High-speed link signal equalization method based on three-dimensional memristor type equalizer

Abstract

The invention discloses a high-speed link signal equalization method based on a three-dimensional memristor type equalizer, and the method comprises the steps: obtaining a distortion signal, generating an ideal receiving signal, and constructing a data set; training the recurrent neural network by using the data set to obtain a trained weight parameter; performing mapping processing on the weight parameters to obtain a conductance value matrix in the three-dimensional memristor type equalizer; deploying the conductance values to the three-dimensional memristor type equalizer according to the conductance value matrix; and finally, inputting the distorted signal into a three-dimensional memristor type equalizer for processing to obtain an optimized signal after distortion compensation. Compared with a traditional feedforward equalizer and decision feedback equalizer method, the method has the advantages that time domain waveform distortion is effectively reduced, high-speed signal integrity is improved, the method has wide applicability, the method can be used for memristor type equalizer training and deployment application of various complex high-speed links, the conductance value is deployed to a three-dimensional memristor cross array, and the method is suitable for training and deployment application of memristor type equalizers of various complex high-speed links. Therefore, the integration density is improved and the occupied area of the equalizer is reduced.

Description

Technical Field

[0001] The present invention relates to signal processing technology in high-speed data transmission links, and particularly to a high-speed link signal equalization method based on a three-dimensional memristive equalizer. Background Art

[0002] With the rapid development of information technology, the demand for data transmission rate is constantly increasing. Especially in the fields of data centers, high-performance computing, and 5G communications, high-speed links are increasingly widely used. However, high-speed data transmission links are vulnerable to factors such as reflection, crosstalk, and delay during high-speed transmission, resulting in signal distortion, which seriously affects the performance and reliability of the system. To solve these problems, signal equalization technology is widely applied in high-speed links to recover and reconstruct distorted signals.

[0003] Traditional signal equalization technologies mainly include feed-forward equalizers, decision feedback equalizers, and continuous-time linear equalizers, etc. These equalizers compensate signals through linear or limited non-linear methods, but their performance is often limited when facing complex non-linear distortion and long-delay dependence. Feed-forward equalizers and continuous-time linear equalizers are difficult to effectively compensate for high-order non-linear distortion due to their limited inherent non-linear representation ability; while the structure of decision feedback equalizers is vulnerable to error propagation in the feedback loop and 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, recurrent neural networks have become an emerging method for signal equalization due to their ability to remember and model sequential data. However, equalizers based on recurrent neural networks face challenges of high resource consumption and excessive power consumption in hardware implementation, which limits their application in actual high-speed links. At the same time, as a new type of non-volatile memory element, memristors have advantages such as low power consumption, high speed, small size, and good support for analog computing, and have been widely studied for the hardware implementation of artificial neural networks. However, existing artificial neural networks are difficult to effectively process changes in sequence length and time dependence due to the lack of memory function, which limits their application in real-time or online processing of continuous data streams.

[0005] Therefore, there is a need to develop a recurrent neural network equalization method more suitable for hardware deployment of memristive brain-like chips for high-speed link equalization. Summary of the Invention

[0006] Aiming at the deficiencies of existing high-speed link signal equalization technologies in non-linear compensation and long-term dependence modeling, the present invention provides a three-dimensional memristive 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 ability of the recurrent neural network and the high-density integration characteristics of the three-dimensional memristor crossbar array, effectively compensating for signal distortion in the high-speed link, and improving signal integrity and transmission reliability.

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

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

[0009] S1. Obtain the distorted signal and generate the ideal received signal, and then construct a data set;

[0010] S2. Use the data set to train the recurrent neural network to obtain the trained weight parameters;

[0011] S3. Obtain the conductance value matrix in the three-dimensional memristive equalizer through mapping processing of the weight parameters;

[0012] S4. Deploy the conductance values to the three-dimensional memristive equalizer according to the conductance value matrix;

[0013] S5. Input the distorted signal into the three-dimensional memristive equalizer for processing to obtain the optimized signal after compensating for the distortion.

[0014] The specific content of S2 is as follows:

[0015] S21. Initialize the weight parameters in the recurrent neural network, and input the data set into the recurrent neural network;

[0016] S22. Subsequently, the recurrent neural network calculates the loss function through the backpropagation algorithm and updates the weight parameters in the network. The loss function is set according to the following formula:

[0017]

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

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

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

[0021] The weight parameters in S2 include: the weight matrix W input from the input layer to the hidden layerxh The weight matrix W from the hidden layer state of the previous time step to the hidden layer state of 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 memristive equalizer is specifically a three-dimensional memristor crossbar array.

[0023] Specifically, S3 is as follows:

[0024] The following steps are performed on each weight matrix or bias vector in the weight parameters. The positive component conductance value matrix and the negative component conductance value matrix obtained by mapping all the weight matrices and bias vectors together serve as the conductance value matrix in the three-dimensional memristive equalizer:

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

[0026] S32. Calculate the scaling factor according to the maximum absolute value |W| max According to the following formula:

[0027]

[0028] where K W represents the scaling factor, G max and G min represent the maximum value and the minimum value of the adjustable conductance range of the memristor respectively, and |W| max represents the maximum absolute value among all the absolute values of the elements of 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 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;

[0030] S34. Map and generate the positive component conductance value matrix and the negative component conductance value matrix from the positive component matrix and the negative component matrix respectively according to the following formula:

[0031] G pos =G min +K W ·W pos

[0032] G neg =G min +K W ·W neg

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

[0034] Specifically, S4 is as follows:

[0035] S41. For each weight matrix or bias vector in the weight parameters, perform the following steps:

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

[0037] S42. After the conductance values of all the weight matrices and bias vectors in the weight parameters are deployed, set the conductance values of the remaining unset memristors to the minimum value of the adjustable conductance range of the memristor.

[0038] A computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the method for high-speed link signal equalization based on a three-dimensional memristive equalizer according to any one of claims 1 to 7.

[0039] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the steps of the method for high-speed link signal equalization based on a three-dimensional memristive equalizer according to any one of claims 1 to 7.

[0040] A computer program product includes a computer program / instructions, and when the computer program / instructions are executed by a processor, it implements the steps of the method for high-speed link signal equalization based on a three-dimensional memristive equalizer according to any one of claims 1 to 7.

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

[0042] First, through the recursive structure and memory unit of the recurrent neural network, the present invention can effectively capture complex non-linear distortions and long-term dependencies in the high-speed link, significantly improve the accuracy and effect of signal equalization, reduce time-domain waveform distortion, and improve high-speed signal integrity.

[0043] Second, the equalizer designed in the present invention can adapt to the changes of various complex high - speed link conditions. By adjusting the conductance values in the memristor cross - array, it ensures the equalization performance for different high - speed link signal transmissions. Description of the Drawings

[0044] Figure 1 Schematic diagram of the unfolded recurrent neural network structure along the time dimension in the embodiment of the present invention;

[0045] Figure 2 Schematic diagram of the three - dimensional memristive brain - like chip used in the embodiment of the present invention;

[0046] Figure 3 Schematic diagram of the cross - array weight deployment in the embodiment of the present invention;

[0047] Figure 4 Schematic diagram of the channel scattering parameters and equalization settings in the embodiment of the present invention;

[0048] Figure 5 Schematic diagram of different equalization methods and the time - domain waveform of the ideal receiver in the implementation of the present invention. Detailed Implementation Manner

[0049] In order to enable those skilled in the art to better understand the method of the present invention and make the above - mentioned objects, features, and advantages of the present invention more clearly understandable, the present invention will be further described in detail below with reference to the embodiments.

[0050] According to the content of the present invention, the complete implementation of the embodiment is as follows:

[0051] The task targeted in this example is to equalize the high - speed link signal through the proposed three - dimensional memristive equalizer. In this example, a recurrent neural network as shown in Figure 1 is used as the algorithm basis of the memristive neural network equalizer. This recurrent neural network contains 32 hidden units. Both the input and output dimensions are 1. After training, the weights of the recurrent neural network are mapped and deployed into the three - dimensional memristive brain - like chip as shown in Figure 2 , and the deployment method is as shown in Figure 3 . The scattering parameters of the high - speed link used are as shown in Figure 4 . Based on this data set, the proposed training and deployment method of the three - dimensional memristive equalizer applicable to high - speed links is implemented in this example, and the effectiveness of the present invention is verified by comparing with the results of traditional feed - forward equalizers and decision - feedback equalizers.

[0052] The high - speed link signal equalization method of the present invention includes the following steps:

[0053] S1. Obtain the distorted signal of the high - speed link, normalize it, generate the corresponding ideal received signal waveform, and then construct a data set;

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

[0055] For a three-dimensional memristive equalizer based on a recurrent neural network applicable to high-speed links, an effective training data set needs to be constructed first. This example uses a pseudo-random binary sequence containing 5000 bits as the input signal, with each bit sampled 16 times to generate 80,000 sampling points. The signal is first normalized to a range between 0 and 1 volt. Then, the normalized data is segmented into subsequences of 100 time steps in a time series, resulting in a total of 800 subsequences, where 20% is used for training and 80% is used for testing.

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

[0057] S2. Use the data set to train the recurrent neural network to generate a predicted output signal through the hidden layer and the linear layer, and obtain the trained weight parameters;

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

[0059] S21. Initialize the weight parameters in the recurrent neural network. The data set (including the distorted signal and the ideal received signal) is input into the recurrent neural network, and a predicted output signal is generated through the hidden layer and the linear layer;

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

[0061] For the He initialization method, the weight matrix W follows 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. Subsequently, the recurrent neural network calculates the loss function through the backpropagation 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] where L represents the loss function, y t is the predicted value generated by the recurrent neural network at time step t based on the current input and the previous state, is the target output at time step t, i.e., the ideal received signal, and t = 1 to T, where 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 ability to model long-term dependencies in the signal sequence. The specific formulas for updating its hidden state and the output layer are as follows:

[0068]

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

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

[0071]

[0072] where L t , y t and h t represent the loss function value, the output value, and the intermediate variable value at time t, respectively, 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 through multiple iterations until the loss function converges, and obtain the trained weight parameters.

[0074] S3. Obtain the conductance value matrix in the three-dimensional memristive equalizer through mapping processing of the weight parameters;

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

[0076] Perform the following steps for each weight matrix or bias vector in the weight parameters. The positive component conductance value matrix and negative component conductance value matrix obtained through mapping processing of all weight matrices and bias vectors together serve as the conductance value matrix in the three-dimensional memristive equalizer:

[0077] S31. Determine the maximum absolute value |W| among all the absolute values of the elements of the weight matrix max ;

[0078] S32. Calculate the scaling factor according to the maximum absolute value |W| max using the following formula:

[0079]

[0080] where 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 respectively, and |W| max represents the maximum absolute value among all the absolute values of the elements of the weight matrix;

[0081] S33. Decompose 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 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.

[0082] S34. Map and generate the positive component conductance value matrix and negative component conductance value matrix from the positive component matrix and negative component matrix respectively using the following formula:

[0083] G pos = G min + K W ·W pos

[0084] G neg = G min + K W ·W neg

[0085] where G pos and G neg represent the positive component conductance value matrix and negative component conductance value matrix respectively, and K WRepresents the scaling factor, G max and G min respectively represent the maximum and minimum values of the adjustable conductance range of the memristor, W pos 、W neg respectively represent the positive component matrix and the negative component matrix.

[0086] The positive component conductance value matrix and the negative component conductance value matrix are respectively the conductance value matrices representing the positive component matrix and the negative component matrix, not 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, so each element of a weight matrix corresponds to using two memristors to represent the positive component and the negative component respectively. That is, a matrix of N by N requires 2N 2 memristors for representation.

[0087] S4. Deploy the conductance values to the three - dimensional memristive equalizer according to the conductance value matrix;

[0088] S41. Perform the following steps for each weight matrix or bias vector in the weight parameters:

[0089] Use the value of each element in the i - th row of the corresponding positive component conductance value matrix of the weight matrix or bias vector as the conductance value to set the memristors corresponding to the (2i - 1)-th row of the three - dimensional memristive equalizer, and use the value of each element in the i - th row of the negative component conductance value matrix as the conductance value to set the memristors corresponding to the 2i - th row of the three - dimensional memristive equalizer, so that the rows of the positive component conductance value matrix corresponding to the weight matrix and the rows of the negative component conductance value matrix are arranged alternately;

[0090] S42. After the conductance value deployment of all the weight matrices and bias vectors in the weight parameters is completed, set the conductance values of the remaining unset memristors to 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 achieve three - dimensional integration and improve the integration density. This memristor cross - array is used to implement the weight connections from the input to the hidden layer, from the hidden layer to the hidden layer, and from the hidden layer to the output layer in the recurrent neural network.

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

[0093] Through the above steps, the trained weights of the recurrent neural network can be effectively mapped and deployed into the memristor cross - array to achieve high - efficiency signal equalization at the hardware level.

[0094] S5. Input the distortion signal of the high-speed link into a three-dimensional memristive equalizer. Utilize the conductance value adjustment characteristic of the memristor array to process and obtain an optimized signal after compensating for the distortion, thereby realizing high-speed signal equalization processing to reconstruct the distorted signal.

[0095] This method requires at least two layers of memristor cross arrays, that is, two layers of memristors and three layers of interconnects. The three-dimensional memristor cross array is multiple stacked memristor layers.

[0096] The present invention aims to solve the signal integrity problems in high-speed communication links caused by reflection, crosstalk, delay, etc. In the training stage, 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 (BPTT) algorithm to adjust the parameters of the recurrent neural network model. Further, the obtained weight matrix is mapped to the conductance values in the three-dimensional memristor cross array, and the above conductance values are deployed in the three-dimensional memristor cross array, thereby improving the integration density and reducing the area occupied by the equalizer.

[0097] To evaluate the effectiveness of the proposed method, it is also compared with traditional feed-forward equalizers and decision-feedback equalizers. The feed-forward and feedback filter coefficients are optimized and determined by commercial software Keysight ADS software, and are configured as 5th order and 4th order respectively. The evaluation results on the test set show that the proposed recurrent neural network-based 3D memristive equalizer achieves a mean square error loss of 7.96×10 -4 , indicating that it can effectively capture and compensate for the non-linear characteristics of the channel. Compared with traditional feed-forward equalizer and decision-feedback equalizer methods, the proposed method shows significant advantages in time-domain waveform recovery. As Figure 5 shown, the reconstructed signal waveform shows clearer transitions and reduced overshoots, significantly improving the signal integrity. These results indicate that the memristive 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 traditional methods. The significant improvement in performance is mainly attributed to the memory ability of the internal state of the recurrent neural network, enabling the model to better capture complex channel distortions and long-range dependencies. Generally speaking, the experimental results fully prove the effectiveness of the present invention.

[0098] The above embodiments are the implementation manners of the present invention, but the implementation manners of the present invention are not limited by the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement manners and are all included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope defined by the claims.

Claims

1. A high-speed link signal equalization method based on a three-dimensional memristive equalizer, characterized in that The method includes the following steps: S1. Obtain a distorted signal and generate an ideal received signal, and then construct a data set; S2. Use the data set to train a recurrent neural network to obtain the trained weight parameters; S3. Obtain the conductance value matrix in the three-dimensional memristive equalizer through mapping processing of the weight parameters; S4. Deploy the conductance values to the three-dimensional memristive equalizer according to the conductance value matrix; S5. Input the distorted signal into the three-dimensional memristive equalizer for processing to obtain an optimized signal with compensated distortion.

2. The high-speed link signal equalization method based on a three-dimensional memristive equalizer according to claim 1, characterized in that: The specific content of S2 is as follows: S21. Initialize the weight parameters in the recurrent neural network, and input the data set into the recurrent neural network; S22. Subsequently, the recurrent neural network calculates the loss function through 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, and y t is the predicted value generated by the recurrent neural network at time step t based on the current input and the previous state, is the target output at time step t, where t = 1 to T and T is the total number of time steps; S23. Repeat S22 until the loss function converges to obtain the trained weight parameters.

3. A high-speed link signal equalization method based on a three-dimensional memristive equalizer according to claim 1, characterized in that: In S21, the He initialization method is used to initialize the weight parameters in the recurrent neural network.

4. A high-speed link signal equalization method based on a three-dimensional memristive equalizer according to claim 1, characterized in that: The weight parameters in S2 include: the weight matrix W input to the hidden layer xh , the weight matrix W from the hidden layer state of the previous time step to the hidden layer state of 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 .

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

6. A high-speed link signal equalization method based on a three-dimensional memristive equalizer according to claim 4, characterized in that: The specific content of S3 is as follows: The following steps are performed on each weight matrix or bias vector in the weight parameters. The positive component conductance value matrix and negative component conductance value matrix obtained through mapping processing of all weight matrices and bias vectors together serve as the conductance value matrix in the three-dimensional memristive equalizer: S31. Determine the maximum absolute value |W| among all the absolute values of the elements of the weight matrix max ; S32. According to the maximum absolute value |W| max The scaling factor is calculated according to the following formula: Among them, K W represents the scaling factor, G max and G min represent the maximum value and the minimum value of the adjustable conductance range of the memristor respectively, and |W| max represents the maximum absolute value among the absolute values of all elements of 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 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; S34. Map and generate the positive component conductance value matrix and negative component conductance value matrix from the positive component matrix and negative component matrix respectively according to the following formula: G pos = G min + K W · W pos G neg = G min + K W · W neg Among them, G pos and G neg represent the positive component conductance value matrix and the negative component conductance value matrix respectively, K W represents the scaling factor, G max and G min represent the maximum value and the minimum value of the adjustable conductance range of the memristor respectively, W pos and W neg represent the positive component matrix and the negative component matrix respectively.

7. A high-speed link signal equalization method based on a three-dimensional memristive equalizer according to claim 1, characterized in that: The specific content of S4 is as follows: S41. The following steps are performed on each weight matrix or bias vector in the weight parameters: Use the value of the i-th row of the positive component conductance value matrix of the weight matrix or bias vector as the conductance value to set the (2i - 1)-th row of memristors in the three-dimensional memristive equalizer, and use the value of the i-th row of the negative component conductance value matrix as the conductance value to set the 2i-th row of memristors in the three-dimensional memristive equalizer; S42. After the conductance values of all weight matrices and bias vectors in the weight parameters are deployed, set the conductance values of the remaining un-set memristors to the minimum value of the adjustable conductance range of the memristors.

8. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for high-speed link signal equalization based on a three-dimensional memristive equalizer according to any one of claims 1 to 7.

9. 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 method for high-speed link signal equalization based on a three-dimensional memristive equalizer according to any one of claims 1 to 7.

10. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, it implements the steps of the method for high-speed link signal equalization based on a three-dimensional memristive equalizer according to any one of claims 1 to 7.

Images