A high-speed link signal equalization method based on a three-dimensional memristor equalizer

Abstract

The application discloses a high-speed link signal equalization method based on a three-dimensional memristor type equalizer, which comprises the following steps: obtaining a distorted signal and generating an ideal received signal, and then constructing a data set; training a recurrent neural network by using the data set to obtain a trained weight parameter; obtaining a conductance value matrix in the three-dimensional memristor type equalizer by mapping the weight parameter; deploying the conductance value to the three-dimensional memristor type equalizer according to the conductance value matrix; and finally inputting the distorted signal into the three-dimensional memristor type equalizer to obtain an optimized signal after compensation distortion. Compared with the traditional feedforward equalizer and decision feedback equalizer method, the application effectively reduces the time domain waveform distortion, improves the high-speed signal integrity, has wide applicability, can be used for the training and deployment application of the memristor type equalizer of various complex high-speed links, and deploys the above-mentioned conductance value to the three-dimensional memristor cross array, so as to improve the integrated density and reduce the area occupied by the equalizer.

Description

TECHNICAL FIELD

[0001] The present application relates to signal processing technology in high-speed data transmission link, and particularly relates to a high-speed link signal equalization method based on a three-dimensional memristor type equalizer. BACKGROUND

[0002] With the rapid development of information technology, the demand for data transmission rate is rising, especially in the fields of data center, high-performance computing and 5G communication, the application of high-speed link is increasingly widespread. However, high-speed data transmission link is easily affected by reflection, crosstalk and time delay and other factors during high-speed transmission, resulting in signal distortion, which seriously affects the performance and reliability of the system. In order to solve these problems, signal equalization technology is widely used in high-speed link to recover and reconstruct distorted signals.

[0003] Traditional signal equalization technology mainly includes feedforward equalizer, decision feedback equalizer and continuous time linear equalizer. These equalizers compensate the signal through linear or limited nonlinear method, but their performance is often limited when facing complex nonlinear distortion and long time delay dependence. Feedforward equalizer and continuous time linear equalizer have limited nonlinear representation ability due to their inherent nonlinearity, and it is difficult to effectively compensate for high-order nonlinear distortion; while the structure of decision feedback equalizer is easily affected by error propagation in the feedback loop, and it performs poorly when processing high complexity signals.

[0004] In recent years, neural network technology has shown excellent performance in the field of signal processing, especially the recurrent neural network with its memory and modeling ability for sequence data has become a new method for signal equalization. However, the equalizer based on recurrent neural network faces the challenges of high resource consumption and high power consumption in hardware implementation, which limits its application in actual high-speed link. At the same time, as a new type of non-volatile storage element, memristor has the advantages of low power consumption, high speed, small size and good support for analog calculation, and has been widely studied for hardware implementation of artificial neural network. However, due to the lack of memory function, the existing artificial neural network is difficult to effectively process the change of sequence length and time dependence, which limits its application in real-time or online processing of continuous data stream.

[0005] Therefore, it is necessary to develop a recurrent neural network equalization method based on memristor type brain-like chip hardware deployment more suitable for high-speed link equalization. SUMMARY

[0006] In view of the deficiencies of the existing high-speed link signal equalization technology in non-linear compensation and long-time dependent modeling, the application provides a three-dimensional memristor type equalizer based on a recurrent neural network, aiming to significantly improve the performance and efficiency of high-speed link signal equalization by combining the memory capacity of the recurrent neural network and the high-density integration characteristics of the three-dimensional memristor cross array, effectively compensating for signal distortion in the high-speed link, and improving signal integrity and transmission reliability.

[0007] The technical method of the application is as follows:

[0008] The high-speed link signal equalization method comprises the following steps:

[0009] S1, obtaining a distorted signal and generating an ideal received signal, and then constructing a data set;

[0010] S2, training the recurrent neural network using the data set to obtain trained weight parameters;

[0011] S3, obtaining a conductance value matrix in the three-dimensional memristor type equalizer through mapping processing of the weight parameters;

[0012] S4, deploying the conductance value to the three-dimensional memristor type equalizer according to the conductance value matrix;

[0013] S5, inputting the distorted signal into the three-dimensional memristor type equalizer to obtain an optimized signal after compensating for distortion.

[0014] The S2 is specifically:

[0015] S21, initializing the weight parameters in the recurrent neural network, and inputting the data set into the recurrent neural network;

[0016] S22, then the recurrent neural network calculates a loss function through a back propagation algorithm and updates the weight parameters in the network, and the loss function is set according to the following formula:

[0017]

[0018] Wherein, L represents the loss function, y t is a prediction value generated by the recurrent neural network at time step t according to the current input and the previous state, is a target output at time step t, t=1 to T, T is the total number of time steps;

[0019] S23, repeating S22 until the loss function converges, to obtain the trained weight parameters.

[0020] The He initialization method is used to initialize the weight parameters in the recurrent neural network in the S21.

[0021] The weight parameters in the S2 include a weight matrix Wxh The weight matrix W from the hidden state at the previous time step to the hidden state at the current time step. hh The weight matrix W from the hidden layer to the output layer hy The bias vector b of the hidden layer h and the bias vector b of the output layer y .

[0022] The three-dimensional memristor equalizer is specifically a three-dimensional memristor cross array.

[0023] Specifically, S3 is:

[0024] For each weight matrix or bias vector in the weight parameters, the following steps are performed: the positive and negative component conductance matrices obtained by mapping all weight matrices and bias vectors are used together as the conductance matrix in the three-dimensional memristor equalizer:

[0025] S31. Determine the maximum absolute value |W| among the absolute values ​​of all elements in the weight matrix. max ;

[0026] S32. Based on the maximum absolute value |W| max The scaling factor is calculated using the following formula:

[0027]

[0028] Among them, K W G represents the scaling factor. max and G min These represent the maximum and minimum values ​​of the adjustable conductance range of the memristor, respectively, |W| max This represents the maximum absolute value among all the absolute values ​​of the elements in the weight matrix;

[0029] S33. Decompose the weight matrix into a positive component matrix and a negative component matrix. The positive component matrix retains the positive values ​​in the weight matrix and sets the non-positive values ​​to 0. The negative component matrix retains the non-positive values ​​in the weight matrix and sets the positive values ​​to 0.

[0030] S34. Generate positive component conductivity matrices and negative component conductivity matrices by mapping the positive component matrices and negative component matrices respectively according to the following formulas:

[0031] G pos =G min +K W ·W pos

[0032] G neg =G min +K W ·W neg

[0033] wherein G pos , G neg respectively represent positive component conductance value matrix and negative component conductance value matrix, K W represents scaling factor, G max and G min respectively represent maximum value and minimum value of adjustable conductance range of memristor, W pos , W neg respectively represent positive component matrix and negative component matrix.

[0034] The S4 specifically is:

[0035] S41, the following steps are performed on each weight matrix or bias vector in the weight parameter:

[0036] the value of the i-th row of the positive component conductance value matrix of the weight matrix or bias vector is set as the conductance value of the (2i-1)-th row memristor of the three-dimensional memristor equalizer, and the value of the i-th row of the negative component conductance value matrix is set as the conductance value of the 2i-th row memristor of the three-dimensional memristor equalizer;

[0037] S42, after the conductance value deployment of all weight matrices and bias vectors in the weight parameter is completed, the remaining unset memristor conductance values are set to the minimum value of the adjustable conductance range of the memristor.

[0038] A computer device comprising a memory and a processor, the memory storing a computer program, and the processor implementing the steps of the high-speed link signal equalization method based on the three-dimensional memristor equalizer of any one of claims 1 to 7 when executing the computer program.

[0039] A computer readable storage medium having a computer program stored thereon, the computer program implementing the steps of the high-speed link signal equalization method based on the three-dimensional memristor equalizer of any one of claims 1 to 7 when executed by a processor.

[0040] A computer program product comprising computer programs / instructions, the computer programs / instructions implementing the steps of the high-speed link signal equalization method based on the three-dimensional memristor equalizer of any one of claims 1 to 7 when executed by a processor.

[0041] Compared with the traditional brain-like chip training method, the method has the following beneficial effects:

[0042] Firstly, the present application can effectively capture the complex nonlinear distortion and long-term dependence relationship in the high-speed link through the recursive structure and memory unit of the recurrent neural network, significantly improve the precision and effect of signal equalization, reduce the time domain waveform distortion and improve the high-speed signal integrity.

[0043] Secondly, the equalizer designed in the application can adapt to the changes of various complex high-speed link conditions, and ensure the equalization performance of different high-speed link signal transmission by adjusting the conductance value in the cross array of the memristor. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 A schematic diagram of the recurrent neural network structure in the embodiment of the application is unfolded along the time dimension;

[0045] Figure 2 A schematic diagram of the three-dimensional memristor type brain chip used in the embodiment of the application is shown;

[0046] Figure 3 A schematic diagram of the cross array weight deployment in the embodiment of the application is shown;

[0047] Figure 4 A schematic diagram of the channel scattering parameters and equalization settings in the embodiment of the application is shown;

[0048] Figure 5 A schematic diagram of different equalization methods and ideal receiver time domain waveforms in the embodiment of the application is shown. DETAILED DESCRIPTION

[0049] In order to make the personnel in the art better understand the method of the application, and make the above-mentioned purposes, features and advantages of the application more clear and easy to understand, the application will be further described in detail below in combination with embodiments.

[0050] According to the application content of the application, the complete embodiment is as follows:

[0051] The task of the present example is to equalize the high-speed link signal through the proposed three-dimensional memristor equalizer. In the present example, a recurrent neural network as shown in Figure 1 is used as the algorithm basis of the memristor neural network equalizer, which contains 32 hidden units. The input and output dimensions are both 1. The trained recurrent neural network weights are mapped and deployed to the three-dimensional memristor type brain chip as shown in Figure 2 The deployment method is shown in Figure 3 . The scattering parameters of the high-speed link used are shown in Figure 4 . The present example implements the proposed three-dimensional memristor type equalizer training and deployment method suitable for high-speed link based on this data set, and verifies the effectiveness of the application by comparing the results of the traditional feedforward equalizer and the decision feedback equalizer.

[0052] The high-speed link signal equalization method of the application comprises the following steps:

[0053] S1, obtaining the distorted signal of the high-speed link for normalization and generating the corresponding ideal received signal waveform, and then constructing a data set;

[0054] The ideal received signal is the original signal that should be received after the transmission link without any interference, distortion or noise. Specifically, the distorted signal of the high-speed link is taken as the input of the recurrent neural network, and the corresponding ideal received signal is taken as the output of the recurrent neural network, and then a data set is constructed;

[0055] For the three-dimensional memristor equalizer suitable for high-speed links based on recurrent neural networks, an effective training data set needs to be constructed first. In this example, a pseudo-random binary sequence containing 5000 bits is used as the input signal, and each bit is sampled 16 times to generate 80,000 sampling points. The signal is first normalized to have a range of 0 to 1 volt. Then, the normalized data is divided into 100 time step subsequences in time sequence, a total of 800 subsequences, of which 20% is used for training and 80% is used for testing.

[0056] The high-speed link is a high-speed communication link for data transmission.

[0057] S2, training the recurrent neural network with the data set to generate a predicted output signal through the hidden layer and the linear layer, and obtaining the trained weight parameters;

[0058] The weight parameters include: a weight matrix W xh input to the hidden layer, a weight matrix W hh from the hidden layer of the previous time step to the hidden layer of the current time step, a weight matrix W hy from the hidden layer to the output layer, a bias vector b h of the hidden layer, and a bias vector b y of the output layer.

[0059] S21, initializing the weight parameters in the recurrent neural network, and inputting the data set (containing the distorted signal and the ideal received signal) into the recurrent neural network to generate a predicted output signal through the hidden layer and the linear layer.

[0060] The He initialization method is used to initialize the weight parameters in the recurrent neural network in S21.

[0061] The weight matrix W of the He initialization method obeys a normal distribution with a mean of 0 and a variance of

[0062]

[0063] where d in is the input dimension of the current layer.

[0064] S22, the recurrent neural network calculates a loss function by a back propagation algorithm and updates the weight parameters in the network to minimize the error between the predicted output and the target signal, the loss function is set according to the following formula:

[0065]

[0066] Wherein, L represents the loss function, y t is the predicted value generated by the recurrent neural network according to the current input and the previous state at time step t, is the target output at time step t, i.e. the ideal received signal, t = 1 to T, T is the total number of time steps;

[0067] The recurrent neural network is used to retain the state information of the previous time step to enhance the modeling ability of long-term dependencies in the signal sequence, and the hidden state update formula and the calculation formula of the output layer are as follows:

[0068]

[0069] Wherein, h t represents the intermediate state variable at time step t, W xh , W hh , W hy are respectively the weight matrix from the input layer to the hidden layer at the current time step, the weight matrix from the hidden layer at the previous time step to the hidden layer at the current time step, and the weight matrix from the hidden layer at the current time step to the output layer, b h , b y are respectively the bias vector of the hidden layer and the bias vector of the output layer, x t and y t are respectively the input and output at time t, and tanh(·) represents the hyperbolic tangent activation function.

[0070] For the weight matrix W in the recurrent neural network, the gradient is calculated according to the following formula:

[0071]

[0072] Wherein, L t , y t and h t are respectively the loss function value, output value and intermediate variable value at time t, represents the partial derivative. Through the above gradient calculation, the gradient descent method can be used to update the weight parameters of the model.

[0073] S23, repeat S22, i.e. update the weight parameters for multiple iterations until the loss function converges, and obtain the trained weight parameters.

[0074] S3, obtaining the conductance value matrix in the three-dimensional memristor equalizer through mapping processing of the weight parameter;

[0075] As shown in Figure 3 , the three-dimensional memristor equalizer is specifically a three-dimensional memristor cross array.

[0076] The following steps are performed on each weight matrix or bias vector in the weight parameter, and the positive component conductance value matrix and the negative component conductance value matrix obtained through mapping processing of all weight matrices and bias vectors are collectively used as the conductance value matrix in the three-dimensional memristor equalizer:

[0077] S31, determining the maximum absolute value |W| max in the absolute values of all elements of the weight matrix;

[0078] S32, calculating the scaling factor according to the maximum absolute value |W| max according to the following formula:

[0079]

[0080] wherein, K W represents the scaling factor, G max and G min represent the maximum and minimum values of the adjustable conductance range of the memristor, |W| max represents the maximum absolute value in the absolute values of all elements of the weight matrix;

[0081] S33, decomposing the weight matrix into a positive component matrix W pos =max(W,0) and a negative component matrix W neg =min(W,0); the positive component matrix retains the positive values in the weight matrix, and the non-positive values are set to 0; the negative component matrix retains the non-positive values in the weight matrix, and the positive values are set to 0.

[0082] S34, mapping the positive component matrix and the negative component matrix to generate a positive component conductance value matrix and a negative component conductance value matrix according to the following formula respectively:

[0083] G pos =G min +K W ·W pos

[0084] G neg =G min +K W ·W neg

[0085] wherein, G pos and G neg represent the positive component conductance value matrix and the negative component conductance value matrix respectively, and K Wdenotes a scaling factor, G max and G min denote the maximum and minimum of the adjustable conductance range of the memristor, respectively, W pos , W neg denote the positive component matrix and the negative component matrix, respectively.

[0086] The positive component conductance value matrix and the negative component conductance value matrix are conductance value matrices representing the positive component matrix and the negative component matrix, respectively, and do not refer to the positive and negative of the conductance value. Since the elements of the weight matrix have positive and negative, and in physics, the conductance value is always non-negative, the elements of a weight matrix are correspondingly represented by two memristors to represent the positive component and the negative component. That is, an N by N matrix needs 2N 2 memristors to represent.

[0087] S4, deploying the conductance values to the three-dimensional memristor equalizer according to the conductance value matrix;

[0088] S41, performing the following steps for each weight matrix or bias vector in the weight parameter:

[0089] setting the corresponding memristor in the 2i-1th row of the three-dimensional memristor equalizer with the value of each element in the ith row of the corresponding positive component conductance value matrix of the weight matrix or bias vector as the conductance value, and setting the corresponding memristor in the 2ith row of the three-dimensional memristor equalizer with the value of each element in the ith row of the negative component conductance value matrix as the conductance value, so that the rows of the positive component conductance value matrix and the rows of the negative component conductance value matrix corresponding to the weight matrix are arranged alternately;

[0090] S42, after the conductance values of all weight matrices and bias vectors in the weight parameter are deployed, setting the conductance values of the remaining memristors that are not set as the minimum value of the adjustable conductance range of the memristor.

[0091] The weight parameters of different network layers of the recurrent neural network are set on different memristor layers to realize three-dimensional integration and improve the integration density. The memristor cross array is used to realize the weight connection of the input to the hidden layer, the hidden layer to the hidden layer, and the hidden layer to the output layer in the recurrent neural network.

[0092] In this embodiment, W hy , b y are located on the same memristor layer, W xh , W hh and b h are located on another memristor layer.

[0093] Through the above steps, the trained weight of the recurrent neural network can be effectively mapped and deployed to the memristor cross array, realizing efficient signal equalization at the hardware level.

[0094] S5, input the distortion signal of the high-speed link into the three-dimensional memristor type equalizer, the conductance value of the memristor array is used to adjust and process to obtain an optimized signal after compensating for distortion, so as to realize high-speed signal equalization processing to reconstruct the distortion signal.

[0095] The method needs at least two layers of memristor cross arrays, that is, two layers of memristor three-layer interconnection lines, and a three-dimensional memristor cross array, that is, a plurality of stacked memristor layers.

[0096] The present application aims to solve the signal integrity problem caused by reflection, crosstalk and delay in high-speed communication link. In the training phase, a recurrent neural network model for signal equalization is constructed, and the gradient of the loss function with respect to the model parameters is calculated by the backpropagation through time algorithm to adjust the recurrent neural network model parameters; further, the obtained weight matrix is mapped to the conductance value in the three-dimensional memristor cross array, and the above conductance value is deployed to the three-dimensional memristor cross array, so as to improve the integrated density and reduce the area occupied by the equalizer.

[0097] In order to evaluate the effectiveness of the proposed method, it is also compared with the traditional feedforward equalizer and decision feedback equalizer, and the feedforward and feedback filter coefficients are determined by the commercial software Keysight ADS software, which are configured as 5th order and 4th order respectively. The evaluation results on the test set show that the proposed 3D memristor type equalizer based on recurrent neural network realizes the mean square error loss of 7.96x10 -4 , which indicates that it can effectively capture and compensate the nonlinear characteristics of the channel. Compared with the traditional feedforward equalizer and decision feedback equalizer method, the proposed method has a significant advantage in time domain waveform recovery, as shown in Figure 5 The reconstructed signal waveform shows clearer transition and reduced overshoot, which significantly improves the integrity of the signal. These results show that the memristor type equalizer based on recurrent neural network not only has a significant improvement in signal-to-noise ratio, but also performs excellently in timing stability, far exceeding the traditional method. The significant improvement in performance is mainly due to the memory ability of the internal state of the recurrent neural network, which enables the model to better capture complex channel distortion and long-range dependence. In summary, the experimental results fully prove the effectiveness of the present application.

[0098] The above embodiments are embodiments of the present application, but the embodiments of the present application are not limited by the above embodiments, and any changes, modifications, substitutions, combinations made without departing from the spirit and principles of the present application. Simplification should be an equivalent replacement method, which is included in the protection scope of the present application. Therefore, the protection scope of the present application should be limited by the protection scope defined in the claims.

Claims

1. A high-speed link signal equalization method based on a three-dimensional memristor equalizer, characterized in that, The method includes the following steps: S1. Acquire the distorted signal and generate the ideal received signal, and then construct the dataset; S2. Use the dataset to train the recurrent neural network to obtain the trained weight parameters; S3. Obtain the conductivity matrix in the three-dimensional memristor equalizer by mapping the weight parameters. S4. Deploy the conductance values ​​to the three-dimensional memristor equalizer according to the conductance matrix; S5. Input the distorted signal into a three-dimensional memristor equalizer to obtain the optimized signal after distortion compensation; Specifically, S2 is: S21. Initialize the weight parameters in the recurrent neural network and input the dataset into the recurrent neural network; S22. Subsequently, the recurrent neural network calculates the loss function using the backpropagation algorithm and updates the weight parameters in the network. The loss function is set according to the following formula: Where L represents the loss function, In time step The recurrent neural network generates predictions based on the current input and the previous state. In time step The target output is defined as follows: t=1 to T, where T is the total number of time steps. S23. Repeat S22 until the loss function converges, and obtain the trained weight parameters.

2. The high-speed link signal equalization method based on a three-dimensional memristor equalizer according to claim 1, characterized in that: The weight parameters in the recurrent neural network are initialized using the He initialization method in S21.

3. The high-speed link signal equalization method based on a three-dimensional memristor equalizer according to claim 1, characterized in that: The weight parameters in S2 include: the weight matrix input to the hidden layer. The weight matrix from the hidden state at the previous time step to the hidden state at the current time step. Weight matrix from hidden layer to output layer Bias vector of hidden layer and the bias vector of the output layer .

4. The high-speed link signal equalization method based on a three-dimensional memristor equalizer according to claim 1, characterized in that: The three-dimensional memristor equalizer is specifically a three-dimensional memristor cross array.

5. The high-speed link signal equalization method based on a three-dimensional memristor equalizer according to claim 3, characterized in that: Specifically, S3 is: For each weight matrix or bias vector in the weight parameters, the following steps are performed: the positive and negative component conductance matrices obtained by mapping all weight matrices and bias vectors are used together as the conductance matrix in the three-dimensional memristor equalizer: S31. Determine the maximum absolute value among the absolute values ​​of all elements in the weight matrix. ; S32, Based on the maximum absolute value The scaling factor is calculated using the following formula: in, Indicates the scaling factor. and These represent the maximum and minimum values ​​of the adjustable conductance range of the memristor, respectively. This represents the maximum absolute value among all the absolute values ​​of the elements in the weight matrix; S33. Decompose the weight matrix into a positive component matrix and a negative component matrix. The positive component matrix retains the positive values ​​in the weight matrix and sets the non-positive values ​​to 0. The negative component matrix retains the non-positive values ​​in the weight matrix and sets the positive values ​​to 0. S34. Generate positive component conductivity matrices and negative component conductivity matrices by mapping the positive component matrices and negative component matrices respectively according to the following formulas: Among them, G pos G neg These represent the positive component conductivity matrix and the negative component conductivity matrix, respectively. Indicates the scaling factor. W represents the minimum value of the adjustable conductance range of the memristor. pos W neg These represent the positive component matrix and the negative component matrix, respectively.

6. The high-speed link signal equalization method based on a three-dimensional memristor equalizer according to claim 1, characterized in that: Specifically, S4 is: S41. Perform the following steps for each weight matrix or bias vector in the weight parameters: The values ​​of the i-th row of the positive component conductance matrix of the weight matrix or bias vector are used as the conductance values ​​to set the 2i-1 row of the memristors in the three-dimensional memristor equalizer, and the values ​​of the i-th row of the negative component conductance matrix are used as the conductance values ​​to set the 2i row of the memristors in the three-dimensional memristor equalizer. S42. After the conductance values ​​of all weight matrices and bias vectors in the weight parameters are deployed, set the remaining unset memristor conductance values ​​to the minimum value of the adjustable conductance range of the memristor.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the high-speed link signal equalization method based on a three-dimensional memristor equalizer as described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the high-speed link signal equalization method based on a three-dimensional memristor equalizer as described in any one of claims 1 to 6.

9. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the high-speed link signal equalization method based on a three-dimensional memristor equalizer as described in any one of claims 1 to 6.

Images