Full-simulation time sequence prediction reserve pool calculation system based on dispersion effect

Abstract

The invention discloses a full-simulation time sequence prediction reserve pool calculation system based on a dispersion effect. The system comprises a light source array module, an input layer module, an all-optical reserve pool layer module and a photoelectric simulation readout layer module, the input layer module is used for receiving an external input signal, converting the external input signal into an electric modulation signal and transmitting the electric modulation signal to the all-optical reserve pool layer module; the all-optical reserve pool layer module loads a modulation signal to an optical signal, forms an optical pulse information sequence through short-term nonlinear memory processing of time sequence information, and transmits the optical pulse information sequence to the photoelectric analog readout layer module; and the photoelectric simulation readout layer module is used for analyzing and reading information in the all-optical storage layer module, obtaining a corresponding weight value through weight training, and executing linear regression operation to obtain an output result. The pulse broadening effect caused by the dispersion device is utilized to realize the storage layer, the linear regression operation of the photoelectric simulation readout layer is realized through the photoelectric simulation readout layer module, and the accuracy and the response speed of time sequence prediction are improved.

Description

Technical Field

[0001] The invention relates to a photon information processing technology, in particular to a full-simulation time series prediction reserve pool computing system based on dispersion effect. Background Art

[0002] With the continuous development of science and technology, time series forecasting technology plays a vital role in many fields, including finance, meteorology, medicine, etc. However, traditional time series forecasting methods often perform poorly when faced with variable, high-dimensional time series data. Specific challenges include reliance on a large amount of historical data, high computational complexity, and model limitations.

[0003] The dispersion effect is an important phenomenon in optics, which refers to the different propagation speeds of light waves of different frequencies in a medium. This feature has been widely used in the fields of spectroscopy and optical communications. The technical background of this patent highlights the innovative idea of ​​introducing the dispersion effect into the time series prediction system, in order to optimize the processing of time series data through optical propagation. In addition, as an emerging computing theory, reservoir computing has broad application prospects. In time series prediction, by constructing a reservoir with rich dynamic behaviors, the complex patterns in time series data can be better captured. At the same time, the reservoir computing has strong fault tolerance and scalability, which provides a guarantee for the stability and reliability of the system. The innovation of this patent is to integrate the dispersion effect with the reservoir computing. By introducing the dispersion effect in the reservoir, the system can respond more sensitively to the time series changes of different frequencies. This fusion brings richer information expression capabilities to the time series prediction system, which is expected to overcome the limitations of the traditional cyclic delay loop method. More importantly, the scalability of the time series prediction system is also an indispensable consideration in practical applications. Therefore, the full-analog time series prediction system based on dispersion effect realizes information processing through optical propagation on the basis of reserve pool calculation, and has strong scalability. This feature enables the system to adapt to time series data of different scales and complexities. Summary of the invention

[0004] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a reserve pool computing system for feedforward propagation of fully simulated timing prediction based on dispersion effect. In order to overcome these limitations, we propose a reserve pool computing system for feedforward propagation of fully simulated timing prediction based on dispersion effect, which aims to improve the accuracy, speed and scalability of timing prediction.

[0005] The technical solution of the present invention is as follows:

[0006] A full-analog time series prediction reserve pool calculation system based on dispersion effect, characterized in that it includes: a light source array module, an input layer module, an all-optical reserve pool layer module and a photoelectric analog readout layer module;

[0007] The light source array module is used to provide light source for the entire system.

[0008] The input layer module is used to receive external input signals and convert them into electrical modulation signals for transmission to the all-optical reservoir layer module;

[0009] The all-optical reservoir layer module is used to receive the modulated signal from the input layer module and load it onto the optical signal of the light source array module, form an optical pulse information sequence through short-term nonlinear memory processing of timing information, and transmit it to the optoelectronic analog readout layer module;

[0010] The photoelectric analog readout layer module is used to parse and read the information in the all-optical storage layer module, obtain corresponding weight values ​​through weight training, and perform linear regression operations to obtain output results.

[0011] Furthermore, the all-optical reservoir layer module at least includes an optoelectronic modulator and a dispersion device; the output end of the light source array module is connected to the first input end of the optoelectronic modulator, the output end of the output layer module is connected to the second input end of the optoelectronic modulator, and the output end of the optoelectronic modulator is connected to the optoelectronic analog readout layer module via the dispersion device, so that the waveform affected by the dispersion is subjected to linear regression via the optoelectronic analog readout layer module.

[0012] Furthermore, the all-optical reservoir layer module also includes an optical amplifier, the input end of which is connected to the output end of the dispersion device, and is used to amplify the gain of the signal and then transmit it to the optoelectronic analog readout layer module.

[0013] Further, the optoelectronic analog readout layer module includes a demultiplexer, a delay array module composed of N delay array units, a variable attenuator array composed of N delay array units, and a combiner;

[0014] The output end of the optical amplifier is connected to the input end of the demultiplexer. The demultiplexer divides light waves of different frequencies into N optical signals. The N output ends of the demultiplexer are connected to the input ends of N delay array units to align the signals of the link. The output end of each delay array unit is connected to the input end of each corresponding delay array unit to realize the multiplication operation of the linear regression operation. The attenuation value of the variable attenuator is used to realize the multiplication factor of the weight. The output end of each delay array unit is connected to the input end of the combiner. After synthesizing one signal, it is transmitted to the photoelectric detector. The photoelectric detector converts the optical signal into an electrical signal, and at the same time adds the aligned optical wave signals of different frequencies, and finally outputs the combined signal to obtain the simulation prediction result.

[0015] Preferably, the light source array module adopts a passive mode-locked laser, an active mode-locked laser, a distributed feedback Bragg semiconductor laser (DFB-LD) and a vertical cavity surface emitting laser (VCSEL).

[0016] Preferably, the input layer module uses but is not limited to intensity modulated signals, such as Gaussian pulse shapes or sinc pulse signals.

[0017] Preferably, the electro-optical modulator adopts but is not limited to a lithium niobate electro-optic modulator, a polymer electro-optic modulator, a silicon-based integrated electro-optic modulator, an acousto-optic modulator or a spatial light modulator.

[0018] Preferably, the combiner is, but not limited to, a wavelength division multiplexer.

[0019] Preferably, the dispersion device includes but is not limited to a prism, a grating, a dispersion prism, an optical fiber, a wave plate, a grating coupler and a dispersion compensator.

[0020] Preferably, the optical amplifier adopts but is not limited to a fiber Raman amplifier, an erbium-doped fiber amplifier, a semiconductor optical amplifier and an erbium-doped semiconductor laser.

[0021] Preferably, the demultiplexer is, but not limited to, a wavelength division multiplexer.

[0022] Preferably, the optical signal delay module adopts but is not limited to an optical fiber delay line, an optical fiber gyroscopic mirror, a waveguide delay device, an optical unit waveguide, a phase modulator and a tunable optical filter.

[0023] Preferably, the adjustable attenuation module adopts but is not limited to an electronic adjustable optical attenuator, a mechanical adjustable optical attenuator, an optical adjustable attenuator, an adjustable grating and a waveguide adjustable attenuator.

[0024] Preferably, the photodetector includes but is not limited to a photodiode, a photoconductive detector and a quantum dot detector.

[0025] Compared with the prior art, the present invention has the following advantages:

[0026] 1) Improve prediction accuracy: By introducing the dispersion effect, the system can more sensitively capture the morphological changes of the timing broadening of the pulse signal, thereby improving the prediction accuracy.

[0027] 2) Efficient information processing: Figure 2 The reservoir network architecture shown is a feedforward propagation method, which allows information to propagate in one direction in the network architecture without a feedback loop. This feedforward propagation network architecture has the advantages of simplicity, stability, training efficiency, and parallelism. Especially when processing static input data, the system exhibits efficient information processing capabilities and can also process complex time series data.

[0028] 3) Real-time and scalability: The system ensures high efficiency in real-time tasks through parallel computing and hardware acceleration, while its scalable design makes it suitable for data processing of different scales.

[0029] 4) This invention highlights the fusion innovation of dispersion effect, reservoir calculation and feedforward propagation, and provides a new idea for solving the limitations of traditional time series prediction methods. By introducing the dispersion effect of optical propagation, we hope to achieve more accurate and efficient time series prediction, thereby promoting scientific and technological progress in multiple fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a schematic block diagram of the implementation of the full-simulation timing prediction reserve pool computing system architecture based on dispersion effect of the present invention.

[0031] Figure 2 The present invention is a flowchart of the overall architecture experiment of the full simulation timing prediction reserve pool computing system architecture based on dispersion effect.

[0032] Figure 3 The present invention is a schematic diagram of the comparison between the predicted output of the timing signal of the fully analog timing prediction reserve pool computing system based on the dispersion effect and the output of the computer digital calculation; wherein, the numerical prediction (a) and the corresponding simulation phase space (b) are performed using the same settings as the experiment; the prediction results are from the experimental data collected by the fully analog timing prediction reserve pool computing system architecture based on the dispersion effect (c), and the phase space of the corresponding simulation data (d). DETAILED DESCRIPTION

[0033] A specific embodiment of the present invention is given below in conjunction with the accompanying drawings. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and process, but the protection scope of the present invention is not limited to the following embodiment.

[0034] The present invention includes two main parts: optical device configuration and data processing algorithm. The first part, optical device configuration: the optical device configuration in the invention includes but is not limited to optical fiber, frequency modulator, grating, etc. The design of these devices takes into account the optimization of dispersion effects to ensure that the pulse-shaped light wave signal can be fully modulated and responded during the optical propagation process. By introducing the dispersion effect, we have achieved efficient processing of time series data. The second part, the introduction of the feedforward propagation network architecture: information can only propagate in one direction in the network, and will not form a circular connection, that is, there is no feedback loop. This architecture is suitable for efficient processing of complex time series data. Figure 2The feedforward reservoir architecture shown is the core part of the present invention, which combines the advantages of reservoir calculation and feedforward propagation. Specifically, the following steps are included: Reservoir construction: Create a reservoir with rich dynamic behavior to provide nonlinear mapping of time series data. Introduction of dispersion effect: The dispersion effect is matched with the state of the reservoir to achieve sensitivity to different frequency components, and its pulse broadening effect forms a memory characteristic in which time series signals are intertwined with each other. Feedforward propagation architecture: Introduce a feedforward propagation structure in the reservoir to achieve hierarchical and level-by-level information transmission. Dynamic adjustment: Dynamically adjust the state of the reservoir according to actual data to adapt to changes in different time series patterns. The present invention aims to introduce optical dispersion effects and combine the theory of feedforward propagation to construct a fully analog time series prediction system to improve the accuracy and real-time performance of time series prediction.

[0035] See also Figure 1 , Figure 1 Schematic diagram of an embodiment of a full-simulation time series prediction reserve pool computing system based on dispersion effect of the present invention, Figure 2 This is the overall architecture diagram of the full simulation time series prediction reserve pool computing system based on dispersion effect. Figure 1 It can be seen that the full-analog timing prediction reserve pool calculation system architecture based on the dispersion effect of the present invention includes a light source array module 1, an input layer module 2, an all-optical reserve pool layer module 3 and an optical readout layer 4. The output end of the light source array module 1 is connected to the first input end of the photoelectric modulator 3-1 in the all-optical reserve pool layer module 3, and the output end of the output layer module 2 is connected to the second input end of the photoelectric modulator 3-1 in the all-optical reserve pool layer module 3. The output end of the photoelectric modulator 3-1 is obtained by the optical dispersion device 3-2 to obtain an output waveform affected by dispersion. Finally, the waveform passes through the optical readout layer module 4, performs a linear regression operation, and finally obtains the output result.

[0036] The light source array module is composed of n light source module units arranged in parallel, the all-optical storage layer module is composed of an optoelectronic modulator and an optical dispersion device, the optical readout layer module is composed of a demultiplexer, a delay array module, an adjustable attenuator array module, a combiner and a photodetector, the delay array module is composed of n optical fiber delay units arranged in parallel, and the adjustable attenuator module is composed of n adjustable attenuation module units arranged in parallel.

[0037] In the light source array module, the n output ends of the light source array module are connected to the input end of the combiner to synthesize one output, one output end of the combiner is connected to the input end of the polarization controller, and the output end of the polarization controller is connected to the first input end of the optoelectronic modulator.

[0038] In the all-optical storage layer module, the output end of the input layer module is connected to the second input end of the photoelectric modulator, the output end of the photoelectric modulator is connected to the input end of the dispersion device, the output end of the dispersion device is connected to the input end of the optical amplifier, and the output end of the optical amplifier is connected to the input end of the photoelectric analog readout layer module.

[0039] In the photoelectric analog readout layer module, the signal input demultiplexer is divided into n-wavelength signals, and each signal passes through the delay array module. Then, the output end of the delay array module is connected to the input end of the corresponding adjustable attenuator, and the n-way input signal of the delay array module is multiplied with the signal of the adjustable attenuation module. The n output ends of the adjustable attenuation array module are connected to the input end of the combiner described later, and the output end of the combiner is connected to the input end of the photoelectric detector. Its function is to add the signals after the delay alignment multiplication, and finally the signal can be obtained from the output of the photoelectric detector.

[0040] exist Figure 2In the overall structure diagram of the full analog timing prediction reserve pool computing system based on the dispersion effect, the optical paths of the signal laser array module 1 are respectively input into the N-way combiner 1-2, and the optical signals of multiple paths are combined into one signal, and then the output end of the combiner 1-2 is connected to the input end of the polarization controller 1-3, and the polarization controller 1-3 is mainly used to adjust and control the polarization state of the light wave, and then the output end of the polarization controller 1-3 is connected to the first input end of the photoelectric modulator 3-1, and the output end of the input layer module 2 is connected to the second input end of the photoelectric modulator 3-1, wherein the photoelectric modulator 3-1 will modulate the electrical pulse modulation signal of the input layer module 2 onto the light, and on the other hand, the photoelectric modulator 3-1 will realize a sin nonlinear adjustment of the input layer module. Then, the output end of the photoelectric modulator 3-1 is connected to the input end of the dispersion device 3-2, and the main function of the dispersion device is to process the dispersion effect caused by the light pulse waveform of different wavelengths. Then, the output end of the dispersion device 3-2 is connected to the input end of the optical amplifier 3-3. The function of the amplification module is to gain amplify the signal. Then, the output end of the optical amplifier 3-3 is connected to the input end 4-1 of the demultiplexer. The demultiplexer 4-1 divides the light waves of different frequencies into N optical signals. Then, the N output ends of the demultiplexer 4-1 are connected to the N input ends of the delay array module 4-3. The delay array module 4-3 is composed of the delay array unit 4-2. The function of the delay array module 4-3 is to perform signal alignment of the link. Then, the N output end of the delay array module 4-3 is connected to the variable attenuator array 4-5. The N input ends are connected, the variable attenuator array 4-5 is composed of the delay array unit 4-4, wherein the function of the variable attenuator array 4-5 is to realize the multiplication operation of the linear regression operation, and the attenuation value of the variable attenuator is used to realize the multiplication factor of the weight, and then the output end of the variable attenuator array 4-5 is connected to the input end of the combiner 4-6 to synthesize a signal, and then the output end of the combiner 4-6 is connected to the input end of the photodetector 4-7, and the photodetector 4-7 performs signal conversion from light to electricity on the one hand, and on the other hand, adds the aligned light wave signals of different frequencies, and finally outputs the combined signal to obtain the simulation prediction result.

[0041] The light source array module can be implemented by, but not limited to, passive mode-locked lasers, active mode-locked lasers, distributed feedback Bragg semiconductor lasers (DFB-LDs), and vertical cavity surface emitting lasers (VCSELs). The sampled signal is, but not limited to, an intensity modulated signal, such as a Gaussian pulse shape or a pulse signal of sinc. The optoelectronic modulator is, but not limited to, a lithium niobate electro-optic modulator, a polymer electro-optic modulator, a silicon-based integrated electro-optic modulator, an acousto-optic modulator, or a spatial light modulator. The combiner is, but not limited to, a wavelength division multiplexer. The dispersion device is, but not limited to, a prism, a grating, a dispersion prism, an optical fiber, a wave plate, a grating coupler, and a dispersion compensator. The optical amplifier is, but not limited to, a fiber Raman amplifier, an erbium-doped fiber amplifier, a semiconductor optical amplifier, and an erbium-doped semiconductor laser. The demultiplexer is, but not limited to, a wavelength division multiplexer. The optical signal delay module adopts but is not limited to optical fiber delay line, optical fiber gyroscopic mirror, waveguide delay, optical unit waveguide, phase modulator and tunable optical filter. The adjustable attenuation module adopts but is not limited to electronic adjustable optical attenuator, mechanical adjustable optical attenuator, optical adjustable attenuator, adjustable grating and waveguide adjustable attenuator. The photoelectric detector adopts but is not limited to photodiode, photoconductive detector and quantum dot detector. The all-optical storage layer module is to make the input pulse morphology signal affected by dispersion, and the timing distortion problem occurs in the transmission, especially the pulse broadening phenomenon caused by dispersion. The specific process is to make corresponding timing changes to the shape of the input pulse signal and change the timing characteristics of the input signal. The photoelectric analog readout layer module is to perform linear regression multiplication and addition operations on the signal. Specifically, the adjustable attenuator is used to adjust the intensity of the input signal. By adjusting the parameters of the adjustable attenuator, the amplitude of the input signal can be attenuated, that is, multiplication operation. The delay line introduces time delay to ensure that the signal can arrive at the right time during the transmission process. In the analog readout layer module, the delay line is used to align the different signals so that they are added at the corresponding time. The final photodetector adds these signals at the corresponding time. In short, the purpose of this module is to process the input signals and finally implement the multiplication and addition operation of the linear regression of these signals by adjusting the intensity and time delay.

[0042] Using the above-mentioned full-simulation timing prediction reserve pool computing system architecture and design method based on dispersion effect, the solution includes the following steps:

[0043] 1) The input layer module inputs the sampled electrical analog signal into the photoelectric modulator. The sampled electrical analog signal is obtained by multiplying the randomly generated mask with the original signal and is loaded onto the optical signal of the light source array module. This process generates a sequence of optical pulses of analog signal information.

[0044] 2) The electrical analog signal is converted into an optical pulse information sequence through the photoelectric modulator, and then passes through the all-optical storage layer module. Subsequently, weight training is performed according to the output response signal affected by dispersion to obtain the corresponding weight value. The weight training method includes but is not limited to the following, and the training weight is shown in the formula: w out =(XgX T +λgI) -1 (XgY), where X is the output of the reserve state, Y is the label matrix, and λ<<1 is a small regularization constant. I is used to create an identity matrix, which is a square matrix with 1 on the main diagonal and 0 elsewhere, aiming to introduce regularization terms to enhance the generalization ability of the model and prevent overfitting.

[0045] 3) In the test phase, the adjustable light attenuation array module is adjusted according to the weight value obtained by training in step 2. Finally, a linear regression operation is performed in the configured analog output layer module, and the formula is as follows: As shown, y(n) represents the output signal, x(n) represents the reserve state signal, N is the number of reserve nodes, and W i is the only parameter that needs to be trained in the system, that is, the weight value obtained by training in step 2. Thus, the final output result is obtained.

[0046] By using a fully analog timing prediction reserve pool computing system based on the dispersion effect, it is possible to realize real-time processing of timing signals in the optical domain, and ultimately obtain accurate timing signal predictions. Taking the Mackey-Glass chaotic signal as an example, accurate results can be obtained through computer numerical simulation and real-time simulation prediction of our architecture. It is worth noting that compared with traditional digital domain numerical prediction methods, the method of the present invention has a faster real-time processing rate and achieves more accurate and efficient fully analog timing predictions. This innovation promotes scientific and technological progress in the field of analog optical computing and provides new possibilities for more advanced timing signal processing and prediction.

Claims

1. A full-simulation time series prediction reserve pool calculation system based on dispersion effect, characterized in that: It comprises: a light source array module (1), an input layer module (2), an all-optical storage pool layer module (3) and a photoelectric analog readout layer module (4); The light source array module (1) is used to provide light source for the entire system. The input layer module (2) is used to receive an external input signal and convert it into an electrical modulation signal for transmission to the all-optical reservoir layer module (3); The all-optical reservoir layer module (3) is used to receive the modulated signal from the input layer module (2), load it onto the optical signal of the light source array module (1), form an optical pulse information sequence through short-term nonlinear memory processing of the timing information, and transmit it to the photoelectric analog readout layer module (4); The photoelectric analog readout layer module (4) is used to parse and read the information in the all-optical storage layer module (3), obtain corresponding weight values ​​through weight training, and perform a linear regression operation to obtain an output result.

2. The full-simulation time series prediction reserve pool computing system based on dispersion effect according to claim 1 is characterized in that: The all-optical storage pool layer module (3) at least includes an optoelectronic modulator (3-1) and a dispersion device (3-2); the output end of the light source array module (1) is connected to the first input end of the optoelectronic modulator (3-1), the output end of the output layer module (2) is connected to the second input end of the optoelectronic modulator (3-1), and the output end of the optoelectronic modulator (3-1) is connected to the optoelectronic analog readout layer module (4) via the dispersion device (3-2), so that the waveform affected by the dispersion is subjected to linear regression via the optoelectronic analog readout layer module (4).

3. The full-simulation time series prediction reserve pool computing system based on dispersion effect according to claim 2 is characterized in that: The all-optical storage pool layer module (3) also includes an optical amplifier (3-3), the input end of which is connected to the output end of the dispersion device (3-2) and is used to amplify the gain of the signal and then transmit it to the optoelectronic analog readout layer module (4).

4. The full-simulation time series prediction reserve pool computing system based on dispersion effect according to claim 3 is characterized in that: The photoelectric analog readout layer module (4) comprises a demultiplexer (4-1), a delay array module (4-3) composed of N delay array units (4-2), a variable attenuator array (4-5) composed of N delay array units (4-4), and a combiner (4-6); The output end of the optical amplifier (3-3) is connected to the input end of the demultiplexer (4-1). The demultiplexer (4-1) divides light waves of different frequencies into N optical signals. The N output ends of the demultiplexer (4-1) are connected to the input ends of N delay array units (4-2) to align the signals of the link. The output end of each delay array unit (4-2) is connected to the input end of each corresponding delay array unit (4-4) to implement the multiplication operation of the linear regression operation. The attenuation value of the variable attenuator is used to implement the multiplication factor of the weight. The output end of each delay array unit (4-4) is connected to the input end of the combiner (4-6). After synthesizing one signal, it is transmitted to the photoelectric detector (4-7). The photoelectric detector (4-7) converts the optical signal into an electrical signal and adds the aligned optical wave signals of different frequencies. Finally, the combined signal is output to obtain a simulation prediction result.

5. The full-simulation time series prediction reserve pool computing system based on dispersion effect according to claim 1 is characterized in that: The light source array module (1) adopts a passive mode-locked laser, an active mode-locked laser, a distributed feedback Bragg semiconductor laser (DFB-LD) and a vertical cavity surface emitting laser (VCSEL).

6. The full-simulation time series prediction reserve pool computing system based on dispersion effect according to claim 1 is characterized in that: The input layer module (2) uses but is not limited to intensity modulated signals, such as Gaussian pulse shapes or sinc pulse signals; the optoelectronic modulator (3-1) uses but is not limited to lithium niobate electro-optic modulators, polymer electro-optic modulators, silicon-based integrated electro-optic modulators, acousto-optic modulators or spatial light modulators; the combiner (1-2) uses but is not limited to wavelength division multiplexers; the dispersion device (3-2) uses but is not limited to prisms, gratings, dispersion prisms, optical fibers, wave plates, grating couplers and dispersion compensators; the optical amplifier (3-3) uses but is not limited to fiber Raman amplifiers, erbium-doped fiber amplifiers, semiconductor optical amplifiers and erbium-doped semiconductor lasers.

7. The full-simulation time series prediction reserve pool computing system based on dispersion effect according to claim 1 is characterized in that: The demultiplexer (4-1) adopts but is not limited to a wavelength division multiplexer; the optical signal delay module (4-3) adopts but is not limited to an optical fiber delay line, an optical fiber gyroscopic mirror, a waveguide delay device, an optical unit waveguide, a phase modulator and a tunable optical filter; the adjustable attenuation module (4-5) adopts but is not limited to an electronically adjustable optical attenuator, a mechanically adjustable optical attenuator, an optically adjustable attenuator, an adjustable grating and a waveguide adjustable attenuator; the photodetector (4-7) adopts but is not limited to a photodiode, a photoconductive detector and a quantum dot detector.

Images