Signal processing method and device and electronic equipment

Abstract

The invention discloses a signal processing method and device and electronic equipment. The method comprises the following steps: acquiring a quadrature amplitude modulation (QAM) signal output by a digital signal processor of a receiving end; a complex number field Transform network is adopted to carry out feature extraction on the QAM signal to obtain a first output, and the complex number field Transform network is used for analyzing the global dependency relationship of the QAM signal; a bidirectional long short-term memory network is adopted to analyze and process the first output to obtain second output, and the bidirectional long short-term memory network is used for capturing forward time sequence dependence information and backward time sequence dependence information of the first output; and determining an original modulation state corresponding to the QAM signal according to the characteristic of each signal in the second output. According to the invention, the technical problems of high signal processing delay and poor equalization capability when a nonlinear equalization scheme in traditional coherent optical transmission is used for dealing with a high-order signal are solved.

Description

Technical Field

[0001] The present application relates to the field of neural networks, and more specifically, to a signal processing method, device, and electronic device. Background Art

[0002] With the explosive growth of emerging technologies such as 5G / 6G, the Metaverse, and computing networks, backbone optical networks face the triple pressure of ultra-high speeds, ultra-long distances, and ultra-high spectral efficiency. To address the Shannon limit, the mainstream backbone network in the industry uses coherent optical transmission with high-order quadrature amplitude modulation (such as 1024QAM and 4096QAM) in conjunction with spatial division multiplexing (SDM) to increase capacity. However, this also leads to an exponential increase in digital signal processing (DSP) complexity, posing a particularly severe challenge to nonlinear equalization algorithms. Nonlinear equalization schemes in traditional coherent optical transmission suffer from high signal processing latency and poor equalization capabilities when dealing with high-order signals.

[0003] To address the above-mentioned problems, no effective solutions have been proposed so far. Summary of the Invention

[0004] The embodiments of the present application provide a signal processing method, device, and electronic device to at least solve the technical problems of high signal processing delay and poor equalization capability of nonlinear equalization schemes in traditional coherent optical transmission when dealing with high-order signals.

[0005] According to one aspect of an embodiment of the present application, a signal processing method is provided, including: obtaining an orthogonal amplitude modulation (QAM) signal output by a digital signal processor at a receiving end; performing feature extraction on the QAM signal using a complex domain Transformer network to obtain a first output, wherein the complex domain Transformer network is used to analyze the global dependency of the QAM signal; performing analysis and processing on the first output using a bidirectional long short-term memory network to obtain a second output, wherein the bidirectional long short-term memory network is used to capture forward timing dependency information and backward timing dependency information of the first output; and determining the original modulation state corresponding to the QAM signal based on the characteristics of each signal in the second output.

[0006] Optionally, a complex domain Transformer network is used to extract features of the QAM signal to obtain a first output, including: obtaining an embedding dimension of the complex domain Transformer network and a dimension index of the embedding dimension; determining a position coding cardinality of the QAM signal based on the embedding dimension and the dimension index; determining a complex rotation position coding of each complex signal in the QAM signal based on the position coding cardinality; determining a complex attention score of each complex signal based on the complex rotation position coding; and determining the first output based on the complex attention score.

[0007] Optionally, based on the position coding cardinality, the complex rotation position coding of each complex signal in the QAM signal is determined, including: obtaining a baseband signal of a single complex signal at time t in the QAM signal, and obtaining a first weight, wherein the baseband signal consists of a real part and an imaginary part, and the distribution obeyed by the first weight is determined according to the embedding dimension; based on the first weight and the baseband signal, each complex signal is mapped to a high-dimensional space to obtain an embedded representation; based on the embedded representation and the position coding cardinality, the complex rotation position coding of each complex signal in the QAM signal is determined.

[0008] Optionally, a complex attention score for each complex signal is determined based on the complex rotated position encoding, including: determining the number of attention heads, query weights, key weights, and value weights in the complex domain Transformer network; determining the key dimension based on the number of heads and the embedding dimension; determining the query matrix based on the complex rotated position encoding and the query weight, determining the key matrix based on the complex rotated position encoding and the key weight, and determining the value matrix based on the complex rotated position encoding and the value weight; determining the complex attention score based on the query matrix, the conjugate transpose of the key matrix, the key dimension, and the value matrix.

[0009] Optionally, determining the first output based on the complex attention score includes: obtaining the attention output corresponding to all attention heads in the complex domain Transformer network, wherein the attention output is determined by the complex attention score; splicing the attention output of each attention head to obtain a splicing result, and projecting the splicing result to obtain a target result; extracting the center position feature of the target result, and determining the first output based on the center position feature, the second weight, the third weight and the imaginary part value of the complex signal, wherein the third weight is determined by the embedding dimension and the modulation format order.

[0010] Optionally, a bidirectional long short-term memory network is used to analyze and process the first output to obtain a second output, including: using a forward long short-term memory network in the bidirectional long short-term memory network to perform forward processing on the first output to obtain a first result; using a backward long short-term memory network in the bidirectional long short-term memory network to perform reverse processing on the first output to obtain a second result; and concatenating the forward hidden state in the first result and the backward hidden state in the second result to obtain the second output.

[0011] Optionally, determining the original modulation state corresponding to the QAM signal based on the characteristics of each signal in the second output includes: mapping the characteristics of each signal in the second output back to the corresponding constellation diagram to obtain a mapping result for each signal, wherein the constellation diagram contains multiple constellation points, and each constellation point represents a different modulation state; determining the distance between the mapping result of each signal and all constellation points, and determining the modulation state represented by the constellation point with the closest distance as the original modulation state of the corresponding signal.

[0012] According to another aspect of an embodiment of the present application, a signal processing device is also provided, including: an acquisition module for acquiring an orthogonal amplitude modulation (QAM) signal output by a digital signal processor at a receiving end; a feature extraction module for extracting features of the QAM signal using a complex domain Transformer network to obtain a first output, wherein the complex domain Transformer network is used to analyze the global dependency of the QAM signal; a processing module for analyzing and processing the first output using a bidirectional long short-term memory network to obtain a second output, wherein the bidirectional long short-term memory network is used to capture forward timing dependency information and backward timing dependency information of the first output; and a determination module for determining the original modulation state corresponding to the QAM signal based on the characteristics of each signal in the second output.

[0013] According to another aspect of the embodiments of the present application, an electronic device is also provided, including: a memory for storing program instructions; a processor, connected to the memory, for executing program instructions to implement the following functions: obtaining an orthogonal amplitude modulated QAM signal output by a digital signal processor at a receiving end; using a complex domain Transformer network to extract features of the QAM signal to obtain a first output, wherein the complex domain Transformer network is used to analyze the global dependency of the QAM signal; using a bidirectional long short-term memory network to analyze and process the first output to obtain a second output, wherein the bidirectional long short-term memory network is used to capture forward timing dependency information and backward timing dependency information of the first output; and determining the original modulation state corresponding to the QAM signal based on the characteristics of each signal in the second output.

[0014] According to another aspect of the embodiments of the present application, a non-volatile storage medium is provided, which includes a stored computer program, wherein a device where the non-volatile storage medium is located executes the above-mentioned signal processing method by running the computer program.

[0015] According to another aspect of the embodiments of the present application, a computer program product is provided, including computer instructions, which implement the above-mentioned signal processing method when executed by a processor.

[0016] In an embodiment of the present application, a quadrature amplitude modulated (QAM) signal output by a digital signal processor at a receiving end is obtained; a complex domain Transformer network is used to extract features of the QAM signal to obtain a first output, wherein the complex domain Transformer network is used to analyze the global dependency of the QAM signal; a bidirectional long short-term memory network is used to analyze and process the first output to obtain a second output, wherein the bidirectional long short-term memory network is used to capture forward timing dependency information and backward timing dependency information of the first output; and the original modulation state corresponding to the QAM signal is determined based on the features of each signal in the second output, thereby achieving the purpose of efficient nonlinear equalization and reducing signal processing delay, thereby realizing the technical effect of enhancing the accuracy of signal recovery, and further solving the technical problems of high signal processing delay and poor equalization capability of nonlinear equalization schemes in traditional coherent optical transmission when dealing with high-order signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0018] Figure 1 is a hardware structure block diagram of a computer terminal for implementing a signal processing method according to an embodiment of the present application;

[0019] Figure 2 is a flowchart of a signal processing method according to an embodiment of the present application;

[0020] Figure 3 is a structural diagram of a complex domain Transformer network according to an embodiment of the present application;

[0021] Figure 4 is a single LSTM recursive structure according to an embodiment of the present application;

[0022] Figure 5 This is a diagram of a 1024QAM coherent optical transmission communication simulation link system architecture according to an embodiment of the present application;

[0023] Figure 6 It is a combined network structure of Tranformer+bi-LSTM according to an embodiment of the present application;

[0024] Figure 7 is a comparison chart of simulation results according to an embodiment of the present application;

[0025] Figure 8 It is a structural diagram of a signal processing device according to an embodiment of the present application. DETAILED DESCRIPTION

[0026] In order to enable those skilled in the art to better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0027] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in a sequence other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0028] First, some nouns or terms that appear in the process of explaining the embodiments of this application are subject to the following explanations:

[0029] CV-Transformer (Complex-Valued Transformer, complex domain Transformer network): is a variant of the Transformer model based on the complex domain. Its core is to expand the real-valued operations in the traditional Transformer to complex operations to enhance the modeling capabilities of complex signals (such as frequency domain and phase-sensitive data).

[0030] Bidirectional Long Short-Term Memory (bidirectional long short-term memory) is a recurrent neural network (RNN) variant that combines forward LSTM and backward LSTM. It is used for sequence modeling tasks such as natural language processing and speech recognition. Its core feature is its ability to simultaneously capture both forward and reverse temporal dependencies of sequence data, thereby more comprehensively understanding contextual information.

[0031] In the field of optical transmission, intersymbol interference (ISI) caused by nonlinear fiber damage and device defects has become a core bottleneck limiting system performance. Nonlinear equalization schemes in traditional coherent optical transmission face severe challenges when dealing with high-order signals. Digital backpropagation algorithms compensate for impairments by inversely solving the fiber transmission equations. While theoretically capable of precise compensation, they require repeated Fourier transforms and nonlinear phase calculations. When transmission distances exceed 80 km, algorithm complexity increases dramatically. For every 10 km increase in distance, computational time increases by approximately 35%. This results in real-time processing delays of up to 15 ms for an 80 km link, far exceeding the 5 ms latency requirement of optical modules. Optical phase conjugation technology relies on link symmetry and requires the insertion of a phase conjugator at the fiber midpoint. In actual deployments, due to factors such as fiber aging and temperature fluctuations, the link loss deviation between the front and back halves often exceeds 0.5 dB / km, causing nonlinearity cancellation efficiency to drop to less than 50%. The computational complexity of the Volterra nonlinear equalizer (VNLE) increases exponentially when performing third-order and higher nonlinearity compensation. While limited memory depth can reduce computational complexity, the equalization capability for high-order signals at the outer constellation points is significantly weakened, particularly at the 20% forward error correction threshold where the error floor effect is noticeable. Therefore, traditional solutions struggle to strike an effective balance between performance and complexity.

[0032] In transmission links, nonlinear impairments in high-order M-QAM signals exhibit spatiotemporal coupling: the distortion of the current symbol arises not only from instantaneous nonlinear effects but also from the interaction of multiple preceding and subsequent symbols. Traditional CNN approaches are limited by the localized perceptual range of convolution kernels and struggle to capture long-range intersymbol interference (ISI).

[0033] In order to solve the problems existing in the related art, the embodiment of the present application provides a signal processing method, which can be run on Figure 1 In the computer terminal shown, the computer terminal is explained below.

[0034] The signal processing method embodiments provided in the embodiments of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Figure 1 FIG1 shows a hardware structure block diagram of a computer terminal for implementing a signal processing method. Figure 1As shown, the computer terminal 10 may include one or more (illustrated by 102a, 102b, ..., 102n in the figure) processors (the processor may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, and a transmission module 106 for communication functions connected via a wired and / or wireless network. In addition, it may also include: a display, a keyboard, a cursor control device, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the I / O interface), a network interface, and a BUS bus. It will be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 1 More or fewer components than shown, or with Figure 1 Different configurations shown.

[0035] It should be noted that the one or more processors and / or other data processing circuits described above may generally be referred to herein as "data processing circuitry." The data processing circuitry may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuitry may be a single, independent processing module, or may be incorporated in whole or in part into any of the other components of the computer terminal 10. As described in the embodiments of the present application, the data processing circuitry serves as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0036] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the signal processing method in the embodiment of the present application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, implementing the above-mentioned signal processing method. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include a memory remotely located relative to the processor, and these remote memories may be connected to the computer terminal 10 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0037] The transmission module 106 is configured to receive or transmit data via a network. A specific example of the aforementioned network may include a wireless network provided by the communications provider of the computer terminal 10. In one embodiment, the transmission module 106 includes a network interface controller (NIC), which can be connected to other network devices via a base station to enable communication with the Internet. In another embodiment, the transmission module 106 may be a radio frequency (RF) module, which is configured to communicate with the Internet wirelessly.

[0038] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the computer terminal 10 .

[0039] It should be noted that, in some optional embodiments, the above Figure 1 The computer terminal shown may include hardware elements (including circuits), software elements (including computer code stored on a computer-readable medium), or a combination of hardware elements and software elements. Figure 1 This is merely one example of a particular embodiment and is intended to illustrate the types of components that may be present in the computer terminal described above.

[0040] In the above-mentioned operating environment, an embodiment of the present application provides an embodiment of a signal processing method. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0041] Figure 2 is a flow chart of a signal processing method according to an embodiment of the present application, such as Figure 2 As shown, the method includes the following steps:

[0042] Step S202: Acquire a quadrature amplitude modulation (QAM) signal output by a digital signal processor at the receiving end.

[0043] In step S202, the digital signal processor (DSP) at the receiving end of the optical communication system is responsible for converting and processing the received optical signal into an electrical signal, and further performing digital signal processing on it to restore the originally transmitted signal information. Quadrature Amplitude Modulation (QAM) is a modulation technique that achieves higher data transmission rates and spectral efficiency within a given bandwidth by simultaneously performing amplitude modulation on both the real part (in-phase, I) and the imaginary part (quadrature, Q) of the signal. After the optical signal is transmitted to the receiving end, the optical receiver converts it into an electrical signal. The DSP then pre-processes these electrical signals, a process that includes but is not limited to downsampling, root-raised cosine matched filtering, orthogonal normalization, and dispersion equalization. The signal passes through a clock recovery module to obtain information before phase recovery, and then undergoes carrier phase recovery to obtain a constellation point signal of lower quality. After the preprocessing step, the DSP will output a complex-valued QAM signal sequence. Each signal point corresponds to a constellation point on the complex plane. These constellation points will have different distributions and quantities depending on the modulation format (such as 16QAM, 64QAM, 1024QAM, etc.).

[0044] Step S204: extract features of the QAM signal using a complex domain Transformer network to obtain a first output, wherein the complex domain Transformer network is used to analyze the global dependency of the QAM signal.

[0045] In step S204, the QAM signal sequence output by the DSP is input into the complex domain Transformer network in complex form. Each signal point in the QAM signal consists of an in-phase component (I) and a quadrature-phase component (Q), which together determine the position of the point on the complex plane. After being fed into the complex embedding layer of the complex domain Transformer network, the signal point is converted into a higher-dimensional complex vector to capture more signal information. The real and imaginary parts undergo independent linear transformations to maintain orthogonality and ensure that the signal's phase information is not lost. The complex domain Transformer network utilizes its core self-attention mechanism to capture relationships between signal points, particularly temporal dependencies. The self-attention layer allows each signal point to pay attention to other points in the entire signal sequence, thereby constructing a more comprehensive signal feature representation. To further enhance temporal modeling capabilities, rotational position encoding is introduced into the complex domain Transformer network. This encodes the position information of each signal point into a phase rotation matrix, which can present continuous phase offsets in the complex feature space, effectively modeling the dispersion accumulation effect in optical fiber transmission. After the signal is processed by multiple encoding layers and decoding layers, the first output is obtained.

[0046] Step S206: Use a bidirectional long short-term memory network to analyze and process the first output to obtain a second output, wherein the bidirectional long short-term memory network is used to capture forward temporal dependency information and backward temporal dependency information of the first output.

[0047] In the above step S206, the traditional LSTM network can only analyze sequence data from front to back (forward), while the bi-LSTM adds an LSTM layer from back to front (reverse) on this basis. This means that the processing of each signal point not only takes into account the signal sequence before it (forward), but also the signal sequence after it (reverse), thereby being able to capture more comprehensive timing information. The first output (signal features after processing from the complex domain Transformer network) serves as the input sequence of the bi-LSTM. The forward LSTM layer processes the signal features from front to back in chronological order, accumulates the influence of historical signals, and captures the forward trend of inter-symbol interference (ISI), which helps to identify the nonlinear effects and dispersion accumulation caused by the influence of the previous signal on the signal point. The reverse LSTM layer starts reverse processing from the end of the signal sequence, collects information about future signals, and helps to predict the interference that subsequent signals may cause to the current signal point, thereby more actively performing signal compensation. After optimization by the Bi-LSTM network, the second output is obtained.

[0048] Step S208: determining the original modulation state corresponding to the QAM signal according to the characteristics of each signal in the second output.

[0049] In the above step S208, the original transmitted binary information is restored by mapping the characteristics of each signal in the second output to the closest constellation point, so as to determine the original modulation state corresponding to the QAM signal.

[0050] Through steps S202 to S208 described above, efficient nonlinear equalization and reduced signal processing delay are achieved, thereby achieving the technical effect of enhancing the accuracy of signal recovery. This further addresses the technical issues of high signal processing delay and poor equalization capabilities of traditional nonlinear equalization schemes in coherent optical transmission when dealing with high-order signals. This is explained below.

[0051] In step S204 of the above-mentioned signal processing method, a complex domain Transformer network is used to perform feature extraction on the QAM signal to obtain a first output, including: obtaining the embedding dimension of the complex domain Transformer network and the dimension index of the embedding dimension; determining the position coding cardinality of the QAM signal based on the embedding dimension and the dimension index; determining the complex rotation position coding of each complex signal in the QAM signal based on the position coding cardinality; determining the complex attention score of each complex signal based on the complex rotation position coding; and determining the first output based on the complex attention score.

[0052] Figure 3 This is a structural diagram of a complex domain Transformer network according to an embodiment of the present application. The complex domain Transformer network is composed of multiple encoders and multiple decoders. For the input data, normalization is used to ensure that the amplitude of the signal is normalized to the unit circle to avoid gradient explosion. The sequence half-length window length (referring to half the length of the window used to capture the dependency between elements in the sequence when performing feature extraction or prediction) is specified to be 16. Considering the computational complexity and the expressive power of the network, the complex domain Transformer network adopts an embedding dimension d of 512. model , set the position encoding cardinality of the input signal to:

[0053]

[0054] Here, k is the dimension index of the embedding dimension.

[0055] In some embodiments of the present application, the embedding dimension is the dimensional size of mapping the input QAM signal to the high-dimensional complex space; when processing complex input, the dimension index is used to distinguish the features of the real part and the imaginary part. For example, for a (512)-dimensional embedding layer, the first (256) dimensions correspond to the real part features, and the second (256) dimensions correspond to the imaginary part features. In the complex domain Transformer network, the position encoding basis generally refers to the parameters used to generate the (complex) rotation position encoding (RoPE), the purpose of which is to encode the absolute position information into a phase rotation matrix to maintain the phase continuity of the signal. The complex attention score is determined by the complex rotation position encoding, and the first output of the complex domain Transformer network is determined based on the complex attention score.

[0056] In the above steps, the complex rotation position coding of each complex signal in the QAM signal is determined based on the position coding cardinality, including: obtaining a baseband signal of a single complex signal at time t in the QAM signal, and obtaining a first weight, wherein the baseband signal is composed of a real part and an imaginary part, and the distribution obeyed by the first weight is determined according to the embedding dimension; based on the first weight and the baseband signal, each complex signal is mapped to a high-dimensional space to obtain an embedded representation; based on the embedded representation and the position coding cardinality, the complex rotation position coding of each complex signal in the QAM signal is determined.

[0057] In some embodiments of the present application, the baseband signal of a single complex signal output by the receiving end DSP at time t is x t =I t +jQ t , j represents the imaginary number, I represents the real part, Q represents the imaginary part, and the baseband signal ( Represents a sequence of N complex elements). Map each complex signal to a high-dimensional space:

[0058]

[0059] Among them, W embed Represents the first weight, weight W embed Independent initialization, obey E t Denotes the embedded representation, l denotes the output dimension of the embedding layer. From this, we can determine the complex rotation position encoding (RoPE):

[0060]

[0061] in, represents complex rotation position encoding, ⊙ represents element-by-element complex multiplication, and after encoding, the feature amplitude remains unchanged, while the phase rotates with the position.

[0062] In the above steps, the complex attention score of each complex signal is determined based on the complex rotation position encoding, including: determining the number of attention heads, query weights, key weights and value weights in the complex domain Transformer network; determining the key dimension based on the number of heads and the embedding dimension; determining the query matrix based on the complex rotation position encoding and the query weight, determining the key matrix based on the complex rotation position encoding and the key weight, and determining the value matrix based on the complex rotation position encoding and the value weight; determining the complex attention score based on the conjugate transpose of the query matrix and the key matrix, the key dimension and the value matrix.

[0063] In some embodiments of the present application, the key / value dimension d of attention is determined k / d v , let d k =d v =d model / h, h = 16 (indicates the number of attention heads). Multi-head division of labor can better focus on various damage types, and the generated matrix Q / K / V matrix is:

[0064]

[0065] Among them, W Q represents the query weight, W K represents the key weight, W V represents the value weight, Q represents the query matrix, K represents the key matrix, and V represents the value matrix.

[0066] The calculation formula of the complex attention score is as follows:

[0067]

[0068] Among them, Attention(Q,K,V) represents the complex attention score and H is the conjugate transpose.

[0069] In the above steps, the first output is determined based on the complex attention score, including: obtaining the attention output corresponding to all attention heads in the complex domain Transformer network, wherein the attention output is determined by the complex attention score; splicing the attention output of each attention head to obtain a splicing result, and projecting the splicing result to obtain a target result; extracting the center position feature of the target result, and determining the first output based on the center position feature, the second weight, the third weight and the imaginary part value of the complex signal, wherein the third weight is determined by the embedding dimension and the modulation format order.

[0070] In some embodiments of the present application, the multi-head attention mechanism in the complex domain Transformer network allows the network to simultaneously focus on signals in the sequence from different perspectives. Each attention head calculates its own set of complex attention scores, reflecting the correlation between the signal points in the sequence. The output of each attention head is spliced ​​together to form a set of multi-head attention outputs, which is the above-mentioned splicing result. The splicing operation expands the width of the signal representation and combines the diverse features extracted by different attention heads. The spliced ​​multi-head attention output passes through a linear projection layer to convert it into a vector of appropriate dimension to meet the input requirements of subsequent processing or the network. The projection step helps to integrate information and reduce dimensions for further processing. The formulas corresponding to splicing and projection are as follows:

[0071]

[0072] Among them, MultiHead represents the target result, Concat represents the splicing process, Head represents the attention output, and W O Denotes the linear projection layer. The hidden layer dimension of the feedforward network is d ff =4d model =2048, the activation function uses ReLU, and a complex gated linear unit is introduced to enhance the nonlinear modeling capability. Then the complex output is merged, and the complex residuals of each layer are added and normalized.

[0073] From the projected target result, extract the feature vector of the center position of the sequence window to obtain the above center position feature When determining the first output based on the center position feature, the second weight, the third weight, and the imaginary part value of the complex signal, it can be determined by the following formula:

[0074] P=Softmax(Re(H L W c )+Im(H L W'c )) (7)

[0075] Among them, P represents the output probability, W c Represents the second weight, W' c represents the third weight, Im represents the imaginary part of the complex signal. M is the modulation format order, 1024. By decomposing 1024QAM into a 32*32 grid and performing coarse and fine two-stage classification, the network's computational complexity can be greatly reduced.

[0076] In step S206 of the above-mentioned signal processing method, a bidirectional long short-term memory network is used to analyze and process the first output to obtain a second output, including: using a forward long short-term memory network in the bidirectional long short-term memory network to perform forward processing on the first output to obtain a first result; using a backward long short-term memory network in the bidirectional long short-term memory network to perform reverse processing on the first output to obtain a second result; and concatenating a forward hidden state in the first result and a backward hidden state in the second result to obtain a second output.

[0077] In some embodiments of the present application, the first output is a sequence of signal features processed by a complex-domain Transformer network, which serves as the input to a bidirectional LSTM network. The forward LSTM (F-LSTM) starts from the start of the sequence and processes the sequence symbol by symbol in the forward time direction. It updates its internal state based on the sequence information of the current symbol and its previous symbols, capturing the forward temporal dependencies of the signal, i.e., the accumulated historical influences. At each time step, the F-LSTM updates its internal memory cell state (through the dynamic interaction of the forget gate, input gate, and output gate) and output state, which reflects the comprehensive characteristics of the current symbol and its historical context. Unlike the F-LSTM, the backward LSTM (B-LSTM) starts from the end of the sequence and processes the sequence symbol by symbol in the reverse time direction. It updates its internal state based on the sequence information of the current symbol and its subsequent symbols, capturing the backward temporal dependencies of the signal, i.e., the expected future influences. At each time step, the B-LSTM similarly updates its internal memory cell state and output state, which reflect the comprehensive characteristics of the current symbol and its future context. At each time step, the first result (forward hidden state) obtained by the F-LSTM is concatenated with the second result (backward hidden state) obtained by the B-LSTM to form a wider feature vector. This concatenation operation essentially integrates the forward and backward dependency information of the signal. The concatenated feature vector sequence constitutes the second output, which contains a more comprehensive representation of the bidirectional temporal dependency information for each signal point.

[0078] Specifically, in optical communication systems, bi-LSTM effectively compensates for inter-symbol interference (ISI) caused by fiber nonlinearity and dispersion by capturing the forward and backward timing dependencies of signals. The network, with its unique gating characteristics, can selectively perform dynamic memory screening on information, effectively alleviating the vanishing gradient problem of recurrent neural networks. Figure 4 Given is a single LSTM recursive structure, where i t ,f t ,o t They represent the input gate, forget gate and output gate of the network structure respectively, and the internal state c t and output h t From the input state x t and the previous state h t-1 The expressions of input, forgetting and output are determined together as follows:

[0079] i t =σ(W i x t +U i h t-1 +b i )

[0080] f t =σ(W f x t +U f h t-1 +b f ) (8)

[0081] o t =σ(W o x t +U o h t-1 +b o )

[0082] Among them, W i 、W f 、W o is the input weight, U i 、U f 、U o is the recursive weight, b i 、b f 、b o is the network bias, where σ is the sigmoid activation function. Figure 4 It can be seen that the internal state memory unit c t The update depends not only on the forget gate output f t , also depends on the current internal state h t , candidate state c t ′, and finally multiplied by the output gate to get the state h passed to the next level t .

[0083] In coherent optical communication systems, the spatiotemporal coupling of intersymbol interference (ISI) requires equalization algorithms to model long-range temporal sequences. A bidirectional long short-term memory (bi-LSTM) network overcomes the limitations of traditional unidirectional recurrent networks by implementing a reverse signal processing mechanism. It constructs two parallel, oppositely oriented signal processing paths: the forward propagation layer extracts historical impairment signatures symbol by symbol according to the actual transmission direction of the optical signal (positive temporal order), while the reverse propagation layer analyzes the pre-distorted interference patterns of future symbols in a mirror-symmetric manner (reverse temporal order). While the two paths share the original input sequence, they each maintain independent memory cell states. A forget gate in the forward layer dynamically filters out invalid historical information beyond 32 symbols, while an input gate in the reverse layer prioritizes phase correlation features over the next 15 symbols. By real-time concatenation and fusion of bidirectional hidden states (with dimensions expanded to 384), the network synchronously integrates past and future impairment signature maps at each moment, forming a joint representation of nonlinearity, dispersion, and noise.

[0084] In step S208 of the above-mentioned signal processing method, determining the original modulation state corresponding to the QAM signal based on the characteristics of each signal in the second output includes: mapping the characteristics of each signal in the second output back to the corresponding constellation diagram to obtain a mapping result for each signal, wherein the constellation diagram includes multiple constellation points, and each constellation point represents a different modulation state; determining the distance between the mapping result of each signal and all constellation points, and determining the modulation state represented by the constellation point with the closest distance as the original modulation state of the corresponding signal.

[0085] In some embodiments of the present application, QAM (Quadrature Amplitude Modulation) is a modulation method that represents data in the form of constellation points on the complex plane. The constellation diagram contains all possible modulation states, and the position of each constellation point corresponds to a specific bit combination. The second output contains optimized signal features, which are obtained by the Bi-LSTM network after fully considering timing dependence and impairments such as dispersion and nonlinear effects. Mapping these optimized features back to the constellation diagram means converting the position of each signal point into a point on the complex plane, which should be close to the position of the constellation point of its original transmission. Calculate the distance between each signal point mapping result in the second output and all constellation points in the constellation diagram, such as the Euclidean distance. For each signal point, find the constellation point closest to it. The modulation state represented by the closest constellation point is determined as the original modulation state of the signal point.

[0086] The embodiment of the present application adopts the complex Transformer as the core feature extraction module, and its self-attention mechanism can establish the global dependency relationship of the full sequence symbols. Specifically, after preprocessing, the complex value signal sequence at the receiving end is projected into the high-dimensional feature space through the complex embedding layer, where the real part and the imaginary part are linearly transformed and orthogonality is maintained. In the complex self-attention calculation process, the query matrix Q, the key matrix K, and the value matrix V are all generated in complex form, and the signal phase information is retained through the conjugate transpose operation. The attention weight calculation adopts the improved complex Softmax function to stabilize the gradient propagation. In order to further enhance the model's perception of the timing characteristics of the signal, the embodiment of the present application also introduces complex rotation position encoding (RoPE) to encode the absolute position information into a phase rotation matrix. This encoding method makes adjacent symbols present continuous phase offsets in the complex feature space, effectively modeling the dispersion accumulation effect in optical fiber transmission. After processing by the multi-level Transformer coding layer, the feature sequence is input into the bidirectional LSTM network for timing modeling. The bidirectional architecture synchronously processes signal sequences using two layers of forward and backward LSTM units. The forward layer captures the cumulative impact of historical symbols, while the backward layer learns the interference pattern of future symbols. Ultimately, the bidirectional hidden states are fused through a splicing operation. Simulation experiments show that in a 1024QAM coherent optical transmission system with an optical signal-to-noise ratio (OSNR) of 25-35dB over 80km of fiber, the complex Transformer + bi-LSTM solution improves bit error rate performance by over 1dB compared to traditional VNLE and CV-NN + bi-LSTM methods, while maintaining the same computational complexity.

[0087] Therefore, the signal processing method provided in the embodiment of the present application, on the basis of solving the uneven defects in traditional coherent optical transmission, further makes up for the shortcomings of the neural network solution, and solves the nonlinear damage problem in the optical transmission link by integrating the global feature extraction advantages of Transformer and the precise timing modeling capabilities of bi-LSTM. When used as an equalizer, LSTM has excellent timing modeling capabilities, but it is overly dependent on the characteristics of the input data, and long sequence information such as high-order QAM signals contains complex features, which are not limited to dispersion accumulation and various nonlinear factors. As data input, it is easy to cause the network to fall into an overfitting state. By introducing a solution for feature pre-extraction of neural networks such as CNN, it is necessary to split the real and imaginary parts of the complex signal, which affects the original phase information of the signal, and the serial processing method makes the entire network run too long, affecting actual use. By introducing a complex domain Transformer for pre-processing, the overall feature optimization processing of long sequences can be achieved, and its self-attention mechanism can model the interaction between symbols of any distance, solve the local convolution kernel limitation of CNN, and the parallel attention mechanism of Transformer greatly improves the training speed compared to the serial CNN convolution operation. Complex signal processing preserves the signal's complete phase information, avoiding the error accumulation caused by IQ component separation. Furthermore, the multi-head attention mechanism simultaneously focuses on different sources of impairment in the transmission system, significantly improving the final equalization effect.

[0088] In order to better understand the solution of the embodiment of the present application, the following Figure 5 The following describes the signal processing flow:

[0089] Step 1: Build a high-order coherent optical transmission system simulation link model.

[0090] Figure 5 This is a diagram of a 1024QAM coherent optical transmission communication simulation link system architecture according to an embodiment of the present application. The detailed process is as follows:

[0091] In the optical signal transmission link, at the transmitter, a binary pseudo-random sequence (PRBS15) is mapped to a QAM signal using Gray coding. To achieve high spectral efficiency, probabilistic shaping technology is used to split the QAM signal into two PAM16 signals for probabilistic shaping coding. The information entropy (the average amount of information describing the uncertainty of the data source or the information content) is set to 8.5 bits / s / Hz. The encoded QAM signal is upsampled and filtered using a root-raised cosine filter to conform to the system's spectral characteristics and reduce inter-symbol interference (ISI). The signal level after shaping and filtering is input into an arbitrary waveform generator (AWG) for digital-to-analog conversion. The converted electrical signal drives a modulator (IQ modulator) to generate a 1024QAM signal, which is then transmitted to a narrow-linewidth laser for transmission. Before entering the optical fiber, the signal is amplified to the specified power by an erbium-doped fiber amplifier (EDFA) and then connected to a variable optical attenuator (VOA) for OSNR (Optical Signal-to-Noise Ratio) adjustment. The simulation experiment link uses 80 km of standard single-mode optical fiber. After transmission through the fiber, the signal is filtered by a narrow-bandwidth optical filter (OBPF). At the receiver, the local oscillator (LO) is used for frequency beats. A mixer restores the original signal, which is then converted to digital form by an analog-to-digital conversion module. The digital signal preprocessing at the receiver includes downsampling, root-raised cosine matched filtering, orthogonal normalization, and dispersion equalization. The signal passes through a clock recovery module to obtain pre-phase information. After carrier phase recovery, the constellation point signal is obtained, which has poorer quality. This signal contains significant nonlinear impairments. This 1024QAM signal serves as the initial input for the proposed scheme, i.e., the quadrature amplitude modulated (QAM) signal output by the digital signal processor at the receiver.

[0092] Step 2: Build the complex domain Transformer network and bi-LSTM network. The process of building the two networks and determining the original modulation state (corresponding to Figure 5 The signal decision in ( ) has been described in detail above and will not be repeated here. Figure 6 The joint network structure of Tranformer+bi-LSTM is shown.

[0093] Step 3: Network training and testing.

[0094] For the nonlinear equalization scheme of the 1024QAM modulation format, the embodiment of the present application adopts the complex domain Transformer combined with the bidirectional LSTM architecture for training and testing. The data set construction covers the range of 25-34dB optical signal-to-noise ratio. For each signal-to-noise ratio level, 10 groups of signals are collected at intervals of 1dB. Each group contains 12,000 complex-valued symbols. By intercepting 16 symbols before and after, a context window of 33 symbols is formed to form an original data set of 15*10*12000. At the same time, random phase noise, polarization mode dispersion and nonlinear phase shift are injected to simulate real damage scenarios. The input data is preprocessed by the complex Transformer, and the global features are extracted by the 16-head self-attention mechanism. The phase continuity is retained by combining the rotation position encoding. The original standard data is used as the reference value for training. After the dimensionality reduction information of the feature extraction is output, it is input into the bidirectional LSTM network. 96 neurons are configured in each direction. The memory window is adaptively adjusted through the dynamic gating mechanism to fuse the forward and backward time series features. During the training phase, a joint amplitude-phase loss function was used. The ADAM optimizer was selected with a cosine annealing learning rate schedule. The initial learning rate was set to 0.004, and the batch size was dynamically adjusted to balance convergence speed and stability. An early stopping mechanism was triggered after 100 consecutive validation loss cycles without improvement. During the deployment phase, the trained model was integrated into the receiving DSP link, and simulation results were obtained using signals within the OSNR range.

[0095] Step 4: Combine neural network test results.

[0096] In the simulation experiment link of 80km of 1024QAM coherent optical transmission, according to the OSNR (Optical Signal-to-Noise Ratio) range of 25-34dB set in step 3, the test data set within the range is applied to the trained network, and the traditional VNLE and CV-NN+bi-LSTM are used as comparison schemes to verify the effect of the scheme in the embodiment of this application. The results of the scheme proposed in the embodiment of this application are as follows Figure 7 As shown, the star point diagram is the bit error rate curve of the traditional VNLE, and the results show that the bit error rate threshold is reached at 29dB; compared with the VNLE solution, the CV-CNN+bi-LSTM neural network solution has a significant improvement in the nonlinear equalization effect, with a 1.3dB performance improvement in the threshold; the solution proposed in the embodiment of the present application is more targeted in feature extraction, and has a 2.5dB performance improvement compared to the VNLE solution. It can be seen from the thermal constellation diagram that the Transformer+bi-LSTM solution based on the complex domain has a better correction effect on the outer circle constellation points, indicating that the nonlinear damage is further alleviated and a more convergent effect is achieved.

[0097] Combined with the above simulation experimental results, it can be seen that the signal processing method provided by the embodiment of the present application has the following advantages: (1) The present application is based on the Transformer network in the complex domain, which retains the phase information of the original data to the maximum extent, avoids the problem of error accumulation caused by the IQ component separation processing, and also maintains the signal phase continuity throughout the process through the complex embedding layer, rotation position encoding and complex self-attention calculation. It has a significant gain effect on the nonlinear processing of ultra-high-order coherent optical 1024QAM signals; (2) The present application adopts a Transformer network based on the attention mechanism, and optimizes the problem of poor nonlinear suppression caused by relying on fixed-size local convolution kernels in traditional neural network solutions by modeling the association of full-sequence symbols. Through long sequence global feature extraction, the multi-head attention mechanism focuses on different damage sources at the same time, which reduces the pressure of feature extraction in subsequent LSTM processing. The two work together to compensate for the accumulated dispersion and nonlinearity, and enhance the robustness of high-order signal nonlinear processing; (3) Compared with traditional convolution operations, this application uses a parallel processing mechanism to greatly improve the training speed, and the output layer also reduces the computational complexity of the network through coarse and fine two-stage classification, which is more conducive to practical application and provides a reference solution for the development of existing optical modules.

[0098] Figure 8 is a structural diagram of a signal processing device according to an embodiment of the present application, such as Figure 8 As shown, the device includes:

[0099] An acquisition module 40 is configured to acquire a quadrature amplitude modulation (QAM) signal output by a digital signal processor at a receiving end;

[0100] a feature extraction module 42 configured to extract features from the QAM signal using a complex domain Transformer network to obtain a first output, wherein the complex domain Transformer network is configured to analyze global dependencies of the QAM signal;

[0101] a processing module 44 configured to analyze and process the first output using a bidirectional long short-term memory network to obtain a second output, wherein the bidirectional long short-term memory network is configured to capture forward temporal dependency information and backward temporal dependency information of the first output;

[0102] The determination module 46 is configured to determine an original modulation state corresponding to the QAM signal according to a feature of each signal in the second output.

[0103] Through the acquisition module, feature extraction module, processing module and determination module in the above signal processing device,

[0104] In the feature extraction module in the above-mentioned signal processing device, the feature extraction module is also used to obtain the embedding dimension of the complex domain Transformer network and the dimension index of the embedding dimension; determine the position coding cardinality of the QAM signal based on the embedding dimension and the dimension index; determine the complex rotation position coding of each complex signal in the QAM signal based on the position coding cardinality; determine the complex attention score of each complex signal based on the complex rotation position coding; and determine the first output based on the complex attention score.

[0105] In the feature extraction module in the above-mentioned signal processing device, the feature extraction module is also used to obtain the baseband signal of a single complex signal in the QAM signal at time t, and to obtain a first weight, wherein the baseband signal is composed of a real part and an imaginary part, and the distribution obeyed by the first weight is determined according to the embedding dimension; based on the first weight and the baseband signal, each complex signal is mapped to a high-dimensional space to obtain an embedded representation; based on the embedded representation and the position coding base, the complex rotation position coding of each complex signal in the QAM signal is determined.

[0106] In the feature extraction module in the above-mentioned signal processing device, the feature extraction module is also used to determine the number of attention heads, query weights, key weights and value weights in the complex domain Transformer network; determine the key dimension based on the number of heads and the embedding dimension; determine the query matrix based on the complex rotation position encoding and the query weight, determine the key matrix based on the complex rotation position encoding and the key weight, and determine the value matrix based on the complex rotation position encoding and the value weight; determine the complex attention score based on the query matrix, the conjugate transpose of the key matrix, the key dimension and the value matrix.

[0107] In the feature extraction module in the above-mentioned signal processing device, the feature extraction module is also used to obtain the attention output corresponding to all attention heads in the complex domain Transformer network, wherein the attention output is determined by the complex attention score; splicing the attention output of each attention head to obtain the splicing result, and projecting the splicing result to obtain the target result; extracting the center position feature of the target result, and determining the first output based on the center position feature, the second weight, the third weight and the imaginary part value of the complex signal, wherein the third weight is determined by the embedding dimension and the modulation format order.

[0108] In the processing module in the above-mentioned signal processing device, the processing module is also used to use the forward long short-term memory network in the bidirectional long short-term memory network to perform forward processing on the first output to obtain a first result; use the backward long short-term memory network in the bidirectional long short-term memory network to perform reverse processing on the first output to obtain a second result; and splice the forward hidden state in the first result and the backward hidden state in the second result to obtain a second output.

[0109] In the determination module in the above-mentioned signal processing device, the determination module is further used to map the characteristics of each signal in the second output back to the corresponding constellation diagram to obtain a mapping result for each signal, wherein the constellation diagram contains multiple constellation points, and each constellation point represents a different modulation state; determine the distance between the mapping result of each signal and all constellation points, and determine the modulation state represented by the constellation point with the closest distance as the original modulation state of the corresponding signal.

[0110] It should be noted that Figure 8 The signal processing apparatus shown is used to perform Figure 2 The signal processing method shown in the figure, therefore the relevant explanations in the above signal processing method are also applicable to the signal processing device, and will not be repeated here.

[0111] An embodiment of the present application also provides an electronic device, which includes a memory and a processor, wherein the memory is used to store program instructions; the processor is connected to the memory and is used to execute program instructions to implement the following functions: obtaining an orthogonal amplitude modulation (QAM) signal output by a digital signal processor at a receiving end; using a complex domain Transformer network to extract features of the QAM signal to obtain a first output, wherein the complex domain Transformer network is used to analyze the global dependency of the QAM signal; using a bidirectional long short-term memory network to analyze and process the first output to obtain a second output, wherein the bidirectional long short-term memory network is used to capture forward timing dependency information and backward timing dependency information of the first output; and determining the original modulation state corresponding to the QAM signal based on the characteristics of each signal in the second output.

[0112] It should be noted that the above electronic equipment is used to perform Figure 2 The signal processing method shown in the figure, therefore the relevant explanations in the above signal processing method are also applicable to the electronic device and will not be repeated here.

[0113] An embodiment of the present application also provides a non-volatile storage medium, which includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the following signal processing method by running the computer program: obtaining an orthogonal amplitude modulation (QAM) signal output by a digital signal processor at a receiving end; using a complex domain Transformer network to extract features of the QAM signal to obtain a first output, wherein the complex domain Transformer network is used to analyze the global dependency of the QAM signal; using a bidirectional long short-term memory network to analyze and process the first output to obtain a second output, wherein the bidirectional long short-term memory network is used to capture forward timing dependency information and backward timing dependency information of the first output; and determining the original modulation state corresponding to the QAM signal based on the characteristics of each signal in the second output.

[0114] It should be noted that the above non-volatile storage medium is used to execute Figure 2 Therefore, the relevant explanations in the above signal processing method are also applicable to the non-volatile storage medium and will not be repeated here.

[0115] An embodiment of the present application further provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the signal processing method in each embodiment of the present application.

[0116] The embodiments of the present application also provide a computer program, which, when executed by a processor, implements the steps of the signal processing method in each embodiment of the present application.

[0117] The serial numbers of the above embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0118] In the above embodiments of the present application, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0119] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only exemplary. For example, the division of the units can be a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0120] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0121] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0122] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk.

[0123] The above is only a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A signal processing method, characterized in that: include: Obtaining a quadrature amplitude modulation (QAM) signal output by a digital signal processor at a receiving end; Performing feature extraction on the QAM signal using a complex domain Transformer network to obtain a first output, wherein the complex domain Transformer network is used to analyze the global dependency of the QAM signal; Using a bidirectional long short-term memory network to analyze and process the first output to obtain a second output, wherein the bidirectional long short-term memory network is used to capture forward temporal dependency information and backward temporal dependency information of the first output; An original modulation state corresponding to the QAM signal is determined according to a characteristic of each signal in the second output.

2. The method according to claim 1, characterized in that A complex domain Transformer network is used to extract features from the QAM signal to obtain a first output, including: Obtain the embedding dimension of the complex domain Transformer network and the dimension index of the embedding dimension; Determining a position coding cardinality of the QAM signal according to the embedding dimension and the dimension index; Determining a complex rotation position code of each complex signal in the QAM signal according to the position code base; Determining a complex attention score for each complex signal based on the complex rotational position encoding; Determine the first output based on the multiple attention scores.

3. The method according to claim 2, characterized in that Determining the complex rotation position code of each complex signal in the QAM signal according to the position code base includes: Obtaining a baseband signal of a single complex signal at time t in the QAM signal, and obtaining a first weight, wherein the baseband signal consists of a real part and an imaginary part, and a distribution obeyed by the first weight is determined according to the embedding dimension; Mapping each complex signal to a high-dimensional space according to the first weight and the baseband signal to obtain an embedded representation; A complex rotational position code of each complex signal in the QAM signal is determined according to the embedded representation and the position code base.

4. The method according to claim 2, characterized in that Determining a complex attention score for each complex signal based on the complex rotation position encoding includes: Determining the number of attention heads, query weights, key weights, and value weights in the complex domain Transformer network; Determining a key dimension based on the number of heads and the embedding dimension; Determine a query matrix based on the complex rotational position code and the query weight, determine a key matrix based on the complex rotational position code and the key weight, and determine a value matrix based on the complex rotational position code and the value weight; The complex attention score is determined based on the query matrix, the conjugate transpose of the key matrix, the key dimension and the value matrix.

5. The method according to claim 2, characterized in that Determining the first output according to the plurality of attention scores includes: Obtaining attention outputs corresponding to all attention heads in the complex-domain Transformer network, wherein the attention outputs are determined by the complex attention scores; Concatenate the attention outputs of each attention head to obtain a concatenated result, and project the concatenated result to obtain a target result; Extract the center position feature of the target result, and determine the first output based on the center position feature, the second weight, the third weight and the imaginary part value of the complex signal, wherein the third weight is determined by the embedding dimension and the modulation format order.

6. The method according to claim 1, characterized in that The first output is analyzed and processed using a bidirectional long short-term memory network to obtain a second output, including: Using a forward long short-term memory network in the bidirectional long short-term memory network to perform forward sequence processing on the first output to obtain a first result; Reverse processing is performed on the first output using a backward long short-term memory network in the bidirectional long short-term memory network to obtain a second result; The forward hidden state in the first result and the backward hidden state in the second result are concatenated to obtain the second output.

7. The method according to claim 1, characterized in that Determining an original modulation state corresponding to the QAM signal according to a characteristic of each signal in the second output includes: Mapping the characteristics of each signal in the second output back to a corresponding constellation diagram to obtain a mapping result for each signal, wherein the constellation diagram includes multiple constellation points, each constellation point representing a different modulation state; The distance between the mapping result of each signal and all constellation points is determined, and the modulation state represented by the constellation point with the closest distance is determined as the original modulation state of the corresponding signal.

8. A signal processing device, characterized in that: include: An acquisition module is used to acquire a quadrature amplitude modulation (QAM) signal output by a digital signal processor at a receiving end; a feature extraction module, configured to extract features from the QAM signal using a complex domain Transformer network to obtain a first output, wherein the complex domain Transformer network is configured to analyze a global dependency of the QAM signal; a processing module, configured to analyze and process the first output using a bidirectional long short-term memory network to obtain a second output, wherein the bidirectional long short-term memory network is configured to capture forward temporal dependency information and backward temporal dependency information of the first output; A determination module is used to determine the original modulation state corresponding to the QAM signal according to the characteristics of each signal in the second output.

9. An electronic device, characterized in that: include: a memory for storing program instructions; A processor, connected to the memory, is used to execute program instructions that implement the following functions: obtaining an orthogonal amplitude modulation (QAM) signal output by a digital signal processor at a receiving end; using a complex domain Transformer network to extract features of the QAM signal to obtain a first output, wherein the complex domain Transformer network is used to analyze the global dependency of the QAM signal; using a bidirectional long short-term memory network to analyze and process the first output to obtain a second output, wherein the bidirectional long short-term memory network is used to capture forward timing dependency information and backward timing dependency information of the first output; and determining the original modulation state corresponding to the QAM signal based on the characteristics of each signal in the second output.

10. A non-volatile storage medium, characterized in that: The non-volatile storage medium includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the signal processing method according to any one of claims 1 to 7 by running the computer program.

11. A computer program product comprising computer instructions, characterized in that When the computer instructions are executed by a processor, the signal processing method according to any one of claims 1 to 7 is implemented.

Images