Search

CN-122021249-A - Optical bus extinction ratio real-time optimization system and method based on deep learning

CN122021249ACN 122021249 ACN122021249 ACN 122021249ACN-122021249-A

Abstract

The invention discloses a deep learning-based real-time optimization system and method for extinction ratio of an optical bus. The method comprises the steps of 1, making a data set required by training a deep learning model, 2, constructing a target key point detection double-channel LSTM model based on deep learning and training, and 3, performing model test on the trained double-channel LSTM model. The invention solves the problem of extinction ratio instability caused by temperature drift and device aging.

Inventors

  • WANG RUI
  • WANG TAO
  • WAN TIANCAI
  • CHEN YINCHAO
  • DENG JUNWEN

Assignees

  • 中国航空工业集团公司成都飞机设计研究所

Dates

Publication Date
20260512
Application Date
20251224

Claims (10)

  1. 1. A real-time optimization method for extinction ratio of optical bus based on deep learning is characterized by real-time acquisition of multidimensional data including at least two of drive current, modulation current, ambient temperature, output optical power and error rate, input of a pre-trained long-and-short-period memory neural network model to generate optimal bias current and modulation current corresponding to target extinction ratio, dynamic adjustment of working state of laser by a drive circuit, suppression of current switching overshoot by means of RC filter network, construction of a closed loop feedback mechanism, and triggering on-line retraining of the model when actual extinction ratio deviation continuously exceeds a set threshold value to enable extinction ratio to be stabilized within a control range.
  2. 2. The method for optimizing the extinction ratio of the optical bus in real time based on deep learning as claimed in claim 1, wherein the method comprises the following steps: step 1, making a data set required by training a deep learning model; Step 2, constructing a target key point detection double-channel LSTM model based on deep learning and training; and 3, performing model test on the trained double-channel LSTM model.
  3. 3. The method for optimizing the extinction ratio of the optical bus in real time based on deep learning as claimed in claim 2, wherein the step 1 is specifically as follows: step 1.1, sampling based on a section 0.8-1.2 times of the threshold current of the laser; Step 1.2, synchronously recording the second derivative of the output light power in the temperature cycle of-40 ℃ to 85 DEG C And bit error rate data, constructing an original data set containing nonlinear distortion characteristics; And 1.3, injecting a temperature step disturbance of +/-5 ℃ and a current Gaussian noise of +/-10% into the original data set to generate a training set.
  4. 4. The method for optimizing the extinction ratio of the optical bus in real time based on deep learning as claimed in claim 2, wherein the step 2 is specifically as follows: Step 2.1, designing a double-channel LSTM network architecture, wherein a main channel is 128 neurons, processing time sequence current and temperature data, an auxiliary channel is 64 neurons, analyzing the correlation characteristics of the optical power change rate and the error rate of a receiving end, and fusing double-channel output through an attention mechanism; Step 2.2, adopting a double-objective loss function in a training stage of the long-term and short-term memory prediction model, and converging by using AdamW optimizers; And 2.3, training a model and regularizing the model.
  5. 5. The method for optimizing the extinction ratio of the optical bus in real time based on deep learning as claimed in claim 4, wherein the step 2.1 is specifically as follows: Step 2.1.1 designing a two-channel LSTM network architecture, wherein the input dimension of the main channel is 4, the hidden layer neuron is 128 for processing the driving current I bias , the modulating current I mod , the temperature T and the bit error rate BER, the input dimension of the auxiliary channel is 2, the hidden layer neuron is 64 for processing the output power P out and the second derivative of the optical power ; Step 2.1.2, attention weight distribution: Where Q is the final state matrix of the main channel, K, V are both the final state matrices of the auxiliary channels, K T is the transpose of the matrix K, and d k is the scaling factor.
  6. 6. The method for optimizing the extinction ratio of an optical bus in real time based on deep learning as claimed in claim 5, wherein d k =64.
  7. 7. The method for optimizing the extinction ratio of an optical bus in real time based on deep learning as claimed in claim 4, wherein in step 2.2, the double objective loss function L is as follows: L=α ·MAE( I pred ,I actual ) + β·ReLU (△BER-10 -6 ) The weight coefficient alpha:beta=3:1, I pred represents a predicted current value, I actual represents an actual current value, delta BER is the error rate fluctuation of a receiving end, MAE is the average absolute value error of a predicted value I pred and a true value I actual , and the expression of a ReLU function is ReLU (x) =max (0, x).
  8. 8. The method for optimizing the extinction ratio of the optical bus in real time based on deep learning as claimed in claim 2, wherein the step 3 is specifically as follows: Step 3.1, performing off-line test by using a training set, and observing whether the predicted bias current output by the model and the extinction ratio corresponding to the predicted bias current are in a required range; Step 3.2, performing on-calibration, namely comparing the actual extinction ratio measured by the built-in photodiode with a target value in real time, and triggering calibration when the temperature change rate is 2 ℃ per ms or the driving current fluctuation is 5% of the driving current I bias under the condition that continuous 10 sampling points meet the condition of I EX actual – EX target I & gt 1dB, wherein EX actual is the actual extinction ratio, and EX target is the target extinction ratio; and 3.3, performing incremental learning.
  9. 9. The method for optimizing the extinction ratio of the optical bus in real time based on deep learning as claimed in claim 8, wherein the step 3.3 is specifically as follows: step 3.3.1, storing 200ms time sequence data before triggering calibration and constructing a training set; Step 3.3.2, updating parameters, namely updating only the weights of an adjustable layer, setting the frozen layer as all weights of an LSTM main channel/auxiliary channel, setting the adjustable layer as a 192 multiplied by 2 matrix, and setting an optimizer as a random gradient of a driving quantity to be reduced; Step 3.3.3 design loss parameter L cabil is as follows: Where I pred denotes a predicted current value, I actual denotes an actual current value, EX err denotes an absolute value of a difference between an actual extinction ratio and a target extinction ratio, EX actual denotes an actual extinction ratio, and EX target denotes a target extinction ratio.
  10. 10. The real-time optimization system for the extinction ratio of the optical bus based on deep learning is characterized by being used for completing the real-time optimization method for the extinction ratio of the optical bus according to any one of claims 1-9, and comprising the following steps: The data acquisition module is used for acquiring multidimensional data; The neural network training module is used for training based on the input multidimensional data to generate an optimal bias current and a modulation current corresponding to the target extinction ratio; The driving circuit adjusts the filter module, dynamically adjusts the working state of the laser based on the optimal bias current and the modulation current, and suppresses current switching overshoot through filtering; and the closed loop feedback module is used for triggering the neural network training module to retrain on line when the actual extinction ratio deviation continuously exceeds a set threshold value, so that the extinction ratio is stabilized in a control range.

Description

Optical bus extinction ratio real-time optimization system and method based on deep learning Technical Field The invention belongs to the technical field of artificial intelligence deep learning, and particularly relates to a real-time optimization system and method for an extinction ratio of an optical bus based on deep learning. Background The deep learning is an artificial intelligent machine learning method based on a neural network, can automatically extract a large amount of data through training a neural network model, and is particularly suitable for solving the problems of real-time modeling and prediction of a complex system. In the field of optical communication, the extinction ratio (Extinction Ratio, ER) is the ratio of logic high-level optical power to low-level optical power, and is a core index for measuring the optical modulation quality of a transmitting end, and directly affects the signal-to-noise ratio and the receiving sensitivity of a link. However, the current optical bus transmitting end generally adopts a fixed table look-up method to control extinction ratio, and the fixed compensation mechanism is difficult to adjust in real time due to nonlinear bending of the laser P-I curve at high temperature or when the device is aged, so that the fluctuation range of the extinction ratio exceeds +/-2 dB, meanwhile, the traditional scheme relies on error rate feedback of the receiving end to carry out hysteresis adjustment, the millisecond response delay is above 25Gbps, the error rate is greatly increased in the high-speed transmission process, and the reliability of the system is reduced. Disclosure of Invention The invention aims to provide a deep learning-based real-time optimization system and method for the extinction ratio of an optical bus. The invention solves the problem of extinction ratio instability caused by temperature drift and device aging. The technical scheme of the invention is that the real-time optimization method of the extinction ratio of the optical bus based on deep learning is characterized by adopting multidimensional data comprising at least two of driving current, modulation current, ambient temperature, output optical power and error rate in real time, inputting a pretrained long-short-period memory neural network model to generate optimal bias current and modulation current corresponding to a target extinction ratio, dynamically adjusting the working state of a laser through a driving circuit, inhibiting current switching overshoot by means of an RC filter network, constructing a closed loop feedback mechanism, and triggering the model to retrain on line when the deviation of the actual extinction ratio continuously exceeds a set threshold value, so that the extinction ratio is stabilized in a control range. The method for optimizing the extinction ratio of the optical bus in real time based on deep learning comprises the following steps: step 1, making a data set required by training a deep learning model; Step 2, constructing a target key point detection double-channel LSTM model based on deep learning and training; and 3, performing model test on the trained double-channel LSTM model. In the optical bus extinction ratio real-time optimization method based on deep learning, the step 1 specifically comprises the following steps: step 1.1, sampling based on a section 0.8-1.2 times of the threshold current of the laser; Step 1.2, synchronously recording the second derivative of the output light power in the temperature cycle of-40 ℃ to 85 DEG C And bit error rate data, constructing an original data set containing nonlinear distortion characteristics; And 1.3, injecting a temperature step disturbance of +/-5 ℃ and a current Gaussian noise of +/-10% into the original data set to generate a training set. In the optical bus extinction ratio real-time optimization method based on deep learning, the step 2 specifically comprises the following steps: Step 2.1, designing a double-channel LSTM network architecture, wherein a main channel is 128 neurons, processing time sequence current and temperature data, an auxiliary channel is 64 neurons, analyzing the correlation characteristics of the optical power change rate and the error rate of a receiving end, and fusing double-channel output through an attention mechanism; Step 2.2, adopting a double-objective loss function in a training stage of the long-term and short-term memory prediction model, and converging by using AdamW optimizers; And 2.3, training a model and regularizing the model. In the foregoing optical bus extinction ratio real-time optimization method based on deep learning, step 2.1 specifically includes the following steps: Step 2.1.1 designing a two-channel LSTM network architecture, wherein the input dimension of the main channel is 4, the hidden layer neuron is 128 for processing the driving current I bias, the modulating current I mod, the temperature T and the bit error rate BER, the input dimension o