Resource element classification method, apparatus, device, medium, and computer program product

Abstract

This application provides a resource element classification method, apparatus, device, medium, and computer program product. The resource element classification method includes: constructing a first time-frequency resource matrix corresponding to a received 5G time-domain signal; and using a ResNet-Unet network to obtain the category of physical REs corresponding to the first time-frequency resource matrix. The ResNet-Unet network represents a network based on a residual network and an encoder-decoder structure. The categories include at least one of: background noise REs, synchronization signal block (SSB) REs, common physical downlink control channel (PDCCH) REs, common physical downlink shared channel (PDSCH) REs, dedicated PDCCH REs, dedicated PDSCH REs, and channel state information reference signal (CSI-RS) REs. This solution can support expanding RE identification granularity, improving RE category identification accuracy, expanding the identification range, improving identification capability, simplifying the scheme process, and reducing process complexity. It effectively solves at least one of the problems of limited identification capability, low accuracy, and complex process in existing resource element classification schemes.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a method, apparatus, device, medium and computer program product for classifying resource elements. Background Technology

[0002] To address the issue of identifying 5G downlink resource elements, existing technologies typically employ the following technical solutions: Scheme 1 is an energy-detection-based NR downlink physical resource element identification scheme. This method obtains the corresponding time-frequency resource matrix after OFDM demodulation of the 5G NR downlink time-domain signal, and calculates the amplitude or power of each resource element. By setting a power threshold, resource elements with power greater than the threshold are identified as signal-carrying resource elements, while those with power less than the threshold are identified as background noise resource elements. This scheme is simple to implement, has low computational complexity, and does not rely on higher-layer protocol parsing. However, it only uses the energy characteristics of resource elements for judgment, resulting in limited granularity. It can only distinguish between signals and noise within resource elements, and cannot further distinguish between different types of physical channels or reference signals. Furthermore, in situations with low signal-to-noise ratios, the energy differences between different signals and between signals and noise decrease, leading to a decline in identification accuracy and limiting the applicability of this type of scheme in complex wireless environments.

[0003] Scheme 2 is an NR downlink physical resource element identification scheme based on protocol parsing. This scheme follows the 5G NR downlink decoding process, sequentially performing synchronization signal detection, physical broadcast channel decoding, and system information parsing on the received downlink signal. By decoding the Synchronization Signal Block (SSB) and obtaining the Master Information Block (MIB) and System Information Block 1 (SIB1), system configuration information related to the synchronization signal block, downlink control channel, and data channel can be obtained. This allows for the determination of the positions of the synchronization signal block, common control channel, and common data channel in the time-frequency resource matrix, thus enabling the identification of the corresponding resource elements. Based on the mapping relationships explicitly defined in the 3GPP protocol, this scheme achieves high identification accuracy under different signal-to-noise ratios. However, this scheme is highly dependent on the complete decoding process and system configuration information. If synchronization, broadcast channel decoding, or system information parsing fails, subsequent resource element identification will be difficult to continue, affecting the robustness of the identification process. Furthermore, this scheme typically can only identify resource element types configured during the system information phase. For resource elements such as dedicated control channels, dedicated data channels, and some reference signals, effective identification is difficult without corresponding configuration information, limiting its overall ability to identify NR downlink physical resource elements. In addition, this scheme requires multi-level channel decoding and protocol processing, resulting in a complex processing flow that makes it difficult to achieve rapid, multi-category, and refined identification of downlink physical resource elements.

[0004] In summary, existing resource element classification schemes suffer from at least one of the following problems: limited identification capability, low accuracy, and complex process. They cannot support rapid identification of multiple categories and refined resource elements at the level of physical channels and reference signals in NR downlink signals. Summary of the Invention

[0005] The purpose of this application is to provide a method, apparatus, equipment, medium, and computer program product for classifying resource elements, so as to solve at least one of the problems of limited identification ability, low accuracy, and complex process in the existing resource element classification schemes.

[0006] To address the aforementioned technical problems, embodiments of this application provide a resource element classification method, including: Construct the first time-frequency resource matrix corresponding to the received 5G time-domain signal; Using the ResNet-Unet network, the category of the physical resource element RE corresponding to the first time-frequency resource matrix is ​​obtained; the ResNet-Unet network represents a network based on a residual network and an encoder-decoder structure; the category includes at least one of the following: background noise RE, synchronization signal block SSB RE, common physical downlink control channel PDCCH RE, common physical downlink shared channel PDSCH RE, dedicated PDCCH RE, dedicated PDSCH RE, and channel state information reference signal CSI-RS RE.

[0007] Optionally, the construction of the first time-frequency resource matrix corresponding to the received 5G time-domain signal includes: Downlink synchronization processing is performed on the received 5G time-domain signal to obtain the first time-domain signal; The first time-domain signal is demodulated using Orthogonal Frequency Division Multiplexing (OFDM) to obtain the second time-frequency resource matrix. The second time-frequency resource matrix is ​​filled according to the data input size of the ResNet-Unet network to obtain the first time-frequency resource matrix.

[0008] Optionally, obtaining the category of the physical resource element (RE) corresponding to the first time-frequency resource matrix using the ResNet-Unet network includes: Using the ResNet-Unet network, the category probability distribution of physical REs corresponding to the first time-frequency resource matrix is ​​obtained; Based on the category probability distribution, the category of the physical RE corresponding to the first time-frequency resource matrix is ​​determined.

[0009] Optional, also includes: Obtain the model training signal and the RE category label corresponding to the model training signal, wherein the model training signal includes: simulation signal and acquired 5G time domain signal; Using the model training signal and RE category labels, a semantic segmentation network is trained to obtain the ResNet-Unet network.

[0010] Optionally, the step of training a semantic segmentation network using the model training signal and RE category labels to obtain the ResNet-Unet network includes: The semantic segmentation network is trained using the first loss function, the model training signal, and the RE category labels to obtain the ResNet-Unet network; the first loss function includes: the Dice loss function and the weight-based focus loss function.

[0011] Optionally, the weights in the focus loss function are the weights corresponding to the true categories of physical REs, and the weights are determined based on the median frequency, which is determined based on the frequency of RE categories.

[0012] Optionally, obtaining the RE category label corresponding to the model training signal includes: Based on the signal generation parameters of the simulated signal, determine the RE category label corresponding to the simulated signal; And / or, based on the physical broadcast channel PBCH decoding information corresponding to the acquired 5G time-domain signal and the first information, determine the RE category label corresponding to the acquired 5G time-domain signal. The first information includes the RE position and the number of REs occupied by each type of signal in the power time-frequency diagram, and the power time-frequency diagram corresponds to the acquired 5G time-domain signal.

[0013] This application embodiment also provides a resource element classification device, including: The first construction module is used to construct the first time-frequency resource matrix corresponding to the received 5G time-domain signal; The first acquisition module is used to acquire the category of physical resource element RE corresponding to the first time-frequency resource matrix using a ResNet-Unet network; the ResNet-Unet network represents a network based on a residual network and an encoder-decoder structure; the category includes at least one of the following: background noise RE, synchronization signal block SSB RE, common physical downlink control channel PDCCH RE, common physical downlink shared channel PDSCH RE, dedicated PDCCH RE, dedicated PDSCH RE, and channel state information reference signal CSI-RS RE.

[0014] Optionally, the construction of the first time-frequency resource matrix corresponding to the received 5G time-domain signal includes: Downlink synchronization processing is performed on the received 5G time-domain signal to obtain the first time-domain signal; The first time-domain signal is demodulated using Orthogonal Frequency Division Multiplexing (OFDM) to obtain the second time-frequency resource matrix. The second time-frequency resource matrix is ​​filled according to the data input size of the ResNet-Unet network to obtain the first time-frequency resource matrix.

[0015] Optionally, obtaining the category of the physical resource element (RE) corresponding to the first time-frequency resource matrix using the ResNet-Unet network includes: Using the ResNet-Unet network, the category probability distribution of physical REs corresponding to the first time-frequency resource matrix is ​​obtained; Based on the category probability distribution, the category of the physical RE corresponding to the first time-frequency resource matrix is ​​determined.

[0016] Optional, also includes: The second acquisition module is used to acquire the model training signal and the RE category label corresponding to the model training signal. The model training signal includes: simulation signal and acquired 5G time domain signal. The first training module is used to train the semantic segmentation network using the model training signal and RE category labels to obtain the ResNet-Unet network.

[0017] Optionally, the step of training a semantic segmentation network using the model training signal and RE category labels to obtain the ResNet-Unet network includes: The semantic segmentation network is trained using the first loss function, the model training signal, and the RE category labels to obtain the ResNet-Unet network; the first loss function includes: the Dice loss function and the weight-based focus loss function.

[0018] Optionally, the weights in the focus loss function are the weights corresponding to the true categories of physical REs, and the weights are determined based on the median frequency, which is determined based on the frequency of RE categories.

[0019] Optionally, obtaining the RE category label corresponding to the model training signal includes: Based on the signal generation parameters of the simulated signal, determine the RE category label corresponding to the simulated signal; And / or, based on the physical broadcast channel PBCH decoding information corresponding to the acquired 5G time-domain signal and the first information, determine the RE category label corresponding to the acquired 5G time-domain signal. The first information includes the RE position and the number of REs occupied by each type of signal in the power time-frequency diagram, and the power time-frequency diagram corresponds to the acquired 5G time-domain signal.

[0020] This application embodiment also provides a resource element classification device, including: a processor; The processor is used to construct a first time-frequency resource matrix corresponding to the received 5G time-domain signal; Using the ResNet-Unet network, the category of the physical resource element RE corresponding to the first time-frequency resource matrix is ​​obtained; the ResNet-Unet network represents a network based on a residual network and an encoder-decoder structure; the category includes at least one of the following: background noise RE, synchronization signal block SSB RE, common physical downlink control channel PDCCH RE, common physical downlink shared channel PDSCH RE, dedicated PDCCH RE, dedicated PDSCH RE, and channel state information reference signal CSI-RS RE.

[0021] Optionally, the construction of the first time-frequency resource matrix corresponding to the received 5G time-domain signal includes: Downlink synchronization processing is performed on the received 5G time-domain signal to obtain the first time-domain signal; The first time-domain signal is demodulated using Orthogonal Frequency Division Multiplexing (OFDM) to obtain the second time-frequency resource matrix. The second time-frequency resource matrix is ​​filled according to the data input size of the ResNet-Unet network to obtain the first time-frequency resource matrix.

[0022] Optionally, obtaining the category of the physical resource element (RE) corresponding to the first time-frequency resource matrix using the ResNet-Unet network includes: Using the ResNet-Unet network, the category probability distribution of physical REs corresponding to the first time-frequency resource matrix is ​​obtained; Based on the category probability distribution, the category of the physical RE corresponding to the first time-frequency resource matrix is ​​determined.

[0023] Optionally, the processor is further configured to: Obtain the model training signal and the RE category label corresponding to the model training signal, wherein the model training signal includes: simulation signal and acquired 5G time domain signal; Using the model training signal and RE category labels, a semantic segmentation network is trained to obtain the ResNet-Unet network.

[0024] Optionally, the step of training a semantic segmentation network using the model training signal and RE category labels to obtain the ResNet-Unet network includes: The semantic segmentation network is trained using the first loss function, the model training signal, and the RE category labels to obtain the ResNet-Unet network; the first loss function includes: the Dice loss function and the weight-based focus loss function.

[0025] Optionally, the weights in the focus loss function are the weights corresponding to the true categories of physical REs, and the weights are determined based on the median frequency, which is determined based on the frequency of RE categories.

[0026] Optionally, obtaining the RE category label corresponding to the model training signal includes: Based on the signal generation parameters of the simulated signal, determine the RE category label corresponding to the simulated signal; And / or, based on the physical broadcast channel PBCH decoding information corresponding to the acquired 5G time-domain signal and the first information, determine the RE category label corresponding to the acquired 5G time-domain signal. The first information includes the RE position and the number of REs occupied by each type of signal in the power time-frequency diagram, and the power time-frequency diagram corresponds to the acquired 5G time-domain signal.

[0027] This application embodiment also provides a resource element classification device, including a memory, a processor, and a program stored in the memory and executable on the processor; when the processor executes the program, it implements the above-described resource element classification method.

[0028] This application also provides a readable storage medium storing a program that, when executed by a processor, implements the steps in the resource element classification method described above.

[0029] This application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the resource element classification method described above.

[0030] The beneficial effects of the above technical solution in this application are as follows: In the above scheme, the resource element classification method constructs a first time-frequency resource matrix corresponding to the received 5G time-domain signal; it uses a ResNet-Unet network to obtain the category of physical resource elements (REs) corresponding to the first time-frequency resource matrix; the ResNet-Unet network represents a network based on a residual network and an encoder-decoder structure; the categories include: background noise REs, synchronization signal block (SSB) REs, common physical downlink control channel (PDCCH) REs, common physical downlink shared channel (PDSCH) REs, dedicated PDCCH REs, dedicated PDSCH REs, and channel state information reference signal (CSI-RS). At least one of the following: capable of supporting the differentiation of different types of physical channels or reference signals to expand the RE identification granularity; capable of improving the RE category identification accuracy under different signal-to-noise ratio conditions based on the ResNet-Unet network; capable of avoiding reliance on configuration information for RE identification, thereby expanding the identification range and improving identification capability; and capable of simplifying the scheme process and reducing process complexity by eliminating the need for multi-level channel decoding and protocol processing; capable of supporting the realization of fast, multi-category, and refined identification of downlink physical resource elements; and capable of effectively solving at least one of the following problems in existing resource element classification schemes: limited identification capability, low accuracy, and complex process. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the resource element classification method according to an embodiment of this application; Figure 2 This is a schematic diagram of the resource element classification implementation architecture in an embodiment of this application; Figure 3 This is a schematic diagram illustrating the specific implementation process of the resource element classification method in this application embodiment; Figure 4 This is a schematic diagram of the ResNet-Unet network architecture according to an embodiment of this application; Figure 5 This is a time-frequency diagram of signal power in an embodiment of this application; Figure 6 This is a schematic diagram of the RE category label in an embodiment of this application; Figure 7 This is a schematic diagram of RE prediction categories in an embodiment of this application; Figure 8 This is a schematic diagram illustrating the performance indicators of different signal-to-noise ratios in embodiments of this application; Figure 9 This is a schematic diagram of the category performance indicators under different signal-to-noise ratios in embodiments of this application. Figure 1 ; Figure 10 This is a schematic diagram of the category performance indicators under different signal-to-noise ratios in embodiments of this application. Figure 2 ; Figure 11 This is a schematic diagram of the category performance indicators under different signal-to-noise ratios in embodiments of this application. Figure 3 ; Figure 12 This is a schematic diagram of the category performance indicators under different signal-to-noise ratios in embodiments of this application. Figure 4 ; Figure 13 This is a schematic diagram of the category performance indicators under different signal-to-noise ratios in embodiments of this application. Figure 5 ; Figure 14 This is a schematic diagram of the resource element classification device according to an embodiment of this application; Figure 15 This is a schematic diagram of the resource element classification device structure according to an embodiment of this application. Detailed Implementation

[0032] To make the technical problems, technical solutions and advantages of this application clearer, a detailed description will be provided below in conjunction with the accompanying drawings and specific embodiments.

[0033] The following is a brief introduction to the relevant content of this plan.

[0034] With the rapid development of 5G NR (Normally Injected) technology, communication systems have achieved significant improvements in data rates, connection scale, and latency. In 5G NR systems, downlink signals typically employ OFDM (Orthogonal Frequency Division Multiplexing) modulation. After OFDM demodulation, the receiver obtains a two-dimensional time-frequency resource grid composed of numerous resource elements (REs) in both the time and frequency domains. Different types of physical channels and reference signals (such as synchronization signal blocks (SSBs), physical downlink control channels (PDCCHs), physical downlink shared channels (PDSCHs), and channel state information reference signals (CSI-RS)) are mapped to the two-dimensional time-frequency resource grid with different time-frequency structures, thereby enabling flexible scheduling and efficient utilization of wireless resources.

[0035] However, with the deployment of 5G networks and the continuous expansion of application scenarios, the demand for analyzing, monitoring, and intelligently processing 5G signals is growing. For example, in applications such as wireless network optimization, spectrum regulation, intelligent sensing, and AI-based wireless communication processing, fine-grained analysis of received 5G signals is often required to identify the signal types carried on different time-frequency resources. In these applications, simply acquiring the complex signal data after OFDM demodulation is often insufficient to meet practical needs; it is usually necessary to further determine the physical channel or reference signal type corresponding to each resource element (RE), i.e., to identify and classify the resource elements in the 5G downlink signal. Therefore, how to accurately identify the resource elements after OFDM demodulation at the receiving end has become an important technical problem in the field of 5G signal processing.

[0036] Based on the above, this application addresses at least one of the problems of limited recognition capability, low accuracy, and complex process in existing resource element classification schemes, and provides a resource element classification method, such as... Figure 1 As shown, it includes: Step 11: Construct the first time-frequency resource matrix corresponding to the received 5G time-domain signal; Step 12: Using the ResNet-Unet network, obtain the category of the physical resource element RE corresponding to the first time-frequency resource matrix; the ResNet-Unet network represents a network based on a residual network and an encoder-decoder structure; the category includes at least one of the following: background noise RE, synchronization signal block SSB RE, common physical downlink control channel PDCCH RE, common physical downlink shared channel PDSCH RE, dedicated PDCCH RE, dedicated PDSCH RE, and channel state information reference signal CSI-RS RE.

[0037] The resource element classification method provided in this application constructs a first time-frequency resource matrix corresponding to the received 5G time-domain signal; it uses a ResNet-Unet network to obtain the category of physical resource elements (REs) corresponding to the first time-frequency resource matrix; the ResNet-Unet network represents a network based on a residual network and an encoder-decoder structure; the categories include: background noise REs, synchronization signal block (SSB) REs, common physical downlink control channel (PDCCH) REs, common physical downlink shared channel (PDSCH) REs, dedicated PDCCH REs, and dedicated PDSCH. It includes at least one of RE and Channel State Information Reference Signal (CSI-RSRE); it can support the differentiation of different types of physical channels or reference signals to expand the RE identification granularity; it can also improve the RE category identification accuracy under different signal-to-noise ratio conditions based on the ResNet-Unet network; it can avoid relying on configuration information for RE identification, thereby expanding the identification range and improving the identification capability; and it no longer requires multi-level channel decoding and protocol processing, thereby simplifying the scheme process and reducing the process complexity; it can support the realization of fast, multi-category fine identification of downlink physical resource elements; and it can effectively solve at least one of the problems of limited identification capability, low accuracy and complex process in existing resource element classification schemes.

[0038] The construction of the first time-frequency resource matrix corresponding to the received 5G time-domain signal includes: performing downlink synchronization processing on the received 5G time-domain signal to obtain a first time-domain signal; performing Orthogonal Frequency Division Multiplexing (OFDM) demodulation on the first time-domain signal to obtain a second time-frequency resource matrix; and performing a padding operation on the second time-frequency resource matrix according to the data input size of the ResNet-Unet network to obtain the first time-frequency resource matrix. This allows for accurate acquisition of the first time-frequency resource matrix. The phrase "performing downlink synchronization processing on the received 5G time-domain signal to obtain a first time-domain signal" may include: performing frequency offset correction on the received 5G time-domain signal through primary synchronization signal (PSS) and secondary synchronization signal (SSS) detection to obtain a corrected signal; performing PBCH detection on the corrected signal to obtain the first time-domain signal; and / or, "performing orthogonal frequency division multiplexing (OFDM) demodulation on the first time-domain signal to obtain a second time-frequency resource matrix" may include: performing physical broadcast channel (PBCH) decoding on the first time-domain signal to obtain decoding parameters (such as necessary system information); performing OFDM demodulation on the first time-domain signal according to the decoding parameters to obtain a second time-frequency resource matrix; and / or, the second time-frequency resource matrix may be implemented as a complex matrix, but is not limited thereto.

[0039] In this embodiment of the application, obtaining the category of the physical resource element (RE) corresponding to the first time-frequency resource matrix using the ResNet-Unet network includes: obtaining the category probability distribution of the physical RE corresponding to the first time-frequency resource matrix using the ResNet-Unet network; and determining the category of the physical RE corresponding to the first time-frequency resource matrix based on the category probability distribution. This allows for the specific acquisition of the category of the physical RE.

[0040] Furthermore, the resource element classification method further includes: acquiring model training signals and corresponding RE category labels, wherein the model training signals include: simulated signals and acquired 5G time-domain signals; and using the model training signals and RE category labels to train a semantic segmentation network to obtain the ResNet-Unet network. This can improve the network training quality and support obtaining a high-accuracy ResNet-Unet network.

[0041] The step of training a semantic segmentation network using the model training signal and RE class labels to obtain the ResNet-Unet network includes: training the semantic segmentation network based on a first loss function, the model training signal, and RE class labels to obtain the ResNet-Unet network; the first loss function includes a Dice loss function and a weight-based focus loss function. This can help alleviate the class imbalance problem of the network and improve the accuracy of boundary recognition.

[0042] In this embodiment, the weights in the focus loss function are the weights corresponding to the true categories of the physical REs. These weights are determined based on the median frequency, which is determined based on the frequency of the RE categories. This allows for obtaining reasonable weights. The phrase "the weights are determined based on the median frequency, which is determined based on the frequency of the RE categories" can also be understood as the weights being determined based on the median frequency determined by the frequency of the RE categories. Category frequency can refer to the frequency of categories in the current batch of training data (training signal), but is not limited to this.

[0043] The process of obtaining the RE category label corresponding to the model training signal includes: determining the RE category label corresponding to the simulation signal based on the signal generation parameters of the simulation signal; and / or, determining the RE category label corresponding to the collected 5G time-domain signal based on the Physical Broadcast Channel (PBCH) decoding information corresponding to the collected 5G time-domain signal and first information, wherein the first information includes the RE position and the number of REs occupied by each type of signal in the power time-frequency diagram, and the power time-frequency diagram corresponds to the collected 5G time-domain signal. This allows for accurate determination of the RE category label corresponding to the model training signal.

[0044] The resource element classification method provided in the embodiments of this application will be illustrated with examples below.

[0045] To address the aforementioned technical problems (such as the coarse identification granularity, reliance on the complete decoding process, and insufficient generalization ability of existing NR downlink physical resource element (RE) identification methods based on energy detection and protocol parsing), this application provides a resource element classification method. Specifically, it can be implemented as a deep learning-based NR (New Radio) downlink physical resource unit identification method. This method can identify different categories of physical resource elements in the downlink signal of a base station in 5G mobile communication technology (NR). It supports high-precision, multi-category identification of downlink physical resource elements by fully utilizing the structural characteristics of the NR downlink signal in the time-frequency resource matrix without relying on the complete NR downlink decoding process. It can also improve the robustness of identification under different signal-to-noise ratio conditions, thereby enhancing the applicability and practical value of the NR downlink physical resource element identification method. Thus, it can effectively support the rapid identification of multiple categories and fine-grained resource elements at the level of physical channels and reference signals in NR downlink signals.

[0046] Optionally, embodiments of this application can implement an end-to-end NR downlink signal resource element identification scheme based on deep learning semantic segmentation. The time-frequency resource matrix of the NR downlink signal can be treated as an image, and REs with different functions (or different categories) can be treated as different categories of "pixels." Then, a deep semantic segmentation network can be trained to achieve pixel-level classification of each RE, thereby bypassing the complex protocol decoding process and directly and precisely identifying various resource elements in the signal. Based on this, the scheme provided by embodiments of this application may include the following parts: First, a hybrid dataset containing simulated signals and live network acquired signals is constructed. For example, the MATLAB 5G toolbox is used to simulate standard NR downlink signals at different signal-to-noise ratios (0~20dB), while simultaneously acquiring live 5G (time domain) signals. After performing the same downlink synchronization and demodulation preprocessing on both types of signals, a standard time-frequency resource matrix and its corresponding category label matrix are generated, thus providing a high-quality and diverse data foundation for model training. Corresponding to the acquisition of model training signals and their corresponding RE category labels, the model training signals include simulated signals and acquired 5G time domain signals. Using the model training signals and RE category labels, a semantic segmentation network is trained to obtain the ResNet-Unet network.

[0047] Secondly, a deep learning model based on an improved ResNet (Residual Network)-Unet (Encoder-Decoder Structure) network was designed and implemented (which can be understood as a ResNet-Unet network model, or simply a ResNet-Unet network). This model uses ResNet50 as the encoder, leveraging its powerful feature extraction capabilities to capture deep semantic information of the signal; the decoder part fuses features from different levels of the encoder through upsampling and skip connections, which can recover detailed information and achieve accurate pixel-level classification. The model's input and output dimensions can be zero-padding to adapt to the specific time-frequency resource matrix size (3276×280) for handling NR signals.

[0048] Finally, a complete training, inference, and performance verification process was established to verify the significant advantages of this scheme in terms of recognition accuracy and robustness on multiple evaluation metrics.

[0049] The implementation of this solution may include the following: Such as Figure 2The overall structure shown implements this scheme, which includes five functional modules: a signal acquisition module, a downlink synchronization and demodulation module, a resource matrix construction module, a deep learning inference module, and a resource element classification output module. The signal acquisition module acquires 5G time-domain signals; the downlink synchronization and demodulation module performs downlink synchronization on the 5G time-domain signals and then performs OFDM demodulation to obtain a complete time-frequency resource matrix; given that the network (in the deep learning inference module) performs five downsampling operations, the network's input tensor must be a multiple of 32. Therefore, in this scheme, the resource matrix construction module zero-padding is used to fill the OFDM-demodulated time-frequency resource matrix and the corresponding label matrix to a size suitable for network processing, which can then be used for model training and / or inference. The deep learning inference module can serve as an NR downlink physical resource element identification network (specifically, a ResNet-Unet network structure) to classify physical resource elements in the NR downlink signal. The resource element classification output module can be used to process the tensor output by the network to obtain a mask matrix. The value of each element in the matrix can be 0 to 6 (corresponding to 7 different categories). During the model training phase, this module can also be used to calculate recognition metrics (i.e., to use a validation set to confirm the model performance).

[0050] The overall workflow of this solution can be described as follows: Figure 3 As shown, it includes: (1) Perform data preprocessing on the original time-domain signals (such as simulated 5G downlink time-domain signals (i.e., simulated signals), acquired 5G downlink time-domain signals, etc.) to complete the downlink synchronization process and obtain the time-frequency resource matrix; specifically, it may include: Frequency offset correction is completed by detecting the primary synchronization signal (PSS) and secondary synchronization signal (SSS) to obtain the corrected signal; then time and frequency synchronization is obtained based on the corrected signal, that is, the target signal is obtained; downlink synchronization processing can be performed on the received 5G time domain signal as described above to obtain the first time domain signal.

[0051] Next, the target signal is decoded using the Physical Broadcast Channel (PBCH) to obtain necessary system information. Then, based on the obtained system information, the target signal is demodulated using OFDM to obtain a two-dimensional time-frequency resource grid, which is a 3276×280 complex matrix (i.e., the time-frequency resource matrix). This matrix represents 3276 subcarriers in the frequency domain and 280 OFDM symbols in the time domain. Each point in this grid corresponds to a resource element (RE). This can be used to obtain the second time-frequency resource matrix by performing Orthogonal Frequency Division Multiplexing (OFDM) demodulation on the first time-domain signal.

[0052] To adapt to the input size requirements of the network, the complex matrix is ​​zero-padded to a size of 3296×288. Then, the real and imaginary parts of the matrix are extracted and stored as a 3296×288×2 three-dimensional array. This three-dimensional array can be used as input to the ResNet-Unet network. Corresponding to the above-mentioned padding operation on the second time-frequency resource matrix according to the data input size of the ResNet-Unet network, the first time-frequency resource matrix is ​​obtained.

[0053] In addition, for the simulated signal, a corresponding label matrix (such as a two-dimensional array of 3296×288, with each element taking a value from 0 to 6, corresponding to any of the 7 categories) can be constructed based on the signal generation parameters of the simulated signal; the RE category label corresponding to the simulated signal can be determined based on the signal generation parameters of the simulated signal mentioned above. For the collected signals, a corresponding tag matrix can be constructed based on the decoding information (the decoding information of the PBCH mentioned above) and the power time-frequency diagram corresponding to the signal (as shown in the figure, the RE positions and RE quantities occupied by each type of signal). The RE category tag corresponding to the collected 5G time-domain signal can be determined based on the physical broadcast channel PBCH decoding information and the first information mentioned above. The first information includes the RE positions and RE quantities occupied by each type of signal in the power time-frequency diagram, and the power time-frequency diagram corresponds to the collected 5G time-domain signal.

[0054] (2) The constructed data (such as the three-dimensional array 3296×288×2 constructed above) is converted into a tensor and fed into the pre-trained network (i.e., the NR downlink physical resource element identification network) for inference. The network outputs a 7×3296×288 three-dimensional tensor, in which the 7 channels can respectively correspond to the predicted probabilities of the 7 types of REs: background noise RE, SSB-RE, common PDCCH (physical downlink control channel)-RE, common PDSCH (physical downlink shared channel)-RE, dedicated PDCCH-RE, dedicated PDSCH-RE, and CSI-RS RE. This corresponds to the above-mentioned use of the ResNet-Unet network to obtain the category probability distribution of the physical REs corresponding to the first time-frequency resource matrix. The ArgMax operation (i.e., taking the classification with the highest probability) is performed on the 7 channels of each RE to obtain the final classification label of each RE. This corresponds to the above-mentioned determination of the category of the physical REs corresponding to the first time-frequency resource matrix based on the category probability distribution. Finally, the system can output multi-category resource element identification results and further calculate the overall index (overall performance index value) and the indexes for each category (performance index values ​​for each category). Here, ArgMax represents the value or position of the independent variable (parameter) that maximizes the objective function. The calculation of the overall index and the indexes for each category can be performed during the model or ResNet-Unet network training phase, but is not performed when using the model or ResNet-Unet network to actually predict RE categories. However, this is not a limitation, and model or network updates can continue.

[0055] The following provides specific examples illustrating the proposed solution.

[0056] (1) On the construction of hybrid datasets; This solution can use simulation and live network data collection to build datasets, ensuring the diversity, standardization, and authenticity of the data.

[0057] For simulated signals, the MATLAB 5G toolbox can be used to generate NR downlink baseband signals at various signal-to-noise ratios (SNRs) of 0:2:20dB, based on typical 5G signal configuration parameters (such as subcarrier spacing, SSB type, common PDCCH and common PDSCH configurations) from the three major telecom operators. These signals can then be used... Figure 3 The simulation process, including PSS and SSS detection, PBCH decoding, and OFDM demodulation, yields a 3276×280 time-frequency resource matrix. Simultaneously, because the simulation environment is controllable, a 3276×280 tag matrix can be generated based on the simulation parameters (i.e., signal generation parameters), where each value represents one of seven RE categories.

[0058] For signals acquired from the existing network (i.e., real signals / actually acquired signals), a receiving antenna and software-defined radio (SDR) equipment (such as the USRP X310) can be used to acquire real NR signals. After the acquired RF signals (i.e., NR signals) are down-converted and converted from analog to digital to baseband signals, they can be imported into the MATLAB environment and processed using the same data preprocessing workflow as the simulated signals to obtain the time-frequency resource matrix. The label matrix for the real signals can be obtained through protocol parsing and / or manual annotation.

[0059] Finally, the simulated signal and the actual sampled signal are merged (the specific method is not limited). Then, the training set, validation set and test set can be divided in a ratio of 7:1.5:1.5. The training set can be used for training the model (or network), the validation set can be used for calculating the performance indicators of the model (or network) during the training process, and the test set can be used for model inference and performance verification after the model (or network) has been trained.

[0060] (2) Design and implementation of the NR downlink physical resource element identification network structure; The network structure adopted in this solution can be as follows: Figure 4 (A schematic diagram of the NR downlink physical resource element identification network structure is shown.) It can be implemented as a semantic segmentation network based on ResNet-Unet. The network input tensor size is 2×3296×288, which is obtained by zero-padding the original data (i.e., the 3276×280 complex matrix obtained above) to create a 3296×288×2 three-dimensional array. Encoder ( Figure 4 The left-hand side of the network can use ResNet50 as the backbone. The input tensor sequentially passes through multiple stages, including CBR (convolution + batch normalization + ReLU activation), Maxpooling, Layer 1, Layer 2, Layer 3, and Layer 4. The spatial size of the feature maps is halved at each stage (e.g., Layer 1 outputs 824×72, Layer 2 outputs 412×36), while the number of channels doubles at each stage (e.g., Layer 1 outputs 256 channels, Layer 2 outputs 512 channels), thus supporting the extraction of multi-scale features encompassing low-level details to high-level semantics. Decoder ( Figure 4The right-hand side of the model can employ an upsampling structure symmetrical to the encoder. Each decoding layer first performs bilinear interpolation upsampling on the lower-resolution feature map from the previous layer, doubling the feature map size (e.g., 103×9 in feat 5 becomes 206×18). Then, the current feature map and the feature map of the same spatial size in the encoder are concatenated (Concat). The fused features are further processed by a "2CR (double convolution)" module consisting of two 3×3 convolutional layers and a ReLU activation function. After multiple levels of upsampling and feature fusion, the feature map is upsampled back to the input size. Finally, a 1×1 convolutional layer (Conv) maps the number of channels to the number of target classes (7), outputting a 7-dimensional vector (c=7) for each spatial location, resulting in the output Y (3296×288×7). If inference is performed, the model output can be converted into the probability of each class using the Softmax function, and then converted into the final classification label of the RE using the ArgMax operation. Figure 4 In the diagram, Up indicates upsampling. Layers 1 through 4 involve residual bottleneck blocks, which contain downsampling modules. In the box containing 2CR, c256 indicates that there are 256 output channels. The other boxes containing 2CR are similar.

[0061] (3) Regarding model training and inference; During model training, a combination of a weight-based Focal loss function and a Dice loss function can be used as the model's loss function. This corresponds to the above-mentioned training of a semantic segmentation network using the first loss function, the model training signal, and RE category labels to obtain the ResNet-Unet network. The first loss function includes the Dice loss function and a weight-based focal loss function. Wherein: 1) Regarding the Focal loss function, it can be that the average of the Focal losses corresponding to each RE is calculated and used as the Focal loss of the model in the current batch; In Focal loss, the summation part of the mean calculation is the sum of the Focal loss of each RE multiplied by the weight of the corresponding class; the "weight of the corresponding class" refers to the weight of the true class to which the RE belongs (which can be the weight of the true class of the physical RE in the above focus loss function). Obtaining the "Focal loss of RE" can include: identifying the true class of RE, obtaining the probability of that true class in the predicted class of RE, and determining the Focal loss of RE based on that probability.

[0062] 2) Regarding the Dice loss function, it can be that the average of the losses corresponding to each RE category is calculated to obtain the Dice loss of the model in the current batch. For example, 1 - the average can be used as the Dice loss of the model in the current batch. The Dice loss function can be used to measure the degree of overlap between the predicted results and the true labels.

[0063] 3) Determining the weights in the Focal loss function may include the following operations: First, based on the true classes of REs in the current batch of training data, calculate the frequency of each RE class. : ; in This represents the total number of training errors (REs) corresponding to the current batch of training data. Let i be the number of REs belonging to the i-th RE category. To prevent the minimum value from being divided by zero.

[0064] Next, the median frequency is calculated. : The median frequency mentioned above is determined based on the frequency of the RE category. Then calculate the initial weights. And smooth: ; Finally, normalization is performed to limit the range, resulting in the weight corresponding to the i-th RE category. : This corresponds to the fact that the aforementioned weights are determined based on the median frequency; Where C is the number of categories, which is 7.

[0065] Furthermore, this scheme assumes the following during training: the optimizer can be Adam, the initial learning rate can be set to 0.0001, the batch size can be set to 8, training can be performed on a server equipped with a GPU (Graphics Processing Unit), and the total training duration is 300 epochs. To prevent overfitting and improve the model's generalization ability, random rotation data augmentation can be performed on the input data during training, but this is not a limitation.

[0066] Furthermore, in practical applications, for the NR downlink signal to be identified, it only needs to be processed... Figure 3The data preprocessing process shown converts the signal into a 2×3296×288 input tensor, which is then fed into the trained model (or network). After forward propagation, the network produces a 7×3296×288 output tensor. Taking the maximum value index of the output along the channel dimension (i.e., using the classification with the highest probability for each RE) yields a 3296×288 two-dimensional classification result matrix. Then, by cropping the padded edges, the final RE classification result with the same size as the original signal's time-frequency resource matrix is ​​restored (i.e., after performing the process up to the end of the time-frequency resource matrix). Figure 3 The "output multi-category resource element recognition results" in the code yields the "recognition category". Based on the above training, this scheme can achieve good recognition results, such as... Figures 5 to 7 (The data sample (signal power time-frequency graph and labels) and the schematic diagram of the output results) demonstrate the model's recognition performance on a test sample. Figure 5 This is the original signal power time-frequency diagram. Figure 6 The actual labels are shown on the right side of the image (the background, SSB, CSI-RS, representing the seven label categories mentioned above). Figure 7 The model prediction results (background, SSB...CSI-RS on the right side of the figure represent the label categories of the above 7 categories) show that the model can accurately segment resource elements of different categories. Figures 5 to 7 The subcarrier index and OFDM symbol in the text represent the subcarrier index and OFDM symbol corresponding to the NR downlink signal.

[0067] (4) Regarding performance verification; This approach allows for the evaluation of models or networks using various metrics, such as pixel accuracy (PA), class-average pixel accuracy (MPA), class-average intersection-over-union (mIoU), frequency-weighted intersection-over-union (FWIoU), and class-average Dice coefficient. Figure 8 (Performance metrics of this scheme under different signal-to-noise ratios) As shown in the graph, the overall performance curves of this scheme under different signal-to-noise ratios show that: under a low signal-to-noise ratio of 0dB, all indicators of this scheme can still maintain a high level, and they steadily increase with the increase of the signal-to-noise ratio, demonstrating good robustness. Furthermore, as... Figures 9 to 13 The graph showing the performance metrics for identifying REs of each category under different signal-to-noise ratios demonstrates that the model (or network) of this scheme has stable identification capabilities for all 7 RE categories. In the graph, Precision represents accuracy, Recall represents recall, F1-Score represents the F1 score, and IoU represents the intersection-over-union ratio.

[0068] Based on the above, the solutions provided in this application involve the following: 1. This scheme defines the NR downlink physical resource element identification task as a semantic segmentation task. The time-frequency resource matrix obtained after OFDM demodulation of the NR downlink signal is regarded as a two-dimensional spatial data structure, and different types of resource elements are mapped to different semantic categories. Multi-category identification is completed at the resource element level through semantic segmentation.

[0069] 2. This solution designs a complete signal processing workflow, including downlink synchronization, OFDM demodulation, time-frequency matrix zero-filling, and network inference. This workflow combines traditional signal processing steps with a deep learning model, realizing automated output from raw time-domain signals to multi-class recognition results.

[0070] 3. To address the issue of significant differences in the proportion of different physical resource elements in the time-frequency matrix, this solution dynamically calculates (i.e., calculates separately for each batch) the weights based on median frequency balance by calculating the proportion of different categories of REs in each batch. The overall loss is calculated by combining weighted Focal Loss and Dice Loss, which not only alleviates the class imbalance problem but also improves the accuracy of boundary recognition.

[0071] 4. This solution combines multi-signal-to-noise ratio simulation signals generated by the MATLAB 5G toolbox with signals acquired from real networks for joint modeling and training, enabling the model to not only conform to standard protocol characteristics but also possess good engineering generalization capabilities.

[0072] In summary, the solution provided in this application has the following advantages: 1. This scheme achieves multi-category and refined identification of NR downlink physical resource elements, and learns the deep spatial texture features of different types of signals on the time-frequency resource grid through a semantic segmentation model, breaking the limitation of traditional energy detection methods that can only determine the presence or absence of signals.

[0073] 2. This solution can maintain a high recognition accuracy without requiring a complete downlink decoding process. By placing the recognition process after physical layer OFDM demodulation and before higher layer protocol parsing, the nonlinear mapping capability of neural networks can be used to directly extract semantics from the time-frequency resource matrix, achieving a "blind recognition" effect.

[0074] 3. This scheme exhibits strong robustness in low signal-to-noise ratio environments, particularly by utilizing Unet's codec and skip connection structure (such as...). Figure 4 The Concat method integrates multi-scale features, which enables the scheme to maintain high recognition accuracy even at low signal-to-noise ratios, significantly improving its applicability in complex wireless environments.

[0075] 4. This solution can simultaneously identify both public and user-specific resources. Existing protocol parsing methods can only identify public channel resources explicitly configured in system messages, and cannot identify user-specific resources or CSI-RS. This solution does not rely on system configuration information, but uses a data-driven approach to uniformly model and identify public and private resources, thus expanding the range of recognizable resource types.

[0076] This application also provides a resource element classification device, such as... Figure 14 As shown, it includes: The first construction module 141 is used to construct the first time-frequency resource matrix corresponding to the received 5G time-domain signal; The first acquisition module 142 is used to acquire the category of physical resource element RE corresponding to the first time-frequency resource matrix using a ResNet-Unet network; the ResNet-Unet network represents a network based on a residual network and an encoder-decoder structure; the category includes at least one of the following: background noise RE, synchronization signal block SSB RE, common physical downlink control channel PDCCH RE, common physical downlink shared channel PDSCH RE, dedicated PDCCH RE, dedicated PDSCH RE, and channel state information reference signal CSI-RS RE.

[0077] The resource element classification device provided in this application constructs a first time-frequency resource matrix corresponding to the received 5G time-domain signal; it uses a ResNet-Unet network to obtain the category of physical resource elements (REs) corresponding to the first time-frequency resource matrix; the ResNet-Unet network represents a network based on a residual network and an encoder-decoder structure; the categories include: background noise REs, synchronization signal block (SSB) REs, common physical downlink control channel (PDCCH) REs, common physical downlink shared channel (PDSCH) REs, dedicated PDCCH REs, and dedicated PDSCH. It includes at least one of RE and Channel State Information Reference Signal (CSI-RSRE); it can support the differentiation of different types of physical channels or reference signals to expand the RE identification granularity; it can also improve the RE category identification accuracy under different signal-to-noise ratio conditions based on the ResNet-Unet network; it can avoid relying on configuration information for RE identification, thereby expanding the identification range and improving the identification capability; and it no longer requires multi-level channel decoding and protocol processing, thereby simplifying the scheme process and reducing the process complexity; it can support the realization of fast, multi-category fine identification of downlink physical resource elements; and it can effectively solve at least one of the problems of limited identification capability, low accuracy and complex process in existing resource element classification schemes.

[0078] The construction of the first time-frequency resource matrix corresponding to the received 5G time-domain signal includes: performing downlink synchronization processing on the received 5G time-domain signal to obtain a first time-domain signal; performing orthogonal frequency division multiplexing (OFDM) demodulation on the first time-domain signal to obtain a second time-frequency resource matrix; and performing a filling operation on the second time-frequency resource matrix according to the data input size of the ResNet-Unet network to obtain the first time-frequency resource matrix.

[0079] In this embodiment of the application, the step of using the ResNet-Unet network to obtain the category of the physical resource element RE corresponding to the first time-frequency resource matrix includes: using the ResNet-Unet network to obtain the category probability distribution of the physical RE corresponding to the first time-frequency resource matrix; and determining the category of the physical RE corresponding to the first time-frequency resource matrix based on the category probability distribution.

[0080] Furthermore, the resource element classification device further includes: a second acquisition module, used to acquire model training signals and RE category labels corresponding to the model training signals, wherein the model training signals include: simulation signals and acquired 5G time-domain signals; and a first training module, used to train a semantic segmentation network using the model training signals and RE category labels to obtain the ResNet-Unet network.

[0081] The step of training a semantic segmentation network using the model training signal and RE class labels to obtain the ResNet-Unet network includes: training a semantic segmentation network using a first loss function, the model training signal, and RE class labels to obtain the ResNet-Unet network; the first loss function includes: the Dice loss function and a weight-based focus loss function.

[0082] In this embodiment of the application, the weights in the focus loss function are the weights corresponding to the true categories of physical REs. The weights are determined based on the median frequency, which is determined based on the frequency of RE categories.

[0083] In this embodiment of the application, obtaining the RE category label corresponding to the model training signal includes: determining the RE category label corresponding to the simulation signal based on the signal generation parameters of the simulation signal; and / or, determining the RE category label corresponding to the collected 5G time-domain signal based on the physical broadcast channel (PBCH) decoding information corresponding to the collected 5G time-domain signal and first information, wherein the first information includes the RE position and the number of REs occupied by each type of signal in the power time-frequency diagram, and the power time-frequency diagram corresponds to the collected 5G time-domain signal.

[0084] The implementation embodiments of the above-mentioned resource element classification method are all applicable to the embodiments of the resource element classification device, and can achieve the same technical effect.

[0085] This application also provides a resource element classification device, such as... Figure 15 As shown, it includes: processor 151; The processor 151 is used to construct a first time-frequency resource matrix corresponding to the received 5G time-domain signal; Using the ResNet-Unet network, the category of the physical resource element RE corresponding to the first time-frequency resource matrix is ​​obtained; the ResNet-Unet network represents a network based on a residual network and an encoder-decoder structure; the category includes at least one of the following: background noise RE, synchronization signal block SSB RE, common physical downlink control channel PDCCH RE, common physical downlink shared channel PDSCH RE, dedicated PDCCH RE, dedicated PDSCH RE, and channel state information reference signal CSI-RS RE.

[0086] The resource element classification device provided in this application constructs a first time-frequency resource matrix corresponding to the received 5G time-domain signal; it uses a ResNet-Unet network to obtain the category of physical resource elements (REs) corresponding to the first time-frequency resource matrix; the ResNet-Unet network represents a network based on a residual network and an encoder-decoder structure; the categories include: background noise REs, synchronization signal block (SSB) REs, common physical downlink control channel (PDCCH) REs, common physical downlink shared channel (PDSCH) REs, dedicated PDCCH REs, and dedicated PDSCH. It includes at least one of RE and Channel State Information Reference Signal (CSI-RSRE); it can support the differentiation of different types of physical channels or reference signals to expand the RE identification granularity; it can also improve the RE category identification accuracy under different signal-to-noise ratio conditions based on the ResNet-Unet network; it can avoid relying on configuration information for RE identification, thereby expanding the identification range and improving the identification capability; and it no longer requires multi-level channel decoding and protocol processing, thereby simplifying the scheme process and reducing the process complexity; it can support the realization of fast, multi-category fine identification of downlink physical resource elements; and it can effectively solve at least one of the problems of limited identification capability, low accuracy and complex process in existing resource element classification schemes.

[0087] The construction of the first time-frequency resource matrix corresponding to the received 5G time-domain signal includes: performing downlink synchronization processing on the received 5G time-domain signal to obtain a first time-domain signal; performing orthogonal frequency division multiplexing (OFDM) demodulation on the first time-domain signal to obtain a second time-frequency resource matrix; and performing a filling operation on the second time-frequency resource matrix according to the data input size of the ResNet-Unet network to obtain the first time-frequency resource matrix.

[0088] In this embodiment of the application, the step of using the ResNet-Unet network to obtain the category of the physical resource element RE corresponding to the first time-frequency resource matrix includes: using the ResNet-Unet network to obtain the category probability distribution of the physical RE corresponding to the first time-frequency resource matrix; and determining the category of the physical RE corresponding to the first time-frequency resource matrix based on the category probability distribution.

[0089] Furthermore, the processor is also configured to: acquire model training signals and RE category labels corresponding to the model training signals, wherein the model training signals include: simulation signals and acquired 5G time-domain signals; and train a semantic segmentation network using the model training signals and RE category labels to obtain the ResNet-Unet network.

[0090] The step of training a semantic segmentation network using the model training signal and RE class labels to obtain the ResNet-Unet network includes: training a semantic segmentation network using a first loss function, the model training signal, and RE class labels to obtain the ResNet-Unet network; the first loss function includes: the Dice loss function and a weight-based focus loss function.

[0091] In this embodiment of the application, the weights in the focus loss function are the weights corresponding to the true categories of physical REs. The weights are determined based on the median frequency, which is determined based on the frequency of RE categories.

[0092] In this embodiment of the application, obtaining the RE category label corresponding to the model training signal includes: determining the RE category label corresponding to the simulation signal based on the signal generation parameters of the simulation signal; and / or, determining the RE category label corresponding to the collected 5G time-domain signal based on the physical broadcast channel (PBCH) decoding information corresponding to the collected 5G time-domain signal and first information, wherein the first information includes the RE position and the number of REs occupied by each type of signal in the power time-frequency diagram, and the power time-frequency diagram corresponds to the collected 5G time-domain signal.

[0093] The implementation embodiments of the above-mentioned resource element classification method are all applicable to the embodiments of the resource element classification device, and can achieve the same technical effect.

[0094] This application embodiment also provides a resource element classification device, including a memory, a processor, and a program stored in the memory and executable on the processor; when the processor executes the program, it implements the above-described resource element classification method.

[0095] The implementation embodiments of the above-mentioned resource element classification method are all applicable to the embodiments of the resource element classification device, and can achieve the same technical effect.

[0096] This application also provides a readable storage medium storing a program that, when executed by a processor, implements the steps in the resource element classification method described above.

[0097] The implementation embodiments of the above-described resource element classification method are all applicable to the embodiments of the readable storage medium and can achieve the same technical effect.

[0098] This application also provides a computer program product, including computer instructions. When the computer instructions are executed by a processor, they implement the various processes of the above-described resource element classification method and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0099] It should be noted that many of the functional components described in this specification are referred to as modules in order to more specifically emphasize the independence of their implementation.

[0100] In this embodiment, the module can be implemented in software so that it can be executed by various types of processors. For example, an identified executable code module may include one or more physical or logical blocks of computer instructions, which may be constructed as objects, procedures, or functions. Nevertheless, the executable code of the identified module does not need to be physically located together, but may include different instructions stored in different bits, which, when logically combined, constitute the module and achieve the module's intended purpose.

[0101] In practice, an executable code module can be a single instruction or many instructions, and can even be distributed across multiple different code segments, different programs, and across multiple memory devices. Similarly, operational data can be identified within the module and can be implemented in any suitable form and organized within any suitable data structure. This operational data can be collected as a single dataset or distributed across different locations (including different storage devices), and can exist, at least in part, solely as electronic signals within the system or network.

[0102] When a module can be implemented using software, considering the current level of hardware technology, modules that can be implemented in software can be implemented using hardware circuits by those skilled in the art to achieve the corresponding functions, without considering cost. These hardware circuits include conventional very-large-scale integrated circuits (VLSI) or gate arrays, as well as existing semiconductors such as logic chips and transistors, or other discrete components. Modules can also be implemented using programmable hardware devices, such as field-programmable gate arrays, programmable array logic, and programmable logic devices.

[0103] The above describes the preferred embodiments of this application. It should be noted that those skilled in the art can make several improvements and modifications without departing from the principles described in this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for classifying resource elements, characterized in that, include: Construct the first time-frequency resource matrix corresponding to the received 5G time-domain signal; Using the ResNet-Unet network, the category of the physical resource element RE corresponding to the first time-frequency resource matrix is ​​obtained; The ResNet-Unet network represents a network based on a residual network and an encoder-decoder structure; the category includes at least one of the following: background noise RE, synchronization signal block SSB RE, common physical downlink control channel PDCCH RE, common physical downlink shared channel PDSCH RE, dedicated PDCCH RE, dedicated PDSCH RE, and channel state information reference signal CSI-RS RE.

2. The resource element classification method according to claim 1, characterized in that, The construction of the first time-frequency resource matrix corresponding to the received 5G time-domain signal includes: Downlink synchronization processing is performed on the received 5G time-domain signal to obtain the first time-domain signal; The first time-domain signal is demodulated using Orthogonal Frequency Division Multiplexing (OFDM) to obtain the second time-frequency resource matrix. The second time-frequency resource matrix is ​​filled according to the data input size of the ResNet-Unet network to obtain the first time-frequency resource matrix.

3. The resource element classification method according to claim 1, characterized in that, The step of using the ResNet-Unet network to obtain the category of the physical resource element (RE) corresponding to the first time-frequency resource matrix includes: Using the ResNet-Unet network, the category probability distribution of physical REs corresponding to the first time-frequency resource matrix is ​​obtained; Based on the category probability distribution, the category of the physical RE corresponding to the first time-frequency resource matrix is ​​determined.

4. The resource element classification method according to claim 1, characterized in that, Also includes: Obtain the model training signal and the RE category label corresponding to the model training signal, wherein the model training signal includes: simulation signal and acquired 5G time domain signal; Using the model training signal and RE category labels, a semantic segmentation network is trained to obtain the ResNet-Unet network.

5. The resource element classification method according to claim 4, characterized in that, The step of training a semantic segmentation network using the model training signal and RE category labels to obtain the ResNet-Unet network includes: The semantic segmentation network is trained using the first loss function, the model training signal, and the RE category labels to obtain the ResNet-Unet network; the first loss function includes: the Dice loss function and the weight-based focus loss function.

6. The resource element classification method according to claim 5, characterized in that, The weights in the focus loss function are the weights corresponding to the true categories of physical REs. These weights are determined based on the median frequency, which is determined based on the frequency of RE categories.

7. The resource element classification method according to claim 4, characterized in that, Obtain the RE category labels corresponding to the model training signals, including: Based on the signal generation parameters of the simulated signal, determine the RE category label corresponding to the simulated signal; And / or, based on the physical broadcast channel PBCH decoding information corresponding to the acquired 5G time-domain signal and the first information, determine the RE category label corresponding to the acquired 5G time-domain signal. The first information includes the RE position and the number of REs occupied by each type of signal in the power time-frequency diagram, and the power time-frequency diagram corresponds to the acquired 5G time-domain signal.

8. A resource element classification device, characterized in that, include: The first construction module is used to construct the first time-frequency resource matrix corresponding to the received 5G time-domain signal; The first acquisition module is used to acquire the category of physical resource element RE corresponding to the first time-frequency resource matrix using the ResNet-Unet network; The ResNet-Unet network represents a network based on a residual network and an encoder-decoder structure; the category includes at least one of the following: background noise RE, synchronization signal block SSB RE, common physical downlink control channel PDCCH RE, common physical downlink shared channel PDSCH RE, dedicated PDCCH RE, dedicated PDSCH RE, and channel state information reference signal CSI-RS RE.

9. A resource element classification device, characterized in that, include: processor; The processor is used to construct a first time-frequency resource matrix corresponding to the received 5G time-domain signal; Using the ResNet-Unet network, the category of the physical resource element RE corresponding to the first time-frequency resource matrix is ​​obtained; The ResNet-Unet network represents a network based on a residual network and an encoder-decoder structure; the category includes at least one of the following: background noise RE, synchronization signal block SSB RE, common physical downlink control channel PDCCH RE, common physical downlink shared channel PDSCH RE, dedicated PDCCH RE, dedicated PDSCH RE, and channel state information reference signal CSI-RS RE.

10. A resource element classification device, comprising a memory, a processor, and a program stored in the memory and executable on the processor; characterized in that, When the processor executes the program, it implements the resource element classification method as described in any one of claims 1 to 7.

11. A readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the resource element classification method as described in any one of claims 1 to 7.

12. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the steps of the resource element classification method as described in any one of claims 1 to 7.

Images