WO2024012185A1 - 信号检测方法及设备、存储介质 - Google Patents

信号检测方法及设备、存储介质 Download PDF

Info

Publication number
WO2024012185A1
WO2024012185A1 PCT/CN2023/102432 CN2023102432W WO2024012185A1 WO 2024012185 A1 WO2024012185 A1 WO 2024012185A1 CN 2023102432 W CN2023102432 W CN 2023102432W WO 2024012185 A1 WO2024012185 A1 WO 2024012185A1
Authority
WO
WIPO (PCT)
Prior art keywords
signal
historical
neural network
real
detection method
Prior art date
Application number
PCT/CN2023/102432
Other languages
English (en)
French (fr)
Inventor
徐晓景
林伟
席志成
芮华
Original Assignee
中兴通讯股份有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 中兴通讯股份有限公司 filed Critical 中兴通讯股份有限公司
Publication of WO2024012185A1 publication Critical patent/WO2024012185A1/zh

Links

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L1/00—Arrangements for detecting or preventing errors in the information received
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L1/00—Arrangements for detecting or preventing errors in the information received
    • H04L1/004—Arrangements for detecting or preventing errors in the information received by using forward error control
    • H04L1/0045—Arrangements at the receiver end
    • H04L1/0047—Decoding adapted to other signal detection operation
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L1/00—Arrangements for detecting or preventing errors in the information received
    • H04L1/20—Arrangements for detecting or preventing errors in the information received using signal quality detector
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D30/00—Reducing energy consumption in communication networks
    • Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks

Definitions

  • the present application relates to the field of communication technology, in particular to a signal detection method, signal detection equipment, computer storage media and computer program products.
  • MIMO Multiple Input Multiple Output
  • OFDM Orthogonal Frequency Division Multiplexing
  • MIMO space division multiplexing technology can make it easy for the system to obtain space diversity gain and channel capacity gain while the system bandwidth and transmission bandwidth remain unchanged.
  • OFDM technology uses multiple orthogonal subcarriers to transmit data in parallel, making each path The data rate is greatly reduced, and a time guard interval is added, so it has strong resistance to multipath interference and frequency selective fading.
  • Signal detection is one of the key technologies of OFDM systems and MIMO systems.
  • the signal detection methods of related technologies mainly include three aspects: linear detection, nonlinear detection and optimal detection. Although they can all realize the signal detection process, the detection performance is low. , cannot meet the growing demand for signal detection; for example, optimal detection algorithms such as maximum likelihood detection algorithms are highly complex because they require a global search of all possible transmitted symbol domains for the received signal; linear detection algorithms include forced Zero algorithm and minimum mean square error detection algorithm, etc., although the computational complexity is not very high, the detection accuracy is not high; nonlinear detection algorithms include spherical decoding algorithm, continuous interference elimination algorithm, etc., which improve detection by increasing detection complexity. Accuracy.
  • Embodiments of the present application provide a signal detection method, signal detection equipment, computer storage media and computer program products, which can improve signal detection performance.
  • inventions of the present application provide a signal detection method.
  • the signal detection method includes:
  • the historical signal samples include historical channel estimates, historical received signals, and historical modulation signals
  • the current channel estimate value and the received signal to be measured are obtained, the current channel estimate value and the received signal to be measured are input into a pre-trained preset neural network for signal detection to obtain the received signal to be measured. signal estimate;
  • the pre-trained preset neural network is obtained by training the preset neural network through at least one set of historical signal samples.
  • embodiments of the present application also provide a signal detection device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the As before The signal detection method.
  • embodiments of the present application also provide a computer-readable storage medium that stores computer-executable instructions, and the computer-executable instructions are used to execute the signal detection method as described above.
  • embodiments of the present application also provide a computer program product.
  • the computer program or the computer instructions are stored in a computer-readable storage medium.
  • the processor of the computer device reads the computer program from the computer-readable storage medium.
  • the computer program or the computer instructions the processor executes the computer program or the computer instructions, so that the computer device performs the signal detection method as described above.
  • the preset neural network is trained to obtain a neural network model that meets the requirements, thereby utilizing the pre-trained neural network
  • the network model replaces the detection model in related technologies for signal detection. Due to the dual-driven approach of the data model, that is, the channel estimation value and the received signal are simultaneously input into the preset neural network for training or detection, which not only makes the preset neural network more efficient It is easy to learn channel correlation features, thereby improving signal detection performance, and it can also significantly reduce the amount of information required for training or detection. It has good environmental adaptability and generalization, so that it can meet the growing needs of signal detection and can make up for Technical gaps in related methods.
  • Figure 1 is a flow chart of a signal detection method provided by an embodiment of the present application.
  • Figure 2 is a flow chart for obtaining at least one set of historical signal samples in a signal detection method provided by an embodiment of the present application
  • Figure 3 is a flow chart before inputting the current channel estimate value and the received signal to be measured into the pre-trained preset neural network for signal detection in the signal detection method provided by one embodiment of the present application;
  • Figure 4 is a flow chart of training a preset neural network to obtain a pretrained preset neural network in the signal detection method provided by an embodiment of the present application;
  • Figure 5 is a flow chart for obtaining the first real number array as a training sample in the signal detection method provided by an embodiment of the present application
  • Figure 6 is a flow chart of obtaining the second real number array as the label of the training sample in the signal detection method provided by one embodiment of the present application;
  • Figure 7 is a flow chart after obtaining the signal estimate value of the historical received signal in the signal detection method provided by an embodiment of the present application.
  • Figure 8 is a flow chart for obtaining the signal estimate value of the received signal to be measured in the signal detection method provided by an embodiment of the present application;
  • Figure 9 is a flow chart for obtaining the third real number array in the signal detection method provided by another embodiment of the present application.
  • Figure 10 is a flow chart after obtaining the real signal estimation result of the received signal to be measured in the signal detection method provided by an embodiment of the present application;
  • Figure 11 is a flow chart for obtaining the complex signal estimation result of the received signal to be measured in the signal detection method provided by an embodiment of the present application;
  • Figure 12 is an execution flow chart of a signal detection method provided by an embodiment of the present application.
  • Figure 13 is a schematic diagram of the principle of a MIMO system based on a preset neural network provided by an embodiment of the present application;
  • Figure 14 is a flow chart before inputting the current channel estimate value and the received signal to be measured into the pre-trained preset neural network for signal detection in the signal detection method provided by another embodiment of the present application;
  • Figure 15 is a flow chart for obtaining the signal estimate value of the received signal to be measured in the signal detection method provided by another embodiment of the present application.
  • Figure 16 is a schematic diagram of the principle of an OFDM system based on a preset neural network provided by an embodiment of the present application;
  • Figure 17 is a schematic diagram of a signal detection device provided by an embodiment of the present application.
  • AI artificial intelligence
  • Deep learning Deep learning
  • the basic idea is to greatly improve the performance of wireless communication systems through the organic integration of wireless communication technology and AI technology.
  • Physical layer AI design includes two mainstream methods: one is end-to-end communication link design based on AI technology, and the other is communication module algorithm design based on AI technology.
  • the communication module algorithm design idea based on AI technology is mainly based on data-driven, that is, one or several communication modules are regarded as an unknown black box and replaced by a deep learning network.
  • This deep learning network mainly relies on massive data training to obtain , the training burden is heavy, time-consuming and labor-intensive, and the application scenario is relatively single, the applicability is not strong, and it cannot meet the current signal detection needs.
  • the signal detection method of one embodiment includes: acquiring at least a set of historical signal samples, where the historical signal samples include historical channel estimates, historical received signals and historical modulation signals; after obtaining the current channel estimate and the received signal to be measured In case of It is obtained by training a preset neural network with a set of historical signal samples.
  • the preset neural network is trained to obtain a neural network model that meets the requirements, thereby utilizing the pre-trained neural network The model replaces the detection model in related technologies for signal detection.
  • Figure 1 is a flow chart of a signal detection method provided by an embodiment of the present application.
  • the signal detection method may include but is not limited to steps S110 to S120.
  • Step S110 Obtain at least one set of historical signal samples, where the historical signal samples include historical channel estimates, historical received signals, and historical modulation signals;
  • Step S120 After obtaining the current channel estimate value and the received signal to be measured, input the current channel estimate value and the received signal to be measured into the pre-trained preset neural network for signal detection, and obtain the signal estimate of the received signal to be measured. value; Wherein, the pre-trained preset neural network is obtained by training the preset neural network through at least one set of historical signal samples.
  • the preset neural network is trained to obtain a neural network model that meets the requirements, thereby using the pre-trained neural network model
  • the data model dual-drive method is adopted, that is, the channel estimation value and the received signal are simultaneously input into the preset neural network for training or detection, which not only makes the preset neural network easier to learn Channel correlation characteristics can improve signal detection performance, and can also significantly reduce the amount of information required for training or detection. It has good environmental adaptability and generalization, thereby meeting the growing needs for signal detection and making up for related methods. technology gaps.
  • the embodiment of the present application is based on data model dual-driver, that is, based on the original technology of the wireless communication system, without changing the structure of the wireless communication system, using deep
  • the learning network replaces the relevant detection module to improve model performance; compared with the data-driven deep learning network that mainly relies on massive data, the data model dual-driven deep learning network in the embodiment of the present application only relies on the communication model or algorithm model. Based on the existing models of the physical layer, it can significantly reduce the amount of information required for training, upgrading or detection, has good environmental adaptability and generalization, and has broad development prospects.
  • the embodiments of the present application can be, but are not limited to, applied in OFDM technology or MIMO technology, for example, in an OFDM receiver under MIMO technology. That is, the preset neural network in the embodiment of the present application can be integrated into In the OFDM receiver, the preset neural network can be trained in the OFDM receiver, and after receiving the relevant information, the signal is detected based on the pretrained preset neural network.
  • the basic principle is consistent with the embodiments of the present application. , will not be described again; it is understood that those skilled in the art can also choose to apply the signal detection method of the embodiment of the present application according to specific application scenarios, which is not limited here. For example, it does not rule out that as the network system further develops, The development has resulted in multiple application scenarios that can be adapted to the signal detection method of the embodiment of the present application.
  • the preset neural network may, but is not limited to, include at least one of the following:
  • the example of the preset neural network given above can be offline training and online deployment, or it can be unified online deployment after completing offline training, or it can be unified online training and deployment. No limitation is made here; in addition to the examples of the preset neural network given above, those skilled in the art can also consider setting up a suitable preset neural network based on the characteristics, needs and other factors of specific application scenarios, which will not be described again here.
  • historical channel estimates and historical received signals are distinguished from current channel estimates and received signals to be measured, that is, the historical channel estimates and historical received signals are used to train a preset neural network, and the current channel estimate The value and the received signal to be tested are the signal parameters to be detected by the preset neural network.
  • the current channel estimate value and the received signal to be measured may also become the new historical channel estimate value and historical received signal.
  • the training of the preset neural network can be dynamic and have a certain degree of migration in time, so that more signal samples can be used to train the preset neural network to achieve better training results.
  • the training process of the preset neural network will be described in detail step by step.
  • the historical channel estimate and the historical received signal can be input into the preset as a set of detection data. Assume God After re-testing in the network, the actual training effect of the preset neural network is judged based on the test results, which is helpful to further determine whether errors may occur in the overall training process of the preset neural network.
  • the historical channel estimate value, historical received signal, current channel estimate value, and received signal to be measured can be obtained by sending a group of signals in sequence, or can be obtained by sending corresponding signals separately. There is no limit here.
  • the specific number of historical signal samples can be set according to specific application scenarios, that is, the training samples for the preset neural network can be set according to specific application scenarios, which is not limited here.
  • step S110 may include, but is not limited to, steps S111 to S113.
  • Step S111 Obtain the historical modulation signal sent by the sending end
  • Step S112 In each wireless fading channel, perform preset signal processing on the historical modulation signal to obtain the received pilot signal, local pilot signal and historical received signal at the receiving end;
  • Step S113 Multiply the received pilot signal by the conjugate of the local pilot signal to obtain a historical channel estimate.
  • each wireless fading channel corresponds to multiple groups of historical signal samples
  • the historical modulated signal undergoes preset signal processing to obtain the received pilot signal, the local pilot signal and the historical received signal at the receiving end, and then the historical channel estimate is obtained by multiplying the received pilot signal by the conjugate of the local pilot signal. , that is to say, due to the change of the channel, the amplitude of the received signal changes randomly and causes signal fading. Therefore, for each wireless fading channel, at least a set of complete and different historical signal samples can be obtained correspondingly, so that a total of multiple Historical signal samples are used for preset neural network training, and the training effect can be greatly guaranteed.
  • the sending end and the receiving end are a set of relative concepts, that is, the sending end corresponds to the initial source of the signal, and the receiving end corresponds to the back-end result of the signal.
  • the specific presentation forms of the sending end and the receiving end are not limited here. It is only used to clearly explain the working principle of the embodiment of the present application. In specific application scenarios, specific sending ends and receiving ends can be set, which is not limited here.
  • the historical modulation signal can be obtained by acquiring at least one set of transmission bit stream signals sent by the transmitting end, and then encoding and modulating the at least one set of transmission bit stream signals. That is, by performing coding on the initial transmission bit stream signal. Pre-processing to obtain historical modulation signals; alternatively, the transmitting end can pre-condition the signal to be processed into a historical modulation signal, so that the historical modulation signal can be obtained directly from the transmitting end.
  • the specific method of multiplying the received pilot signal by the conjugate of the local pilot signal to obtain the historical channel estimate can be multiple, and is not limited here.
  • the least squares LS channel estimation algorithm may be used to multiply the received pilot signal by the conjugate of the local pilot signal to obtain the historical channel estimate, or other traditional communication algorithms common in the field may be used, but is not limited to Implementation.
  • the least squares method is a mathematical model applied in the fields of error estimation, uncertainty, system identification and prediction, forecasting and other data processing fields. It is well known to those skilled in the art and will not be described in detail here. .
  • the preset signal processing method for historical modulation signals can be multiple, which is not limited here.
  • the first output signal can be obtained by performing fast Fourier transform on the historical modulation signal and adding a cyclic prefix, and then passing the first output signal through different wireless fading channels to obtain a different second output signal, and then The second output signal undergoes decyclic prefixing and inverse Fourier transformation to achieve signal restoration, and finally the received pilot signal, local pilot signal and historical received signal;
  • the historical modulated signal can be input into the preset network model, and the historical modulated signal is demodulated and equalized through the preset network model, and finally the received pilot signal, local Pilot signal and historical received signal.
  • step S120 when there are multiple groups of historical signal samples, step S120 also includes but is not limited to step S130 before step S120.
  • Step S130 For each group of historical signal samples, use historical channel estimates and historical received signals as training samples, use historical modulation signals as labels of the training samples, and train the preset neural network to obtain a pretrained preset neural network.
  • the received signal is represented by Y
  • Y Data represents the historical received signal
  • Y Pilot represents the received pilot signal
  • the local pilot signal is X Pilot .
  • the LS channel estimate is Together with Y Data , input it into the preset neural network for training to obtain the final estimated value of the sent data.
  • Step S130 may include but is not limited to steps S131 to S133.
  • Step S131 Combine the obtained historical channel estimation value and the real part and imaginary part of the historical received signal to obtain the first real number array as the training sample;
  • Step S132 Combine the real part and imaginary part of the acquired historical modulation signal to obtain a second real number array as the label of the training sample;
  • Step S133 Input the first real number array and the second real number array to the preset neural network for training, and obtain the signal estimate value of the historical received signal.
  • the first real number array as a training sample is obtained by combining the historical channel estimate value, the real part and the imaginary part of the historical received signal, and the real part and the imaginary part of the historical modulation signal are obtained and combined to obtain the training sample.
  • the second real number array of the label that is, the input form of the historical channel estimate value, the historical received signal and the historical modulation signal is uniformly set to a format that meets the input requirements of the preset neural network, so that the preset neural network can use the real number array according to the real number array.
  • Training is performed to obtain the signal estimate value of the historical received signal, and in specific application scenarios, the corresponding real part and imaginary part values can respectively represent the values of the antenna port and the receiving antenna, that is, the form of a real number array can better adapt to Neural network operations.
  • the default neural network in addition to the above-mentioned real number neural network, can also be of other types.
  • the input to the default neural network can change as the type of the default neural network changes, that is, That is to say, steps S131 to S133 in the above embodiment are not used to limit the content input to the default neural network, and can also be determined according to specific requirements. Select the appropriate training input content for the scenario, which is not limited here.
  • Step S131 may include but is not limited to steps S1311 to S1312.
  • Step S1311 Obtain the first real part and the first imaginary part of the historical channel estimate value, and obtain the second real part and the second imaginary part of the historical received signal;
  • Step S1312 Arrange and combine the first real part, the first imaginary part, the second real part and the second imaginary part in a preset order to obtain a first real number array as a training sample.
  • the corresponding historical channel estimate value and the historical received signal can be generated respectively. real number set, and then arrange and combine the corresponding first real part, first imaginary part, second real part and second imaginary part in each real number set according to the preset order, then we can obtain the historical channel estimation value and the historical received signal. First array of real numbers.
  • the preset order of arrangement is not limited, and may be but is not limited to being arranged according to the real part and imaginary part corresponding to the historical channel estimate value and the historical received signal, that is, the real part and imaginary part of the historical channel estimate value Arrange them in adjacent positions, and arrange the real and imaginary parts of the historical received signals in adjacent positions, or they can be arranged sequentially according to specific application scenarios, etc.
  • Step S132 when the historical modulation signal includes a modulation signal stream, and the modulation signal stream includes multiple layer modulation signals, step S132 is further described.
  • Step S132 may include but is not limited to steps S1321 to S1322.
  • Step S1321 Obtain the real part and imaginary part of each layer modulation signal
  • Step S1322 Arrange and combine the real parts and imaginary parts of the modulation signals of each layer in a preset order to obtain a second real number array as the label of the training sample.
  • the historical modulated signal when the historical modulated signal includes a modulated signal stream, and the modulated signal stream includes multiple layers of modulated signals, it means that the historical modulated signal at this time is equivalent to a signal layer.
  • the real part and imaginary part of each layer of modulated signals part, which is equivalent to obtaining the entire real part and imaginary part of the historical modulation signal, and then the real part and imaginary part of the modulation signal of each layer can be arranged and combined in a preset order to obtain the second real number array as the label of the training sample.
  • the preset order of arrangement is not limited, and may be but is not limited to being arranged according to the real part and imaginary part of each layer modulation signal, that is, the real part and imaginary part of each layer modulation signal are arranged adjacent to each other. position, while arranging the real and imaginary parts of adjacent layer modulation signals in adjacent positions, etc., or they can be arranged sequentially according to specific application scenarios.
  • step S133 also includes but is not limited to steps S134 to S135, where steps S134 to S135 are used to determine the pre-trained preset neural network.
  • Step S134 Obtain the current value of the preset loss function corresponding to the preset neural network based on the signal estimate value of the historical received signal and the second real number array;
  • Step S135 When the current value of the preset loss function meets the preset convergence condition, it is determined that the pretrained preset neural network is obtained.
  • the loss function is calculated through the signal estimate value obtained by training and the predetermined second real number array, and based on the calculated value, it is judged whether it meets the preset convergence condition, so that when it is determined that the calculated value does not meet the preset convergence condition , realize further training and optimization of the preset neural network to obtain a new signal estimate, and so on, and then calculate a new loss function value through the new signal estimate and make a judgment again, so that the preset neural network can be continuously optimized.
  • the final signal estimate value obtained by training continuously approaches the transmitted signal and because it can flexibly support data of different lengths According to the bit source and different modulation methods, it can be more reliably applied in scenarios related to calculating loss functions, and finally a preset neural network that meets the requirements can be trained.
  • the preset loss function includes at least one of the following:
  • MSE Mean Square Error
  • RMSE Root Mean Squared Error
  • MAE Mean Absolute Error
  • the preset neural network is a real number neural network, so it needs to be Separate it from the real part and imaginary part of Y Data to form a set of real number arrays, such as: Among them, Re( ⁇ ) and Im( ⁇ ) are the operations of taking the real part and taking the imaginary part respectively. Because in MIMO systems There are P antenna ports and R receiving antenna values, and Y Data has R receiving antenna values, so the real array obtained previously also contains the real and imaginary parts of the P antenna ports and R receiving antennas. Then this real number array is used as the training input of the default neural network.
  • the transmission bit stream sent by the transmitter is encoded and modulated to generate a transmission modulation signal X Data (i.e., historical modulation signal) containing multi-layer MIMO signals.
  • X Data i.e., historical modulation signal
  • K T ⁇ L layer
  • L layer is the number of layers of MIMO signal
  • T is the number of resource elements (Resource Element, RE) occupied by each layer modulation signal
  • this one-dimensional real number array X' Data is used as a training sample label
  • Step S120 includes but is not limited to steps S121 to S122.
  • Step S121 Combine the obtained current channel estimate value and the real part and imaginary part of the received signal to be measured to obtain a third real number array
  • Step S122 Input the third real number array to the pre-trained preset neural network for signal detection, and obtain the real number signal estimation result of the received signal to be tested.
  • the third real number array as the detection input is obtained by combining the current channel estimate value and the real part and imaginary part of the received signal to be measured, thereby providing the preset neural network with corresponding input parameters that meet the requirements, that is to say , uniformly set the input form of the current channel estimate value and the received signal to be measured to a format that meets the input requirements of the preset neural network, so that the preset neural network can detect the real number signal estimation result of the received signal to be measured based on the real number array.
  • the default neural network in addition to the above output real signal estimation results, can also output other types.
  • the input to the default neural network can change with the output type of the default neural network. Changes are not limited here.
  • Step S121 includes but is not limited to steps S1211 to S1212.
  • Step S1211 Obtain the third real part and the third imaginary part of the current channel estimate value, and obtain the fourth real part and the fourth imaginary part of the received signal to be measured;
  • Step S1212 Arrange and combine the third real part, the third imaginary part, the fourth real part and the fourth imaginary part in a preset order to obtain a third real number array.
  • the current channel estimation value and the reception signal to be tested can be generated respectively.
  • the set of real numbers corresponding to the signal, and then the corresponding third real part, third imaginary part, fourth real part and fourth imaginary part in each set of real numbers are arranged and combined in a preset order to obtain the current channel estimate and the value to be measured.
  • the preset order of arrangement is not limited, and may be but is not limited to being arranged according to the real part and imaginary part corresponding to the current channel estimate value and the received signal to be measured, that is, the real part and imaginary part of the current channel estimate value are arranged respectively. parts are arranged in adjacent positions, and the real and imaginary parts of the received signal to be measured are arranged in adjacent positions, or they can be arranged sequentially according to specific application scenarios, etc.
  • steps after step S122 further explains the steps after step S122.
  • the steps after step S122 also include but are not limited to steps S123 to S124.
  • Step S123 Perform complex conversion processing on the real signal estimation result of the received signal to be tested, to obtain the complex signal estimation result of the received signal to be measured;
  • Step S124 Demodulate and decode the complex signal estimation result to obtain an output signal.
  • the complex signal estimation result of the received signal to be measured is obtained from the real signal estimation result of the received signal to be measured, so that the complex signal estimation result can be further demodulated and decoded to obtain the final output signal.
  • step S123 may include, but is not limited to, steps S1231 to S1232.
  • Step S1231 Perform complex conversion on the real part and imaginary part of each layer signal in the real signal estimation array to obtain the complex conversion data corresponding to each layer signal;
  • Step S1232 Arrange and combine each complex conversion data in a preset order to obtain a complex signal estimate of the received signal to be measured. count results.
  • the real number signal estimation result includes a real number signal estimation array
  • the real number signal estimation array includes the real parts and imaginary parts of multiple layer signals arranged in sequence
  • the values of each layer signal in the real number signal estimation array can be calculated separately.
  • the real part and the imaginary part are complex converted to obtain the complex converted data corresponding to each layer signal. Then, by arranging and combining each complex converted data, the complex signal estimation result of the received signal to be measured can be obtained accurately and reliably.
  • the preset neural network is used to implement online signal detection. After the preset neural network training is completed, it can be deployed at the receiving end in the MIMO system for online signal detection.
  • complex( ⁇ ) represents the operation of composing a complex-valued data, using the converted complex-valued data Then perform subsequent demodulation and decoding to obtain the final output signal.
  • a MIMO signal detector and an LS channel estimation module based on a neural network are provided.
  • L 2 data layers
  • P 2 antenna ports
  • T 2 physical transmit antennas
  • R 2 physical receive antennas.
  • the frequency domain signal Y at the receiving end can be expressed as formula (2):
  • ;N is additive Gaussian white noise
  • the dimension is R ⁇ 1, that is, 2 ⁇ 1.
  • the dimensions of the frequency domain received signal Y Data and the received pilot signal Y Pilot are also R ⁇ 1, that is, 2 ⁇ 1, and the dimension of the local pilot signal X Pilot at the receiving end is P ⁇ 1, that is, 2 ⁇ 1.
  • the frequency domain received signal and the received pilot signal occupy a total of 2 OFDM symbols, then the received pilot signal occupies the first OFDM symbol, and the frequency domain received signal occupies the second OFDM symbol.
  • the neural network-based MIMO signal detection method proposed in this example focuses on how to detect the transmit signal X from the frequency domain signal Y at the receiving end.
  • the specific steps are as follows:
  • Step 1 Determine the structure of the neural network-based MIMO signal detector. First input X Pilot and Y Pilot into the LS channel estimation sub-module, and use formula (1) in the above example 1 to obtain the LS channel estimation value Its dimensions are P ⁇ R, which is 2 ⁇ 2. Then Together with Y Data , it is input to the MIMO signal detector based on the neural network to obtain the final estimated value of the transmitted data. Among them, the LS channel estimation sub-module can be constructed by relying on methods in related technologies.
  • the MIMO signal detector based on neural networks uses a convolutional neural network (Convolutional Neural Network, CNN).
  • Step 2 Construct the training input of the convolutional neural network MIMO signal detector.
  • add It is separated from the real part and imaginary part of Y Data to form a set of real number arrays with a dimension of 120 ⁇ 12 and arranged in order.
  • Step 3 After generating the labels of the training samples of the MIMO signal detector, perform model training.
  • the transmitter generates the transmit modulation signal An array of real numbers serves as training samples.
  • K T ⁇ L layer
  • the dimension of this one-dimensional real number array X′ Data is 1 ⁇ 480, which is used as the label of the training sample. Under different fading channels, multiple sets of training samples and labels are generated and input into CNN for training to obtain the estimated value of the transmitted signal. Then it is used together with the label X′ Data to calculate the MSE loss function, as shown in the following formula (3):
  • n is the number of batch size samples for each training.
  • the Adam optimizer is used to train the CNN based on the MSE loss function. After the loss function converges to a certain extent, the network training is completed. At this time, the model parameters are saved for subsequent online signal detection.
  • Step 4 Implement online MIMO signal detection.
  • the CNN training After the CNN training is completed, it can be deployed at the receiving end in the OFDM system for online signal detection.
  • detecting online signals follow the same processing flow in steps 1 and 2.
  • the output of the CNN that is, the estimated value of the transmitted signal, can be obtained.
  • the estimated value of the transmitted signal includes The corresponding real number array, and then complete the conversion of the real number array to the balanced complex value data, as follows:
  • complex( ⁇ ) represents the operation of composing a complex-valued data, using the converted complex-valued data Then perform subsequent demodulation and decoding to obtain the final output signal.
  • the signal detection method of the embodiment of the present application outputs the data after equalization and delayer mapping, and uses the signal modulated by the sending end as the training label, and uses MSE as the loss function, so that the estimated value continuously approaches the sending value.
  • signal and can flexibly support data bit streams of different lengths and different modulation methods, and has good versatility.
  • step S120 when there are multiple groups of historical signal samples and the preset neural network includes an enhanced channel estimation subnetwork and a channel equalization subnetwork, step S120 also includes but is not limited to steps S140 to S150.
  • Step S140 For each group of historical signal samples, input the historical channel estimation value into the enhanced channel estimation sub-network for channel enhancement estimation training to obtain the first estimation result;
  • Step S150 Input the first estimation result and historical received signals into the channel equalization sub-network for channel equalization training to obtain a pre-trained preset neural network.
  • the preset neural network includes two sub-networks: enhanced channel estimation sub-network and channel equalization sub-network.
  • the sub-networks each have independent performance, so for each group of historical signal samples, the historical channel estimate values can be input into the enhanced channel estimation sub-network for channel enhancement estimation training to obtain the first estimate result, and then the first estimate
  • the results and historical received signals are input into the channel equalization sub-network for channel equalization training to obtain a pre-trained preset neural network.
  • the preset neural network set up with integrated functions it can achieve the same effective and reliable training effect.
  • the enhanced channel estimation sub-network and the channel equalization sub-network can be trained together, or can be trained separately, or those skilled in the art can also select corresponding training methods according to specific application scenarios to train the enhanced channel estimation sub-network.
  • the network and channel equalization sub-network are trained, etc., which are not limited here.
  • step S120 includes but is not limited to steps S125 to S126.
  • Step S125 Input the current channel estimation value into the pre-trained enhanced channel estimation sub-network to perform channel enhancement estimation and obtain the second estimation result;
  • Step S126 Input the second estimation result and the received signal to be tested into the pre-trained channel equalization sub-network to perform channel equalization, and obtain the signal estimate value of the received signal to be measured.
  • the second prediction result is directly obtained by inputting the current channel estimate value into the pre-trained enhanced channel estimation sub-network for channel enhancement estimation, and then the second prediction result is combined with the to-be-tested
  • the received signal is input into the pre-trained channel equalization sub-network for channel equalization, and the signal estimate value of the received signal to be measured can be obtained.
  • steps S140 to S150 and steps S125 to S126 are similar to the relevant steps in the previous embodiments.
  • the only difference is that the preset neural network in the previous embodiments has an integrated structure, while the one in this application The preset neural network is divided into two sub-networks. Through the cooperation of the two sub-networks, the functions of a single preset neural network can be completely realized. Since the detailed principles and processes of the preset neural network have been explained in detail in the previous embodiments, this article Steps of Embodiments Embodiments may refer to the detailed principles and processes of the preset neural network in the foregoing embodiments. To avoid redundancy, they will not be described again here.
  • a MIMO signal detector and an LS channel estimation module based on a neural network are provided.
  • L 4 data layers
  • P 4 antenna ports
  • T 4 physical transmit antennas
  • R 4 physical receive antennas.
  • the frequency domain signal Y at the receiving end can be expressed as formula (4):
  • ;N is additive Gaussian white noise
  • the dimension is R ⁇ 1, that is, 4 ⁇ 1.
  • the dimensions of the frequency domain received signal Y Data and the received pilot signal Y Pilot are also R ⁇ 1, that is, 4 ⁇ 1, and the dimension of the local pilot signal X Pilot at the receiving end is P ⁇ 1, that is, 4 ⁇ 1.
  • the frequency domain received signal and the received pilot signal occupy a total of 4 OFDM symbols
  • the received pilot signal occupies the first OFDM symbol
  • the frequency domain received signal occupies the remaining 3 OFDM symbols.
  • the neural network-based MIMO signal detection method proposed in this example focuses on how to detect the transmit signal X from the frequency domain signal Y at the receiving end.
  • the specific steps are as follows:
  • Step 1 Determine the structure of the neural network-based MIMO signal detector. First input X Pilot and Y Pilot into the LS channel estimation sub-module, and use formula (1) in the above example 1 to obtain the LS channel estimation value Its dimensions are P ⁇ R, which is 4 ⁇ 4. Then Together with Y Data , it is input to the MIMO signal detector based on the neural network to obtain the final estimated value of the transmitted data. Among them, the LS channel estimation sub-module can be constructed by relying on methods in related technologies.
  • the MIMO signal detector based on neural networks contains two sub-networks: one is the enhanced channel estimation sub-network, which is used to enhance channel estimation and remove noise and interference. The other is the channel equalization sub-network, which is used for channel equalization and detection of MIMO signals. Both sub-networks use residual neural networks with channel attention mechanisms.
  • Step 2 Construct the training input of the enhanced channel estimation sub-network.
  • This real number array H′ LS is used as the training input of the enhanced channel estimation subnetwork. H′ LS is input into the enhanced channel estimation subnetwork, and the output H′′ is obtained through the residual neural network, with a dimension of 240 ⁇ 32.
  • Y Data occupies 3 OFDM symbols, and the dimension of Y Data is 240*24. Then, this real number array with a dimension of 240 ⁇ 56 is used as the training input of the channel equalization subnetwork.
  • Step 3 After generating the labels of the training samples of the MIMO signal detector, perform model training.
  • the transmitter generates the transmit modulation signal Similar operations in step 2 obtain training samples for the enhanced channel estimation subnetwork and channel equalization subnetwork respectively.
  • K T ⁇ L layer
  • the dimension of this one-dimensional real number array X′ Data is 1 ⁇ 5760, which is used as the label of the training sample.
  • the transmitted signal estimate is obtained from the output of the channel equalization subnetwork. Then it is used together with the label X′ Data to calculate the Euclidean distance as the loss function, as shown in the following formula (5):
  • n is the number of batch size samples for each training.
  • the loss function based on Euclidean distance uses the Adam optimizer to train the residual neural network. After the loss function converges to a certain extent, the network training is determined to be completed. At this time, the model parameters are saved for subsequent online signal detection.
  • Step 4 Implement online MIMO signal detection. After the residual neural network training is completed, it can be deployed on the receiving end of the OFDM system for online signal detection. When detecting online signals, follow the same processing flow in steps 1 and 2. Similarly, first obtain the input parameters of the residual neural network and then perform online detection. The output of the residual neural network, that is, the estimated value of the transmitted signal, can be obtained. The estimated value of the sent signal includes the corresponding real number array, and then the conversion of the real number array to the equalized complex value data is completed, as follows:
  • complex( ⁇ ) represents the operation of composing a complex-valued data, using the converted complex-valued data Then perform subsequent demodulation and decoding to obtain the final output signal.
  • the signal detection method of the embodiment of the present application outputs the data after equalization and delayer mapping, and uses the signal modulated by the transmitter as the training label, and uses the Euclidean distance as the loss function, so that the estimated value is continuously approximated. While sending signals, it can flexibly support data bit streams of different lengths and different modulation methods, and has good versatility.
  • one embodiment of the present application also discloses a signal detection device 100, including: at least one processor 110; at least one memory 120 for storing at least one program; when at least one program is at least When executed, a processor 110 implements the signal detection method as in any previous embodiment.
  • an embodiment of the present application also discloses a computer-readable storage medium in which computer-executable instructions are stored, and the computer-executable instructions are used to execute the signal detection method as in any of the previous embodiments.
  • an embodiment of the present application also discloses a computer program product, which includes a computer program or computer instructions.
  • the computer program or computer instructions are stored in a computer-readable storage medium.
  • the processor of the computer device reads the computer program from the computer-readable storage medium.
  • the computer program or computer instructions are obtained, and the processor executes the computer program or computer instructions, so that the computer device performs the signal detection method as in any of the previous embodiments.
  • Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, Digital Versatile Disk (DVD) or other optical disk storage, magnetic cassettes, tapes, disk storage or other magnetic storage devices, or may Any other medium used to store the desired information and that can be accessed by a computer.
  • communication media typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media .

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Quality & Reliability (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

本申请公开了一种信号检测方法及设备、存储介质。其中，信号检测方法，包括：获取至少一组历史信号样本，其中，历史信号样本包括历史信道估计值、历史接收信号和历史调制信号；在获取到当前信道估计值和待测接收信号的情况下，将当前信道估计值和待测接收信号输入到预训练的预设神经网络进行信号检测，得到待测接收信号的信号估计值；其中，预训练的预设神经网络为通过至少一组历史信号样本对预设神经网络进行训练而得到。

Description

信号检测方法及设备、存储介质
相关申请的交叉引用
本申请基于申请号为202210831551.7、申请日为2022年07月15日的中国专利申请提出，并要求该中国专利申请的优先权，该中国专利申请的全部内容在此引入本申请作为参考。
技术领域
本申请涉及通信技术领域，尤其是一种信号检测方法、信号检测设备、计算机存储介质及计算机程序产品。
背景技术
多输入多输出(Multiple In Multiple Out，MIMO)空分复用技术和正交频分复用(Orthogonal Frequency Division Multiplexing，OFDM)技术是无线通信系统中极具竞争力的技术。MIMO空分复用技术可以使得系统在系统带宽和发射带宽不变的情况下容易获得空间分集增益和信道的容量增益，OFDM技术采用多个正交的子载波并行传输数据，使得每一路上的数据速率大大降低，并且加入了时间保护间隔，因此具有较强的抗多径干扰和频率选择性衰落的能力。
信号检测是OFDM系统、MIMO系统的关键技术之一，相关技术的信号检测方式主要包括线性检测、非线性检测和最优检测这三个方面，虽然均能够实现信号检测过程，但检测性能较低，无法满足日益增长的信号检测需求；例如，最优检测算法如最大似然检测算法，由于需要针对接收信号对所有可能的发送符号域进行全局搜索，因此复杂度较高；线性检测算法包括迫零算法和最小均方误差检测算法等，虽然计算复杂度不算很高，但检测精度并不高；非线性检测算法包括球形解码算法、连续干扰消除算法等，通过增加检测复杂度以提高检测精度。
发明内容
本申请实施例提供了一种信号检测方法、信号检测设备、计算机存储介质及计算机程序产品，能够提升信号检测性能。
第一方面，本申请实施例提供了一种信号检测方法，所述信号检测方法包括：
获取至少一组历史信号样本，其中，所述历史信号样本包括历史信道估计值、历史接收信号和历史调制信号；
在获取到当前信道估计值和待测接收信号的情况下，将所述当前信道估计值和所述待测接收信号输入到预训练的预设神经网络进行信号检测，得到所述待测接收信号的信号估计值；
其中，预训练的所述预设神经网络为通过至少一组所述历史信号样本对所述预设神经网络进行训练而得到。
第二方面，本申请实施例还提供了一种信号检测设备，包括：存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现如前面 所述的信号检测方法。
第三方面，本申请实施例还提供了一种计算机可读存储介质，存储有计算机可执行指令，所述计算机可执行指令用于执行如前面所述的信号检测方法。
第四方面，本申请实施例还提供了一种计算机程序产品，计算机程序或所述计算机指令存储在计算机可读存储介质中，计算机设备的处理器从所述计算机可读存储介质读取所述计算机程序或所述计算机指令，所述处理器执行所述计算机程序或所述计算机指令，使得所述计算机设备执行如前面所述的信号检测方法。
本申请实施例中，基于获取到的包括历史信道估计值、历史接收信号和历史调制信号的历史信号样本，对预设神经网络进行训练而得到符合要求的神经网络模型，从而利用预训练的神经网络模型替代相关技术中的检测模型进行信号检测，由于采用数据模型双驱动的方式，即同时输入信道估计值和接收信号到预设神经网络中进行训练或检测，不仅可以使得预设神经网络更容易学习信道关联特征，从而提升信号检测性能，而且还能够显著减少训练或检测所需的信息量，具有良好的环境自适应性和泛化性，从而能够满足日益增长的信号检测需求，可以弥补相关方法中的技术空白。
附图说明
图1是本申请一个实施例提供的信号检测方法的流程图；
图2是本申请一个实施例提供的信号检测方法中，获取至少一组历史信号样本的流程图；
图3是本申请一个实施例提供的信号检测方法中，将当前信道估计值和待测接收信号输入到预训练的预设神经网络进行信号检测之前的流程图；
图4是本申请一个实施例提供的信号检测方法中，对预设神经网络进行训练而得到预训练的预设神经网络的流程图；
图5是本申请一个实施例提供的信号检测方法中，得到作为训练样本的第一实数数组的流程图；
图6是本申请一个实施例提供的信号检测方法中，得到作为训练样本的标签的第二实数数组的流程图；
图7是本申请一个实施例提供的信号检测方法中，得到历史接收信号的信号估计值之后的流程图；
图8是本申请一个实施例提供的信号检测方法中，得到待测接收信号的信号估计值的流程图；
图9是本申请另一个实施例提供的信号检测方法中，得到第三实数数组的流程图；
图10是本申请一个实施例提供的信号检测方法中，得到待测接收信号的实数信号估计结果之后的流程图；
图11是本申请一个实施例提供的信号检测方法中，得到待测接收信号的复数信号估计结果的流程图；
图12是本申请一个实施例提供的信号检测方法的执行流程图；
图13为本申请一个实施例提供的基于预设神经网络的MIMO系统的原理示意图；
图14是本申请另一个实施例提供的信号检测方法中，将当前信道估计值和待测接收信号输入到预训练的预设神经网络进行信号检测之前的流程图；
图15为本申请另一个实施例提供的信号检测方法中，得到待测接收信号的信号估计值的流程图；
图16为本申请一个实施例提供的基于预设神经网络的OFDM系统的原理示意图；
图17是本申请一个实施例提供的信号检测设备的示意图。
具体实施方式
为了使本申请的目的、技术方法及优点更加清楚明白，以下结合附图及实施例，对本申请进行进一步详细说明。应当理解，此处所描述的具体实施例仅用以解释本申请，并不用于限定本申请。
需要说明的是，虽然在流程图中示出了逻辑顺序，但是在某些情况下，可以以不同于流程图中的顺序执行所示出或描述的步骤。说明书和权利要求书及上述附图中的术语“第一”、“第二”等是用于区别类似的对象，而不必用于描述特定的顺序或先后次序。
近年来，人工智能(Artificial Intelligence，AI)技术特别是深度学习在计算机视觉、自然语言处理、语音识别等多个领域获得了巨大成功。AI技术使能的智能无线通信被认为是未来6G发展主流方向之一，其基本思想是通过无线通信技术与AI技术的有机融合，大幅度提升无线通信系统的性能。物理层AI设计包含了两种主流方法：一种是基于AI技术的端到端通信链路设计，一种是基于AI技术的通信模块算法设计。目前，基于AI技术的通信模块算法设计思路主要是基于数据驱动，即将某个或某几个通信模块看成一个未知的黑盒子，利用深度学习网络取而代之，该深度学习网络主要依赖海量数据训练得到，训练负担很重，较为耗时耗力，且应用场景较为单一，适用性不强，无法满足当前的信号检测需求。
基于此，本申请提供了一种信号检测方法、信号检测设备、计算机存储介质及计算机程序产品。其中一个实施例的信号检测方法，包括：获取至少一组历史信号样本，其中，历史信号样本包括历史信道估计值、历史接收信号和历史调制信号；在获取到当前信道估计值和待测接收信号的情况下，将当前信道估计值和待测接收信号输入到预训练的预设神经网络进行信号检测，得到待测接收信号的信号估计值；其中，预训练的预设神经网络为通过至少一组历史信号样本对预设神经网络进行训练而得到。该实施例中，基于获取到的包括历史信道估计值、历史接收信号和历史调制信号的历史信号样本，对预设神经网络进行训练而得到符合要求的神经网络模型，从而利用预训练的神经网络模型替代相关技术中的检测模型进行信号检测，由于采用数据模型双驱动的方式，即同时输入信道估计值和接收信号到预设神经网络中进行训练或检测，不仅可以使得预设神经网络更容易学习信道关联特征，从而提升信号检测性能，而且还能够显著减少训练或检测所需的信息量，具有良好的环境自适应性和泛化性，从而能够满足日益增长的信号检测需求，可以弥补相关方法中的技术空白。
下面结合附图，对本申请实施例作进一步阐述。
如图1所示，图1是本申请一个实施例提供的信号检测方法的流程图，该信号检测方法可以包括但不限于步骤S110至S120。
步骤S110：获取至少一组历史信号样本，其中，历史信号样本包括历史信道估计值、历史接收信号和历史调制信号；
步骤S120：在获取到当前信道估计值和待测接收信号的情况下，将当前信道估计值和待测接收信号输入到预训练的预设神经网络进行信号检测，得到待测接收信号的信号估计值； 其中，预训练的预设神经网络为通过至少一组历史信号样本对预设神经网络进行训练而得到。
本步骤中，基于获取到的包括历史信道估计值、历史接收信号和历史调制信号的历史信号样本，对预设神经网络进行训练而得到符合要求的神经网络模型，从而利用预训练的神经网络模型替代相关技术中的检测模型进行信号检测，由于采用数据模型双驱动的方式，即同时输入信道估计值和接收信号到预设神经网络中进行训练或检测，不仅可以使得预设神经网络更容易学习信道关联特征，从而提升信号检测性能，而且还能够显著减少训练或检测所需的信息量，具有良好的环境自适应性和泛化性，从而能够满足日益增长的信号检测需求，可以弥补相关方法中的技术空白。
可以看出，相比于相关技术的基于数据驱动的深度学习网络，本申请实施例基于数据模型双驱动，即是在无线通信系统原有技术基础上，不改变无线通信系统的结构，利用深度学习网络代替相关检测模块以提升模型性能；相比于基于数据驱动的深度学习网络主要依赖海量数据，本申请实施例中的基于数据模型双驱动的深度学习网络只需依靠通信模型或者算法模型，以物理层已有模型为基础，可以显著减少训练、升级或检测所需的信息量，具有良好的环境自适应性和泛化性，具备广阔的发展前景。
在一实施例中，本申请实施例可以但不限于应用在OFDM技术或MIMO技术中，例如应用在MIMO技术下的OFDM接收机中，即可以将本申请实施例中的预设神经网络集成于OFDM接收机中，可以在OFDM接收机中对该预设神经网络进行训练，并在接收到相关信息之后，基于预训练的预设神经网络对信号进行检测，其基本原理与本申请实施例一致，不再赘述；可以理解地是，本领域技术人员也可以根据具体应用场景来选择应用本申请实施例的信号检测方法的方式，此处并未限定，例如，不排除随着网络系统的进一步发展而衍生出多个可以适配于本申请实施例的信号检测方法的应用场景等。
在一实施例中，预设神经网络可以但不限于包括如下至少一：
深度神经网络；
卷积神经网络；
残差卷积神经网络；
带注意力机制的残差卷积神经网络。
上述给出的预设神经网络的示例可以为线下训练、线上部署的，也可以为统一在线下训练完成之后再进行线上部署的，也可以为统一在线上训练并完成部署的，此处不作限定；除了上述给出的预设神经网络的示例之外，本领域技术人员也可以根据具体应用场景的特点、需求等因素以考虑设置合适的预设神经网络，此处不再赘述。
在一实施例中，历史信道估计值和历史接收信号与当前信道估计值和待测接收信号形成区分，即历史信道估计值和历史接收信号的作用在于用来训练预设神经网络，当前信道估计值和待测接收信号则为待预设神经网络进行检测的信号参数，但随着训练过程的持续进行，当前信道估计值和待测接收信号也可能成为新的历史信道估计值和历史接收信号，也就是说，对于预设神经网络的训练可以为动态化的，在时间上具有一定的迁移性，以便于将更多的信号样本用于训练预设神经网络，以达到更好的训练效果，后续实施例中将针对预设神经网络的训练过程逐步展开详细描述。
在一实施例中，待通过至少一组历史信号样本对预设神经网络进行训练而得到预训练的预设神经网络之后，可以将历史信道估计值和历史接收信号作为一组检测数据输入到预设神 经网络中进行再检测，根据检测结果判断预设神经网络的实际训练效果如何，有利于进一步判断预设神经网络的整体训练过程是否可能出现错误。
在一实施例中，历史信道估计值、历史接收信号以及当前信道估计值、待测接收信号可以通过一组信号按顺序发送而获取到的，也可以由各自对应的信号分别发送而获取到，此处并未限定。
在一实施例中，历史信号样本的具体个数可以根据具体应用场景进行设置，也即对于预设神经网络的训练样本可以根据具体应用场景进行设置，此处并未限定。
如图2所示，本申请的一个实施例，当历史信号样本为多组，且每个无线衰落信道对应于多组历史信号样本时，步骤S110可以但不限于包括步骤S111至步骤S113。
步骤S111：获取由发送端发送的历史调制信号；
步骤S112：在每个无线衰落信道中，对历史调制信号进行预设信号处理而在接收端得到接收导频信号、本地导频信号和历史接收信号；
步骤S113：将接收导频信号乘以本地导频信号的共轭，得到历史信道估计值。
本步骤中，由于历史信号样本为多组，且每个无线衰落信道对应于多组历史信号样本，因此通过获取由发送端发送的历史调制信号，使得能够在各个无线衰落信道中，分别地对历史调制信号进行预设信号处理，从而在接收端得到接收导频信号、本地导频信号和历史接收信号，进而通过将接收导频信号乘以本地导频信号的共轭而得到历史信道估计值，也就是说，由于信道的变化导致接收信号的幅度发生随机变化而产生信号衰落，因此对于每个无线衰落信道均可以对应地得到至少一组完整、不同的历史信号样本，这样总共能够生成多个历史信号样本以供预设神经网络进行训练，训练效果能够得到极大保证。
在一实施例中，发送端和接收端是一组相对概念，即发送端对应于信号的初始来源，接收端对应于信号的后端结果，此处并非限定发送端和接收端的具体呈现形式，只是用于清楚地说明本申请实施例的工作原理，在具体应用场景中可以设置具体的发送端和接收端，此处并未限定。
在一实施例中，获取由发送端发送的历史调制信号的方式可以为多种，此处限定。例如，可以通过获取由发送端发送的至少一组发送比特流信号，进而对至少一组发送比特流信号进行编码、调制而得到历史调制信号，也就是说，通过对初始的发送比特流信号进行预处理而得到历史调制信号；或者，可以由发送端预先将待处理的信号调节为历史调制信号，这样就可以直接从发送端获取到历史调制信号等。
在一实施例中，将接收导频信号乘以本地导频信号的共轭而得到历史信道估计值的具体方式可以为多种，此处并未限定。例如，可以但不限于采用最小二乘LS信道估计算法将接收导频信号乘以本地导频信号的共轭而得到历史信道估计值，或者可以但不限于采用本领域常见的其他传统通信算法所实现，其中，最小二乘法是一种在误差估计、不确定度、系统辨识及预测、预报等数据处理领域得到应用的一种数学模型，为本领域技术人员所熟知，此处对其不作赘述。
在一实施例中，对历史调制信号进行预设信号处理的方式可以为多种，此处并未限定。例如，可以为对历史调制信号进行快速傅里叶变换、添加循环前缀之后得到第一输出信号，然后将该第一输出信号分别通过不同的无线衰落信道而得到不同的第二输出信号，然后对第二输出信号进行去循环前缀、逆傅里叶变换以实现信号还原，即可最终得到接收导频信号、 本地导频信号和历史接收信号；又如，可以将历史调制信号输入到预设网络模型中，通过预设网络模型对该历史调制信号进行解调及均衡处理，最终得到接收导频信号、本地导频信号和历史接收信号。
如图3所示，本申请的一个实施例，当历史信号样本为多组时，步骤S120之前还包括但不限于步骤S130。
步骤S130：对于每组历史信号样本，以历史信道估计值和历史接收信号作为训练样本、以历史调制信号作为训练样本的标签，对预设神经网络进行训练而得到预训练的预设神经网络。
本步骤中，以历史信道估计值和历史接收信号作为训练样本，使得预设神经网络在训练时更容易学习到不同的无线衰落信道的特征，以获得更加良好的训练效果，并且无需采用其他训练样本即可实现完整的训练流程，因此训练所需参数和时间均会减少，并且以与历史信道估计值和历史接收信号相关联的历史调制信号作为训练样本的标签，能够提升预设神经网络训练的稳定性。
以下给出一个具体示例以说明上述各实施例的工作原理。
示例一：
首先，需要确定预设神经网络，假设在具有N个接收天线的MIMO系统中，接收信号由Y表示，YData表示历史接收信号，YPilot表示接收导频信号，本地导频信号为XPilot。先通过XPilot和YPilot进行信道估计值计算，用如下公式(1)得到LS信道估计值
然后将LS信道估计值和YData一起输入到预设神经网络进行训练而得到最终的发送数据估计值
如图4所示，本申请的一个实施例，对步骤S130进行进一步说明，步骤S130可以包括但不限于步骤S131至S133。
步骤S131：对获取到的历史信道估计值、历史接收信号的实部和虚部进行组合，得到作为训练样本的第一实数数组；
步骤S132：对获取到的历史调制信号的实部和虚部进行组合，得到作为训练样本的标签的第二实数数组；
步骤S133：将第一实数数组和第二实数数组输入到预设神经网络进行训练，得到历史接收信号的信号估计值。
本步骤中，通过获取历史信道估计值、历史接收信号的实部和虚部以组合得到作为训练样本的第一实数数组，以及通过获取历史调制信号的实部和虚部进行组合得到作为训练样本的标签的第二实数数组，也就是说，统一将历史信道估计值、历史接收信号和历史调制信号的输入形式设置为符合预设神经网络输入要求的格式，以便于预设神经网络根据实数数组进行训练得到历史接收信号的信号估计值，并且在具体的应用场景中，相应的实部和虚部的值可以分别表征天线端口和接收天线的值，即通过实数数组的形式能够更好地适应神经网络的运算。
在一实施例中，除了上述的实数神经网络之外，预设神经网络还可以为其他类型，相应地，对于预设神经网络的输入可以随着预设神经网络的类型变化而变化，也就是说，上述实施例中的步骤S131至S133并不作为对输入到预设神经网络的内容的限定，还可以根据具体 场景选择适宜的训练输入内容，此处不作限定。
如图5所示，本申请的一个实施例，对步骤S131进行进一步说明，步骤S131可以包括但不限于步骤S1311至S1312。
步骤S1311：获取历史信道估计值的第一实部和第一虚部，以及获取历史接收信号的第二实部和第二虚部；
步骤S1312：按预设顺序排列组合第一实部、第一虚部、第二实部和第二虚部，得到作为训练样本的第一实数数组。
本步骤中，通过分别获取历史信道估计值的第一实部和第一虚部，以及获取历史接收信号的第二实部和第二虚部，可以分别生成历史信道估计值和历史接收信号对应的实数集合，进而按照预设顺序排列组合各个实数集合中相应的第一实部、第一虚部、第二实部和第二虚部，就能够得到包括历史信道估计值和历史接收信号的第一实数数组。
在一实施例中，排列的预设顺序不限定，可以但不限于根据历史信道估计值和历史接收信号对应的实部和虚部分别进行排列，也即将历史信道估计值的实部和虚部排列在相邻位置，以及将历史接收信号的实部和虚部排列在相邻位置，或者可以根据具体应用场景进行顺序排列等。
如图6所示，本申请的一个实施例，当历史调制信号包括调制信号流，调制信号流包括多个层调制信号时，对步骤S132进行进一步说明，步骤S132可以包括但不限于步骤S1321至S1322。
步骤S1321：获取各个层调制信号的实部和虚部；
步骤S1322：按预设顺序排列组合各个层调制信号的实部和虚部，得到作为训练样本的标签的第二实数数组。
本步骤中，当历史调制信号包括调制信号流，调制信号流包括多个层调制信号，说明此时的历史调制信号相当于为一个信号层，通过获取其中的各个层调制信号的实部和虚部，就相当于获取到了历史调制信号的整个的实部和虚部，然后可以按照预设顺序排列组合各个层调制信号的实部和虚部而得到作为训练样本的标签的第二实数数组。
在一实施例中，排列的预设顺序不限定，可以但不限于根据各个层调制信号的实部和虚部分别进行排列，也即将每个层调制信号的实部和虚部排列在相邻位置，同时将相邻的层调制信号的实部和虚部排列在相邻位置等，或者可以根据具体应用场景进行顺序排列等。
如图7所示，本申请的一个实施例，步骤S133之后还包括但不限于步骤S134至S135,其中，步骤S134至S135用于确定得到预训练的预设神经网络。
步骤S134：根据历史接收信号的信号估计值和第二实数数组，得到预设神经网络对应的预设损失函数的当前数值；
步骤S135:当预设损失函数的当前数值符合预设收敛条件，确定得到预训练的预设神经网络。
本步骤中，通过训练得到的信号估计值和预确定的第二实数数组计算得到损失函数并基于计算数值判断是否符合预设收敛条件，以便于在确定计算数值不符合预设收敛条件的情况下，实现对于预设神经网络的进一步训练优化而得到新的信号估计值，以此类推，再通过新的信号估计值计算得到新的损失函数值并再次进行判断，从而能够不断地优化预设神经网络，使得最终训练得到的信号估计值不断逼近发送信号，而且由于可以灵活地支持不同长度的数 据比特源和不同的调制方式，因此可以更可靠地应用在计算损失函数的相关场景中，最终能够训练得到符合要求的预设神经网络。
在一实施例中，预设损失函数包括如下至少之一：
均方误差(Mean Square Error，MSE)损失函数；
均方根误差(Root Mean Squared Error，RMSE)损失函数；
平均绝对值误差(Mean Absolute Error，MAE)损失函数；
欧氏距离损失函数；
余弦距离损失函数；
MSE损失函数、RMSE损失函数、MAE损失函数、欧氏距离损失函数和余弦距离损失函数的线性加权函数。
上述损失函数为本领域技术人员所熟知，在此对其不作赘述；或者，本领域技术人员还可以根据具体应用场景为预设神经网络设置相应的预设损失函数，此处并未限定。
以下给出具体示例以说明上述各实施例的工作原理。
示例二：
首先，基于MIMO系统构建预设神经网络的训练输入，该预设神经网络为实数神经网络，所以需要把和YData的实部和虚部分开，构成一组实数数组，如： 其中Re(·)和Im(·)分别是取实部和取虚部的操作。由于MIMO系统中有P个天线端口和R个接收天线上的值，YData有R个接收天线上的值，所以前面得到的实数数组中也包含了P个天线端口和R个接收天线上的实部和虚部的值，那么这个实数数组就作为预设神经网络的训练输入。
然后，进一步产生预设神经网络的训练样本的标签并进行模型训练。参照图12，通过发送端发出的发送比特流，通过编码和调制后生成了包含多层MIMO信号的发送调制信号XData(即历史调制信号)，经过无线衰落信道之后，基于神经网络(即预训练的预设神经网络，下同，不再赘述)的MIMO信号检测器得到接收导频信号YPilot、接收信号YData和本地导频信号XPilot，按照类似于上一步骤的操作，可以得到实数数组 以之作为训练样本，另外发送调制信号XData同样按照实部和虚部依次排列，组成一个一维实数数组，如下所示：
其中，K＝T·Llayer，Llayer为MIMO信号的层数，T为每个层调制信号所占的资源粒子(Resource Element，RE)数，这个一维实数数组X′Data即作为训练样本的标签，可以看出，对于多层MIMO信号而言，由于是采用上述方式统一排列的，因此并不需要关注其具体长度或调制方式，也就是说，本示例可以灵活地支持不同长度的数据比特流和不同的调制方式。以此类推，在不同的无线衰落信道下，生成多组训练样本和标签，设定每组训练样本和标签的排列顺序和标签数据相同，并输入到预设神经网络中进行训练，得到发送信号估计值然后将该发送信号估计值与标签X′Data一起计算损失函数，待损失函数收敛到一定程度后即可以确定完成网络训练。
如图8所示，本申请的一个实施例，对步骤S120进行进一步说明，步骤S120包括但不限于步骤S121至S122。
步骤S121：对获取到的当前信道估计值和待测接收信号的实部和虚部进行组合，得到第三实数数组；
步骤S122:将第三实数数组输入到预训练的预设神经网络进行信号检测，得到待测接收信号的实数信号估计结果。
本步骤中，通过获取当前信道估计值和待测接收信号的实部和虚部以组合得到作为检测输入的第三实数数组，从而为预设神经网络提供对应符合要求的输入参量，也就是说，统一将当前信道估计值和待测接收信号的输入形式设置为符合预设神经网络输入要求的格式，以便于预设神经网络根据实数数组进行检测得到待测接收信号的实数信号估计结果。
在一实施例中，除了上述输出的实数信号估计结果之外，预设神经网络还可以输出为其他类型，相应地，对于预设神经网络的输入可以随着预设神经网络的输出类型变化而变化，此处并未限定。
如图9所示，本申请的一个实施例，对步骤S121进行进一步说明，步骤S121包括但不限于步骤S1211至S1212。
步骤S1211：获取当前信道估计值的第三实部和第三虚部，以及获取待测接收信号的第四实部和第四虚部；
步骤S1212:按预设顺序排列组合第三实部、第三虚部、第四实部和第四虚部，得到第三实数数组。
本步骤中，通过分别获取当前信道估计值的第三实部和第三虚部，以及获取待测接收信号的第四实部和第四虚部，可以分别生成当前信道估计值和待测接收信号对应的实数集合，进而按照预设顺序排列组合各个实数集合中相应的第三实部、第三虚部、第四实部和第四虚部，就能够得到包括当前信道估计值和待测接收信号的第三实数数组。
在一实施例中，排列的预设顺序不限定，可以但不限于根据当前信道估计值和待测接收信号对应的实部和虚部分别进行排列，也即将当前信道估计值的实部和虚部排列在相邻位置，以及将待测接收信号的实部和虚部排列在相邻位置，或者可以根据具体应用场景进行顺序排列等。
如图10所示，本申请的一个实施例，对步骤S122之后的步骤实施例进行进一步说明，步骤S122之后还包括但不限于步骤S123至S124。
步骤S123：对待测接收信号的实数信号估计结果进行复数转换处理，得到待测接收信号的复数信号估计结果；
步骤S124：对复数信号估计结果进行解调、译码，得到输出信号。
本步骤中，通过对待测接收信号的实数信号估计结果得到待测接收信号的复数信号估计结果，以便于通过该复数信号估计结果进行进一步地进行解调和译码，从而得到最终的输出信号。
如图11所示，本申请的一个实施例，在实数信号估计结果包括实数信号估计数组，实数信号估计数组包括按顺序排列的多个层信号的实部和虚部的情况下，对步骤S123进行进一步说明，步骤S123可以但不限于包括步骤S1231至S1232。
步骤S1231：分别对实数信号估计数组中的各个层信号的实部和虚部进行复数转换，得到各个层信号对应的复数转换数据；
步骤S1232:按预设顺序排列组合各个复数转换数据，得到待测接收信号的复数信号估 计结果。
本步骤中，由于实数信号估计结果包括实数信号估计数组，且实数信号估计数组包括按顺序排列的多个层信号的实部和虚部，因此可以分别对实数信号估计数组中的各个层信号的实部和虚部进行复数转换，从而得到每个层信号对应的复数转换数据，进而通过对各个复数转换数据进行排列组合，能够准确可靠地得到待测接收信号的复数信号估计结果。
以下给出一个具体示例以说明上述各实施例的工作原理。
示例三：
如图12所示，利用预设神经网络实施在线信号检测，预设神经网络训练完成之后即可部署在MIMO系统中的接收端进行在线信号检测。当在线信号检测时，可以参照示例一和示例二中的方式类似的得到预设神经网络的输入参数，然后预设神经网络进行在线检测，得到预设神经网络的输出即发送信号的估计值，然后完成实数数组到均衡后复数值数据的转换，如下操作所示：
其中，complex(·)表示组成一个复数值数据的操作，利用转换得到的复数值数据再进行后续传解调和译码，得到最终的输出信号。
以下给出一个完整流程示例以说明上述各实施例的工作原理。
示例四：
如图13所示，在一个MIMO技术下的OFDM系统中，设置有基于神经网络(即本申请实施例中的预设神经网络)的MIMO信号检测器和LS信道估计模块，此时有L＝2个数据层，P＝2个天线端口和T＝2个物理发送天线，R＝2个物理接收天线。接收端频域信号Y可以表示为公式(2)：
其中，X是这2个数据层调制后的发送信号，维度是L×1，即2×1；H是发送端和接收端之间的无线信道矩阵，维度是T×R，即2×2；N是加性高斯白噪声，维度是R×1，即2×1。频域接收信号YData和接收导频信号YPilot的维度也是R×1，即2×1，接收端的本地导频信号XPilot的维度为P×1，即2×1。假设频域接收信号和接收导频信号一共占了2个OFDM符号，那么接收导频信号则占了第一个OFDM符号，频域接收信号占了第二个OFDM符号。
本示例提出的基于神经网络的MIMO信号检测方法就是着眼于如何从接收端频域信号Y中检测出发送信号X，其具体步骤如下：
步骤1：确定基于神经网络的MIMO信号检测器的结构。先把XPilot和YPilot输入到LS信道估计子模块，用上述示例一中的公式(1)得到LS信道估计值其维度为P×R，即2×2。然后将和YData一起输入到基于神经网络的MIMO信号检测器，得到最终的发送数据估计值其中，LS信道估计子模块可以依靠相关技术中的方法构建，基于神经网络的MIMO信号检测器采用一个卷积神经网络(Convolutional Neural Network，CNN)。
步骤2：构建卷积神经网络MIMO信号检测器的训练输入。按照上述示例一中的类似方式，将和YData的实部和虚部分开，构成一组维度为120×12的实数数组并按照顺序排列。由 于MIMO系统中有P＝2个天线端口和R＝2个接收天线上的值，YData有R＝2个接收天线上的值，如果MIMO信号占用了10个资源块(每个资源块有12个子载波)，即NRB＝10个，那么一共占了120个子载波，这个实数数组的维度即为120×12，这个实数数组就作为卷积神经网络的训练输入。
步骤3：待产生MIMO信号检测器的训练样本的标签之后，进行模型训练。发送端生成发送调制信号XData(包含2个层MIMO信号)，经过无线衰落信道后得到接收端的导频信号YPilot、接收信号YData和本地导频信号XPilot，按照步骤2的类似操作得到一个实数数组作为训练样本。
另外，发送调制信号XData按照实部和虚部依次排列，先排第一个层的调制信号，再排第二个层的调制信号，最终组成一个一维实数数组，如下所示：
其中，K＝T·Llayer，Llayer＝2为MIMO信号的层数，T＝120为MIMO信号每个层调制信号所占的RE数。那么这个一维实数数组X′Data的维度为1×480，作为训练样本的标签。在不同的衰落信道下，生成多组训练样本和标签，并输入到CNN中进行训练，得到发送信号估计值然后将其和标签X′Data一起计算MSE损失函数，如下面公式(3)所示：
其中，n是每次训练的batch size样本数。当模型训练时，基于MSE损失函数利用Adam优化器对CNN进行训练，待损失函数收敛到一定程度后确定完成网络训练，此时保存模型参数，用于后续在线信号检测。
步骤4：实施在线MIMO信号检测，CNN训练完成之后即可部署在OFDM系统中的接收端进行在线信号检测。当在线信号检测时，同样按照步骤1和步骤2中的处理流程，类似地先得到CNN的输入参数再进行在线检测，可以得到CNN的输出即发送信号的估计值，该发送信号的估计值包括相应的实数数组，然后完成实数数组到均衡后复数值数据的转换，如下操作：
其中，complex(·)表示组成一个复数值数据的操作，利用转换得到的复数值数据再进行后续传解调和译码，得到最终的输出信号。
可以看出，本申请实施例的信号检测方法，其输出是均衡后解层映射后的数据，并采用发送端调制后的信号作为训练标签，并使用MSE作为损失函数，使得估计值不断逼近发送信号，同时可以灵活地支持不同长度的数据比特流和不同的调制方式，具有良好的泛用性。
如图14所示，本申请的一个实施例，当历史信号样本为多组，且预设神经网络包括增强信道估计子网络和信道均衡子网络时，步骤S120之前还包括但不限于步骤S140至S150。
步骤S140：对于每组历史信号样本，将历史信道估计值输入到增强信道估计子网络进行信道增强估计训练，得到第一预估结果；
步骤S150:将第一预估结果和历史接收信号输入到信道均衡子网络中进行信道均衡训练而得到预训练的预设神经网络。
本步骤中，预设神经网络包括增强信道估计子网络和信道均衡子网络这两个子网络，每 个子网络分别具备单独的性能，因此对于每组历史信号样本来说，可以分别将历史信道估计值输入到增强信道估计子网络进行信道增强估计训练得到第一预估结果，再将第一预估结果和历史接收信号输入到信道均衡子网络中进行信道均衡训练而得到预训练的预设神经网络，相比于功能一体化设置的预设神经网络，能够达到同样有效可靠的训练效果。
在一实施例中，增强信道估计子网络和信道均衡子网络可以一起进行训练，也可以单独地分别进行训练，或者本领域技术人员也可以根据具体应用场景选择相应的训练方式对增强信道估计子网络和信道均衡子网络进行训练等，此处并未限定。
如图15所示，本申请的一个实施例，步骤S120包括但不限于步骤S125至S126。
步骤S125：将当前信道估计值输入到预训练的增强信道估计子网络进行信道增强估计，得到第二预估结果；
步骤S126:将第二预估结果和待测接收信号输入到预训练的信道均衡子网络中进行信道均衡，得到待测接收信号的信号估计值。
本步骤中，当需要进行信号检测时，通过将当前信道估计值输入到预训练的增强信道估计子网络进行信道增强估计而直接得到第二预估结果，然后将第二预估结果和待测接收信号输入到预训练的信道均衡子网络中进行信道均衡，就能够得到待测接收信号的信号估计值，通过分别测试计算信道预估结果和最终的信号估计值，有利于提高预设神经网络的信号检测精度。
需要说明的是，步骤S140至S150、步骤S125至S126的工作原理与前述实施例中的相关步骤所类似，区别仅在于前述实施例中的预设神经网络为一体式结构，而本申请中的预设神经网络分为两个子网络，通过两个子网络的配合可以完整地实现单个的预设神经网络的功能，由于前述实施例中对于预设神经网络的详细原理及流程已经详细说明，因此本实施例的步骤实施例可以参照前述实施例中对于预设神经网络的详细原理及流程，为免冗余，在此不作赘述。
以下给出另一个完整流程示例以说明上述各实施例的工作原理。
示例五：
如图16所示，在一个MIMO技术下的OFDM系统中，设置有基于神经网络(即本申请实施例中的预设神经网络)的MIMO信号检测器和LS信道估计模块，此时有L＝4个数据层，P＝4个天线端口和T＝4个物理发送天线，R＝4个物理接收天线。接收端频域信号Y可以表示为公式(4)：
其中，X是这4个数据层调制后的发送信号，维度是L×1，即4×1；H是发送端和接收端之间的无线信道矩阵，维度是T×R，即4×4；N是加性高斯白噪声，维度是R×1，即4×1。频域接收信号YData和接收导频信号YPilot的维度也是R×1，即4×1，接收端的本地导频信号XPilot的维度为P×1，即4×1。假设频域接收信号和接收导频信号一共占了4个OFDM符号，那么接收导频信号则占了第一个OFDM符号，频域接收信号占了剩余3个OFDM符号。
本示例提出的基于神经网络的MIMO信号检测方法就是着眼于如何从接收端频域信号Y中检测出发送信号X，其具体步骤如下：
步骤1：确定基于神经网络的MIMO信号检测器的结构。先把XPilot和YPilot输入到LS信道估计子模块，用上述示例一中的公式(1)得到LS信道估计值其维度为P×R，即4×4。然后将和YData一起输入到基于神经网络的MIMO信号检测器，得到最终的发送数据估计值其中，LS信道估计子模块可以依靠相关技术中的方法构建，基于神经网络的MIMO信号检测器包含2个子网络：一个为增强信道估计子网络，用于信道估计增强，起到去除噪声、干扰的作用，提升信道估计值的质量；另一个为信道均衡子网络，用于信道均衡，进行MIMO信号的检测，这2个子网络都采用带有通道注意力机制的残差神经网络。
步骤2：构建增强信道估计子网络的训练输入，按照上述示例一中的类似方式，将的实部和虚部分开，构成一组维度为240×32的实数数组并按照顺序排列。由于MIMO系统中有P＝4个天线端口和R＝4个接收天线上的值，YData有R＝4个接收天线上的值。如果MIMO信号占用了20个资源块(每个资源块有12个子载波)，即NRB＝20个，那么一共占了240个子载波，并且导频占了一个OFDM符号，因此这个实数数组的维度即为240×32，这个实数数组H′LS就作为增强信道估计子网络的训练输入，把H′LS输入到增强信道估计子网络中，经过残差神经网络得到输出H″，其维度为240×32。
然后，构建信道均衡子网络的训练输入，把H″和YData的实部和虚部分开，构成一组实数数组，按照下面顺序排列，其维度为240×56：
这里YData占了3个OFDM符号，YData的维度是240*24。那么，这个维度为240×56的实数数组就作为信道均衡子网络的训练输入。
步骤3：待产生MIMO信号检测器的训练样本的标签之后，进行模型训练。发送端生成发送调制信号XData(包含3个OFDM符号上的4个层MIMO信号)，经过无线衰落信道后得到接收端的导频信号YPilot、接收信号YData和本地导频信号XPilot，按照步骤2的类似操作分别得到增强信道估计子网络和信道均衡子网络的训练样本。
另外，发送调制信号XData按照实部和虚部依次排列，先排第一层的调制信号，再排第二层的调制信号，最终组成一个一维实数数组，如下所示：
其中，K＝T·Llayer，Llayer＝4为MIMO信号的层数，T＝240*3为MIMO信号每个层层调制信号所占的RE数。那么这个一维实数数组X′Data的维度为1×5760，作为训练样本的标签。在不同的衰落信道下，生成多组训练样本和标签，并输入到增强信道估计子网络和信道均衡子网络进行整体训练，由信道均衡子网络输出而得到发送信号估计值然后将其和标签X′Data一起计算欧氏距离作为损失函数，如下面公式(5)所示：
其中，n是每次训练的batch size样本数。当模型训练时，基于欧氏距离的损失函数利用Adam优化器对残差神经网络进行训练，待损失函数收敛到一定程度后确定完成网络训练，此时保存模型参数，用于后续在线信号检测。
步骤4：实施在线MIMO信号检测，残差神经网络训练完成之后即可部署在OFDM系统中的接收端进行在线信号检测。当在线信号检测时，同样按照步骤1和步骤2中的处理流程，类似地先得到残差神经网络的输入参数再进行在线检测，可以得到残差神经网络的输出即发送信号的估计值，该发送信号的估计值包括相应的实数数组，然后完成实数数组到均衡后复数值数据的转换，如下操作：
其中，complex(·)表示组成一个复数值数据的操作，利用转换得到的复数值数据再进行后续传解调和译码，得到最终的输出信号。
可以看出，本申请实施例的信号检测方法，其输出是均衡后解层映射后的数据，并采用发送端调制后的信号作为训练标签，并使用欧式距离作为损失函数，使得估计值不断逼近发送信号，同时可以灵活地支持不同长度的数据比特流和不同的调制方式，具有良好的泛用性。
另外，如图17所示，本申请的一个实施例还公开了一种信号检测设备100，包括：至少一个处理器110；至少一个存储器120，用于存储至少一个程序；当至少一个程序被至少一个处理器110执行时实现如前面任意实施例中的信号检测方法。
另外，本申请的一个实施例还公开了一种计算机可读存储介质，其中存储有计算机可执行指令，计算机可执行指令用于执行如前面任意实施例中的信号检测方法。
此外，本申请的一个实施例还公开了一种计算机程序产品，包括计算机程序或计算机指令，计算机程序或计算机指令存储在计算机可读存储介质中，计算机设备的处理器从计算机可读存储介质读取计算机程序或计算机指令，处理器执行计算机程序或计算机指令，使得计算机设备执行如前面任意实施例中的信号检测方法。
本领域普通技术人员可以理解，上文中所公开方法中的全部或某些步骤、系统可以被实施为软件、固件、硬件及其适当的组合。某些物理组件或所有物理组件可以被实施为由处理器，如中央处理器、数字信号处理器或微处理器执行的软件，或者被实施为硬件，或者被实施为集成电路，如专用集成电路。这样的软件可以分布在计算机可读介质上，计算机可读介质可以包括计算机存储介质(或非暂时性介质)和通信介质(或暂时性介质)。如本领域普通技术人员公知的，术语计算机存储介质包括在用于存储信息(诸如计算机可读指令、数据结构、程序模块或其他数据)的任何方法或技术中实施的易失性和非易失性、可移除和不可移除介质。计算机存储介质包括但不限于RAM、ROM、EEPROM、闪存或其他存储器技术、CD-ROM、数字多功能盘(DVD)或其他光盘存储、磁盒、磁带、磁盘存储或其他磁存储装置、或者可以用于存储期望的信息并且可以被计算机访问的任何其他的介质。此外，本领域普通技术人员公知的是，通信介质通常包含计算机可读指令、数据结构、程序模块或者诸如载波或其他传输机制之类的调制数据信号中的其他数据，并且可包括任何信息递送介质。

Claims (20)

  1. 一种信号检测方法，包括：
    获取至少一组历史信号样本，其中，所述历史信号样本包括历史信道估计值、历史接收信号和历史调制信号；
    在获取到当前信道估计值和待测接收信号的情况下，将所述当前信道估计值和所述待测接收信号输入到预训练的预设神经网络进行信号检测，得到所述待测接收信号的信号估计值；
    其中，预训练的所述预设神经网络为通过至少一组所述历史信号样本对所述预设神经网络进行训练而得到。
  2. 根据权利要求1所述的信号检测方法，其中，所述历史信号样本为多组，每个无线衰落信道对应于多组所述历史信号样本；所述获取至少一组历史信号样本，包括：
    获取由发送端发送的历史调制信号；
    在每个所述无线衰落信道中，对所述历史调制信号进行预设信号处理而在接收端得到接收导频信号、本地导频信号和历史接收信号；
    将所述接收导频信号乘以所述本地导频信号的共轭，得到历史信道估计值。
  3. 根据权利要求2所述的信号检测方法，其中，基于如下步骤通过至少一组所述历史信号样本对所述预设神经网络进行训练而得到预训练的所述预设神经网络：
    对于每组所述历史信号样本，以所述历史信道估计值和所述历史接收信号作为训练样本、以所述历史调制信号作为所述训练样本的标签，对所述预设神经网络进行训练而得到预训练的所述预设神经网络。
  4. 根据权利要求2所述的信号检测方法，其中，当所述预设神经网络包括增强信道估计子网络和信道均衡子网络时，基于如下步骤通过至少一组所述历史信号样本对所述预设神经网络进行训练而得到预训练的所述预设神经网络：
    对于每组所述历史信号样本，将所述历史信道估计值输入到所述增强信道估计子网络进行信道增强估计训练，得到第一预估结果；
    将所述第一预估结果和所述历史接收信号输入到所述信道均衡子网络中进行信道均衡训练而得到预训练的所述预设神经网络。
  5. 根据权利要求4所述的信号检测方法，其中，所述将所述当前信道估计值和所述待测接收信号输入到预训练的预设神经网络进行信号检测，得到所述待测接收信号的信号估计值，包括：
    将所述当前信道估计值输入到预训练的所述增强信道估计子网络进行信道增强估计，得到第二预估结果；
    将所述第二预估结果和所述待测接收信号输入到预训练的所述信道均衡子网络中进行信道均衡，得到所述待测接收信号的信号估计值。
  6. 根据权利要求3所述的信号检测方法，其中，所述以所述历史信道估计值和所述历史接收信号作为训练样本、以所述历史调制信号作为所述训练样本的标签，对所述预设神经网络进行训练，包括：
    对获取到的所述历史信道估计值、所述历史接收信号的实部和虚部进行组合，得到作为训练样本的第一实数数组；
    对获取到的所述历史调制信号的实部和虚部进行组合，得到作为所述训练样本的标签的第二实数数组；
    将所述第一实数数组和所述第二实数数组输入到所述预设神经网络进行训练，得到所述历史接收信号的信号估计值。
  7. 根据权利要求6所述的信号检测方法，其中，所述对获取到的所述历史信道估计值、所述历史接收信号的实部和虚部进行组合，得到作为训练样本的第一实数数组，包括：
    获取所述历史信道估计值的第一实部和第一虚部，以及获取所述历史接收信号的第二实部和第二虚部；
    按预设顺序排列组合所述第一实部、所述第一虚部、所述第二实部和所述第二虚部，得到作为训练样本的第一实数数组。
  8. 根据权利要求6所述的信号检测方法，其中，所述历史调制信号包括调制信号层，所述调制信号层包括多个层调制信号；所述对获取到的所述历史调制信号的实部和虚部进行组合，得到作为所述训练样本的标签的第二实数数组，包括：
    获取各个所述层调制信号的实部和虚部；
    按预设顺序排列组合各个所述层调制信号的实部和虚部，得到作为所述训练样本的标签的第二实数数组。
  9. 根据权利要求6所述的信号检测方法，其中，所述得到所述历史接收信号的信号估计值之后，基于如下步骤确定得到预训练的所述预设神经网络：
    根据所述历史接收信号的信号估计值和所述第二实数数组，得到所述预设神经网络对应的预设损失函数的当前数值；
    当所述预设损失函数的当前数值符合预设收敛条件，确定得到预训练的所述预设神经网络。
  10. 根据权利要求9所述的信号检测方法，其中，所述预设损失函数包括如下至少之一：
    均方误差MSE损失函数；
    均方根误差RMSE损失函数；
    平均绝对值误差MAE损失函数；
    欧氏距离损失函数；
    余弦距离损失函数；
    MSE损失函数、RMSE损失函数、MAE损失函数、欧氏距离损失函数和余弦距离损失函数的线性加权函数。
  11. 根据权利要求1所述的信号检测方法，其中，所述将所述当前信道估计值和所述待测接收信号输入到预训练的预设神经网络进行信号检测，得到所述待测接收信号的信号估计值，包括：
    对获取到的所述当前信道估计值和所述待测接收信号的实部和虚部进行组合，得到第三实数数组；
    将所述第三实数数组输入到预训练的预设神经网络进行信号检测，得到所述待测接收信号的实数信号估计结果。
  12. 根据权利要求11所述的信号检测方法，其中，所述对获取到的所述当前信道估计值和所述待测接收信号的实部和虚部进行组合，得到第三实数数组，包括：
    获取所述当前信道估计值的第三实部和第三虚部，以及获取所述待测接收信号的第四实部和第四虚部；
    按预设顺序排列组合所述第三实部、所述第三虚部、所述第四实部和所述第四虚部，得到第三实数数组。
  13. 根据权利要求11所述的信号检测方法，其中，所述得到所述待测接收信号的实数信号估计结果之后，还包括：
    对所述待测接收信号的实数信号估计结果进行复数转换处理，得到所述待测接收信号的复数信号估计结果；
    对所述复数信号估计结果进行解调、译码，得到输出信号。
  14. 根据权利要求13所述的信号检测方法，其中，所述实数信号估计结果包括实数信号估计数组，所述实数信号估计数组包括按顺序排列的多个层信号的实部和虚部；所述对所述待测接收信号的实数信号估计结果进行复数转换处理，得到所述待测接收信号的复数信号估计结果，包括：
    分别对所述实数信号估计数组中的各个所述层信号的实部和虚部进行复数转换，得到各个所述层信号对应的复数转换数据；
    按预设顺序排列组合各个所述复数转换数据，得到所述待测接收信号的复数信号估计结果。
  15. 根据权利要求2所述的信号检测方法，其中，所述将所述接收导频信号乘以所述本地导频信号的共轭，得到历史信道估计值，包括：
    采用最小二乘LS信道估计算法将所述接收导频信号乘以所述本地导频信号的共轭，得到历史信道估计值。
  16. 根据权利要求2所述的信号检测方法，其中，所述获取由发送端发送的历史调制信号，包括：
    获取由发送端发送的至少一组发送比特流信号；
    对至少一组所述发送比特流信号进行编码、调制，得到历史调制信号。
  17. 根据权利要求1所述的信号检测方法，其中，所述预设神经网络包括如下至少一：
    深度神经网络；
    卷积神经网络；
    残差卷积神经网络；
    带注意力机制的残差卷积神经网络。
  18. 一种信号检测设备，包括：
    至少一个处理器；
    至少一个存储器，用于存储至少一个程序；
    当至少一个所述程序被至少一个所述处理器执行时实现如权利要求1至17任意一项所述的信号检测方法。
  19. 一种计算机可读存储介质，其中存储有处理器可执行的程序，所述处理器可执行的程序被处理器执行时用于实现如权利要求1至17任意一项所述的信号检测方法。
  20. 一种计算机程序产品，包括计算机程序或计算机指令，所述计算机程序或所述计算机指令存储在计算机可读存储介质中，计算机设备的处理器从所述计算机可读存储介质读取所 述计算机程序或所述计算机指令，所述处理器执行所述计算机程序或所述计算机指令，使得所述计算机设备执行如权利要求1至17任意一项所述的信号检测方法。
PCT/CN2023/102432 2022-07-15 2023-06-26 信号检测方法及设备、存储介质 WO2024012185A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210831551.7A CN117439714A (zh) 2022-07-15 2022-07-15 信号检测方法及设备、存储介质
CN202210831551.7 2022-07-15

Publications (1)

Publication Number Publication Date
WO2024012185A1 true WO2024012185A1 (zh) 2024-01-18

Family

ID=89535519

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/102432 WO2024012185A1 (zh) 2022-07-15 2023-06-26 信号检测方法及设备、存储介质

Country Status (2)

Country Link
CN (1) CN117439714A (zh)
WO (1) WO2024012185A1 (zh)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP1908241A2 (en) * 2005-07-13 2008-04-09 LG Electronics Inc. Method for transmitting/receiving signal having spread training symbol in mobile communication system
CN109921882A (zh) * 2019-02-20 2019-06-21 深圳市宝链人工智能科技有限公司 一种基于深度学习的mimo解码方法、装置及存储介质
CN111464465A (zh) * 2020-03-11 2020-07-28 重庆邮电大学 一种基于集成神经网络模型的信道估计方法
CN113890799A (zh) * 2021-10-28 2022-01-04 华南理工大学 一种基于域对抗网络的水声通信信道估计与信号检测方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP1908241A2 (en) * 2005-07-13 2008-04-09 LG Electronics Inc. Method for transmitting/receiving signal having spread training symbol in mobile communication system
CN109921882A (zh) * 2019-02-20 2019-06-21 深圳市宝链人工智能科技有限公司 一种基于深度学习的mimo解码方法、装置及存储介质
CN111464465A (zh) * 2020-03-11 2020-07-28 重庆邮电大学 一种基于集成神经网络模型的信道估计方法
CN113890799A (zh) * 2021-10-28 2022-01-04 华南理工大学 一种基于域对抗网络的水声通信信道估计与信号检测方法

Also Published As

Publication number Publication date
CN117439714A (zh) 2024-01-23

Similar Documents

Publication Publication Date Title
Weng et al. Deep cascading network architecture for robust automatic modulation classification
Zhang et al. Artificial intelligence-aided receiver for a CP-free OFDM system: Design, simulation, and experimental test
JP5854694B2 (ja) 受信装置、受信方法、及び受信プログラム
CN111628833B (zh) 基于卷积神经网络的mimo天线数目估计方法
US20140140425A1 (en) Mimo communication system signal detection method
Ji et al. Modulation recognition in maritime multipath channels: A blind equalization-aided deep learning approach
Ye et al. Bilinear convolutional auto-encoder based pilot-free end-to-end communication systems
Zhang et al. Lightweight automatic modulation classification via progressive differentiable architecture search
An et al. Robust automatic modulation classification in low signal to noise ratio
Dehri et al. Blind digital modulation classification for STBC‐OFDM system in presence of CFO and channels estimation errors
Li et al. Deep learning for OFDM channel estimation in impulsive noise environments
WO2024012185A1 (zh) 信号检测方法及设备、存储介质
Liu et al. RecNet: Deep learning-based OFDM receiver with semi-blind channel estimation
Yıldırım et al. Deep receiver design for multi-carrier waveforms using cnns
KR20200024072A (ko) 기계학습 기반 신호 검출 장치 및 방법
US20230216724A1 (en) Context Aware Data Receiver for Communication Signals Based on Machine Learning
Seo et al. Adversarial, yet friendly signal design for secured wireless communication
Delgado et al. Deep learning decoder for MIMO communications with impulsive noise
Yang et al. Deep learning based detection and channel tracking for mimo systems
Lijun et al. A signal demodulation algorithm based on generative adversarial networks
KR101937559B1 (ko) Mimo-ofdm 시스템을 이용한 선형 근사화 신호 검출 장치 및 그 방법
Wang et al. A spatiotemporal multi-stream learning framework based on attention mechanism for automatic modulation recognition
Ko et al. Unsupervised machine intelligence for automation of multi-dimensional modulation
Xin et al. Dmcnet: Data-driven multi-Pilot convolution neural network for MIMO-OFDM receiver
KR102576415B1 (ko) 다중 인덱스 변조를 수행하는 직교 주파수 분할 다중화 시스템에서 수신기의 메시지 신호 검출 방법 및 수신기

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23838679

Country of ref document: EP

Kind code of ref document: A1