Method and apparatus for signal equalization in high-speed communication systems

Abstract

The present application provides a method and apparatus for signal equalization in a high-speed communication system, wherein the method for signal equalization in a high-speed communication system in a first device comprises: sending a preamble code to a second device; receiving neural network setting information sent by the second device, wherein the neural network setting information comprises time step information and hidden layer size information; and obtaining neural network parameter information by training a gliding recurrent neural network according to the neural network setting information, wherein the gliding recurrent neural network comprises an input layer, at least one hidden layer and an output layer, and during the training process, the input layer converts a serial input sequence into a gliding input sequence according to the time step information, and the at least one hidden layer performs equalization training based on the gliding input sequence. The present application can obtain better performance through new SRNN-based equalization training, and due to its low complexity, it is easy to be better implemented in hardware such as FPGA, ASIC, SoC, etc.

Description

Technical Field

[0001] The present application relates to the field of communication technology, and in particular to a technical solution for signal equalization of a high-speed communication system based on a sliding recurrent neural network (SLIDING-RNN, SRNN). Background Art

[0002] In recent decades, with the rapid development of high-speed wireless services such as 5G and higher, wireless transmission of uncompressed Ultra High Definition Television (UHDTV), real-time holograms, etc., the demand for data rates has grown exponentially. In order to meet the bandwidth requirements for providing higher data rates, adaptive and flexible signal equalization has also become one of the core technologies in high-speed communication systems. For example, Passive Optical Network (PON) is recognized as the most promising broadband access technology. At the latest ITU-T G.HSP (High Speed ​​PON) standardization meeting, adaptive and flexible signal equalization has become one of the core technologies for the upcoming 25G-PON aftermarket.

[0003] In recent years, as an emerging technology, nonlinear equalizer based on neural network (NN) is regarded as a powerful equalization tool that can compensate for transmission impairments in the field of optical communications, including linear and nonlinear distortions. Fully connected neural network (FCNN) is the most widely used one, which is used as an equalizer in many works and shows better performance than some conventional algorithms. Compared with traditional equalization, the performance advantage of FCNN-based nonlinearity can be attributed to the nonlinear activation functions (AF), the depth of NN structure and the characteristics of input data.

[0004] Taking PON as an example, Figure 1 It is a flowchart of equalization based on FCNN in the prior art. Figure 2 FIG. 1 is a schematic diagram of a training process of an FCNN-based equalization in the prior art. The most common FCNN consists of an input layer with many data features, multiple hidden layers with multiple regular neurons, and an output layer. For the sake of clarity, Figure 1 and Figure 2In the example, FCNN including hidden layer 1 (FC1) and hidden layer 2 (FC2) is taken as an example. After receiving the preamble from OLT (Optical Line Terminal), ONU (Optical Network Unit) measures ISI (Inter-Symbol Interference) and sets NN to (Ni, h1, h2), where Ni, h1, and h2 represent the input data features, the size of FC1, and the size of FC2, respectively. FC1 includes normal neurons (N1, N2, N3), and FC2 includes normal neurons (N5, N6). Figure 1 and Figure 2 ,The key steps of FCNN-based equalization in the prior art can be described as follows: 1) Serial to parallel conversion (many parallel features) by 1 symbol shift at the input layer: First, the received serial signal must be converted from serial to many parallel input features, e.g., the upcoming sequence (X t-2 ,X t-1 ,X t ,X t+1 ,X t+2 ,…) can be converted to: Input 1(X t-2 ,X t-1 ,…), input 2(X t-1 ,X t ,…), input 3(X t ,X t+1 ,…) and input 4(X t+1 ,X t+2 ,…),.... 2) Then, each converted input is fully connected to FC1. 3) Next, the parallel input features of the input layer are firstly connected to the weight matrix W Ni,h1 After multiplication, it is passed to the FC1 layer (including AF), and then passed through the weight matrix W h1,h2 After multiplication, it is passed to the FC2 layer (including AF) and finally combined with the weight matrix W h2,1 After multiplication, it is output in the output layer O1 (including AF), where Figure 2 It can be seen that due to the serial-to-parallel approach, there are multiple inputs at the same time (such as Figure 2 Although FCNN-based equalization is an effective method to solve the nonlinear distortion in HSP, however, traditional FCNN cannot learn channel states and requires a large number of parallel input features, resulting in complex ASIC hardware design. Summary of the invention

[0005] The purpose of this application is to provide a new mechanism for signal equalization in high-speed communication systems. Specifically, a method and device for signal equalization in high-speed communication systems are provided. This application can achieve signal equalization in high-speed communication systems (such as high-speed PON, 5G and higher versions of wireless communication systems) through a new sliding-based recurrent neural network.

[0006] According to an embodiment of the present application, a method for signal equalization in a high-speed communication system is provided, which is applied to a first device, wherein the method includes:

[0007] Sending a preamble to a second device;

[0008] Receiving neural network setting information sent by the second device, wherein the neural network setting information includes time step information and hidden layer size information;

[0009] According to the neural network setting information, the neural network parameter information is obtained by using a gliding recurrent neural network training, wherein the gliding recurrent neural network includes an input layer, at least one hidden layer and an output layer, and during the training process, the input layer converts a serial input sequence into a gliding input sequence according to the time step information, and the at least one hidden layer performs balanced training based on the gliding input sequence.

[0010] According to another embodiment of the present application, a method for signal equalization in a high-speed communication system is provided, wherein the method includes:

[0011] receiving a preamble from a first device;

[0012] Measuring inter-symbol interference according to the preamble code, and determining neural network setting information according to the inter-symbol interference, wherein the neural network setting information includes time step information and hidden layer size information;

[0013] The neural network setting information is sent to the first device, so that the first device obtains neural network parameter information by using gliding recurrent neural network training according to the neural network setting information.

[0014] According to another embodiment of the present application, a first device for signal equalization in a high-speed communication system is provided, wherein the first device includes:

[0015] means for sending a preamble to a second device;

[0016] means for receiving neural network setting information sent by the second device, wherein the neural network setting information includes time step information and hidden layer size information;

[0017] A device for obtaining neural network parameter information by using a gliding recurrent neural network training according to the neural network setting information, wherein the gliding recurrent neural network includes an input layer, at least one hidden layer and an output layer, during the training process, the input layer converts a serial input sequence into a gliding input sequence according to the time step information, and the at least one hidden layer performs balanced training based on the gliding input sequence.

[0018] According to another embodiment of the present application, a second device for signal equalization in a high-speed communication system is provided, wherein the second device includes:

[0019] means for receiving a preamble from a first device;

[0020] A device for measuring inter-symbol interference according to the preamble code and determining neural network setting information according to the inter-symbol interference, wherein the neural network setting information includes time step information and hidden layer size information;

[0021] A device for sending the neural network setting information to the first device so that the first device obtains neural network parameter information by using gliding recurrent neural network training according to the neural network setting information.

[0022] According to another embodiment of the present application, a method for signal equalization in a high-speed communication system is also provided, wherein the method includes:

[0023] The first device sends a preamble code to the second device;

[0024] The second device receives a preamble from the first device;

[0025] The second device measures inter-symbol interference according to the preamble code, and determines neural network setting information according to the inter-symbol interference, wherein the neural network setting information includes time step information and hidden layer size information;

[0026] The second device sends the neural network setting information to the first device;

[0027] The first device receives the neural network setting information sent by the second device;

[0028] The first device obtains neural network parameter information by using the gliding recurrent neural network training according to the neural network setting information, wherein the gliding recurrent neural network includes an input layer, at least one hidden layer and an output layer. During the training process, the input layer converts a serial input sequence into a gliding input sequence according to the time step information, and the at least one hidden layer performs balanced training based on the gliding input sequence.

[0029] According to another embodiment of the present application, a first device is further provided, wherein the first device includes:

[0030] A memory for storing one or more programs;

[0031] One or more processors connected to the memory,

[0032] When the one or more programs are executed by the one or more processors, the one or more processors are caused to perform the following operations:

[0033] Sending a preamble to a second device;

[0034] Receiving neural network setting information sent by the second device, wherein the neural network setting information includes time step information and hidden layer size information;

[0035] According to the neural network setting information, the neural network parameter information is obtained by using a gliding recurrent neural network training, wherein the gliding recurrent neural network includes an input layer, at least one hidden layer and an output layer, and during the training process, the input layer converts a serial input sequence into a gliding input sequence according to the time step information, and the at least one hidden layer performs balanced training based on the gliding input sequence.

[0036] According to another embodiment of the present application, a second device is further provided, wherein the second device includes:

[0037] A memory for storing one or more programs;

[0038] One or more processors connected to the memory,

[0039] When the one or more programs are executed by the one or more processors, the one or more processors are caused to perform the following operations:

[0040] receiving a preamble from a first device;

[0041] Measuring inter-symbol interference according to the preamble code, and determining neural network setting information according to the inter-symbol interference, wherein the neural network setting information includes time step information and hidden layer size information;

[0042] The neural network setting information is sent to the first device, so that the first device obtains neural network parameter information by using the gliding cyclic neural network training according to the neural network setting information.

[0043] Compared with the prior art, the present application has the following advantages: a new balancing concept based on a sliding recurrent neural network is proposed. By setting the time step information on the second device side, the neural network setting information including the time step information and the hidden layer size information can be sent to the first device side, so that the neural network parameter information can be obtained by training the sliding recurrent neural network according to the neural network setting information in the first device. Since the input layer of the sliding recurrent neural network can convert the serial input sequence into a sliding input sequence according to the time step information, there is only one input at the same time during the training process, thereby the complex weight matrix W of the FCNN can be converted into a sliding input sequence. Ni,h1 Simplified to a simple W 1,h1 The resulting low complexity makes SRNN-based equalization easier to implement in hardware such as FPGA (field programmable gate array), ASIC (application-specific integrated circuit), SoC (system-on-chip), and better performance can be achieved through new SRNN-based equalization training. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings:

[0045] Figure 1 It is a flowchart of equalization based on FCNN in the prior art;

[0046] Figure 2 It is a schematic diagram of the training process of FCNN-based equalization in the prior art;

[0047] Figure 3 A schematic flow chart of a method for signal equalization in a high-speed communication system in a first device according to an embodiment of the present application is shown;

[0048] Figure 4 A schematic flow chart of a method for signal equalization in a high-speed communication system in a second device according to an embodiment of the present application is shown;

[0049] Figure 5 A schematic diagram of a process for implementing high-speed PON signal equalization based on SRNN according to an example of the present application is shown;

[0050] Figure 6 A schematic diagram of a training process for implementing high-speed PON signal equalization based on SRNN according to an example of the present application is shown;

[0051] Figure 7 A simulation schematic diagram of an example of the present application is shown;

[0052] Figure 8 Shown based on Figure 7 The simulation results of an example of

[0053] Fig. 9 Shown based on Figure 7 Simulation results of another example;

[0054] Fig.10 A schematic diagram showing the structure of a first device for signal equalization in a high-speed communication system in a first device according to an embodiment of the present application is shown;

[0055] Fig.11 A schematic diagram showing the structure of a second device for signal equalization in a high-speed communication system in a second device according to an embodiment of the present application is shown;

[0056] Fig.12 An exemplary system is shown that can be used to implement the various embodiments described in this application.

[0057] The same or similar reference numerals in the drawings represent the same or similar components. DETAILED DESCRIPTION

[0058] It should be mentioned before discussing the exemplary embodiments in more detail that some exemplary embodiments are described as processes or methods depicted as flow charts. Although the flow charts describe the operations as sequential processes, many of the operations therein can be implemented in parallel, concurrently or simultaneously. In addition, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but can also have additional steps not included in the accompanying drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0059] The "device" referred to in this context refers to an intelligent electronic device that can perform predetermined processing processes such as numerical calculations and / or logical calculations by running predetermined programs or instructions. It may include a processor and a memory, and the processor executes program instructions pre-stored in the memory to perform the predetermined processing process, or the predetermined processing process is performed by hardware such as an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), a digital signal processor (DSP), etc., or is achieved by a combination of the above two.

[0060] The methods discussed later in this document (some of which are illustrated by flow charts) may be implemented by hardware, software, firmware, middleware, microcode, hardware description languages, or any combination thereof. When implemented by software, firmware, middleware, or microcode, program code or code segments to perform the necessary tasks may be stored in a machine or computer readable medium (such as a storage medium). (One or more) processors may perform the necessary tasks.

[0061] The specific structural and functional details disclosed herein are merely representative and are for the purpose of describing exemplary embodiments of the present application. However, the present application can be implemented in many alternative forms and should not be interpreted as being limited to only the embodiments set forth herein.

[0062] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0063] The terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments. Unless the context clearly indicates otherwise, the singular forms "one", "one" and "item" used herein are also intended to include plural numbers. It should also be understood that the terms "include" and / or "comprise" used herein specify the existence of stated features, integers, steps, operations, units and / or components, without excluding the existence or addition of one or more other features, integers, steps, operations, units, components and / or combinations thereof.

[0064] It should also be noted that, in some alternative implementations, the functions / actions mentioned may occur in a different order than that indicated in the accompanying drawings. For example, depending on the functions / actions involved, two figures shown in succession may actually be performed substantially simultaneously or may sometimes be performed in the reverse order.

[0065] The present application found that for high-speed communication systems, FCNN-based equalization faces many challenges. For example, for 50G+HSP, the fundamental challenge of FCNN-based equalization is that at least the following three aspects must be addressed: 1) Some researchers have shown that in experiments using pseudo-random bit sequences, the NN actually understands the PRBS generation pattern rather than predicting the channel state; 2) When the nonlinearity and inter-symbol interference (ISI) of 50G+HSP using bandwidth-limited optical devices (such as 20GHz-class cost-effective optical devices) are very severe, the bit error rate (BER) performance of FCNN with a moderate NN size cannot meet the HD-FEC threshold of 0.0038; 3) The performance of FCNN mainly depends on the large input data features, and the complex weight matrix W Ni,h1 (Ni and h1 are the input feature size and the number of neurons in the first hidden layer respectively) greatly increases the complexity of ASIC hardware.

[0066] The problem to be solved by this application is how to break the above dilemma and maintain its superior performance. This application provides a new concept of equalization, which can have an advantage in performance when the input size is 1 through a unique sliding method. This application first proposes a new SRNN to mine the serial dependencies between signals through time for signal equalization in high-speed communication systems. It should be noted that the scheme of this application is applicable to any high-speed communication system that requires signal equalization, such as 50G+HSP, 5G or higher versions of wireless communication systems. The scheme of this application is mainly implemented by a first device and a second device in a high-speed communication system; in some embodiments, the first device is an optical line terminal in a high-speed PON, and the second device is an optical network unit in the high-speed PON. By making the optical line terminal and the optical network unit perform the method described in this application, signal equalization in the high-speed PON can be achieved; in some embodiments, the first device is a base station in a high-speed wireless communication system (such as gNB in ​​5G), and the second device is a user equipment in the high-speed wireless communication system. By making the base station and the user equipment perform the method described in this application, signal equalization in a high-speed wireless communication system can be achieved.

[0067] Figure 3 A method for signal equalization in a high-speed communication system in a first device according to an embodiment of the present application is shown. The method according to this embodiment includes step S11, step S12, and step S13. In step S11, the first device sends a preamble to the second device; in step S12, the first device receives the neural network setting information sent by the second device, wherein the neural network setting information includes time step information and hidden layer size information; in step S13, the first device obtains neural network parameter information by using a sliding recurrent neural network training according to the neural network setting information, wherein the sliding recurrent neural network includes an input layer, at least one hidden layer, and an output layer. During the training process, the input layer converts a serial input sequence into a sliding input sequence according to the time step information, and the at least one hidden layer performs equalization training based on the sliding input sequence.

[0068] In step S11, the first device sends a preamble to the second device. For example, the first device is an optical line terminal in a high-speed PON, the second device is an optical network unit in the high-speed PON, and the optical line terminal sends a preamble to the optical network unit.

[0069] In step S12, the first device receives the neural network setting information sent by the second device, wherein the neural network setting information includes time step information and hidden layer size information. In some embodiments, the neural network setting information is determined by the second device based on the measured ISI. In some embodiments, the neural network setting information is determined by the second device based on the measured ISI and in combination with other information (such as bandwidth, etc.). In some embodiments, the neural network setting information also includes other setting information related to the neural network. In some embodiments, the time step information (TimeSteps, also referred to as "Ts" in the context) is also the TimeSteps in the sliding neural network, which is used for the conversion of the serial input sequence to the sliding input sequence. The specific implementation method will be described in detail in the subsequent embodiments. In some embodiments, the hidden layer size information is used to indicate the size of each hidden layer included in the neural network. For example, the sliding neural network includes two hidden layers, and the hidden layer size information includes the number of RNN neurons in the first hidden layer and the number of normal neurons in the second hidden layer.

[0070] In step S13, the first device uses the gliding recurrent neural network training to obtain neural network parameter information according to the neural network setting information, wherein the gliding recurrent neural network includes an input layer, at least one hidden layer and an output layer. During the training process, the input layer converts the serial input sequence into a gliding input sequence according to the time step information, and the at least one hidden layer performs equalization training based on the gliding input sequence. In some embodiments, the first device constructs a gliding recurrent neural network according to the received neural network setting information, and obtains neural network parameter information by using the gliding recurrent neural network training, so as to realize signal equalization of the high-speed communication system based on the neural network parameter information obtained by training. In some embodiments, the neural network parameter information includes any NN parameters related to signal equalization of the high-speed communication system, such as weights, biases, etc.

[0071] In some embodiments, the conversion from the serial input sequence to the sliding input sequence is performed by a unique sliding method in the input layer. The number of input features in each sliding input sequence obtained by sliding is the same (the size of the time step information is also used as the size of the sliding input sequence). As an example, assuming that the time step information Ts is 4, the upcoming serial input sequence (X t-2 , X t-1 , X t , X t+1 , X t+2 , ...) is converted into many sliding input sequences by sliding, such as sliding input sequence 1 (X t-2 , X t-1 , X t, X t+1 ), slide input sequence 2 (X t-1 , X t , X t+1 , X t+2 ), etc., where each sliding input sequence contains 4 input features.

[0072] In some embodiments, the at least one hidden layer includes one or more hidden layers with RNN neurons ("hidden layers with RNN neurons" may also be referred to as "RNN layers" in the context). In some embodiments, the at least one hidden layer also includes one or more hidden layers with normal neurons ("hidden layers with normal neurons" may also be referred to as "FC" layers in the context). In some embodiments, the at least one hidden layer includes only one RNN layer (such as a gliding recurrent neural network includes in sequence: an input layer, an RNN layer, a hidden layer with normal neurons, and an output layer), and the input layer obtains a gliding input sequence through serial-to-gliding conversion and is input into the RNN layer. The RNN layer has only one input at the same time, that is, only one input feature in the gliding input sequence is input into the RNN layer at the same time. In some embodiments, the at least one hidden layer includes multiple RNN layers, and the sliding input sequence obtained by serial to sliding conversion in the input layer is input to the next RNN layer, and the RNN layer has only one input at the same time, wherein, if the next hidden layer after an RNN layer is still an RNN layer, all states in the RNN layer are connected to the next RNN layer, and if the next hidden layer after an RNN layer is a hidden layer with normal neurons, only the last state in the RNN layer is connected to the hidden layer with normal neurons. In the prior art, FCNN uses a parallel input method, so there are multiple input features at the same time, which requires the use of a complex weight matrix W Ni,h1 (used to process many features in parallel), and this application is based on the newly proposed sliding recurrent neural network, which converts the serial input sequence into a sliding input sequence at the input layer, so that during the training process, only one input feature of the sliding input sequence is input to the RNN layer of the next layer of the input layer at the same time. Therefore, it can be simplified to use a simple weight matrix W 1,h1 The resulting low complexity makes SRNN-based equalization easy to implement in hardware such as FPGA, ASIC, and SoC.

[0073] In some embodiments, the at least one hidden layer includes a first hidden layer having RNN neurons and a second hidden layer having normal neurons; during the training process, each element in the sliding input sequence is recursively operated with the corresponding state of each RNN neuron in the first hidden layer in turn. For example, the at least one hidden layer includes a hidden layer layer1 having RNN neurons (R1, R2, R3) located after the input layer, and multiple sliding input sequences converted in the input layer will be sequentially input into the layer1, and each element in each sliding input sequence will be sequentially recursively operated with the corresponding state in R1, R2 and R3. For example, the sliding input sequence (X t-2 , X t-1 , X t , X t+1 ) t-2 Connect to state S in R1 t-2 (That is, X t-2 Need to be with S t-2 operation), X t-1 Connect to state S in R1 t-1 (That is, X t-1 Need to be with S t-1 operation), X t Connect to state S in R1 t (That is, X t Need to be with S t operation), X t+1 Connect to state S in R1 t+1 (That is, X t+1 Need to be with S t+1 Similarly, the sliding input sequence (X t-2 , X t-1 , X t , X t+1 ) are also connected to R2 and R3 (S t-2 , S t-1 , S t , S t+1 ) and perform recursive operations on the corresponding states in the gliding recurrent neural network, which will not be repeated here. Different from the conventional equalizer, the present application uses RNN neurons with state-dependent memory in the gliding recurrent neural network.

[0074] In some embodiments, after each element in the sliding input sequence completes the recursive operation with the corresponding state in each RNN neuron in the RNN layer (that is, after a specific Ts), the RNN layer retains the last state, and only the last state of the RNN layer is input to each normal neuron in the second hidden layer. For example, continuing the above example, after a specific Ts, R1, R2 and R3 can only output the last state S t+1 (S t=sigmoid(W h1,h1 S t-1 +W 1,h1 X t )), and output S t+1 is connected to every normal neuron in the next hidden layer. Then, the final state of the first hidden layer output is combined with the weight matrix W h1 ,h2 After multiplication, the output is sent to the second hidden layer (including AF), and finally combined with the weight matrix W h2,1 After multiplication, the output is sent to the output layer O1 (including AF).

[0075] In some embodiments, the method further includes: the first device sends the neural network parameter information obtained through training to the second device; the first device receives a ready signal from the second device and sends a payload to the second device. Based on this, signal equalization in a high-speed communication system can be achieved. It should be noted that in the process of training the neural network parameter information, the first device needs to complete the training based on the time step information set by the second device, and when the training is completed, only the neural network parameter information obtained through training needs to be directly used without considering the time step information.

[0076] Figure 4 A flow chart of a method for signal equalization in a high-speed communication system in a second device according to an embodiment of the present application is shown. The method according to this embodiment includes step S21, step S22, and step S23. In step S21, the second device receives a preamble from the first device; in step S22, the second device obtains inter-symbol interference based on the preamble measurement, and determines neural network setting information based on the inter-symbol interference, wherein the neural network setting information includes time step information and hidden layer size information; in step S23, the second device sends the neural network setting information to the first device, so that the first device obtains neural network parameter information by using the gliding recurrent neural network training according to the neural network setting information.

[0077] In step S21, the second device receives a preamble from the first device. For example, the optical network unit receives a preamble from an optical line terminal.

[0078] In step S22, the second device measures the inter-symbol interference according to the preamble code, and determines the neural network setting information according to the inter-symbol interference, wherein the neural network setting information includes time step information and hidden layer size information. In some embodiments, the second device determines the neural network setting information according to the inter-symbol interference and further combines other information (such as bandwidth, etc.). The neural network setting information has been described in detail in the above embodiments and will not be repeated here.

[0079] In step S23, the second device sends the neural network setting information to the first device, so that the first device obtains the neural network parameter information by using the gliding cyclic neural network training according to the neural network setting information. The implementation method of the first device obtaining the neural network parameter information by using the gliding cyclic neural network training according to the neural network setting information has been described in detail in the above embodiment and will not be repeated here.

[0080] In some embodiments, the method also includes: a second device receives the neural network parameter information sent by the first device; the second device configures the first device according to the neural network parameter information; the second device sends a ready signal to the first device; and the second device receives the payload sent by the first device.

[0081] The present application also provides a method for signal equalization in a high-speed communication system, wherein the method includes: a first device sends a preamble code to a second device; the second device receives the preamble code from the first device; the second device measures inter-symbol interference according to the preamble code, and determines neural network setting information according to the inter-symbol interference, wherein the neural network setting information includes time step information and hidden layer size information; the second device sends the neural network setting information to the first device; the first device receives the neural network setting information sent by the second device; the first device uses a gliding recurrent neural network training to obtain neural network parameter information according to the neural network setting information, wherein the gliding recurrent neural network includes an input layer, at least one hidden layer and an output layer, and during the training process, the input layer converts a serial input sequence into a gliding input sequence according to the time step information, and the at least one hidden layer performs equalization training based on the gliding input sequence. The relevant operations have been described in detail in the above embodiments and will not be repeated here.

[0082] Figure 5 A schematic diagram of a process for implementing high-speed PON signal equalization based on SRNN in an example of the present application is shown. The SRNN in this example includes an input layer, a hidden layer Hidden1 with RNN neurons (R1, R2, R3), a hidden layer Hidden2 with normal neurons (N5, N6), and an output layer. Figure 5The SRNN-based workflow shown is as follows: 1) OLT sends a preamble code to ONU; 2) ONU measures ISI and determines that the neural network setting information is (Ts, h1, h2); 3) ONU sends (Ts, h1, h2) to OLT; 4) further includes operations a) and b), wherein a) serial to glide conversion, b) SRNN-based equalization training; 5) OLT sends the trained neural network parameter information (NN params) to ONU; 6) ONU is configured according to the neural network parameter information; 7) ONU sends a ready instruction to OLT; 8) OLT sends a payload to ONU.

[0083] Figure 6 A schematic diagram of a training process for realizing high-speed PON signal equalization based on SRNN is shown in an example of the present application, that is, Figure 5 The specific implementation process of step 4) is shown in FIG. Wherein, operation a) is used to perform a serial to sliding conversion (Serial to Sliding), Figure 6 The two sliding sequences Sliding1 and Sliding2 after conversion are shown, and Sliding1 includes (X t-2 , X t-1 , X t , X t+1 ), Sliding2 includes (X t-1 , X t , X t+1 , X t+2 ),Depend on Figure 6 It can be seen that due to the sliding mode, there is only one input (Input1). Among them, step b) is used for SRNN-based equalization training, hidden layer 1 (Hidden1: RNN) includes three RNN neurons (R1, R2, R3), and each input data in each sliding input sequence is recursively operated with the corresponding state in each RNN neuron in the first hidden layer in turn. For simplicity, Figure 6 Only the process of calculating the corresponding states in Sliding1 and R1 is shown in FIG. t-2 Connect to state S in R1 t-2 , connect X t-1 Connect to state S in R1 t-1 , X t Connect to state S in R1 t , X t+1 Connect to state S in R1 t+1 , after a certain Ts (Ts=4 in this example), each RNN retains the last state, and the RNN layer only retains the last state S t+1is input into hidden layer 2 (Hidden2: FC2) (i.e., the final state S in each RNN neuron t+1 All are connected to the normal neurons N5 and N6 in Hidden2. The final state of each RNN neuron is calculated by the weight matrix W h1,h2 After multiplication, it is passed to N5 and N6 in Hidden2, and finally passed through the weight matrix W h2,1 After multiplication, it is output at the output layer. Figure 6 It can be seen that since there is only one input at the same time (that is, the input size is 1), the weight matrix of Hidden1 is simplified to W 1,h1 .

[0084] In this application, TensorFlow and VPItransmissionMaker are used TM The Optical system was simulated to verify the SRNN concept proposed in this application. The simulation principle is as follows Figure 7 As shown, the present application is simulated on a 50Gb / s / λNRZ IM-DD PON to measure the performance of the solution of the present application. The NRZ signal is modulated on a 20G-class C-band directly modulated laser (Directly Modulated Laser, DML), and after 20 km SSMF transmission, the receiver sensitivity is measured using a variable optical attenuator (Variable Optical Attenuator, VOA). The optical signal is detected by a 20G-class photodiode (PD). Then, the detected signal is sampled by a digital sampling oscilloscope (DSO) with a 59GHz bandwidth and a 160-GSa / s sampling rate. Finally, the resampled signal is processed using TensorFlow based on the SRNN-based equalization proposed in this application.

[0085] like Figure 8 The simulation results of TensorFlow shown in FIG. 1 show that the BER performance of the SRNN proposed in this application mainly depends on the key TimeSteps (Ts). After detailed simulation, Ts is optimized to 13 ( Figure 8 The A' of SRNN is 0.0005). Fig. 9 As shown, the SRNN proposed in this application has only 1 input size (thus converting the complex matrix W in FCNN into 20,15 Convert to simple W 1,15), and with the same size (15,3,1) NN, its BER performance can be enhanced from A: 0.0033 (FCNN) to B: 6e-5 (SRNN, due to the inherent serial dependency in time), so the ROP is -10dBm, which is 10 times higher than FCNN. Furthermore, this application also reduces the NN structure to (8,3,1), and the conclusion is that the BER performance of the SRNN proposed in this application can be enhanced from C: 0.0210 (the FCNN of [13in, (8,3,1)] cannot reach the threshold of HD-FEC) to D: 0.0005 (through [1in, (8,3,1)] SRNN, even better than E: 0.0010 achieved by the larger [30in, (30,3,1)] FCNN).

[0086] For FPGA hardware complexity, from the perspective of input features, the hardware complexity of the same NN structure (15,3,1) can be reduced from the complexity matrix W 20,15 Simplified to simple W 1,15 Of course, the hardware complexity of FPGA based SRNN should be carefully evaluated, including the design of state-dependent memory (additional W h1,h1 As additional complexity). After a detailed comprehensive evaluation, the overall complexity of the new SRNN with [1in, (8, 3, 1)] and Ts = 13 (the overall complexity does not consider bias, the number of multiplications is: (1 + 8) * 8 + 8 * 3 + 3 * 1 = 99, which includes the additional complexity W 8,8 ) is lower than the overall complexity of the FCNN with [13in, (8, 3, 1)] (the overall complexity does not consider bias, the number of multiplications is: 13*8+8*3+3*1=131), and the new SRNN has better BER performance (such as Fig. 9 C to D), especially at high ROP (greater than -12dBm).

[0087] The SRNN-based equalization concept proposed in this application can be used as an upgrade of the current 25G / λPON in the future 50G+HSP (e.g. Figure 3 As shown in ), SRNN-based equalization can support 50G+ / λ and support cost-effective low-grade 20GHz optical devices, supporting higher performance and better power budget.

[0088] According to the solution of the present application, a new SRNN-based equalization concept is proposed. By setting the time step information on the second device end, the neural network setting information including the time step information and the hidden layer size information can be sent to the first device, so that the neural network parameter information can be obtained by training the sliding recurrent neural network according to the neural network setting information in the first device. Since the input layer of the sliding recurrent neural network can convert the serial input sequence into a sliding input sequence according to the time step information, there is only one input at the same time during the training process, thereby the complex weight matrix W of the FCNN can be converted into a sliding input sequence. Ni,h1 Simplified to a simple W 1,h1 The resulting low complexity makes SRNN-based equalization easier to implement in hardware such as FPGA (field programmable gate array), ASIC (application-specific integrated circuit), SoC (system-on-chip), and better performance can be achieved through new SRNN-based equalization training.

[0089] Fig.10 The schematic diagram of the structure of a first device for high-speed passive optical network signal equalization in a first device according to an embodiment of the present application is shown. The first device 1 comprises a first unit 11 , a second unit 12 and a third unit 13 .

[0090] The first unit 11 is used to send a preamble to the second device. For example, the first device is an optical line terminal in a high-speed PON, the second device is an optical network unit in the high-speed PON, and the first unit 11 in the optical line terminal sends a preamble to the optical network unit.

[0091] The second unit 12 is used to receive the neural network setting information sent by the second device, wherein the neural network setting information includes time step information and hidden layer size information. In some embodiments, the neural network setting information is determined by the second device based on the measured ISI. In some embodiments, the neural network setting information is determined by the second device based on the measured ISI and in combination with other information (such as bandwidth, etc.). In some embodiments, the neural network setting information also includes other setting information related to the neural network. In some embodiments, the time step information (TimeSteps, also referred to as "Ts" in the context) is also the TimeSteps in the sliding neural network, which is used for the conversion of the serial input sequence to the sliding input sequence. The specific implementation method will be described in detail in the subsequent embodiments. In some embodiments, the hidden layer size information is used to indicate the size of each hidden layer included in the neural network. For example, the sliding neural network includes two hidden layers, and the hidden layer size information includes the number of RNN neurons in the first hidden layer and the number of normal neurons in the second hidden layer.

[0092] The third unit 13 is used to obtain neural network parameter information by using the gliding recurrent neural network training according to the neural network setting information, wherein the gliding recurrent neural network includes an input layer, at least one hidden layer and an output layer. During the training process, the input layer converts the serial input sequence into a gliding input sequence according to the time step information, and the at least one hidden layer performs equalization training based on the gliding input sequence. In some embodiments, the third unit 13 constructs a gliding recurrent neural network according to the received neural network setting information, and obtains neural network parameter information by using the gliding recurrent neural network training, so as to realize signal equalization of the high-speed communication system based on the neural network parameter information obtained by training. In some embodiments, the neural network parameter information includes any NN parameters related to signal equalization of the high-speed communication system, such as weights, biases, etc.

[0093] In some embodiments, the conversion from the serial input sequence to the sliding input sequence is performed by a unique sliding method in the input layer. The number of input features in each sliding input sequence obtained by sliding is the same (the size of the time step information is also used as the size of the sliding input sequence). As an example, assuming that the time step information Ts is 4, the upcoming serial input sequence (X t-2 , X t-1 , X t , X t+1 , X t+2 , ...) is converted into many sliding input sequences by sliding, such as sliding input sequence 1 (X t-2 , X t-1 , X t , X t+1 ), slide input sequence 2 (X t-1 , X t , X t+1 , X t+2 ), etc., where each sliding input sequence contains 4 input features.

[0094] In some embodiments, the at least one hidden layer includes one or more hidden layers with RNN neurons ("hidden layers with RNN neurons" may also be referred to as "RNN layers" in the context). In some embodiments, the at least one hidden layer also includes one or more hidden layers with normal neurons ("hidden layers with normal neurons" may also be referred to as "FC" layers in the context). In some embodiments, the at least one hidden layer includes only one RNN layer (such as a gliding recurrent neural network includes in sequence: an input layer, an RNN layer, a hidden layer with normal neurons, and an output layer), and the input layer obtains a gliding input sequence through serial-to-gliding conversion and is input into the RNN layer. The RNN layer has only one input at the same time, that is, only one input feature in the gliding input sequence is input into the RNN layer at the same time. In some embodiments, the at least one hidden layer includes multiple RNN layers, and the sliding input sequence obtained by serial to sliding conversion in the input layer is input to the next RNN layer, and the RNN layer has only one input at the same time, wherein, if the next hidden layer after an RNN layer is still an RNN layer, all states in the RNN layer are connected to the next RNN layer, and if the next hidden layer after an RNN layer is a hidden layer with normal neurons, only the last state in the RNN layer is connected to the hidden layer with normal neurons. In the prior art, FCNN uses a parallel input method, so there are multiple input features at the same time, which requires the use of a complex weight matrix W Ni,h1 (used to process many features in parallel), and this application is based on the newly proposed sliding recurrent neural network, which converts the serial input sequence into a sliding input sequence at the input layer, so that during the training process, only one input feature of the sliding input sequence is input to the RNN layer of the next layer of the input layer at the same time. Therefore, it can be simplified to use a simple weight matrix W 1,h1 The resulting low complexity makes SRNN-based equalization easy to implement in hardware such as FPGA, ASIC, and SoC.

[0095] In some embodiments, the at least one hidden layer includes a first hidden layer having RNN neurons and a second hidden layer having normal neurons; during the training process, each element in the sliding input sequence is recursively operated with the corresponding state of each RNN neuron in the first hidden layer in turn. For example, the at least one hidden layer includes a hidden layer layer1 having RNN neurons (R1, R2, R3) located after the input layer, and multiple sliding input sequences converted in the input layer will be sequentially input into the layer1, and each element in each sliding input sequence will be sequentially recursively operated with the corresponding state in R1, R2 and R3. For example, the sliding input sequence (X t-2 , X t-1 , X t, X t+1 ) t-2 Connect to state S in R1 t-2 (That is, X t-2 Need to be with S t-2 operation), X t-1 Connect to state S in R1 t-1 (That is, X t-1 Need to be with S t-1 operation), X t Connect to state S in R1 t (That is, X t Need to be with S t operation), X t+1 Connect to state S in R1 t+1 (That is, X t+1 Need to be with S t+1 Similarly, the sliding input sequence (X t-2 , X t-1 , X t , X t+1 ) are also connected to R2 and R3 (S t-2 , S t-1 , S t , S t+1 ) and perform recursive operations on the corresponding states in the gliding recurrent neural network, which will not be repeated here. Different from the conventional equalizer, the present application uses RNN neurons with state-dependent memory in the gliding recurrent neural network.

[0096] In some embodiments, after each element in the sliding input sequence completes the recursive operation with the corresponding state in each RNN neuron in the RNN layer (that is, after a specific Ts), the RNN layer retains the last state, and only the last state of the RNN layer is input to each normal neuron in the second hidden layer. For example, continuing the above example, after a specific Ts, R1, R2 and R3 can only output the last state S t+1 (S t =sigmoid(W h1,h1 S t-1 +W 1,h1 X t )), and output S t+1 is connected to every normal neuron in the next hidden layer. Then, the final state of the first hidden layer output is combined with the weight matrix W h1 ,h2 After multiplication, the output is sent to the second hidden layer (including AF), and finally combined with the weight matrix W h2,1 After multiplication, the output is sent to the output layer O1 (including AF).

[0097] In some embodiments, the first device 1 further includes: a device for sending the neural network parameter information obtained through training to the second device (not shown); a device for receiving a ready signal from the second device and sending a payload to the second device (not shown). Based on this, signal equalization in a high-speed communication system can be achieved. It should be noted that in the process of training the neural network parameter information, the first device needs to complete the training based on the time step information set by the second device, and when the training is completed, only the neural network parameter information obtained through training needs to be used directly without considering the time step information.

[0098] Fig.11 The schematic diagram of the structure of the second device for high-speed passive optical network signal equalization in the second equipment according to one embodiment of the present application is shown. The second device 2 includes a fourth unit 21 , a fifth unit 22 and a sixth unit 23 .

[0099] The fourth unit 21 is used to receive a preamble from the first device. For example, the fourth unit 21 in the optical network unit receives a preamble from an optical line terminal.

[0100] The fifth unit 22 is used to measure the inter-symbol interference according to the preamble code, and determine the neural network setting information according to the inter-symbol interference, wherein the neural network setting information includes time step information and hidden layer size information. In some embodiments, the fifth unit 22 determines the neural network setting information according to the inter-symbol interference and further combines other information (such as bandwidth, etc.). The neural network setting information has been described in detail in the above embodiments and will not be repeated here.

[0101] The sixth unit 23 is used to send the neural network setting information to the first device, so that the first device obtains the neural network parameter information by using the gliding cyclic neural network training according to the neural network setting information. The implementation method of the first device obtaining the neural network parameter information by using the gliding cyclic neural network training according to the neural network setting information has been described in detail in the above embodiment and will not be repeated here.

[0102] In some embodiments, the second device 2 also includes: a device for receiving the neural network parameter information sent by the first device (not shown in the figure); a device for configuring the first device according to the neural network parameter information (not shown in the figure); a device for sending a ready signal to the first device (not shown in the figure); and a device for receiving a payload sent by the first device (not shown in the figure).

[0103] According to the solution of the present application, a new SRNN-based equalization concept is proposed. By setting the time step information on the second device end, the neural network setting information including the time step information and the hidden layer size information can be sent to the first device, so that the neural network parameter information can be obtained by training the sliding recurrent neural network according to the neural network setting information in the first device. Since the input layer of the sliding recurrent neural network can convert the serial input sequence into a sliding input sequence according to the time step information, there is only one input at the same time during the training process, thereby the complex weight matrix W of the FCNN can be converted into a sliding input sequence. Ni,h1 Simplified to a simple W 1,h1 The resulting low complexity makes SRNN-based equalization easier to implement in hardware such as FPGA (field programmable gate array), ASIC (application-specific integrated circuit), SoC (system-on-chip), and better performance can be achieved through new SRNN-based equalization training.

[0104] Fig.12 An exemplary system that can be used to implement the various embodiments described in this application is shown. In some embodiments, the system 1000 can be used as any processing device in the embodiments of the present application. In some embodiments, the system 1000 may include one or more computer-readable media (e.g., system memory or NVM / storage device 1020) with instructions and one or more processors (e.g., (one or more) processors 1005) coupled to the one or more computer-readable media and configured to execute instructions to implement modules to perform the actions described in this application.

[0105] For one embodiment, the system control module 1010 may include any suitable interface controller to provide any suitable interface to at least one of the processor(s) 1005 and / or any suitable device or component in communication with the system control module 1010 .

[0106] The system control module 1010 may include a memory controller module 1030 to provide an interface to the system memory 1015. The memory controller module 1030 may be a hardware module, a software module, and / or a firmware module.

[0107] The system memory 1015 may be used, for example, to load and store data and / or instructions for the system 1000. For one embodiment, the system memory 1015 may include any suitable volatile memory, such as a suitable DRAM. In some embodiments, the system memory 1015 may include double data rate type four synchronous dynamic random access memory (DDR4 SDRAM).

[0108] For one embodiment, system control module 1010 may include one or more input / output (I / O) controllers to provide interfaces to NVM / storage device 1020 and communication interface(s) 1025 .

[0109] For example, NVM / storage device 1020 may be used to store data and / or instructions. NVM / storage device 1020 may include any suitable non-volatile memory (e.g., flash memory) and / or may include any suitable non-volatile storage device(s) (e.g., one or more hard disk drives (HDDs), one or more compact disk (CD) drives, and / or one or more digital versatile disk (DVD) drives).

[0110] NVM / storage device 1020 may include storage resources that are physically part of the device on which system 1000 is installed, or it may be accessible to the device without being part of the device. For example, NVM / storage device 1020 may be accessed via communication interface(s) 1025 over a network.

[0111] Communication interface(s) 1025 may provide an interface for system 1000 to communicate over one or more networks and / or with any other suitable devices. System 1000 may wirelessly communicate with one or more components of a wireless network in accordance with any of one or more wireless network standards and / or protocols.

[0112] For one embodiment, at least one of the processor(s) 1005 may be packaged together with the logic of one or more controllers (e.g., memory controller module 1030) of the system control module 1010. For one embodiment, at least one of the processor(s) 1005 may be packaged together with the logic of one or more controllers of the system control module 1010 to form a system-in-package (SiP). For one embodiment, at least one of the processor(s) 1005 may be integrated on the same die with the logic of one or more controllers of the system control module 1010. For one embodiment, at least one of the processor(s) 1005 may be integrated on the same die with the logic of one or more controllers of the system control module 1010 to form a system on chip (SoC).

[0113] In various embodiments, the system 1000 may be, but is not limited to: a server, a workstation, a desktop computing device, or a mobile computing device (e.g., a laptop computing device, a handheld computing device, a tablet computer, a netbook, etc.). In various embodiments, the system 1000 may have more or fewer components and / or a different architecture. For example, in some embodiments, the system 1000 includes one or more cameras, a keyboard, a liquid crystal display (LCD) screen (including a touch screen display), a non-volatile memory port, multiple antennas, a graphics chip, an application specific integrated circuit (ASIC), and a speaker.

[0114] The present application also provides a first device, wherein the first device includes: a memory for storing one or more programs; and one or more processors connected to the memory, so that when the one or more programs are executed by the one or more processors, the one or more processors execute the method for signal equalization in a high-speed communication system in the first device described in the present application.

[0115] The present application also provides a second device, wherein the second device includes: a memory for storing one or more programs; and one or more processors connected to the memory, so that when the one or more programs are executed by the one or more processors, the one or more processors execute the method for signal equalization in a high-speed communication system in the second device described in the present application.

[0116] The present application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program can be executed by a processor to implement the method for signal equalization in a first device for a high-speed communication system as described in the present application.

[0117] The present application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program can be executed by a processor to implement the method for signal equalization in a second device for a high-speed communication system as described in the present application.

[0118] The present application also provides a computer program product. When the computer program product is executed by a device, the device executes the method for signal equalization in a first device for a high-speed communication system described in the present application.

[0119] The present application also provides a computer program product. When the computer program product is executed by a device, the device executes the method for signal equalization in a second device for a high-speed communication system described in the present application.

[0120] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and that the present application can be implemented in other specific forms without departing from the spirit or basic features of the present application. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive, and the scope of the present application is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present application. Any figure mark in the claims should not be regarded as limiting the claims involved. In addition, it is obvious that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices stated in the system claim can also be implemented by one unit or device through software or hardware. The words first, second, etc. are used to indicate names, and do not indicate any particular order.

Claims

1. A method for signal equalization in a high-speed communication system, applied to a first device, wherein: The method comprises: Sending a preamble to a second device; Receiving neural network setting information sent by the second device, wherein the neural network setting information includes time step information and hidden layer size information; According to the neural network setting information, neural network parameter information is obtained by using a sliding recurrent neural network training, wherein the sliding recurrent neural network includes an input layer, at least one hidden layer and an output layer, during the training process, the input layer converts a serial input sequence into a sliding input sequence according to the time step information, and the at least one hidden layer performs balanced training based on the sliding input sequence; The serial input sequence is converted into a plurality of sliding input sequences by converting the serial input sequence into a sliding input sequence, each sliding input sequence contains the same number of input features, and the number of input features contained in each sliding input sequence corresponds to the size of the time step information.

2. The method according to claim 1, wherein: The method further comprises: Sending the neural network parameter information obtained through training to the second device; A ready signal is received from the second device, and a payload is sent to the second device.

3. The method according to claim 1 or 2, wherein: The at least one hidden layer includes one or more hidden layers having RNN neurons.

4. The method according to claim 3, wherein: The at least one hidden layer also includes one or more hidden layers having normal neurons.

5. The method according to claim 1 or 2, wherein: The first device is an optical line terminal in a high-speed passive optical network, and the second device is an optical network unit in the high-speed passive optical network.

6. The method according to claim 1 or 2, wherein: The first device is a base station in a high-speed wireless communication system, and the second device is a user equipment in the high-speed wireless communication system.

7. A method for signal equalization in a high-speed communication system, applied to a second device, wherein: The method comprises: receiving a preamble from a first device; Measuring inter-symbol interference according to the preamble code, and determining neural network setting information according to the inter-symbol interference, wherein the neural network setting information includes time step information and hidden layer size information; Sending the neural network setting information to the first device, so that the first device obtains neural network parameter information by using the gliding recurrent neural network training according to the neural network setting information; The sliding recurrent neural network includes an input layer, at least one hidden layer and an output layer. During the training process, the input layer converts the serial input sequence into a sliding input sequence according to the time step information, and the at least one hidden layer performs balanced training based on the sliding input sequence. The serial input sequence is converted into a plurality of sliding input sequences by converting the serial input sequence into a sliding input sequence, each sliding input sequence contains the same number of input features, and the number of input features contained in each sliding input sequence corresponds to the size of the time step information.

8. The method according to claim 7, wherein: The method further comprises: Receiving the neural network parameter information sent by the first device; configuring the first device according to the neural network parameter information; sending a ready signal to the first device; Receive the payload sent by the first device.

9. A first device for signal equalization in a high-speed communication system in a first device, wherein: The first device comprises: means for sending a preamble to a second device; means for receiving neural network setting information sent by the second device, wherein the neural network setting information includes time step information and hidden layer size information; A device for obtaining neural network parameter information by using a gliding recurrent neural network training according to the neural network setting information, wherein the gliding recurrent neural network includes an input layer, at least one hidden layer and an output layer, during the training process, the input layer converts a serial input sequence into a gliding input sequence according to the time step information, and the at least one hidden layer performs equalization training based on the gliding input sequence; The serial input sequence is converted into a plurality of sliding input sequences by converting the serial input sequence into a sliding input sequence, each sliding input sequence contains the same number of input features, and the number of input features contained in each sliding input sequence corresponds to the size of the time step information.

10. The first device according to claim 9, wherein: The first device also includes: Means for sending the neural network parameter information obtained through training to the second device; Means for receiving a ready signal from the second device and sending a payload to the second device.

11. The first device according to claim 9 or 10, wherein: The at least one hidden layer includes one or more hidden layers having RNN neurons.

12. A second device for signal equalization in a high-speed communication system in a second device, wherein: The second device comprises: means for receiving a preamble from a first device; A device for measuring inter-symbol interference according to the preamble code and determining neural network setting information according to the inter-symbol interference, wherein the neural network setting information includes time step information and hidden layer size information; means for sending the neural network setting information to the first device so that the first device obtains neural network parameter information by using a gliding recurrent neural network training according to the neural network setting information; The sliding recurrent neural network includes an input layer, at least one hidden layer and an output layer. During the training process, the input layer converts the serial input sequence into a sliding input sequence according to the time step information, and the at least one hidden layer performs balanced training based on the sliding input sequence. The serial input sequence is converted into a plurality of sliding input sequences by converting the serial input sequence into a sliding input sequence, each sliding input sequence contains the same number of input features, and the number of input features contained in each sliding input sequence corresponds to the size of the time step information.

13. A method for signal equalization in a high-speed communication system, wherein: The method comprises: The first device sends a preamble code to the second device; The second device receives a preamble from the first device; The second device measures inter-symbol interference according to the preamble code, and determines neural network setting information according to the inter-symbol interference, wherein the neural network setting information includes time step information and hidden layer size information; The second device sends the neural network setting information to the first device; The first device receives the neural network setting information sent by the second device; The first device obtains neural network parameter information by using a gliding recurrent neural network training according to the neural network setting information, wherein the gliding recurrent neural network includes an input layer, at least one hidden layer and an output layer, and during the training process, the input layer converts a serial input sequence into a gliding input sequence according to the time step information, and the at least one hidden layer performs balanced training based on the gliding input sequence; The serial input sequence is converted into a plurality of sliding input sequences by converting the serial input sequence into a sliding input sequence, each sliding input sequence contains the same number of input features, and the number of input features contained in each sliding input sequence corresponds to the size of the time step information.

14. A first device, wherein: The first device comprises: A memory for storing one or more programs; One or more processors connected to the memory, When the one or more programs are executed by the one or more processors, the one or more processors are caused to perform the following operations: Sending a preamble to a second device; Receiving neural network setting information sent by the second device, wherein the neural network setting information includes time step information and hidden layer size information; According to the neural network setting information, neural network parameter information is obtained by using a sliding recurrent neural network training, wherein the sliding recurrent neural network includes an input layer, at least one hidden layer and an output layer, during the training process, the input layer converts a serial input sequence into a sliding input sequence according to the time step information, and the at least one hidden layer performs balanced training based on the sliding input sequence; The serial input sequence is converted into a plurality of sliding input sequences by converting the serial input sequence into a sliding input sequence, each sliding input sequence contains the same number of input features, and the number of input features contained in each sliding input sequence corresponds to the size of the time step information.

15. A second device, wherein: The second device comprises: A memory for storing one or more programs; One or more processors connected to the memory, When the one or more programs are executed by the one or more processors, the one or more processors are caused to perform the following operations: receiving a preamble from a first device; Measuring inter-symbol interference according to the preamble code, and determining neural network setting information according to the inter-symbol interference, wherein the neural network setting information includes time step information and hidden layer size information; Sending the neural network setting information to the first device, so that the first device obtains neural network parameter information by using the gliding recurrent neural network training according to the neural network setting information; The sliding recurrent neural network includes an input layer, at least one hidden layer and an output layer. During the training process, the input layer converts the serial input sequence into a sliding input sequence according to the time step information, and the at least one hidden layer performs balanced training based on the sliding input sequence. The serial input sequence is converted into a plurality of sliding input sequences by converting the serial input sequence into a sliding input sequence, each sliding input sequence contains the same number of input features, and the number of input features contained in each sliding input sequence corresponds to the size of the time step information.

Images