Eye diagram degradation factor quantification diagnosis method and system based on convolutional neural network

Abstract

The application discloses a kind of based on convolutional neural network's eye diagram deterioration factor's quantitative diagnosis method, mainly solve the problems of existing technology in eye diagram analysis relies on artificial experience, multi-factor coupling is difficult to separate and fault attribution is fuzzy.The scheme is: constructing high-speed link parameterized simulation model, respectively mapping reflection, loss and crosstalk as load resistance, dielectric loss tangent and line spacing etc.Parameter, and generate large-scale eye diagram sample data set using collaborative simulation framework;For each eye diagram sample, construct the health degree score label system including reflection, loss, crosstalk three dimensions;Convolutional neural network model is constructed and trained, learn the mapping relationship of eye diagram image to three-dimensional health degree score;The eye diagram to be tested is input to the trained model, and the three health degree scores are compared by output, to identify and quantify the leading factors causing eye diagram deterioration.The application can realize the automatic, standardization intelligent diagnosis of high-speed link eye diagram, accurately locate the problem, and can be used for high-speed link design verification and signal integrity optimization.

Description

Technical Field

[0001] This invention belongs to the field of electronic technology, and further relates to a quantitative diagnostic method and system for eye diagram degradation factors. It can be used to perform automated and quantitative analysis of eye diagrams of high-speed interconnect links, accurately identify and quantify the dominant factors causing eye diagram degradation, including reflection, crosstalk, and loss, and provide efficient and reliable intelligent diagnostic basis for the design verification and signal integrity optimization of high-speed links. Background Technology

[0002] With the trend towards higher frequencies and smaller sizes in high-speed digital systems, ensuring signal integrity (SI) of interconnect links has become a core challenge in electronic design. Eye diagrams, as a key visualization tool for evaluating the quality of high-speed links, directly reflect the combined impact of various SI issues such as reflections, crosstalk, and losses on signal quality. Traditional eye diagram analysis methods rely on engineers' experience to qualitatively judge parameters such as eye height and eye width.

[0003] However, when multiple influencing factors are superimposed and coupled in real-world scenarios, manual interpretation makes it difficult to accurately separate the contributions of each factor, leading to ambiguous attribution. Consequently, SI design optimization becomes untargeted, inefficient, and costly.

[0004] US Patent Application No. 9633164B2 discloses a "Signal Integrity Waveform Decomposition Analysis System and Method," which quantifies the contribution of each loop to eye diagram closure by identifying resonant loops in a hardware interconnect system and calculating the impact of each loop on the overall system response. While this method can attribute eye diagram closure to specific physical locations, its process is complex, relying on precise S-parameter models and complex mathematical decompositions. Furthermore, it does not directly utilize information from the eye diagram image itself, resulting in extremely high computational costs and failing to meet the demands of rapid iterative design.

[0005] Patent document CN107342810A discloses a "Deep Learning Intelligent Eye Diagram Analysis Method Based on Convolutional Neural Networks," which is the first to apply CNN to eye diagram analysis. It trains the network to automatically extract features from eye diagram images to identify preset communication system parameters. While this method achieves end-to-end automated analysis, its output consists of general performance indicators rather than directly corresponding to the fundamental physical factors causing signal degradation in high-speed digital circuits, such as reflection, loss, and crosstalk. Therefore, the interpretability of its diagnostic results and its direct guidance for circuit design are limited.

[0006] In their paper "Eye Diagram Analysis with Deep Neural Networks for Signal Integrity Applications", Miao Weiyang et al. proposed a method using a CNN model to directly predict physical parameters related to reflection, loss, and crosstalk from a single eye diagram, so as to directly output the predicted physical parameter values ​​of impedance, dielectric loss tangent, and spacing. However, this method lacks the evaluation of eye diagram quality and has limited practical guiding significance.

[0007] Patent document CN202510765420.7 discloses an "LPDDR4X simulation method and system based on statistical theory," which constructs a time-domain simulation platform, combines impulse response and crosstalk noise analysis, and uses statistical methods to calculate eye diagram height to address the shortcomings of traditional methods in assessing eye height. While this method can accurately quantify eye height, its core lies in deterministic simulation based on a physical model and noise superposition, without utilizing machine learning models to learn features from eye diagram images. Therefore, it cannot achieve automated diagnosis of unknown eye diagrams or quantitative ranking of degradation factors such as reflection, loss, and crosstalk. Summary of the Invention

[0008] The purpose of this invention is to provide an intelligent eye diagram diagnosis method based on supervised learning, in order to solve the problems of fault attribution relying on experience, difficulty in separating multiple coupled factors, and inability to quantify the contribution of each factor in existing eye diagram analysis, and to realize automated and standardized diagnosis of high-speed link eye diagrams and quantitative ranking of dominant degradation factors.

[0009] The technical solution to achieve the objective of this invention is: to use a convolutional neural network to learn the nonlinear mapping relationship between distortion patterns and the influence mechanisms of various factors from eye diagram images that have been attributed, which includes:

[0010] 1. A quantitative diagnostic method for eye diagram degradation factors based on convolutional neural networks, characterized in that it includes:

[0011] (1) Construct a high-speed link parameterized simulation model as a reference template;

[0012] (2) Based on the deep integration of Python scripts and ADS simulation software, an ADS-Python co-simulation framework is constructed to automatically modify simulation parameters, calculate the health score of three deterioration factors for each sample, and generate large-scale eye diagram datasets in batches.

[0013] (3) Based on the reference template, use the co-simulation framework to generate a dataset with three-dimensional health score labels;

[0014] (4) Use the dataset to train a convolutional neural network to learn the mapping relationship between each eye diagram image and the three types of health scores: reflection, loss, and crosstalk, and obtain a trained eye diagram convolutional neural network diagnostic model.

[0015] (5) Input the eye diagram to be diagnosed into the trained eye diagram convolutional neural network diagnostic model to obtain the three predicted health scores for reflection, loss and crosstalk, and realize the intelligent diagnosis and quantitative ranking results of the dominant factors of eye diagram deterioration.

[0016] Furthermore, in (2), an ADS-Python co-simulation framework is constructed based on the deep integration of Python scripts and ADS simulation software. Its implementation includes:

[0017] 2a) Establish a main control unit using Python as the scripting language, responsible for parameter parsing, dynamic reconstruction of the netlist, task scheduling, and data post-processing;

[0018] 2b) Establish a computing core with ADS as the simulation engine to be responsible for executing the underlying link transient edge simulation tasks;

[0019] 2c) The main control unit and the computing core interact via file streams, i.e., Python transmits the netlist file through a terminal...

[0020] The input is fed into ADS to drive the simulation. After ADS completes the calculation, it generates a data file and sends it back to Python for reading and parsing.

[0021] Furthermore, in step (3), a standard dataset with three-dimensional health score labels is generated based on a reference template using a co-simulation framework. The implementation of this includes:

[0022] 3a) Utilize the ADS-Python co-simulation framework to automatically scan three key parameters: termination resistance R, dielectric loss Tan D, and line spacing S, generating parameter combinations that cover both single and composite factors;

[0023] 3b) Using the ideal baseline eye diagram as the anchor point, obtain the eye height H of the baseline eye diagram. ref Wide eyes W ref and level L ref ,

[0024] Then extract the eye height H from the eye diagram to be tested. meas Wide eyes W meas Overshoot ratio OS meas and level offset |L meas -L ref |;

[0025] 3c) Truncation operator via normalization The formulas calculate the four sub-scores: amplitude integrity, timing integrity, overshoot penalty, and level stability.

[0026] 3d) Based on the above four sub-scores, calculate the final health score using a weighted summation. ;

[0027] 3e) For the three single-factor scenarios where only R, Tan D, and S are changed, their health scores are denoted as scoreR, scoreTanD, and scoreS, respectively.

[0028] The scoreS is calculated, and the three scores are ranked to obtain the dominant deterioration factor and the relative severity of multiple factors; for composite interference scenarios that simultaneously change R, Tan D, and S, the comprehensive health score is also calculated.

[0029] 3f) For each scenario in each case, waveform data is obtained by modifying the netlist and calling the ADS engine to perform time-domain simulation. The data is then converted into a probability density distribution heatmap to generate an eye diagram. Scores and rankings are calculated according to preset rules. Finally, the scores and rankings are used as labels to classify and store the eye diagrams in order to construct a standard dataset for training CNN models.

[0030] 2. A quantitative diagnostic system for eye diagram degradation factors based on convolutional neural networks, characterized in that it includes:

[0031] The high-speed link parameterized simulation model block is used to map the three types of signal integrity problems in high-speed links—reflection, loss, and crosstalk—to quantifiable physical parameters such as load resistance R, dielectric loss tangent Tan D, and line spacing S, respectively, as reference templates for sample training.

[0032] The ADS-Python co-simulation module is used to perform simulations on specified parameter combinations, calculate three-dimensional health scores, and add corresponding labels.

[0033] The dataset generation module is used to generate large-scale standard eye diagram datasets in batches as training samples for the model.

[0034] The model training module is used to train a convolutional neural network (CNN) diagnostic model, learning the nonlinear mapping relationship between eye diagram image features and three-dimensional health scores of reflection, loss, and crosstalk.

[0035] The intelligent diagnosis and ranking module is used to output predicted health scores for three dimensions: reflection, crosstalk, and loss. Based on the scores, it automatically identifies and quantifies the dominant factors that cause the decline in eye diagram quality.

[0036] Compared with the prior art, the present invention has the following advantages:

[0037] Firstly, since the present invention outputs a health score in three dimensions—reflection, loss, and crosstalk—it is more in line with engineering optimization intuition. Engineers can directly identify the dominant degradation factors and determine the optimization priority based on the principle that "the lower the score, the more severe the degradation" without secondary interpretation, thus making up for the deficiency of existing technologies that cannot directly determine the optimization direction.

[0038] Secondly, the training model of this invention has made key improvements based on the ResNet-18 backbone network, which can adapt to the special needs of SI analysis. Specifically, firstly, a multi-task learning model architecture is designed, which outputs three health scores simultaneously through a single network, forcing the network to learn the feature representations that distinguish different physical effects; secondly, by designing a special loss function, MSE can be explicitly decomposed into the sum of the prediction errors of three scores: reflection, loss, and crosstalk. Moreover, this loss function can force the model to optimize its ability to identify three different interference factors in a balanced manner, thereby improving the diagnostic accuracy of the model in scenarios with multiple coupled factors.

[0039] Thirdly, this invention constructs a three-dimensional health labeling system based on an ideal eye diagram and comprehensively calculates key indicators such as eye height, eye width, overshoot, and signal level for each physical effect. This enables the model training to systematically evaluate single or compound degradation scenarios. The three final output scores can be directly quantified, compared, and ranked, clearly revealing the relative contribution of each factor to the overall signal degradation.

[0040] Fourth, because this invention is based on an interpretable quantization algorithm anchored by an ideal eye diagram, the health score is not a "black box" output. Instead, it compares the measured indicators of the eye diagram under test with the ideal benchmark and introduces normalization and weighted calculation, so that each health score has a clear physical meaning and calculation logic, which greatly enhances the credibility and acceptability of the model results. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating the implementation of the quantitative diagnostic method for eye diagram degradation factors based on convolutional neural networks according to the present invention.

[0042] Figure 2 This is a simulation topology diagram of the high-speed link built in the method of this invention;

[0043] Figure 3 This is an ideal reference eye diagram for an example link in the method of this invention;

[0044] Figure 4 This is a diagram of the convolutional neural network architecture used in the method of this invention;

[0045] Figure 5 This is a block diagram of the quantitative diagnostic system for eye diagram degradation factors based on convolutional neural networks according to the present invention. Detailed Implementation

[0046] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative effort should all fall within the protection scope of the present invention.

[0047] Example 1: A quantitative diagnostic method for eye diagram degradation factors based on convolutional neural networks.

[0048] Reference Figure 1 This example includes the following steps:

[0049] Step 1: Construct a parameterized simulation model for the high-speed link.

[0050] This step will construct a parameterized simulation model for the high-speed link. By mapping the three types of signal integrity problems in the high-speed link—reflection, loss, and crosstalk—to quantifiable physical parameters such as load resistance R, dielectric loss tangent Tan D, and line spacing S, these parameters will serve as reference templates for subsequent sample training.

[0051] 1.1) The high-speed link to be analyzed is set as a typical T-type topology, which includes a main control chip and multiple external DDR3 chips. Its core feature is that the clock, address and command signals of the chip in the link have strictly equal trace lengths to each DDR3 chip. Since this topology is extremely sensitive to signal reflection and crosstalk effects, it is an ideal test object for verifying the effectiveness of intelligent diagnostic algorithms.

[0052] 1.2) In the Keysight ADS simulation environment, build an equivalent circuit model based on the above T-shaped topology, such as... Figure 2 As shown, it includes a transmitter, a T-type transmission line link, coupled transmission lines, terminating resistors, an eye diagram detector, and a receiver. R2 is the terminating resistor of the T-type topology, Tan D is the loss tangent of the transmission line material used in the link, and the coupling transmission line spacing S is the spacing between the attacking line and the victim line in the crosstalk structure. This model reproduces the end-to-end transmission behavior of high-speed signals on the DDR3 bus.

[0053] 1.3) To obtain a reference baseline eye diagram for subsequent eye diagram scoring, in Figure 2 Based on this, crosstalk structures were removed, R2 was adjusted to terminate at 50Ω, material TanD was set to 0.017, and conditions such as the stack-up structure and transmission line length were locked. The characteristic impedance was controlled by adjusting the transmission line width. Finally, the eye diagram generated under these conditions was used as the reference eye diagram for this T-type topology. Figure 3 As shown, the ideal link's reference eye diagram has an eye height of 0.226 V, an eye width of 0.5 ns, and steady-state "1" and "0" levels of 0.495 V and 0.253 V, respectively. This reference eye diagram will serve as a benchmark for subsequent quantization of signal degradation and for constructing a dataset labeling system.

[0054] Step 2: Build the ADS-Python co-simulation framework.

[0055] The ADS-Python co-simulation framework uses a Python program as the main control unit to coordinate parameter generation, simulation scheduling, and data post-processing. It uses the ADS simulation engine hpeesofsim as the computational core to execute the underlying high-speed link simulation tasks. Its implementation includes:

[0056] 2.1) Establish a main control unit using Python as the scripting language, responsible for parameter parsing, dynamic reconstruction of the netlist, task scheduling, and data post-processing;

[0057] 2.2) Establish a computing core with ADS as the simulation engine to be responsible for executing the underlying link transient edge simulation tasks;

[0058] 2.3) The main control unit and the computing core interact via file stream. Specifically, Python inputs the netlist file to ADS through the terminal to drive the simulation. After ADS completes the calculation, it generates a data file and feeds it back to Python for reading and parsing.

[0059] Step 3: Use a co-simulation framework to generate a standard dataset with three-dimensional health score labels.

[0060] This step uses the ADS-Python co-simulation framework to perform batch simulations and generates corresponding labels for the samples using a three-dimensional health rating system to generate a dataset for training convolutional neural networks. Its implementation includes:

[0061] 3.1) Generate full parameter combination:

[0062] 3.1.1) The range of variation for three key parameters is defined by a Python program within the ADS-Python co-simulation framework:

[0063] The parameters of the terminating resistor R2 vary from 30 to 300 Ω.

[0064] The dielectric loss tangent, Tan, varies from 0.01 to 0.05.

[0065] The coupling line spacing parameter S_mclin varies from 50 to 500 mil.

[0066] 3.1.2) Perform a Cartesian product operation on the above parameter range to generate a complete parameter combination, and then randomly shuffle it from...

[0067] 2000 samples were selected as the data source for this model training.

[0068] 3.2) Generate the corresponding netlist based on the parameter combination:

[0069] 3.2.1) For each extracted parameter combination, an independent Case directory is automatically created, and four sub-cases are created under it.

[0070] Scene directories: Scene R_only, which modifies only the terminating resistance; Scene TanD_only, which modifies only the material loss tangent; Scene S_only, which modifies only the line spacing; and Scene All, which modifies all three parameters simultaneously. Each scene directory will copy a basic netlist template file.

[0071] 3.2.2) Modify parameters in the netlist file for the corresponding scenario using regular expression matching:

[0072] In the R_only scenario, only the terminating resistor R2 is modified;

[0073] In the TanD_only scenario, only the material loss tangent angle Tan D is modified;

[0074] In the S_only scenario, only the line spacing S is modified;

[0075] All scenarios will modify the above three parameters simultaneously;

[0076] 3.3) Establish a three-dimensional health scoring system and generate scores:

[0077] 3.3.1) Using the ideal baseline eye diagram as the anchor point, obtain the eye height H of the baseline eye diagram. ref Wide eyes W ref and level L ref Then extract the eye height H of the eye map to be tested in each scene. meas Wide eyes W meas Overshoot ratio OS meas and level offset |L meas -L ref |;

[0078] 3.3.2) For each scenario, use the normalization truncation operator. The formulas calculate the following four sub-scores respectively:

[0079] Amplitude integrity: ,

[0080] Timing integrity: ,

[0081] Overshoot penalty: ,

[0082] Level stability: ;

[0083] 3.3.3) Based on the above four sub-scores, a weighted summation is used to calculate the final health score for each scenario. :

[0084] ,

[0085] in, It is a weighting factor for amplitude integrity. It is a weighting factor for temporal integrity. It is the weighting coefficient of the overshoot penalty. These are weighting coefficients for level stability. Their values ​​are determined by the characteristics of the link, with higher weighting coefficients for more sensitive characteristics.

[0086] 3.3.4) For the three single-factor scenarios that only change R, Tan D, and S, their health scores are recorded as scoreR, scoreTanD, and scoreS, respectively. The three scores are then ranked to obtain the dominant deterioration factor and the relative severity of multiple factors. For the composite interference scenario that changes R, Tan D, and S simultaneously, the comprehensive health score scoreAll is also calculated.

[0087] 3.4) Perform batch simulations to obtain the dataset:

[0088] 3.4.1) Once the netlist for each scenario is ready, the write_bat function is called to automatically generate a Windows batch script. This script is responsible for configuring the ADS environment variables and calling hpeesofsim to execute the simulation task for the corresponding netlist.

[0089] 3.4.2) After the simulation is completed, the script automatically calls the dsdump command to convert the dataset file containing eye diagram time domain data generated by ADS into plain text format;

[0090] 3.4.3) The Python program reads the above plain text file, parses out the time and density data and saves it as a CSV file. At the same time, it uses the above three-dimensional health rating system to calculate the score and uses the score and ranking results as the label for the scene.

[0091] 3.4.4) The program generates a 300×300 two-dimensional histogram based on the data in the CSV file. After normalization by column, the probability density distribution is obtained and plotted as a heat map. The eye map is then classified and stored according to the label. Finally, the program generates a normalized matrix of the same dimension to construct a standard dataset for training CNN models.

[0092] Step 4: Construct a convolutional neural network diagnostic model.

[0093] Reference Figure 4 This step involves building a convolutional neural network diagnostic model based on the existing ResNet-18 convolutional neural network as the backbone. The model takes an eye diagram as input and a 3D health score as output. This includes:

[0094] 4.1) A ResNet-18 convolutional neural network consisting of an initial convolutional layer, four residual block groups, and a global average pooling layer is selected as the core backbone, wherein:

[0095] The initial convolutional layer consists of a 7×7 convolutional layer with a stride of 2 and 64 output channels, and a 3×3 max pooling layer with a stride of 2 and 64 output channels, from top to bottom, used to extract the most basic low-level features from the original input image.

[0096] The four residual blocks, from top to bottom, include: two Layer 1 residual blocks with a step size of 1 and 64 output channels, two Layer 2 residual blocks with a step size of 2 and 128 output channels, two Layer 3 residual blocks with a step size of 2 and 256 output channels, and two Layer 4 residual blocks with a step size of 2 and 512 output channels. These residual blocks are used to progressively extract increasingly abstract and higher-level semantic features from the image, starting from the lower-level features.

[0097] The global average pooling layer is used to compress the high-dimensional feature map output by the previous residual block group into a fixed-length vector, and finally outputs a 512-dimensional feature vector matrix FT.

[0098] 4.2) In order to adapt the model's mapping from eye diagram features to three-dimensional health score output, a fully connected embedding layer and a linear regression layer are added sequentially after the above network model to form a convolutional neural network diagnostic model.

[0099] Step 5: Train the convolutional neural network diagnostic model.

[0100] 5.1) Divide the 2000 eye diagram samples obtained from step 3 simulation into a training set containing 1600 samples and a validation set containing 400 samples in an 8:2 ratio;

[0101] 5.2) The eye diagrams of the training set are converted into two-dimensional probability histograms as input data for the model. Specifically, the sampling points are uniformly concentrated on a two-dimensional plane composed of time and voltage amplitude, and a 300×300 uniform grid is divided. The number of sample points falling into each grid cell is counted and global normalization is performed to obtain the two-dimensional probability histogram matrix P. Considering that the backbone network ResNet-18 accepts 3-channel input by default, while matrix P is single-channel data, the single-channel matrix P needs to be copied and expanded in the channel dimension to construct the final input tensor X with dimensions of 3×300×300.

[0102] 5.3) Define the mean squared error loss function L for the convolutional neural network diagnostic model to supervise training quality. This function explicitly decomposes the total error into three components: reflection, loss, and crosstalk, forcing the network to learn the feature distributions of these three interference factors in a balanced manner during backpropagation. Its expression is:

[0103] ;

[0104] in, This is a reflection of the health score vector. This is the health loss score vector. This is the crosstalk health score vector. The model prediction for reflection. The model predicts the loss. , where B is the model prediction for crosstalk, and B is the batch size.

[0105] 5.4) Set the initial learning rate to 2×10⁻ 4 The weight decay is 1×10⁻ 4 The batch size is 32 and the number of training rounds is 80.

[0106] 5.5) The specific training process for a convolutional neural network is as follows:

[0107] 5.5.1) Forward propagation and score prediction:

[0108] The preprocessed training set input tensor X is fed into the convolutional neural network diagnostic model. The data first passes through an initial convolutional layer, reducing the feature map size to 150×150. Then, it passes through a max pooling layer to obtain a 75×75 feature map. Next, the feature map passes through four residual block groups, and the size is finally reduced to 10×10. After that, it is compressed by a global average pooling layer, outputting a 512-dimensional feature vector matrix FT. This feature vector is then compressed into a 256-dimensional latent feature vector through a fully connected embedding layer. Finally, it outputs a three-dimensional health score of reflection, loss, and crosstalk through a linear regression output layer.

[0109] 5.5.2) Loss Calculation and Parameter Update:

[0110] With the goal of minimizing the mean squared error loss function of the L model, after obtaining the predicted score, the loss value between the predicted score and the true label is calculated. The gradient is calculated using the backpropagation algorithm, and the convolutional kernel weights and fully connected layer parameters are iteratively updated through the optimizer, so that the predicted score of the model in the next round is closer to the true label.

[0111] 5.5.3) Validation, Evaluation, and Optimal Model Saving:

[0112] After each training set iteration, pause network parameter updates, input the eye diagram of the validation set into the current model for forward propagation, and calculate its mean squared error loss L and prediction accuracy on the validation set.

[0113] The system monitors the validation set performance in real time: if the validation set loss in the current round is the lowest in history, the current model parameters are updated and saved; if no lower loss occurs, the original optimal parameters are retained.

[0114] 5.5.4) Repeat steps 5.5.1) to 5.5.3) to allow the model to continuously learn features and optimize parameters on the training set until the preset upper limit of 80 rounds is reached and the loss on the validation set tends to converge. Then terminate the iteration process and obtain the trained convolutional neural network diagnostic model.

[0115] Step 6: Intelligent diagnosis and ranking of the dominant factors in eye diagram degradation.

[0116] The actual eye diagram to be diagnosed is input into the trained convolutional neural network model, and the model can output a predicted health score for three dimensions: reflection, crosstalk, and loss.

[0117] By comparing the three scores based on the principle that "the lower the score, the more severe the degradation," the dominant degradation factors leading to eye diagram closure can be identified, and the severity of all factors can be quantified and ranked, thus providing engineers with a basis for link optimization.

[0118] It should be noted that the step numbers in the specification and claims of this invention are only for the purpose of clearly describing the embodiments of this invention and facilitating understanding, and their order is not limited.

[0119] Example 2: A quantitative diagnostic system for eye diagram degradation factors based on convolutional neural networks.

[0120] Reference Figure 5This example includes: a high-speed link parameterization simulation module 1, an ADS-Python co-simulation module 2, a dataset generation module 3, a model training module 4, and an intelligent diagnosis and ranking module 5. Specifically: the ADS-Python co-simulation module 2 includes: a circuit simulation submodule 21, an eye diagram sample generation submodule 22, and an intelligent diagnosis and ranking module 23; the model training module 3 includes: a CNN model construction submodule 31, a CNN model training submodule 32, and a CNN model supervision submodule 33; and the intelligent diagnosis and ranking module 5 includes: an eye diagram diagnosis submodule 51 and a result output submodule 52.

[0121] During system operation, the high-speed link parameterization simulation module 1 first constructs a reference template encompassing three types of signal integrity issues: reflection, loss, and crosstalk, mapping physical parameters to load resistance R, dielectric loss tangent Tan D, and line spacing S. Subsequently, the ADS-Python co-simulation module 2 schedules batch simulation tasks based on the template generated by module 1. The circuit simulation submodule 21 scans key parameters and executes the simulation, while the eye diagram sample generation submodule 22 converts the waveforms obtained from module 21's simulation into thermal eye diagrams. The three-dimensional health label construction submodule 23 calculates and generates health score labels for reflection, loss, and crosstalk for each eye diagram sample obtained by module 22. The dataset generation module 3 converts the time-domain waveform data generated by module 2 into two-dimensional probability histograms and classifies and stores them according to labels, ultimately forming a large-scale standard dataset. The model training module 4 uses the dataset for network learning. The CNN model construction submodule 41 first uses a CNN model to build a diagnostic network model with ResNet-18 as the core backbone. Then, the CNN model training submodule 42 performs end-to-end iterative training on the network model built in module 41. The CNN model supervision submodule 43 supervises the convergence of the model using a validation set after each training cycle of module 42 to obtain the trained network model. Finally, the intelligent diagnosis and ranking module 5 uses the trained model to process the eye diagram to be tested. The eye diagram diagnosis submodule 51 first scores and predicts the eye diagram to be tested. Then, the result output submodule 52 quantifies and ranks the deterioration factors according to the scores, thereby outputting the dominant factors of eye diagram deterioration and the basis for optimization.

[0122] It should be noted that the above functional modules can be implemented, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, they can be implemented, in whole or in part, as program instruction products. A program instruction product includes one or a set of program instructions. When the program instructions are loaded and executed on a computer, the described process or function is generated, in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The program instructions can be stored in a computer-readable and writable storage medium, or transferred from one computer's readable and writable storage medium to another.

[0123] In this embodiment, the direct coupling or communication connection between the modules can be achieved through indirect coupling or communication connection via interfaces, devices, or modules. The functional modules and sub-modules in this embodiment can dynamically reside within a single processing unit, or each module can exist physically independently, or two or more modules can dynamically reside within a single processing unit. When these dynamic components are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable and writable storage medium. This storage medium can be a memory, disk, or optical disc, etc.

[0124] The effects of this invention will be further illustrated below with simulation experiments:

[0125] 1. Simulation conditions:

[0126] The simulation experiment hardware platform is: Intel(R) Core(TM) Ultra 7 265K (3.90 GHz) processor and 32GB of memory.

[0127] The simulation experiment software platform consists of: Windows 11 operating system, Keysight ADS2024 software, and Python 3.9.0.

[0128] Typical T-type topology in simulation experiments, such as Figure 2 As shown, it includes a main control chip and multiple external DDR3 chips.

[0129] 2. Simulation content and result analysis:

[0130] Under the above simulation conditions, the convolutional network diagnostic model trained by this invention is used to simulate three test sample eye diagrams, Case 1, Case 2, and Case 3, by setting composite random parameters on the link in step one. The sampling points of each eye diagram are uniformly concentrated on a two-dimensional plane composed of time and voltage amplitude, and a 300×300 uniform grid is divided and global normalized to obtain a two-dimensional probability histogram matrix P. Then, the single-channel matrix P is copied and extended in the channel dimension to construct a final input tensor X with a dimension of 3×300×300. This tensor is then used as the input of the convolutional network diagnostic model. Through data compression and vector extraction at each layer of the model, inference is completed to finally obtain the output of the three-dimensional health score, which is compared with the actual score. The results are shown in Table 1.

[0131] Table 1 Comparison of predicted scores and actual scores from the model output

[0132]

[0133] As can be seen from Table 1:

[0134] For Case 1 sample with physical parameters R=32.6 Ω, Tan D=0.014, and S=170 mil, the health scores of reflection, loss, and crosstalk predicted by the model were 64.528, 92.307, and 84.953, respectively. Compared with the calculated actual health scores of 64.804, 92.071, and 84.310, the relative errors were 0.426%, 0.256%, and 0.763%, respectively.

[0135] For Case 2 samples with physical parameters R=147.3 Ω, Tan D=0.026, and S=233.2 mil, the model predicted health scores for reflection, loss, and crosstalk to be 89.779, 93.302, and 88.826, respectively. Compared with the calculated actual health scores of 89.167, 91.063, and 87.750, the relative errors were 0.686%, 2.459%, and 1.226%, respectively.

[0136] For Case 3 samples with physical parameters R=278.8 Ω, Tan D=0.039, and S=70.2 mil, the model predicted health rates of reflection, loss, and crosstalk to be 77.966, 88.917, and 77.200, respectively. Compared with the calculated actual health rates of 77.883, 88.819, and 76.667, the relative errors were 0.107%, 0.110%, and 0.695%, respectively.

[0137] The above results show that the predicted results and the actual calculated results of the invented method are in high agreement. The relative prediction error of each sample is generally small, with the maximum error not exceeding 2.5%, and the error of most samples is controlled within 1%. Furthermore, the ranking results are completely consistent. This demonstrates that the convolutional neural network model constructed using the method of this invention can accurately predict eye diagram health under different degradation factors and can quantify and rank the severity of these degradation factors. It can provide intelligent basis for the localization and design optimization of SI problems in high-speed links, exhibiting excellent accuracy and reliability.

Claims

1. A quantitative diagnostic method for eye diagram degradation factors based on convolutional neural networks, characterized in that, include: (1) Construct a high-speed link parameterized simulation model as a reference template; (2) Based on the deep integration of Python scripts and ADS simulation software, an ADS-Python co-simulation framework is constructed to automatically modify simulation parameters, calculate the health score of three deterioration factors for each sample, and generate large-scale eye diagram datasets in batches. (3) Based on the reference template, use the co-simulation framework to generate a dataset with three-dimensional health score labels; (4) Use the dataset to train a convolutional neural network to learn the mapping relationship between each eye diagram image and the three types of health scores: reflection, loss, and crosstalk, and obtain a trained eye diagram convolutional neural network diagnostic model. (5) Input the eye diagram to be diagnosed into the trained eye diagram convolutional neural network diagnostic model to obtain the three predicted health scores for reflection, loss and crosstalk, and realize the intelligent diagnosis and quantitative ranking results of the dominant factors of eye diagram deterioration.

2. The method according to claim 1, characterized in that, The high-speed link parameterized simulation model constructed in (1) as a reference template includes the following implementation: 1a) The three types of physical effects that cause eye diagram degradation—reflection, loss, and crosstalk—are mapped to physical parameters that can be independently adjusted in circuit design, namely, termination impedance R, dielectric loss tangent Tan D, and line spacing S. 1b) Based on the above mapping relationship, establish a simulation circuit model in ADS simulation software that includes three variable parameters: termination impedance R, dielectric loss tangent TanD, and line spacing S.

3. The method according to claim 1, characterized in that, In part (2), an ADS-Python co-simulation framework is constructed based on the deep integration of Python scripts and ADS simulation software. Its implementation includes: 2a) Establish a main control unit using Python as the scripting language, responsible for parameter parsing, dynamic reconstruction of the netlist, task scheduling, and data post-processing; 2b) Establish a computing core with ADS as the simulation engine to be responsible for executing the underlying link transient edge simulation tasks; 2c) The main control unit and the computing core interact via file stream. Specifically, Python inputs the netlist file to ADS through the terminal to drive the simulation. After ADS completes the calculation, it generates a data file and feeds it back to Python for reading and parsing.

4. The method according to claim 1, characterized in that, The process in (3) involves generating a standard dataset with three-dimensional health score labels based on a reference template using a co-simulation framework. The implementation of this dataset includes: 3a) Utilize the ADS-Python co-simulation framework to automatically scan three key parameters: termination resistance R, dielectric loss Tan D, and line spacing S, generating parameter combinations that cover both single and composite factors; 3b) Using the ideal baseline eye diagram as the anchor point, obtain the eye height H of the baseline eye diagram. ref Wide eyes W ref and level L ref Then extract the eye height H from the eye diagram to be tested. meas Wide eyes W meas Overshoot ratio OS meas and level offset |L meas -L ref |; 3c) Truncation operator via normalization The formulas calculate the following four sub-scores respectively: Amplitude integrity: , Timing integrity: , Overshoot penalty: , Level stability: ; 3d) Based on the above four sub-scores, calculate the final health score using a weighted summation. : ； in, It is a weighting factor for amplitude integrity. It is a weighting factor for temporal integrity. It is the weighting coefficient of the overshoot penalty. These are weighting coefficients for level stability. Their values ​​will be determined by the characteristics of the link, with higher weighting coefficients for more sensitive characteristics. 3e) For the three single-factor scenarios that only change R, Tan D, and S, their health scores are recorded as scoreR, scoreTanD, and scoreS, respectively. The three scores are then ranked to obtain the dominant deterioration factor and the relative severity of multiple factors. For the composite interference scenario that changes R, Tan D, and S simultaneously, the overall health score is calculated in the same way. 3f) For each scenario in each case, waveform data is obtained by modifying the netlist and calling the ADS engine to perform time-domain simulation. The data is then converted into a probability density distribution heatmap to generate an eye diagram. Scores and rankings are calculated according to preset rules. Finally, the scores and rankings are used as labels to classify and store the eye diagrams in order to construct a standard dataset for training CNN models.

5. The method according to claim 1, characterized in that, The implementation of training a convolutional neural network using a dataset in (4) includes: 4a) Divide the standard dataset into a training set and a validation set according to a preset ratio, use the corresponding three-dimensional health score as the supervision label, and expand the single-channel eye diagram probability density matrix into three-channel data after copying it in the channel dimension to meet the model input requirements. 4b) Using ResNet-18 as the core backbone, a fully connected embedding layer and a three-output linear regression layer are connected after it, corresponding to the three health scores of output reflection, loss and crosstalk, respectively, to form a dedicated convolutional neural network model. 4c) Define the mean squared error loss function for a dedicated convolutional neural network model. The sum of prediction errors for the three-dimensional health score is expressed as: ； in, This is a reflection of the health score vector. This is the health loss score vector. This is the crosstalk health score vector. The model prediction for reflection. The model predicts the loss. B represents the model prediction value for crosstalk, and B is the batch size. 4d) Input the training set into a dedicated convolutional neural network model for end-to-end iterative training: that is, by minimizing the mean squared error loss function. This enables the network to learn a nonlinear mapping relationship from eye diagram image features to a three-dimensional health score of reflection, loss, and crosstalk; 4e) After each training cycle, use the validation set to monitor the model's convergence and generalization ability to prevent overfitting; 4f) Repeat the above training and validation process until the model's loss function on the validation set is satisfactory. The convolutional neural network diagnostic model is obtained by converging to the minimum value, which is the training objective.

6. A quantitative diagnostic system for eye diagram degradation factors based on convolutional neural networks, characterized in that, include: The high-speed link parameterized simulation module is used to map the three types of signal integrity problems in high-speed links—reflection, loss, and crosstalk—to quantifiable physical parameters such as load resistance R, dielectric loss tangent Tan D, and line spacing S, respectively, as reference templates for sample training. The ADS-Python co-simulation module is used to perform simulations on specified parameter combinations, calculate three-dimensional health scores, and add corresponding labels. The dataset generation module is used to generate large-scale standard eye diagram datasets in batches as training samples for the model. The model training module is used to train a convolutional neural network diagnostic model and learn the nonlinear mapping relationship between eye diagram image features and three-dimensional health scores of reflection, loss, and crosstalk. The intelligent diagnosis and ranking module is used to output predicted health scores for three dimensions: reflection, crosstalk, and loss. Based on the scores, it automatically identifies and quantifies the dominant factors that cause the decline in eye diagram quality.

7. The system according to claim 6, characterized in that, The ADS-Python co-simulation module includes: The circuit simulation submodule is used to scan three key parameters: termination resistance R, dielectric loss Tan D, and line spacing S, generate parameter combinations covering single and composite factors, modify the netlist accordingly, call the ADS simulation engine to perform batch simulation, and obtain time-domain waveform data. The eye diagram sample generation submodule is used to convert the time-domain waveform data obtained from the simulation into a probability density distribution and draw it as a thermal eye diagram image. The 3D health label construction submodule is used to generate health score labels in three dimensions: reflection, loss, and crosstalk for each eye diagram sample.

8. The system according to claim 6, characterized in that, The model training module includes: The CNN model building submodule is used to build a convolutional neural network diagnostic model with ResNet-18 as the core backbone, followed by a fully connected embedding layer and a three-output linear regression layer. The CNN model training submodule is used to input the training set into the convolutional neural network diagnostic model for end-to-end iterative training. The CNN model supervision submodule is used to supervise the convergence and generalization ability of the model using the validation set after each training cycle.

9. The system according to claim 6, characterized in that, The intelligent diagnosis and ranking module includes: The eye diagram diagnosis submodule is used to predict the health score of the input eye diagram to be diagnosed; The results output submodule is used to quantify and sort the predicted scores, obtain the dominant factors of eye diagram deterioration, and output the results.

Images