CN107342810A - Deep learning Brilliant Eyes figure analysis method based on convolutional neural networks - Google Patents

Abstract

The invention discloses a kind of deep learning Brilliant Eyes figure analysis method based on convolutional neural networks, it is related to technical field of photo communication, wherein carrying out performance evaluation to eye pattern by building simultaneously training convolutional neural networks module, comprises the following steps：Obtain eye pattern training dataset；Eye pattern is pre-processed；CNN modules are trained to carry out feature extraction；The eye pattern of required analysis is inputted to the CNN modules that training is completed after pretreatment and carries out pattern-recognition and performance evaluation；Export analysis result.Depth learning technology based on convolutional neural networks is applied in eye Diagram Analysis by the present invention, solve the problems, such as directly handle initial data in traditional eye pattern performance evaluation, manual intervention need to be carried out, the intellectuality and automation of eye pattern original image information analysis are realized using CNN, can be as the eye pattern software processing module of oscillograph and the eye Diagram Analysis module of simulation software, and then be embedded into tester and carry out intelligent signal analysis and performance monitoring.

Description

Deep Learning Intelligent Eye Diagram Analysis Method Based on Convolutional Neural Network

technical field

The invention relates to the technical field of optical communication, in particular to a deep learning intelligent eye diagram analysis method based on a convolutional neural network.

Background technique

Machine learning (ML) techniques provide powerful tools to solve problems in many fields such as natural language processing, data mining, speech recognition, and image recognition. At the same time, machine learning technology has also been widely used in the field of optical communication, which has greatly promoted the development of intelligent systems. Current research mainly focuses on the use of different machine learning algorithms for optical performance monitoring (OPM) and nonlinear damage compensation. The machine learning algorithms used include expected maximum (EM), random forest, backpropagation artificial neural network (BP -ANN), K-Nearest Neighbors (KNN) and Support Vector Machines (SVM), etc. However, all of the above machine learning algorithms have their own limitations in the capability of feature extraction. More specifically, machine learning models cannot directly process natural data in its raw form, so considerable domain expertise and engineering skills have to be designed before applying algorithms to convert raw data into suitable internal representations or feature vectors. , so that the subsystem can detect the pattern of the input data. Therefore, it is hoped that more advanced machine learning algorithms can be developed that not only directly process the raw data, but also automatically detect the desired features.

Recently, deep learning has become a hot research topic with the aim of bringing machine learning closer to the goal of artificial intelligence (AI). Deep learning can be understood as a deep neural network with multiple non-linear layers that learns features from data through a self-learning process rather than manual design by human engineers. One of the most famous breakthroughs in deep learning is Google DeepMind's computer program "AlphaGo", which for the first time defeated a professional player at a board game by self-learning. In addition, as a current research hotspot, deep learning has made significant progress in various application areas such as unmanned aerial vehicles, medical diagnosis, and sentiment analysis. However, to the best of our knowledge, there is little research work based on deep learning in the field of optical communication systems.

At the same time, in the field of optical communication, the current modulation format recognition and OSNR, CD, linear impairment, nonlinear impairment and other performance index estimation techniques cannot directly process the original data, but must artificially extract the corresponding features, requiring a large number of human intervention. Therefore, it is hoped that the eye diagram can be used to use more advanced technology to carry out intelligent analysis of various performances, without manual intervention, to achieve accurate measurement, without the need for instant processing of data statistics, and to realize the intelligence and automation of performance analysis using the eye diagram.

Contents of the invention

The purpose of the present invention is to apply deep learning technology to the field of optical communication, provide an intelligent and reliable deep learning intelligent eye diagram analysis method based on convolutional neural network, and solve the problem that the original image data cannot be directly processed in the traditional eye diagram performance analysis. The shortcomings of manual intervention are needed, and the intelligence and automation of performance analysis on the original image of the eye diagram are realized.

In order to achieve the above object, the present invention discloses a deep learning intelligent eye diagram analysis method based on convolutional neural network, which applies deep learning technology based on convolutional neural network to eye diagram analysis, and utilizes convolutional neural network to analyze the eye diagram. Carry out multiple performance analysis, described method comprises the following steps: step one, obtain the eye pattern training data set of required analysis; Step two, eye pattern image preprocessing; Step three, training convolutional neural network (CNN) module to eye Extract features from the graph; Step 4, input the required analysis eye diagram into the trained CNN module for pattern recognition and performance analysis; Step 5, output the analysis results.

Preferably, the multiple properties to be analyzed in the eye diagram are modulation format, optical signal-to-noise ratio (OSNR), dispersion (CD), linear impairment and nonlinear impairment.

Preferably, in the first step of obtaining the eye diagram training set, the training data sets under various performance indicators of the eye diagram are collected, wherein each set of data in the training data set is input as an eye diagram image and output as a specific performance The specific indicator information pair constitutes.

Preferably, in the eye diagram preprocessing step 2, the color eye diagram image in the training data set acquired in the step 1 is converted into a grayscale image, and the obtained eye diagram grayscale image is subjected to downsampling processing.

Preferably, in the feature extraction step 3 of the training CNN module, the preprocessed eye diagram in the step 2 is input into the constructed CNN module, and after the training process is performed based on the training data, the CNN module automatically Extract features from eye diagram images and construct relationships between features and different properties.

Preferably, in step 4 of the pattern recognition and performance analysis of the CNN module, the preprocessed eye diagram to be analyzed is input into the trained CNN module, and the CNN module performs pattern recognition on the input eye diagram, and through its Previous learning experiences perform performance analysis on the current input eye diagram.

Preferably, in the fifth step of outputting analysis results, the information output by the CNN module includes various performances to be analyzed, and analysis results of different performances can be obtained from the output information.

Preferably, the structure of the CNN module mainly includes: an input layer, n convolutional layers (C1, C2, ..., Cn), n pooling layers (P1, P2, ..., Pn), m fully connected Layers (F1, F2, ..., Fm), an output layer, wherein the input of the input layer is a preprocessed eye pattern image, and the input layer is connected to the convolutional layer C1; the convolutional layer C1 contains k1 A convolution kernel with a size of a1×a1, the input layer image obtains k1 feature maps through the convolution layer C1, and then transfers the obtained feature maps to the pooling layer P1; the pooling layer P1 uses b1×b1 Pooling the feature maps generated by the convolutional layer C1 to obtain corresponding k1 sampled feature maps, and then transfer the obtained feature maps to the next convolutional layer C2; the n convolutional layers Layers and pooling layer pairs are sequentially connected to continuously extract deep-level sampling features of the image. The last pooling layer Pn is connected to the fully connected layer F1. Among them, the convolutional layer Ci contains ki layers of size ai×ai The convolution kernel, the sampling size of the pooling layer Pj is bj×bj, Ci represents the i-th convolutional layer, and Pj represents the j-th pooling layer; the fully connected layer F1 is obtained by the last pooling layer Pn Each pixel represents a neuron node of the fully connected layer F1, and all neuron nodes of the F1 layer are connected with the neurons of the next fully connected layer F2 The nodes are fully connected; connected sequentially through m fully connected layers, and the last fully connected layer Fm is fully connected with the output layer; the output layer outputs node information of different performances of the eye diagram to be analyzed.

Preferably, the node information output by the output layer is an L-bit binary bit sequence, wherein the N different performances are respectively represented by L1, L2, ..., LN bit binary bit information, and the Li bit is used to represent the first Li kinds of different index information of i properties, where L=L1+L2+...+LN.

Preferably, the CNN-based eye diagram processing algorithm will be used as the eye diagram software processing module of the oscilloscope or the eye diagram analysis module of the simulation software, and then embedded in the test instrument for intelligent signal analysis and performance monitoring.

The beneficial effect of the present invention is that: the present invention solves the disadvantages of traditional eye diagram analysis, applies the deep learning technology based on convolutional neural network to eye diagram analysis, uses convolutional neural network to perform various performance analysis on eye diagrams, and applies The present invention can directly process the original image data of the eye diagram without manual intervention for feature extraction, realize the intelligence and automation of eye diagram performance analysis, and then can be used as the eye diagram software processing module of the oscilloscope or the eye diagram analysis of the simulation software Module, embedded into the test instrument for intelligent signal analysis and performance monitoring.

Description of drawings

Fig. 1 shows the flowchart of the deep learning intelligent eye pattern analysis method based on the convolutional neural network of the present invention;

Fig. 2 shows a schematic structural diagram of a deep learning intelligent eye diagram analysis based on a convolutional neural network according to an embodiment of the present invention;

Fig. 3 shows different modulation formats and partial eye diagram images of different OSNRs collected by an embodiment of the present invention;

FIG. 4 shows a schematic diagram of the accuracy of OSNR estimated under different modulation formats according to an embodiment of the present invention;

FIG. 5 shows a schematic diagram of a comparison of accuracy of eye diagram performance analysis between CNN and other machine learning algorithms under different modulation formats according to an embodiment of the present invention.

detailed description

The specific implementation manners of the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. The following examples are used to illustrate the present invention, but are not intended to limit the protection scope of the present invention.

As shown in Figure 1, the deep learning intelligent eye diagram analysis method based on the convolutional neural network proposed by the present invention applies the deep learning technology based on the convolutional neural network to the eye diagram analysis, and uses the convolutional neural network to analyze the eye diagram. A variety of performance analysis, including the following steps: Step 1, obtain the eye diagram training data set required for analysis; Step 2, eye diagram image preprocessing; Step 3, train the convolutional neural network (CNN) module to extract features from the eye diagram ; Step 4, the required analysis eye pattern is input to the trained CNN module for pattern recognition and performance analysis; Step 5, the output analysis result.

In this embodiment, the eye diagram performance to be analyzed is modulation format and OSNR.

In the first step of obtaining the eye diagram training data set, a basic simulation system is established based on VPI Transmission Maker 9.0, and optical signals of four different modulation formats are generated by pseudo-random binary sequences, which are: 4PAM, RZ-DPSK, NRZ- OOK, RZ-OOK. The four modulation formats are all based on direct detection, and the transmitted information is reflected in the amplitude of the signal, which is suitable for subsequent eye diagram analysis. An erbium-doped fiber amplifier (EDFA) was used in the simulated system to add amplified spontaneous emission (ASE) noise to the optical signal, and a variable optical attenuator (VOA) was used to adjust the OSNR from 10 to 25 dB in 1 dB steps . In order to simulate the real optical signal as much as possible, a chromatic dispersion (CD) simulator is added to the system, so that the eye diagram generated by the simulation can better reflect the real situation. For the optical signals of four different modulation formats in this embodiment, the 4PAM, NRZ-OOK and RZ-OOK signals are directly detected by the photodetector (PD), while the RZ-DPSK signal is detected by the combined delay interferometer (DI). A balanced photodetector (BPD) is used for detection. After synchronous sampling, a digital signal containing four kinds of signal strength information is obtained. In order to obtain a more realistic visual effect, this embodiment uses a special eye pattern generation module in the oscilloscope to convert the received digital signal into a corresponding eye pattern image.

Based on the simulation system, this embodiment stipulates that each modulation format generates 16 different OSNR values ( ), collect 100 eye diagram images in "jpg" format with a pixel size of 900×1200 for each OSNR value of each modulation format, here, each OSNR value of each modulation format and its corresponding The eye pattern images of are used as a set of training data, so the entire training data set includes a total of 6400 (1600×4) sets of training data.

In the second step of the eye image preprocessing step, in order to reduce the amount of calculation and enhance the generalization ability, the eye image image collected in the step 1 is converted into a gray image after the gray scale conversion, and the following steps are performed: Sampling reduces the pixel size of the original eye diagram to 28×28, and finally the processed training dataset is input into the established CNN module. As shown in Figure 3, different eye diagrams can exhibit different modulation formats, and if the observed eye diagram is carefully analyzed visually, it can also be seen that the eye diagram has a first-order approximate relationship with the OSNR value.

The training CNN module is carried out in the feature extraction step 3, wherein the eye pattern training data set of the CNN module is input, and each eye pattern image is in one-to-one correspondence with a label vector composed of 20 bits, and the first 4 bits of the label vector Represents different modulation formats (4PAM: 0001, RZ-DPSK: 0010, NRZ-OOK: 0100, RZ-OOK: 1000), the last 16 bits represent different OSNR values (10dB: 0000000000000001, 11dB: 0000000000000010, ..., 25dB: 1000000000000000). During the training process, the CNN module gradually extracts effective features of the input eye image. At the same time, in order to minimize the error between the ideal label vector and the actual output label vector, the CNN module uses the method of gradient descent through backpropagation to gradually adjust the parameters of its kernel.

Fig. 2 represents the structural representation of the intelligent eye diagram analysis based on the convolutional neural network of a specific embodiment of the present invention, the structure of the CNN module mainly includes the following parts: an input layer, two convolutional layers (C1, C2) , two pooling layers (P1, P2), a fully connected layer (F1), and an output layer. The preprocessed 28×28 eye pattern image is input into the CNN module as the input layer and connected to the convolutional layer C1; the input eye pattern image passes through the convolutional layer C1 containing 6 convolution kernels with a size of 5×5 to obtain 6 feature maps with a size of 24×24, and then transfer the obtained feature maps to the pooling layer P1; the pooling layer P1 performs maximum pooling on the 6 feature maps with a sampling size of 2×2, and obtains the corresponding 6 size It is a 12×12 sampled feature map, and then the obtained feature map is sent to the convolutional layer C2; the convolutional layer C2 contains 12 convolution kernels with a size of 5×5, and the 6 features obtained by the pooling layer P1 The image is passed through the convolutional layer C2 to obtain 12 feature maps with a size of 8×8, and then the obtained feature maps are sent to the pooling layer P2; the pooling layer P2 also uses a sampling size of 2×2 to generate 12 feature maps with a size of 4×4 are maximally pooled to obtain the corresponding 12 sampled feature maps, and then the obtained feature maps are sent to the fully connected layer F1; the pixels of all feature maps obtained by the pooling layer P2 Mapped to a one-dimensional fully connected layer F1, each pixel represents a neuron node of the fully connected layer F1, each neuron node of the fully connected layer F1 is fully connected to the output layer; the final output layer outputs the eye Node information for graph properties.

Among them, the convolutional layer is the core component of the CNN module. The parameters in this layer consist of a set of convolutional kernels that have a small local receptive field but extend to the full depth of the eye image. During forward propagation, each kernel is convolved with pixels across the width and height of the eye image, outputting a two-dimensional plane called the feature map generated from that kernel. Unlike classical convolution in mathematics, operations in CNNs are discrete convolutions, which can be viewed as matrix multiplications. The convolution kernel can be regarded as a feature detector. Through the convolution kernel, the CNN module can learn its unique features from the input image. At the same time, in order to build a more effective model, multiple convolution kernels are generally required to detect Multiple features to produce multiple feature maps in convolutional layers. After the feature extraction of the convolutional layer, the pooling layer will combine semantically similar features into a corresponding one. The typical pooling method is to calculate the maximum value of the local unit block in a feature map and perform sub-sampling of the feature map. . In this embodiment, each sub-sampling unit obtains input from a 2×2 unit area in the convolutional feature map, and uses the maximum value of these inputs as a pooled value to form a pooled feature map.

In step 4 of the CNN module pattern recognition and performance analysis, 4 different modulation formats that need to be analyzed after preprocessing, each modulation format has a range of The eye pattern image with the OSNR value (with 1dB as the step size) is input into the CNN module that has been trained above. As shown in the figure, the performance analysis of the modulation format and OSNR is performed, and the analysis results are output in the form of a 20-bit bit vector.

In the fifth step of outputting the analysis result, the first 4 bits of the 20-bit bit vector output by the CNN module can obtain the modulation format information of the analyzed eye diagram, and the last 16 bits can obtain the corresponding OSNR value.

In order to show the accuracy of the analysis of the method proposed in the present invention, Fig. 4 shows the estimation accuracy of the CNN module for OSNR under different modulation formats and different iterations. Clearly, the accuracies of the four modulation formats all increase with the number of iterations of the CNN module. CNN modules trained with different iterations have different performance recognition capabilities. In this embodiment, when the number of iterations exceeds 31, the CNN module under the four modulation formats has an accuracy of estimating the corresponding OSNR to 100%, that is, the analyzed performance has no wrong result.

Meanwhile, to demonstrate the advantages of the present invention, CNN is compared with four other well-known machine learning algorithms, namely decision tree, KNN, BP-ANN and SVM. The estimation accuracy of each algorithm for OSNR under different modulation formats is shown in Figure 5 in the form of a histogram. It can be seen from the figure that CNN has obvious advantages over the other four algorithms. Among them, the decision tree algorithm has fast processing speed and small memory requirements, and these advantages also lead to low estimation accuracy; KNN algorithm usually has good estimation accuracy in low dimensions, but may produce large deviation; the SVM algorithm has great advantages in estimation accuracy and memory usage, it only needs a few support vectors, but it is essentially just a binary classifier, so it needs multiple OSNR values Although the BP neural network is also developed from the neural network, it lacks the ability to extract features, requires a large amount of training data to achieve better results, and is prone to fall into local minimum and overfitting. Compared with the above algorithms, CNN is less sensitive to the variance of input data, and the constructed network is more powerful, which can avoid over-fitting to a large extent, and can automatically extract the characteristics of input data, especially in image processing. At the same time, due to the advantages of local receptive fields, weight distribution, sub-sampling, etc., CNN can achieve the best accuracy at a moderate computational cost.

In summary, the method proposed by the present invention applies deep learning technology based on convolutional neural network to eye diagram analysis, which can be effectively used as the eye diagram software processing module of the oscilloscope or the eye diagram analysis module of the simulation software, and then embedded in the test Intelligent signal analysis and performance monitoring are carried out in the instrument to realize the automation and intelligence of eye diagram analysis.

The above embodiments are only used to illustrate the present invention, and do not limit the protection scope of the present invention. For those skilled in the relevant fields, the present invention may have various modifications and changes. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims (10)

1. A deep learning intelligent eye diagram analysis method based on a convolutional neural network, characterized in that, the deep learning technology based on a convolutional neural network is applied to the eye diagram analysis, and the convolutional neural network is used to perform multiple eye diagram analysis. Performance analysis, the method includes the following steps: Step 1. Obtain the eye diagram training data set to be analyzed; Step 2, eye pattern image preprocessing; Step 3, training the convolutional neural network (CNN) module to extract features from the eye diagram; Step 4. Input the required analysis eye pattern into the trained CNN module for pattern recognition and performance analysis; Step five, output the analysis result. 2. the deep learning intelligent eye diagram analysis method based on convolutional neural network according to claim 1, is characterized in that, the multiple performances that need to be analyzed in the eye diagram are modulation format, optical signal-to-noise ratio (OSNR) , chromatic dispersion (CD), linear impairment and nonlinear impairment. 3. the deep learning intelligent eye diagram analysis method based on convolutional neural network according to claim 1, is characterized in that, in described eye diagram training set acquisition step one, under the various performance different index situations of collection eye diagram A training data set, wherein each set of data in the training data set is composed of a pair of specific index information whose input is an eye pattern image and whose output is a specific performance. 4. the deep learning intelligence eye pattern analysis method based on convolutional neural network according to claim 1, is characterized in that, in described eye pattern preprocessing step 2, the color in the training data set that acquires in described step 1 The eye pattern image is converted into a grayscale image, and the obtained eye pattern grayscale image is down-sampled. 5. the deep learning intelligent eye diagram analysis method based on convolutional neural network according to claim 1, is characterized in that, described training CNN module carries out feature extraction step 3, with the eye after the pretreatment in described step 2 The image is input into the constructed CNN module, and after the training process is performed based on the training data, the CNN module automatically extracts features from the eye diagram image, and constructs the relationship between the features and different performances. 6. the deep learning intelligence eye pattern analysis method based on convolutional neural network according to claim 1, is characterized in that, in described CNN module pattern recognition and performance analysis step 4, the eye pattern of the required analysis of preprocessing Input the trained CNN module, the CNN module performs pattern recognition on the input eye diagram, and performs performance analysis on the current input eye diagram through its previous learning experience. 7. the deep learning intelligent eye pattern analysis method based on convolutional neural network according to claim 1, is characterized in that, in the described output analysis result step 5, the information output by the CNN module includes each required analysis. different properties, and the analysis results of different properties can be obtained from the output information. 8. the deep learning intelligent eye diagram analysis method based on convolutional neural network according to claim 1, is characterized in that, the structure of described CNN module mainly comprises: an input layer, n convolution layers (C1, C2, ..., Cn), n pooling layers (P1, P2, ..., Pn), m fully connected layers (F1, F2, ..., Fm), an output layer; Wherein, the input of the input layer is a preprocessed eye diagram image, and the input layer is connected to the convolution layer C1; The convolutional layer C1 contains k ₁ convolution kernels with a size of a ₁ ×a ₁ , the input layer image is passed through the convolutional layer C1 to obtain k ₁ feature maps, and then the obtained feature maps are sent to the pooling layer P1; The pooling layer P1 pools the feature maps generated by the convolution layer C1 with a sampling size of b ₁ ×b ₁ to obtain corresponding k ₁ sampled feature maps, and then transmits the obtained feature maps to The next convolutional layer C2; The sequential connection of the n convolutional layers and the pooling layer pairs continuously extracts the deep-level sampling features of the image, and the last pooling layer Pn is connected with the fully connected layer F1, wherein the convolutional layer Ci contains k _i A convolution kernel with a size of a _i × a _i , the sampling size of the pooling layer Pj is b _j × b _j , Ci represents the i-th convolutional layer, and Pj represents the j-th pooling layer; The fully connected layer F1 is a one-dimensional layer formed by mapping pixels of all k _n feature maps obtained by the last pooling layer Pn, and each pixel represents a neuron node of the fully connected layer F1, All neuron nodes of the F1 layer are fully connected with the neuron nodes of the next fully connected layer F2; Connect sequentially through m fully connected layers, and the last fully connected layer Fm is fully connected with the output layer; The output layer outputs node information of different performances of the eye diagram to be analyzed. 9. the deep learning intelligent eye pattern analysis method based on convolutional neural network according to claim 8, is characterized in that, the node information of described output layer output is the binary bit sequence of L position, and wherein, described N different The performance of each is represented by L ₁ , L ₂ , ..., L _N -bit binary bit information, and L _i bits are used to represent L _i different index information of the i-th performance, where L=L ₁ +L ₂ +... +L _N . 10. the deep learning intelligent eye pattern analysis method based on convolutional neural network according to claim 1, is characterized in that, the eye pattern processing algorithm based on CNN will be used as the eye pattern analysis of the eye pattern software processing module of oscilloscope or simulation software module, and then embedded into the test instrument for intelligent signal analysis and performance monitoring.