pump light to obtain a second pump light; the second seed light and the second pump light are transmitted in opposite directions and meet in the optical computing medium, resulting in stimulated Raman scattering and achieving all-optical multiplication; subsequently, through a second dispersive medium, the calculation results of different wavelengths are translated, aligned, and superimposed in the time domain, and the convolution calculation results are read out through a photodetector; this invention utilizes the optical computing medium, combined with the broadband characteristics of broadband ultrashort pulses and the nonlinear gain mechanism of stimulated Raman scattering, to achieve highly robust, low-cost, and highly parallel all-optical convolution calculation.

[0045] Furthermore, the group delay dispersion introduced by the first dispersive medium is the first segment dispersion, and the group delay dispersion introduced by the second dispersive medium is the second segment dispersion. The second segment dispersion and the first segment dispersion satisfy a preset condition so that the second segment dispersion compensates for or adjusts the wavelength-time delay introduced by the first segment dispersion.

[0046] The preset conditions are that the dispersion of the second segment is less than that of the first segment, or the dispersion of the second segment is equal in magnitude but opposite in sign to that of the first segment.

[0047] Furthermore, when the second dispersion is less than the first dispersion, the second dispersion partially compensates for the wavelength-time delay introduced by the first dispersion. The signals of different wavelengths in the stimulated Raman scattering light are convolved and summed in a sliding window manner with relative delays. The convolution calculation system operates in convolution operation mode.

[0048] Furthermore, when the second dispersion segment is equal in magnitude and opposite in sign to the first dispersion segment, the time-domain pulse is Fourier transformed back to the original pulse width, achieving full-time-domain compression. Signals of different wavelengths in the stimulated Raman scattering light are coherently superimposed at the same time to perform vector dot product operations. At this time, the convolution calculation system operates in matrix-vector multiplication and addition mode.

[0049] Furthermore, the convolution calculation system also includes a mode switching module; the mode switching module is used to adjust the second segment of dispersion to switch between the convolution operation mode and the matrix-vector multiplication and addition mode.

[0050] Specifically, when the signal-to-noise ratio of stimulated Raman scattering is limited, the second dispersion segment is adjusted to be equal in magnitude and opposite in sign to the first dispersion segment, thereby switching the convolution calculation system from convolution operation mode to matrix-vector multiplication and addition mode.

[0051] The convolution computation system based on stimulated Raman scattering and time-frequency mapping provided in this invention utilizes the broadband characteristics of broadband ultrashort pulses to achieve a linear mapping between wavelength and time through a first dispersive medium. A second seed light, obtained by carrying different elements of the convolution kernel data to be computed at different wavelengths in the first seed light, and a second pump light, obtained by loading the matrix data vector to be computed onto the first pump light, undergo stimulated Raman scattering in the optical computing medium. This allows multiple multiplication operations to be processed in parallel using the wavelength dimension of light within a single physical channel, greatly improving the computational density.

[0052] The convolution calculation system based on stimulated Raman scattering and time-frequency mapping in optical fiber provided by this invention utilizes the stimulated Raman scattering effect occurring in the optical computing medium to achieve "multiplication" (i.e., the nonlinear gain of the intensity of the second pump light and the second seed light), and utilizes the delay characteristics of the second dispersive medium to achieve "addition" (i.e., the recombination or integration of calculation results of different wavelengths in the time domain), so that the entire convolution operation (multiplication and addition operations) is completed entirely during optical transmission without the need for electronic devices.

[0053] The convolution calculation system based on stimulated Raman scattering and time-frequency mapping provided in this invention operates during photon flight, and its delay is limited only by the fiber length and the speed of light, resulting in extremely low delay. Since the entire convolution calculation system mainly utilizes the nonlinear effects of the optical computing medium (optical fiber) and dispersive optical fiber, it does not require the switching of high-power electronic transistors. Therefore, the convolution calculation system of this invention fundamentally solves the high energy consumption problem of traditional electronic computing.

[0054] In the convolutional computation system based on fiber stimulated Raman scattering and time-frequency mapping provided in this embodiment of the invention, the convolutional kernel data to be computed (modulated by broadband ultrashort pulses) is updated in real time and at high speed by a first electro-optic modulator, and the matrix data vector to be computed (modulated by a first pump light) is updated in real time and at high speed by a second electro-optic modulator. Compared with optical computing architectures with fixed optical paths or etched on chips, the convolutional computation system of this invention has extremely strong software-defined capabilities and can adapt to different neural network weights or general computing needs such as image processing.

[0055] The core computing components (optical computing medium, first dispersion medium and second dispersion medium) of the convolution computing system based on fiber stimulated Raman scattering and time-frequency mapping provided in this embodiment of the invention have extremely mature manufacturing processes. They do not require expensive extreme ultraviolet lithography machines or complex on-chip integrated photonics processes, and have the potential for large-scale, low-cost deployment.

[0056] When performing calculations using the convolution calculation system based on fiber stimulated Raman scattering and time-frequency mapping of this invention, the data is calculated while it is being transmitted through the optical fiber, breaking the traditional "store-and-compute separation" architecture and significantly reducing system latency. The entire calculation process does not require complex photonic integrated chips, but mainly utilizes commercial optical fiber devices, which have good environmental stability and low cost. By increasing the spectral bandwidth, the dimension of the convolution kernel can be easily increased, thereby improving parallel computing capabilities.

[0057] like Figure 6 As shown, in a second aspect, embodiments of the present invention also provide a convolution calculation method based on fiber stimulated Raman scattering and time-frequency mapping. This method is implemented using the convolution calculation system based on fiber stimulated Raman scattering and time-frequency mapping described in any of the preceding claims. The method includes: Step 101: Perform a first-stage time-domain stretching on the received broadband ultrashort pulse through the first dispersive medium to obtain the first seed light after wavelength and time linear mapping, and use the received continuous light or long pulse as the first pump light.

[0058] Step 102: Modulate the convolution kernel data to be calculated onto the spectral intensity of the first seed light using the first electro-optic modulator to obtain the second seed light after data loading, and modulate the matrix data vector to be calculated onto the time-domain waveform of the first pump light using the second electro-optic modulator to obtain the second pump light.

[0059] Step 103: The second seed light and the second pump light are injected into the optical computing medium using a counter-transmission method to perform parallel multiplication calculations and output stimulated Raman scattering light.

[0060] Step 104: The stimulated Raman scattered light is sequentially subjected to a second-stage time-domain stretching through a second dispersive medium and then converted from an optical signal to an electrical signal through a photodetector to obtain the convolution calculation result.

[0061] The convolution calculation method based on fiber stimulated Raman scattering and time-frequency mapping provided in this embodiment of the invention can be implemented using any of the above-mentioned convolution calculation systems based on fiber stimulated Raman scattering and time-frequency mapping, thus achieving the same effect as the above-mentioned convolution calculation systems based on fiber stimulated Raman scattering and time-frequency mapping.

[0062] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the method described above can be referred to the corresponding process in the aforementioned system embodiments, and will not be repeated here.

[0063] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions within the technical scope disclosed in the present invention should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A convolution calculation system based on fiber-stimulated Raman scattering and time-frequency mapping, characterized in that, It includes a pulsed laser source, a first dispersive medium, a first electro-optic modulator, a monochromatic laser source, a second electro-optic modulator, an optical computing medium, a second dispersive medium, and a photodetector; The broadband ultrashort pulse generated by the pulsed laser source is subjected to a first-stage time-domain stretching in the first dispersive medium to obtain the first seed light after wavelength and time linear mapping. The convolution kernel data to be calculated is modulated onto the spectral intensity of the first seed light by the first electro-optic modulator to obtain the second seed light after data loading; The continuous light or long pulse generated by the monochromatic laser source is used as the first pump light. The matrix data vector to be calculated is modulated onto the time domain waveform of the first pump light by the second electro-optic modulator to obtain the second pump light. The second seed light and the second pump light are transmitted in opposite directions in the optical computing medium. By utilizing the stimulated Raman scattering effect of light in the medium, the parallel multiplication operation of the convolution kernel data to be calculated and the matrix data vector to be calculated is realized, and stimulated Raman scattering light is output. The stimulated Raman scattered light undergoes a second-stage time-domain stretching in the second dispersive medium to obtain a time-domain modulated optical signal. The time-domain modulated optical signal is processed into an electrical signal by the photodetector to obtain the convolution calculation result; Wherein, the group delay dispersion introduced by the first dispersive medium is the first segment of dispersion, and the group delay dispersion introduced by the second dispersive medium is the second segment of dispersion; The second segment dispersion and the first segment dispersion meet a preset condition so that the second segment dispersion compensates for or adjusts the wavelength-time delay introduced by the first segment dispersion; the preset condition is that the second segment dispersion is less than the first segment dispersion, or the second segment dispersion and the first segment dispersion are equal in magnitude but opposite in sign.

2. The convolution calculation system based on fiber-stimulated Raman scattering and time-frequency mapping according to claim 1, characterized in that, When the second segment dispersion is less than the first segment dispersion, the second segment dispersion partially compensates for the wavelength-time delay introduced by the first segment dispersion. The signals of different wavelengths in the stimulated Raman scattering light are convolved and summed in a sliding window manner with relative delays. The convolution calculation system operates in convolution operation mode.

3. The convolution calculation system based on fiber stimulated Raman scattering and time-frequency mapping according to claim 2, characterized in that, When the second segment of dispersion is equal in magnitude and opposite in sign to the first segment of dispersion, signals of different wavelengths in the stimulated Raman scattering light are coherently superimposed at the same time to perform vector dot product operation, and the convolution calculation system operates in matrix-vector multiplication and addition mode.

4. The convolution calculation system based on fiber stimulated Raman scattering and time-frequency mapping according to claim 3, characterized in that, The convolution calculation system also includes a mode switching module; the mode switching module is used to adjust the second segment of dispersion to switch between the convolution operation mode and the matrix-vector multiplication and addition mode.

5. The convolution calculation system based on fiber stimulated Raman scattering and time-frequency mapping according to claim 1, characterized in that, The pulsed laser source is a laser or light source capable of linear chirping; the optical computing medium is a waveguide medium with nonlinear effects that can generate stimulated Raman scattering light.

6. The convolution calculation system based on fiber stimulated Raman scattering and time-frequency mapping according to claim 5, characterized in that, The pulsed laser source is a femtosecond pulsed laser; the optical computing medium is an optical fiber with nonlinear effects that can generate stimulated Raman scattering light.

7. The convolution calculation system based on fiber stimulated Raman scattering and time-frequency mapping according to claim 1, characterized in that, Both the first dispersive medium and the second dispersive medium are dispersive optical fibers.

8. The convolution calculation system based on fiber stimulated Raman scattering and time-frequency mapping according to claim 1, characterized in that, The center wavelength of the first pump light is set at the Raman gain peak frequency shift of the center wavelength of the broadband ultrashort pulse.

9. A convolution calculation method based on stimulated Raman scattering in optical fibers and time-frequency mapping, characterized in that, The method is implemented using the convolution calculation system based on fiber stimulated Raman scattering and time-frequency mapping as described in any one of claims 1 to 8, and the method includes: The received broadband ultrashort pulses are subjected to a first-stage time-domain stretching through a first dispersive medium to obtain the first seed light after wavelength-time linear mapping, and the received continuous light or long pulses are used as the first pump light. The convolution kernel data to be calculated is modulated onto the spectral intensity of the first seed light by the first electro-optic modulator to obtain the second seed light after data loading, and the matrix data vector to be calculated is modulated onto the time domain waveform of the first pump light by the second electro-optic modulator to obtain the second pump light. The second seed light and the second pump light are injected into the optical computing medium in a counter-transmission manner to perform parallel multiplication calculations, and stimulated Raman scattering light is output. The stimulated Raman scattered light is sequentially subjected to a second-stage time-domain stretching through a second dispersive medium and then converted from an optical signal to an electrical signal by a photodetector to obtain the convolution calculation result. Wherein, the group delay dispersion introduced by the first dispersive medium is the first segment of dispersion, and the group delay dispersion introduced by the second dispersive medium is the second segment of dispersion; The second segment dispersion and the first segment dispersion meet a preset condition so that the second segment dispersion compensates for or adjusts the wavelength-time delay introduced by the first segment dispersion; the preset condition is that the second segment dispersion is less than the first segment dispersion, or the second segment dispersion and the first segment dispersion are equal in magnitude but opposite in sign.

Images