Small sample data enhancement method and system suitable for optical communication neural network equalization

Abstract

The application discloses a small sample data enhancement method and system suitable for optical communication neural network equalization, and relates to the technical field of optical communication signal processing. The method comprises the following steps: extracting a training sample window from a received signal and determining a training label to form an original training sample; according to reference symbol information of the training label, the training sample window is classified into a same category set; according to the amplitude characteristics of a received symbol at a center position of the training sample window, the category set is divided into subsets; according to an enhancement control parameter, an expanded training sample is extracted from each subset, and the original training sample is merged to obtain an enhanced training set, so that the problem that the related technology causes the bit error rate performance of an optical communication system to deteriorate when the training sample is limited is solved; and the optical communication neural network equalizer can learn the signal characteristics reflected by different amplitude states under the same symbol level under the condition of a small sample, the representation ability of the damage characteristics of a received signal is improved, and the equalization performance of the optical communication system is improved.

Description

Technical Field

[0001] This application relates to the field of optical communication signal processing technology, and in particular to a small sample data enhancement method and system suitable for optical communication neural network equalization. Background Technology

[0002] Optical communication systems include a transmitter, an optical fiber transmission link, and a receiver. With the development of neural network technology, the digital signal processor at the receiver of optical communication systems uses a neural network equalizer instead of a traditional linear equalizer.

[0003] When the number of labeled samples available for training neural network equalizers is limited, related technologies are constrained by the matching degree between the source and target domains or rely on pseudo-labels whose reliability is difficult to guarantee. Neither can directly solve the problems of insufficient training sample coverage and low sample utilization. They cannot fully cover the channel distortion characteristics and noise disturbances in different signal amplitude regions, which degrades the bit error rate performance of optical communication systems.

[0004] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] This application provides a small-sample data enhancement method and optical communication system suitable for optical communication neural network equalization. Without relying on any external data or external models, by operating only on the limited training data of the optical communication system itself, training samples that fully cover the channel distortion characteristics and noise disturbances in different signal amplitude regions can be obtained. This enables the optical communication neural network equalizer to learn the signal characteristics reflected by different amplitude states under the same symbol level under small-sample conditions, improves the ability to characterize the impairment characteristics of the received signal, and thus improves the equalization performance of the optical communication system.

[0006] To solve the above-mentioned technical problems, this application provides the following technical solution: This application provides a small-sample data augmentation method suitable for equalization of neural networks in optical communication, including: Multiple training sample windows are extracted from the received signal of the optical communication system. The known transmitted training sequence is used as the training label. The original training sample is composed of each training sample window and its corresponding training label. Based on the reference symbol information corresponding to each training label, training sample windows with the same reference symbol information are divided into the same category set, resulting in multiple category sets; Within each category set, the training sample window is divided into multiple amplitude subsets based on the amplitude representation value of the training sample window; the amplitude representation value is used to characterize the amplitude features of the received symbol at the center position of the training sample window. Based on the preset enhancement control parameters, training sample windows are extracted from each amplitude subset as expanded training samples; the number of samples extracted from each amplitude subset is related to the number of samples in the amplitude subset in its class set. The augmented training samples are merged with the original training samples to generate an enhanced training set for training an optical communication neural network equalizer.

[0007] This application also provides an optical communication system, including a transmitter, an optical fiber transmission link, and a receiver, wherein the receiver includes a processor; the processor is configured to execute a computer program to implement the steps of the above-described small sample data augmentation method applicable to optical communication neural network equalization.

[0008] The advantage of the technical solution provided in this application is that fiber dispersion, nonlinear effects and their coupling, as well as noise and other factors in optical communication systems, will cause the received signals corresponding to different reference symbols to exhibit different amplitude distributions. For the same reference symbol, the received signal amplitude at the center of its corresponding training sample window will also differ due to the different combinations of preceding and following symbols, channel impairments, and noise effects, thus forming different amplitude states with a certain distribution range within that reference symbol category. Based on the reference symbol information indicated by the training labels, the original training samples are divided into multiple category sets, so that sample windows within the same category set correspond to the same reference symbol information, and different category sets correspond to different reference symbol information. This avoids the mixing of sample windows with different reference symbol information in subsequent processing, which could lead to label errors. Then, based on the amplitude representation value that represents the amplitude feature of the received symbol at the center position of the training sample window, the sample windows within each category set are further divided into multiple amplitude subsets, so that sample windows with different amplitude features under the same reference symbol information are separated into different amplitude subsets. Then, according to the proportion of the number of samples in each amplitude subset to the number of original samples in that category, the number of augmented sample windows generated from each amplitude subset is determined, and a corresponding number of training sample windows are extracted from each subset as augmented samples. This ensures that sample windows in different amplitude state regions under each reference symbol information can participate in the generation of augmented samples, and the number of sample windows extracted from different regions is related to the original number of regions. The augmented samples are merged with the original samples to form an enhanced training set. This fully utilizes training samples of different amplitude states corresponding to different reference symbols, expands the number of effective training samples, and improves the ability of the training data to represent the characteristics of the received signal. This allows the equalization model to more fully learn the mapping relationship between the received signal and the transmitted symbol during actual channel transmission. The classification operation ensures that the labels of the augmented samples are consistent with the original labels, while the hierarchical and extraction operations ensure that the augmented samples originate from real received signals without introducing artificial interpolation or noise, preserving the channel impairment characteristics and physical statistical properties of the samples. Finally, the optical communication neural network equalizer trained using this enhanced training set achieves a more comprehensive learning of the characteristics of the received signal under limited original training samples, outputting more accurate symbol estimates and thus reducing the symbol decision error rate at the receiver. Simultaneously, because it achieves equalization performance comparable to training with a large number of samples with fewer original training samples, it reduces the requirement for the length of the training sequence in the optical communication system, lowers the training sequence overhead required for equalizer training, and improves the effective information transmission rate of the optical communication system.

[0009] In addition, this application also provides a corresponding optical communication system for a small sample data augmentation method applicable to optical communication neural network equalization, which has corresponding advantages.

[0010] The technical features mentioned above, those to be mentioned below, and those shown individually in the accompanying drawings can be arbitrarily combined, as long as the combined technical features are not contradictory. All feasible combinations of features are the technical content explicitly described in this application. Any one of the multiple sub-features contained in the same statement can be applied independently, without necessarily being applied together with other sub-features.

[0011] It should be understood that the above general description and the following detailed description are merely exemplary and do not limit this application. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart illustrating a small-sample data augmentation method for optical communication neural network equalization provided in this application; Figure 2 This is a schematic diagram of the structure of the optical communication system provided in this application in an exemplary embodiment; Figure 3 A schematic diagram of the hardware framework applicable to the small sample data augmentation method for optical communication neural network equalization provided in this application; Figure 4 A comparative diagram showing the training length required for different optical communication neural network equalizers provided in this application to achieve a 20% soft-decision forward error correction threshold; Figure 5 A schematic diagram illustrating the bit error rate performance of the optical communication neural network equalizer with the same enhancement control parameters but different training lengths provided in this application. Detailed Implementation

[0014] To enable those skilled in the art to better understand the technical solutions of this application, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. The terms "first," "second," "third," "fourth," etc., used in the specification and the aforementioned drawings are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. The term "exemplary" means "serving as an example, embodiment, or illustration." Any embodiment described herein as "exemplary" is not necessarily to be construed as superior to or better than other embodiments.

[0015] In optical communication systems, as the signal rate and modulation order increase, the received signal is affected by factors such as fiber dispersion, device bandwidth limitation, modulator nonlinearity, photodetector response, and noise, which aggravates the distortion of the received signal and increases the difficulty of symbol decision and equalization.

[0016] The optical communication neural network equalizer is a neural network model applied to signal processing at the receiver end of an optical communication system. It is used to learn the channel impairment compensation relationship based on the received signal to achieve equalization and restoration of the damaged signal. The neural network equalizer can learn the mapping relationship between the received signal and the target symbol through a data-driven approach, thereby compensating for linear and nonlinear impairments in the system. However, its equalization performance usually depends to a large extent on the amount and representativeness of the training data and the degree of coverage of signal distortion characteristics.

[0017] In practical high-speed optical communication systems, training data typically consists of an input window formed by the sampled signals from the receiver and its corresponding transmitted symbol labels. When the number of training samples is limited, the training samples do not adequately cover the system channel state, symbol distortion characteristics, and noise disturbances. This makes it difficult for the neural network model to fully learn the effective mapping relationship between the received signal and the transmitted symbol. In this case, the optical communication neural network equalizer is prone to problems such as decreased training stability, insufficient generalization ability, and degraded equalization performance. Increasing the number of original training samples can improve the sufficiency of training data, but longer training sequences will consume more transmission symbols, reducing the effective payload ratio. This impact is even more severe in scenarios that require frequent equalizer training and updates, such as burst transmissions, multi-user access, or rapid changes in channel state.

[0018] To address the dependence of optical communication neural network equalizers on a large amount of target link training data, related technologies employ methods such as transfer learning, multi-source transfer learning, unsupervised and semi-supervised learning, and knowledge distillation. Transfer learning typically trains a neural network model using data from the source system or similar link conditions, then transfers the model parameters to the target system and fine-tunes them using a small amount of target link training data. Multi-source transfer learning, building upon single-source domain transfer learning, further utilizes data from multiple source systems or various link conditions to construct a source domain dataset, or fuses features learned from multiple source models, providing richer prior information for the target system. Unsupervised equalization methods do not directly rely on known transmitted symbols in the target link; instead, they utilize the statistical characteristics of unlabeled received signals, constellation structure, symbol decision results, or self-generated pseudo-labels to construct training targets, and train the neural network through blind optimization, iterative decision-making, or label updates. Semi-supervised equalization methods combine a small amount of labeled training data with a large amount of unlabeled received signals, reducing the need for known training symbols in the target link through pseudo-label generation, consistency constraints, or joint training. Knowledge distillation methods typically utilize high-performance teacher networks to provide soft labels, feature information, or training guidance to student networks. Some optical communication equalization methods process unlabeled received signals through teacher models and use the results to guide student network training.

[0019] However, when the source system and the target system differ in signal rate, modulation format, device response, link impairments, or received power, the features learned in the source model may not be directly applicable to the target system, leading to a decline in transfer performance. The target system still requires a certain amount of training data for readjustment. Here, the target link refers to the physical optical transmission path currently in operation that needs equalization compensation; the target system refers to the hardware platform and digital signal processing algorithm architecture deployed at the receiving end of this link for signal processing; and the source system refers to the old system used to pre-train the model. While multi-source transfer learning can utilize model information from multiple source systems or multiple link conditions, it requires building and maintaining multiple source models or multiple sets of source domain data, increasing the complexity of the training process and system implementation. Furthermore, if the correlation between the source system and the target system is insufficient, the improvement of the target system's equalization performance by multi-source information will be limited. Unsupervised and semi-supervised equalization methods use unlabeled received signals, self-generated pseudo-labels, symbol decision results, or blind equalizer outputs. When there are errors in the initial symbol decision or blind equalization result, the generated training information may be inaccurate, and these errors may accumulate during subsequent training, affecting the learning effect of the neural network and the final equalization performance. Knowledge distillation methods require the pre-construction of a reliable teacher model, relying on it to provide soft labels or feature guidance. Its effectiveness is affected by the performance of the teacher model and its adaptation to the target link, and additional training or acquisition of the teacher model is also required. Therefore, the aforementioned related technologies, from the perspective of network model or training strategy, aim to reduce the neural network's dependence on large amounts of training data. However, they rely on source domain data, pre-trained source models, teacher models, or self-generated training information, without directly addressing the problem of insufficient training with limited training samples on the target link. Under limited original training sample conditions, the training samples do not adequately cover channel distortion characteristics and noise disturbances, leading to decreased training stability and degraded equalization performance of the optical communication neural network equalizer.

[0020] To address the problems of insufficient training data, inadequate sample coverage, limited model learning ability, and degraded bit error rate performance in neural network equalizers under small-sample training conditions in optical communication systems, this application provides a data augmentation method suitable for training neural network equalizers. Starting from existing training data, the training sample window is divided into the same category set according to the reference symbol information of the training labels. Within each category set, subsets are formed based on amplitude representation values. Expanded training samples are extracted from each subset according to enhancement control parameters and then merged with the original training samples to generate an enhanced training set for training the optical communication neural network equalizer. This improves the effectiveness and representativeness of limited training samples without collecting a large amount of additional training data, enhances the coverage of channel impairment features by the training samples, and enables the neural network to more fully learn the mapping relationship between received signals and transmitted symbols, thereby improving the training stability and generalization ability of the equalizer. After introducing the technical solution of this application, various non-limiting embodiments of this application are described in detail below with reference to the accompanying drawings and specific implementation details.

[0021] Please see first. Figure 1 According to the method for small-sample data augmentation applicable to optical communication neural network equalization provided in this application, it can be implemented as a computer program product, installed and run in the digital signal processor at the receiving end of the optical communication system, for constructing training sample windows, generating labels and performing data augmentation processing on the received signal, and constructing an augmented training set before training the optical communication neural network equalizer, thereby supporting the training of the optical communication neural network equalizer. In some embodiments of the method, the method includes the following steps: S101: Extract multiple training sample windows from the received signal of the optical communication system, use the known transmitted training sequence as the training label, and form the original training sample from each training sample window and its corresponding training label.

[0022] The received signal refers to the discrete sampling point sequence obtained by photoelectric detection and sampling at the optical communication receiver. The training sample window is a fixed-length data window composed of continuous sampling points from the receiver, serving as the input sample for the optical communication neural network equalizer. The training label is the reference symbol information corresponding to the training sample window, determined by the synchronized known training sequence, representing the reference symbol information of the transmitted symbol corresponding to the center position of the window. The original training sample refers to the training sample composed of the training sample window and its corresponding training label.

[0023] In this step, the receiver of the optical communication system performs analog-to-digital conversion and sampling on the electrical signal output by the photodetector to obtain a sequence of sampled points of the received signal. Multiple fixed-length sampled point sequence segments are extracted from this sequence, each serving as a training sample window, which represents the input features of the neural network equalizer. For example, the extraction of training sample windows can employ non-overlapping block truncation, where training sample windows do not overlap and are directly segmented into fixed-length segments of the received signal; alternatively, fixed-length truncation based on the symbol center can be used, where the sampled points corresponding to each reference symbol in the synchronized received signal are used as the window center, and a fixed number of sampled points are truncated forward and backward to form the training sample window. For sampled segments whose edges cannot meet the full window length, zero-padding is applied to the missing portions. Simultaneously, the transmitter inserts a known symbol sequence into the data frame as a training sequence. After symbol synchronization, the receiver uses this known transmitted training sequence to determine the training label corresponding to each training sample window. For example, the training label can be determined based on the symbol information in the known transmitted training sequence, or it can be determined based on the mapping relationship between the bits and modulation symbols in the known transmitted training sequence. For ease of description, the original training sample is defined as the training sample consisting of each training sample window and its corresponding training label.

[0024] S102: Based on the reference symbol information corresponding to each training label, training sample windows with the same reference symbol information are divided into the same category set, resulting in multiple category sets.

[0025] Among them, reference symbol information refers to the symbol information carried by the transmitted symbols in the known transmitted training sequence, and the category set refers to the training sample set obtained by grouping training sample windows with the same reference symbol information according to the reference symbol information indicated by the training label.

[0026] In this step, all original training samples are traversed, and the training label corresponding to each training sample window is read. The symbol identity is determined based on the reference symbol information indicated by the label. Training sample windows with the same symbol identity are grouped into the same category set, allowing subsequent processing to be performed independently within each symbol category. For example, for PAM4 (four-level pulse-amplitude modulation), the reference symbol information is the symbol level carried by the transmitted symbols in the known transmission training sequence. Correspondingly, the reference symbol information corresponds to four symbol levels. Therefore, the original training samples will be divided into four category sets, corresponding to symbol levels -3, -1, 1, and 3, respectively. Of course, in addition to classifying by symbol value, they can also be classified by the amplitude range of the received signal, by symbol polarity, or by the constellation point position corresponding to the modulation format, as long as the information used for classification can uniquely determine the transmitted symbol identity corresponding to the window. This application does not impose any limitations on this. All training sample windows within the same category set correspond to the same reference symbol information, while different category sets correspond to different reference symbol information. Through this division, subsequent S103 is performed independently within each category set, which isolates training sample windows with different reference symbol information from each other, avoids the mixing of samples of different reference symbol categories during the enhancement process and prevents label errors, and ensures that the correspondence between the labels of the augmented training samples generated in S104 and the original training samples remains unchanged.

[0027] S103: Within each category set, the training sample window is divided into multiple amplitude subsets based on the amplitude representation value of the training sample window.

[0028] The amplitude representation value characterizes the amplitude features of the received symbol at the center of the training sample window, reflecting the distribution of the actual received amplitudes across different training sample windows within the same reference symbol information. An amplitude subset refers to a dataset partitioned according to the amplitude representation value within a certain category set. For ease of description, the total number of training sample windows in the original training samples can be defined as the original sample count, the total number of training sample windows in the category set can be defined as the category sample count, the total number of training sample windows in the amplitude subset can be defined as the amplitude subset sample count, and the total number of training sample windows in the expanded training samples can be defined as the expanded sample count.

[0029] After obtaining multiple category sets in the previous step, this step obtains the amplitude representation value of each training sample window within each category set. Then, based on the magnitude of the amplitude representation value, the training sample windows within that category set are divided into multiple amplitude subsets. Training sample windows with different amplitude characteristics under the same reference symbol information are separated into different amplitude subsets. For example, this step can bin the samples with equal probability to make the number of samples in each amplitude subset equal, or bin them directly according to the physical amplitude range, or automatically divide the subsets using a clustering method. The number of bins in a sample subset is not limited to three; it can be two or four, depending on the granularity of the amplitude range division. Furthermore, the partitioning operation of the training sample windows within each category set can be performed in parallel, sequentially, or in a random order; this application does not impose any limitations on this.

[0030] S104: Based on the preset enhancement control parameters, extract training sample windows from each amplitude subset as expanded training samples.

[0031] The enhancement control parameter refers to a preset parameter used to control the required number of augmented samples. This parameter can be an empirical threshold, which can be flexibly set by those skilled in the art according to actual conditions. It can be configured as a fixed value or a value that can be flexibly adjusted according to actual conditions. The augmented training samples refer to the training samples consisting of training sample windows extracted from each amplitude subset and their corresponding training labels. Since the correspondence between the training sample windows and their corresponding training labels remains unchanged during the extraction process, the extracted training sample windows and their corresponding training labels together constitute the augmented training samples.

[0032] In this step, the number of training sample windows to be extracted for each amplitude subset is determined based on the enhancement control parameters. The number of training sample windows extracted for each amplitude subset can be the same or different, but the number extracted for each amplitude subset should be related to the number of samples in its respective category set. Then, training sample windows are extracted from each amplitude subset according to this number as expanded training samples. The extraction process uses the complete window and its label as the processing unit, without splitting or modifying the sampling points inside the window, thereby maintaining the channel impairment characteristics and physical statistical properties contained in the original training samples. For example, this step can also uniformly apply the same enhancement ratio to all categories, that is, the number of training sample windows extracted for each amplitude subset is equal to its original number multiplied by the same enhancement ratio; or different enhancement ratios can be set according to the scarcity of training sample windows contained in each amplitude subset, so that amplitude subsets with scarcer samples obtain higher enhancement ratios; or the number of expanded samples can be set first and then the number of training sample windows to be extracted for each amplitude subset can be deduced in reverse, all of which do not affect the implementation of this application.

[0033] S105: Merge the augmented training samples with the original training samples to generate an enhanced training set for training the optical communication neural network equalizer.

[0034] In optical communication, a neural network equalizer refers to a neural network model applied to signal processing at the receiver end of an optical communication system. It learns the mapping relationship between the received signal and the transmitted symbols to achieve equalization and restoration of damaged signals. Its input is a training sample window consisting of a sequence of sampling points from the received signal, and its output is an estimate of the corresponding transmitted symbol. The network structure of the optical communication neural network equalizer can be a feedforward neural network, a convolutional neural network, or other neural network structures suitable for signal equalization. The augmented training set refers to the training dataset used to train the optical communication neural network equalizer, obtained by merging the augmented training samples with the original training samples.

[0035] In this step, the expanded training samples extracted from each amplitude subset are combined with the original training samples from S101 to obtain the enhanced training set. For example, the receiver obtains a total of 400 original training sample windows and 300 expanded training sample windows from its known training sequences. These 300 expanded windows are combined with the original 400 training sample windows to obtain an enhanced training set containing 700 training sample windows. The enhanced training set only resamples the existing real receiving windows by category and amplitude, without interpolation, noise addition, perturbation, or artificial synthesis within the windows. Therefore, it still originates from the real received signals of the optical communication system, retaining the channel impairment characteristics and physical statistical properties contained in the original training samples, and the correspondence between each training window and its training label remains consistent before and after merging. Exemplarily, this step can also randomly shuffle the merged samples or arrange them alternately by category to make the distribution of training sample windows of different symbol categories and different amplitude states within each mini-batch more balanced.

[0036] In the technical solutions provided in this application, fiber dispersion, nonlinear effects and their coupling, as well as noise and other factors in the optical communication system, will cause the received signals corresponding to different reference symbols to exhibit different amplitude distributions. For the same reference symbol, the received signal amplitude at the center of its corresponding training sample window will also differ due to the different combinations of preceding and following symbols, channel impairments, and noise effects, thus forming different amplitude states with a certain distribution range within that reference symbol category. Based on the reference symbol information indicated by the training labels, the original training samples are divided into multiple category sets, so that sample windows within the same category set correspond to the same reference symbol information, and different category sets correspond to different reference symbol information. This avoids the mixing of sample windows with different reference symbol information in subsequent processing, which could lead to label errors. Then, based on the amplitude representation value that represents the amplitude feature of the received symbol at the center position of the training sample window, the sample windows within each category set are further divided into multiple amplitude subsets, so that sample windows with different amplitude features under the same reference symbol information are separated into different amplitude subsets. Then, according to the proportion of the number of samples in each amplitude subset to the number of original samples in that category, the number of augmented sample windows generated from each amplitude subset is determined, and a corresponding number of training sample windows are extracted from each subset as augmented samples. This ensures that sample windows in different amplitude state regions under each reference symbol information can participate in the generation of augmented samples, and the number of sample windows extracted from different regions is related to the original number of regions. The augmented samples are merged with the original samples to form an enhanced training set. This fully utilizes training samples of different amplitude states corresponding to different reference symbols, expands the number of effective training samples, and improves the ability of the training data to represent the characteristics of the received signal. This allows the equalization model to more fully learn the mapping relationship between the received signal and the transmitted symbol during actual channel transmission. The classification operation ensures that the labels of the augmented samples are consistent with the original labels, while the hierarchical and extraction operations ensure that the augmented samples originate from real received signals without introducing artificial interpolation or noise, preserving the channel impairment characteristics and physical statistical properties of the samples. Finally, the optical communication neural network equalizer trained using this enhanced training set achieves a more comprehensive learning of the characteristics of the received signal under limited original training samples, outputting more accurate symbol estimates and thus reducing the symbol decision error rate at the receiver. Simultaneously, because it achieves equalization performance comparable to training with a large number of samples with fewer original training samples, it reduces the requirement for the length of the training sequence in the optical communication system, lowers the training sequence overhead required for equalizer training, and improves the effective information transmission rate of the optical communication system.

[0037] In the above embodiments, there is no limitation on how to extract multiple training sample windows from the received signal of the optical communication system. In order to avoid the problem that the training sample windows are difficult to completely and orderly cover the timing of the received signal, and that samples in the beginning region of the sequence are easily lost, this embodiment also provides an exemplary implementation method for extracting training sample windows, which may include the following steps: Symbol synchronization is performed on the received signal to determine the corresponding sampling position of each transmitted symbol in the received signal; a sliding step size is determined based on the interval between the corresponding sampling positions of adjacent transmitted symbols in the received signal, and the sliding step size is used as the movement interval between the window center positions of adjacent training sample windows; the window length of the training sample window is determined based on the number of input nodes of the optical communication neural network equalizer; multiple sampling point sequence segments are extracted from the received signal using a sliding window method according to the sliding step size and window length, and used as each training sample window; for sampling point sequence segments that cannot form a complete window at the beginning position of the received signal, zero padding is used to complete the window length before being used as a training sample window.

[0038] Symbol synchronization refers to the timing alignment of the received signal with its corresponding transmitted symbol to establish the correspondence between the sampling positions in the received signal and the transmitted symbols. A known transmitted training sequence refers to a training sequence inserted or generated at the transmitting end in a known manner, whose transmitted symbol content is known. The sliding step size is determined by the interval between the corresponding sampling positions of adjacent transmitted symbols in the received signal. The window length refers to the number of sampling points contained in a single training sample window, determined by the number of input nodes of the optical communication neural network equalizer. The number of input nodes refers to the number of input data points used by the optical communication neural network equalizer to receive a single training sample. This number can be adjusted according to the channel memory length and the inter-symbol interference range, and is generally set to cover the number of sampling points corresponding to the symbol at the center of the window and several symbols before and after it. A sampling point sequence segment refers to a continuous sequence of sampling points extracted from the received signal.

[0039] In this embodiment, the received signal is first symbol synchronized to determine the corresponding sampling position of each transmitted symbol in the received signal, and the received data point at the corresponding sampling position is used as the center data point of the training sample window. Since there is one symbol period between two adjacent transmitted symbols, the data point interval corresponding to one symbol period is used as the sliding step size, causing the center position of adjacent training sample windows to move sequentially by one symbol period. For each center data point, a corresponding number of consecutive sampling points are intercepted forward and backward according to a preset window length, and these, together with the center data point, form a training sample window. The window center is moved sequentially according to the sliding step size until the traversal of the corresponding positions of each symbol in the received signal is completed. The amplitude value of each sampling point in a single training sample window serves as the input data for the optical communication neural network equalizer. For example, when the number of input nodes of the optical communication neural network equalizer is 13, the window length is set to 13 sampling points. The i-th training sample window, centered on the i-th received data point, intercepts 6 sampling points forward and backward, which, together with the received data point at the center position, form a training sample window containing 13 sampling points. If a complete window cannot be formed at the starting position of the received signal, zeros are padded to the missing sampling positions so that the length of the training sample window after zero padding reaches the preset window length. This preserves the training samples located at the edge of the received signal and avoids further reducing the limited number of training samples by directly discarding incomplete windows.

[0040] For example, the receiver of an optical communication system includes at least a photodetector, a transimpedance amplifier, an analog-to-digital converter (ADC), and a digital signal processor (DSP). The photodetector converts the received optical signal into an analog current signal, the transimpedance amplifier converts the current signal into a voltage signal and amplifies it, and the ADC samples this voltage signal at a fixed sampling clock frequency, outputting a digitized sequence of sample points. The value of each sample point in this sequence represents the instantaneous amplitude of the received signal at that moment. The sample point sequence is input to the DSP for further processing. In the DSP, the received signal is represented as a received sequence, and a sliding window method is used to construct a digital signal from this received sequence. A training sample window, window length Set as the number of input nodes for the optical communication neural network equalizer, the th training sample window The center position corresponds to the first Each received data point can be represented as: ,in, For positions at the beginning of the sequence where a complete window cannot be directly formed, zero padding is used. Each training sample window corresponds to a training label, and the... training sample window The corresponding training labels can be represented as The original training samples obtained thus can be represented as: ,in, This represents the set of training sample windows in the original training samples. This represents the corresponding set of training labels. Subsequent data augmentation processes are performed on a complete training sample window and its training labels, without splitting or modifying the sampling points within the training sample window.

[0041] As can be seen from the above, this embodiment determines the sliding step size based on the interval between the corresponding sampling positions of adjacent transmitted symbols in the received signal, causing the center positions of adjacent training sample windows to move sequentially according to the interval, and corresponding to the corresponding sampling positions of each transmitted symbol in the received signal in time sequence, thereby sequentially constructing training sample windows corresponding to each transmitted symbol, so that the sampling point sequence within the window can characterize the inter-symbol interference features between the center symbol and the symbols before and after it. The window length is determined by the number of input nodes of the optical communication neural network equalizer, so that the window dimension matches the input dimension of the optical communication neural network equalizer; by padding the sequence segments that cannot form a complete window at the starting position with zeros, the starting region of the sequence can also form a complete window.

[0042] For example, the reference symbol information in the above embodiments can be further limited to symbol level. Accordingly, the process of dividing the training sample window according to the reference symbol information corresponding to each training label is as follows: Based on the sign level corresponding to each training label, training sample windows with the same sign level are divided into the same category set.

[0043] The symbol level refers to the signal amplitude value of the transmitted symbol under its modulation format. Taking the PAM4 signal as an example, its transmitted symbol can take four symbol levels, namely -3, -1, 1, and 3.

[0044] The above embodiments do not limit how to divide the training sample window within each category set into multiple amplitude subsets. To avoid insufficient coverage of channel impairment states by the training set due to excessive concentration of augmented samples in a certain amplitude state region, this embodiment also provides an exemplary method for generating amplitude subsets, which may include the following: Within each category set, the training sample windows are sorted according to the magnitude of their amplitude representation values; the sorted training sample windows are then divided into three equal parts, which are designated as the first amplitude subset, the second amplitude subset, and the third amplitude subset, respectively. In this embodiment, for a certain category set, each training sample window in the category set has an amplitude representation value, which represents the amplitude feature of the received symbol at the center position of the window. Within each category set, the training sample windows in the same category set can be sorted in descending order of amplitude representation value, or they can be sorted in ascending order of amplitude representation value. The sorted training sample windows are then divided into three equal parts according to their quantity, which are respectively designated as the first amplitude subset, the second amplitude subset, and the third amplitude subset. For example, in a category set with a symbol level of 1, there are 9 training sample windows. The amplitude representation values ​​of each training sample window are ordered from smallest to largest as 0.45, 0.51, 0.56, 0.63, 0.69, 0.73, 0.78, 0.85, and 1.05. The first amplitude subset may contain the first three, namely the training sample windows corresponding to 0.45, 0.51, and 0.56; the second amplitude subset may contain the middle three, namely the training sample windows corresponding to 0.63, 0.69, and 0.73; and the third amplitude subset may contain the last three, namely the training sample windows corresponding to 0.78, 0.85, and 1.05. It should be noted that dividing into three equal parts is only an exemplary implementation method of this embodiment. In practical applications, the sorted training sample windows may also be divided into two or four equal parts, or other numbers of amplitude subsets. The specific number of stratifications can be adjusted according to the total amount of original samples and the desired distribution of augmented samples.

[0045] For example, the original training samples are divided into groups according to the symbol levels corresponding to the training labels. The nth category set, the nth The sample index set corresponding to each category set and sample set They are respectively represented as , .in, This indicates the number of training sample windows in the set of this category. Indicates the first Four symbol levels. For example, for a PAM4 signal, the training labels can correspond to four symbol levels: -3, -1, 1, and 3. In the... A collection of categories Within the training sample windows, the samples are sorted from smallest to largest according to their amplitude representation values. Based on this sorting order, the samples in this category are approximately divided into the first amplitude subset. Second amplitude subset and the third amplitude subset Three subsets. If the first The total number of training window samples in each category set, i.e., the number of category samples, is The total number of training window samples contained in each of the three amplitude subsets, which is also the number of samples in the amplitude subset, are respectively ,but .

[0046] In this embodiment, through the sorting and equal division operations described above, sample windows with different amplitude characteristics are separated into different amplitude subsets within the same category set. The purpose of this processing is to ensure that samples of different amplitude states can participate in the generation of enhanced samples during subsequent resampling, avoiding insufficient coverage of channel impairment states in the training set due to excessive concentration of enhanced samples in a certain amplitude state region.

[0047] Based on the above embodiments, the amplitude representation value, as a numerical value used to reflect the amplitude information of the received signal carried by the training sample window, can be the amplitude value of the received sampling point at the center of the window, or it can be the average amplitude value, peak amplitude value, or root mean square amplitude value of all sampling points within the window. This embodiment uses the sampling value at the center of the training sample window as an example to illustrate the amplitude representation value. Before sorting the training sample windows in the category set according to the magnitude of the amplitude representation value, the center position of each training sample window can be obtained, the amplitude value of the target received data point at the center position can be extracted, and the amplitude value of the target received data point can be used as the amplitude representation value of the corresponding training sample window.

[0048] In this embodiment, the center position of each training sample window can be obtained first. The center position is the received sampling position corresponding to the corresponding transmitted symbol after symbol synchronization. Then, the amplitude value of the target received data point at the center position is extracted, and this amplitude value is used as the amplitude representation value of the corresponding training sample window. For example, if a training sample window is a 13-point sequence centered on the corresponding sampling point of the i-th transmitted symbol in the received signal, then the center position of the training sample window is the corresponding sampling point, and the target received data point is the received sampling point at the center position. If the amplitude value of the corresponding sampling point of the i-th transmitted symbol in the received signal is 1.4, then the amplitude representation value of the training sample window is 1.4. The amplitude value of the received data point at the center position of the window is used as the amplitude representation value because the training label represents the symbol category of the transmitted symbol corresponding to the center position of the window, and the received sampling point at the center position can better reflect the received amplitude characteristics of the transmitted symbol after channel transmission. For the same reference symbol category, different training samples may have different received signal amplitudes at their center positions due to differences in the combination of preceding and following symbols, inter-symbol interference, noise, and other channel impairments, thus exhibiting different amplitude states. Therefore, using the amplitude value of the sampling point at the center position as the amplitude representation value can characterize the amplitude state differences of different training samples within the same reference symbol category, providing a basis for subsequent sorting and stratification according to the amplitude representation value.

[0049] Based on the above embodiments, considering the training convergence stability of the optical communication neural network equalizer, this embodiment further specifies that the training sample window included in the enhanced training set is randomly shuffled, which may include the following: The augmented training samples are merged with the original training samples to form an initial augmented training set. The order of the training sample windows in the initial augmented training set is randomly adjusted to avoid consecutive arrangement of training sample windows with the same symbol category or similar amplitude states. This ensures that training sample windows with different symbol categories and different amplitude states are fully mixed in the initial augmented training set to obtain an augmented training set, which is then used to train the optical communication neural network equalizer.

[0050] The initial augmented training set refers to the set of training samples after the expanded training samples and the original training samples have been merged, but whose order has not yet been adjusted. The augmented training set is the set of training samples obtained by adjusting the order of the training sample windows contained in the initial augmented training set.

[0051] In this embodiment, when merging to generate the augmented training set, the augmented training samples are first merged with the original training samples to form an initial augmented training set; then the arrangement order of each training sample window in the initial augmented training set is randomly adjusted to avoid training sample windows with the same symbol category or similar amplitude states being arranged consecutively, so that training sample windows with different symbol categories and different amplitude states are fully mixed in the training set to obtain the final augmented training set. The enhanced training set is input into the optical communication neural network equalizer for training. If the optical communication neural network equalizer is a feedforward neural network equalizer, and the feedforward neural network equalizer is a regression equalizer, the output layer of the feedforward neural network equalizer uses a linear activation function to output a continuous estimate of the transmitted symbol level, and uses the symbol level amplitude of the training label as the regression target to calculate the loss. When the feedforward neural network equalizer is a classification equalizer, its output layer uses a softmax activation function to output the probability corresponding to each symbol category, and converts the symbol level corresponding to the training label into a discrete category label. The cross-entropy loss is calculated based on the discrete category label and the category probability output by the network. After training, the trained feedforward neural network equalizer is used to process the received signal to be equalized. For a regression equalizer, the corresponding transmitted symbol estimation result is determined based on the continuous estimate of the network output; for a classification equalizer, the corresponding symbol category is determined based on the probability corresponding to each symbol category output by the network.

[0052] As can be seen from the above, this embodiment randomly adjusts the order of the merged initial augmented training set, so that samples of different symbol categories and different amplitude states are fully mixed in the training set. This avoids the expanded training samples being arranged in a concentrated manner according to category or amplitude subset order, improves the randomness of small batch samples during training, and is conducive to improving the convergence stability of neural network training.

[0053] Based on the above embodiments, in order to avoid excessive concentration of expanded training samples in a certain category or local amplitude region, this embodiment also provides an implementation method for reasonably distributing the training sample windows extracted from each amplitude subset among each symbol level category and each amplitude layer, which may include the following: Based on the augmentation control parameters and the number of class samples in each class set, the number of class augmentations for each class set is determined. According to the proportion of the number of samples in each amplitude subset to the number of class samples in its class set, the number of class augmentations for each class set is allocated to its amplitude subsets to obtain the target extraction number for each amplitude subset. According to the target extraction number for each amplitude subset, a corresponding number of training sample windows are extracted from the corresponding amplitude subset.

[0054] Here, the number of class samples refers to the number of training sample windows contained in a certain class set, and the number of class augmentations refers to the number of training sample windows generated for a certain class set that need to be resampled. The number of target extractions refers to the number of training sample windows allocated to a certain amplitude subset that need to be extracted from that amplitude subset.

[0055] In this embodiment, when extracting training sample windows from each amplitude subset, the number of class enhancements for each class set is first determined based on the enhancement control parameters and the number of class samples in each class set. Then, according to the proportion of the number of samples in each amplitude subset to the number of samples in its respective class set, the number of class enhancements for each class set is allocated to its respective amplitude subsets, resulting in the target extraction number for each amplitude subset. Finally, according to the target extraction number for each amplitude subset, the corresponding number of training sample windows are extracted from the corresponding amplitude subset. Taking an example containing four symbol level categories, if the enhancement control parameter is 2, the number of class samples in the four class sets are 100, 80, 120, and 90, respectively, and the number of class enhancements for each category are 200, 160, 240, and 180. Taking a class set with 100 class samples as an example, the number of samples in its three amplitude subsets are 40, 30, and 30, respectively, and the sum of these three is 100, which is equal to the number of class samples for that category. The target extraction quantity for each of the three amplitude subsets is equal to the number of augmented categories multiplied by the proportion of the sample size in each amplitude subset to 100. That is, the first amplitude subset is 200 × 40 ÷ 100 = 80, the second amplitude subset is 200 × 30 ÷ 100 = 60, and the third amplitude subset is 200 × 30 ÷ 100 = 60. In this way, the final augmented training samples are distributed among the symbol level categories and amplitude layers according to the original distribution ratio, avoiding over-concentration in a certain category or local amplitude region.

[0056] As a more convenient implementation, the calculation formulas for extracting training sample windows of different amplitude subsets can be stored locally in advance, and then the corresponding values ​​can be directly calculated by calling these calculation formulas. For example, to make the first... The first, second, and third amplitude subsets within a class all participate in the sample augmentation generation process. The number of category enhancements required for each category to be resampled can be expressed as: ,Will The distribution is based on the proportion of each amplitude subset within the category, and correspondingly, the... The number of training sample windows that need to be resampled from the first amplitude subset of the category can be expressed by the following formula: Calculations show that This represents the rounding operation, the first... The number of training sample windows that need to be resampled from the second amplitude subset in the class can be determined by the relational expression. Calculated; the first The number of training sample windows that need to be resampled from the third amplitude subset in the class can be determined by the relational formula. The calculations show that, correspondingly These represent the target number of samples to be extracted from the first, second, and third amplitude subsets, respectively.

[0057] As can be seen from the above, this embodiment determines the augmentation quantity of each category according to the augmentation control parameters and the number of category samples, and allocates the required augmentation quantity to each amplitude subset according to the proportion of the number of samples in each amplitude subset to the number of category samples in the corresponding category set to obtain the target extraction quantity. This allows the augmented training samples to be distributed in each symbol level category and each amplitude layer according to the original distribution ratio, thereby avoiding excessive concentration of augmented training samples and making the final augmented training set distribution more reasonable.

[0058] Based on the above embodiments, this embodiment further defines the method for determining the number of category enhancements, which may include the following: Multiply the preset enhancement control parameters by the number of category samples in each category set; round the product result and use the rounded result as the number of category enhancements for the corresponding category set.

[0059] In this embodiment, the rounding operation refers to rounding the value down or to the nearest integer to ensure that the category enhancement quantity is an integer. The enhancement control parameter is expressed as follows: It is used to control the number of training sample windows that need to be expanded. .like If , then no data augmentation will be performed. For the , The number of category augmentations needed for each category to be generated by resampling It can be represented as: .

[0060] As can be seen from the above, this embodiment obtains the number of class enhancements by multiplying the enhancement control parameters by the number of samples in each class and rounding down. This allows the number of training sample windows extracted for each class corresponding to different reference symbol information to be determined by the original number of that class and the enhancement control parameters, thereby making the enhancement scale controllable and matching the original data volume of each class.

[0061] Based on the above embodiments, this embodiment further defines the implementation method for extracting the training sample window from the amplitude subset, which may include the following: The target number of samples to be drawn from each amplitude subset is compared with the number of samples in the corresponding amplitude subset. If the target number of samples to be drawn does not exceed the number of samples in the amplitude subset, a random sampling method without replacement is used to draw a training sample window equal to the target number of samples from the corresponding amplitude subset. If the target number of samples to be drawn exceeds the number of samples in the amplitude subset, all training sample windows in the corresponding amplitude subset are selected repeatedly in integer rounds, and then random sampling without replacement is performed within the remaining number of samples that is less than one round.

[0062] Random sampling without replacement refers to a sampling method in which the sampled sample is not returned to the population after each draw, so that the same sample is drawn at most once. Integer round repeated selection refers to the process of drawing all training sample windows in the amplitude subset in multiple rounds when the target number of samples exceeds the number of samples in the amplitude subset.

[0063] In this embodiment, random sampling is performed within each amplitude subset according to the corresponding target sampling quantity. If the target sampling quantity does not exceed the sample size of the amplitude subset, a random sampling method without replacement is used to draw a training sample window equal to the target sampling quantity from the corresponding amplitude subset. In this case, each sample is selected at most once, which can cover a portion of the samples within the amplitude subset. If the target sampling quantity exceeds the sample size of the amplitude subset, all training sample windows in the amplitude subset are selected repeatedly in integer rounds, that is, all samples in the amplitude subset are completely drawn multiple times. Then, random sampling without replacement is performed within the remaining sampling quantity of less than one round, so that a large augmentation scale requirement can still be met when the amplitude subset samples are insufficient.

[0064] For example, for the first Each amplitude subset of the category set, starting from the first amplitude subset Randomly selected from A training sample window constitutes the expanded sample set corresponding to the first amplitude subset. From the second amplitude subset Randomly selected from Each training sample window constitutes an expanded sample set corresponding to the second amplitude subset. From the third amplitude subset Randomly selected from Each training sample window constitutes an expanded sample set corresponding to the third amplitude subset. The training sample windows obtained from the three amplitude subsets are merged to obtain the first... Expanded training samples corresponding to the category set It can be represented as Since resampling always occurs within the same category set, therefore the... Class The training labels for all training sample windows are still 0. The corresponding set of extended tags It can be represented as .in, Indicates length is A row vector consisting entirely of 1s. Indicates length is A row vector consisting entirely of 1s. Indicates length is A row vector of all 1s. After performing the above intra-class hierarchical resampling on all categories, the augmented training samples for each category are obtained. The augmented training samples of all categories are merged with the original training samples to obtain the augmented training set, and the training sample window set within the augmented training set. It can be represented as Enhance the training label set corresponding to the training sample window set within the training set. It can be represented as .

[0065] As can be seen from the above, this embodiment first compares the target extraction quantity with the amplitude subset sample quantity. If the quantity does not exceed the limit, random sampling without replacement is used. If the quantity exceeds the limit, all samples are selected repeatedly in integer rounds and then the remaining quantity of samples is extracted without replacement. This ensures that the extraction of each amplitude subset can cover the samples within the amplitude subset and can expand the quantity by repeated selection when needed, thereby meeting the needs of samples of each category and amplitude layer under different enhancement scales.

[0066] This application also provides a corresponding apparatus for a small-sample data augmentation method suitable for optical communication neural network equalization, further enhancing the practicality of the method. The apparatus can be described from both a functional module perspective and a hardware perspective. The small-sample data augmentation apparatus for optical communication neural network equalization provided in this application is described below. This apparatus is used to implement the small-sample data augmentation method for optical communication neural network equalization provided in this application. In this embodiment, the small-sample data augmentation apparatus for optical communication neural network equalization may include or be divided into one or more program modules. These one or more program modules are stored in a storage medium and executed by one or more processors to complete the small-sample data augmentation method for optical communication neural network equalization disclosed in the embodiment. The program module referred to in this embodiment is a series of computer program instruction segments capable of performing specific functions, which is more suitable than the program itself for describing the execution process of the small-sample data augmentation apparatus for optical communication neural network equalization in the storage medium. The following description will specifically introduce the functions of each program module in this embodiment. The small-sample data augmentation apparatus for optical communication neural network equalization described below can be referred to in correspondence with the small-sample data augmentation method for optical communication neural network equalization described above.

[0067] From the perspective of functional modules, the small sample data enhancement device for optical communication neural network equalization provided in this embodiment may include: The original training sample generation module is used to extract multiple training sample windows from the received signals of the optical communication system, use the known transmitted training sequence as the training label, and the original training sample is composed of each training sample window and its corresponding training label. The classification module is used to divide training sample windows with the same reference symbol information into the same category set according to the reference symbol information corresponding to each training label, thus obtaining multiple category sets; The amplitude partitioning module is used to divide the training sample window in each category set into multiple amplitude subsets based on the amplitude representation value of the training sample window; the amplitude representation value is used to represent the amplitude state of the received signal corresponding to the training sample window.

[0068] The expanded sample determination module is used to extract training sample windows from each amplitude subset according to preset enhancement control parameters, and use them as expanded training samples; the number of samples extracted from each amplitude subset is related to the number of samples in the amplitude subset in its class set; The augmented dataset generation module is used to merge the augmented training samples with the original training samples to generate an augmented training set for training an optical communication neural network equalizer.

[0069] The small-sample data augmentation device for optical communication neural network equalization mentioned above is described from the perspective of functional modules. Furthermore, this application also provides an electronic device, described from a hardware perspective. This electronic device includes a memory and a processor. The memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any of the above embodiments of the small-sample data augmentation method for optical communication neural network equalization.

[0070] It is understood that if the small-sample data augmentation method applicable to optical communication neural network equalization in the above embodiments is implemented as a software functional unit and sold or used as an independent product, it can be stored in a non-volatile storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the related technology, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and executes all or part of the steps of the methods in the various embodiments of this application. The aforementioned storage medium includes, but is not limited to, various media capable of storing program code, such as: USB flash drive, mobile hard disk, read-only memory (ROM), random access memory (RAM), electrically erasable programmable ROM, register, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, removable disk, CD-ROM, magnetic disk, or optical disk. Based on this, this application also provides a non-volatile storage medium on which a computer program is stored, which is configured to execute the steps in any of the above embodiments of the small-sample data augmentation method applicable to optical communication neural network equalization when running.

[0071] It is understood that if the small-sample data augmentation method for optical communication neural network equalization in the above embodiments is implemented as a software functional unit and sold or used as an independent product, the computer software product may not need to be stored in a physical storage medium. For example, it can be directly transmitted to a computer or other device with information processing capabilities via a wired or wireless network to execute all or part of the steps of the methods in the various embodiments of this application. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the related technology, or all or part of the technical solution, can be embodied in the form of a software product. Based on this, this application also provides a computer program product, which stores a computer program that is configured to execute the steps in any of the above embodiments of the small-sample data augmentation method for optical communication neural network equalization when running.

[0072] Finally, this application also provides an optical communication system, please refer to [link to relevant documentation]. Figure 2 It may include a transmitter 201, an optical fiber transmission link 202, and a receiver 203.

[0073] The transmitter 201 may include, for example, a signal generator and an electro-optic modulator. The signal generator generates a binary data stream containing a known training sequence and maps this data stream into a multi-level modulated symbol sequence. The electro-optic modulator converts the modulated symbol sequence into an optical signal, such as using a Mach-Zehnder modulator or a direct-modulation laser, loading the electrical signal onto an optical carrier to generate a modulated optical signal. An optical fiber transmission link 202 connects the transmitter 201 and the receiver 203 for transmitting the optical signal. The optical fiber transmission link may include, for example, an optical fiber and an optional optical amplifier. When the optical signal propagates in the optical fiber, it is affected by fiber dispersion and nonlinear effects, resulting in waveform distortion and inter-symbol interference. The optical amplifier is used to compensate for signal attenuation during optical fiber transmission and may be an erbium-doped fiber amplifier. The receiver 203, in addition to including a processor 2031, may also include a photodetector and an analog-to-digital converter. The photodetector converts the received optical signal into an analog electrical signal. The analog-to-digital converter samples the analog electrical signal at a set sampling clock frequency, converting the analog electrical signal into a digitized sequence of sample points. The processor 2031, when executing a computer program, implements the steps of the small-sample data augmentation method for optical communication neural network equalization described in any of the above embodiments. After generating an augmentation training set, it trains an optical communication neural network equalizer using the augmentation training set, and uses the trained optical communication neural network equalizer to perform equalization processing on the received signal to be equalized, outputting a symbol estimate. The processor 2031 may be a digital signal processor, a field-programmable gate array, an application-specific integrated circuit, or other hardware device with data processing capabilities.

[0074] By employing the aforementioned data augmentation method, the optical communication system can achieve better balanced performance under limited training sample conditions, reduce training sequence overhead, and improve the system's effective information transmission rate.

[0075] To further illustrate the application of the technical solution of this application in an optical communication system, this embodiment also uses a 50GBaud PAM4 short-range optical communication system for verification. The hardware framework of the experimental system is as follows: Figure 3 As shown, it may include the following: The transmitter generates a random bit sequence and maps it to PAM4 symbols. After upsampling, pulse shaping, and resampling, the sequence is input to an arbitrary waveform generator. The electrical PAM4 signal output from the arbitrary waveform generator drives a directly modulated laser (DML) to modulate the optical intensity, generating a modulated optical signal, which is then transmitted to the receiver via a 10 km fiber optic link. At the receiver, the received optical power is adjusted to 5 dBm using an adjustable optical attenuator. The received signal is then obtained through photodetector and high-speed sampling. The digital signal processor at the receiver performs resampling, matched filtering, clock synchronization, training sample window construction, tag generation, data augmentation, neural network equalization, symbol decision, bit demapping, and bit error rate calculation. The data augmentation process employs... Figure 1 The method shown generates an enhanced training set, which is then used to train an optical communication neural network equalizer.

[0076] In this embodiment, two types of network structures are selected: Feedforward Neural Network (FNN) and Convolutional Neural Network (CNN). For each type of network, two output formats—regression and classification—are constructed for validation. The initial learning rate for training the optical communication neural network equalizer is 1×10⁻⁶. -3 The learning rate decay factor is 0.9, the learning rate decay period is 10 training epochs, the maximum number of training epochs is set to 50, and the mini-batch size is 100 samples. The feedforward neural network consists of one input layer, two fully connected hidden layers, and one output layer, with each of the two fully connected hidden layers containing 64 neurons. The convolutional neural network consists of one input layer, two one-dimensional convolutional layers, two fully connected hidden layers, and one output layer. Each one-dimensional convolutional layer contains four convolutional kernels of length 2, and each of the two fully connected hidden layers contains 64 neurons. The input window length for both the feedforward and convolutional neural networks is set to 13.

[0077] To fairly compare the balanced performance before and after data augmentation, the training random seed was reset to the same value before and after each trial, thus reducing differences caused by network initialization and other training-related random operations. In the case of data augmentation, a separate data augmentation random seed was used, controlling the random resampling process separately from the training random seed. Ten trials were conducted using different training seeds, with each augmentation experiment using a different augmentation seed. The bit error rate (BER) result was obtained by averaging the results of the ten trials. Figure 3 The experimental system shown was used to perform performance tests on the technical solution of this application. The test results are as follows: Figure 4 and Figure 5 As shown.

[0078] Figure 4 This paper presents the training sequence lengths required for different equalizers to reach the 20% soft-decision forward error correction (SD-FEC) threshold under different enhancement control parameters. It can be concluded that the training sequence lengths required for each equalizer to reach the 20% SD-FEC threshold generally decrease with increasing enhancement control parameters. Specifically, when the enhancement control parameter is 2.4, compared with the case without data augmentation, the number of training symbols required for feedforward neural network regression, feedforward neural network classification, convolutional neural network regression, and convolutional neural network classification to reach this threshold is reduced by approximately 62.5%, 43.8%, 56.4%, and 31.3%, respectively. These results demonstrate that the technical solution of this application can reduce the number of original training symbols required by the optical communication neural network equalizer while maintaining the target equalization performance.

[0079] Figure 5 The bit error rate performance of feedforward neural networks and convolutional neural network regression equalizers with different training sequence lengths is presented, with an enhancement control parameter of 2.4. Figure 5It can be seen that as the length of the training sequence increases, the overall bit error rate (BER) of each optical communication neural network equalizer shows a downward trend, indicating that more sufficient training samples help improve the learning ability and equalization performance of the optical communication neural network equalizer to channel impairments. Further comparison of the results with and without the proposed technical solution shows that, under the same training length, the BER of each optical communication neural network equalizer using the proposed technical solution is reduced. When the training sequence is short, the original training samples do not adequately cover the characteristics of channel impairments and nonlinear distortion. The proposed technical solution expands the number of training samples through structured resampling and increases the participation of samples of different symbol categories and different amplitude regions within the same category in the network training process, thereby improving the generalization ability of the optical communication neural network equalizer. When the original training sequence is long, the training samples cover the signal statistical characteristics more fully, and the additional performance gain brought by data augmentation gradually decreases. Therefore, the BER performance before and after data augmentation tends to be similar. These test results show that the proposed technical solution can improve the utilization efficiency of limited training samples and improve the neural network equalization performance under small sample conditions.

[0080] Any of the components, modules, units, parts, methods, and operations described herein can be implemented using software, firmware, hardware (e.g., fixed logic circuitry), manual processing, or any combination thereof. Alternatively or additionally, any functionality described herein can be performed at least in part by one or more hardware logic components, such as, but not limited to, a central processing unit (CPU), a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), an application-specific standard product (ASSP), a system-on-a-chip (SoC), a complex programmable logic device (CPLD), a microprocessor (MCU), etc. The systems, computing devices, or apparatuses described herein encompass a wide range of means, devices, and machines for processing data, including, for example, one or more programmable processors, computers, SoCs, or combinations thereof. The apparatus may also include code that creates an execution environment for the computer program in question, such as code constituting processor firmware, a protocol stack, a database management system, an operating system, a cross-platform runtime environment, a virtual machine, or one or more combinations thereof. The aforementioned computer program (also known as a program, software, software application, app, script, or code) can be written in any form of programming language, including compiled or interpreted languages, declarative or procedural languages, and can be deployed in any form, including as a standalone program or as a module, component, subroutine, object, or other unit suitable for a computing environment.

[0081] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this specification can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0082] The foregoing has provided a detailed description of a small-sample data enhancement method and optical communication system suitable for neural network equalization in optical communication. The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. Several improvements and modifications can be made to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for augmenting small sample data suitable for equalization of neural networks in optical communication, characterized in that, include: Multiple training sample windows are extracted from the received signal of the optical communication system. The known transmitted training sequence is used as the training label. The original training sample is composed of each training sample window and its corresponding training label. Based on the reference symbol information corresponding to each training label, training sample windows with the same reference symbol information are divided into the same category set, resulting in multiple category sets; Within each category set, the training sample window is divided into multiple amplitude subsets based on the amplitude representation value of the training sample window. The amplitude characterization value is used to characterize the amplitude features of the received symbols at the center position of the training sample window; Based on the preset enhancement control parameters, training sample windows are extracted from each amplitude subset as expanded training samples; the number of samples extracted from each amplitude subset is related to the number of samples in the amplitude subset in its class set. The augmented training samples are merged with the original training samples to generate an enhanced training set for training an optical communication neural network equalizer.

2. The small-sample data augmentation method for optical communication neural network equalization according to claim 1, characterized in that, Multiple training sample windows are extracted from the received signals of the optical communication system, including: Symbol synchronization is performed on the received signal to determine the corresponding sampling position of each transmitted symbol in the received signal, and the received signal sampling point at the corresponding sampling position is respectively used as the center position data point of each training sample window; The sliding step size is determined based on the interval between the corresponding sampling positions of adjacent transmitted symbols in the received signal, and the sliding step size is used as the movement interval between the window center positions of adjacent training sample windows. The window length of the training sample window is determined based on the number of input nodes of the optical communication neural network equalizer. According to the sliding step size and the window length, multiple sampling point sequence segments are extracted from the received signal using a sliding window method, which serve as each training sample window. For sampling point sequences that cannot form a complete window at the starting position of the received signal, they are padded with zeros to the window length and then used as training sample windows.

3. The small-sample data augmentation method for optical communication neural network equalization according to claim 1, characterized in that, Based on the reference symbol information corresponding to each training label, training sample windows with the same reference symbol information are grouped into the same category set, including: Based on the sign level corresponding to each training label, training sample windows with the same sign level are divided into the same category set.

4. The method for small-sample data augmentation suitable for equalization of optical communication neural networks according to claim 1, characterized in that, Based on the amplitude representation values ​​of the training sample windows, the training sample windows within each category set are divided into multiple amplitude subsets, including: Within each category set, the training sample windows within each category set are sorted from smallest to largest according to the magnitude representation value; The sorted training sample window is divided into three equal parts according to the number of samples, which are respectively used as the first amplitude subset, the second amplitude subset, and the third amplitude subset; Wherein, the amplitude representation value of the first amplitude subset is less than the amplitude representation value of the second amplitude subset, and the amplitude representation value of the second amplitude subset is less than the amplitude representation value of the third amplitude subset.

5. The method for small-sample data augmentation suitable for equalization of optical communication neural networks according to claim 4, characterized in that, Before sorting the training sample windows within the category set according to the magnitude representation values ​​from smallest to largest, the process also includes: Obtain the center position of each training sample window, where the center position of each training sample window corresponds to the corresponding sampling position of each transmitted symbol in the received signal; Extract the amplitude value of the target received data point at the center position, and use the amplitude value of the target received data point as the amplitude representation value of the corresponding training sample window.

6. The method for small-sample data augmentation suitable for equalization of optical communication neural networks according to claim 1, characterized in that, The augmented training samples are merged with the original training samples to generate an enhanced training set for training an optical communication neural network equalizer, including: The augmented training samples are merged with the original training samples to form an initial enhanced training set; The order of the training sample windows in the initial enhanced training set is randomly adjusted to avoid consecutive arrangement of training sample windows with the same symbol category or similar amplitude states, so that training sample windows with different symbol categories and different amplitude states are fully mixed in the initial enhanced training set to obtain an enhanced training set, which is then used to train an optical communication neural network equalizer.

7. The method for small-sample data augmentation suitable for equalization of neural networks in optical communication according to any one of claims 1 to 6, characterized in that, Training sample windows are extracted from each amplitude subset, including: Based on the enhancement control parameters and the number of category samples in each category set, the number of category enhancements for each category set is determined. Based on the proportion of the number of samples in each amplitude subset to the number of samples in the category set to its respective category set, the category enhancement quantity of each category set is allocated to its amplitude subsets to obtain the target extraction quantity of each amplitude subset; According to the target extraction number of each amplitude subset, the corresponding number of training sample windows are extracted from the corresponding amplitude subset.

8. The method for small-sample data augmentation suitable for equalization of optical communication neural networks according to claim 7, characterized in that, Based on the enhancement control parameters and the number of category samples in each category set, the number of category enhancements for each category set is determined, including: Multiply the preset enhancement control parameters by the number of class samples in each class set; The product result is rounded down, and the rounded result is used as the number of category augmentations for the corresponding category set.

9. The method for small-sample data augmentation suitable for equalization of optical communication neural networks according to claim 7, characterized in that, Based on the target extraction quantity for each amplitude subset, a corresponding number of training sample windows are extracted from the corresponding amplitude subset, including: Compare the number of targets extracted from each amplitude subset with the number of samples in the corresponding amplitude subset; If the target sampling quantity does not exceed the sample quantity of the amplitude subset, a training sample window equal to the target sampling quantity is extracted from the corresponding amplitude subset using a random sampling method without replacement. If the target number of samples exceeds the number of samples in the amplitude subset, based on the required number of samples, first, all training sample windows in the corresponding amplitude subset are repeatedly selected in integer rounds, and then the remaining number of training sample windows that are less than one round are randomly selected from the amplitude subset without replacement.

10. An optical communication system, characterized in that, It includes a transmitter, an optical fiber transmission link, and a receiver, wherein the receiver includes a processor; The processor is configured to implement the steps of the small sample data augmentation method for optical communication neural network equalization as described in any one of claims 1 to 9 when executing a computer program.

Images