CN115102674A - Bi-LSTM network-based high-speed link eye pattern prediction method - Google Patents
Bi-LSTM network-based high-speed link eye pattern prediction method Download PDFInfo
- Publication number
- CN115102674A CN115102674A CN202210691524.4A CN202210691524A CN115102674A CN 115102674 A CN115102674 A CN 115102674A CN 202210691524 A CN202210691524 A CN 202210691524A CN 115102674 A CN115102674 A CN 115102674A
- Authority
- CN
- China
- Prior art keywords
- term memory
- data
- batch
- short term
- lstm network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L1/00—Arrangements for detecting or preventing errors in the information received
- H04L1/20—Arrangements for detecting or preventing errors in the information received using signal quality detector
- H04L1/205—Arrangements for detecting or preventing errors in the information received using signal quality detector jitter monitoring
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B17/00—Monitoring; Testing
- H04B17/30—Monitoring; Testing of propagation channels
- H04B17/391—Modelling the propagation channel
- H04B17/3912—Simulation models, e.g. distribution of spectral power density or received signal strength indicator [RSSI] for a given geographic region
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L43/00—Arrangements for monitoring or testing data switching networks
- H04L43/08—Monitoring or testing based on specific metrics, e.g. QoS, energy consumption or environmental parameters
- H04L43/0852—Delays
- H04L43/087—Jitter
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Theoretical Computer Science (AREA)
- Biophysics (AREA)
- Data Mining & Analysis (AREA)
- Electromagnetism (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Environmental & Geological Engineering (AREA)
- Computational Linguistics (AREA)
- Spectroscopy & Molecular Physics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Quality & Reliability (AREA)
- Dc Digital Transmission (AREA)
Abstract
The invention discloses a Bi-LSTM network-based high-speed link eye pattern prediction method, which comprises the following specific steps: constructing a bidirectional long-short term memory Bi-LSTM network, generating a training set containing time step information, training the bidirectional long-short term memory Bi-LSTM network by adopting a state transfer method, and generating an eye pattern by utilizing the trained bidirectional long-short term memory Bi-LSTM network. Because the invention constructs a training set containing time step information, the time correlation information of the discrete time sequence is fully considered, the defect that the time correlation information of the discrete time sequence is ignored in the prior art is overcome, and the eye pattern can be predicted more accurately.
Description
Technical Field
The invention belongs to the technical field of physics, and further relates to a high-speed link prediction eye diagram method based on a Bidirectional Long Short-Term Memory Bi-LSTM (Bidirectional Long Short-Term Memory) network in the technical field of electric digital data processing. The method can be used for predicting the output eye pattern when a high-speed, high-density and miniaturized system (hereinafter referred to as a high-speed link) inputs any length and any code pattern sequence, obtaining the eye pattern under a certain error rate and realizing the performance evaluation of the high-speed link. The method can be applied to the design analysis of signal integrity in the technical field of electrical digital data processing.
Background
At present, the development of electronic systems towards high speed, high density and miniaturization leads to increasingly serious signal integrity problems such as reflection, crosstalk, ground bounce, jitter and inter Symbol interference (isi), which have a serious influence on the output signal waveform of the high speed link. Therefore, considering analysis of the signal waveforms in the product or even superimposing the output signal waveforms as an eye diagram during the design process is crucial for the performance assessment of the high speed link and for solving the signal integrity problem. The current methods for predicting the performance of the high-speed link system mainly include a transient simulation method and a unit impulse response method.
The transient simulation method is a traditional eye pattern method for measuring a certain bit error rate. Firstly, adding an input signal at an input end; then, acquiring a signal at a receiving end of the oscilloscope by using the oscilloscope; and finally, accumulating and superposing signal waveforms in a longer time into an eye pattern through afterglow characteristics of the oscilloscope. To obtain an eye pattern at a certain bit error rate, for example at a bit error rate of 1e-6, i.e. the probability of the inner contour of the eye pattern occurring is 1e-6, it is necessary to simulate a pattern of at least 1e6 bits. The method has the disadvantages that if the eye pattern with low bit error rate is obtained, the number of bits of the code pattern needing to be simulated is very large, and the simulation speed is very slow. For example, if a bit error rate of 1e-12 is desired, then at least 1e12 bit patterns are needed to simulate, the simulation time being on the order of days.
The patent document "an eye pattern estimation method under additive noise interference, device and storage medium" (patent application No. CN202210103888.6, application publication No. CN 114124318a) of shenzhen dinghecheng intellectual property agency limited company discloses a method for predicting a statistical eye pattern by using a unit impulse response method. The method comprises the following implementation steps: firstly, acquiring unit impulse response and outputting sampling data; secondly, processing to obtain a first signal probability density function; thirdly, generating an output noise probability density function; fourthly, processing to obtain a second signal probability density function; and fifthly, generating a statistical eye pattern, wherein the statistical eye pattern contains the profile information of the eye pattern with any bit error rate. However, the method still has three disadvantages: firstly, if the high-speed link contains jitter, because the predicted statistical eye pattern output by the method only considers the unit impulse response and ignores the jitter contained in the high-speed link, the eye height and eye width data of the predicted statistical eye pattern has a relative error of at least 5% with the transient simulation result, and the predicted outline of the statistical eye pattern has a large deviation with the transient simulation result; secondly, if the high-speed link contains nonlinear factors, such as saturation nonlinearity caused by an amplifier, the nonlinear factors not only affect the current code bit, but also affect the subsequent code bits. Because the statistical eye diagram output by the method only considers the unit impulse response and neglects the influence of the saturation nonlinearity on the subsequent code bits, the predicted eye height and eye width data of the statistical eye diagram has a relative error of at least more than 5% with the transient simulation result, and the predicted outline of the statistical eye diagram has a larger deviation with the transient simulation result; thirdly, if the rising edge response and the falling edge response of the high-speed link are not symmetrical, stray glitches can occur between consecutive bits according to the superposition of a linear system. Because the predicted statistical eye diagram output by the method only considers the unit impulse response, the edge response of the unit impulse response is completely symmetrical, and the condition that the rising edge response and the falling edge response are asymmetrical is ignored, the relative error of the eye height and eye width data of the predicted statistical eye diagram and the transient simulation result is more than 5%, and the outline of the predicted statistical eye diagram has larger deviation with the transient simulation result.
Disclosure of Invention
The invention aims to provide a Bi-LSTM network-based high-speed link eye pattern prediction method for solving the problems that the transient simulation technology in the prior art needs too long number of simulated code patterns for obtaining a low bit error rate eye pattern, so that the simulation time is too long, and the unit impulse response method in the prior art only considers the unit impulse response of the current code bit, ignores the influence of jitter and nonlinear factors in a high-speed link and asymmetry of rising edge response and falling edge response on the subsequent code bit, and further causes inaccurate eye pattern prediction.
The idea for achieving the purpose of the invention is that the simulation time of the invention is shorter, and the data set required by training is less, so the training time can be controlled within 5 minutes, and the time for finally obtaining the eye pattern under low code rate is shorter. For example, the eye pattern with the error rate of 1e-7 is obtained, the total time required by the method is 32.55 minutes, the time comprises the time for obtaining a training data set through simulation, the training data set preprocessing time, the training time of a bidirectional long-short term memory Bi-LSTM network and the time required for predicting the eye pattern, and the problems that the simulation time is 180.7 minutes and the required time is too long for obtaining the eye pattern with the error rate of 1e-7 by the transient simulation technology in the prior art are solved. In addition, each group of data in the training data set generated by the invention is composed of a series of Pseudo-Random Binary Sequence PRBS (Pseudo-Random Binary Sequence) and response waveforms output after the Pseudo-Random Binary Sequence PRBS is input into a high-speed link, so that the training data set not only considers the current code bit, but also considers the subsequent code bits, and therefore, the problem of inaccurate prediction of the eye pattern caused by the fact that the jitter and nonlinear factors in the high-speed link and the influence of asymmetry of rising edge response and falling edge response on the subsequent code bits are ignored in the unit impulse response simulation method in the prior art is solved. Experiments show that relative errors between the eye height and the eye width of the eye pattern of the high-speed link predicted by the method and the eye height and the eye width of the eye pattern obtained by transient simulation are 0% and 1.27% respectively; and the relative errors between the eye height and the eye width of the eye pattern predicted by the simulation method of the unit impulse response in the prior art and the eye height and the eye width of the eye pattern obtained by actual transient simulation are respectively 11.81 percent and 11.37 percent. The method has the advantage of high accuracy.
The invention has the following implementation steps:
step 1, constructing a bidirectional long-short term memory Bi-LSTM network:
step 1.1, constructing a bidirectional long-short term memory Bi-LSTM network formed by connecting a first bidirectional long-short term memory network layer, a second bidirectional long-short term memory network layer and a full connection layer in series;
step 1.2, the number of hidden units of the first bidirectional long-short term memory network layer and the second bidirectional long-short term memory network layer is set to be 16, and an activation function used in each hidden unit is realized by adopting a hyperbolic tangent function; sequentially setting the dimension of the input data of the first and second bidirectional long-short term memory network layers to 10 × 10 × L 1 10 × 10 × 32, the dimensions of the output data are all set to 10 × 10 × 32; the dimension of input data of a full connection layer is set to be 10 multiplied by 32, and the dimension of output data is set to be 10 multiplied by L 2 (ii) a Wherein L is 1 At least 50; l is 2 ＝2×UI 1 ，UI 1 Is the number of data points contained within a unit interval UI.
Step 2, generating a training set containing time step information:
step 2.1, setting the time step of the bidirectional long-short term memory Bi-LSTM network as 10 and the training batch size as 10;
step 2.2, generating an N-bit pseudo random binary sequence PRBS, inputting the sequence into a high-speed link for simulation, and obtaining a discrete time sequence;
step 2.3, uniform interpolation is carried out on the discrete time sequence at equal intervals, so that each unit interval UI contains N insert Point to obtain an interpolated waveform sequence, where N insert At least 50; carrying out amplitude normalization processing on the waveform sequence after interpolation to obtain a waveform sequence after amplitude normalization;
step 2.4, taking each L continuous pseudo-random binary code of the N-bit pseudo-random binary sequence PRBS as a group, carrying out sliding window interception on the pseudo-random binary sequence, and setting the moving step length of a window to be 2 to obtain an input sample matrix;
step 2.5, normalize every L in the waveform sequence by the amplitude 2 The continuous data points are a group, the waveform sequence after the amplitude normalization is subjected to sliding window processing, and the moving step length of the window is set to be L win Obtaining an output sample matrix; wherein L is win ＝L sam1 ×200、L sam ＝(N-L)/2+1；
Step 2.6, delete L at the end of input sample matrix del Performing row data to obtain an input sample matrix; deleting L at the end of the output sample matrix del Performing row data to obtain an output sample matrix; wherein L is del ＝L sam -L sam /(10×10)；
Step 2.7, performing dimension increasing operation on the input sample matrix according to the time step of the long-term and short-term memory Bi-LSTM network in every 10 rows to obtain a three-dimensional input sample matrix; grouping the three-dimensional input sample matrix according to the batch size and the time sequence to obtain batch input data;
step 2.8, performing dimension increasing operation on the output sample matrix according to the time step of the long-term and short-term memory Bi-LSTM network in every 10 rows to obtain a three-dimensional output sample matrix; grouping the three-dimensional output sample matrix according to the batch size and the time sequence to obtain batch output data;
step 2.9, forming a training set by all batch input data and batch output data;
step 3, training the Bi-directional long-short term memory Bi-LSTM network by adopting a state transfer method:
step 3.1, setting the biases of the first bidirectional long-short term memory network layer and the second bidirectional long-short term memory network layer to be 0, and initializing the weights by using a Glorot initialization method;
step 3.2, inputting the training set into the Bi-directional long-short term memory Bi-LSTM network, and iteratively updating the parameters of the network by adopting a state transfer method until the mean square error value calculated by the loss function is converged to obtain the well-trained Bi-directional long-short term memory Bi-LSTM network;
step 4, generating an eye diagram by using the trained bidirectional long-short term memory Bi-LSTM network:
step 4.1, generating an M-bit pseudo random binary sequence PRBS, and processing the M-bit pseudo random binary sequence by adopting the same processing method for the N-bit pseudo random binary sequence PRBS in the step 2 to obtain batch input data; inputting batch input data into a trained bidirectional long-short term memory Bi-LSTM network to obtain a three-dimensional output test matrix;
step 4.2, reducing the dimension of the three-dimensional output test matrix according to the time step of the bidirectional long-short term memory Bi-LSTM network set in the step 2.2 to obtain a two-dimensional matrix; splicing each row of data in the two-dimensional matrix together to obtain a one-dimensional output waveform sequence; carrying out inverse amplitude normalization processing on the one-dimensional output waveform sequence to obtain a one-dimensional waveform sequence after inverse amplitude normalization;
step 4.3, every L of the one-dimensional waveform sequence after the inverse amplitude normalization 2 And carrying out primary interception on the data points, overlapping all the intercepted data points, and carrying out point tracing drawing on the data points after the overlapping is finished to obtain the eye diagram.
Compared with the prior art, the invention has the following advantages:
firstly, because the data required by the network trained by the invention is composed of a small amount of pseudo random binary sequence PRBS and response waveforms input by the small amount of pseudo random binary sequence PRBS and output by high-speed link simulation, the problem that the transient simulation technology in the prior art needs to simulate a large amount of pseudo random binary sequence PRBS to obtain an eye pattern with low error rate, so that the simulation time is too long is solved. The method has the advantage of high efficiency of predicting the eye pattern of the high-speed link.
Secondly, each group of data in the training data set generated by the invention consists of a string of pseudo random binary sequence PRBS and a response waveform output after the pseudo random binary sequence PRBS is input into the high-speed link, not only the current code bit but also the subsequent code bit are considered, and the accuracy of predicting the eye pattern is improved. The problem that the eye pattern prediction is inaccurate by a unit impulse response simulation method in the prior art is solved. The method and the device improve the accuracy of predicting the eye diagram of the high-speed link.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a high-speed link diagram built in ADS according to the present invention.
Detailed Description
The invention is described in further detail below with reference to the figures and examples.
The specific steps of the present invention will be described in further detail with reference to fig. 1 and the embodiment.
Step 1, constructing a bidirectional long-short term memory Bi-LSTM network.
Step 1.1, a bidirectional long-short term memory Bi-LSTM network which is formed by connecting a first bidirectional long-short term memory network layer, a second bidirectional long-short term memory network layer and a full connection layer in series is built.
Step 1.2, the number of the hidden units of the first bidirectional long-term memory network layer and the second bidirectional short-term memory network layer is set to be 16, and the activation functions used in each hidden unit are hyperbolic tangent functions. Sequentially setting the dimension of the input data of the first and second bidirectional long-short term memory network layers to 10 × 10 × L 1 10 × 10 × 32, the dimensions of the output data are set to 10 × 10 × 32. The dimension of input data of a full connection layer is set to be 10 multiplied by 32, and the dimension of output data is set to be 10 multiplied by L 2 . Wherein L is 1 At least 50 to ensure that the input data of the first bi-directional long short term memory network layer contains intersymbol interference (ISI) information of the high speed link, L 2 2 × UI, where UI is the number of points contained within one unit interval UI. L in the embodiment of the present invention 1 Set to 80 and the number of points contained within one unit interval UI is 100.
And 2, generating a training set containing time step information.
And 2.1, setting the time step of the bidirectional long-short term memory Bi-LSTM network as 10 and setting the training Batch Size (Batch Size) as 10.
Step 2.2, generating 16000 bit pseudo random binary sequence PRBS, inputting the sequence into a high-speed link built in a simulation software advanced Design system ADS (advanced Design System) for simulation, and obtaining a discrete time sequence, wherein each unit interval UI of the sequence contains 32 data points and 512000 data points in total.
The high-speed link map built in the ADS of the present invention is described in further detail with reference to fig. 2.
In the embodiment of the invention, the bit rate of the high-speed link built in the ADS is 16Gbps, and a channel simulation (ChannelSim) simulation method is used. The high-speed link consists of a sending end, a transmission line 1, a signal repeater, a transmission line 2 and a receiving end. The PRBS is input at the transmitting end of the high-speed link, and the discrete time sequence is obtained at the receiving end. The sending end uses an Input/Output Buffer Information Specification-algorithm Modeling Interface IBIS-AMI (Input/Output Buffer Information Specification-Algorithmic Modeling Interface) model. The transmission lines 1 and 2 use s (scattering) parametric models with different insertion losses. The receiving end uses an input/output buffer area information specification-algorithm modeling interface IBIS-AMI model. The signal repeater models the interface IBIS-AMI model using input/output buffer information specification-algorithms. The transmitting end and the transmission line 1 are in differential interconnection, the transmission line 1 and the signal repeater are in differential interconnection, the signal repeater and the transmission line 2 are in differential interconnection, and the transmission line 2 and the receiving end are in differential interconnection.
And 2.3, performing uniform interpolation on the discrete time sequence at equal intervals, so that each unit interval UI at least comprises 50 points, and the time intervals between each data point are the same, thereby ensuring the accuracy of predicting the eye diagram of the high-speed link. The bit rate of the high speed link in an embodiment of the invention is 16Gbps, i.e. the unit interval UI per bit is 6.25e-11 seconds. To achieve 100 data points in each unit interval UI, the time interval between each data point in each unit interval UI needs to be 6.25e to 13 seconds, that is, a data point at two adjacent times is uniformly interpolated according to the interval of 6.25e to 13 seconds in a discrete time sequence, so as to obtain an interpolated waveform sequence containing 1599997 data points. And carrying out amplitude normalization processing on the waveform sequence after interpolation to obtain the waveform sequence after amplitude normalization.
And 2.4, taking every 80-bit continuous pseudo-random binary code of the 16000-bit pseudo-random binary sequence PRBS as a group, carrying out sliding window interception on the pseudo-random binary sequence, and setting the moving step length of the window to be 2. The 7961 sets of sequences resulting from the sliding window are organized into an input sample matrix having dimensions 7961 x 80.
And 2.5, taking every 200 continuous data points in the waveform sequence after amplitude normalization as a group, and performing sliding window processing on the waveform sequence after amplitude normalization, wherein the moving step length of a window is set to 8000. The 7961 sets of sequences from the sliding window are organized into an output sample matrix with dimensions 7961 x 200.
And 2.6, in order to ensure that the row number of the input sample matrix can be divided by the result of multiplying the time step by the size of the training batch, deleting 61 rows of data at the tail of the input sample matrix to obtain the input sample matrix with the dimensionality of 7900 multiplied by 80. In order to ensure that the row number of the output sample matrix can be divided by the result of multiplying the time step length by the size of the training batch, 61 rows of data at the tail of the output sample matrix are deleted, and the output sample matrix with the dimensionality of 7900 x 200 is obtained.
And 2.7, performing dimensionality increasing operation on the input sample matrix every 10 rows according to the time step of the long-term and short-term memory network to obtain a three-dimensional input sample matrix with the dimensionality of 790 multiplied by 10 multiplied by 80. And dividing the three-dimensional input sample matrix into 79 groups of batch input data with the dimension of 10 multiplied by 80 according to the batch size and the time sequence, and numbering each group of data from the 1 st group to the 79 th group.
And 2.8, performing dimensionality increasing operation on the output sample matrix according to the time step of the long-term and short-term memory network in every 10 rows to obtain a three-dimensional output sample matrix with the dimensionality of 790 multiplied by 10 multiplied by 200. The three-dimensional output sample matrix is divided into 79 groups of batch output data with the dimension of 10 multiplied by 200 according to the batch size and the time sequence, and each group of data is numbered from the 1 st group to the 79 th group.
And 2.9, forming a training set by the batch input data and the batch output data of all groups.
And 3, training the Bi-directional long-short term memory Bi-LSTM network by adopting a state transfer method.
And 3.1, setting the bias of the first bidirectional long-short term memory network layer and the bias of the second bidirectional long-short term memory network layer to be 0, and initializing the weights by using a Glorot initialization method.
And 3.2, inputting the training set into the bidirectional long-short term memory Bi-LSTM network, and iteratively updating the parameters of the network by adopting a state transfer method until the mean square error value calculated by the loss function is converged to obtain the trained bidirectional long-short term memory Bi-LSTM network. The state transmission method refers to that in the parameter iterative updating process, each group of input data and output data in the training set are sequenced according to the time sequence, and are sequentially input into the Bi-directional long-short term memory Bi-LSTM network for parameter iterative updating.
In the embodiment of the invention, the loss function of the bidirectional long-short term memory Bi-LSTM network uses a Mean-Square Error (MSE) function, and the optimization algorithm of the loss function uses an adaptive Moment estimation adam (adaptive motion estimation) algorithm. And after the parameters of the network are iteratively updated for 300 times, the mean square error value calculated by the loss function is converged.
And 4, generating an eye diagram by using the trained bidirectional long-short term memory Bi-LSTM network.
And 4.1, generating a 1e5 bit pseudo-random binary sequence PRBS, and processing the 1e5 bit pseudo-random binary sequence by adopting the same processing method for the 16000 bit pseudo-random binary sequence PRBS in the step 2 to obtain 499 groups of batch input data with the dimension of 10 multiplied by 80. Inputting batch input data into a trained bidirectional long-short term memory Bi-LSTM network, and outputting a three-dimensional output test matrix with the dimensionality of 4990 multiplied by 10 multiplied by 200.
And 4.2, reducing the dimension of the three-dimensional output test matrix according to the time step of the bidirectional long-short term memory Bi-LSTM network set in the step 2.2 to obtain a two-dimensional matrix with the dimension of 49900 x 200. And splicing each row of data in the two-dimensional matrix together to obtain a one-dimensional output waveform sequence containing 9980000 data points. And carrying out inverse amplitude normalization processing on the one-dimensional output waveform sequence to obtain a one-dimensional waveform sequence after inverse amplitude normalization.
And 4.3, intercepting the one-dimensional waveform sequence after the inverse amplitude normalization once every 200 data points according to two time intervals UI, overlapping all the intercepted data, and drawing points of the data points after the overlapping is completed to obtain the eye diagram.
The effect of the present invention is further explained by combining the simulation experiment as follows:
1. simulation conditions are as follows:
the hardware platform of the simulation experiment of the invention is as follows: the processor is Intel (R) Xeon (R) Platinum 8269CY, the dominant frequency is 2.50GHZ, and the memory is 64 GB.
The software platform of the simulation experiment of the invention is as follows: windows 10 operating system, ADS2020 software, MATLAB R2020b software, Python 3.9, tensoflow 2.7.0.
2. Simulation content and result analysis thereof:
the simulation experiment of the invention is a method for predicting an eye pattern by adopting the technology of the invention, the transient simulation technology and the prior art (based on the unit impulse response technology). Respectively inputting 1e5 bit pseudo random binary sequences PRBS, and comparing the time required by each method for predicting the eye pattern and the accuracy of the predicted eye pattern.
In a simulation experiment, the method for predicting the eye pattern based on the unit impulse response in the prior art is used in the patent document "an eye pattern measuring method under additive noise interference and a device thereof and a storage medium" (patent application No. CN202210103888.6, application publication No. CN 114124318a) applied by shenzhen dingcheng honesty intellectual property agency limited company.
The results of predicting the eye height and the eye width of the eye pattern using the transient simulation technique, the unit impulse response technique, and the technique of the present invention are shown in table 1.
TABLE 1 eye height and eye width results summary of different methods for predicting eye diagram
| Method type | Eye width (ps) | Relative error (%) | Eye height (mv) | Relative error (%) |
| Transient simulation | 46.25 | -- | 90.6 | -- |
| Method based on unit impulse response | 40.63 | 12.16 | 80.3 | 11.37 |
| The method of the invention | 46.25 | 0 | 89.5 | 1.27 |
It can be seen from table 1 that the relative error of the eye height of the eye pattern obtained by the method for predicting an eye pattern using the technique of the present invention and the method for predicting an eye pattern using the transient simulation technique is 0%, and the relative error of the eye width is 1.27%. The method for predicting the eye pattern based on the unit impulse response technology and the method for predicting the eye pattern based on the transient simulation technology have the relative error of the eye height of 11.37% and the relative error of the eye width of 12.16%. The method for predicting the eye pattern by the technology has higher accuracy compared with the method for predicting the eye pattern by the technology.
Using the transient simulation technique, the time alignment ratios used to predict the eye pattern at different bit error rates, i.e., different pattern lengths, using the technique of the present invention are shown in table 2.
TABLE 2 comparison of time spent by each method for different pattern lengths
| Length of code pattern | Transient simulation (min) | Method of the invention (min) |
| 1e5 bits | 3.5 | 0.3 |
| 1e6 bits | 20.3 | 3.38 |
| 5e6 bits | 89.5 | 16.9 |
| 1e7 bits | 180.7 | 32.55 |
As can be seen from table 2, the method of predicting an eye pattern using the technique of the present invention consumes less time than the method of predicting an eye pattern using the transient simulation technique.
Claims (3)
1. A high-speed link prediction eye diagram method based on a Bi-LSTM network is characterized in that a training set containing time step information is generated, and a bidirectional long-short term memory Bi-LSTM network is trained by adopting a state transfer method; the method comprises the following specific steps:
step 1, constructing a bidirectional long-short term memory Bi-LSTM network:
step 1.1, constructing a bidirectional long-short term memory Bi-LSTM network formed by connecting a first bidirectional long-short term memory network layer, a second bidirectional long-short term memory network layer and a full connection layer in series;
step 1.2, the number of hidden units of the first bidirectional long-short term memory network layer and the second bidirectional long-short term memory network layer is set to be 16, and the activation function used in each hidden unit is realized by adopting a hyperbolic tangent function; sequentially setting the dimension of the input data of the first and second bidirectional long-short term memory network layers to 10 × 10 × L 1 10 × 10 × 32, the dimensions of the output data are all set to 10 × 10 × 32; the dimension of input data of a full connection layer is set to be 10 multiplied by 32, and the dimension of output data is set to be 10 multiplied by L 2 (ii) a Wherein L is 1 At least 50; l is 2 ＝2×UI 1 ，UI 1 Is the total number of data points contained within one unit interval UI;
step 2, generating a training set containing time step information:
step 2.1, setting the time step length of the bidirectional long-short term memory Bi-LSTM network as 10, and setting the size of a training batch as 10;
step 2.2, generating an N-bit pseudo random binary sequence PRBS, inputting the sequence into a high-speed link for simulation, and obtaining a discrete time sequence;
step 2.3, performing uniform interpolation at equal intervals on the discrete time sequence to enable each unit interval UI to contain N insert Point to obtain an interpolated waveform sequence, where N insert At least 50; carrying out amplitude normalization processing on the waveform sequence after interpolation to obtain amplitude normalizationA transformed waveform sequence;
step 2.4, pseudo-random binary sequence PRBS per L with N bits 1 The method comprises the steps that a pseudo-random binary code with continuous bits is a group, sliding window interception is carried out on a pseudo-random binary sequence, the moving step length of a window is set to be 2, and an input sample matrix is obtained;
step 2.5, normalizing every L in the waveform sequence by the amplitude 2 The continuous data points are a group, the waveform sequence after amplitude normalization is subjected to sliding window processing, and the moving step length of the window is set to be L win Obtaining an output sample matrix; wherein L is win ＝L sam1 ×200、L sam ＝(N-L)/2+1；
Step 2.6, delete L at the end of input sample matrix del Performing row data to obtain an input sample matrix; deleting L at the end of the output sample matrix del Performing row data to obtain an output sample matrix; wherein L is del ＝L sam -L sam /(10×10)；
Step 2.7, performing dimension increasing operation on the input sample matrix according to the time step of the long-term and short-term memory Bi-LSTM network in every 10 rows to obtain a three-dimensional input sample matrix; grouping the three-dimensional input sample matrix according to the batch size and the time sequence to obtain batch input data;
step 2.8, performing dimension increasing operation on the output sample matrix according to the time step of the long-term and short-term memory Bi-LSTM network in every 10 rows to obtain a three-dimensional output sample matrix; grouping the three-dimensional output sample matrix according to the batch size and the time sequence to obtain batch output data;
step 2.9, forming a training set by all batch input data and batch output data;
step 3, training the Bi-directional long-short term memory Bi-LSTM network by adopting a state transfer method:
step 3.1, setting the bias of the first bidirectional long-term memory network layer and the bias of the second bidirectional long-term memory network layer to be 0, and initializing the weights by using a Glorot initialization method;
step 3.2, inputting the training set into the Bi-directional long-short term memory Bi-LSTM network, and iteratively updating the parameters of the network by adopting a state transfer method until the mean square error value calculated by the loss function is converged to obtain the well-trained Bi-directional long-short term memory Bi-LSTM network;
step 4, generating an eye diagram by using the trained bidirectional long-short term memory Bi-LSTM network:
step 4.1, generating an M-bit pseudo random binary sequence PRBS, and processing the M-bit pseudo random binary sequence by adopting the same processing method for the N-bit pseudo random binary sequence PRBS in the step 2 to obtain batch input data; inputting the batch input data into a trained bidirectional long-short term memory Bi-LSTM network to obtain a three-dimensional output test matrix;
step 4.2, reducing the dimension of the three-dimensional output test matrix according to the time step of the bidirectional long-short term memory Bi-LSTM network set in the step 2.2 to obtain a two-dimensional matrix; splicing each row of data in the two-dimensional matrix together to obtain a one-dimensional output waveform sequence; carrying out inverse amplitude normalization processing on the one-dimensional output waveform sequence to obtain a one-dimensional waveform sequence after inverse amplitude normalization;
4.3, every L of the one-dimensional waveform sequence after the inverse amplitude normalization 2 And carrying out primary interception on the data points, overlapping all the intercepted data points, and carrying out point tracing drawing on the data points after the overlapping is finished to obtain the eye diagram.
2. The Bi-LSTM network-based eye pattern prediction method for high-speed links according to claim 1, wherein the state transmission method in step 3.2 is to sort each batch of input data and batch of output data in the training set according to the chronological order during the iterative update of parameters, and input the sorted batch of input data and batch of output data into the Bi-directional long-short term memory Bi-LSTM network in sequence for iterative update of parameters.
3. The Bi-LSTM network-based eye pattern prediction method for high-speed links according to claim 1, wherein the loss function in step 3.2 is:
wherein the content of the first and second substances,
means for representing the mean squared error value, N, of the batch output data in the f-th training set and the batch output data in the f-th training set predicted by the Bi-LSTM network in the q-th iteration S Represents the length of the f batch output data in the training set, n represents the index number of the data point of the f batch output data in the training set, Σ represents the summation operation, y n Representing the nth data point in the f-th batch of output data in the training set,
a predicted value representing the nth data point of the ith batch of output data in the training set.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210691524.4A CN115102674B (en) | 2022-06-17 | 2022-06-17 | Bi-LSTM network-based high-speed link eye diagram prediction method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210691524.4A CN115102674B (en) | 2022-06-17 | 2022-06-17 | Bi-LSTM network-based high-speed link eye diagram prediction method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115102674A true CN115102674A (en) | 2022-09-23 |
| CN115102674B CN115102674B (en) | 2023-08-22 |
Family
ID=83291591
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210691524.4A Active CN115102674B (en) | 2022-06-17 | 2022-06-17 | Bi-LSTM network-based high-speed link eye diagram prediction method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115102674B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115941129A (en) * | 2022-10-17 | 2023-04-07 | 西安电子科技大学 | High-speed link eye diagram implementation method based on linear fitting error matrix |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108170529A (en) * | 2017-12-26 | 2018-06-15 | 北京工业大学 | A kind of cloud data center load predicting method based on shot and long term memory network |
| CN108647643A (en) * | 2018-05-11 | 2018-10-12 | 浙江工业大学 | A kind of packed tower liquid flooding state on-line identification method based on deep learning |
| CN110233661A (en) * | 2019-04-28 | 2019-09-13 | 国网安徽省电力有限公司信息通信分公司 | Shot and long term Memory Neural Networks training method, channel parameter adjust system and method |
| CN110932809A (en) * | 2019-11-20 | 2020-03-27 | 北京邮电大学 | Optical fiber channel model simulation method and device |
| CN111541466A (en) * | 2020-04-20 | 2020-08-14 | 北京航空航天大学 | Unmanned aerial vehicle data link design method based on random modulation hopping |
| CN112838909A (en) * | 2020-12-30 | 2021-05-25 | 西安电子科技大学 | Communication interference detection method based on Gaussian eye pattern texture entropy characteristics |
| CN113687433A (en) * | 2021-09-15 | 2021-11-23 | 湖南师范大学 | Bi-LSTM-based magnetotelluric signal denoising method and system |
| US20220007198A1 (en) * | 2020-07-01 | 2022-01-06 | Celona, Inc. | Method and Apparatus for Monitoring and Predicting Channel Availability and Preemptively Reconfiguring the Network in a Spectrum Controlled Network |
| CN114124318A (en) * | 2022-01-28 | 2022-03-01 | 深圳市鼎阳科技股份有限公司 | Eye pattern measuring and calculating method and device under additive noise interference and storage medium |
-
2022
- 2022-06-17 CN CN202210691524.4A patent/CN115102674B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108170529A (en) * | 2017-12-26 | 2018-06-15 | 北京工业大学 | A kind of cloud data center load predicting method based on shot and long term memory network |
| CN108647643A (en) * | 2018-05-11 | 2018-10-12 | 浙江工业大学 | A kind of packed tower liquid flooding state on-line identification method based on deep learning |
| CN110233661A (en) * | 2019-04-28 | 2019-09-13 | 国网安徽省电力有限公司信息通信分公司 | Shot and long term Memory Neural Networks training method, channel parameter adjust system and method |
| CN110932809A (en) * | 2019-11-20 | 2020-03-27 | 北京邮电大学 | Optical fiber channel model simulation method and device |
| CN111541466A (en) * | 2020-04-20 | 2020-08-14 | 北京航空航天大学 | Unmanned aerial vehicle data link design method based on random modulation hopping |
| US20220007198A1 (en) * | 2020-07-01 | 2022-01-06 | Celona, Inc. | Method and Apparatus for Monitoring and Predicting Channel Availability and Preemptively Reconfiguring the Network in a Spectrum Controlled Network |
| CN112838909A (en) * | 2020-12-30 | 2021-05-25 | 西安电子科技大学 | Communication interference detection method based on Gaussian eye pattern texture entropy characteristics |
| CN113687433A (en) * | 2021-09-15 | 2021-11-23 | 湖南师范大学 | Bi-LSTM-based magnetotelluric signal denoising method and system |
| CN114124318A (en) * | 2022-01-28 | 2022-03-01 | 深圳市鼎阳科技股份有限公司 | Eye pattern measuring and calculating method and device under additive noise interference and storage medium |
Non-Patent Citations (4)
| Title |
|---|
| HUAN HUAN ZHANG; ZHAO SHENG XUE; XIN YI LIU; PING LI; LIJUN JIANG; GUANG MING SHI SCHOOL OF ARTIFICIAL INTELLIGENCE, XIDIAN UNIVER: "Optimization of High-Speed Channel for Signal Integrity With Deep Genetic Algorithm", 《IEEE TRANSACTIONS ON ELECTROMAGNETIC COMPATIBILITY ( VOLUME: 64, ISSUE: 4, AUGUST 2022)》 * |
| XIUQIN CHU; WENTING GUO; JUN WANG; FENG WU; YUHUAN LUO KEY LABORATORY OF HIGH SPEED CIRCUIT DESIGN AND EMC MINISTRY OF EDUCATION, : "Fast and Accurate Estimation of Statistical Eye Diagram for Nonlinear High-Speed Links", 《 IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS ( VOLUME: 29, ISSUE: 7, JULY 2021)》 * |
| 罗玉焕: "基于D优化设计与神经网络的快速多边沿响应法", 《中国优秀硕士学位论文全文数据库》 * |
| 黄君浩;贺辉;: "基于LSTM的眼动行为识别及人机交互应用", 《计算机系统应用》, no. 03 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115941129A (en) * | 2022-10-17 | 2023-04-07 | 西安电子科技大学 | High-speed link eye diagram implementation method based on linear fitting error matrix |
| CN115941129B (en) * | 2022-10-17 | 2024-04-16 | 西安电子科技大学 | High-speed link eye diagram implementation method based on linear fitting error matrix |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115102674B (en) | 2023-08-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US9058442B2 (en) | Incorporating noise and/or jitter into waveform generation | |
| US9047425B2 (en) | Time-domain signal generation | |
| CN112508243B (en) | Training method and device for multi-fault prediction network model of power information system | |
| Venema et al. | A stochastic iterative amplitude adjusted Fourier transform algorithm with improved accuracy | |
| CN101341488A (en) | Synthesizing current source driver model for analysis of cell characteristics | |
| CN109188410B (en) | Distance calibration method, device and equipment under non-line-of-sight scene | |
| CN113687433B (en) | Bi-LSTM-based magnetotelluric signal denoising method and system | |
| CN115102674B (en) | Bi-LSTM network-based high-speed link eye diagram prediction method | |
| CN110932809A (en) | Optical fiber channel model simulation method and device | |
| CN113761805B (en) | Controllable source electromagnetic data denoising method, system, terminal and readable storage medium based on time domain convolution network | |
| Goay et al. | Transient simulations of high-speed channels using CNN-LSTM with an adaptive successive halving algorithm for automated hyperparameter optimizations | |
| CN112115821A (en) | Multi-signal intelligent modulation mode identification method based on wavelet approximate coefficient entropy | |
| CN115097398A (en) | Radar anti-interference signal recovery method based on cross-domain signal low-loss recovery network | |
| CN113676266B (en) | Channel modeling method based on quantum generation countermeasure network | |
| CN111398928B (en) | Method for calculating detection threshold of synthetic ultra-narrow pulse radar based on resampling algorithm | |
| CN109309536B (en) | Method for reducing approximation complexity of Nakagami inverse CDF function | |
| CN115983094A (en) | Logging curve generation method based on S-CNN-Bi-GRU network, processing terminal and readable storage medium | |
| Jiao et al. | Method for accurate and efficient eye diagram prediction of nonlinear high-speed links | |
| CN113361204B (en) | Radar radiation source PRI modulation identification method based on SAE | |
| Ge et al. | A new MCMC algorithm for blind Bernoulli-Gaussian deconvolution | |
| CN115941129B (en) | High-speed link eye diagram implementation method based on linear fitting error matrix | |
| Goay et al. | Temporal convolutional networks for transient simulation of high-speed channels | |
| Yang et al. | Deep-learning-based bouncing-order prediction for propagation channel characterization using graph modeling | |
| CN112821971A (en) | Time-varying channel signal detection method based on countermeasure learning | |
| CN113516168B (en) | Multidimensional electric quantity continuous time sequence generation method based on generation countermeasure network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |