CN-122021758-A - Optical neural network topology self-adaptive mode division multiplexing communication system and training method

CN122021758ACN 122021758 ACN122021758 ACN 122021758ACN-122021758-A

Abstract

The application relates to a mode division multiplexing communication system with self-adaptive optical neural network topology and a training method, which are applied to the mode division multiplexing communication system and are used for solving the problem that the transmission or calculation performance is reduced due to dynamic coupling crosstalk of a space mode caused by environmental disturbance. The method comprises the steps of monitoring optical neural network output in real time, generating a state matrix representing mode crosstalk, extracting matrix characteristics, constructing an environment vector, deciding and generating a reconstruction action for controlling an adjustable optical sub-device according to the vector through a pre-trained deep reinforcement learning network, dynamically adjusting network physical topology by a driving device to compensate the crosstalk, and finally optimizing a strategy network on line based on a reconstructed performance evaluation result to form a closed loop. The optical neural network in the mode division multiplexing has the online self-adaptive capacity, can continuously inhibit the dynamic mode crosstalk, and ensures the stability and the high performance of the system in actual deployment.

Inventors

LIU BO
HU LIUYANG
LIN WEI
LI CHANGJIN
JIA FAN
LIU HAIFENG
ZHANG HAO

Assignees

南开大学

Dates

Publication Date: 20260512
Application Date: 20260209

Claims (8)

1. The optical neural network topology self-adaptive mode division multiplexing training method is characterized by comprising the following steps of: S1, performing mode coupling state monitoring on a light field signal at an output end of an optical neural network, analyzing mode power and phase information in the signal to calculate a coupling relation between modes, and generating a mode coupling matrix in a current environment; S2, extracting environmental features from the mode coupling matrix, extracting non-diagonal elements of the mode coupling matrix to form crosstalk features, calculating matrix overall variances, splicing crosstalk feature vectors and variance values into comprehensive feature representations, and generating environmental feature vectors; S3, performing reinforcement learning strategy reasoning on the environment feature vector based on a pre-trained deep neural network, performing nonlinear transformation on the environment feature vector through a plurality of full-connection layers and an activation function of the deep neural network, mapping a final layer output value to a physical parameter range of an adjustable optical sub-device, and generating a topology reconstruction action vector; s4, performing photoelectric signal conversion and driving on the topology reconstruction motion vector, converting digital control parameters into analog voltage signals, driving a phase modulator, and generating a reconstructed physical topology of the optical neural network; And S5, based on the physical topological row mode coupling state monitoring of the optical neural network, a mode coupling matrix of the next period is obtained, the mode coupling matrix of the next period and the topological reconstruction motion vector are input into a preset rewarding function to be calculated to obtain a scalar rewarding value, and the strategy gradient updating is carried out on the deep neural network by combining the environment feature vector and the topological reconstruction motion vector, so that an updated deep neural network is generated.
2. The method according to claim 1, wherein S1 comprises: S11, performing photoelectric conversion and digital sampling processing on an optical field signal at the output end of the obtained optical neural network, converting the optical signal into an electric signal, performing high-speed analog-to-digital conversion, and generating a parallel digital signal sequence; S12, carrying out modal amplitude and phase estimation processing on the digital signal sequence, solving complex field distribution of each mode through a digital signal processor, and generating complex amplitude information of each optical mode; And S13, performing cross-correlation calculation processing on the complex amplitude information, analyzing the relation between energy transfer and phase delay among different modes, and generating a mode coupling matrix for representing the coupling strength among the modes.
3. The method according to claim 1, wherein S2 comprises: s21, extracting off-diagonal elements of the mode coupling matrix, removing diagonal elements of the matrix to filter self-coupling components, and generating a non-ideal coupling deviation matrix; s22, carrying out vectorization and statistical feature calculation on the non-ideal coupling deviation matrix, expanding the non-ideal coupling deviation matrix into a one-dimensional sequence, and calculating the overall fluctuation degree of the non-ideal coupling deviation matrix to generate vectorization deviation sequence and overall variance value; And S23, performing splicing and transposition processing on the vectorized deviation sequence and the overall variance value, combining the one-dimensional sequence and the scalar variance into a comprehensive characteristic column vector, and generating an environment characteristic vector.
4. The method according to claim 1, wherein S3 comprises: S31, performing strategy network forward computation on the environment feature vector, extracting abstract features layer by layer through a plurality of full-connection layers and nonlinear activation functions, and generating high-dimensional feature representation; S32, carrying out output layer normalization on the high-dimensional characteristic representation, mapping the characteristic to a target dimension by using a linear output layer, and limiting a value range within a preset range by using a hyperbolic tangent function to generate an un-normalized original action parameter; and S33, carrying out vector packaging and formatting processing on the unnormalized original action parameters, and combining control parameters of all the adjustable devices according to a preset sequence to generate a topological reconstruction action vector capable of directly driving hardware.
5. The method according to claim 1, wherein S4 comprises: s41, carrying out instruction analysis on the topology reconstruction motion vector, and outputting a digital voltage instruction set corresponding to a plurality of phase modulators; S42, performing digital-to-analog conversion and signal amplification processing on the digital voltage instruction set, converting a digital voltage value into high-precision analog voltage, amplifying the high-precision analog voltage to a level sufficient for driving a phase modulator, and generating an analog voltage driving signal; And S43, performing pulse modulation and motor driving processing on the analog voltage driving signal, converting a digital angle instruction into a pulse width modulation signal, and applying the pulse width modulation signal to a corresponding phase modulator to generate a reconstructed physical topology of the optical neural network.
6. The method according to any one of claims 1-5, wherein S5 comprises: S51, based on the physical topological row mode coupling state monitoring of the optical neural network, a mode coupling matrix of the next period is obtained, a preset rewarding function is combined to calculate rewarding values of the mode coupling matrix of the next period and the topological reconstruction motion vector, the crosstalk suppression degree and the energy cost of the reconstruction motion are comprehensively evaluated, and scalar rewarding values are generated; s52, performing experience element group construction processing on the environment feature vector, the topology reconstruction motion vector and the scalar rewards value to form a complete state-motion-rewards data pair, and generating a new training sample; and S53, carrying out strategy gradient update processing on the historical accumulated training samples and the new training samples, randomly sampling a batch of data from the experience playback buffer area, calculating strategy gradients to update network weights, and generating an updated deep neural network.
7. The method of claim 6, wherein the predetermined expression of the bonus function is: Wherein, the For a scalar prize value, For the mode coupling matrix of the next cycle, Is a matrix of units which is a matrix of units, Is the Frobenius norm of the matrix, The motion vector is reconstructed for the topology and, Is the L1 norm of the vector which, For the mode coupling matrix in the current environment, 、、 Weight coefficients for controlling crosstalk penalties, action cost penalties, and performance improvement rewards.
8. An optical neural network topology adaptive mode division multiplexing communication system, the communication system comprising: The mode coupling monitoring module is used for carrying out mode coupling state monitoring on the light field signal at the output end of the optical neural network, analyzing the mode power and the phase information in the signal to calculate the coupling relation between modes and generate a mode coupling matrix in the current environment; The environment feature extraction module is used for extracting environment features of the mode coupling matrix, extracting non-diagonal elements of the mode coupling matrix to form crosstalk features, calculating matrix overall variances, splicing crosstalk feature vectors and variance values into comprehensive feature representations, and generating environment feature vectors; The reinforcement learning reasoning module is used for performing reinforcement learning strategy reasoning on the environment feature vector based on a pre-trained deep neural network, performing nonlinear transformation on the environment feature vector through a plurality of full-connection layers and an activation function of the deep neural network, mapping a final layer output value to a physical parameter range of an adjustable optical sub-device, and generating a topology reconstruction action vector; the topology reconstruction driving module is used for carrying out photoelectric signal conversion and driving on the topology reconstruction motion vector, converting digital control parameters into analog voltage signals and driving the phase modulator to generate a reconstructed physical topology of the optical neural network; A strategy gradient updating module, configured to monitor the physical topology line mode coupling state of the optical neural network to obtain a mode coupling matrix of a next period, input the mode coupling matrix of the next period and the topology reconstruction motion vector into a preset reward function to calculate to obtain a scalar reward value, and carrying out strategy gradient update on the deep neural network by combining the environment feature vector and the topology reconstruction motion vector, and generating an updated deep neural network.

Description

Optical neural network topology self-adaptive mode division multiplexing communication system and training method Technical Field The invention relates to the technical field of optical machine learning and intelligent optical communication, in particular to an optical neural network topology self-adaptive mode division multiplexing communication system and a training method. Background The mode division multiplexing technology is an important implementation manner of space division multiplexing, and the mode division multiplexing technology is used for transmitting data in parallel by taking different mutually orthogonal spatial modes in optical fibers as independent channels, so that the transmission capacity of an optical communication system is remarkably improved. In recent years, the technology is also introduced into the design of an optical neural network, and parallel optical computation is performed by using different modes, so as to construct a high-speed and high-throughput photon computing architecture. However, in practical deployment, the performance of the optical neural network is significantly affected by dynamic changes of the external environment. Most of the existing optical neural network training methods optimize network parameters based on a preset and fixed physical topological structure. The method can obtain good effect in a static ideal laboratory environment, but in a real scene, environmental factors such as temperature fluctuation, mechanical vibration and the like can cause waveguide characteristic change, so that unpredictable dynamic drift of coupling relations among modes occurs. The dynamic mode crosstalk causes the trained static model to fail rapidly, the calculation error is increased, and the system performance is obviously reduced. Disclosure of Invention In view of the above, an object of the present invention is to provide an optical neural network topology-adaptive mode division multiplexing communication system and training method that can autonomously maintain optimal performance of an optical neural network in a dynamic environment. The invention adopts the following scheme: in a first aspect, the present invention provides a method for training topology-adaptive mode division multiplexing of an optical neural network, including the following steps: S1, performing mode coupling state monitoring on a light field signal at an output end of an optical neural network, analyzing mode power and phase information in the signal to calculate a coupling relation between modes, and generating a mode coupling matrix in a current environment; S2, extracting environmental features of the mode coupling matrix, extracting non-diagonal elements of the mode coupling matrix to form crosstalk features, calculating matrix overall variances, splicing crosstalk feature vectors and variance values into comprehensive feature representations, and generating environmental feature vectors; S3, performing reinforcement learning strategy reasoning on the environment feature vector based on the pre-trained deep neural network, performing nonlinear transformation on the environment feature vector through a plurality of full-connection layers and an activation function of the deep neural network, mapping a final layer output value to a physical parameter range of the adjustable optical sub-device, and generating a topology reconstruction action vector; S4, performing photoelectric signal conversion and driving on the topology reconstruction motion vector, converting digital control parameters into analog voltage signals, driving a phase modulator, and generating a reconstructed physical topology of the optical neural network; S5, based on optical neural network physical topology line mode coupling state monitoring to obtain a mode coupling matrix of the next period, inputting the mode coupling matrix of the next period and the topology reconstruction motion vector into a preset rewarding function to calculate to obtain a scalar rewarding value, and carrying out strategy gradient updating on the deep neural network by combining the environment feature vector and the topology reconstruction motion vector to generate an updated deep neural network. In one embodiment, the S1 of the adaptive mode division multiplexing training method for optical neural network topology provided by the present invention specifically includes the following steps: S11, performing photoelectric conversion and digital sampling processing on an optical field signal at the output end of the obtained optical neural network, converting the optical signal into an electric signal, performing high-speed analog-to-digital conversion, and generating a parallel digital signal sequence; S12, carrying out modal amplitude and phase estimation processing on the digital signal sequence, and solving complex field distribution of each mode through a digital signal processor to generate complex amplitude information of each optical mode;