A collaborative analysis method for high-speed transmitter power supply signal integrity

Abstract

The present invention belongs to the field of high-speed communication link analysis, and discloses a high-speed transmitter power signal integrity collaborative analysis method, including: step 1, collecting training data through circuit simulation; step 2, building a power signal integrity collaborative analysis model based on contrast mask learning and feature modulation; step 3, model training combining supervised learning and contrast learning; step 4, model reasoning based on efficient parallel decoding. The present invention can efficiently analyze the performance of high-speed transmitter circuits under the influence of power supply noise. By combining contrast learning and supervised learning, the robustness of feature representation is enhanced, and the prediction accuracy of noise signals is improved.

Description

Technical Field

[0001] The present invention relates to the field of high-speed communication link analysis, and in particular to a high-speed transmitter power supply signal integrity collaborative analysis method. Background Art

[0002] As emerging applications such as data centers, artificial intelligence (AI), and 5G mobile networks continue to increase their demand for high-performance computing and data transmission, modern high-performance systems have placed higher demands on efficient, low-latency transmission links. High-speed serial links play a vital role in these systems, especially as chiplet architecture and high-density heterogeneous integration technologies become increasingly popular, and the increase in data transmission rates poses new challenges to system performance.

[0003] High-speed serial links usually consist of a transmitter (TX), a transmission channel, and a receiver (RX), and their transmission rate is closely related to system performance. As one of the key components in the link, the transmitter is responsible for generating and sending high-frequency signals to carry digital information. Since the quality of the transmitter signal directly affects the link performance, any deterioration in signal quality may lead to errors, data loss, or reduced throughput. As modern chip-based systems place increasing emphasis on power and thermal efficiency, lower power supply voltages compress the noise margin and increase the transmitter's sensitivity to power supply noise (PSN). The interaction between power supply noise and the input signal further exacerbates the impact of noise on the transmitter's behavior. Therefore, ensuring the performance of the transmitter and improving signal integrity (SI) and power integrity (PI) are critical to achieving high-bandwidth, high-speed data transmission.

[0004] The existing technology has several deficiencies in the power signal integrity analysis of the transmitter in the high-speed link, mainly including the following points: 1. The high-speed transmitter has a strong nonlinear characteristic, especially the influence of parasitic effects in the high-speed link makes the nonlinear characteristics of the transmitter further aggravated, and the output signal shows long-term dependence and complex time correlation. Existing models, such as recurrent neural networks (RNN) and long short-term memory networks (LSTM), fail to fully capture this strong nonlinear behavior and cannot effectively model the complex dynamic behavior of the transmitter in the high-speed link. 2. The high sensitivity of the transmitter to dynamic power noise causes signal distortion such as timing jitter and amplitude variation. The existing methods fail to effectively simulate and deal with the impact of power noise on the performance of the transmitter, resulting in the inability to accurately predict signal distortion. 3. The power noise and the input signal are interdependent and form complex interactions through shared paths and coupling ports. Traditional models often fail to fully consider this interactive effect and ignore the combined impact of the interaction between power noise and input signals, resulting in significant errors in the prediction of the output signal.

[0005] Based on the above problems, in order to perform efficient power signal integrity collaborative analysis on transmitter circuits in high-speed communication links, the present invention proposes a high-speed transmitter power signal integrity collaborative analysis method based on contrast mask learning and feature modulation. Summary of the invention

[0006] The object of the present invention is to provide a high-speed transmitter power supply signal integrity collaborative analysis method, which can accurately capture the interaction between power supply noise and input signals and capture the highly nonlinear behavior of transmitter circuits, and realize efficient power supply signal integrity collaborative analysis of transmitter circuits to solve the problems raised in the above-mentioned background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A high-speed transmitter power supply signal integrity collaborative analysis method comprises the following steps:

[0009] Step 1, collecting training data through circuit simulation: in the circuit simulation software, input different input signals and power supply signals to the transmitter circuit, and collect the output signal of the transmitter circuit;

[0010] Step 2, construct a power signal integrity collaborative analysis model based on contrastive mask learning and feature modulation: establish a power signal integrity collaborative analysis model based on contrastive mask learning and feature modulation, the power signal integrity collaborative analysis model includes 3 sequence encoders, 1 non-sequence link encoder and 1 non-autoregressive decoder; the power signal integrity collaborative analysis model encodes input features and mask outputs, and fuses the input feature encoding to form a comprehensive input feature embedding vector, and compares the input feature embedding vector and the output embedding vector to ensure that more robust input features are learned, so as to achieve accurate prediction of the transmitter noisy output signal through the non-autoregressive decoder to achieve power signal integrity collaborative analysis;

[0011] Step 3: Model training combining supervised learning and contrastive learning: In the model training phase, contrastive learning is combined with mask data modeling in non-autoregressive decoding to perform end-to-end optimization;

[0012] Step 4, model reasoning based on efficient parallel decoding: In the reasoning stage, the power signal integrity co-analysis model generates output sequences through an efficient parallel decoding mechanism; this process adopts non-autoregressive decoding and masked data modeling strategies.

[0013] Furthermore, the step 1 comprises:

[0014] Step 1.1, provide input signal: The input signal is a binary symbol stream, represented as , where each symbol is "0" or "1"; the input data is transmitted through the transmitter circuit and represented by a series of trapezoidal waveforms. The waveform characteristics include amplitude ,cycle and the ratio of transition time to period ;Providing an input signal to an input port of a transmitter circuit;

[0015] Step 1.2, providing a power signal with power noise: using a quadrilateral waveform to generate power noise, where the power noise has a randomized sub-period and amplitude; providing the power signal with noise to a power port of the transmitter;

[0016] Step 1.3, specify link parameters: Link parameters include equalizer coefficients , the transmitter load capacitance , S parameters of transmission lines , the equivalent resistance of the receiving end and pull-up voltage ;

[0017] Step 1.4, obtain the transmitter simulation output signal: input different input signals, power supply signals with power supply noise, and specified link parameters into the transmitter circuit through circuit simulation software to obtain the transmitter output signal ;

[0018] Step 1.5, discretization of output signal: Discretization is performed, and each voltage value is mapped to the closest discrete level using the corresponding step size; this process converts the continuous voltage value of the output signal into a series of discrete categories, and ultimately each discrete voltage value represents a category.

[0019] Furthermore, the step 2 comprises:

[0020] Step 2.1, encoding input features and masking outputs: using the power signal integrity collaborative analysis model to encode input features and output targets, the input features of the power signal integrity collaborative analysis model include input signal sequence, noisy power signal sequence and link parameters, the output features of the power signal integrity collaborative analysis model are output signal sequences, feature encoding is performed on the input features to obtain input feature embedding vectors, the input feature embedding vectors include input signal embedding vectors, power signal embedding vectors and link parameter embedding vectors, masking and feature encoding are performed on the output features to obtain output signal embedding vectors;

[0021] Step 2.2, feature fusion module based on FiLM module: use FiLM module to fuse input signal embedding vector, power signal embedding vector and link parameter embedding vector to generate a unified latent feature representation;

[0022] Step 2.3, contrastive learning: Contrastive learning is used to optimize the embedding representation of the power signal integrity co-analysis model, and the similarity between the input features and the output features is improved by minimizing the contrast loss;

[0023] Step 2.4, non-autoregressive decoding based on masked data modeling: The non-autoregressive decoder adopts the standard Transformer architecture and combines masked data modeling for non-autoregressive decoding; the non-autoregressive decoder receives the context vector and the mask input, and outputs a decoding result with the same length as the mask input sequence.

[0024] Furthermore, the step 2.1 comprises:

[0025] Step 2.1.1, feature encoding of input sequence: For the input sequence, define the length as The input signal sequence , the length is Power signal sequence ; Among them, I t =(t=1,...,n) means Input signal value at the moment, P t =(t=1,...,n) means The power signal value at the moment; the sequence encoder in the power signal integrity collaborative analysis model is divided into an input signal sequence encoder and a power signal sequence encoder. The input signal sequence and power signal sequence The corresponding input signal sequence encoder and power signal sequence encoder are respectively input; the input signal sequence Processed by the input signal sequence encoder, the feature representation is obtained ; Power signal sequence Processed by power signal sequence encoder, the characteristic representation is obtained ;

[0026] Step 2.1.2, feature encoding of link parameters: Link parameters In the encoding process, the geometric features V of the input signal are combined h , t p , r t Perform joint encoding; enhance input scalar features It will be processed by a non-sequential link encoder; each scalar feature of the input is first standardized, and the mean of the feature in the training data set is subtracted from each feature value, and then the result is divided by the standard deviation of the feature in the training data set; each standardized scalar feature will be further processed by a multilayer perceptron; each feature will be encoded by a multilayer perceptron with the same structure but different parameters; the structure of the multilayer perceptron includes two hidden layers, each hidden layer contains 16 neurons; each hidden layer uses ReLU as the activation function; the output layer of the multilayer perceptron is processed using a linear activation function, and the output is converted into a dimension of The eigenvector of

[0027] Step 2.1.3, feature encoding of S parameters: For the S parameters of the circuit, first apply a small constant offset to it to make its value positive, and then perform a logarithmic transformation; the S parameter matrix of each frequency point will be divided into real and imaginary parts, and each part will be used as a 2-channel input for feature extraction through a convolutional neural network; the convolutional neural network contains two convolutional layers, the first convolutional layer uses 16 filters, the second convolutional layer uses 32 filters, and each convolutional layer uses a ReLU activation function; the output of the convolutional layer is flattened and converted to Embedding vector of dimension;

[0028] Step 2.1.4, mask processing of output signal sequence: use random mask strategy to process the output signal sequence; the length is The output signal sequence is defined as , the sequence is masked according to a certain ratio; the number of masks is between 1 and The mask position is randomly determined by uniform distribution, and the signal value of the mask position is replaced by a special marker symbol. <mask>, this symbol is also considered a classification label;

[0029] Step 2.1.5, feature encoding of the output signal sequence after masking: the output sequence after masking Input to the output signal sequence encoder for processing to obtain the masked embedded vector representation .

[0030] Furthermore, the step 2.2 comprises:

[0031] Step 2.2.1: Timing alignment and fusion of input signal embedding and power signal embedding: embed the input signal and power signal embedding Align in the time dimension; first embed the input signal into and power signal embedding Splicing is performed along the feature dimension, that is:

[0032] ,

[0033] in Indicates that and Concatenate along the feature dimension; then use a linear layer to project the concatenated result back to the model dimension , and obtain the fused feature representation ;

[0034] Step 2.2.2, use link parameter embedding to modulate the fusion features of the input signal and the power signal: Use link parameter embedding right Modulation is performed to embed circuit characteristics and adjust the adaptability of the power signal integrity collaborative analysis model to signal changes under different circuit configurations; first, a multi-layer perceptron is used to adjust the link parameter embedding To match Length , and then adjust and Passed together to the FiLM module for feature modulation; the FiLM module transforms from Generate scaling factor in and the offset coefficient , used to modulate the fusion features:

[0035] ,

[0036] in, and There are two fully connected layers, which are used to generate scaling parameters and offset parameters; scaling parameters and offset parameters are applied to fusion features The modulated feature representation is obtained by element-wise multiplication and addition. :

[0037] ,

[0038] in, Represents element-wise multiplication operation;

[0039] Step 2.2.3, further feature extraction and temporal dependency capture: In order to further extract information and capture the In the temporal dependency, a multi-head self-attention mechanism is applied:

[0040] E fused =MultiHeadAttention(E FiLM )

[0041] MultiHeadAttention represents the multi-head self-attention mechanism; finally, the fusion features processed by the self-attention mechanism It is further refined by a feed-forward neural network to generate a fused input feature vector of the circuit state :

[0042] ,

[0043] represents a feedforward neural network; the fused input feature vector The context vector is input into the non-autoregressive decoder as the context vector of the non-autoregressive decoder.

[0044] Furthermore, the step 2.3 includes:

[0045] Step 2.3.1, calculate the global representation of input features and output features: For each batch, first use average pooling to pool the input features and output features Pooling is performed to obtain global feature representation and ; Assume the batch size is , then for each sample and ,calculate:

[0046] ,

[0047] Represents the average pooling operation, through which the feature sequence of each sample is mapped into a global representation of a fixed dimension;

[0048] Step 2.3.2, calculate contrast loss: calculate the similarity of each pair of samples:

[0049] ,

[0050] in is the input feature and output features The dot product of Input features and samples The similarity between the output features of The calculation is done using the following formula:

[0051] ,

[0052] in, It is a sample The dot product of the input and output features of , indicating the similarity of the matching input-output pairs; is the temperature scaling parameter used to adjust the smoothness of the loss; It is to calculate the similarity of non-matching sample pairs, with the aim of reducing the similarity with non-matching samples.

[0053] Further, the step 2.4 comprises:

[0054] Step 2.4.1, input embedding and position encoding: The context vector received by the non-autoregressive decoder is the fused feature , the input sequence is the masked output sequence ; First of all, The sequence is embedded to convert each category into a fixed-dimensional vector representation. Subsequently, positional encoding is added to preserve the position information of the sequence. The positional encoding is generated using sine and cosine functions to ensure that the power signal integrity co-analysis model can identify the relative relationship between each position in the sequence:

[0055] ,

[0056] in is the embedding dimension; is the index of the embedding dimension, indicating the different dimensions of the positional encoding; is the index of the current position in the sequence;

[0057] Step 2.4.2, multi-head self-attention mechanism: The masked input sequence embedding X is transformed linearly to generate query, key and value vectors:

[0058] ,

[0059] in , and They are the linear transformation matrices of the query vector Q, key vector K, and value vector V respectively; the dot product of the query vector and the key vector will be used as the attention score, and the attention weight is calculated by applying the softmax function after the scaling factor. :

[0060] ,

[0061] in Represents the key vector The transpose of Represents the key vector The output of the attention function of dimension By querying the vector With key vector Transpose Dot product, and divide by the scaling factor After that, the score is calculated; the score is converted into weight through the softmax function and combined with the value vector Multiply them together to generate a weighted output, which is the result of the cross-attention calculation;

[0062] The multi-head attention mechanism calculates multiple different attention heads independently and finally merges the output results of each head together:

[0063]

[0064] in ,

[0065] in Indicated in The linear transformation matrix of the query vector Q, key vector K, and value vector V used in the attention head; represents the output of each attention head, It means that after concatenating the output vectors of all attention heads, a linear transformation matrix is ​​used. The information of these attention heads is fused to finally obtain the output of the multi-head attention mechanism. ; After residual connection and layer normalization, the output of the multi-head attention layer is finally obtained;

[0066] Step 2.4.3, Multi-head Cross-Attention Mechanism: Query Vector Output from the non-autoregressive decoder , and the key vector Sum value vector It comes from the fused input feature vector as the context vector :

[0067] ,

[0068] The linear transformation matrix of the query vector Q, key vector K and value vector V in the multi-head cross attention mechanism is The input vector and context vector Mapped to query vectors , key vector Sum value vector ; Query vector and key vector After the dot product of is adjusted by the scaling factor, the attention weight is calculated using the softmax function:

[0069] ,

[0070] Cross-attention also uses a multi-head structure, each head performs calculations independently, and finally merges the results:

[0071]

[0072] in ,

[0073] Each of the heads Through different linear transformation matrices , , The query vector , the key vector Sum value vector Mapped to their respective processing spaces; after the calculation is completed, the output of each attention head is merged through the splicing operation and linearly transformed Merge to generate the final cross-attention output; finally, the output is residually connected and layer normalized;

[0074] Step 2.4.4, Bitwise Feedforward Neural Network: The output of the attention layer will be passed into a bitwise feedforward neural network, which consists of two linear layers and a nonlinear activation function ReLU, to further process the feature representation obtained from the self-attention layer to increase the expressiveness of the power signal integrity collaborative analysis model;

[0075] Step 2.4.5, linear output: Finally, the linear layer generates the same sequence as the input The output of this linear layer is a probability distribution at each position for generating outputs that match the target sequence.

[0076] Furthermore, the step 3 comprises:

[0077] Step 3.1, Parallel decoding of masked output sequence: The non-autoregressive decoder uses the masked output sequence As input, fusion features are used as context information; with this input, the non-autoregressive decoder generates the complete output sequence ,in is the predetermined output sequence length; after the non-autoregressive decoder completes decoding, each position of the output sequence The values ​​of are updated;

[0078] Step 3.2, calculate the cross entropy loss: The cross entropy loss function is used to measure the difference between the output sequence predicted by the power signal integrity co-analysis model and the true label; the goal of the power signal integrity co-analysis model is to learn how to predict the missing output, and the cross entropy loss is only calculated for the mask position:

[0079] ,

[0080] in, represents the cross entropy loss, represents a set of mask positions, Indicates that given an input mask sequence and contextual features Next, location Output at The predicted probability of

[0081] Step 3.3, loss function calculation: The total loss function is composed of the cross entropy loss and contrastive learning loss in mask data modeling; by adjusting an adjustable weight parameter , combining cross entropy loss and contrastive learning loss for optimization; the total loss function The calculation formula is:

[0082] ,

[0083] In the formula, is the contrastive learning loss;

[0084] Step 3.4, model optimization: Use the stochastic gradient descent variant Adam method to optimize the total loss function; optimize the model parameters through the back propagation algorithm to make the prediction of the mask position more accurate, and at the same time optimize the model through contrast learning to model the relationship between input and output.

[0085] Furthermore, the step 4 comprises:

[0086] Step 4.1, initialize the input sequence and context vector: At the beginning of inference, the non-autoregressive decoder receives a fully masked input sequence, that is, , where all output positions are replaced by special <mask>Mark; the input sequence of the mask is used as the initial input of the non-autoregressive decoder; in addition, the input features are encoded by their respective encoders and fused using the feature fusion module based on the FiLM module to obtain the fused input feature vector As a context vector for a non-autoregressive decoder;

[0087] Step 4.2, non-autoregressive decoding: based on the masked input sequence and context vector , the non-autoregressive decoder generates the complete output sequence in parallel.

[0088] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention provides a power signal integrity collaborative analysis method based on contrast mask decoder and FiLM feature modulation, which can efficiently analyze the performance of high-speed transmitter circuits under the influence of power supply noise. By combining contrastive learning and supervised learning, the robustness of feature representation is enhanced and the prediction accuracy of noise signals is improved. Mask data modeling supports structured learning of long sequence signals, while the non-autoregressive decoder enables the reasoning process to be carried out efficiently in parallel, significantly improving the decoding efficiency. The feature fusion module of the present invention effectively captures the interaction between signals and circuits and provides a comprehensive circuit state representation. Compared with traditional circuit simulation software, the present invention shows higher efficiency in power signal integrity and transmission signal analysis, and can quickly and accurately perform power signal integrity collaborative analysis of high-speed transmitters. The method of the present invention greatly improves the efficiency of circuit design, provides designers with an efficient and reliable tool, supports circuit design optimization, and improves the performance and reliability of high-speed serial links. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 It is a flow chart of the present invention.

[0090] Figure 2 It is a schematic diagram of the overall structure of the power signal integrity collaborative analysis model of the present invention.

[0091] Figure 3 It is a structural schematic diagram of the feature fusion module based on the FiLM module of the present invention.

[0092] Figure 4 It is a comparison curve diagram of the result of the present invention when predicting a single output signal and the simulation result of traditional simulation software.

[0093] Figure 5 It is a comparison diagram between the eye diagram prediction result of the present invention and the simulation result of traditional simulation software. DETAILED DESCRIPTION

[0094] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0095] like Figure 1-Figure 5 As shown, the specific implementation steps of the high-speed transmitter power signal integrity collaborative analysis method based on contrast mask learning and feature modulation of the present invention are:

[0096] Step 1: Collect training data through circuit simulation: In the circuit simulation software, input different input signals and power supply signals to the transmitter circuit, and collect the output signal of the transmitter circuit. The circuit simulation software can use HSPICE, etc. The specific steps are as follows:

[0097] Step 1.1, provide input signal: The input signal is a binary symbol stream, represented as , where each symbol "0" or "1"; the input data is transmitted through the transmitter circuit and represented by a series of trapezoidal waveforms. The waveform characteristics include amplitude ,cycle and the ratio of transition time to period ; Provides input signal to the input port of the transmitter circuit.

[0098] Step 1.2, provide a power signal with power noise: Generate power noise using a quadrilateral waveform with randomized sub-cycles and amplitudes (sub-cycle durations are 0.05 to 0.2 times , the maximum amplitude fluctuation can reach The noisy power signal is supplied to the power port of the transmitter.

[0099] Step 1.3, specify link parameters: Link parameters include equalizer coefficients , the transmitter load capacitance , S parameters of transmission lines , the equivalent resistance of the receiving end and pull-up voltage These parameters affect the transmission characteristics of the signal in the circuit, which in turn affects the output signal of the transmitter.

[0100] Step 1.4, obtain the transmitter simulation output signal: input different input signals, power supply signals with power supply noise, and specified link parameters into the transmitter circuit through circuit simulation software to obtain the transmitter output signal .

[0101] Step 1.5, discretization of output signal: Discretization is performed, and each voltage value is mapped to the closest discrete level using the corresponding step size; this process converts the continuous voltage value of the output signal into a series of discrete categories, and ultimately each discrete voltage value represents a category.

[0102] Step 2: Construct a power signal integrity collaborative analysis model based on contrast mask learning and feature modulation: Figure 2 As shown in the figure, a power signal integrity collaborative analysis model based on contrastive mask learning and feature modulation is established, which includes 3 sequence encoders, 1 non-sequence link encoder and 1 non-autoregressive decoder. The power signal integrity collaborative analysis model encodes the input features and mask outputs, and fuses the input feature encoding to form a comprehensive input feature embedding vector. The input feature embedding vector and the output embedding vector are contrastively learned to ensure that more robust input features are learned, so that accurate prediction of the transmitter noisy output signal can be achieved through the non-autoregressive decoder to achieve power signal integrity collaborative analysis. The specific steps are as follows:

[0103] Step 2.1, encoding input features and masking output: using the power signal integrity collaborative analysis model to encode input features and output targets, the input features of the power signal integrity collaborative analysis model include input signal sequence, noisy power signal sequence and link parameters, the output features of the power signal integrity collaborative analysis model are output signal sequences, feature encoding is performed on the input features to obtain input feature embedding vectors, the input feature embedding vectors include input signal embedding vectors, power signal embedding vectors and link parameter embedding vectors, masking and feature encoding are performed on the output features to obtain output signal embedding vectors. The specific implementation is as follows:

[0104] Step 2.1.1, feature encoding of input sequence: For the input sequence, define the length as The input signal sequence , the length is Power signal sequence ; Among them, I t =(t=1,...,n) means Input signal value at the moment, P t =(t=1,...,n) means The power signal value at the moment; the sequence encoder in the power signal integrity collaborative analysis model is divided into an input signal sequence encoder and a power signal sequence encoder. The input signal sequence and power signal sequence The corresponding input signal sequence encoder and power signal sequence encoder are respectively input; the input signal sequence Processed by the input signal sequence encoder, the feature representation is obtained ; Power signal sequence Processed by power signal sequence encoder, the characteristic representation is obtained .

[0105] Step 2.1.2, feature encoding of link parameters: In order to enhance the generalization ability of the model, link parameters In the encoding process, the geometric features V of the input signal are combined h , t p , r t Perform joint encoding; enhance input scalar features It will be processed by a non-sequential link encoder; each scalar feature of the input is first standardized, and the mean of the feature in the training dataset is subtracted from each feature value, and then the result is divided by the standard deviation of the feature in the training dataset; each standardized scalar feature will be further processed by a multi-layer perceptron (MLP). Specifically, each feature will be encoded by a multi-layer perceptron with the same structure but different parameters; the structure of the multi-layer perceptron includes two hidden layers, each hidden layer contains 16 neurons; each hidden layer uses ReLU as the activation function; the output layer of the multi-layer perceptron is processed using a linear activation function, and the output is converted into a dimension of The feature vector of .

[0106] Step 2.1.3, feature encoding of S parameters: For the S parameters of the circuit, first apply a small constant offset to it to make its value positive, and then perform a logarithmic transformation to improve numerical stability; the S parameter matrix of each frequency point will be divided into real and imaginary parts, and each part will be used as a 2-channel input for feature extraction through a convolutional neural network (CNN); the convolutional neural network contains two convolutional layers, the first convolutional layer uses 16 filters, the second convolutional layer uses 32 filters, and each convolutional layer uses a ReLU activation function; the output of the convolutional layer is flattened and converted to through a fully connected layer Embedding vector of dimension .

[0107] Step 2.1.4, mask processing of output signal sequence: use random mask strategy to process the output signal sequence; the length is The output signal sequence is defined as , the sequence is masked according to a certain ratio; the number of masks is between 1 and The mask position is randomly determined by uniform distribution, and the signal value of the mask position is replaced by a special marker symbol. <mask>, this symbol is also considered a classification label.

[0108] Step 2.1.5, feature encoding of the output signal sequence after masking: the output sequence after masking Input to the output signal sequence encoder for processing to obtain the masked embedded vector representation .

[0109] Step 2.2, feature fusion module based on FiLM module: Figure 3 As shown, a feature fusion module based on the FiLM module is used to fuse the input signal embedding vector, the power signal embedding, and the link parameter embedding vector to generate a unified latent feature representation. Among them, the FiLM module (Feature-wise Linear Modulation) is a neural network module, which is mainly used to dynamically adjust the activation function of each layer in the network according to the input features, thereby improving the model's sensitivity to input differences‌‌12. The FiLM module adjusts the output of each layer through linear transformation, so that the model can better capture subtle differences in the input data, thereby improving the accuracy and generalization ability of the model. The specific steps are as follows:

[0110] Step 2.2.1: Timing alignment and fusion of input signal embedding and power signal embedding: embed the input signal and power signal embedding Alignment is performed in the time dimension. Since the input signal and the power signal interact with each other and jointly affect the output signal, the input signal is first embedded in and power signal embedding Splicing is performed along the feature dimension, that is:

[0111] ,

[0112] in Indicates that and Concatenate along the feature dimension; then use a linear layer to project the concatenated result back to the model dimension , and obtain the fused feature representation .

[0113] Step 2.2.2, use link parameter embedding to modulate the fusion features of the input signal and the power signal: Use link parameter embedding right Modulation is performed to embed circuit characteristics and adjust the adaptability of the power signal integrity collaborative analysis model to signal changes under different circuit configurations. First, a multi-layer perceptron is used to adjust the link parameter embedding To match Length , and then adjust and Passed together to the FiLM module for feature modulation; the FiLM module transforms from Generate scaling factor in and the offset coefficient , used to modulate the fusion features:

[0114] ,

[0115] in, and There are two fully connected layers, which are used to generate scaling parameters and offset parameters; scaling parameters and offset parameters are applied to fusion features The modulated feature representation is obtained by element-wise multiplication and addition. :

[0116] ,

[0117] in, Represents an element-wise multiplication operation.

[0118] Step 2.2.3, further feature extraction and temporal dependency capture: In order to further extract information and capture the In the temporal dependency, a multi-head self-attention mechanism is applied:

[0119] E fused =MultiHeadAttention(E FiLM )

[0120] MultiHeadAttention represents a multi-head self-attention mechanism. This mechanism can effectively capture the interactive information of the signal across time steps. Finally, the fusion features processed by the self-attention mechanism are It is further refined by a feed-forward neural network to generate a fused input feature vector of the circuit state :

[0121] ,

[0122] represents a feed-forward neural network.

[0123] This feature indicates It contains complex interactive information between input signals, power signals, and link parameters, and can provide rich feature representation for subsequent analysis. The context vector is input into the non-autoregressive decoder as the context vector of the non-autoregressive decoder.

[0124] Step 2.3, contrastive learning: Use contrastive learning to optimize the embedded representation of the power signal integrity collaborative analysis model, and improve the similarity between input features and output features by minimizing the contrast loss. The specific steps are as follows:

[0125] Step 2.3.1, calculate the global representation of input features and output features: For each batch, first use average pooling to pool the input features and output features Pooling is performed to obtain global feature representation and Specifically, assuming the batch size is , then for each sample and ,calculate:

[0126] ,

[0127] Represents the average pooling operation. Through the pooling operation, the feature sequence of each sample is mapped to a global representation of a fixed dimension, which helps to capture the overall characteristics of the input and output.

[0128] Step 2.3.2, calculate contrast loss: The purpose of contrast loss is to minimize the similarity between the same input-output pairs and maximize the similarity between different input-output pairs. Specifically, calculate the similarity of each pair of samples:

[0129] ,

[0130] in is the input feature and output features The dot product of Input features and samples The similarity between the output features of . Contrastive loss function The calculation is done using the following formula:

[0131] ,

[0132] in, It is a sample The dot product of the input and output features of , indicating the similarity of the matching input-output pairs; is a temperature scaling parameter used to adjust the smoothness of the loss and is usually set to a small value to control the sensitivity of the model to similarity; It is to calculate the similarity of non-matching sample pairs, with the aim of reducing the similarity with non-matching samples.

[0133] Step 2.4, non-autoregressive decoding based on masked data modeling: The non-autoregressive decoder uses the standard Transformer architecture and combines masked data modeling for non-autoregressive decoding. The non-autoregressive decoder receives the context vector and mask input, and outputs a decoding result with the same length as the mask input sequence. The specific steps are as follows:

[0134] Step 2.4.1, input embedding and position encoding: The context vector received by the non-autoregressive decoder is the fused feature , the input sequence is the masked output sequence ; First of all, The sequence is embedded to convert each category into a fixed-dimensional vector representation. Subsequently, positional encoding is added to preserve the position information of the sequence. The positional encoding is generated using sine and cosine functions to ensure that the power signal integrity co-analysis model can identify the relative relationship between each position in the sequence:

[0135] ,

[0136] in is the embedding dimension, usually equal to the model dimension; is the index of the embedding dimension, indicating the different dimensions of the positional encoding; is the index of the current position in the sequence.

[0137] Step 2.4.2, Multi-head Self-Attention Mechanism: The masked input sequence embedding X is transformed linearly to generate query (Q), key (K) and value (V) vectors:

[0138] ,

[0139] in , and They are the linear transformation matrices of the query vector Q, key vector K, and value vector V respectively; the dot product of the query vector and the key vector will be used as the attention score, and the attention weight is calculated by applying the softmax function after the scaling factor. :

[0140] ,

[0141] in Represents the key vector The transpose of Represents the key vector The output of the attention function of dimension By querying the vector With key vector Transpose Dot product, and divide by the scaling factor After that, the score is calculated; the score is converted into weight through the softmax function and combined with the value vector Multiply them together to generate a weighted output, which is the result of the cross-attention calculation.

[0142] The multi-head attention mechanism calculates multiple different attention heads independently and finally merges the output results of each head together:

[0143]

[0144] in ,

[0145] in Indicated in The linear transformation matrix of the query vector Q, key vector K, and value vector V used in the attention head; represents the output of each attention head, It means that after concatenating the output vectors of all attention heads, a linear transformation matrix is ​​used. The information of these attention heads is fused to finally obtain the output of the multi-head attention mechanism. ; After residual connection and layer normalization, the output of the multi-head attention layer is finally obtained.

[0146] Step 2.4.3, Multi-head Cross-Attention Mechanism: Query Vector Output from the non-autoregressive decoder , and the key vector Sum value vector It comes from the fused input feature vector as the context vector :

[0147] ,

[0148] The linear transformation matrix of the query vector Q, key vector K and value vector V in the multi-head cross attention mechanism is The input vector and context vector Mapped to query vectors , key vector Sum value vector ; Query vector and key vector After the dot product of is adjusted by the scaling factor, the attention weight is calculated using the softmax function:

[0149] ,

[0150] Cross-attention also uses a multi-head structure, each head performs calculations independently, and finally merges the results:

[0151]

[0152] in ,

[0153] Each of the heads Through different linear transformation matrices , , The query vector , the key vector Sum value vector Mapped to their respective processing spaces; after the calculation is completed, the output of each attention head is merged through the splicing operation and linearly transformed Merge to generate the final cross-attention output; finally, the output is residually connected and layer normalized.

[0154] Step 2.4.4, bitwise feedforward neural network (FNN): The output of the attention layer will be passed into an FNN, which consists of two linear layers and a nonlinear activation function ReLU, to further process the feature representation obtained from the self-attention layer to increase the expressiveness of the power signal integrity collaborative analysis model.

[0155] Step 2.4.5, linear output: Finally, the linear layer generates the same sequence as the input The output of this linear layer is a probability distribution at each position for generating outputs that match the target sequence.

[0156] Step 3: Model training combining supervised learning and contrastive learning: In the model training phase, we combined contrastive learning with mask data modeling in non-autoregressive decoding to perform end-to-end optimization. The specific steps are as follows:

[0157] Step 3.1, Parallel decoding of masked output sequence: The non-autoregressive decoder uses the masked output sequence As input, fusion features are used as context information; with this input, the non-autoregressive decoder generates the complete output sequence ,in is the predetermined output sequence length. For the specific process, refer to step 2.4. After the non-autoregressive decoder completes decoding, each position of the output sequence The values ​​are updated.

[0158] Step 3.2, calculate the cross entropy loss: The cross entropy loss function is used to measure the difference between the output sequence predicted by the power signal integrity co-analysis model and the true label. The goal of the power signal integrity co-analysis model is to learn how to predict the missing output, and the cross entropy loss is only calculated for the mask position:

[0159] ,

[0160] in, represents the cross entropy loss, represents a set of mask positions, Indicates that given an input mask sequence and contextual features Next, location Output at The predicted probability of .

[0161] Step 3.3, loss function calculation: The total loss function is composed of the cross entropy loss and contrastive learning loss in mask data modeling. By adjusting an adjustable weight parameter , the cross entropy loss and contrastive learning loss are combined for optimization. The total loss function The calculation formula is:

[0162] ,

[0163] In the formula, is the contrastive learning loss.

[0164] Step 3.4, model optimization: Use the stochastic gradient descent variant Adam method to optimize the total loss function. Through the back propagation algorithm, optimize the model parameters to make the prediction of the mask position more accurate, and at the same time optimize the model to model the relationship between input and output through contrast learning.

[0165] Step 4, model reasoning based on efficient parallel decoding: In the reasoning stage, the power signal integrity collaborative analysis model generates an output sequence through an efficient parallel decoding mechanism. This process still uses non-autoregressive decoding and mask data modeling strategies, which can effectively improve the reasoning speed and ensure high-quality output. The specific steps are as follows:

[0166] Step 4.1, initialize the input sequence and context vector: At the beginning of inference, the model receives a fully masked input sequence, that is, , where all output positions are replaced by special <mask>Mark. This masked input sequence is used as the initial input of the non-autoregressive decoder. In addition, the input features are encoded by their respective encoders and fused using a feature fusion module based on the FiLM module to obtain a context vector optimized by contrastive learning. .

[0167] Step 4.2, non-autoregressive decoding: based on the masked input sequence and context vector , the non-autoregressive decoder generates the complete output sequence in parallel.

[0168] In order to verify the effectiveness of the method of the present invention in the collaborative analysis of power signal integrity, we conducted tests under multiple groups of different signal conditions. The pulse period of the experimental signal is 80-150ps, and the signal conversion time ratio is 5%-20%. The experimental results show that the method of the present invention is significantly better than the traditional SPICE simulation software in efficiency and has high accuracy. For single link prediction, the average absolute error of the method of the present invention is less than 3.3mV, the relative error is between 0.28% and 0.5%, and 99% of the relative errors are less than 2.7%. Figure 4 The comparison between the predicted output signal affected by a single power supply noise and the SPICE simulation results is shown, indicating that the method of the present invention can highly match the SPICE simulation results. In terms of simulation efficiency, the average reasoning time of the method of the present invention is 79 to 108 times higher than that of SPICE, showing extremely high processing efficiency. For the prediction of the eye diagram, by randomly generating 1k groups of 4-bit signals, under 10 combinations of input and link parameters, the average relative errors of the prediction results of the method of the present invention in amplitude, eye height and eye width are 0.59% to 2.24%, 0.10% to 0.30% and 0.24% to 1.65%, respectively. Compared with traditional SPICE simulation, the method of the present invention not only achieves significantly high accuracy in eye diagram prediction accuracy, but also provides 100 to 134 times acceleration in simulation efficiency. Figure 5 The comparison between the eye diagram generated by the method of the present invention and the eye diagram generated by SPICE simulation is shown, and a high degree of consistency is shown. The method of the present invention can significantly reduce the consumption of computing resources while maintaining the prediction accuracy, realize efficient collaborative analysis of power signal integrity, and significantly improve the simulation efficiency in the circuit design process.

[0169] In summary, the high-speed transmitter power signal integrity collaborative analysis method based on contrast mask learning and feature modulation proposed in the present invention has significant accuracy and efficiency. This method can effectively support the analysis of high-speed transmission links, improve data transmission quality, reduce bit error rate, meet the needs of modern high-performance electronic systems for efficient and accurate design, promote the development of electronic design automation tools, and promote the design and optimization of high-performance electronic systems and interconnection links, and promote technological progress in the field of high-speed communications.

[0170] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.< / mask> < / mask> < / mask> < / mask>

Claims

1. A high-speed transmitter power supply signal integrity collaborative analysis method, characterized in that: The steps include: Step 1, collecting training data through circuit simulation: in the circuit simulation software, input different input signals and power supply signals to the transmitter circuit, and collect the output signal of the transmitter circuit; Step 2, construct a power signal integrity collaborative analysis model based on contrastive mask learning and feature modulation: establish a power signal integrity collaborative analysis model based on contrastive mask learning and feature modulation, the power signal integrity collaborative analysis model includes 3 sequence encoders, 1 non-sequence link encoder and 1 non-autoregressive decoder; the power signal integrity collaborative analysis model encodes input features and mask outputs, and fuses the input feature encoding to form a comprehensive input feature embedding vector, and compares the input feature embedding vector and the output embedding vector to ensure that more robust input features are learned, so as to achieve accurate prediction of the transmitter noisy output signal through the non-autoregressive decoder to achieve power signal integrity collaborative analysis; include: Step 2.1, encoding input features and masking outputs: using the power signal integrity collaborative analysis model to encode input features and output targets, the input features of the power signal integrity collaborative analysis model include input signal sequence, noisy power signal sequence and link parameters, the output features of the power signal integrity collaborative analysis model are output signal sequences, feature encoding is performed on the input features to obtain input feature embedding vectors, the input feature embedding vectors include input signal embedding vectors, power signal embedding vectors and link parameter embedding vectors, masking and feature encoding are performed on the output features to obtain output signal embedding vectors; Step 2.2, feature fusion module based on FiLM module: use FiLM module to fuse input signal embedding vector, power signal embedding vector and link parameter embedding vector to generate a unified latent feature representation; Step 2.3, contrastive learning: Contrastive learning is used to optimize the embedding representation of the power signal integrity co-analysis model, and the similarity between the input features and the output features is improved by minimizing the contrast loss; Step 2.4, non-autoregressive decoding based on mask data modeling: The non-autoregressive decoder uses the standard Transformer architecture and combines mask data modeling for non-autoregressive decoding; the non-autoregressive decoder receives the context vector and the mask input, and outputs a decoding result with the same length as the mask input sequence; Step 3: Model training combining supervised learning and contrastive learning: In the model training phase, contrastive learning is combined with mask data modeling in non-autoregressive decoding to perform end-to-end optimization; Step 4, model reasoning based on efficient parallel decoding: In the reasoning stage, the power signal integrity co-analysis model generates output sequences through an efficient parallel decoding mechanism; this process adopts non-autoregressive decoding and masked data modeling strategies.

2. A high-speed transmitter power supply signal integrity collaborative analysis method according to claim 1, characterized in that: The step 1 comprises: Step 1.1, provide input signal: The input signal is a binary symbol stream, represented as , where each symbol "0" or "1"; the input data is transmitted through the transmitter circuit and represented by a series of trapezoidal waveforms. The waveform characteristics include amplitude ,cycle and the ratio of transition time to period ;Providing an input signal to an input port of a transmitter circuit; Step 1.2, providing a power signal with power noise: generating power noise with a quadrilateral waveform, where the power noise has a randomized sub-period and amplitude; providing the power signal with noise to a power port of the transmitter; Step 1.3, specify link parameters: Link parameters include equalizer coefficients , the transmitter load capacitance , S parameters of transmission lines , the equivalent resistance of the receiving end and pull-up voltage ; Step 1.4, obtain the transmitter simulation output signal: input different input signals, power supply signals with power supply noise, and specified link parameters into the transmitter circuit through circuit simulation software to obtain the transmitter output signal ; Step 1.5, discretization of output signal: Discretization is performed, and each voltage value is mapped to the closest discrete level using the corresponding step size; this process converts the continuous voltage value of the output signal into a series of discrete categories, and ultimately each discrete voltage value represents a category.

3. A high-speed transmitter power supply signal integrity collaborative analysis method according to claim 2, characterized in that: The step 2.1 comprises: Step 2.1.1, feature encoding of input sequence: For the input sequence, define the length as The input signal sequence , the length is Power signal sequence ; Among them, I t =(t=1,...,n) means Input signal value at the moment, P t =(t=1,...,n) means The power signal value at the moment; the sequence encoder in the power signal integrity collaborative analysis model is divided into an input signal sequence encoder and a power signal sequence encoder. The input signal sequence and power signal sequence The corresponding input signal sequence encoder and power signal sequence encoder are respectively input; the input signal sequence Processed by the input signal sequence encoder, the feature representation is obtained ; Power signal sequence Processed by power signal sequence encoder, the characteristic representation is obtained ; Step 2.1.2, feature encoding of link parameters: Link parameters In the encoding process, the geometric features V of the input signal are combined h , t p , r t Perform joint encoding; enhance input scalar features It will be processed by a non-sequential link encoder; each scalar feature of the input is first standardized, and the mean of the feature in the training data set is subtracted from each feature value, and then the result is divided by the standard deviation of the feature in the training data set; each standardized scalar feature will be further processed by a multilayer perceptron; each feature will be encoded by a multilayer perceptron with the same structure but different parameters; the structure of the multilayer perceptron includes two hidden layers, each hidden layer contains 16 neurons; each hidden layer uses ReLU as the activation function; the output layer of the multilayer perceptron is processed using a linear activation function, and the output is converted into a dimension of The eigenvector of Step 2.1.3, feature encoding of S parameters: For the S parameters of the circuit, first apply a small constant offset to it to make its value positive, and then perform a logarithmic transformation; the S parameter matrix of each frequency point will be divided into real and imaginary parts, and each part will be used as a 2-channel input for feature extraction through a convolutional neural network; the convolutional neural network contains two convolutional layers, the first convolutional layer uses 16 filters, the second convolutional layer uses 32 filters, and each convolutional layer uses a ReLU activation function; the output of the convolutional layer is flattened and converted to Embedding vector of dimension; Step 2.1.4, mask processing of output signal sequence: use random mask strategy to process the output signal sequence; the length is The output signal sequence is defined as , the sequence is masked according to a certain ratio; the number of masks is between 1 and The mask position is randomly determined by uniform distribution, and the signal value of the mask position is replaced by a special marker symbol. <mask>, this symbol is also considered a classification label;< / mask> Step 2.1.5, feature encoding of the masked output signal sequence: the masked output sequence Input to the output signal sequence encoder for processing to obtain the masked embedded vector representation .

4. A high-speed transmitter power supply signal integrity collaborative analysis method according to claim 3, characterized in that: The step 2.2 comprises: Step 2.2.1: Timing alignment and fusion of input signal embedding and power signal embedding: embed the input signal and power signal embedding Align in the time dimension; first embed the input signal into and power signal embedding Splicing is performed along the feature dimension, that is: ， in Indicates that and Concatenate along the feature dimension; then use a linear layer to project the concatenated result back to the model dimension , and obtain the fused feature representation ; Step 2.2.2, use link parameter embedding to modulate the fusion features of the input signal and the power signal: Use link parameter embedding right Modulation is performed to embed circuit characteristics and adjust the adaptability of the power signal integrity collaborative analysis model to signal changes under different circuit configurations; first, a multi-layer perceptron is used to adjust the link parameter embedding To match Length , and then adjust and Passed together to the FiLM module for feature modulation; the FiLM module transforms from Generate scaling factor in and the offset coefficient , used to modulate the fusion features: ， in, and There are two fully connected layers, which are used to generate scaling parameters and offset parameters; the scaling parameters and offset parameters are applied to the fusion features. The modulated feature representation is obtained by element-wise multiplication and addition. : ， in, Represents element-wise multiplication operation; Step 2.2.3, further feature extraction and temporal dependency capture: In order to further extract information and capture In the temporal dependency, a multi-head self-attention mechanism is applied: E fused =MultiHeadAttention(E FiLM ) MultiHeadAttention represents the multi-head self-attention mechanism; finally, the fusion features processed by the self-attention mechanism It is further refined by a feed-forward neural network to generate a fused input feature vector of the circuit state : ， represents a feedforward neural network; the fused input feature vector The context vector is input into the non-autoregressive decoder as the context vector of the non-autoregressive decoder.

5. A high-speed transmitter power supply signal integrity collaborative analysis method according to claim 4, characterized in that: The step 2.3 includes: Step 2.3.1, calculate the global representation of input features and output features: For each batch, first use average pooling to pool the input features and output features Pooling is performed to obtain global feature representation and ; Assume the batch size is , then for each sample and ,calculate: ， Represents the average pooling operation, through which the feature sequence of each sample is mapped into a global representation of a fixed dimension; Step 2.3.2, calculate contrast loss: calculate the similarity of each pair of samples: ， in is the input feature and output features The dot product of Input features and samples The similarity between the output features of The calculation is done using the following formula: ， in, It is a sample The dot product of the input and output features of , indicating the similarity of the matching input-output pairs; is the temperature scaling parameter used to adjust the smoothness of the loss; It is to calculate the similarity of non-matching sample pairs, with the aim of reducing the similarity with non-matching samples.

6. A high-speed transmitter power supply signal integrity collaborative analysis method according to claim 5, characterized in that: The step 2.4 comprises: Step 2.4.1, input embedding and position encoding: The context vector received by the non-autoregressive decoder is the fused feature , the input sequence is the masked output sequence ; First of all, The sequence is embedded to convert each category into a fixed-dimensional vector representation. Subsequently, positional encoding is added to preserve the position information of the sequence. The positional encoding is generated using sine and cosine functions to ensure that the power signal integrity co-analysis model can identify the relative relationship between each position in the sequence: ， in is the embedding dimension; is the index of the embedding dimension, indicating the different dimensions of the positional encoding; is the index of the current position in the sequence; Step 2.4.2, multi-head self-attention mechanism: The masked input sequence embedding X is transformed linearly to generate query, key and value vectors: ， in , and They are the linear transformation matrices of the query vector Q, key vector K, and value vector V respectively; the dot product of the query vector and the key vector will be used as the attention score, and the attention weight is calculated by applying the softmax function after the scaling factor. : ， in Represents the key vector The transpose of Represents the key vector The output of the attention function of dimension By querying the vector With key vector Transpose Dot product, and divide by the scaling factor After that, the score is calculated; the score is converted into weight through the softmax function and combined with the value vector Multiply them together to generate a weighted output, which is the result of the cross-attention calculation; The multi-head attention mechanism calculates multiple different attention heads independently and finally merges the output results of each head together: in , in Indicated in The linear transformation matrix of the query vector Q, key vector K, and value vector V used in the attention head; represents the output of each attention head, It means that after concatenating the output vectors of all attention heads, a linear transformation matrix is ​​used. The information of these attention heads is fused to finally obtain the output of the multi-head attention mechanism. ; After residual connection and layer normalization, the output of the multi-head attention layer is finally obtained; Step 2.4.3, Multi-head Cross-Attention Mechanism: Query Vector Output from the non-autoregressive decoder , and the key vector Sum value vector It comes from the fused input feature vector as the context vector : ， The linear transformation matrix of the query vector Q, key vector K and value vector V in the multi-head cross attention mechanism is The input vector and context vector Mapped to query vectors , key vector Sum value vector ; Query vector and key vector After the dot product of is adjusted by the scaling factor, the attention weight is calculated using the softmax function: ， Cross-attention also uses a multi-head structure, each head performs calculations independently, and finally merges the results: in , Each of the heads Through different linear transformation matrices , , The query vector , the key vector Sum value vector Mapped to their respective processing spaces; after the calculation is completed, the output of each attention head is merged through the splicing operation and linearly transformed Merge to generate the final cross-attention output; finally, the output is residually connected and layer normalized; Step 2.4.4, Bitwise Feedforward Neural Network: The output of the attention layer will be passed into a bitwise feedforward neural network, which consists of two linear layers and a nonlinear activation function ReLU, to further process the feature representation obtained from the self-attention layer to increase the expressiveness of the power signal integrity collaborative analysis model; Step 2.4.5, linear output: Finally, the linear layer generates the same sequence as the input The output of this linear layer is a probability distribution at each position for generating outputs that match the target sequence.

7. The high-speed transmitter power supply signal integrity collaborative analysis method according to claim 1, characterized in that: The step 3 comprises: Step 3.1, Parallel decoding of masked output sequence: The non-autoregressive decoder uses the masked output sequence As input, fusion features are used as context information; with this input, the non-autoregressive decoder generates the complete output sequence ,in is the predetermined output sequence length; after the non-autoregressive decoder completes decoding, each position of the output sequence The values ​​of are updated; Step 3.2, calculate the cross entropy loss: The cross entropy loss function is used to measure the difference between the output sequence predicted by the power signal integrity co-analysis model and the true label; the goal of the power signal integrity co-analysis model is to learn how to predict the missing output, and the cross entropy loss is only calculated for the mask position: ， in, represents the cross entropy loss, represents a set of mask positions, Indicates that given an input mask sequence and contextual features Next, location Output at The predicted probability of Step 3.3, loss function calculation: The total loss function is composed of the cross entropy loss and contrastive learning loss in mask data modeling; by adjusting an adjustable weight parameter , combining cross entropy loss and contrastive learning loss for optimization; the total loss function The calculation formula is: ， In the formula, is the contrastive learning loss; Step 3.4, model optimization: Use the stochastic gradient descent variant Adam method to optimize the total loss function; optimize the model parameters through the back propagation algorithm to make the prediction of the mask position more accurate, and at the same time optimize the model through contrast learning to model the relationship between input and output.

8. The high-speed transmitter power supply signal integrity collaborative analysis method according to claim 1, characterized in that: The step 4 comprises: Step 4.1, initialize the input sequence and context vector: At the beginning of inference, the non-autoregressive decoder receives a fully masked input sequence, that is, , where all output positions are replaced by special <mask>Mark; the input sequence of the mask is used as the initial input of the non-autoregressive decoder; in addition, the input features are encoded by their respective encoders and fused using the feature fusion module based on the FiLM module to obtain the fused input feature vector As a context vector for a non-autoregressive decoder;< / mask> Step 4.2, non-autoregressive decoding: based on the masked input sequence and context vector , the non-autoregressive decoder generates the complete output sequence in parallel.

Images