Equalization method based on enhanced reservoir calculation

Abstract

The invention relates to the technical field of optical communication, and particularly discloses an equalization method based on enhanced reserve pool calculation, and according to an enhanced reserve pool used in the method, a sliding window is added to an input layer of an original reserve pool, and output feedback is added to an output layer. The use of a sliding window enables the reserve pool to take into account inter-coding interference between adjacent symbols. Feedback is output so that trained readings are not only fed back to the reservoir, but also associated with readings at the next moment, which expands the functionality of the original reservoir so that it depends on fixed random weights from input to output, and can also dynamically adapt to tasks.

Description

Technical Field

[0001] The present invention relates to the field of optical communication technology, and in particular to a balancing method based on enhanced reserve pool calculation. Background Art

[0002] With the emergence and popularization of a large number of emerging applications such as cloud computing, the Internet of Things, high-definition television, and virtual reality, short-distance optical fiber communication systems covering metropolitan area networks, data center networks, and access networks have become the main carriers for big data transmission in the 5G and Internet of Things era. Although coherent transceivers have appeared for short-distance transmission, intensity modulation direct detection (IMDD) provides a more favorable trade-off between performance and complexity. Therefore, intensity modulation direct detection is still mainly used in short-distance optical systems. The main limitation of IMDD optical transmission is the power fading effect caused by the interaction between chromatic dispersion (CD) and direct detection, which limits its transmission distance and capacity. In addition, low-cost devices such as optical modulators, photodiodes (PDs), and transimpedance amplifiers (TIAs) also produce nonlinear distortion. In order to address these defects, various equalization techniques have been explored to compensate for dispersion and mitigate the impact of device nonlinearity in the system. Optical domain equalization techniques mainly include dispersion-compensating fiber, chirped optical Bragg gratings, ring resonators, etc. However, optical domain equalization schemes are less used due to their high complexity and high cost. Traditional equalization techniques in the electrical domain include linear feedforward equalizer (FFE), decision feedback equalizer (DFE), maximum likelihood sequence estimator (MLSE), and Volterra equalizer. Recently, neural network-based equalization techniques have received increasing attention worldwide. However, these neural network equalizers provide high equalization capabilities at the cost of complex network optimization. For example, they require backpropagation of the entire network to adjust all model parameters during the training phase, which is usually time-consuming and energy-consuming. Summary of the invention

[0003] In order to solve the problems existing in the prior art, the present invention provides a balancing method based on enhanced reserve pool calculation, which solves the problems mentioned in the above background technology.

[0004] To achieve the above object, the present invention provides the following technical solution: a balancing method based on enhanced reserve pool calculation, comprising the following steps:

[0005] S1, generate an OOK signal with a bandwidth of 32GBaud, and perform 2 times sampling on each symbol of the OOK signal;

[0006] S2, the sampled OOK signal is pulse shaped by a root raised cosine filter SRRC with a roll-off factor of 0.1;

[0007] S3, modulating the signal through a modulator to convert the electrical signal into an optical signal;

[0008] S4, injecting the optical signal into the optical channel composed of a standard single-mode optical fiber SSMF;

[0009] S5. At the receiving end, the photodetector PD square rate detection converts the optical signal into an electrical signal;

[0010] S6, performing matched filtering through a root raised cosine filter SRRC;

[0011] S7, using the trained enhanced reserve pool to equalize the signal;

[0012] S8. Finally, calculate the bit error rate of the equalized signal.

[0013] Preferably, in step S7, the enhanced reserve pool adds a sliding window to the input layer of the original reserve pool and adds output feedback to the output layer. The enhanced reserve pool includes three layers, namely, an input layer, a reservoir layer and an output layer.

[0014] Preferably, the number of neurons in the input layer, the reservoir layer and the output layer are M, N and L respectively; the reservoir layer, as the core structure of the enhanced reserve pool RC, is composed of a large number of sparse and randomly connected neurons with short-term memory capability, and adopts an N×N dimensional matrix W res To represent the connection status between neurons;

[0015] The state equation inside the reservoir is x[n] = α·f(W in ·[1;u[n]]+W res ·x[n-1]+W back ·[1;y[n-1]])+(1-α)·x[n-1], where α represents the leakage rate and f represents the activation function. Tanh is selected as the activation function; W in The dimension is N×(M+1), which is the weight matrix connecting the input layer and the reservoir; W back is the feedback weight matrix from the output layer to the reservoir, with a dimension of N×(L+1); the W in , W res , W back The weights of these three matrices are randomly generated and remain unchanged after generation;

[0016] The output y[n] of the reservoir is obtained by linearly transforming the input signal u[n], the reservoir state x[n], and the previous moment y[n-1] of the output signal; the output equation of the reservoir is in, The dimension is L×(M+1), indicating a direct connection between the input layer and the output layer; The weights and interconnections between the N reservoir states and the L network outputs are defined; The dimension is L×(L+1), which is the self-feedback weight matrix of the output layer, which associates the output signal y[n-1] at the previous moment with the output signal y[n] at the current moment.

[0017] Preferably, the sliding window is used before the signal enters the input layer. Adding the sliding window can make multiple symbols of the signal enter the reserve pool at the same time, and the sliding window is used to consider the inter-coding interference between adjacent symbols; the enhanced reserve pool includes a self-feedback weight matrix from the output layer to the output layer and the feedback weight matrix W from the output layer to the storage layer back , through these two feedbacks, the interference of subsequent symbols on the equalized symbols is reduced; the sliding window length is 3 symbol periods, the sliding window moves one symbol period each time, and one symbol period contains multiple samples. These samples are used to enable the neural network to learn the characteristics of the symbols.

[0018] Preferably, the enhanced reserve pool is trained, and the training includes:

[0019] Initialize the reservoir by setting its initial state to 0, W in and W back The weights in are obtained from a uniform distribution between -1 and 1, and the sparsity represents the proportion of non-zero elements in the reservoir matrix to the total elements; according to the size of the sparsity, a binary distributed reservoir weight matrix W that obeys a normal distribution is generated res , and rescaled according to the size of the spectral radius, the sparsity and spectral radius are set to 0.8 and 0.8 respectively;

[0020] In the training phase, only the output weights need to be trained, the mean square error MSE is defined as the target loss function, and the ridge regression RR least squares method is used to solve the linear regression problem. The training equation is:

[0021]

[0022] Where A = [A[1], A[2], …A[k]], A[n] = [1; u[n]; x[n]; 1; y[n-1]], n = [1, …k], k represents the size of the training set;

[0023] Target set matrix Y target =[Y target [1],Y target 2],…Y target [k]], λ is the ridge regularization factor, I is an identity matrix of size M+N+L+2;

[0024] Input signal u[n], according to the formula x[n] = α·f(W in ·[1;u[n]]+Wres ·x[n-1]+W back ·[1;y[n-1]])+(1-α)·x[n-1] and The output signal y[n] is calculated.

[0025] The beneficial effects of the present invention are as follows: the present invention proposes a short-distance transmission equalization method based on enhanced reserve pool calculation for an optical fiber transmission system of OOK modulated IMDD, and adding a sliding window at the input layer can take into account the inter-coding interference between adjacent symbols. Feedback is added at the output layer so that the trained readings are not only fed back to the reservoir, but also associated with the readings at the next moment, which expands the function of the original reserve pool. It no longer relies solely on fixed random weights from input to output, but can also dynamically adapt to tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a structural diagram of an enhanced storage pool in an embodiment of the present invention;

[0027] Figure 2 Schematic diagram of a sliding window in an embodiment of the present invention, (a) is a moving scheme of the sliding window, and (b) is a schematic diagram of the internal structure of the sliding window;

[0028] Figure 3 Flow chart of numerical simulation of the present invention, (a) is simulation configuration, (b) is spectrum response curve of simulation system;

[0029] Figure 4 Schematic diagram of the comparison between the enhanced storage tank of the present invention and the original storage tank under numerical simulation, (a) is 15 km, (b) is 25 km;

[0030] Figure 5 Flowchart of experimental verification of the present invention, (a) is the experimental configuration, (b) is the spectrum response curve of the simulation and experimental system;

[0031] Figure 6 Schematic diagram of the comparison of experiments and simulations between the enhanced reserve pool of the present invention and the original reserve pool when the number of hidden layers is 1800, (a) is 15 km, and (b) is 25 km. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] See also Figure 1-Figure 2 The present invention provides a technical solution: an equalization method based on enhanced reserve pool calculation, comprising: generating an OOK signal, sampling each symbol of the OOK signal, and pulse shaping the sampled OOK signal through a root raised cosine filter; converting the signal into an analog signal through a digital-to-analog converter DAC; modulating the signal through a modulator, converting the electrical signal into an optical signal, and injecting the optical signal into an optical channel composed of a standard single-mode optical fiber; converting the optical signal into an electrical signal at the receiving end through a photoelectric detector square rate detection, and converting the signal into a digital domain through an analog-to-digital converter ADC, and then matching filtering through a root raised cosine filter; equalizing the signal using a trained enhanced reserve pool to obtain an output signal, and finally calculating the bit error rate of the equalized signal; specifically comprising the following steps:

[0034] S1, generate an OOK signal with a bandwidth of 32GBaud, and perform 2 times sampling on each symbol of the OOK signal;

[0035] S2, the sampled OOK signal is pulse shaped by a root raised cosine filter SRRC with a roll-off factor of 0.1;

[0036] S3, modulating the signal through a modulator to convert the electrical signal into an optical signal;

[0037] S4, injecting the optical signal into the optical channel composed of a standard single-mode optical fiber SSMF;

[0038] S5. At the receiving end, the photodetector PD square rate detection converts the optical signal into an electrical signal;

[0039] S6, performing matched filtering through a root raised cosine filter SRRC;

[0040] S7, using the trained enhanced reserve pool to equalize the signal;

[0041] S8. Finally, calculate the bit error rate of the equalized signal.

[0042] Furthermore, the enhanced storage pool adds a sliding window to the input layer of the original storage pool and adds output feedback to the output layer. The enhanced storage pool includes three layers, such as Figure 1 As shown, that is, the input layer, reservoir (hidden layer) and output layer, all solid lines in the figure are randomly generated weights, and all dotted lines represent trained weights.

[0043] Furthermore, the number of neurons in the input layer, reservoir layer and output layer are M, N and L respectively; the reservoir layer, as the core structure of the enhanced reserve pool RC, is composed of a large number of sparse and randomly connected neurons with short-term memory capability, and defines an N×N dimensional matrix W resTo represent the connection status between neurons.

[0044] Reservoir status Updated at time n, the state equation inside the reservoir is x[n] = α·f(W in ·[1;u[n]]+W res ·x[n-1]+W back ·[1;y[n-1]])+(1-α)·x[n-1], where α represents the leakage rate and f represents the activation function. Here, tanh is selected as the activation function; W in The dimension is N×(M+1), which is the weight matrix connecting the input layer and the reservoir; W back is the feedback weight matrix from the output layer to the reservoir, with a dimension of N×(L+1); the bias components in the input layer and the output layer are represented by M+1 and L+1, respectively, to enhance the nonlinearity inside the reservoir. in , W res , W back The weights of these three matrices are randomly generated and remain unchanged after generation. The nonlinear dynamics of x[n] consists of three parts: the first part is the newly injected input information u[n]; the second part is the influence of the node state x[n-1] at the previous moment; and the third part is the feedback y[n-1] of the output signal at the previous moment.

[0045] Output of the reserve pool It is obtained by linearly transforming the input signal u[n], the reservoir state x[n], and the output signal y[n-1] at the previous moment; the output equation of the reservoir is: in, The dimension is L×(M+1), indicating a direct connection between the input layer and the output layer; The weights and interconnections between the N reservoir states and the L network outputs are defined; The dimension is L×(L+1), which is the self-feedback weight matrix of the output layer, which associates the output signal y[n-1] at the previous moment with the output signal y[n] at the current moment.

[0046] As shown in the internal state equation and output equation of the reservoir, the weight W is considered respectively back The impact of the output signal on the current state of the reservoir and its weight Impact on current output. are newly introduced parameters to be trained, which can be processed in the same training phase of the original reservoir. The trained readings are not only fed back to the reservoir, but also associated with the readings at the next moment. This extends the functionality of the original reservoir, as it no longer relies solely on fixed random weights from input to output, but can also dynamically adapt to the task. The enhanced reservoir can overcome the computational limitations of memory decay in the presence of feedback without having to fall into a chaotic state. In a sense, feedback gives them the general ability to simulate more accurate computations.

[0047] Further, such as Figure 2 As shown in the figure, the sliding window is used before the signal enters the input layer. Previously, the signal enters the reserve pool symbol by symbol. Adding a sliding window allows multiple symbols of the signal to enter the reserve pool at the same time. Fiber dispersion and square-law PD photoelectric conversion make the IMDD channel nonlinear and memristive, which means that inter-code interference is caused by the previous and next symbols. A sliding window is applied to consider the inter-code interference between adjacent symbols; considering that the enhanced reserve pool contains a self-feedback weight matrix from the output layer to the output layer and the feedback weight matrix W from the output layer to the storage layer back , through these two feedbacks, the interference of subsequent symbols on the equalization symbol can be reduced, so its equalization point is at the first position. Figure 2 As shown in (a), the sliding window length is 3 symbol periods, and the sliding window moves one symbol period each time. Figure 2 As shown in (b), a symbol period contains multiple samples, which help the neural network learn the characteristics of the symbol. These samples are used to enable the neural network to learn the characteristics of the symbol.

[0048] Furthermore, the enhanced reserve pool is trained. The training is divided into two stages, namely state selection and weight calculation. The training includes:

[0049] First, initialize the reservoir by setting its initial state to 0, W in and W back The weights in are drawn from a uniform distribution between -1 and 1, and the sparsity represents the proportion of non-zero elements in the reservoir matrix to the total elements. Based on the size of the sparsity, a binary distributed reservoir weight matrix W that follows a normal distribution is generated. res , and rescaled according to the size of the spectral radius. The sparsity and spectral radius are set to 0.8 and 0.8 respectively;

[0050] In the training phase, only the output weights need to be trained, the mean square error MSE is defined as the target loss function, and the ridge regression RR least squares method is used to solve the linear regression problem. The training equation is:

[0051]

[0052] Where A = [A[1], A[2], …A[k]], A[n] = [1; u[n]; x[n]; 1; y[n-1]], n = [1, …k], k represents the size of the training set;

[0053] Target set matrix Y target =[Y target [1],Y target [2],…Y target [k]], λ is the ridge regularization factor, I is an identity matrix of size M+N+L+2;

[0054] Input signal u[n], according to the formula x[n] = α·f(W in ·[1;u[n]]+W res ·x[n-1]+W back ·[1;y[n-1]])+(1-α)·x[n-1] and The output signal y[n] is calculated.

[0055] Numerical simulation and results

[0056] like Figure 3 As shown in (a), the numerical simulation is performed using the method of the present invention, and the simulation configuration is as follows Figure 3 As shown in (a), Figure 3 As shown in (b), the system spectrum response shows that a frequency attenuation zero point appears at a transmission distance of 25 km. In order to numerically compare the performance of the enhanced reserve pool in calculating the equalized signal, the original reserve pool is used as the object of comparison. The number of neurons in the hidden layer of the original reserve pool is consistent with the number of neurons in the enhanced reserve pool. Figure 4 As shown, Figure 4 (a) and (b) show the bit error rate variation at different received optical powers for 15 km and 25 km transmission respectively. The star curve is an optical back-to-back transmission without dispersion effect. It can be seen that for 15 km transmission, when the number of reservoir neurons Nr = 300, the performance of both equalizers has a receiver sensitivity drop of about 1 dB. When Nr is increased to 1800, the performance of both reserve pool-based equalizers is close to that of optical back-to-back transmission. When the dispersion effect is strong at 25 km, the enhanced reserve pool equalization performance is better than the original reserve pool. When the number of neurons Nr = 300, the enhanced reserve pool has a receiver sensitivity improvement of more than 2 dB compared to the original reserve pool. This is because the enhanced reserve pool contains a weight matrix and W backThe feedback scheme can learn more features of the training information to overcome the computational limitations of memory decay. In addition, increasing the number of hidden layer neurons can improve the balancing performance of the reserve pool. By increasing the number of neurons from 300 to 1800, the original reserve pool shows similar performance to the enhanced reserve pool with only 300 neurons.

[0057] Experimental verification and results

[0058] like Figure 5 (a) shows, the corresponding experimental configuration is as follows Figure 5 As shown in (a), the DSP at the transmitting end is performed offline in Matlab software to generate a 32G baud rate OOK signal, which is then upsampled and filtered. An arbitrary waveform generator is used for digital-to-analog conversion. The electrical signal then passes through a Mach-Zehnder modulator to generate an optical signal with a power of 0.1dBm. Passing through a single-mode optical fiber, the received optical power is controlled by a variable optical attenuator and an erbium-doped fiber amplifier to achieve the ideal optical power. At the receiving end, a photodetector with a 3dB bandwidth of 50GHz is used to convert the optical signal into an electrical signal, which is then sampled at 80GSa / s by a digital storage oscilloscope for analog-to-digital conversion. In the receiving end DSP, signal equalization and downsampling are performed offline. As shown Figure 5 As shown in (b), it can be seen from the figure that the 15km transmission shows a small dispersion effect, and the dispersion after 25km transmission causes the signal spectrum to periodically attenuate.

[0059] The experiment further verified the balancing performance of the enhanced reserve pool. Figure 6 As shown, Figure 6 (a) and (b) show the experimental results of bit error rate and received optical power of the enhanced reserve pool and the original reserve pool at 15 km and 25 km transmission respectively when the number of hidden layers is 1800. At 15 km, due to the weak dispersion effect, the performance of the enhanced reserve pool and the original reserve pool is similar. Figure 5 As shown in (b), although the experimental and simulated channel responses show the shape of an analog low-pass filter, the simulated curve is flatter and the amplitude attenuation within the signal bandwidth is smaller. Therefore, the required received optical power in the experiment is 1dB higher than the simulation to reach the 7% FEC threshold. At the same time, the frequency zero point at 25km is about 2GHz lower than the simulation. This frequency shift may be caused by the nonlinearity of the device in the experimental transmission. However, this effect can be mitigated by an enhanced reserve pool. As shown in Figure 6 As shown in (b), after enhanced reserve pool equalization, the experimental performance of dispersion effect and nonlinearity shows a similar trend to the simulation involving only dispersion effect. However, the original reserve pool cannot equalize this nonlinearity, and the experimental receiver sensitivity is 1dB higher than the simulation.

[0060] In several embodiments provided in the embodiments of the present invention, it should be understood that the disclosed apparatus and method can also be implemented in other ways. The apparatus and method embodiments described above are merely schematic. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the apparatus, method and computer program product according to multiple embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, program segment or a part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0061] In addition, the functional modules in the various embodiments of the present invention may be integrated together to form an independent part, or each module may exist independently, or two or more modules may be integrated to form an independent part.

[0062] If the function is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, an electronic device, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, RandomAccess Memory), a disk or an optical disk. It should be noted that, in this article, the term "include", "comprise" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements that are not explicitly listed, or also includes elements inherent to such a process, method, article or device. Without more constraints, an element defined by the phrase "comprising a..." does not exclude the existence of other identical elements in the process, method, article or apparatus comprising the element.

[0063] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The singular forms "a", "said" and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings.

[0064] It should be understood that the term "and / or" used in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0065] The word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to determining" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)", depending on the context.

[0066] The "first\second" mentioned in the embodiments is only to distinguish similar objects, and does not represent a specific order for the objects. It is understandable that the "first\second" can be interchanged with the specific order or sequence where permitted. It should be understood that the objects distinguished by "first\second" can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than those illustrated or described herein.

[0067] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A balancing method based on enhanced reserve pool calculation, characterized in that: The steps include: S1, generate an OOK signal with a bandwidth of 32GBaud, and perform 2 times sampling on each symbol of the OOK signal; S2, the sampled OOK signal is pulse shaped by a root raised cosine filter SRRC with a roll-off factor of 0.1; S3, modulating the signal through a modulator to convert the electrical signal into an optical signal; S4, injecting the optical signal into the optical channel composed of a standard single-mode optical fiber SSMF; S5. At the receiving end, the photodetector PD square rate detection converts the optical signal into an electrical signal; S6, performing matched filtering through a root raised cosine filter SRRC; S7, using the trained enhanced reserve pool to equalize the signal; S8. Finally, calculate the bit error rate of the equalized signal.

2. The balancing method based on enhanced reserve pool calculation according to claim 1, characterized in that: In step S7, the enhanced reservoir adds a sliding window to the input layer of the original reservoir and adds output feedback to the output layer. The enhanced reservoir includes three layers, namely, an input layer, a reservoir layer, and an output layer.

3. The balancing method based on enhanced reserve pool calculation according to claim 2, characterized in that: The number of neurons in the input layer, reservoir and output layer are M, N and L respectively; the reservoir, as the core structure of the enhanced reserve pool RC, is composed of a large number of sparse and randomly connected neurons with short-term memory capability, using an N×N dimensional matrix W res To represent the connection status between neurons; The state equation inside the reservoir is x[n] = α·f(W in ·[1;u[n]]+W res ·x[n-1]+W back ·[1;y[n-1]])+(1-α)·x[n-1], where α represents the leakage rate and f represents the activation function. Tanh is selected as the activation function; W in The dimension is N×(M+1), which is the weight matrix connecting the input layer and the reservoir; W back is the feedback weight matrix from the output layer to the reservoir, with a dimension of N×(L+1); the W in , W res , W back The weights of these three matrices are randomly generated and remain unchanged after generation; The output y[n] of the reservoir is obtained by linearly transforming the input signal u[n], the reservoir state x[n], and the previous moment y[n-1] of the output signal; the output equation of the reservoir is in, The dimension is L×(M+1), indicating a direct connection between the input layer and the output layer; The weights and interconnections between the N reservoir states and the L network outputs are defined; The dimension is L×(L+1), which is the self-feedback weight matrix of the output layer, which associates the output signal y[n-1] at the previous moment with the output signal y[n] at the current moment.

4. The balancing method based on enhanced reserve pool calculation according to claim 2, characterized in that: The sliding window is used before the signal enters the input layer. Adding the sliding window allows multiple symbols of the signal to enter the reserve pool at the same time. The sliding window is used to consider the inter-coding interference between adjacent symbols. The enhanced reserve pool contains a self-feedback weight matrix from the output layer to the output layer. and the feedback weight matrix W from the output layer to the storage layer back , through these two feedbacks, the interference of subsequent symbols on the equalized symbols is reduced; the sliding window length is 3 symbol periods, the sliding window moves one symbol period each time, and one symbol period contains multiple samples. These samples are used to enable the neural network to learn the characteristics of the symbols.

5. The balancing method based on enhanced reserve pool calculation according to claim 1, characterized in that: The enhanced reserve pool is trained, including: Initialize the reservoir by setting its initial state to 0, W in and W back The weights in are obtained from a uniform distribution between -1 and 1, and the sparsity represents the proportion of non-zero elements in the reservoir matrix to the total elements; according to the size of the sparsity, a binary distributed reservoir weight matrix W that obeys a normal distribution is generated res , and rescaled according to the size of the spectral radius, the sparsity and spectral radius are set to 0.8 and 0.8 respectively; In the training phase, only the output weights need to be trained, the mean square error MSE is defined as the target loss function, and the ridge regression RR least squares method is used to solve the linear regression problem. The training equation is: Where A = [A[1], A[2], …A[k]], A[n] = [1; u[n]; x[n]; 1; y[n-1]], n = [1, …k], k represents the size of the training set; Target set matrix Y target =[Y target [1],Y target [2],…Y target [k]], λ is the ridge regularization factor, I is an identity matrix of size M+N+L+2; Input signal u[n], according to the formula x[n] = α·f(W in ·[1;u[n]]+W res ·x[n-1]·+W back ·[1;y[n-1]])+(1-α)·x[n-1] and The output signal y[n] is calculated.

Images