An optical fiber nonlinear equalization method, system, device and storage medium

Abstract

The present invention discloses a fiber optic nonlinear equalization method, system, device and storage medium, belonging to the field of optical transmission technology. The method includes: obtaining the M-QAM signal sequence received at the receiving end; inputting the M-QAM signal sequence into a trained bidirectional long short-term memory neural network based on a memristor array to output a score matrix; using the Viterbi algorithm to obtain the optimal label sequence corresponding to each input feature sequence according to the score matrix and the training score matrix obtained during the training process. The intermediate label values of all the optimal label sequences form the label values of the M-QAM signal sequence, and the data categories corresponding to the label values are subjected to constellation mapping to obtain the original signal sequence transmitted by the transmitting end, thus completing the nonlinear equalization. The present invention uses a memristor array to implement synaptic weights, greatly reducing the complexity of the network, and uses the bidirectional feature to process the current input information by combining the information of the previous and subsequent sequences, improving the signal quality and transmission performance.

Description

Technical Field

[0001] The present invention relates to a method, system, device and storage medium for optical fiber nonlinear equalization, belonging to the technical field of optical transmission. Background Art

[0002] Since the invention of optical fiber communication technology in the 1960s, due to its unique characteristics, optical fiber communication technology has become the only communication system that can support global communication and has developed rapidly. In today's communication networks, optical fiber communication has become an essential part. With people's demand for high-bandwidth services, the requirements for information transmission links are also getting higher and higher, and great efforts are needed to cope with the exponentially growing capacity demand brought by the next-generation mobile networks and high-bandwidth Internet applications. The optical fiber communication community predicted a surge in capacity demand a decade ago and began to focus on researching technologies in this area. Now, coherent optical communication technology has become one of the main communication technologies in the backbone network.

[0003] In recent years, communication technologies have been continuously updated, data capacities have been continuously expanded, distances have been continuously increased, and the nonlinearity in optical fibers has become increasingly prominent. After 20 years of high-level research, several solutions have been determined, including combinations of higher-order modulation formats, including probability and / or geometry, and wavelength division multiplexing (WDM)-assisted spatial division multiplexing or bandwidth expansion to other bands such as the o-band. Therefore, no matter what method is used to increase the channel capacity, the ultimate main limiting factor is still the nonlinear Shannon transmission limit. In long-distance high-bandwidth optical networks, this limitation is mainly due to the nonlinearity formed by Kerr-induced optical fibers within and between channels and their interaction with the amplified spontaneous emission noise amplified by cascaded optical amplifiers. Currently, the main methods for nonlinear equalization are optical phase compensation (OPC) in relays, digital backpropagation algorithm (DBP), and inverse Volterra series transfer function (IVSTF).

[0004] The application of OPC requires wavelength converters and is not compatible with dynamically configured elastic optical networks; DBP has high complexity, and it is unrealistic to use it to simulate multi-channel channel transmission; IVSTF is essentially to alleviate the nonlinearity within the channel and does not meet the requirements of the current development of multi-channel optical communication. Summary of the Invention

[0005] The purpose of the present invention is to provide a method, system, device and storage medium for optical fiber nonlinear equalization to solve the problems of high complexity and low adaptability in the prior art.

[0006] To achieve the above purpose, the present invention is implemented by adopting the following technical solutions:

[0007] In the first aspect, the present invention provides a method for optical fiber nonlinear equalization, including:

[0008] Obtain the M-QAM signal sequence received by the receiving end;

[0009] Input the M-QAM signal sequence into the trained bidirectional long short-term memory neural network based on the memristive array, and output the score matrix;

[0010] According to the score matrix and the training score matrix obtained during the training process, use the Viterbi algorithm to obtain the optimal label sequence corresponding to each input feature sequence. The middle label values of all the optimal label sequences form the label values of the M-QAM signal sequence. Map the data categories corresponding to the label values to obtain the original signal sequence sent by the sending end, and complete the non-linear equalization.

[0011] Combined with the first aspect, further, the bidirectional long short-term memory neural network based on the memristive array is trained by the following method:

[0012] Construct a training data set with real labels;

[0013] Divide the training data set into several batches, and perform batch training on the bidirectional long short-term memory neural network based on the memristive array. When training, use the Adam optimization algorithm to optimize the parameters in the network.

[0014] Combined with the first aspect, further, the bidirectional long short-term memory neural network based on the memristive array includes an input layer, a hidden layer, and an output layer. In the input layer, extract the input feature sequences of each symbol in the M-QAM signal sequence, send each input feature sequence to the hidden layer to be processed by the bidirectional long short-term memory unit, and convert the processing result into a score matrix. The output layer outputs the score matrix.

[0015] Combined with the first aspect, further, extracting the input feature sequences of each symbol in the M-QAM signal sequence in the input layer includes:

[0016] For the i-th symbol in the M-QAM signal sequence, combine the i-th symbol with k symbols before and after the i-th symbol to obtain the input feature sequence of the i-th symbol.

[0017] Combined with the first aspect, further, the bidirectional long short-term memory unit includes an input gate, a forget gate, and an output gate;

[0018] The input gate is used to control the proportion of the input information at the current moment to be recorded into the memory state;

[0019] The forget gate is used to control the proportion of the memory state at the previous moment to be included in the memory state at the current moment;

[0020] The output gate is used to control the proportion of the memory state at the current moment to be recorded into the hidden layer state at the current moment;

[0021] The equation representation of the bidirectional long short-term memory unit is as follows:

[0022]

[0023] Among them, is the initial proportional input value of the input gate, is the proportional value of the previous moment of the input gate, is the proportional input value of the forget gate, is the proportional input value of the output gate, x t is the input information at the current moment, h t is the hidden layer state at the current moment, h t-1 is the hidden layer state of the previous moment, W a 、W i 、W f 、W o are the weight parameter matrices of the input signal at the current moment, U a 、U i 、U f 、U o are the weight parameter matrices of the hidden layer state of the previous moment, b a 、b i 、b f 、b o are the bias parameter matrices;

[0024]

[0025]

[0026] Among them, σ() is the Sigmoid function, is the Hadamard product, c t-1 is the memory state of the previous moment.

[0027] In the second aspect, the present invention also provides an optical fiber nonlinear equalization system, including:

[0028] A signal acquisition module: used to acquire the M-QAM signal sequence received by the receiving end;

[0029] A signal processing module: used to input the M-QAM signal sequence into the trained bidirectional long short-term memory neural network based on the memristor array and output a score matrix;

[0030] A signal conversion module: used to obtain the optimal label sequence corresponding to each input feature sequence using the Viterbi algorithm according to the score matrix and the training score matrix obtained during the training process. The intermediate label values of all the optimal label sequences form the label values of the M-QAM signal sequence, and the data categories corresponding to the label values are subjected to constellation mapping to obtain the original signal sequence transmitted by the transmitting end, thereby completing nonlinear equalization.

[0031] In a third aspect, the present invention further provides an optical fiber non - linear equalization device, including a processor and a storage medium;

[0032] The storage medium is used for storing instructions;

[0033] The processor is configured to operate according to the instructions to execute the steps of the method according to any one of the first aspect.

[0034] In a fourth aspect, the present invention further provides a computer - readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the method according to any one of the first aspect are implemented.

[0035] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0036] An optical fiber non - linear equalization method, system, device and storage medium provided by the present invention, through a bidirectional long - short - term memory neural network constructed by a memristor array. The memristor has a two - terminal storage function. It performs calculations through physical laws at the same position where information is stored. In terms of power consumption and inference latency, it has significant advantages compared with the corresponding devices of CMOS (Complementary Metal - Oxide - Semiconductor). The bidirectional long - short - term memory neural network based on the memristor array is an artificial intelligence neural network with high efficiency. It can not only achieve the non - linear equalization of optical fibers, but also use the memristor array to implement synaptic weights, which can greatly reduce the complexity of the network, store a large number of parameters, thereby improving the memory computing ability. Under the great development of future integrated circuits, it can also realize the construction of neural networks for actual circuits; the bidirectional long - short - term memory units in the bidirectional long - short - term memory neural network utilize the bidirectional characteristics to simultaneously process pre - order and post - order information, respectively using two long - short - term memory units to capture historical information and future information, realizing the processing of the current input information by combining pre - order and post - order information, significantly improving the signal quality and realizing the improvement of the system transmission performance. Utilizing the bidirectional characteristics can also improve the efficiency and accuracy of data processing and can effectively handle the inter - symbol interference caused by non - linearity; in the present invention, the memristors are formed into an array to form a memristive circuit with multiple rows and columns. By training and inferring the bidirectional long - short - term memory neural network based on the memristor array, regression and classification problems are solved. During the training and inference process, all matrix multiplications and updates are physically implemented on the memristor crossbar interfacing with digital computing. This solution not only enhances the anti - interference ability of the entire optical transmission system and improves the transmission performance of the system, but also can be combined with other multi - carrier multiplexing technologies to improve the adaptability of the present invention and maximize the performance of the optical communication system. Description of the Drawings

[0037] Figure 1 It is a flowchart of an optical fiber nonlinear equalization method provided by an embodiment of the present invention;

[0038] Figure 2 It is a schematic diagram of a coherent optical communication system provided by an embodiment of the present invention;

[0039] Figure 3 It is a structural diagram of a memristive array provided by an embodiment of the present invention;

[0040] Figure 4 It is a schematic diagram of a bidirectional long short-term memory neural network based on a memristive array provided by an embodiment of the present invention. Specific implementation manners

[0041] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and cannot be used to limit the protection scope of the present invention.

[0042] Embodiment 1

[0043] An optical fiber nonlinear equalization method provided by the present invention is applied in a large-capacity long-distance coherent optical communication system in this embodiment (the principle is as Figure 2 shown). Using DSP technology at the receiving end can effectively perform equalization and compensation, thereby improving the signal quality and enhancing the transmission performance of the coherent optical communication system; it mainly includes the following steps: orthogonalization and normalization of I / Q signals and imbalance compensation, dispersion compensation, polarization mode dispersion compensation, frequency offset estimation, carrier phase recovery, nonlinear equalization, and system performance analysis.

[0044] The principle of the coherent optical communication system is as Figure 2 shown. The transmitting end consists of an external cavity laser (ECL), an IQ modulator, and an arbitrary waveform generator (AWG). The modulated signal is transmitted through an optical fiber link and an erbium-doped fiber amplifier (EDFA) and then reaches the receiving end. At the receiving end, after the local oscillator signal (LO) and the received signal are mixed by a 90° optical mixer, photoelectric detection is then performed, and finally digital signal processing (DSP) demodulation is carried out to recover the original signal.

[0045] At the receiving end DSP, a neural network algorithm is used for nonlinear processing to improve the signal quality and optimize the system transmission performance. In the nonlinear equalization part, it is implemented in the form of a bidirectional long short-term memory neural network based on a memristive array. The weight values of the previous moment and the current moment are input to the hidden layer for weight ratio calculation, and then the weight values of each ratio are input to the bidirectional long short-term memory unit for predicting and equalizing the data of the future moment, and finally the required equalized value is output.

[0046] There are approximately 10 in the human brain 14-15neuronal synapses and 10 11 neurons. They are all interconnected by neuronal synapses, and the topological structure formed by neurons and synapses endows the neural network with the potential to process a vast amount of information. In this patent, different information is converted into voltage signals and loaded onto different array wires, causing the memristance values of the corresponding memristors to change. The information in the form of voltage is transformed into memristance values and stored in the crossbar array. Thus, it can be seen that the memristor crossbar array circuit structure can store important weight information. In this structure, the vertical and horizontal wires of the memristor crossbar can represent different discrete values of input and output, and the two-dimensional matrix formed by the memristance values of all the memristors in the memristor crossbar in sequence can represent logical relationships. In this patent, the memory property of the core sequence of the bidirectional long short-term memory unit at different times is realized by the memristor array, and its structure is as Figure 3 shown.

[0047] For multiple shapes, any logical membership function can be realized by a spin memristor crossbar array. Assume that μ() is the domain name of the fuzzy set, representing the system function. This patent constructs a spin memristor crossbar array structure, where x i is the system input value, and μ(x i ) is the system output value. Define A() and B() as the corresponding membership functions, and x is the input signal. Sampling and quantization are performed with a finite resolution. Under most lossless conditions, the input signal passes through a quantizer to obtain quantized information, such that the quantization interval is equal to 1. The input variable is a discrete integer value, i.e., the high and low levels of the signal, and the corresponding membership function is also a discrete value. This patent uses a 5×5 memristor array, and each column in the memristor array represents the value of the input variable. The memristance value of the memristor at the cross point should be set to the corresponding value next to the cross point.

[0048] As Figure 1 shown, a fiber optic nonlinear equalization method provided by an embodiment of the present invention includes the following steps:

[0049] S1. Obtain the M-QAM signal sequence received at the receiving end.

[0050] In this embodiment, the signal sequence is a 16-QAM signal sequence.

[0051] S2. Input the M-QAM signal sequence into the trained bidirectional long short-term memory neural network based on the memristor array to output a score matrix.

[0052] In this embodiment, the structure of the bidirectional long short-term memory neural network based on the memristor array is as Figure 4 shown, where the current time is the t-th moment, the current input information is x t , the state of the hidden layer at the previous moment is h t-1 , is a bidirectional long short-term memory unit, and the memory state at the previous moment is c t-1 , c t is the memory state updated at time t, h t is the hidden layer state at the current moment, that is, the output of the bidirectional long short-term memory unit at the current moment.

[0053] The bidirectional long short-term memory unit is mainly composed of three gates: an input gate, a forget gate, and an output gate. Different gate structures have different functions. The input gate is used to control the proportion of the input information x t recorded in the memory state c t The forget gate is used to control the proportion of the memory state c t-1 recorded in the memory state c t at the current moment; the output gate is used to control the proportion of the memory state c t recorded in the hidden layer state h t at the current moment;

[0054] Its equation representation is as follows:

[0055]

[0056] Among them, is the initial proportion input value of the input gate, is the proportion value of the input gate at the previous moment, is the forget gate proportion input value, is the output gate proportion input value, x t is the input information at the current moment, h t is the hidden layer state at the current moment, h t-1 is the hidden layer state at the previous moment, W a , W i , W f , W o is the weight parameter matrix of the input signal at the current moment, U a , U i , U f , U o is the weight parameter matrix of the hidden layer state at the previous moment, b a , b i , b f , b o is the bias parameter matrix;

[0057]

[0058]

[0059] Among them, σ() is the Sigmoid function, is the Hadamard product, c t-1 is the memory state at the previous moment.

[0060] Perform non - linear equalization on the received 16 - QAM, perform a labeling task on the sequence. The bidirectional long - short - term memory neural network combined with the memristive array can significantly improve the sequence labeling performance. Send the input feature sequence to the bidirectional long - short - term memory neural network of the memristive array. The forward long - short - term memristive array bidirectional long - short - term memory unit processes the sequence from the front end of the input sequence to the back, and the backward process is to process the data from the end of the data forward.

[0061] In the hidden layer, the output result of the bidirectional long - short - term memory neural network is transformed into the label space; the constellation diagram of the M - QAM modulation format signal has M standard constellation points. According to these M standard constellation points, the M - QAM signal is divided into M different data categories, and each standard constellation point corresponds to a data category, numbered from 1 to M as the label of the data category. Therefore, the dimension of the original label space is M - dimensional. Since in the label set, the special start label and end label of the sequence should also be added, the result output by the hidden layer is a score matrix obtained by transforming according to the output result of the bidirectional long - short - term memory neural layer. The constellation diagram of the 16 - QAM modulation format has 16 standard constellation points, and the signal is divided into 16 different data categories according to these 16 standard constellation points.

[0062] The bidirectional long - short - term memory neural network based on the memristive array is trained by the following method:

[0063] Construct a training data set with real labels;

[0064] Divide the training data set into several batches, and perform batch training on the bidirectional long - short - term memory neural network based on the memristive array. When training, use the Adam optimization algorithm to optimize the parameters in the network.

[0065] The bidirectional long - short - term memory neural network based on the memristive array includes an input layer, a hidden layer, and an output layer. The processing flow of the M - QAM signal sequence is as follows: In the input layer, extract the input feature sequence of each symbol in the M - QAM signal sequence, send each input feature sequence to the hidden layer for processing by the bidirectional long - short - term memory unit, and convert the processing result into a score matrix. The output layer outputs the score matrix.

[0066] Extract the input feature sequence of each symbol in the M - QAM signal sequence in the input layer, including:

[0067] For the i - th symbol in the M - QAM signal sequence, combine the i - th symbol with k symbols before and after the i - th symbol to obtain the input feature sequence of the i - th symbol.

[0068] S3. Using the Viterbi algorithm with the score matrix and the training score matrix obtained during the training process, obtain the optimal label sequence corresponding to each input feature sequence. The intermediate label values of all the optimal label sequences form the label values of the M-QAM signal sequence. Perform constellation mapping on the data categories corresponding to the label values to obtain the original signal sequence transmitted by the transmitter (i.e., the original M-QAM signal sequence transmitted by the transmitter), thus completing the nonlinear equalization.

[0069] The present invention proposes a coherent optical communication system based on a memristor array bidirectional long short-term memory neural network. Its bidirectional characteristics can improve the efficiency and accuracy of data processing and can effectively handle the inter-symbol interference caused by nonlinearity. The memristors are integrated into an array to form a memristive circuit with multiple rows and columns. By connecting a fully connected network to a recurrent LSTM (Long short-term memory) network, this multi-layer network based on LSTM is trained and inferred to solve regression and classification problems. During the training and inference processes, all matrix multiplications and updates are physically implemented on the memristor crossbar interfaced with digital computing. This solution not only enhances the anti-interference ability of the system and improves the transmission performance of the system, but also can be combined with other multi-carrier multiplexing technologies to maximize the performance of the optical communication system. However, traditional LSTM models have high complexity and high computational load. Due to limited memory capacity and limited data communication, there are bottlenecks in the computational power of a large number of parameters. Implementing with a memristor array and using the memristor array to implement synaptic weights in LSTM can greatly reduce the complexity of the network and can store a large number of parameters, thereby improving the in-memory computing ability.

[0070] Embodiment 2

[0071] An optical fiber nonlinear equalization system provided by an embodiment of the present invention includes:

[0072] A signal acquisition module: used to acquire the M-QAM signal sequence received by the receiver;

[0073] A signal processing module: used to input the M-QAM signal sequence into a trained bidirectional long short-term memory neural network based on a memristor array and output a score matrix;

[0074] A signal conversion module: used to obtain the optimal label sequence corresponding to each input feature sequence using the Viterbi algorithm with the score matrix and the training score matrix obtained during the training process. The intermediate label values of all the optimal label sequences form the label values of the M-QAM signal sequence. Perform constellation mapping on the data categories corresponding to the label values to obtain the original signal sequence transmitted by the transmitter, thus completing the nonlinear equalization.

[0075] Embodiment 3

[0076] An optical fiber nonlinear equalization device provided by an embodiment of the present invention includes a processor and a storage medium;

[0077] The storage medium is used to store instructions;

[0078] The processor is used to operate according to the instructions to execute the steps of the following method:

[0079] Obtain the M-QAM signal sequence received by the receiving end;

[0080] Input the M-QAM signal sequence into a trained bidirectional long short-term memory neural network based on a memristive array, and output a score matrix;

[0081] Use the Viterbi algorithm to obtain the optimal label sequence corresponding to each input feature sequence according to the score matrix and the training score matrix obtained during the training process. The intermediate label values of all the optimal label sequences form the label value of the M-QAM signal sequence. Perform constellation mapping on the data categories corresponding to the label values to obtain the original signal sequence sent by the sending end, and complete the nonlinear equalization.

[0082] Embodiment 4

[0083] A computer-readable storage medium provided by an embodiment of the present invention stores a computer program thereon. When the program is executed by a processor, it implements the steps of the following method:

[0084] Obtain the M-QAM signal sequence received by the receiving end;

[0085] Input the M-QAM signal sequence into a trained bidirectional long short-term memory neural network based on a memristive array, and output a score matrix;

[0086] Use the Viterbi algorithm to obtain the optimal label sequence corresponding to each input feature sequence according to the score matrix and the training score matrix obtained during the training process. The intermediate label values of all the optimal label sequences form the label value of the M-QAM signal sequence. Perform constellation mapping on the data categories corresponding to the label values to obtain the original signal sequence sent by the sending end, and complete the nonlinear equalization.

[0087] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0088] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in one or more flows and / or blocks. Figure 1 one or more flows and / or blocks Figure 1 or a device for implementing the functions specified in one or more blocks.

[0089] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device that implements the functions specified in one or more flows and / or blocks. Figure 1 one or more flows and / or blocks Figure 1 or a device for implementing the functions specified in one or more blocks.

[0090] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more flows and / or blocks. Figure 1 one or more flows and / or blocks Figure 1 or a device for implementing the functions specified in one or more blocks.

[0091] The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the technical principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A fiber optic nonlinear equalization method, characterized in that, Including: Obtain the M-QAM signal sequence received by the receiving end; Input the M-QAM signal sequence into the trained bidirectional long short-term memory neural network based on a memristor array to output a score matrix; According to the score matrix and the training score matrix obtained during the training process, use the Viterbi algorithm to obtain the optimal label sequence corresponding to each input feature sequence. The middle label values of all the optimal label sequences form the label values of the M-QAM signal sequence. Perform constellation mapping on the data categories corresponding to the label values to obtain the original signal sequence sent by the sending end, completing non-linear equalization; The bidirectional long short-term memory neural network based on a memristor array includes an input layer, a hidden layer, and an output layer. For the i-th symbol in the M-QAM signal sequence, combine the i-th symbol with k symbols before and after the i-th symbol to obtain the input feature sequence of the i-th symbol. Send each input feature sequence to the hidden layer to be processed by bidirectional long short-term memory units, and convert the processing result into a score matrix. The output layer outputs the score matrix; The bidirectional long short-term memory unit includes an input gate, a forget gate, and an output gate; The input gate is used to control the proportion of the input information at the current moment to be recorded into the memory state; The forget gate is used to control the proportion of the memory state at the previous moment to be included in the memory state at the current moment; The output gate is used to control the proportion of the memory state at the current moment to be recorded into the hidden layer state at the current moment; The equation representation of the bidirectional long short-term memory unit is as follows: Among them, is the initial proportional input value of the input gate, is the proportional value of the input gate at the previous moment, is the proportional input value of the forget gate, is the proportional input value of the output gate, x t is the input information at the current moment, h t is the hidden layer state at the current moment, h t-1 is the hidden layer state at the previous moment, W a 、W i 、W f 、W o are the weight parameter matrices of the input signal at the current moment, U a 、U i 、U f 、U o are the weight parameter matrices of the hidden layer state at the previous moment, b a 、b i 、b f 、b o are the bias parameter matrices; where σ() is the Sigmoid function, ⊙ is the Hadamard product, and c t-1 is the memory state at the previous moment.

2. The optical fiber nonlinear equalization method according to claim 1, characterized in that, The bidirectional long short-term memory neural network based on a memristor array is trained by the following method: Construct a training data set with real labels; Divide the training data set into several batches, and perform batch training on the bidirectional long short-term memory neural network based on a memristor array. When training, use the Adam optimization algorithm to optimize the parameters in the network.

3. A fiber optic nonlinear equalization system, characterized in that, Including: Signal acquisition module: used to obtain the M-QAM signal sequence received by the receiving end; Signal processing module: used to input the M-QAM signal sequence into the trained bidirectional long short-term memory neural network based on a memristor array to output a score matrix; Signal conversion module: used to obtain the optimal label sequence corresponding to each input feature sequence according to the score matrix and the training score matrix obtained during the training process. The middle label values of all the optimal label sequences form the label values of the M-QAM signal sequence. Perform constellation mapping on the data categories corresponding to the label values to obtain the original signal sequence sent by the sending end, completing non-linear equalization; Among them, the bidirectional long short-term memory neural network based on a memristor array includes an input layer, a hidden layer, and an output layer. For the i-th symbol in the M-QAM signal sequence, combine the i-th symbol with k symbols before and after the i-th symbol to obtain the input feature sequence of the i-th symbol. Send each input feature sequence to the hidden layer to be processed by bidirectional long short-term memory units, and convert the processing result into a score matrix. The output layer outputs the score matrix; The bidirectional long short-term memory unit includes an input gate, a forget gate, and an output gate; The input gate is used to control the proportion of the input information at the current moment to be recorded into the memory state; The forget gate is used to control the proportion of the memory state at the previous moment that is included in the memory state at the current moment; The output gate is used to control the proportion of the memory state at the current moment that is included in the hidden layer state at the current moment; The equation representation of the bidirectional long short-term memory unit is as follows: Among them, is the initial proportional input value of the input gate, is the previous moment's proportional value of the input gate, is The forgetting gate ratio input value, is the output gate ratio input value, x t is the input information at the current moment, h t is the hidden layer state at the current moment, h t-1 is the hidden layer state at the previous moment, W a 、W i 、W f 、W o is the weight parameter matrix of the input signal at the current moment, U a 、U i 、U f 、U o is the weight parameter matrix of the hidden layer state at the previous moment, b a 、b i 、b f 、b o is the bias parameter matrix; Among them, σ() is the Sigmoid function, ⊙ is the Hadamard product, and c t-1 is the memory state at the previous moment.

4. An optical fiber nonlinear equalization device, characterized in that, It includes a processor and a storage medium; The storage medium is used to store instructions; The processor is used to operate according to the instructions to execute the steps of the method according to any one of claims 1 to 2.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 2.

Images