Signal monitoring model training method, signal monitoring method, and electronic device

Abstract

The application discloses a signal monitoring model training method, a signal monitoring method and electronic equipment, and relates to the technical field of signal processing. The signal monitoring model embeds a depth separable convolution layer, a multi-scale parallel structure and a compression-excitation attention mechanism module. The depth separable convolution layer first expands the feature channel space, the multi-scale parallel structure simultaneously extracts multi-dimensional signal features, and the compression-excitation attention mechanism module strengthens effective features to cover more feature requirements of modulation types and avoid adaptation to only a single feature modulation format. Moreover, by aggregating the depth of multi-scale features and enhancing the channel dimension of features, richer deep transmission features can be extracted, and complex features of high-order modulation signals can be accurately captured and strengthened, thereby improving the adaptability to high-order modulation format signals, and the accuracy of signal transmission feature monitoring can be improved when the signal monitoring model is used for monitoring.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, and in particular to a training method for a signal monitoring model, a signal monitoring method, and an electronic device. Background Technology

[0002] The accuracy of signal transmission feature monitoring, such as signal-to-noise ratio (SNR) and modulation format identification (MFI), is crucial for the optimization and development of signal transmission scenarios. Related monitoring methods rely on manual extraction of signal features (such as amplitude, phase, and spectral characteristics), requiring complex hardware (such as high-speed oscilloscopes and spectrum analyzers). The influence of human factors leads to low detection efficiency and accuracy. Further, related technologies utilize convolutional neural networks or lightweight networks based on basic structures to identify signal transmission features. However, these methods are designed for specific signal features (such as amplitude only, or only a single type of time / frequency domain feature), and their network structures can only adapt to the corresponding modulation format, failing to cover more modulation types. Furthermore, the extraction depth and dimensionality are insufficient, resulting in limited adaptability to higher-order modulation format signals and limited accuracy in signal transmission feature monitoring.

[0003] Therefore, improving the accuracy of signal transmission characteristic monitoring is a technical problem that urgently needs to be solved by those in this field. Summary of the Invention

[0004] This invention provides a training method for a signal monitoring model, a signal monitoring method, and an electronic device, to at least solve the problem of low accuracy in signal transmission feature monitoring in related technologies.

[0005] This invention provides a training method for a signal monitoring model, comprising:

[0006] The spatial distribution feature map of the sample signal and the actual signal transmission features are input into the initial feature extraction module to obtain a preliminary feature map; the sample signal contains multiple types of modulation signals with different modulation orders.

[0007] The initial feature map is expanded by using depthwise separable convolutional layers to obtain expanded multi-channel feature maps. The expanded multi-channel feature maps are then extracted using a multi-scale parallel structure to obtain multi-branch feature maps. Finally, the multi-branch feature maps are spliced ​​together to obtain spliced ​​feature maps.

[0008] The concatenated feature map is enhanced by a compression-excitation attention mechanism module to obtain the enhanced concatenated feature map, and the predicted signal transmission features are output through the output module.

[0009] The loss value is determined based on the predicted signal transmission characteristics and the corresponding signal transmission characteristics of the sample signal. If the loss value meets the preset requirements, the training of the signal monitoring model is considered complete.

[0010] The present invention also provides a signal monitoring method, comprising:

[0011] Acquire the target signal and generate a spatial distribution feature representation map corresponding to the target signal based on the target signal;

[0012] Input the spatial distribution feature representation map corresponding to the target signal into the target signal monitoring model;

[0013] The target signal monitoring model outputs the signal transmission characteristics corresponding to the target signal; wherein, the target signal monitoring model is a model trained using the training method of the signal monitoring model described above.

[0014] The present invention also provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the training method of the above-described signal monitoring model, or the steps of the signal monitoring method.

[0015] The beneficial effects of this invention lie in the fact that the signal monitoring model comprises an initial feature extraction module, a depthwise separable convolutional layer, a multi-scale parallel structure, a compression-excitation attention mechanism module, and an output module connected in sequence. First, the initial feature extraction module extracts features from the spatial distribution feature representation map of the sample signal and the actual signal transmission features to obtain a preliminary feature map. Then, the depthwise separable convolutional layer expands the channels of the preliminary feature map. Next, the multi-scale parallel structure extracts the expanded multi-channel feature map to obtain a multi-branch feature map, and the multi-branch feature maps are then concatenated to obtain a concatenated feature map. After obtaining the concatenated feature map, the compression-excitation attention mechanism module further enhances the features of the concatenated feature map. The enhanced concatenated feature map is then output by the output module to show the predicted signal transmission features. Finally, when the loss value meets the preset requirements, the training of the signal monitoring model is considered complete. Thus, this method achieves the training of the signal monitoring model; furthermore, during the training process, the depthwise separable convolutional layer, the multi-scale parallel structure, and the compression-excitation attention mechanism module are utilized. The depthwise separable convolutional layer first expands the feature channel space, and the multi-scale parallel structure simultaneously extracts multi-dimensional signal features (such as amplitude, time domain, and frequency domain). The compression-excitation attention mechanism module enhances the effective features, thereby covering the feature requirements of more modulation types and solving the problem of only being able to adapt to a single feature modulation format. Moreover, by deeply aggregating multi-scale features and enhancing channel-dimensional features, it is possible to extract richer deep transmission features and accurately capture and enhance the complex features of high-order modulation signals, thereby improving the adaptability to high-order modulation format signals. As a result, when using this signal monitoring model for monitoring, the accuracy of signal transmission feature monitoring can be improved.

[0016] In addition, the present invention also provides a signal monitoring method and an electronic device, which have the same or corresponding technical features as the training method of the signal monitoring model mentioned above, and have the same effect. Attached Figure Description

[0017] To more clearly illustrate the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a training method for a signal monitoring model provided in an embodiment of the present invention;

[0019] Figure 2 A structural diagram of a coherent optical fiber communication system built based on optical communication system simulation software is provided for an embodiment of the present invention;

[0020] Figure 3 A schematic diagram of a signal monitoring model provided in an embodiment of the present invention;

[0021] Figure 4 This is a schematic diagram of the overall structure of a signal monitoring model provided in an embodiment of the present invention;

[0022] Figure 5 A schematic diagram of a multi-scale parallel structure provided in an embodiment of the present invention;

[0023] Figure 6 A schematic diagram of a compression-excitation attention mechanism module provided in an embodiment of the present invention;

[0024] Figure 7 A flowchart illustrating a method for monitoring the optical signal-to-noise ratio and identifying the modulation format of an elastic optical network, provided in an embodiment of the present invention;

[0025] Figure 8 This is a schematic diagram illustrating the variation of optical signal-to-noise ratio monitoring accuracy with training rounds, provided as an embodiment of the present invention.

[0026] Figure 9 A schematic diagram illustrating the accuracy of actual and predicted modulation format categories in an embodiment of the present invention;

[0027] Figure 10 This is a flowchart of a signal monitoring method provided in an embodiment of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0029] It should be noted that, in the description of this invention, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. The terms "first," "second," etc., used in this invention are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0030] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0031] It should be noted that the training method for the signal monitoring model provided by this invention can be used in multiple signal processing-related fields, such as optical communication, radar, wireless communication, and sonar.

[0032] To enable those skilled in the art to better understand the training method of the signal monitoring model provided by the present invention, the following description is provided in conjunction with specific embodiments and accompanying drawings. Figure 1 A flowchart illustrating a training method for a signal monitoring model provided in an embodiment of the present invention is shown below. Figure 1 As shown, the method includes:

[0033] S10: Input the spatial distribution feature representation map of the sample signal and the actual signal transmission features into the initial feature extraction module to obtain a preliminary feature map; wherein, the sample signal contains multiple types of modulation signals with different modulation orders;

[0034] S11: Expand the channels of the initial feature map using depthwise separable convolutional layers to obtain expanded multi-channel feature maps, and extract the expanded multi-channel feature maps using multi-scale parallel structures to obtain multi-branch feature maps, and then stitch the multi-branch feature maps together to obtain stitched feature maps.

[0035] S12: The concatenated feature map is enhanced using the compression-excitation attention mechanism module to obtain the enhanced concatenated feature map, and the predicted signal transmission features are output through the output module.

[0036] S13: Determine the loss value based on the predicted signal transmission characteristics and the corresponding signal transmission characteristics of the sample signal, and determine the completion of training of the signal monitoring model when the loss value meets the preset requirements.

[0037] Spatial distribution feature representation maps can be constellation diagrams, amplitude histograms, and asynchronous delay tap diagrams, etc. Because constellation diagrams can intuitively and accurately represent the amplitude and phase information of the modulated signal, as well as the signal's noise interference and distortion, they can directly reflect the core characteristics of the modulation format, facilitating rapid identification of the signal's modulation type. Therefore, constellation diagrams can be preferentially chosen as spatial distribution feature representation maps and input into the signal monitoring model.

[0038] In some embodiments, obtaining the spatial distribution feature map of the sample signal and the actual signal transmission features includes:

[0039] Obtain the signal type; the signal type includes at least the signal type corresponding to optical fiber communication and the signal type corresponding to mobile communication;

[0040] Determine the target communication scenario based on the signal type;

[0041] The spatial distribution feature representation map of sample signals and the actual signal transmission features are obtained from the target communication scenario.

[0042] Specifically, when the signal type is the same as that corresponding to optical fiber communication, and the target communication scenario is a communication scenario based on a coherent optical fiber communication system, the spatial distribution characteristic representation diagram is an optical constellation diagram.

[0043] To obtain the spatial distribution characteristics of sample signals and the actual signal transmission characteristics in a communication scenario based on a coherent optical fiber communication system, a coherent optical fiber communication system was first constructed. Figure 2 This is a structural diagram of a coherent optical fiber communication system built based on optical communication system simulation software, provided as an embodiment of the present invention. The coherent optical fiber communication system includes a transmitter, a transmission link, and a receiver. The transmitter includes a first continuous wave laser 1 and an in-phase / quadrature modulator 2; the transmission link is sequentially provided with an optical fiber 3, an optical fiber amplifier 4, and a signal-to-noise ratio setting module 5; the receiver is sequentially provided with a second continuous wave laser 6, a coherent demodulator 7, and a constellation diagram monitor 8.

[0044] In a communication scenario based on a coherent optical fiber communication system, obtaining the spatial distribution feature map of the sample signal and the actual signal transmission features from the target communication scenario includes:

[0045] At the transmitting end, a pseudo-random binary sequence is acquired and converted into multiple types of modulated electrical signals with different modulation orders; the first continuous wave laser is controlled to generate an optical carrier; the multiple types of modulated electrical signals and the optical carrier are modulated by an in-phase / quadrature modulator to obtain a modulated optical signal;

[0046] The modulated optical signal is transmitted through optical fiber, and during the transmission process, it is amplified by an optical fiber amplifier and the signal-to-noise ratio is set by a signal-to-noise ratio setting module to obtain an optical signal with the target signal-to-noise ratio.

[0047] At the receiving end, a coherent demodulator receives the optical signal with the target signal-to-noise ratio and the local oscillator light from the second continuous-wave laser. Using the local oscillator light as a reference light, the signal light with the target signal-to-noise ratio is coherently demodulated to obtain the demodulated electrical signal. A constellation diagram monitor collects the optical constellation diagram data of the demodulated electrical signal to obtain a spatial distribution feature representation map of the sample signal and the actual signal transmission characteristics. Furthermore, setting the signal-to-noise ratio (SNR) through the SNR setting module includes: setting the SNR range based on the signal type; sequentially setting the SNR within the SNR range according to a preset step size in the SNR setting module; and stopping the SNR setting when it is detected that the SNR to be set exceeds the SNR range, in order to obtain spatial distribution feature representation maps of multiple types of modulated signals with different modulation orders under various SNRs.

[0048] There are no limitations on the preset step size and signal-to-noise ratio (SNR) range. For example, if the preset step size is 1 dB, in a communication scenario based on a coherent optical fiber communication system, the SNR (specifically, the optical signal-to-noise ratio, OSNR) range is 15 dB to 24 dB. Since the signal quality is good even with excessively high SNR, there is no need to perform OSNR detection. Therefore, this embodiment stops setting the SNR when it exceeds the range, thus avoiding invalid detection.

[0049] like Figure 2 As shown, at the transmitting end, the pseudo-random binary sequence (PRBS) is converted into Quadrature Phase Shift Keying (QPSK), 8-Phase Shift Keying (8PSK), 16-Quadrature Amplitude Modulation (16QAM), and 32-Quadrature Amplitude Modulation (32QAM) electrical signals. A first continuous wave laser with a center wavelength of 1550 nm, a linewidth of 0.1 MHz, and a power of 10 dBm generates an optical carrier. The electrical signals are modulated with the optical carrier by an in-phase / quadrature modulator (I / Q modulator). The modulated optical signal is transmitted through 80km of standard single-mode fiber (SMF) with an attenuation coefficient of 0.2dB / km and a dispersion coefficient of 16.75ps / nm / km. During transmission, the optical signal is amplified by a fiber amplifier (such as an erbium-doped fiber amplifier, DFA) with a gain of 16dB and a noise figure of 4dB, and the OSNR is adjusted using a signal-to-noise ratio (SNR) setting module. At the receiving end, the signal passes through an optical bandpass filter in a coherent demodulator. A second continuous-wave laser serves as a local oscillator for coherent detection. The photodetector in the coherent demodulator converts the optical signal into an electrical signal, which is then demodulated using coherent demodulation. Finally, a constellation diagram monitor is used to acquire the constellation diagram data of the signal. The entire system adjusts the OSNR of the four signals from 15dB to 24dB in 1dB increments to obtain constellation diagrams. The collected constellation diagram data is used for training and testing of the signal monitoring model, and finally, MFI and OSNR are monitored.

[0050] Compared to methods that use historical spatial distribution feature representation maps and historical signal transmission features as spatial distribution feature representation maps and actual signal transmission features corresponding to sample signals, the method provided in this embodiment uses a coherent optical fiber communication system to simulate and obtain spatial distribution feature representation maps and actual signal transmission features corresponding to multiple types of modulation signals with different modulation orders in a communication scenario based on the coherent optical fiber communication system. This enables the rapid acquisition of spatial distribution feature representation maps and actual signal transmission features of a large number of sample signals of various types.

[0051] The preceding text described how to obtain spatial distribution feature maps and actual signal transmission characteristics of sample signals in a coherent optical fiber communication system scenario. The following section explains how to obtain spatial distribution feature maps and actual signal transmission characteristics of sample signals in a mobile communication scenario.

[0052] In mobile communication scenarios, the spatial distribution characteristic representation diagram is an in-phase / orthogonal constellation diagram. The signal type is the signal type corresponding to mobile communication.

[0053] The spatial distribution feature representation map of the sample signal and the actual signal transmission features obtained from the target communication scenario include:

[0054] A mobile communication simulation scenario is built using a network simulator, and the core parameters of the mobile communication simulation scenario are configured; among them, the core parameters include at least the signal transmission distance adjustment range and the movement speed adjustment range;

[0055] In the established mobile communication simulation scenario, the control signal transmission distance changes at preset intervals within the signal transmission distance adjustment range, and the moving speed changes at preset intervals within the moving speed adjustment range, synchronously generating and transmitting multiple types of modulated electrical signals with different modulation orders.

[0056] During signal transmission, the signal-to-noise ratio is set through the signal-to-noise ratio setting module to obtain an electrical signal with the target signal-to-noise ratio;

[0057] The in-phase / orthogonal constellation diagram data of the electrical signal with the target signal-to-noise ratio are collected by the constellation diagram monitor to obtain the spatial distribution characteristic map of the sample signal and the actual signal transmission characteristics.

[0058] Setting the signal-to-noise ratio via the signal-to-noise ratio setting module includes:

[0059] The signal-to-noise ratio range is set based on the signal type;

[0060] In the signal-to-noise ratio (SNR) setting module, the SNR is set sequentially within the SNR range according to a preset step size; when the SNR to be set is detected to be outside the SNR range, the setting of the SNR is stopped in order to obtain the spatial distribution feature representation map of multiple types of modulation signals with different modulation orders under various SNRs.

[0061] In a mobile communication scenario, the preset step size is 1dB, and the signal-to-noise ratio (SNR) range is 5dB to 30dB. A network simulator is used to build the mobile communication scenario, simulating signal transmission at different distances (e.g., 100m to 500m) and speeds (e.g., 0km / h to 120km / h). Constellation diagrams of four signal types—QPSK, 16QAM, 64QAM, and 256QAM—within the 5dB to 30dB SNR range are collected to construct a dataset.

[0062] In this embodiment, by constructing a mobile communication simulation scenario, the spatial distribution feature representation map and actual signal transmission feature of multiple types of modulation signals with different modulation orders under the mobile communication scenario are simulated, so that a large number of sample signals of various types can be quickly obtained as spatial distribution feature representation maps and actual signal transmission features.

[0063] The above describes how to obtain the spatial distribution feature map and actual signal transmission features of sample signals under different scenarios. After obtaining the spatial distribution feature map and actual signal transmission features of the sample signals, before inputting them into the initial feature extraction module, the following steps are also included:

[0064] The spatial distribution feature map of the sample signal and the actual signal transmission feature are divided into a training set and a test set;

[0065] The spatial distribution feature map of the sample signal and the actual signal transmission features are input into the initial feature extraction module, including:

[0066] The spatial distribution feature representation map of the sample signals obtained from the training set and the actual signal transmission features are input into the initial feature extraction module.

[0067] This section uses a coherent optical fiber communication system as an example to illustrate data acquisition and partitioning, specifically the acquisition of spatial distribution feature maps and actual signal transmission characteristics (sample datasets) of sample signals. As in the above embodiment, the OSNR range of the acquired QPSK, 8PSK, 16QAM, and 32QAM signals is 15dB to 24dB, with a step size of 1dB. Significant differences exist between the constellation diagrams of signals with different modulation formats; for the same modulation format, constellation diagrams with low OSNR values ​​are more blurred and diffuse, while those with high OSNR values ​​are more focused and clearer. For each OSNR value of each modulation format signal, if 100 constellation diagram data points are acquired, a total of 4 × 10 × 100 = 4000 data points are generated, where 4 and 10 represent the 4 modulation formats and 10 OSNR values, respectively. Furthermore, the acquired constellation diagram data is randomly shuffled and divided into training and test sets, with a ratio of 4:1, i.e., 3200 training images and 800 test images.

[0068] The preceding text described how to obtain the spatial distribution feature representation map of the sample signal and the actual signal transmission characteristics. The signal monitoring model is then trained using the spatial distribution feature representation map of the sample signal and the actual signal transmission characteristics. First, a signal monitoring model is constructed, which is a multi-scale channel attention network model.

[0069] Figure 3 This is a schematic diagram of a signal monitoring model provided in an embodiment of the present invention, as shown below. Figure 3 As shown, the signal monitoring model includes an initial feature extraction module 9, a deep convolutional layer 10, a multi-scale parallel structure 11, a compression-excitation attention mechanism module 12, and an output module 13, connected in sequence. Specifically, to improve the accuracy of the signal monitoring results, a deep convolutional layer, a multi-scale parallel structure, and a compression-excitation attention mechanism module are embedded sequentially before the initial feature extraction module and the output module.

[0070] The initial feature extraction module consists of an input layer, a convolutional layer (Conv), a batch normalization (BN) layer, an activation function layer (Rectified Linear Unit, ReLU), and a max pooling layer, which are connected in sequence.

[0071] The output module includes a flatten layer, a fully connected layer (FC), and an output layer.

[0072] In some embodiments, a depthwise convolutional layer (DWconv) sequentially comprises channel-wise convolutional units and pointwise convolutional units; the number of kernels in the pointwise convolutional units is greater than the number of channels in the initial feature map.

[0073] Expanding channels in the initial feature map using depthwise separable convolutional layers includes:

[0074] Spatial feature enhancement extraction is performed by independently convolving each channel of the input preliminary feature map using channel-wise convolution units.

[0075] The feature map extracted after spatial feature enhancement is input into the pointwise convolutional unit;

[0076] The feature map is expanded by performing cross-channel convolution operations on the feature map through pointwise convolution units.

[0077] In the method provided in this embodiment, the feature channel space is expanded through depthwise separable convolutional layers, and the complexity of depthwise separable convolution is lower than that of traditional convolution.

[0078] In some embodiments, the multi-scale parallel structure includes multiple branches and a feature channel splicing layer connected to all branches; the multiple branches include at least a first branch, a second branch, a third branch, and a pooling branch; the first branch includes a first-size convolutional layer; the second branch includes a first-size convolutional layer and a second-size convolutional layer connected in sequence; the third branch includes a first-size convolutional layer and a third-size convolutional layer connected in sequence; the first size is smaller than the second size, and the second size is smaller than the third size. For example, the first size is 1×1, the second size is 3×3, and the third size is 5×5.

[0079] The augmented multi-channel feature map is extracted using a multi-scale parallel structure to obtain a multi-branch feature map. These multi-branch feature maps are then concatenated to obtain a concatenated feature map, which includes:

[0080] For the expanded multi-channel feature map, feature extraction operations are performed in parallel through multiple branches to obtain multiple feature maps extracted by multiple branches; wherein, the multiple feature maps include at least a first type feature map, a second type feature map, a third type feature map, and a fourth type feature map;

[0081] Multiple feature maps are stitched together using a feature channel stitching layer to obtain a stitched feature map.

[0082] In the method provided in this embodiment, a spliced ​​feature map is obtained based on multiple branches and a multi-scale parallel structure of a feature channel splicing layer connected to multiple branches, which realizes the fusion of information from different receptive fields in the same layer; and the large-size convolutional branches are first reduced in dimensionality by small-size convolution, which can reduce the amount of computation.

[0083] In some embodiments, the compressed-incentivized attention mechanism module includes a global average pooling layer, a first fully connected layer, a first activation function layer, a second fully connected layer, a second activation function layer, and a channel weight scaling layer connected in sequence.

[0084] The concatenated feature map is augmented using a compression-excitation attention mechanism module, resulting in the following augmented feature map:

[0085] Global pooling is performed on the stitched feature map through a global average pooling layer to compress its spatial dimensionality;

[0086] Based on the output of global pooling, the dimensions are adjusted sequentially through a fully connected layer, and then processed by an activation function to extract channel correlation information;

[0087] The activation function is used to normalize the channel association information and generate the weights corresponding to each channel.

[0088] The channel weight scaling layer multiplies the weights of each channel with the corresponding channel features of the stitched feature map channel by channel to obtain the stitched feature map after feature enhancement.

[0089] The method provided in this embodiment enhances the features of the stitched feature map based on a compression-excitation attention mechanism module comprising a globally average pooling layer, a first fully connected layer, a first activation function layer, a second fully connected layer, a second activation function layer, and a channel weight scaling layer connected in sequence. By deeply aggregating multi-scale features and enhancing features along the channel dimension, richer deep transmission features can be extracted, and complex features of high-order modulation signals can be accurately captured and enhanced. This improves the adaptability to high-order modulation format signals, thereby increasing the accuracy of signal transmission feature monitoring when using this signal monitoring model.

[0090] To enable those skilled in the art to better understand the structure of the signal monitoring model, the overall structure of the signal monitoring model is explained below. Figure 4 This is a schematic diagram of the overall structure of a signal monitoring model provided in an embodiment of the present invention, as shown below. Figure 4 As shown, the structure sequentially includes an input layer (84×84×3), a convolutional layer (3×3Conv, 32), a normalization and activation function layer, a pooling layer (2×2 Maxpool), a depthwise separable convolutional layer (3×3 DWconv), a multi-scale parallel structure, a compression-activation attention mechanism module (SE Layer), a max pooling layer (2×2 Maxpool), a flattening layer, a fully connected layer (FC), and an output layer (such as output OSNR and MFI).

[0091] Figure 5This is a schematic diagram of a multi-scale parallel structure provided in an embodiment of the present invention, as shown below. Figure 5 As shown, after the depthwise separable convolutional layer, it splits into four branches. From top to bottom, the first branch is a 1×1 convolution, 8 (i.e., 1×1Conv, 8, indicating that a 5×5 kernel is used to perform convolution operations on the input feature map, and the final output is a feature map with 8 channels) and a 3×3 convolution, 16 (i.e., 3×3Conv, 16); the second branch is a 1×1 convolution, 16 (i.e., 1×1Conv, 16); the third branch is a 1×1 convolution, 4 (i.e., 1×1Conv, 4) and a 5×5 convolution, 8 (i.e., 5×5Conv, 8); the fourth branch is a 3×3 maxpooling layer and a 1×1 convolution, 8 (i.e., 1×1Conv, 8). The four branches are then output to the compression-stimulation attention mechanism module through a feature channel concatenation layer (Filter Concatenation).

[0092] Figure 6 This is a schematic diagram of a compression-excitation attention mechanism module provided in an embodiment of the present invention, as shown below. Figure 6 As shown, after the multi-scale parallel structure, the system sequentially connects a global average pooling layer, a first fully connected layer (FC), a first activation function layer (ReLU), a second fully connected layer (FC), a second activation function layer (Sigmoid), and a channel weight scaling layer (Scale), and finally outputs to the output module.

[0093] Specifically, the signal monitoring model first adjusts the constellation map to an 84×84×3 image as input. The constellation map first passes through a traditional convolutional layer with a kernel size of 3×3 and a kernel count of 32. After batch normalization, ReLU activation, and a 2×2 max-pooling layer, it outputs a 42×42×32 image. Next, it enters the depthwise separable convolution module: first, 3×3 channel-wise convolutions are used to independently extract spatial information, and then 1×1 pointwise convolutions are used to fuse cross-channel information, increasing the number of channels from 32 to 48 while maintaining the feature map size of 42×42. Compared to traditional convolutions, depthwise separable convolutions have lower complexity.

[0094] The system then employs a multi-scale parallel structure: four branches (1×1, 3×3, and 5×5 convolutions and 3×3 max pooling) are run in parallel on the 48-channel feature maps. The 3×3 and 5×5 branches are first reduced in dimensionality by 1×1 convolutions to reduce computation. The four branches output 16, 16, 8, and 8 channels respectively, which are then concatenated to form a 48-channel system, achieving the fusion of local, mid-range, and large receptive field information within the same layer.

[0095] The compression-excitation attention mechanism module performs a compression-excitation operation on the 48 channels of the multi-scale parallel structure output, using global average pooling and fully connected layers for dimensionality reduction and expansion. Sigmoid is used to generate channel weights, which are then multiplied channel-by-channel with the original features to output a 42×42×48 feature map. Finally, the 42×42×48 feature map is reduced to 21×21×48 by a 2×2 max-pooling layer, then flattened, and finally passed through two fully connected layers to output signal transmission features (such as OSNR monitoring and MFI results).

[0096] This embodiment implements the construction of a signal monitoring model. The spatial distribution feature representation map of the sample signals in the training set obtained above, along with the actual signal transmission features, are input into the signal monitoring model for iterative training. A loss value is determined based on the predicted signal transmission features and the corresponding signal transmission features of the sample signals. If the loss value meets a preset requirement (e.g., is less than a preset loss value), the training of the signal monitoring model is considered complete.

[0097] For signal monitoring model training, the entire network was trained by the Adam optimizer through backpropagation and mini-batch stochastic gradient descent, with the learning rate set to 0.001.

[0098] To verify the robustness of the model, after obtaining the trained signal monitoring model based on the training set, the following steps are also included:

[0099] The spatial distribution feature representation map of the sample signals obtained from the test set is input into the trained signal monitoring model;

[0100] The trained signal monitoring model outputs a spatial distribution feature representation map of the sample signal, corresponding to the test signal transmission characteristics.

[0101] If the difference between the test signal transmission characteristics and the actual signal transmission characteristics is less than a preset difference, the trained signal monitoring model is determined as the final signal monitoring model.

[0102] Furthermore, when the difference between the test signal transmission characteristics and the actual signal transmission characteristics is greater than or equal to a preset difference, cluster analysis is performed on the samples in the test set whose difference exceeds a preset threshold to determine the signal transmission parameter range (such as signal-to-noise ratio range, modulation type) corresponding to the abnormal difference samples; based on the signal transmission parameter range, training samples of the corresponding range are added to the training set.

[0103] The signal monitoring model is fine-tuned using the supplemented training set until the difference between the model prediction results on the test set and the actual signal transmission characteristics (i.e., labeled data) is less than the preset difference. The trained signal monitoring model is then determined as the final signal monitoring model.

[0104] In the method provided in this embodiment, the trained signal monitoring model is trained using a test set. For cases with large prediction differences, the signal monitoring model is further fine-tuned using training samples, thereby improving the robustness of the signal monitoring model.

[0105] The following section uses an optical communication scenario as an example to illustrate the entire process of monitoring the optical signal-to-noise ratio and identifying the modulation format in a flexible optical network. Figure 7 A flowchart illustrating a method for monitoring the optical signal-to-noise ratio and identifying the modulation format of an elastic optical network, as provided in an embodiment of the present invention, is shown below. Figure 7 As shown, the method includes:

[0106] S14: A coherent optical fiber communication system built based on optical communication system simulation software;

[0107] S15: Obtain constellation diagrams of multiple types of modulation signals with different modulation orders, and divide them into training and test sets;

[0108] S16: Construct a signal monitoring model;

[0109] S17: Iteratively train the signal monitoring model;

[0110] S18: Test the signal monitoring model using the test set to obtain the optical signal-to-noise ratio monitoring and modulation format identification results.

[0111] Specifically, various types of modulation signals with different modulation orders include QPSK, 8PSK, 16QAM and 32QAM as described above.

[0112] The signal monitoring model integrates depthwise separable convolution, multi-scale parallel structure, and a squeeze-excitation channel attention mechanism for adaptive feature weighting from constellation diagrams. A coherent communication system was built using optical communication system simulation software to acquire constellation diagram data for QPSK, 8PSK, 16QAM, and 32QAM signals for training the signal monitoring model. Simulation results show that the proposed signal monitoring model can effectively achieve OSNR monitoring and modulation format identification based on constellation diagram data. Within the OSNR range of 15dB to 24dB, the signal monitoring model achieves OSNR monitoring accuracies of 100%, 100%, 99.5%, and 98.0% for 56Gbaud QPSK, 8PSK, 16QAM, and 32QAM modulated signals, respectively; simultaneously, the model achieves 100% accuracy in identifying the modulation format of all four signals. These results demonstrate the superior performance of the proposed signal monitoring model.

[0113] The trained signal monitoring model is tested using a test set, and the test results are recorded and visualized. Figure 8This is a schematic diagram illustrating the change in optical signal-to-noise ratio (OSNR) monitoring accuracy with training epochs, provided as an embodiment of the present invention. It shows how the accuracy of the signal monitoring model in monitoring the OSNR of four signals—QPSK, 8PSK, 16QAM, and 32QAM—changes with the training epochs. Figure 8 The horizontal axis represents the number of training epochs, and the vertical axis represents the optical signal-to-noise ratio (OSNR) monitoring accuracy for each epoch. Within the OSNR range of 15 to 24 dB, the signal monitoring model achieves OSNR monitoring accuracies of 100%, 100%, 99.5% for 56 Gbaud QPSK, 8PSK, 16QAM, and 32QAM modulated signals, respectively. Figure 8 In the data, the curves corresponding to quadrature phase shift keying (QPSK), octal phase shift keying (OCS), and hexadecimal quadrature amplitude modulation (QAM) overlap after 80 training iterations, achieving a high accuracy of 98.0%. Furthermore, after 60 epochs, the OSNR monitoring accuracy for the four signals reached over 90%, and after 90 epochs, the OSNR monitoring accuracy for the four signals stabilized at over 98%. This demonstrates that the signal monitoring model of this invention exhibits excellent performance in OSNR monitoring and possesses strong generalization ability.

[0114] Figure 9 This is a schematic diagram illustrating the accuracy of the actual and predicted categories of a modulation format according to an embodiment of the present invention. Figure 9 This document details the confusion matrix generated by the signal monitoring model for modulation format identification of four signals using constellation diagram data from the test set, with a training set of 3200 samples and a test set of 800 samples. The confusion matrix (…) Figure 9 The numbers in each element (i.e., cell) of the first 4 rows and 4 columns represent the number of datasets and the recognition rate, respectively. For example, the number of quadrature phase shift keying (QPSK) signals is 200, with a recognition rate of 25.0%; the number of octal phase shift keying (OCK) signals is 200, with a recognition rate of 25.0%; the number of hexadecimal quadrature amplitude modulation (QAM) signals is 200, with a recognition rate of 25.0%; and the number of base-32 QAM signals is 200, with a recognition rate of 25.0%. The signal monitoring model achieves 100% accuracy in recognizing the modulation formats of four signals: QPSK, 8PSK, 16QAM, and 32QAM, demonstrating high precision and indicating that the signal monitoring model of this invention has excellent performance in modulation format recognition.

[0115] In summary, the signal monitoring model proposed in this invention integrates depthwise separable convolution, multi-scale parallel structure, and SE channel attention mechanism, enabling efficient extraction and weighting of constellation diagram features. This achieves high-precision OSNR monitoring and MFI for various types and high-rate high-order modulation format signals. Essentially, the signal monitoring model proposed in this invention uses multi-scale feature fusion and adaptive channel weighting mechanisms to perform high-precision analysis and parameter training on signal representations with spatial distribution characteristics (such as imaged constellation diagrams). The optical signal-to-noise ratio monitoring and modulation format identification method of the signal monitoring model exhibits excellent performance, including wide applicability to signal modulation types, strong adaptability to high-order modulation format signals, and high accuracy in optical signal-to-noise ratio monitoring and modulation format identification for various signals.

[0116] The above describes a training method for a signal monitoring model. This embodiment also provides a signal monitoring method. Figure 10 A flowchart of a signal monitoring method provided in an embodiment of the present invention is shown below. Figure 10 As shown, the method includes:

[0117] S19: Acquire the target signal and generate a spatial distribution feature representation map corresponding to the target signal based on the target signal;

[0118] S20: Input the spatial distribution feature representation map corresponding to the target signal into the target signal monitoring model;

[0119] S21: Output the signal transmission characteristics corresponding to the target signal through the target signal monitoring model.

[0120] The target signal monitoring model is a model trained using the training method for the signal monitoring model described above.

[0121] To facilitate practical application of the target signal monitoring model for signal monitoring, the target communication scenario in this implementation is based on a coherent optical fiber communication system, and the target signal monitoring model is integrated into the coherent receiver. Embedding the target signal monitoring model into the coherent receiver provides technical support for the automated management of optical networks.

[0122] The target communication scenario is a mobile communication scenario, and the target signal monitoring model is integrated into the edge computing node or terminal-side software radio device. It receives I / Q data in real time and outputs monitoring results via a Universal Serial Bus (USB) interface.

[0123] In the method provided in this embodiment, signal monitoring is performed based on the signal monitoring model trained above, which improves the accuracy of signal monitoring.

[0124] Embodiments of the present invention also provide a training apparatus for a signal monitoring model. The training apparatus for the signal monitoring model provided in this embodiment includes:

[0125] The first input module is used to input the spatial distribution feature representation map of the sample signal and the actual signal transmission features into the initial feature extraction module to obtain a preliminary feature map; wherein, the sample signal contains multiple types of modulation signals with different modulation orders;

[0126] The expansion and stitching module is used to expand the channels of the initial feature map using depthwise separable convolutional layers to obtain expanded multi-channel feature maps, and to extract the expanded multi-channel feature maps using a multi-scale parallel structure to obtain multi-branch feature maps, and to stitch the multi-branch feature maps to obtain stitched feature maps.

[0127] The feature enhancement and output module is used to perform feature enhancement processing on the spliced ​​feature map using the compression-excitation attention mechanism module to obtain the feature-enhanced spliced ​​feature map, and output the predicted signal transmission features through the output module.

[0128] The first determining module is used to determine the loss value based on the predicted signal transmission characteristics and the corresponding signal transmission characteristics of the sample signal, and to determine the completion of training of the signal monitoring model when the loss value meets the preset requirements.

[0129] In some embodiments, the training apparatus for the signal monitoring model includes a first acquisition module for acquiring a spatial distribution feature representation map of the sample signal and actual signal transmission features.

[0130] The first acquisition module includes:

[0131] The second acquisition module acquires the signal type; wherein the signal type includes at least the signal type corresponding to optical fiber communication and the signal type corresponding to mobile communication;

[0132] The second determining module is used to determine the target communication scenario based on the signal type;

[0133] The third acquisition module is used to acquire the spatial distribution feature representation map of the sample signal and the actual signal transmission features from the target communication scenario.

[0134] In some embodiments, the third acquisition module includes:

[0135] The modulation module is used at the transmitting end to acquire a pseudo-random binary sequence and convert it into multiple types of modulated electrical signals with different modulation orders; control the first continuous wave laser to generate an optical carrier; and modulate the multiple types of modulated electrical signals and the optical carrier through an in-phase / quadrature modulator to obtain a modulated optical signal.

[0136] The transmission module is used to transmit the modulated optical signal through optical fiber, and during the transmission process, it amplifies the signal through an optical fiber amplifier and sets the signal-to-noise ratio through a signal-to-noise ratio setting module to obtain an optical signal with a target signal-to-noise ratio.

[0137] The demodulation and acquisition module is used at the receiving end to receive the optical signal with the target signal-to-noise ratio and the local oscillator light of the second continuous wave laser through a coherent demodulator. Using the local oscillator light as a reference light, the module performs coherent demodulation on the signal light with the target signal-to-noise ratio to obtain the demodulated electrical signal. The module also acquires the optical constellation diagram data of the demodulated electrical signal through a constellation diagram monitor to obtain the spatial distribution characteristic characterization map of the sample signal and the actual signal transmission characteristics.

[0138] In some embodiments, the third acquisition module includes:

[0139] The setup and configuration module is used to build a mobile communication simulation scenario using a network simulator and configure the core parameters of the mobile communication simulation scenario; among which, the core parameters include at least the signal transmission distance adjustment range and the movement speed adjustment range;

[0140] The generation and transmission module is used to control the signal transmission distance to change at preset intervals within the signal transmission distance adjustment range and the movement speed to change at preset intervals within the movement speed adjustment range in the built mobile communication simulation scenario, and synchronously generate and transmit multiple types of modulated electrical signals with different modulation orders.

[0141] The setting module is used to set the signal-to-noise ratio (SNR) during signal transmission to obtain an electrical signal with the target SNR.

[0142] The acquisition module is used to acquire in-phase / orthogonal constellation diagram data of the electrical signal with the target signal-to-noise ratio through the constellation diagram monitor, so as to obtain the spatial distribution characteristic characterization map of the sample signal and the actual signal transmission characteristics.

[0143] In some embodiments, the training apparatus for the signal monitoring model further includes:

[0144] The partitioning module is used to divide the spatial distribution feature representation map of the sample signal and the actual signal transmission feature into training set and test set.

[0145] The first input module includes:

[0146] The second input module is used to input the spatial distribution feature representation map of the sample signals obtained from the training set and the actual signal transmission features into the initial feature extraction module.

[0147] The training device for the signal monitoring model also includes:

[0148] The third input module is used to input the spatial distribution feature representation map of the sample signals obtained from the test set into the trained signal monitoring model;

[0149] The fourth acquisition module is used to output the test signal transmission characteristics corresponding to the spatial distribution feature representation map of the sample signal through the trained signal monitoring model.

[0150] The third determination module is used to determine the trained signal monitoring model as the final signal monitoring model when the difference between the detected test signal transmission characteristics and the actual signal transmission characteristics is less than a preset difference.

[0151] In some embodiments, the expansion and stitching module specifically includes an expansion module and a stitching module. The expansion module is used to expand the channels of the initial feature map using depthwise separable convolutional layers.

[0152] The expansion module specifically includes:

[0153] The convolution and extraction module is used to independently perform convolution operations on each channel of the input preliminary feature map through channel-by-channel convolution units in order to perform spatial feature enhancement extraction.

[0154] The fourth input module is used to input the feature map extracted after spatial feature enhancement into the pointwise convolutional unit;

[0155] The channel expansion convolution module is used to perform cross-channel convolution operations on the feature map through pointwise convolution units, so as to expand the channels of the initial feature map.

[0156] In some embodiments, the stitching module is used to extract the expanded multi-channel feature map using a multi-scale parallel structure to obtain a multi-branch feature map, and stitch the multi-branch feature maps to obtain a stitched feature map.

[0157] The splicing module specifically includes:

[0158] The feature extraction module is used to perform feature extraction operations in parallel through multiple branches on the expanded multi-channel feature map to obtain multiple feature maps extracted by multiple branches; wherein, the multiple feature maps include at least a first type feature map, a second type feature map, a third type feature map and a fourth type feature map;

[0159] The splicing submodule is used to splice multiple feature maps through the feature channel splicing layer to obtain a spliced ​​feature map.

[0160] In some embodiments, the feature enhancement and output module includes a feature enhancement module. The feature enhancement module is used to perform feature enhancement processing on the stitched feature map using a compression-excitation attention mechanism module to obtain a feature-enhanced stitched feature map.

[0161] The feature enhancement module specifically includes:

[0162] The global pooling operation module is used to perform global pooling operations on the stitched feature map through a global average pooling layer to compress its spatial dimensionality;

[0163] The adjustment and extraction module is used to adjust the dimensions of the output based on global pooling by passing it through a fully connected layer, and then processing it through an activation function to extract channel correlation information.

[0164] The weight generation module is used to normalize the channel association information using an activation function and generate the weights corresponding to each channel.

[0165] The weight processing module is used to multiply the weights of each channel with the corresponding channel features of the stitched feature map channel by channel through the channel weight scaling layer to obtain the stitched feature map after feature enhancement.

[0166] This embodiment also provides a signal monitoring device, including:

[0167] The fifth acquisition module is used to acquire the target signal and generate a spatial distribution feature representation map corresponding to the target signal based on the target signal;

[0168] The fifth input module is used to input the spatial distribution feature representation map corresponding to the target signal into the target signal monitoring model;

[0169] The signal transmission feature acquisition module is used to output the signal transmission features corresponding to the target signal through the target signal monitoring model.

[0170] For a description of the training device for the signal monitoring model and the features in the corresponding embodiments of the signal monitoring device, please refer to the relevant description of the corresponding embodiments of the training method for the signal monitoring model, which will not be repeated here.

[0171] Embodiments of the present invention also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above-described training method embodiments for a signal monitoring model.

[0172] Embodiments of the present invention also provide a computer-readable storage medium storing a computer program configured to execute the steps in the training method embodiments of any of the signal monitoring models described above when running.

[0173] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0174] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in the training method embodiments of any of the signal monitoring models described above.

[0175] Embodiments of the present invention also provide another computer program product, including a non-volatile computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in the training method embodiments of any of the signal monitoring models described above.

[0176] 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 herein 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 implementations should not be considered beyond the scope of this invention.

[0177] The training method for a signal monitoring model, the signal monitoring method, and the electronic device provided by this invention have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only intended to help understand the method and core ideas of this invention. It should be noted that those skilled in the art can make various improvements and modifications to this invention without departing from its principles, and these improvements and modifications also fall within the protection scope of this invention.

Claims

1. A training method for a signal monitoring model, characterized in that, include: The spatial distribution feature map of the sample signal and the actual signal transmission features are input into the initial feature extraction module to obtain a preliminary feature map; wherein, the sample signal contains multiple types of modulation signals with different modulation orders; The initial feature map is expanded by using depthwise separable convolutional layers to obtain an expanded multi-channel feature map. The expanded multi-channel feature map is then extracted using a multi-scale parallel structure to obtain a multi-branch feature map. Finally, the multi-branch feature maps are spliced ​​together to obtain a spliced ​​feature map. The spliced ​​feature map is enhanced using a compression-excitation attention mechanism module to obtain an enhanced spliced ​​feature map, and the predicted signal transmission features are output through an output module. The loss value is determined based on the predicted signal transmission characteristics and the corresponding signal transmission characteristics of the sample signal, and the training of the signal monitoring model is completed when the loss value meets the preset requirements. The depthwise separable convolutional layer sequentially comprises channel-wise convolutional units and pointwise convolutional units; the number of convolutional kernels in the pointwise convolutional units is greater than the number of channels in the preliminary feature map; Expanding channels in the initial feature map using depthwise separable convolutional layers includes: The channel-wise convolution unit independently performs convolution operations on each channel of the input preliminary feature map to perform spatial feature enhancement extraction. The feature map extracted after spatial feature enhancement is input into the pointwise convolutional unit; The pointwise convolutional unit performs cross-channel convolution operations on the feature map to expand the channels of the initial feature map; The multi-scale parallel structure includes multiple branches and a feature channel splicing layer connected to all of the multiple branches; The augmented multi-channel feature map is extracted using a multi-scale parallel structure to obtain a multi-branch feature map. The multi-branch feature maps are then concatenated to obtain a concatenated feature map, which includes: For the expanded multi-channel feature map, feature extraction operations are performed in parallel through the multiple branches to obtain multiple feature maps extracted by the multiple branches; wherein, the multiple feature maps include at least a first type feature map, a second type feature map, a third type feature map, and a fourth type feature map; The multiple feature maps are stitched together using the feature channel stitching layer to obtain the stitched feature map; The compressed-incentivized attention mechanism module includes a global average pooling layer, a first fully connected layer, a first activation function layer, a second fully connected layer, a second activation function layer, and a channel weight scaling layer connected in sequence. The concatenated feature map is enhanced using a compression-excitation attention mechanism module, resulting in an enhanced concatenated feature map including: The global average pooling layer performs a global pooling operation on the stitched feature map to compress its spatial dimension. Based on the output of the global pooling, the dimensions are adjusted sequentially through a fully connected layer, and then processed by an activation function to extract channel correlation information. The channel association information is normalized using an activation function to generate weights for each channel; The channel weight scaling layer multiplies each channel weight with the corresponding channel feature of the stitched feature map channel by channel to obtain the enhanced stitched feature map.

2. The training method for the signal monitoring model according to claim 1, characterized in that, Obtaining the spatial distribution feature map of the sample signal and the actual signal transmission features includes: Obtain the signal type; wherein, the signal type includes at least the signal type corresponding to optical fiber communication and the signal type corresponding to mobile communication; Determine the target communication scenario based on the signal type; The spatial distribution feature map of the sample signal and the actual signal transmission features are obtained from the target communication scenario.

3. The training method for the signal monitoring model according to claim 2, characterized in that, The spatial distribution feature representation diagram is an optical constellation diagram, the signal type is the signal type corresponding to optical fiber communication, and the target communication scenario is a communication scenario based on a coherent optical fiber communication system. The coherent optical fiber communication system includes a transmitter, a transmission link, and a receiver; The transmitting end includes a first continuous wave laser and an in-phase / quadrature modulator; The transmission link is sequentially equipped with an optical fiber, an optical fiber amplifier, and a signal-to-noise ratio setting module. The receiving end is sequentially equipped with a second continuous wave laser, a coherent demodulator, and a constellation monitor.

4. The training method for the signal monitoring model according to claim 3, characterized in that, The spatial distribution feature representation map of the sample signal and the actual signal transmission features obtained from the target communication scenario include: At the transmitting end, a pseudo-random binary sequence is acquired and converted into multiple types of modulated electrical signals with different modulation orders; the first continuous wave laser is controlled to generate an optical carrier; the multiple types of modulated electrical signals and the optical carrier are modulated by the in-phase / quadrature modulator to obtain a modulated optical signal; The modulated optical signal is transmitted through the optical fiber, and during the transmission process, it is amplified by the optical fiber amplifier and the signal-to-noise ratio is set by the signal-to-noise ratio setting module to obtain an optical signal with the target signal-to-noise ratio. At the receiving end, the coherent demodulator receives the optical signal with the target signal-to-noise ratio and the local oscillator light of the second continuous wave laser. Using the local oscillator light as the reference light, the signal light with the target signal-to-noise ratio is coherently demodulated to obtain the demodulated electrical signal. The constellation diagram monitor collects the optical constellation diagram data of the demodulated electrical signal to obtain the spatial distribution characteristic characterization map of the sample signal and the actual signal transmission characteristics.

5. The training method for the signal monitoring model according to claim 2, characterized in that, The spatial distribution feature representation diagram is an in-phase / orthogonal constellation diagram; the signal type is the signal type corresponding to mobile communication; the target communication scenario is a mobile communication scenario. The spatial distribution feature representation map of the sample signal and the actual signal transmission features obtained from the target communication scenario include: A mobile communication simulation scenario is built using a network simulator, and the core parameters of the mobile communication simulation scenario are configured; wherein, the core parameters include at least the signal transmission distance adjustment range and the movement speed adjustment range; In the established mobile communication simulation scenario, the control signal transmission distance changes at preset intervals within the range of the signal transmission distance adjustment, and the moving speed changes at preset intervals within the range of the moving speed adjustment, synchronously generating and transmitting multiple types of modulated electrical signals with different modulation orders; During signal transmission, the signal-to-noise ratio is set through the signal-to-noise ratio setting module to obtain an electrical signal with the target signal-to-noise ratio; The in-phase / orthogonal constellation diagram data of the electrical signal with the target signal-to-noise ratio are collected by the constellation diagram monitor to obtain the spatial distribution characteristic map of the sample signal and the actual signal transmission characteristics.

6. The training method for the signal monitoring model according to claim 4 or 5, characterized in that, After obtaining the spatial distribution feature map and actual signal transmission features of the sample signal, before inputting these features into the initial feature extraction module, the process also includes: The spatial distribution feature map of the sample signal and the actual signal transmission feature are divided into a training set and a test set; The spatial distribution feature map of the sample signal and the actual signal transmission features are input into the initial feature extraction module, including: The spatial distribution feature representation map of the sample signals obtained from the training set and the actual signal transmission features are input into the initial feature extraction module; After confirming the completion of training the signal monitoring model, the following steps are also included: The spatial distribution feature representation map of the sample signals obtained from the test set is input into the trained signal monitoring model; The trained signal monitoring model outputs a spatial distribution feature representation map of the sample signal, corresponding to the test signal transmission characteristics. If the difference between the test signal transmission characteristics and the actual signal transmission characteristics is less than a preset difference, the trained signal monitoring model is determined as the final signal monitoring model.

7. The training method for the signal monitoring model according to claim 1, characterized in that, The plurality of branches includes at least a first branch, a second branch, a third branch, and a pooling branch; the first branch includes a first-size convolutional layer; the second branch includes a first-size convolutional layer and a second-size convolutional layer connected in sequence; the third branch includes a first-size convolutional layer and a third-size convolutional layer connected in sequence; the first size is smaller than the second size, and the second size is smaller than the third size.

8. A signal monitoring method, characterized in that, include: Acquire the target signal and generate a spatial distribution feature representation map corresponding to the target signal based on the target signal; Input the spatial distribution feature representation map corresponding to the target signal into the target signal monitoring model; The target signal monitoring model outputs the signal transmission characteristics corresponding to the target signal; wherein, the target signal monitoring model is a model trained using the training method of the signal monitoring model as described in any one of claims 1 to 7; the target communication scenario is a communication scenario based on a coherent optical fiber communication system, and the target signal monitoring model is integrated in a coherent receiver; the target communication scenario is a mobile communication scenario, and the target signal monitoring model is integrated in an edge computing node or a terminal-side software radio device. The depthwise separable convolutional layer sequentially comprises channel-wise convolutional units and pointwise convolutional units; the number of convolutional kernels in the pointwise convolutional units is greater than the number of channels in the preliminary feature map; Expanding channels in the initial feature map using depthwise separable convolutional layers includes: The channel-wise convolution unit independently performs convolution operations on each channel of the input preliminary feature map to perform spatial feature enhancement extraction. The feature map extracted after spatial feature enhancement is input into the pointwise convolutional unit; The pointwise convolutional unit performs cross-channel convolution operations on the feature map to expand the channels of the initial feature map; The multi-scale parallel structure includes multiple branches and a feature channel splicing layer connected to all of the multiple branches; The augmented multi-channel feature map is extracted using a multi-scale parallel structure to obtain a multi-branch feature map. The multi-branch feature maps are then concatenated to obtain a concatenated feature map, which includes: For the expanded multi-channel feature map, feature extraction operations are performed in parallel through the multiple branches to obtain multiple feature maps extracted by the multiple branches; wherein, the multiple feature maps include at least a first type feature map, a second type feature map, a third type feature map, and a fourth type feature map; The multiple feature maps are stitched together using the feature channel stitching layer to obtain the stitched feature map; The compressed-incentivized attention mechanism module includes a global average pooling layer, a first fully connected layer, a first activation function layer, a second fully connected layer, a second activation function layer, and a channel weight scaling layer connected in sequence. The concatenated feature map is enhanced using a compression-excitation attention mechanism module, resulting in an enhanced concatenated feature map including: The global average pooling layer performs a global pooling operation on the stitched feature map to compress its spatial dimension. Based on the output of the global pooling, the dimensions are adjusted sequentially through a fully connected layer, and then processed by an activation function to extract channel correlation information. The channel association information is normalized using an activation function to generate weights for each channel; The channel weight scaling layer multiplies each channel weight with the corresponding channel feature of the stitched feature map channel by channel to obtain the enhanced stitched feature map.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the training method of the signal monitoring model as described in any one of claims 1 to 7, or the signal monitoring method as described in claim 8, when executing the computer program.

Images