A deep learning-based digital communication bit synchronization method

Abstract

The application discloses a kind of digital communication bit synchronization methods based on deep learning, it is related to signal processing field.The application first constructs known training data and contrast label, then constructs neural network model, training data is used as input data, the output result is compared with contrast label, to update the internal parameter of neural network model, carry out deep learning training, save neural network model parameter after training ends;Digital baseband signal is input into trained neural network model when actual communication, obtains the moment of maximum correlation result.The application uses data-driven deep learning algorithm to adaptively synchronize the maximum correlation result of unknown receiving signal amplitude, and the receiving amplitude of digital baseband signal is not sensitive, so as to improve the robustness of algorithm in practical application, reduce false synchronization and miss synchronization probability, obtain better communication quality.

Description

Technical Field

[0001] This invention relates to the field of signal processing, specifically to a digital communication bit synchronization method based on deep learning, which is applicable to various communication systems where baseband received signals are unstable due to channel environment and background noise, and where adaptive judgment of synchronization waveform correlation results is required. Background Technology

[0002] Digital communication synchronization technology refers to the use of various synchronization techniques in digital communication systems to ensure data communication tasks. Bit synchronization is crucial for baseband and bandwidth transmission in digital communication, ensuring that the receiver samples and decides on the baseband digital signal at the optimal sampling time. There are two methods for bit synchronization: pilot insertion and direct extraction of self-synchronization. Because the direct extraction of self-synchronization has limitations such as inapplicability to all modulation schemes, inapplicability to high symbol rate burst communication, and complex digital processing design, pilot insertion is currently the most commonly used method for bit synchronization in digital communication systems. Traditional pilot-based bit synchronization methods transmit pilots in certain fixed time slots, utilizing the single-peak characteristic of the pilot autocorrelation function for symbol synchronization or positioning. The autocorrelation of the pilot sequence increases with the number of bits in the sequence; its output reaches its maximum only when all bits enter the recognizer. If even one bit is misaligned, the output drops significantly. Better autocorrelation characteristics in the pilot sequence lead to easier recognition and a lower probability of misjudgment. Traditional pilot insertion methods capture the maximum correlation result by setting an autocorrelation threshold. However, in practice, sudden and drastic changes in channel gain can cause the actual correlation result to be lower than the autocorrelation threshold, resulting in missed synchronization, or the sidelobe correlation result to be higher than the autocorrelation threshold, resulting in false synchronization. The lower the worst-case signal-to-noise ratio requirement of the communication system, the lower the autocorrelation threshold. Therefore, for pilot synchronization signals with unknown amplitude, it is difficult to determine the capture of the maximum correlation result by setting a threshold, leading to bit synchronization failure. Summary of the Invention

[0003] In view of this, this invention proposes a deep learning-based digital communication bit synchronization method. This invention can adaptively determine the maximum correlation result when the amplitude of the received signal is unknown during communication, automatically adapting the bit synchronization result to changes in channel gain. This improves the robustness of the algorithm in practical applications, reduces the probability of false synchronization and missed synchronization, and achieves better communication quality.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] A deep learning-based digital communication bit synchronization method includes the following steps:

[0006] Step 1: Input multiple known digital baseband signals containing all pilot sequences, some pilot sequences, and none pilot sequences into the recognizer. The recognizer outputs the correlation result waveform corresponding to each digital baseband signal. In addition, construct the 1 and 0 sequence corresponding to each digital baseband signal, where 1 indicates that the moment is the moment of maximum correlation result, and 0 indicates that the moment is not the moment of maximum correlation result.

[0007] Step 2: Construct a neural network model. The input to the neural network model is the correlation result waveform output by the recognizer, and the output is the 1 and 0 sequence representing the moment of maximum correlation. Use the correlation result waveform obtained in Step 1 to train the neural network model. During the training process, calculate the loss value of the 1 and 0 sequence output by the neural network model relative to the 1 and 0 sequence constructed in Step 1 using the loss function of cross-entropy or mean squared error. Based on the loss value, update the parameters of the neural network model using the Adam backpropagation algorithm until the loss value converges. End the training and save the parameters of the neural network model to complete the deep learning training.

[0008] Step 3: In actual digital signal communication, the received signal at the signal receiving end is oversampled and sent to the recognizer to obtain the correlation result waveform of the received signal. Then, the correlation result waveform is sent to the trained neural network model. If the sequence output by the neural network model contains 1, that is, it contains the maximum correlation moment, then the digital communication bit synchronization is completed. Otherwise, the recognizer continues to output the correlation result waveform of the received signal at the next moment and sends it to the neural network model for processing.

[0009] Furthermore, in step 1, the pilot sequence is continuously inserted within a certain range of the digital baseband signal.

[0010] Furthermore, in step 2, the input layer of the neural network model is a fully connected layer, which is used to connect the input data with the intermediate layer. The intermediate layer is composed of three convolutional layers connected in series, which is used to map the input layer data into data that is convenient for the output layer to process. The output layer is a sigmoid function, which is used to map the intermediate layer data into a sequence of 1s and 0s.

[0011] Furthermore, in step 2, the input layer of the neural network model is a fully connected layer, which is used to associate the input data with the intermediate layer. The intermediate layer is composed of three convolutional layers connected in series, which is used to map the input layer data into data that is convenient for the output layer to process. The output layer is a Softmax function, which is used to map the intermediate layer data into a sequence of 1s and 0s.

[0012] Furthermore, in step 2, the input layer of the neural network model is a fully connected layer, which is used to associate the input data with the intermediate layer. The intermediate layer is a dense convolutional neural network with a depth of 8, which is used to map the input layer data into data that is convenient for the output layer to process. The output layer is a sigmoid function, which is used to map the intermediate layer data into a sequence of 1s and 0s.

[0013] Furthermore, in step 2, the input layer of the neural network model is a fully connected layer, which is used to associate the input data with the intermediate layer. The intermediate layer is a dense convolutional neural network with a depth of 8, which is used to map the input layer data into data that is convenient for the output layer to process. The output layer is a Softmax function, which is used to map the intermediate layer data into a sequence of 1s and 0s.

[0014] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art:

[0015] 1. This invention does not require setting a threshold for the relevant results. It can adaptively determine whether the maximum relevant result exists and the location of the maximum relevant result based on the output of the recognizer, thus ensuring the robustness of bit synchronization calculation.

[0016] 2. The deep learning neural network used in this invention has no special limitations and can be adjusted according to specific needs, which facilitates algorithm expansion.

[0017] 3. After the deep learning training is completed, the network model of this invention is fixed, that is, the calculation process is fixed. It can be implemented through different programming languages ​​and hardware computing platforms. It requires modification of the existing communication system, which is convenient for promotion and porting.

[0018] 4. This invention can adaptively change the deep learning objective based on changes in training data and output labels, such as changing the channel requirements, and has good generalization ability. Attached Figure Description

[0019] Figure 1 This is a flowchart of the digital communication bit synchronization method in an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of the neural network input and output when the maximum correlation result is included in an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of the neural network input and output when there is no maximum correlation result in the embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram of the neural network model in an embodiment of the present invention.

[0023] Figure 5 This is a schematic diagram of deep learning training in an embodiment of the present invention.

[0024] Figure 6 This is a schematic diagram of the Sigmoid function in an embodiment of the present invention.

[0025] Figure 7 This is a schematic diagram showing the relationship between the successful acquisition rate and signal-to-noise ratio of pilot signals with a fixed received amplitude under different threshold settings in an embodiment of the present invention.

[0026] Figure 8 This is a schematic diagram showing the relationship between the successful acquisition rate and signal-to-noise ratio of pilot signals with random received amplitude under different threshold settings in an embodiment of the present invention.

[0027] Figure 9 This is a schematic diagram illustrating the relationship between the successful acquisition rate and signal-to-noise ratio of pilot signals with fixed received amplitude based on a deep learning model in an embodiment of the present invention.

[0028] Figure 10 This is a schematic diagram illustrating the relationship between the successful acquisition rate and signal-to-noise ratio of pilot signals with random received amplitude based on a deep learning model in an embodiment of the present invention. Detailed Implementation

[0029] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0030] A deep learning-based digital communication bit synchronization method, such as Figure 1 As shown, it includes the following steps:

[0031] Step 1: Input multiple known digital baseband signals containing all pilot sequences, some pilot sequences, and none pilot sequences into the recognizer. The recognizer outputs the correlation result waveform corresponding to each digital baseband signal. In addition, construct the 1 and 0 sequence corresponding to each digital baseband signal, where 1 indicates that the moment is the moment of maximum correlation result, and 0 indicates that the moment is not the moment of maximum correlation result.

[0032] Specifically, Matlab was used to generate multiple known digital baseband signals with all pilot sequences inserted, some pilot sequences inserted, and no pilot sequences inserted as training data, along with multiple corresponding 1 and 0 sequences as label data. The pilot sequences were continuously inserted within a certain range of the digital baseband signal.

[0033] The pilot sequence mapping constellation diagram is BPSK raised cosine shaping. The AWGN channel signal-to-noise ratio is set to -30dB to 15dB, and the amplitude range of the received signal is set to 2048 to 256. The waveforms of the identifier correlation results at the receiver under different signal-to-noise ratios and received amplitudes are generated.

[0034] Step 2: Construct a neural network model. The input to the neural network model is the correlation result waveform output by the recognizer, and the output is a sequence of 1s and 0s representing the moment of maximum correlation. Train the neural network model using the correlation result waveform obtained in Step 1, as follows: Figure 5 As shown. During training, the loss value of the 1 and 0 sequence output by the neural network model is calculated relative to the 1 and 0 sequence constructed in step 1 using the loss function of cross-entropy or mean squared error. Based on the loss value, the parameters of the neural network model are updated using the Adam backpropagation algorithm until the loss value converges, the training ends and the parameters of the neural network model are saved, and the deep learning training is completed.

[0035] Specifically, the input layer of the neural network model is a fully connected layer, used to establish a connection between the input data and the intermediate layers. The intermediate layers consist of three concatenated convolutional layers, used to map the input layer data into data that is easier for the output layer to process, such as... Figure 6 As shown, the output layer uses the Sigmoid function to map the intermediate layer data into a sequence of 1s and 0s.

[0036] like Figure 4 As shown, the input layer of the neural network model can be a fully connected layer, used to connect the input data with the intermediate layers. The intermediate layers consist of three concatenated convolutional layers, used to map the input layer data into data that is easy for the output layer to process. The output layer is a Softmax function, used to map the intermediate layer data into a sequence of 1s and 0s.

[0037] Alternatively, the input layer of the neural network model can be a fully connected layer, used to establish a connection between the input data and the intermediate layers. The intermediate layers are dense convolutional neural networks of depth 8, used to map the input layer data into data that is easy for the output layer to process. The output layer is a sigmoid function, used to map the intermediate layer data into a sequence of 1s and 0s.

[0038] Alternatively, the input layer of the neural network model can be a fully connected layer, used to establish a connection between the input data and the intermediate layers. The intermediate layers are dense convolutional neural networks of depth 8, used to map the input layer data into data that is convenient for the output layer to process. The output layer is a softmax function, used to map the intermediate layer data into a sequence of 1s and 0s.

[0039] Step 3: In actual digital signal communication, the received signal at the signal receiving end is oversampled and sent to the recognizer to obtain the correlation waveform of the received signal. Then, the correlation waveform is sent to the trained neural network model. If the sequence output by the neural network model contains 1, such as... Figure 2 As shown, if the maximum correlation time is included, then digital communication bit synchronization is complete; otherwise, as... Figure 3 As shown, the recognizer continues to output the waveform of the received signal at the next moment and sends it to the neural network model for processing.

[0040] Specifically, the neural network development is carried out using Keras, a Python-based application programming interface. Therefore, the model parameters are directly saved as an h5 file in hierarchical data format using its editor, PyCharm Professional. This file can be opened and the data visualized using Matlab tools for subsequent implementation and porting operations.

[0041] Simulation of a deep learning-based digital communication bit synchronization method:

[0042] First, for a fixed digital signal with a received amplitude of 2048, the bit synchronization success acquisition rate as a function of the signal-to-noise ratio was simulated under a Gaussian channel. The results are as follows: Figure 9 As shown.

[0043] Secondly, random signals with amplitudes of 2048, 1792, 1536, 1280, 1024, 768, 512, and 256 were received, and their bit synchronization success acquisition rate as a function of signal-to-noise ratio was simulated under a Gaussian channel. The results are as follows: Figure 10 As shown.

[0044] Simulations were performed on traditional bit synchronization methods, assuming a desired bit synchronization success rate of 100% (successful capture is indicated by the detected maximum correlation result being the actual maximum correlation result), and a minimum tolerable signal-to-noise ratio of -15dB.

[0045] First, for a fixed digital signal with a received amplitude of 2048, the bit synchronization success acquisition rate was simulated as a function of signal-to-noise ratio and a fixed correlation threshold under a Gaussian channel. The correlation thresholds 1 to 4 decrease sequentially, and are 1.0 × 10⁻⁴ respectively. 5 3.0×10 5 5.0×10 5 and 7.0×10 5 The result is as follows Figure 7 As shown.

[0046] This demonstrates that when the relevant thresholds are set from 2 to 4, the success rate is 100% when the signal-to-noise ratio is greater than or equal to -15dB.

[0047] Secondly, for random signals with received amplitudes of 2048, 1792, 1536, 1280, 1024, 768, 512, and 256, the bit synchronization success acquisition rate as a function of signal-to-noise ratio and a fixed correlation threshold was simulated under a Gaussian channel. The correlation thresholds 1 to 5 decrease sequentially, and are 1.0 × 10⁻⁵ respectively. 5 1.8×10 5 2.6×10 5 5×10 5 and 7×105 The result is as follows Figure 8 As shown.

[0048] The simulation results show that the system cannot achieve a 100% success rate under any of the set threshold conditions. This further demonstrates that setting a fixed correlation threshold is unreliable.

[0049] contrast Figure 7 and Figure 9 It can be seen that the success rate of acquisition is better than that of the fixed threshold method under low signal-to-noise ratio, and the success rate is 100% under medium and high signal-to-noise ratio. This shows that the deep learning-based digital communication bit synchronization method can adapt to the amplitude of the pilot signal reception, demonstrating the effectiveness and robustness of the deep learning-based digital communication bit synchronization method.

[0050] contrast Figure 8 and Figure 10 This demonstrates that, compared to traditional methods, the deep learning-based digital communication bit synchronization method does not require setting a fixed correlation threshold and has a 100% success rate when the signal-to-noise ratio is higher than -21dB. In other words, it is not sensitive to the received amplitude of the pilot signal and can be applied to burst communication, reflecting the robustness of the deep learning-based digital communication bit synchronization method to the received amplitude of the pilot signal.

[0051] Therefore, this method utilizes a data-driven deep learning algorithm to train a neural network model using large-scale data. By processing the nonlinearity of the data model, it breaks away from traditional bit synchronization methods derived from linear, stationary mathematical models with Gaussian statistical properties at the physical layer. It can adaptively determine the maximum correlation result when the amplitude of the received signal is unknown during communication.

[0052] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and does not limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and scope of the present invention as defined in the appended claims should be included within the protection scope of the present invention.

Claims

1. A digital communication bit synchronization method based on deep learning, characterized in that, Includes the following steps: Step 1: Input multiple known digital baseband signals containing all pilot sequences, some pilot sequences, and none pilot sequences into the recognizer, and the recognizer outputs the relevant result waveform for each digital baseband signal; In addition, a sequence of 1s and 0s corresponding to each digital baseband signal is constructed, where 1 indicates that the moment is the moment of maximum correlation result, and 0 indicates that the moment is not the moment of maximum correlation result; Step 2: Construct a neural network model. The input to the neural network model is the correlation result waveform output by the recognizer, and the output is the 1 and 0 sequence representing the moment of maximum correlation. Use the correlation result waveform obtained in Step 1 to train the neural network model. During the training process, calculate the loss value of the 1 and 0 sequence output by the neural network model relative to the 1 and 0 sequence constructed in Step 1 using the loss function of cross-entropy or mean squared error. Based on the loss value, update the parameters of the neural network model using the Adam backpropagation algorithm until the loss value converges. End the training and save the parameters of the neural network model to complete the deep learning training. Step 3: In actual digital signal communication, the received signal at the signal receiving end is oversampled and sent to the recognizer to obtain the correlation result waveform of the received signal. Then, the correlation result waveform is sent to the trained neural network model. If the sequence output by the neural network model contains 1, that is, it contains the maximum correlation moment, then the digital communication bit synchronization is completed. Otherwise, the recognizer continues to output the correlation result waveform of the received signal at the next moment and sends it to the neural network model for processing.

2. The deep learning-based digital communication bit synchronization method according to claim 1, characterized in that, In step 1, the pilot sequence is continuously inserted within a certain range of the digital baseband signal.

3. The deep learning-based digital communication bit synchronization method according to claim 1, characterized in that, In step 2, the input layer of the neural network model is a fully connected layer, which is used to associate the input data with the intermediate layers. The intermediate layers are composed of three concatenated convolutional layers, which are used to map the input layer data into data that is convenient for the output layer to process. The output layer is a sigmoid function, which is used to map the intermediate layer data into a sequence of 1s and 0s.

4. The deep learning-based digital communication bit synchronization method according to claim 1, characterized in that, In step 2, the input layer of the neural network model is a fully connected layer, which is used to connect the input data with the intermediate layers. The intermediate layers are composed of three concatenated convolutional layers, which are used to map the input layer data into data that is easy for the output layer to process. The output layer is a Softmax function, which is used to map the intermediate layer data into a sequence of 1s and 0s.

5. The deep learning-based digital communication bit synchronization method according to claim 1, characterized in that, In step 2, the input layer of the neural network model is a fully connected layer, which is used to associate the input data with the intermediate layers. The intermediate layers are dense convolutional neural networks with a depth of 8, which are used to map the input layer data into data that is easy for the output layer to process. The output layer is a sigmoid function, which is used to map the intermediate layer data into a sequence of 1s and 0s.

6. The digital communication bit synchronization method based on deep learning according to claim 1, characterized in that, In step 2, the input layer of the neural network model is a fully connected layer, which is used to associate the input data with the intermediate layers. The intermediate layers are dense convolutional neural networks with a depth of 8, which are used to map the input layer data into data that is convenient for the output layer to process. The output layer is a Softmax function, which is used to map the intermediate layer data into a sequence of 1s and 0s.

Images