A fiber tensor computation system and method based on dispersion delay and wavelength division multiplexing

Abstract

The embodiment of the application provides a kind of optical fiber tensor calculation system and method based on dispersion delay and wavelength division multiplexing, the system includes: multiple color light source is used to generate multiple color light containing N different wavelength channels;First wavelength division multiplexer is used to divide multiple color light into N different wavelength optical signals;First order electro-optic modulator is used to load the loading matrix B obtained after cyclic rearrangement;Second wavelength division multiplexer is used to combine N different wavelength optical signals into multiple color light;Optical beam splitter is used to M multiple color light signals;Second order electro-optic modulator is used to load each column element in matrix T into M output branches according to τ time interval sequence;Dispersion device is used to delay different wavelength optical signals;Photoelectric detector is used to output product matrix.The scheme improves the parallelism and calculation efficiency of matrix multiplication, realizes the lossless summation of product term in optical domain, and greatly reduces the calculation overhead and system complexity.

Description

Technical Field

[0001] This solution relates to the field of optical computing technology, and in particular to an optical fiber tensor computing system and method based on dispersion delay and wavelength division multiplexing. Background Technology

[0002] Matrix multiplication is a fundamental operation in tensor computation, image processing, signal processing, and scientific computing. Matrix multiplication involves multiplying corresponding matrix elements and summing the product terms according to their row and column positions in the output matrix. As the size of the matrix and the amount of data to be processed increase, the loading, multiplication, and summation of matrix elements require greater parallel processing capabilities.

[0003] Optical computing leverages the parallel transmission and modulation characteristics of optical signals to provide a new approach to matrix operations. Wavelength division multiplexing (WDM) can utilize multiple channels with different wavelengths to carry data in parallel, and electro-optic modulators can load matrix elements as time-domain optical intensity signals. Optical computing can significantly reduce the energy requirements of multiplication and addition operations in matrix calculations.

[0004] However, optical computing suffers from the following technical shortcomings that urgently need to be addressed: First, the mapping between matrix elements and optical resources lacks flexibility and uniformity. Existing methods typically design fixed wavelength allocation tables and time slot division rules for matrices of specific dimensions. Each wavelength channel corresponds to a fixed column of the input matrix, and each time slot corresponds to a fixed column of the output matrix. When the matrix size changes, not only must the number of wavelength channels be reallocated, but also the time slot width and dispersion parameters must be adjusted. The mapping logic under different scales is incompatible, and there is a lack of a unified mathematical mapping framework to describe matrix multiplication of arbitrary dimensions. This results in the system being unable to dynamically adapt to different computing tasks, severely restricting the versatility and deployment efficiency of optical computing processors.

[0005] Secondly, the timing misalignment and synchronization difficulties introduced by dispersive delays are significant challenges. The delay difference generated by dispersive devices for different wavelengths is proportional to the wavelength spacing and transmission distance. To correctly align product terms according to the output position, the absolute delay of each wavelength channel must be accurately calculated. However, in practical systems, as the number of wavelength channels increases, the cumulative delay difference between the longest and shortest wavelengths can easily exceed the width of a single time slot, causing product terms that originally belonged to the same output position to fall into adjacent time slots, resulting in incorrect summation objects. Multiple product terms may partially overlap in time, introducing unexpected partial cumulative interference. In addition, the temperature sensitivity of dispersive devices can cause delay drift, further exacerbating alignment errors. Existing technologies lack quantitative constraints and dynamic calibration mechanisms for delay differences, time slot widths, and wavelength spacing, making it difficult to maintain correct temporal alignment under different matrix sizes or environmental changes.

[0006] In summary, how to establish a matrix multiplication optical computing method with both flexible mapping and precise timing alignment capabilities in a complex system that couples wavelength division multiplexing, time-domain modulation, and dispersion delay is a technical challenge that urgently needs to be solved in this field. Summary of the Invention

[0007] This solution aims to at least address the technical problems existing in the prior art. Therefore, the first aspect of this invention proposes an optical fiber tensor calculation system based on dispersion delay and wavelength division multiplexing, used to calculate the product C=AT of an N-order weight matrix A and an N×M input matrix T. The system includes: Polychromatic light source, used to generate polychromatic light containing N different wavelength channels; The first wavelength division multiplexer is connected to the polychromatic light source and is used to divide the polychromatic light into N optical signals of different wavelengths; N first-stage electro-optic modulators are respectively disposed on the N output branches of the first wavelength division multiplexer, and are used to load the loading matrix B obtained by cyclic rearrangement of the weight matrix A; the cyclic rearrangement is used to distribute the N weight elements belonging to the same output row in the weight matrix A along a cyclic diagonal in the loading matrix B. The second wavelength division multiplexer is connected to the output of the N first-stage electro-optic modulators and is used to combine the N optical signals of different wavelengths into polychromatic light. A 1×M optical beam splitter is connected to the output of the second wavelength division multiplexer to split the combined polychromatic light into M paths, thereby obtaining M polychromatic light signals. M second-stage electro-optic modulators are respectively disposed on the M output branches of the 1×M optical beam splitter to receive the M polychromatic light signals and load each column of the input matrix T. The elements of each column of the matrix T are loaded into the M output branches in sequence according to the time interval τ to obtain M product light signals. M dispersive devices are respectively set after each second-stage electro-optic modulator to generate different group delays for optical signals of different wavelengths, so that the product optical signals belonging to the same output result are aligned in time to obtain aligned optical signals. M photodetectors are respectively positioned after each dispersive device to receive the alignment light signal, perform photoelectric conversion and optical power superposition on the alignment light signal, and output the corresponding column of the product matrix C=AT.

[0008] Optionally, the center wavelengths of the N wavelength channels generated by the polychromatic light source are respectively to The N output branches of the first wavelength division multiplexer correspond one-to-one with the N wavelength channels.

[0009] Optionally, the loading matrix B satisfies ,in This represents a modulo operation with a period of N; the i-th first-stage electro-optic modulator is used to load B(i,j) in the j-th source time slot, and completes the synchronous loading of N×N elements in the loading matrix B within a time range of N×τ, where τ is the time slot width of a single data element.

[0010] Optionally, the m-th second-stage electro-optic modulator is used to load the m-th column element T(j,m) of the input matrix T in the j-th source time slot, and to complete the synchronous loading of N×M elements in the input matrix T within a time range of N×τ, where m=1,2,…,M and j=1,2,…,N.

[0011] Optionally, the dispersive device is a dispersive fiber or a fiber Bragg grating, and the dispersive device makes the wavelength... Optical signal generation group delay The time difference between the arrival of adjacent wavelength optical signals at the corresponding photodetector is τ, where i = 1, 2, ..., N.

[0012] Optionally, the m-th photodetector is used to receive and superimpose the product optical signals of N wavelength channels in the output time slot corresponding to the r-th row of the product matrix C. The N product optical signals correspond to the product of each element in the r-th row of the weight matrix A and the corresponding element in the m-th column of the input matrix T, and the superposition result is C(r,m). The M photodetectors output the M columns of the product matrix C, where r=1,2,…,N and m=1,2,…,M.

[0013] Optionally, the system further includes a synchronization controller, which is connected to the N first-stage electro-optic modulators and the M second-stage electro-optic modulators respectively, and is used to provide a unified time slot synchronization signal to all electro-optic modulators, and control the loading matrix B and the input matrix T to be loaded synchronously according to the same source time slot sequence.

[0014] A second aspect of this invention proposes a method for calculating fiber tensors based on dispersion delay and wavelength division multiplexing, the method comprising: After the polychromatic light source generates polychromatic light containing N different wavelength channels, the polychromatic light is divided into N optical signals of different wavelengths using a first wavelength division multiplexer, and each optical signal is passed through N first-stage electro-optic modulators respectively. The weight matrix A is rearranged into a loading matrix B according to the cyclic shift rule, and the loading matrix B is loaded through the N first-stage electro-optic modulators; The N optical signals after the first-stage electro-optic modulation are combined into polychromatic light using a second wavelength division multiplexer. The polychromatic light is then divided into M channels, and the resulting M polychromatic light signals are passed through M second-stage electro-optic modulators respectively. The M second-stage electro-optic modulators are used to load each column of the input matrix T, and the elements of each column of the matrix T are loaded into the M output branches in sequence according to the time interval τ to obtain M product optical signals. The M-channel product optical signals are passed through a dispersive device, which delays the optical signals of different wavelengths in the M-channel product optical signals in different ways, so that the product optical signals belonging to the same output result are aligned in time to obtain aligned optical signals. By measuring the temporal information of the light intensity of the aligned light signal received by M photodetectors, the product of the weight matrix A and the input matrix T is obtained.

[0015] Optionally, rearranging the weight matrix A into a loading matrix B according to a cyclic shift rule includes: Obtain the N-order weight matrix A, whose elements are denoted as A(r,j), where r=1,2,…,N are row indices and j=1,2,…,N are column indices; According to the cyclic shift rule, each column of the weight matrix A is cyclically shifted to obtain the loading matrix B; where the element B(i,j) in the loading matrix B is B(1+mod(ji,N),j), and mod(·) represents the modulo operation with a period of N; The output loading matrix B contains N elements belonging to the same output row r that satisfy r = 1 + mod (ji, N) and are distributed along the same cyclic diagonal. This matrix is ​​used in conjunction with the dispersion delay to align the corresponding product optical signals in the same output time slot.

[0016] The embodiments of the present invention have the following beneficial effects: This invention provides an optical fiber tensor calculation system based on dispersion delay and wavelength division multiplexing. The system includes: a polychromatic light source for generating polychromatic light containing N different wavelength channels; a first wavelength division multiplexer connected to the polychromatic light source for splitting the polychromatic light into N optical signals of different wavelengths; N first-stage electro-optic modulators, respectively disposed on the N output branches of the first wavelength division multiplexer, for loading a loading matrix B obtained by cyclic rearrangement of a weight matrix A; the cyclic rearrangement is used to distribute the N weight elements belonging to the same output row in the weight matrix A along a cyclic diagonal in the loading matrix B; a second wavelength division multiplexer connected to the outputs of the N first-stage electro-optic modulators for combining the N optical signals of different wavelengths into polychromatic light; and a 1×M optical beamsplitter connected to the output of the second wavelength division multiplexer. The output connection is used to split the combined polychromatic light into M paths, resulting in M ​​polychromatic light signals; M second-stage electro-optic modulators are respectively set on the M output branches of the 1×M optical beam splitter, used to receive the M polychromatic light signals and load each column of the input matrix T, loading each column element of the matrix T sequentially into the M output branches according to the time interval τ, to obtain M product light signals; M dispersive devices are respectively set after each second-stage electro-optic modulator, used to generate different group delays for light signals of different wavelengths, so that the product light signals belonging to the same output result are aligned in time, to obtain aligned light signals; M photodetectors are respectively set after each dispersive device, used to receive the aligned light signals and perform photoelectric conversion and optical power superposition on the aligned light signals, outputting the corresponding column of the product matrix C=AT. This scheme utilizes N channels with different wavelengths to carry weighted data in parallel, which is then distributed to M output branches to load each column of the input matrix, improving the parallelism and computational efficiency of matrix multiplication. Through cyclic rearrangement and dispersion delay, the product optical signals are automatically aligned and superimposed, achieving lossless summation of the product terms in the optical domain. Multiplication and addition operations are naturally completed in the optical domain, eliminating the need for electronic multiplication and addition units. The electronic part is only responsible for driving and control, significantly reducing computational overhead and system complexity. Attached Figure Description

[0017] Figure 1 A schematic diagram of the optical path structure of an optical fiber tensor calculation system based on dispersion delay and wavelength division multiplexing is provided for an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating how the original matrix A is rearranged into a loaded matrix B according to a cyclic shift rule when N=3, as provided in an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating a timing loading relationship and output time slot alignment, provided as an embodiment of the present invention. Figure 4 The flowchart illustrates the steps of a fiber tensor calculation method based on dispersion delay and wavelength division multiplexing, as provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present solution, and not all embodiments. Based on the embodiments of the present solution, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present solution.

[0019] 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.

[0020] This invention targets a vertically arranged multi-band camera imaging system. Using one image as a reference and combining it with the target observation distance, dynamic registration parameters are calculated through depth parallax compensation. Furthermore, a strategy of inverse region of interest mapping is employed to process only the necessary local regions, significantly reducing computational overhead.

[0021] Figure 1 This is a schematic diagram of the optical path structure of an optical fiber tensor calculation system based on dispersion delay and wavelength division multiplexing, provided as an embodiment of the present invention.

[0022] like Figure 1 As shown, the system is used to calculate the product C=AT of an N-order weight matrix A and an N×M input matrix T. The system includes: Polychromatic light source, used to generate polychromatic light containing N different wavelength channels; The first wavelength division multiplexer is connected to the polychromatic light source and is used to divide the polychromatic light into N optical signals of different wavelengths; N first-stage electro-optic modulators are respectively disposed on the N output branches of the first wavelength division multiplexer, and are used to load the loading matrix B obtained by cyclic rearrangement of the weight matrix A; the cyclic rearrangement is used to distribute the N weight elements belonging to the same output row in the weight matrix A along a cyclic diagonal in the loading matrix B. The second wavelength division multiplexer is connected to the output of the N first-stage electro-optic modulators and is used to combine the N optical signals of different wavelengths into polychromatic light. A 1×M optical beam splitter is connected to the output of the second wavelength division multiplexer to split the combined polychromatic light into M paths, thereby obtaining M polychromatic light signals. M second-stage electro-optic modulators are respectively disposed on the M output branches of the 1×M optical beam splitter to receive the M polychromatic light signals and load each column of the input matrix T. The elements of each column of the matrix T are loaded into the M output branches in sequence according to the time interval τ to obtain M product light signals. M dispersive devices are respectively set after each second-stage electro-optic modulator to generate different group delays for optical signals of different wavelengths, so that the product optical signals belonging to the same output result are aligned in time to obtain aligned optical signals. M photodetectors are respectively positioned after each dispersive device to receive the alignment light signal, perform photoelectric conversion and optical power superposition on the alignment light signal, and output the corresponding column of the product matrix C=AT.

[0023] The system in this embodiment of the invention is an all-optical matrix multiplication accelerator. It maps the matrix multiplication C=A×T multiplication and addition operation to the optical domain through four steps: "wavelength parallel loading - time synchronization modulation - dispersion delay alignment - photoelectric detection summation", without having to perform electronic multiplication and addition operations element by element.

[0024] Specifically, let matrix A be an N-order matrix, matrix T be an N×M matrix, and the target output be matrix C:

[0025] in, This represents the element in the r-th row and m-th column of matrix C. This represents the element in the r-th row and j-th column of matrix A, where j is the wavelength channel index. This represents the element in the j-th row and m-th column of matrix T. This indicates that the summation is performed on all N wavelength channels.

[0026] The computational flow and optical path structure of this embodiment include a polychromatic light source, a first wavelength division multiplexer for beam splitting, N first-stage electro-optic modulators for loading matrix B, a second wavelength division multiplexer for beam combining, a 1×M optical beam splitter, M second-stage electro-optic modulators for loading matrix T, M segments of dispersive optical fiber, and M photodetectors. The two-stage electro-optic modulators are synchronously loaded according to a unified time slot, and the photodetectors sample in the corresponding output time slot.

[0027] Specifically, a polychromatic light source generates channels containing N different wavelengths ( , , …, ( ) polychromatic light.

[0028] The first wavelength division multiplexer decomposes the polychromatic light according to wavelength and distributes it to N output branches. The first wavelength division multiplexer distributes N wavelength channels to N first-stage branches, and each branch contains a first-stage electro-optic modulator.

[0029] N first-stage electro-optic modulators constitute the first-stage modulation of the system, completing the data loading of weight matrix A. Weight matrix A is rearranged into loading matrix B according to a cyclic shift rule, satisfying:

[0030] in, This represents the element in the i-th row and j-th column of matrix B, mod( ) represents the modulo operation. This indicates finding the remainder when (ji) is divided by N.

[0031] The purpose of rearrangement is to distribute the N weight elements (A(r,1)...A(r,N)) belonging to the same output row r along a "circular diagonal" in matrix B, in preparation for subsequent dispersion alignment.

[0032] Within N consecutive time slots (each time slot width is τ), the i-th first-stage electro-optic modulator is sequentially loaded with B(i,1), B(i,2), …, B(i,N), that is, at wavelengths… λ i The data of the i-th row of matrix B is loaded in chronological order. Within N source time slots, N modulators jointly complete the loading of N×N elements of B; during continuous calculation, the loaded frames are sent at Nτ periods.

[0033] The second wavelength division multiplexer recombines the N optical signals of different wavelengths that have undergone the first stage modulation into a single polychromatic beam. At this point, each wavelength channel of the polychromatic beam carries its corresponding weighted data sequence.

[0034] The combined polychromatic light is split into M equal paths by a 1×M beam splitter and distributed to M parallel output branches. These M branches correspond to the M columns of the output matrix C, meaning each branch independently calculates one column of the output matrix.

[0035] M second-stage electro-optic modulators form the second-stage modulation of the system, completing the data loading of the input matrix T. The second-stage electro-optic modulator of the m-th branch loads the elements of the m-th column of the input matrix T sequentially in the same N consecutive time slots: T(1,m), T(2,m), …, T(N,m).

[0036] After two stages of modulation, the wavelength in the m-th branch λ iThe optical power in the j-th time slot is proportional to the product B(i,j)×T(j,m). Since B(i,j)=A(r,j) (where r=1+mod(ji,N)), the multiplication operation of the corresponding matrix elements is actually implemented, that is, the multiplication is completed during the light intensity modulation process.

[0037] Each second-stage electro-optic modulator is followed by a dispersive fiber or a dispersive device with the same function. The core function of the dispersive device is to produce different propagation delays for optical signals of different wavelengths, and the delay is linearly related to the wavelength.

[0038] The system is equipped with dispersive devices to adjust the wavelength. λ i Generates group delay:

[0039] in, This represents the total group delay after wavelength λi passes through the dispersive device; This represents the basic fixed time delay, which is the common time deviation that all wavelengths have; N represents the total number of wavelength channels, a total of N wavelengths, and i represents the wavelength channel index number. This represents the time delay difference between adjacent wavelengths.

[0040] That is, the time difference between adjacent wavelengths (such as λi and λi+1) arriving at the photodetector is exactly τ. This is equivalent to rearranging the time-domain signals of each wavelength channel.

[0041] The alignment mechanism is as follows: the N weight elements corresponding to a certain output row r are distributed along the circular diagonal of matrix B. Taking N=3 as an example, r=1 corresponds to B(1,1), B(2,2), and B(3,3), and these three products are located in time slot 1 of wavelength λ1, time slot 2 of wavelength λ2, and time slot 3 of wavelength λ3, respectively. After dispersion delay, λ1 is delayed by 2τ, λ2 by τ, and λ3 by 0, and all three arrive at the detector at exactly one output time slot.

[0042] A photodetector is placed after each dispersive device to complete the final summation and photoelectric conversion.

[0043] A photodetector performs square-law detection on optical signals, and its output photocurrent is proportional to the incident light power. When the product of optical signals from multiple wavelength channels arrives simultaneously, the detector automatically superimposes the optical power of each wavelength channel to obtain C(r,m).

[0044] As an optional embodiment, the center wavelengths of the N wavelength channels generated by the polychromatic light source are respectively to The N output branches of the first wavelength division multiplexer correspond one-to-one with the N wavelength channels.

[0045] As an optional embodiment, the loading matrix B satisfies ,in This represents a modulo operation with a period of N; the i-th first-stage electro-optic modulator is used to load B(i,j) in the j-th source time slot, and completes the synchronous loading of N×N elements in the loading matrix B within a time range of N×τ, where τ is the time slot width of a single data element.

[0046] This invention transforms the weight matrix A into a wavelength-time loading matrix B. Row index i in B corresponds to the wavelength. And for the i-th first-stage electro-optic modulator, column index j corresponds to the source time slot, and B satisfies: .

[0047] in, This represents the element in the i-th row and j-th column of matrix B, mod( ) represents the modulo operation. This indicates finding the remainder when (ji) is divided by N.

[0048] make ,but For a given output row r and wavelength i, the source time slots that need to participate in the summation are: Therefore, within one N time slot period, the N product terms belonging to the same output row r are distributed along a cyclic diagonal of B.

[0049] Figure 2 This is a schematic diagram illustrating how the original matrix A is rearranged into a loaded matrix B according to a cyclic shift rule when N=3, as provided in an embodiment of the present invention.

[0050] like Figure 2 As shown, when N=3, the original matrix A is rearranged according to the circular shift rule to B=[a11,a22,a33;a31,a12,a23;a21,a32,a13]. Output line r=1 corresponds to B(1,1), B(2,2), and B(3,3), which are a11, a12, and a13 respectively; output line r=2 corresponds to B(3,1), B(1,2), and B(2,3), which are a21, a22, and a23 respectively; output line r=3 corresponds to B(2,1), B(3,2), and B(1,3), which are a31, a32, and a33 respectively.

[0051] This invention combines cyclic rearrangement of the weight matrix with dispersion delay to align and superimpose product optical signals corresponding to the same output result at the photodetector according to a predetermined timing sequence, thereby achieving optical domain summation of matrix element product terms without loss.

[0052] As an optional embodiment, the m-th second-stage electro-optic modulator is used to load the m-th column element T(j,m) of the input matrix T in the j-th source time slot, and complete the synchronous loading of N×M elements in the input matrix T within a time range of N×τ, where m=1,2,…,M and j=1,2,…,N.

[0053] In this embodiment of the invention, in each source time slot j with a width of τ, the second-stage electro-optic modulator on the m-th output branch loads the j-th element T(j,m) of the m-th column of the input matrix T. Since there are M such second-stage electro-optic modulators working in parallel, in the same source time slot j, the M modulators load T(j,1) to T(j,M) respectively, thus completing the parallel loading of all M elements in the j-th row of the input matrix T.

[0054] As the source time slot changes continuously from j=1 to j=N, each second-stage electro-optic modulator sequentially loads all N elements of its assigned column. All M modulators complete the loading of N×M elements within a total duration of N×τ. All modulators are controlled by the same synchronization clock and simultaneously switch to the next row of data at the boundary of each time slot. This achieves the ordered loading of all elements of the input matrix T in the time dimension and parallel broadcasting between each output branch, providing time-matched data input for subsequent multiplication operations with the weight matrix elements carried by each wavelength channel in the optical domain.

[0055] As an optional embodiment, the dispersive device is a dispersive fiber or a fiber Bragg grating, and the dispersive device causes the wavelength to... Optical signal generation group delay The time difference between the arrival of adjacent wavelength optical signals at the corresponding photodetector is τ, where i = 1, 2, ..., N.

[0056] In this embodiment of the invention, the core function of the dispersive device is to generate different propagation delays for light signals of different wavelengths, and the delay is linearly related to the wavelength.

[0057] Dispersive fibers utilize the inherent light dispersion properties of the fiber material to achieve delay control. Fiber Bragg gratings, on the other hand, are devices inscribed with periodic refractive index modulations in the fiber core; essentially, they are wavelength-selective mirrors.

[0058] The system is equipped with dispersive devices to adjust the wavelength. λ i Generates group delay:

[0059] That is, adjacent wavelengths (such as λ) i With λ i+1The time difference between the arrival times at the photodetector is exactly τ. This is equivalent to rearranging the time-domain signals of each wavelength channel.

[0060] Figure 3 This is a schematic diagram illustrating a timing loading relationship and output time slot alignment, provided as an embodiment of the present invention.

[0061] like Figure 3 As shown, taking wavelengths λ1, λ2, and λ3 as examples, the three wavelength channels generate relative delays of 2τ, τ, and 0, respectively. For the first column of matrix T, the first photodetector receives and superimposes the data in the first output time slot. , and Received and superimposed in the second output time slot , and Received and superimposed in the third output time slot , and The three elements of the first column of the output matrix are obtained sequentially. The other output branches load the corresponding columns of matrix T in the same way, and the corresponding columns of the output matrix are obtained by the corresponding photodetectors.

[0062] As an optional embodiment, the m-th photodetector is used to receive and superimpose the product optical signals of N wavelength channels in the output time slot corresponding to the r-th row of the product matrix C. The N product optical signals correspond to the product of each element in the r-th row of the weight matrix A and the corresponding element in the m-th column of the input matrix T, and the superposition result is C(r,m). M photodetectors output the M columns of the product matrix C, where r=1,2,…,N and m=1,2,…,M.

[0063] After dispersion delay, the product optical signals carried by the N wavelength channels that were originally staggered on the time axis are rearranged to the same moment, which corresponds to the output time slot of the r-th row of the product matrix C. At this time, the N optical signals received by the detector correspond to the products of the N elements in the r-th row of the weight matrix A and the N corresponding elements in the m-th column of the input matrix T, namely A(r,1)T(1,m), A(r,2)T(2,m), ..., A(r,N)T(N,m). Since the output photocurrent of the photodetector is proportional to the total incident optical power, the optical power of these N product optical signals is automatically physically superimposed on the photosensitive surface of the detector, and the summation operation can be achieved without additional energy consumption. The superposition result is... .

[0064] As an optional embodiment, the system further includes a synchronization controller, which is connected to the N first-stage electro-optic modulators and the M second-stage electro-optic modulators respectively, for providing a unified time slot synchronization signal to all electro-optic modulators and controlling the loading matrix B and the input matrix T to be loaded synchronously according to the same source time slot sequence.

[0065] Specifically, the synchronization controller provides a unified time slot synchronization signal to all electro-optic modulators, thereby ensuring that the N first-stage electro-optic modulators and the M second-stage electro-optic modulators strictly follow the same source time slot division. That is, within a time slot of width τ, the N first-stage electro-optic modulators synchronously load all N elements of the j-th column of the loading matrix B, while the M second-stage electro-optic modulators synchronously load all M elements of the j-th row of the input matrix T, so that the two-stage modulation maintains a one-to-one correspondence in time.

[0066] Compared with the prior art, the present invention has at least the following beneficial effects: (1) The present invention utilizes N different wavelength channels to carry the weight matrix data after cyclic rearrangement in parallel, and distributes the polychromatic light after beam combining to M output branches, respectively loading each column of the input matrix, which is beneficial to improving the parallelism and computational efficiency of matrix multiplication.

[0067] (2) The present invention combines the cyclic rearrangement of the weight matrix with the dispersion delay to make the product light signals corresponding to the same output result aligned and superimposed at the photodetector according to a predetermined time sequence, thus realizing the optical domain summation of the product terms of matrix elements without loss.

[0068] (3) The core matrix multiplication and addition process of the present invention is completed in the two-stage optical intensity modulation, optical signal transmission, dispersion delay and photoelectric detection process. The computation optical path can be constructed using wavelength division multiplexer, optical beam splitter, electro-optic modulator, dispersion fiber and photoelectric detector, which helps to reduce the complexity of system implementation.

[0069] (4) The present invention loads the weight matrix data into different wavelength channels and enables multiple output branches to share the polychromatic light after the first-level modulation. Each output branch only needs to load the corresponding column of the input matrix. The core matrix multiplication and addition process does not require setting up electronic multiplication and addition units item by item. The electronic part is mainly used for modulation driving, timing control and detection result reading, which helps to reduce electronic calculation and data transmission overhead.

[0070] In summary, the fiber tensor calculation system based on dispersion delay and wavelength division multiplexing provided in this embodiment of the invention includes: a polychromatic light source for generating polychromatic light containing N different wavelength channels; a first wavelength division multiplexer connected to the polychromatic light source for splitting the polychromatic light into N optical signals of different wavelengths; N first-stage electro-optic modulators respectively disposed on the N output branches of the first wavelength division multiplexer for loading a loading matrix B obtained by cyclic rearrangement of weight matrix A; the cyclic rearrangement is used to distribute the N weight elements belonging to the same output row in weight matrix A along a cyclic diagonal in loading matrix B; a second wavelength division multiplexer connected to the output ends of the N first-stage electro-optic modulators for combining the N optical signals of different wavelengths into polychromatic light; and a 1×M optical beamsplitter connected to the second wavelength division multiplexer. The output terminal is connected to split the combined polychromatic light into M paths, resulting in M ​​polychromatic light signals; M second-stage electro-optic modulators are respectively set on the M output branches of the 1×M optical beam splitter, used to receive the M polychromatic light signals and load each column of the input matrix T, loading each column element of the matrix T sequentially into the M output branches according to the time interval τ, to obtain M product light signals; M dispersive devices are respectively set after each second-stage electro-optic modulator, used to generate different group delays for light signals of different wavelengths, so that the product light signals belonging to the same output result are aligned in time, to obtain aligned light signals; M photodetectors are respectively set after each dispersive device, used to receive the aligned light signals and perform photoelectric conversion and optical power superposition on the aligned light signals, outputting the corresponding column of the product matrix C=AT. This scheme utilizes N channels with different wavelengths to carry weighted data in parallel, which is then distributed to M output branches to load each column of the input matrix, improving the parallelism and computational efficiency of matrix multiplication. Through cyclic rearrangement and dispersion delay, the product optical signals are automatically aligned and superimposed, achieving lossless summation of the product terms in the optical domain. Multiplication and addition operations are naturally completed in the optical domain, eliminating the need for electronic multiplication and addition units. The electronic part is only responsible for driving and control, significantly reducing computational overhead and system complexity.

[0071] Figure 4 The flowchart illustrates the steps of a fiber tensor calculation method based on dispersion delay and wavelength division multiplexing, as provided in an embodiment of the present invention.

[0072] like Figure 4 As shown, the method includes: Step 101: After the polychromatic light source generates polychromatic light containing N different wavelength channels, the polychromatic light is divided into N optical signals of different wavelengths using the first wavelength division multiplexer, and each optical signal is passed through N first-stage electro-optic modulators respectively. Step 102: Rearrange the weight matrix A into a loading matrix B according to the cyclic shift rule, and load the loading matrix B through the N first-stage electro-optic modulators; Step 103: Use the second wavelength division multiplexer to combine the N optical signals after the first stage electro-optic modulation into polychromatic light, then divide the polychromatic light into M paths, and let the resulting M polychromatic light signals pass through M second stage electro-optic modulators respectively. Step 104: Load each column of the input matrix T through the M second-stage electro-optic modulators respectively, and load each column element of the matrix T into the M output branches in sequence according to the time interval τ to obtain M product optical signals; Step 105: Pass the M-channel product optical signals through a dispersive device. The dispersive device applies different delays to the optical signals of different wavelengths in the M-channel product optical signals, so that the product optical signals belonging to the same output result are aligned in time to obtain aligned optical signals. Step 106: By measuring the time-domain information of the light intensity of the aligned light signal received by M photodetectors, obtain the product of the weight matrix A and the input matrix T.

[0073] In steps 101-106, the present invention maps the matrix multiplication C=A×T multiplication and addition operation to the optical domain through four steps: “wavelength parallel loading—time synchronization modulation—dispersion delay alignment—photoelectric detection summation”, without having to perform electronic multiplication and addition operations element by element.

[0074] Specifically, let matrix A be an N-order matrix, matrix T be an N×M matrix, and the target output be matrix C: .

[0075] The computational flow and optical path structure of this embodiment include a polychromatic light source, a first wavelength division multiplexer for beam splitting, N first-stage electro-optic modulators for loading matrix B, a second wavelength division multiplexer for beam combining, a 1×M optical beam splitter, M second-stage electro-optic modulators for loading matrix T, M segments of dispersive optical fiber, and M photodetectors. The two-stage electro-optic modulators are synchronously loaded according to a unified time slot, and the photodetectors sample in the corresponding output time slot.

[0076] Specifically, a polychromatic light source generates channels containing N different wavelengths ( , , …, ( ) polychromatic light.

[0077] The first wavelength division multiplexer decomposes the polychromatic light according to wavelength and distributes it to N output branches. The first wavelength division multiplexer distributes N wavelength channels to N first-stage branches, and each branch contains a first-stage electro-optic modulator.

[0078] N first-stage electro-optic modulators constitute the first-stage modulation of the system, completing the data loading of weight matrix A. Weight matrix A is rearranged into loading matrix B according to a cyclic shift rule, satisfying: The purpose of rearrangement is to distribute the N weight elements (A(r,1)…A(r,N)) belonging to the same output row r along a "circular diagonal" in matrix B, in preparation for subsequent dispersion alignment.

[0079] Within N consecutive time slots (each time slot width is τ), the i-th first-stage electro-optic modulator is sequentially loaded with B(i,1), B(i,2), …, B(i,N), that is, at wavelengths… λ i The data of the i-th row of matrix B is loaded in chronological order. Within N source time slots, N modulators jointly complete the loading of N×N elements of B; during continuous calculation, the loaded frames are sent at Nτ periods.

[0080] The second wavelength division multiplexer recombines the N optical signals of different wavelengths that have undergone the first stage modulation into a single polychromatic beam. At this point, each wavelength channel of the polychromatic beam carries its corresponding weighted data sequence.

[0081] The combined polychromatic light is split into M equal paths by a 1×M beam splitter and distributed to M parallel output branches. These M branches correspond to the M columns of the output matrix C, meaning each branch independently calculates one column of the output matrix.

[0082] M second-stage electro-optic modulators form the second-stage modulation of the system, completing the data loading of the input matrix T. The second-stage electro-optic modulator of the m-th branch loads the elements of the m-th column of the input matrix T sequentially in the same N consecutive time slots: T(1,m), T(2,m), …, T(N,m).

[0083] After two stages of modulation, the wavelength in the m-th branch λ i The optical power in the j-th time slot is proportional to the product B(i,j)×T(j,m). Since B(i,j)=A(r,j) (where r=1+mod(ji,N)), the multiplication operation of the corresponding matrix elements is actually implemented, that is, the multiplication is completed during the light intensity modulation process.

[0084] Each second-stage electro-optic modulator is followed by a dispersive fiber or a dispersive device with the same function. The core function of the dispersive device is to produce different propagation delays for optical signals of different wavelengths, and the delay is linearly related to the wavelength.

[0085] The system is equipped with dispersive devices to adjust the wavelength. λ i Generates group delay:

[0086] That is, adjacent wavelengths (such as λ) i With λ i+1 The time difference between the arrival times at the photodetector is exactly τ. This is equivalent to rearranging the time-domain signals of each wavelength channel.

[0087] The alignment mechanism is as follows: the N weight elements corresponding to a certain output row r are distributed along the circular diagonal of matrix B. Taking N=3 as an example, r=1 corresponds to B(1,1), B(2,2), and B(3,3), and these three products are located in time slot 1 of wavelength λ1, time slot 2 of wavelength λ2, and time slot 3 of wavelength λ3, respectively. After dispersion delay, λ1 is delayed by 2τ, λ2 by τ, and λ3 by 0, and all three arrive at the detector at exactly one output time slot.

[0088] A photodetector is placed after each dispersive device to complete the final summation and photoelectric conversion.

[0089] A photodetector performs square-law detection on optical signals, and its output photocurrent is proportional to the incident light power. When the product of optical signals from multiple wavelength channels arrives simultaneously, the detector automatically superimposes the optical power of each wavelength channel to obtain C(r,m).

[0090] As an optional embodiment, rearranging the weight matrix A into a loading matrix B according to a cyclic shift rule includes: Step 1021: Obtain the N-order weight matrix A, whose elements are denoted as A(r,j), where r=1,2,…,N are row indices and j=1,2,…,N are column indices; Step 1022: Circularly shift each column of the weight matrix A according to the cyclic shift rule to obtain the loading matrix B; where the element B(i,j) in the loading matrix B is B(1+mod(ji,N),j), and mod(·) represents the modulo operation with a period of N; Step 1023: Output loading matrix B, in which N elements belonging to the same output row r satisfy r=1+mod(ji,N) and are distributed along the same cyclic diagonal, which is used to cooperate with the dispersion delay to align the corresponding product optical signals in the same output time slot.

[0091] In steps 1021-1023, let matrix A be an N-order matrix, matrix T be an N×M matrix, and the target output be matrix C: .

[0092] This invention transforms the weight matrix A into a wavelength-time loading matrix B. Row index i in B corresponds to the wavelength. And for the i-th first-stage electro-optic modulator, column index j corresponds to the source time slot, and B satisfies: .

[0093] make ,but For a given output row r and wavelength i, the source time slots that need to participate in the summation are: Therefore, within one N time slot period, the N product terms belonging to the same output row r are distributed along a cyclic diagonal of B.

[0094] In summary, the fiber tensor calculation method based on dispersion delay and wavelength division multiplexing provided in this embodiment of the invention includes: after a polychromatic light source generates polychromatic light containing N different wavelength channels, the polychromatic light is divided into N optical signals of different wavelengths using a first wavelength division multiplexer, and each optical signal is passed through N first-stage electro-optic modulators; the weight matrix A is rearranged into a loading matrix B according to a cyclic shift rule, and the loading matrix B is loaded through the N first-stage electro-optic modulators; the N optical signals after first-stage electro-optic modulation are combined into polychromatic light using a second wavelength division multiplexer, and the polychromatic light is then divided into M paths, and the resulting M paths are... The colored light signals are passed through M second-stage electro-optic modulators. Each column of the input matrix T is loaded through the M second-stage electro-optic modulators, and each element of the column in matrix T is sequentially loaded into the M output branches according to a time interval τ, resulting in M ​​product light signals. These M product light signals are then passed through a dispersive device, which applies different delays to the light signals of different wavelengths in the M product light signals, aligning the product light signals belonging to the same output result in time, thus obtaining aligned light signals. The product of the weight matrix A and the input matrix T is obtained by measuring the temporal information of the light intensity of the aligned light signals received by M photodetectors. This scheme utilizes N channels with different wavelengths to carry weighted data in parallel, which is then distributed to M output branches to load each column of the input matrix, improving the parallelism and computational efficiency of matrix multiplication. Through cyclic rearrangement and dispersion delay, the product optical signals are automatically aligned and superimposed, achieving lossless summation of the product terms in the optical domain. Multiplication and addition operations are naturally completed in the optical domain, eliminating the need for electronic multiplication and addition units. The electronic part is only responsible for driving and control, significantly reducing computational overhead and system complexity.

[0095] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0096] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0097] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0098] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A fiber tensor computation system based on dispersion delay and wavelength division multiplexing, used to calculate the product C=AT of an N-order weight matrix A and an N×M input matrix T, characterized in that, The system includes: Polychromatic light source, used to generate polychromatic light containing N different wavelength channels; The first wavelength division multiplexer is connected to the polychromatic light source and is used to divide the polychromatic light into N optical signals of different wavelengths; N first-stage electro-optic modulators are respectively disposed on the N output branches of the first wavelength division multiplexer, and are used to load the loading matrix B obtained by cyclic rearrangement of the weight matrix A; the cyclic rearrangement is used to distribute the N weight elements belonging to the same output row in the weight matrix A along a cyclic diagonal in the loading matrix B. The second wavelength division multiplexer is connected to the output of the N first-stage electro-optic modulators and is used to combine the N optical signals of different wavelengths into polychromatic light. A 1×M optical beam splitter is connected to the output of the second wavelength division multiplexer to split the combined polychromatic light into M paths, thereby obtaining M polychromatic light signals. M second-stage electro-optic modulators are respectively disposed on the M output branches of the 1×M optical beam splitter to receive the M polychromatic light signals and load each column of the input matrix T. The elements of each column of the matrix T are loaded into the M output branches in sequence according to the time interval τ to obtain M product light signals. M dispersive devices are respectively set after each second-stage electro-optic modulator to generate different group delays for optical signals of different wavelengths, so that the product optical signals belonging to the same output result are aligned in time to obtain aligned optical signals. M photodetectors are respectively positioned after each dispersive device to receive the alignment light signal, perform photoelectric conversion and optical power superposition on the alignment light signal, and output the corresponding column of the product matrix C=AT.

2. The system according to claim 1, characterized in that, The center wavelengths of the N wavelength channels generated by the polychromatic light source are respectively to The N output branches of the first wavelength division multiplexer correspond one-to-one with the N wavelength channels.

3. The system according to claim 1, characterized in that, The loading matrix B satisfies ,in This represents a modulo operation with a period of N; the i-th first-stage electro-optic modulator is used to load B(i,j) in the j-th source time slot, and completes the synchronous loading of N×N elements in the loading matrix B within a time range of N×τ, where τ is the time slot width of a single data element.

4. The system according to claim 1, characterized in that, The m-th second-stage electro-optic modulator is used to load the m-th column element T(j,m) of the input matrix T in the j-th source time slot, and to complete the synchronous loading of N×M elements in the input matrix T within the time range of N×τ, where m=1,2,…,M and j=1,2,…,N.

5. The system according to claim 1, characterized in that, The dispersive device is a dispersive fiber or a fiber Bragg grating, and the dispersive device makes the wavelength... Optical signal generation group delay The time difference between the arrival of adjacent wavelength optical signals at the corresponding photodetector is τ, where i = 1, 2, ..., N.

6. The system according to claim 1, characterized in that, The m-th photodetector is used to receive and superimpose the product optical signals of N wavelength channels in the output time slot corresponding to the r-th row of the product matrix C. The N product optical signals correspond to the product of each element in the r-th row of the weight matrix A and the corresponding element in the m-th column of the input matrix T, and the superposition result is C(r,m). The M photodetectors output the M columns of the product matrix C, where r=1,2,…,N and m=1,2,…,M.

7. The system according to claim 1, characterized in that, The system also includes a synchronization controller, which is connected to the N first-level electro-optic modulators and the M second-level electro-optic modulators respectively. The synchronization controller is used to provide a unified time slot synchronization signal to all electro-optic modulators and control the loading matrix B and the input matrix T to be loaded synchronously according to the same source time slot sequence.

8. A method for calculating fiber tensors based on dispersion delay and wavelength division multiplexing, characterized in that, The method includes: After the polychromatic light source generates polychromatic light containing N different wavelength channels, the polychromatic light is divided into N optical signals of different wavelengths using a first wavelength division multiplexer, and each optical signal is passed through N first-stage electro-optic modulators respectively. The weight matrix A is rearranged into a loading matrix B according to the cyclic shift rule, and the loading matrix B is loaded through the N first-stage electro-optic modulators; The N optical signals after the first-stage electro-optic modulation are combined into polychromatic light using a second wavelength division multiplexer. The polychromatic light is then divided into M channels, and the resulting M polychromatic light signals are passed through M second-stage electro-optic modulators respectively. The M second-stage electro-optic modulators are used to load each column of the input matrix T, and the elements of each column of the matrix T are loaded into the M output branches in sequence according to the time interval τ to obtain M product optical signals. The M-channel product optical signals are passed through a dispersive device, which delays the optical signals of different wavelengths in the M-channel product optical signals in different ways, so that the product optical signals belonging to the same output result are aligned in time to obtain aligned optical signals. By measuring the temporal information of the light intensity of the aligned light signal received by M photodetectors, the product of the weight matrix A and the input matrix T is obtained.

9. The method according to claim 8, characterized in that, The step of rearranging the weight matrix A into a loading matrix B according to the cyclic shift rule includes: Obtain the N-order weight matrix A, whose elements are denoted as A(r,j), where r=1,2,…,N are row indices and j=1,2,…,N are column indices; According to the cyclic shift rule, each column of the weight matrix A is cyclically shifted to obtain the loading matrix B; where the element B(i,j) in the loading matrix B is B(1+mod(ji,N),j), and mod(·) represents the modulo operation with a period of N; The output loading matrix B contains N elements belonging to the same output row r that satisfy r = 1 + mod (ji, N) and are distributed along the same cyclic diagonal. This matrix is ​​used in conjunction with the dispersion delay to align the corresponding product optical signals in the same output time slot.

Images