Signal identification method and device, computer equipment, storage medium and program product

Abstract

The invention relates to a signal identification method and device, computer equipment, a storage medium and a program product. The method comprises the following steps: acquiring a time-frequency feature image corresponding to a to-be-identified signal; extracting signal key features in the time-frequency feature image; extracting spatial topological features in the signal key features; and predicting the signal attribute of the to-be-identified signal according to the spatial topological characteristics. According to the method, the key features and the spatial position information of the to-be-identified signal can be captured at the same time by combining the extraction process of the key features of the signal and the extraction process of the spatial topological features, so that the to-be-identified signal can be accurately identified, and the signal identification precision is improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a signal recognition method, apparatus, computer equipment, storage medium, and program product. Background Technology

[0002] With the development of communication technology, a massive number of transmission sources coexist, including cellular base stations, wireless LAN access points, low-power wide-area network devices, radar systems, and various non-cooperative drone links, creating an extremely complex heterogeneous electromagnetic environment. Issues such as illegal covert communication or co-channel interference from sensors pose serious challenges to signal spectrum monitoring and signal sensing processes.

[0003] Traditional signal processing methods, such as energy detection, matched filtering, or cyclic stationary feature extraction, are typically used to identify signals. However, these methods often rely on prior signal knowledge in low signal-to-noise ratio or non-line-of-sight transmission environments, resulting in low signal recognition accuracy. Summary of the Invention

[0004] Therefore, it is necessary to provide a signal recognition method, device, computer equipment, storage medium, and program product that can improve the accuracy of signal recognition in response to the above-mentioned technical problems.

[0005] In a first aspect, this application provides a signal recognition method, including:

[0006] Obtain the time-frequency feature image corresponding to the signal to be identified;

[0007] Extract key signal features from time-frequency feature images;

[0008] Extracting spatial topological features from key features of the signal;

[0009] Based on spatial topological features, predict the signal attributes of the signal to be identified.

[0010] In one embodiment, extracting key signal features from a time-frequency feature image includes: performing dimensionality upscaling on the time-frequency feature image to obtain a first feature image; determining the channel attention weights corresponding to each channel based on the number of channels in the first feature image; and performing channel attention enhancement on the first feature image based on the channel attention weights corresponding to each channel to obtain key signal features.

[0011] In one embodiment, determining the channel attention weights corresponding to each channel based on the number of channels in the first feature image includes: extracting global features from the first feature image to obtain channel descriptors of the signal to be identified; for each channel of the first feature image, determining the kernel size of the target convolution used by the channel based on the number of channels; and extracting features from the channel descriptors based on the target convolution corresponding to each channel to determine the channel attention weights corresponding to each channel.

[0012] In one embodiment, the time-frequency feature image is subjected to dimensionality upscaling to obtain a first feature image, including: resizing the time-frequency feature image; normalizing the resized time-frequency feature image to obtain a single-channel grayscale image; and upscaling the single-channel grayscale image to obtain the first feature image.

[0013] In one embodiment, extracting spatial topological features from key signal features includes: extracting spatial distribution features of key signal features under different receptive fields; determining spatial attention weights corresponding to key signal features based on spatial distribution features under different receptive fields; and performing spatial attention enhancement on key signal features based on spatial attention weights to obtain spatial topological features.

[0014] In one embodiment, predicting the signal attributes of the signal to be identified based on spatial topological features includes: extracting common features of the spatial topological features from multiple prediction tasks to obtain common features of the tasks; and predicting the signal attributes of the signal to be identified under multiple prediction tasks based on the common features of the tasks. The signal attributes under multiple prediction tasks include at least two of the following: signal category under classification prediction task, center frequency offset under center frequency offset prediction task, and signal bandwidth under signal bandwidth prediction task.

[0015] Secondly, this application also provides a signal identification device, comprising:

[0016] The acquisition module is used to acquire the time-frequency feature image corresponding to the signal to be identified;

[0017] The first extraction module is used to extract key signal features from the time-frequency feature image;

[0018] The second extraction module is used to extract spatial topological features from key features of the signal;

[0019] The prediction module is used to predict the signal attributes of the signal to be identified based on spatial topological features.

[0020] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0021] Obtain the time-frequency feature image corresponding to the signal to be identified;

[0022] Extract key signal features from time-frequency feature images;

[0023] Extracting spatial topological features from key features of the signal;

[0024] Based on spatial topological features, predict the signal attributes of the signal to be identified.

[0025] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0026] Obtain the time-frequency feature image corresponding to the signal to be identified;

[0027] Extract key signal features from time-frequency feature images;

[0028] Extracting spatial topological features from key features of the signal;

[0029] Based on spatial topological features, predict the signal attributes of the signal to be identified.

[0030] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0031] Obtain the time-frequency feature image corresponding to the signal to be identified;

[0032] Extract key signal features from time-frequency feature images;

[0033] Extracting spatial topological features from key features of the signal;

[0034] Based on spatial topological features, predict the signal attributes of the signal to be identified.

[0035] The aforementioned signal recognition method, apparatus, computer equipment, storage medium, and program product can acquire a time-frequency feature image corresponding to the signal to be recognized; extract key signal features from the time-frequency feature image; extract spatial topological features from the key signal features; and predict the signal attributes of the signal to be recognized based on the spatial topological features. In the above process, by first extracting the key signal features from the time-frequency feature image, noise can be suppressed to a certain extent, and key features containing signal energy can be screened out. Then, by extracting the spatial topological features from the key signal features, the perception of the signal's location can be achieved. By combining the extraction process of key signal features and the extraction process of spatial topological features, the key features and spatial location information of the signal to be recognized can be captured simultaneously, thereby enabling accurate recognition of the signal and improving signal recognition accuracy. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a diagram illustrating the application environment of the signal recognition method in one embodiment;

[0038] Figure 2 This is a flowchart illustrating a signal recognition method in one embodiment;

[0039] Figure 3 This is a flowchart illustrating the key feature extraction steps of a signal in one embodiment;

[0040] Figure 4 This is a flowchart illustrating the spatial topology feature extraction steps in one embodiment;

[0041] Figure 5 This is a flowchart illustrating the signal attribute prediction step in one embodiment;

[0042] Figure 6 This is a flowchart illustrating the signal recognition method in another embodiment;

[0043] Figure 7 Here is a block diagram of the signal recognition model in one embodiment;

[0044] Figure 8 This is a structural block diagram of a signal recognition device in one embodiment;

[0045] Figure 9 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0047] The signal recognition method provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or located on the cloud or other network servers. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, projection devices, etc. Portable wearable devices can include smartwatches, smart bracelets, head-mounted devices, etc. Head-mounted devices can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. Server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0048] In one exemplary embodiment, such as Figure 2 As shown, a signal recognition method is provided, which can be applied to... Figure 1 Taking the server in the example of [the example], the following is included:

[0049] S210, acquire the time-frequency feature image corresponding to the signal to be identified.

[0050] The signals to be identified may include wireless signals (such as Wireless Fidelity (WiFi) signals, Bluetooth signals, radar pulse signals, fifth-generation mobile communication technology new radio (5G NR) signals, and long-range radio (LoRa) signals), underwater acoustic signals (such as signals emitted by underwater targets such as submarines and marine life), vibration signals from industrial equipment (such as vibration signals from equipment such as motors, bearings, or gearboxes), and biological signals (such as physiological signals such as electrocardiograms, electroencephalograms, or electromyograms).

[0051] Among them, the time-frequency feature image can be understood as a two-dimensional time-frequency diagram obtained by converting the signal to be identified.

[0052] The following section uses a wireless signal as an example to describe the signal identification method in detail.

[0053] In some embodiments, the original complex baseband signal, i.e., the signal to be identified, can be acquired using a broadband radio frequency receiver. The original complex baseband signal can be understood as a one-dimensional discrete signal whose frequency varies with time. The original complex baseband signal is denoted as... The complex form can be represented as: ;in, Indicates in-phase components; The quadrature components are represented by ; n represents the sampling point index according to time; j represents the imaginary unit. Then, the one-dimensional original complex baseband signal is converted into a two-dimensional time-frequency feature image through short-time Fourier transform.

[0054] In some embodiments, to reduce spectral leakage, a Hamming window or a Lechman window can be used to window and truncate the signal to be identified. Specifically, a window length of [insert window length here] can be used. The sliding window is used to truncate the signal to be identified.

[0055] In some embodiments, the time-frequency feature image obtained based on the short-time Fourier transform can be represented as:

[0056]

[0057] Where n represents the sampling point index according to time; m represents the time frame index; and k represents the frequency index. H represents the window function; H represents the step length. The value represents the length of the sliding window; j represents the imaginary unit. The elements in the time-frequency matrix S, composed of pixels in the generated time-frequency feature image, are... , where represents the energy density at time m and frequency k.

[0058] S220 extracts key signal features from the time-frequency feature image.

[0059] Among them, key signal features can be understood as channel features that represent signal energy.

[0060] In some embodiments, background white noise in the time-frequency feature image can be suppressed and channel features containing signal energy can be screened out.

[0061] S230 extracts spatial topological features from key signal features.

[0062] Spatial topological features can be understood as characteristics that represent the distribution of signals.

[0063] In some embodiments, the feature distribution within the receptive field can be obtained by convolution processing the key features of the signal, thereby obtaining spatial topological features.

[0064] S240, predicts the signal attributes of the signal to be identified based on spatial topological features.

[0065] The signal attributes may include at least one of the following: signal type, center frequency offset, and signal bandwidth.

[0066] In some embodiments, when the signal attributes include at least two of signal type, center frequency offset, and signal bandwidth, the signal attributes under each prediction task head can be obtained using multiple prediction task heads.

[0067] In the above signal recognition method, a time-frequency feature image corresponding to the signal to be recognized can be obtained; key signal features in the time-frequency feature image can be extracted; spatial topological features in the key signal features can be extracted; and the signal attributes of the signal to be recognized can be predicted based on the spatial topological features. In the above process, by first extracting the key signal features in the time-frequency feature image, noise can be suppressed to a certain extent, and key features containing signal energy can be screened out. Then, by extracting the spatial topological features in the key signal features, the perception of the signal's location can be realized. By combining the extraction process of key signal features and the extraction process of spatial topological features, the key features and spatial location information of the signal to be recognized can be captured simultaneously, thereby enabling accurate recognition of the signal and improving the signal recognition accuracy.

[0068] Based on the technical solutions of the above embodiments, this application also provides another optional embodiment, in which the signal key feature extraction step of S220 is refined.

[0069] See Figure 3 The signal key feature extraction steps shown include:

[0070] S310, perform dimensionality upscaling on the time-frequency feature image to obtain the first feature image.

[0071] In some embodiments, the time-frequency feature image can be up-dimension processed to obtain a multi-channel first feature image.

[0072] In some embodiments, to adapt to the input requirements of subsequent processing of time-frequency feature images and reduce computational load, the time-frequency feature images can be resized. Specifically, the generated time-frequency feature images can be resized to a preset size, such as H×W, using bilinear interpolation; where H represents the height of the image and W represents the width of the image. Subsequently, the resized time-frequency feature images can be normalized to map the pixel values ​​of each pixel in the time-frequency feature images to the [0,1] interval, resulting in a single-channel grayscale image.

[0073] In some embodiments, the normalization formula can be:

[0074]

[0075] in, Represents a single-channel grayscale image tensor; S represents the pixel value of the current pixel. Indicates the minimum pixel value; This represents the maximum pixel value.

[0076] Then, the single-channel grayscale image can be upscaled to obtain the first feature image.

[0077] In some embodiments, a preprocessing network can be used to upscale a single-channel grayscale image. This preprocessing network may include convolutional layers, normalization layers, and activation function layers. The single-channel grayscale image is passed through these layers to obtain a multi-channel first feature image.

[0078] In some embodiments, the operation process is as follows:

[0079]

[0080] in, Represents the first feature image; Represents a single-channel grayscale image tensor; (·) indicates convolution processing using a standard 3×3 convolutional layer; BN(·) indicates batch normalization processing. (·) indicates that non-linear processing is performed using the Swish activation function.

[0081] S320, determine the channel attention weights corresponding to each channel based on the number of channels in the first feature image.

[0082] In some embodiments, global features of the first feature image can be extracted to obtain channel descriptors of the signal to be identified. For each channel of the first feature image, the kernel size of the target convolution used by the channel is determined according to the number of channels; based on the target convolution corresponding to each channel, features are extracted from the channel descriptor to determine the channel attention weights corresponding to each channel.

[0083] The channel descriptor can be understood as an aggregation of spatial information of the first feature image.

[0084] In some embodiments, global average pooling can be performed on the first feature image to extract global features and obtain the channel descriptor of the signal to be identified.

[0085] In some embodiments, a channel descriptor can be represented as:

[0086]

[0087] in, Indicates the channel descriptor; H represents the height of the first feature image; W represents the width of the first feature image; C represents the number of channels; i represents the height value of the current pixel in the first feature image; j represents the width value of the current pixel in the first feature image.

[0088] In some embodiments, the kernel size of the target convolution used by a channel can be determined based on the number of channels and the correspondence between the number of channels and the kernel size.

[0089] In some embodiments, the kernel size is positively correlated with the number of channels, and the relationship between the number of channels and the kernel size can be expressed as:

[0090]

[0091] Where k represents the kernel size; C represents the number of channels; b and b represent hyperparameters, respectively. This indicates taking the nearest odd number.

[0092] In some embodiments, for each channel, features can be extracted from the channel descriptor based on the target convolution corresponding to that channel to determine the channel attention weights. ;in, This indicates that convolution processing is performed using the target convolution.

[0093] In some embodiments, the attention weights can be non-linearly processed using the Sigmoid activation function to obtain attention weights for reweighting each channel of the first feature map. ;in, This indicates that the Sigmoid activation function is used for non-linear processing.

[0094] S330: Based on the channel attention weights corresponding to each channel, channel attention enhancement is performed on the first feature image to obtain key signal features.

[0095] In some embodiments, the channel attention weights corresponding to each channel can be used to reweight each channel of the first feature image to obtain key signal features, thereby suppressing background white noise and filtering out key channel features containing signal energy.

[0096] In this embodiment, the channel attention weights corresponding to each channel are determined based on the number of channels in the first feature image. Different channel attention weights can be assigned to feature maps of different channels, thereby highlighting key features containing signal energy and effectively suppressing background white noise.

[0097] Based on the technical solutions of the above embodiments, this application also provides another optional embodiment, in which the spatial topological feature extraction step of S230 is refined.

[0098] See Figure 4 The spatial topological feature extraction steps shown include:

[0099] S410 extracts the spatial distribution characteristics of key signal features under different receptive fields.

[0100] In some embodiments, the spatial distribution features of key signal features under different receptive fields can be extracted using a Receptive-Field Attention Convolution (RFAConv) network. The RFAConv network can simulate a large receptive field through multiple convolutions of different sizes or dilated convolutions to extract the spatial distribution features of key signal features under different receptive fields.

[0101] S420 determines the spatial attention weights corresponding to key signal features based on the spatial distribution characteristics under different receptive fields.

[0102] In some embodiments, spatial distribution features under different receptive fields can be fused to obtain spatial attention weights corresponding to key signal features.

[0103] S430, based on spatial attention weights, performs spatial attention enhancement on key signal features to obtain spatial topological features.

[0104] In some embodiments, spatial attention weights can be used to reweight key signal features to obtain spatial topological features, which can then focus on specific regions in the time-frequency feature image where the signal to be identified exists, such as the frequency band edges of the signal to be identified, thereby enhancing the boundary perception capability that is crucial for bandwidth and center frequency estimation.

[0105] In some embodiments, spatial topological features can be obtained according to the following formula:

[0106]

[0107] in, Represents spatial topological features; This represents the first feature image used for channel attention enhancement, i.e., the key features of the signal; This represents the spatial distribution features obtained after convolution processing; This means converting spatial distribution characteristics into a probability distribution.

[0108] In this embodiment, the spatial distribution features under different receptive fields, i.e. spatial features at different scales, can be used to assign more precise attention weights to each spatial location, thereby highlighting key locations in the image and laying the foundation for more accurate identification of the signal to be identified in the future.

[0109] Based on the technical solutions of the above embodiments, this application also provides another optional embodiment, in which the signal attribute prediction step of S240 is refined.

[0110] See Figure 5 The signal attribute prediction steps shown include:

[0111] S510 extracts common features from spatial topological features across multiple prediction tasks to obtain common features for each task.

[0112] Among them, the spatial topological features obtained through channel attention enhancement and spatial attention enhancement contain rich semantic information.

[0113] In some embodiments, convolutional layers can be used to extract common features from spatial topological features, and then global average pooling and flattening can be used to obtain one-dimensional common features of the task.

[0114] S520 predicts the signal attributes of the signal to be identified under multiple prediction tasks based on the common characteristics of the tasks.

[0115] In some embodiments, a multi-prediction task head can be set based on a hard parameter sharing mechanism, and common features of the tasks can be input into the multi-prediction task head to predict the signal attributes of the signal to be identified under the multi-prediction task.

[0116] Among them, the signal attributes under the multi-prediction task include at least two of the following: signal category under the classification prediction task, center frequency offset under the center frequency offset prediction task, and signal bandwidth under the signal bandwidth prediction task.

[0117] In some embodiments, the multi-prediction task head may include a classification head, a frequency offset regression head, and a bandwidth regression head.

[0118] In some embodiments, the classification head is used to predict the signal class in a classification prediction task. The classification head may consist of a fully connected layer followed by a Softmax function that outputs the predicted probability for each signal class.

[0119]

[0120] in, This represents the common features of the tasks obtained after global average pooling; This represents the weight matrix of the fully connected layer in a classification prediction task; This represents the bias term vector for the classification prediction task; This is the normalized exponential function, used to map the output to a probability distribution.

[0121] In some embodiments, the frequency offset regression head is used to predict the center frequency offset in a center frequency offset prediction task. The frequency offset regression head consists of independent fully connected layers and directly outputs estimated values ​​of physical parameters.

[0122] The predicted value of the center frequency offset can be obtained using the following formula:

[0123]

[0124] in, This represents the predicted value of the center frequency offset; This represents the common features of the tasks obtained after global average pooling; The weight matrix representing the frequency offset regression head; This represents the bias term of the frequency offset regression head.

[0125] In some embodiments, the bandwidth regression head is used to predict the signal bandwidth in a signal bandwidth prediction task. The bandwidth regression head consists of independent fully connected layers and directly outputs estimated values ​​of physical parameters.

[0126] The predicted value of the signal bandwidth can be obtained using the following formula:

[0127]

[0128] in, This represents the predicted value of the signal bandwidth. This represents the common features of the tasks obtained after global average pooling; The weight matrix representing the bandwidth regression head; This indicates the bias term of the bandwidth regression header.

[0129] In the above process, high-dimensional feature vectors can be transformed through linear mapping. This is mapped to specific physical scalar values. The design of multiple prediction task heads allows multiple prediction tasks to share most of the network parameters, greatly saving computational resources. The hard parameter sharing mechanism enables the gradients of different prediction tasks to mutually assist in learning features, thereby achieving feature enhancement.

[0130] Based on the technical solutions of the above embodiments, this application also provides another optional embodiment, in which the signal recognition method is described in detail.

[0131] See Figure 6The signal recognition method shown includes:

[0132] S601, acquire the time-frequency feature image corresponding to the signal to be identified.

[0133] S602 adjusts the size of the time-frequency feature image.

[0134] S603 performs normalization processing on the resized time-frequency feature image to obtain a single-channel grayscale image.

[0135] S604 performs dimensionality upscaling on a single-channel grayscale image to obtain the first feature image.

[0136] S605, extract the global features of the first feature image to obtain the channel descriptor of the signal to be identified.

[0137] S606, for each channel of the first feature image, determine the kernel size of the target convolution used by the channel based on the number of channels.

[0138] S607: Based on the target convolution corresponding to each channel, feature extraction is performed on the channel descriptor to determine the channel attention weight corresponding to each channel.

[0139] S608, based on the channel attention weights corresponding to each channel, performs channel attention enhancement on the first feature image to obtain key signal features.

[0140] S609 extracts the spatial distribution characteristics of key signal features under different receptive fields.

[0141] S610 determines the spatial attention weights corresponding to key signal features based on the spatial distribution characteristics under different receptive fields.

[0142] S611, based on spatial attention weights, performs spatial attention enhancement on key signal features to obtain spatial topological features.

[0143] S612 extracts common features from spatial topological features across multiple prediction tasks to obtain common features for each task.

[0144] S613, based on the common characteristics of the tasks, predict the signal attributes of the signal to be identified under multiple prediction tasks.

[0145] Among them, the signal attributes under the multi-prediction task include at least two of the following: signal category under the classification prediction task, center frequency offset under the center frequency offset prediction task, and signal bandwidth under the signal bandwidth prediction task.

[0146] In one exemplary embodiment, such as Figure 7As shown, a signal recognition model is provided. After obtaining the time-frequency feature image corresponding to the signal to be recognized, the signal recognition model can be used to perform steps S220~S240 on the time-frequency feature image to achieve signal recognition.

[0147] See Figure 7 The signal recognition model shown may include a preprocessing Stem network, a shallow feature extraction network, a deep feature extraction network, and a prediction network.

[0148] The preprocessing Stem network is a ConvBNAct (k=3, s=2) network, which can include a standard convolutional layer with a kernel size of 3×3, a stride of 2, and 32 output channels, a batch normalization (BN) layer, and a Swish activation function layer. Here, k represents the kernel size and s represents the stride.

[0149] Among them, the shallow feature extraction network embeds parameters. This is a Mobile Inverted Bottleneck Convolution (MBConv) network for an Efficient Channel Attention (ECA) network with a value of 2 and b=1. Specifically, it can include network blocks 1 through 3. Block 1 has a spread ratio of 1, a kernel size of 2×2, and a stride of 1. Block 1 can be represented as MBConv1. — ECA (k=2, s=1). The expansion ratio of block2 is set to 6, the kernel size is 2×2, and the stride is 2. Block2 can be represented as MBConv6. — ECA (k=2, s=2). The expansion ratio of block3 is set to 6, the kernel size is 3×3, and the stride is 2; block3 can be represented as MBConv6. — ECA (k=3, s=2).

[0150] The deep feature extraction network is a Receptive-Field Attention Convolution (RFAConv) network that combines an ECA network. — The ECA network, specifically block 4, has a 3×3 kernel size and a stride of 2. Block 4 can be represented as RFAConv. — ECA (k=3, s=2).

[0151] The deep feature extraction network can sequentially include a ConvBNAct(1) subnetwork, an RFAConv(3) subnetwork, an ECA subnetwork, a one-dimensional convolutional Conv(1×1) subnetwork, a BN layer, a regularized Dropout layer, and a residual connection layer. Among them, the RFAConv(3) subnetwork can include multiple convolutional layers of different sizes.

[0152] The aforementioned effective channel attention (ECA) network in the shallow feature extraction network, combined with the receptive field attention (RFA) network in the deep feature extraction network, can be fully adapted to the physical characteristics of the time-frequency map of wireless signals. Specifically, the ECA network in the shallow feature extraction network can suppress background noise, significantly improving signal detection sensitivity even in low signal-to-noise ratio environments such as -10dB. The RFA network in the deep feature extraction network can effectively address the limitation of handling broadband and frequency-hopping signals due to restricted receptive fields.

[0153] The prediction network can include ConvBNAct top (k=1, s=1) sub-network, Global Average Pooling & Flatten sub-network, prediction task head (including classification head and regression head).

[0154] The signal recognition model obtained through pruning and structural reorganization techniques achieves significant compression of model parameters, resulting in approximately 0.12 million parameters and a floating-point operation capacity of approximately 0.03 GFLOPs (30 million floating-point operations per second; G represents billions), with the model weight file size around 0.5MB. This lightweight design eliminates reliance on computational resources, enabling deployment on low-cost, low-power IoT terminals, drone-borne chips, or handheld spectrum analyzers, significantly lowering the hardware barrier for large-scale spectrum monitoring networks and providing technical support for the deployment of high-density monitoring nodes in environments with signal recognition requirements.

[0155] In the process of performing signal recognition on time-frequency feature images using a signal recognition model, the channel grayscale image can be input into the ConvBNAct (k=3, s=2) network to obtain the first feature image through dimensionality upscaling. The first feature image is then input into blocks 1, 2, and 3 sequentially to obtain key signal features. These key signal features are then input into block 4 to obtain spatial topological features. Finally, the spatial topological features are input into ConvBNAct. topA (k=1, s=1) sub-network is used to obtain common task features. These common task features are then processed by a Global Average Pooling & Flatten sub-network to obtain one-dimensional common task features, which are then simultaneously input into multiple prediction task heads to obtain the signal attributes of the signal to be identified under multiple prediction tasks.

[0156] In the above embodiments, by utilizing a hard parameter sharing mechanism, signal classification, center frequency estimation, and bandwidth estimation can be completed simultaneously during a single signal identification process. Compared to the scheme of executing multiple independent single-task models serially, this approach effectively reduces device computational power consumption while improving system response speed.

[0157] In some embodiments, during the training of a signal recognition model, in order to balance classification accuracy and parameter estimation accuracy, a composite loss function can be constructed according to the following formula:

[0158]

[0159] in, Represents the composite loss function; Represents the signal category loss function; This represents the regression loss function for the center frequency offset; This represents the signal bandwidth regression loss function; The weights represent the signal category loss function. This represents the weights corresponding to the center frequency offset regression loss function; This represents the weights corresponding to the signal bandwidth regression loss function.

[0160] In some embodiments, the signal category loss function can be specifically expressed as:

[0161]

[0162] Where C represents the number of categories; y represents the indicator variable of the true label corresponding to the i-th signal category; if the signal to be identified belongs to the i-th category, then If the signal to be identified does not belong to the i-th class, then... =0; This represents the predicted probability of the signal category.

[0163] In some embodiments, Huber Loss is used to represent the center frequency offset regression loss function and the signal bandwidth regression loss function, which can be specifically expressed by the following formula:

[0164]

[0165] in, Indicates the predicted value; Represents the actual value; This represents the threshold parameter.

[0166] In the above embodiments, the introduction of the Huber loss function enhances the signal recognition model's ability to resist burst noise and outliers in the signal to be identified. This loss function mechanism enables the signal recognition model to avoid overfitting to abnormal samples during training, and instead focus on the statistical regularity of the signal. As a result, it can maintain stable prediction output when facing unknown channel fading, multipath effects, or co-channel interference, effectively reducing the false alarm rate and false negative rate.

[0167] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0168] Based on the same inventive concept, this application also provides a signal recognition device for implementing the signal recognition method described above. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations in one or more signal recognition device embodiments provided below can be found in the limitations of the signal recognition method described above, and will not be repeated here.

[0169] In one exemplary embodiment, such as Figure 8 As shown, a signal recognition device is provided, comprising: an acquisition module 810, a first extraction module 820, a second extraction module 830, and a prediction module 840, wherein:

[0170] The acquisition module 810 is used to acquire the time-frequency feature image corresponding to the signal to be identified;

[0171] The first extraction module 820 is used to extract key signal features from the time-frequency feature image;

[0172] The second extraction module 830 is used to extract spatial topological features from key features of the signal;

[0173] The prediction module 840 is used to predict the signal attributes of the signal to be identified based on spatial topological features.

[0174] In one embodiment, extracting key signal features from a time-frequency feature image includes: performing dimensionality upscaling on the time-frequency feature image to obtain a first feature image; determining the channel attention weights corresponding to each channel based on the number of channels in the first feature image; and performing channel attention enhancement on the first feature image based on the channel attention weights corresponding to each channel to obtain key signal features.

[0175] In one embodiment, the first extraction module 820 is specifically used to: extract global features of the first feature image to obtain channel descriptors of the signal to be identified; for each channel of the first feature image, determine the kernel size of the target convolution used by the channel according to the number of channels; and extract features from the channel descriptors according to the target convolution corresponding to each channel to determine the channel attention weights corresponding to each channel.

[0176] In one embodiment, the first extraction module 820 is specifically used for: resizing the time-frequency feature image; normalizing the resized time-frequency feature image to obtain a single-channel grayscale image; and upscaling the single-channel grayscale image to obtain a first feature image.

[0177] In one embodiment, the second extraction module 830 is specifically used to: extract the spatial distribution features of key signal features under different receptive fields; determine the spatial attention weights corresponding to the key signal features based on the spatial distribution features under different receptive fields; and perform spatial attention enhancement on the key signal features based on the spatial attention weights to obtain spatial topological features.

[0178] In one embodiment, the prediction module 840 is specifically used to: extract common features of spatial topological features for multiple prediction tasks to obtain common features of the tasks; and predict the signal attributes of the signal to be identified under multiple prediction tasks based on the common features of the tasks; wherein the signal attributes under multiple prediction tasks include at least two of the following: signal category under classification prediction task, center frequency offset under center frequency offset prediction task, and signal bandwidth under signal bandwidth prediction task.

[0179] Each module in the aforementioned signal recognition device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0180] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores time-frequency characteristic images, key signal features, spatial topological features, and signal attributes. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a signal recognition method.

[0181] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0182] In one exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the signal recognition method provided in any of the above embodiments or to implement the following steps:

[0183] Obtain the time-frequency feature image corresponding to the signal to be identified;

[0184] Extract key signal features from time-frequency feature images;

[0185] Extracting spatial topological features from key features of the signal;

[0186] Based on spatial topological features, predict the signal attributes of the signal to be identified.

[0187] In one embodiment, when the processor executes the computer program, it further performs the following steps: performing dimensionality upscaling on the time-frequency feature image to obtain a first feature image; determining the channel attention weight corresponding to each channel based on the number of channels in the first feature image; and performing channel attention enhancement on the first feature image based on the channel attention weight corresponding to each channel to obtain key signal features.

[0188] In one embodiment, when the processor executes the computer program, it further performs the following steps: extracting global features of the first feature image to obtain channel descriptors of the signal to be identified; for each channel of the first feature image, determining the kernel size of the target convolution used by the channel based on the number of channels; and extracting features from the channel descriptors based on the target convolution corresponding to each channel to determine the channel attention weights corresponding to each channel.

[0189] In one embodiment, when the processor executes the computer program, it further performs the following steps: resizing the time-frequency feature image; normalizing the resized time-frequency feature image to obtain a single-channel grayscale image; and upscaling the single-channel grayscale image to obtain a first feature image.

[0190] In one embodiment, when the processor executes the computer program, it further performs the following steps: extracting the spatial distribution features of key signal features under different receptive fields; determining the spatial attention weights corresponding to the key signal features based on the spatial distribution features under different receptive fields; and performing spatial attention enhancement on the key signal features based on the spatial attention weights to obtain spatial topological features.

[0191] In one embodiment, when the processor executes the computer program, it further performs the following steps: extracting common features of spatial topological features for multiple prediction tasks to obtain common features of the tasks; predicting the signal attributes of the signal to be identified under multiple prediction tasks based on the common features of the tasks; wherein the signal attributes under multiple prediction tasks include at least two of the following: signal category under classification prediction task, center frequency offset under center frequency offset prediction task, and signal bandwidth under signal bandwidth prediction task.

[0192] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When executed by a processor, the computer program implements the steps of the signal recognition method provided in any of the above embodiments or implements the following steps:

[0193] Obtain the time-frequency feature image corresponding to the signal to be identified;

[0194] Extract key signal features from time-frequency feature images;

[0195] Extracting spatial topological features from key features of the signal;

[0196] Based on spatial topological features, predict the signal attributes of the signal to be identified.

[0197] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: performing dimensionality upscaling on the time-frequency feature image to obtain a first feature image; determining the channel attention weight corresponding to each channel based on the number of channels in the first feature image; and performing channel attention enhancement on the first feature image based on the channel attention weight corresponding to each channel to obtain key signal features.

[0198] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: extracting global features of the first feature image to obtain channel descriptors of the signal to be identified; for each channel of the first feature image, determining the kernel size of the target convolution used by the channel based on the number of channels; extracting features from the channel descriptors based on the target convolution corresponding to each channel, and determining the channel attention weights corresponding to each channel.

[0199] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: resizing the time-frequency feature image; normalizing the resized time-frequency feature image to obtain a single-channel grayscale image; and upscaling the single-channel grayscale image to obtain a first feature image.

[0200] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: extracting the spatial distribution features of key signal features under different receptive fields; determining the spatial attention weights corresponding to the key signal features based on the spatial distribution features under different receptive fields; and performing spatial attention enhancement on the key signal features based on the spatial attention weights to obtain spatial topological features.

[0201] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: extracting common features of spatial topological features for multiple prediction tasks to obtain common features of the tasks; predicting the signal attributes of the signal to be identified under multiple prediction tasks based on the common features of the tasks; wherein the signal attributes under multiple prediction tasks include at least two of the following: signal category under classification prediction task, center frequency offset under center frequency offset prediction task, and signal bandwidth under signal bandwidth prediction task.

[0202] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the signal recognition method provided in any of the above embodiments, or implements the following steps:

[0203] Obtain the time-frequency feature image corresponding to the signal to be identified;

[0204] Extract key signal features from time-frequency feature images;

[0205] Extracting spatial topological features from key features of the signal;

[0206] Based on spatial topological features, predict the signal attributes of the signal to be identified.

[0207] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: performing dimensionality upscaling on the time-frequency feature image to obtain a first feature image; determining the channel attention weight corresponding to each channel based on the number of channels in the first feature image; and performing channel attention enhancement on the first feature image based on the channel attention weight corresponding to each channel to obtain key signal features.

[0208] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: extracting global features of the first feature image to obtain channel descriptors of the signal to be identified; for each channel of the first feature image, determining the kernel size of the target convolution used by the channel based on the number of channels; extracting features from the channel descriptors based on the target convolution corresponding to each channel, and determining the channel attention weights corresponding to each channel.

[0209] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: resizing the time-frequency feature image; normalizing the resized time-frequency feature image to obtain a single-channel grayscale image; and upscaling the single-channel grayscale image to obtain a first feature image.

[0210] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: extracting the spatial distribution features of key signal features under different receptive fields; determining the spatial attention weights corresponding to the key signal features based on the spatial distribution features under different receptive fields; and performing spatial attention enhancement on the key signal features based on the spatial attention weights to obtain spatial topological features.

[0211] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: extracting common features of spatial topological features for multiple prediction tasks to obtain common features of the tasks; predicting the signal attributes of the signal to be identified under multiple prediction tasks based on the common features of the tasks; wherein the signal attributes under multiple prediction tasks include at least two of the following: signal category under classification prediction task, center frequency offset under center frequency offset prediction task, and signal bandwidth under signal bandwidth prediction task.

[0212] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0213] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0214] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A signal recognition method, characterized in that, The method includes: Obtain the time-frequency feature image corresponding to the signal to be identified; Extract key signal features from the time-frequency feature image; Extract spatial topological features from the key features of the signal; Based on the spatial topological features, predict the signal attributes of the signal to be identified.

2. The method according to claim 1, characterized in that, The extraction of key signal features from the time-frequency feature image includes: The time-frequency feature image is subjected to dimensionality upscaling to obtain the first feature image; Based on the number of channels in the first feature image, determine the channel attention weight corresponding to each channel; Based on the channel attention weights corresponding to each channel, channel attention enhancement is performed on the first feature image to obtain the key features of the signal.

3. The method according to claim 2, characterized in that, The step of determining the channel attention weights corresponding to each channel based on the number of channels in the first feature image includes: Extract the global features of the first feature image to obtain the channel descriptor of the signal to be identified; For each channel of the first feature image, the kernel size of the target convolution used by the channel is determined according to the number of channels in the channel; Based on the target convolution corresponding to each channel, feature extraction is performed on the channel descriptor to determine the channel attention weight corresponding to each channel.

4. The method according to claim 2, characterized in that, The step of performing dimensionality upscaling on the time-frequency feature image to obtain the first feature image includes: The time-frequency feature image is resized; The resized time-frequency feature image is normalized to obtain a single-channel grayscale image; The single-channel grayscale image is subjected to dimensionality upscaling to obtain the first feature image.

5. The method according to claim 1, characterized in that, The extraction of spatial topological features from the key features of the signal includes: Extract the spatial distribution characteristics of the key features of the signal under different receptive fields; Based on the spatial distribution characteristics under different receptive fields, determine the spatial attention weights corresponding to the key features of the signal; Based on the spatial attention weights, spatial attention enhancement is performed on the key features of the signal to obtain the spatial topological features.

6. The method according to any one of claims 1-5, characterized in that, The step of predicting the signal attributes of the signal to be identified based on the spatial topological features includes: Common features of multiple prediction tasks are extracted from the spatial topological features to obtain common features of the tasks; Based on the common features of the tasks, predict the signal properties of the signal to be identified under multiple prediction tasks; The signal attributes under the multiple prediction tasks include at least two of the following: signal category under the classification prediction task, center frequency offset under the center frequency offset prediction task, and signal bandwidth under the signal bandwidth prediction task.

7. A signal recognition device, characterized in that, The device includes: The acquisition module is used to acquire the time-frequency feature image corresponding to the signal to be identified; The first extraction module is used to extract key signal features from the time-frequency feature image; The second extraction module is used to extract spatial topological features from the key features of the signal; The prediction module is used to predict the signal attributes of the signal to be identified based on the spatial topological features.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images