An end-to-end hand-generated constant-amplitude telegraph decoding system based on deep learning

Abstract

The invention discloses an end-to-end hand-sent constant-amplitude telegraph decoding system based on deep learning, the system comprises: a preprocessing module for denoising a signal and obtaining its time-frequency matrix; an end-to-end recognition model for extracting features from the time-frequency matrix obtained after preprocessing, sequentially comprising an FC-N layer, a Bi-LSTM layer, a fully connected layer, a Softmax layer and a CTC layer, and obtaining an output sequence of each frame after model processing, and then obtaining a predicted message through integration. The invention can accurately decode hand-sent constant-amplitude telegraphs.

Description

Technical Field

[0001] The present invention belongs to the technical field of hand-transmitted constant-amplitude telegraph decoding, and in particular relates to an end-to-end hand-transmitted constant-amplitude telegraph decoding system based on deep learning. Background Art

[0002] Continuous Wave (CW) is essentially an amplitude keying modulation signal, which generally uses Morse code to send information. It controls the transmitter through the key to generate short signals "." and long signals "_", and uses different combinations to represent different characters, thus forming phrases and sentences. CW signals have the advantages of narrow signal bandwidth, strong anti-interference ability, and simple coding method. As an important shortwave wireless communication method, it has been active in aviation, maritime, military communications and other fields.

[0003] Constant amplitude telegraph decoding is used to convert CW signals into corresponding messages. At present, there are two main methods for decoding constant amplitude telegraphs. One is the statistical judgment method, the most typical of which is based on traditional machine learning methods, such as support vector machines, clustering, etc. It first extracts the time-frequency features of the signal to obtain the time-frequency diagram of the signal, and then identifies the points, dashes and different intervals through classification algorithms such as support vector machines and clustering. Finally, the identified dot-dash sequence is converted into a message according to the codebook. The other is the CRNN method based on deep learning, which uses CNN (Convolution Neural Networks) and RNN (Recurrent Neural Networks) cascade to extract the features and contextual information of the Morse code, and then uses the fully connected layer and the CTC (Connectionist Temporal Classification) loss function as the objective function to obtain the posterior probability of the output sequence, and then obtain the predicted message.

[0004] Both of the above decoding methods have certain defects. For traditional machine learning methods, only the characteristics of the time domain code length of the Morse signal are used, and the associated information of the Morse code context is ignored. When the code length jitters, the effectiveness of this type of method decreases significantly. For the latter, although the model can achieve a good convergence effect, it does not consider the channel transmission characteristics of the CW signal, does not preprocess the CW signal, and directly calculates the time-frequency matrix of the signal. At the same time, this method uses CNN to extract features from the time-frequency matrix. While extracting frequency domain features, it also reduces the dimension of the time domain and does not expand the number of channels during the convolution operation, which causes a loss of time domain information to a certain extent, thereby weakening the use of RNN for the time domain features of the Morse code. Summary of the invention

[0005] The present invention aims to provide an end-to-end hand-transmitted constant-amplitude telegraph decoding system based on deep learning, so as to solve the technical problems that the traditional machine learning decoding method is not robust to the long jitter of the hand-transmitted telegraph code and the time domain feature loss of Morse code in the existing neural network decoding method.

[0006] In order to solve the above technical problems, the specific technical solutions of the present invention are as follows:

[0007] An end-to-end hand-transmitted constant-amplitude telegraph decoding system based on deep learning, including a preprocessing module and an end-to-end recognition model;

[0008] The preprocessing module includes signal limiting and short-time Fourier transform; after the input CW signal is processed by the limiting and short-time Fourier transform, a time-frequency matrix of the signal is obtained, and the time dimension of the time-frequency matrix is ​​the number of frames obtained by the short-time Fourier transform;

[0009] The end-to-end recognition model includes an FC-N layer, a Bi-LSTM layer, a fully connected layer, a Softmax layer and a CTC layer; after the time-frequency matrix of the Morse signal is obtained, it is processed by the FC-N layer, the Bi-LSTM layer, the fully connected layer, the Softmax layer and the CTC layer to obtain the probability distribution of the output sequence of each frame, and the predicted output of each frame is integrated into a predicted message according to the probability distribution.

[0010] Furthermore, the FC-N layer includes a fully connected layer, and the output of each fully connected layer is activated by a Relu function. In the process of feature extraction using the FC-N layer, the time domain feature dimension can be kept unchanged, and the features of the Morse code can be extracted in the frequency domain, so that the subsequent network can make full use of the time domain information of the Morse code.

[0011] Furthermore, the Bi-LSTM layer includes a multi-layer Bi-LSTM structure.

[0012] The end-to-end hand-transmitted constant-amplitude telegraph decoding system based on deep learning of the present invention has the following advantages:

[0013] The end-to-end recognition model of the present invention adopts the FC-N-Bi-LSTM-CTC framework to directly transcribe the CW signal to be recognized into a message without the need for table lookup mapping operations. At the same time, the time dimension is kept unchanged during the feature extraction process, and the contextual association information of the Morse code can be fully utilized. In addition, CTC is used to automatically solve the alignment problem between the input sequence and the predicted label, and the output of each frame is determined by the probability distribution of the model output, and finally the output sequence corresponding to each frame is integrated into a predicted message. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a structural principle diagram of an end-to-end hand-transmitted constant-amplitude telegraph decoding system based on deep learning of the present invention;

[0015] Figure 2 It is a schematic diagram of CW signal preprocessing of the present invention;

[0016] Figure 3 It is the structural principle diagram of FC-N of the present invention;

[0017] Figure 4 It is a schematic diagram of the output of the end-to-end recognition model of the present invention;

[0018] Figure 5 It is a schematic diagram of the processing process of the end-to-end hand-transmitted constant-amplitude telegraph decoding system based on deep learning of the present invention; DETAILED DESCRIPTION

[0019] In order to better understand the purpose, structure and function of the present invention, the following is a further detailed description of an end-to-end hand-transmitted constant-amplitude telegraph decoding system based on deep learning in conjunction with the accompanying drawings.

[0020] like Figure 1 As shown, the end-to-end constant amplitude telegraph decoding system based on deep learning of the present invention comprises:

[0021] Preprocessing, including signal limiting and short-time Fourier transform in sequence. After the input constant amplitude telegraph (CW) signal is processed by the limiting and short-time Fourier transform, the time-frequency matrix of the signal is obtained;

[0022] The end-to-end recognition model includes FC-N (Full Connection-N, N fully connected) layers, Bi-LSTM layers (Bi-directional Long Short-Term Memory, bidirectional long short-term memory network), fully connected layers, Softmax layers and CTC (Connectionist Temporal Classification, connection temporal classification) layers. After the two-dimensional time-frequency features of the Morse signal are extracted, the probability distribution of the output of different characters on each frame is obtained after being processed by the FC-N layer, Bi-LSTM layer, fully connected layer, Softmax layer and CTC layer. Finally, the output sequence on each frame is integrated into the final prediction label according to the probability distribution.

[0023] For the preprocessing module, in the present invention, Figure 2 As shown, the input of the model is a CW signal. The input signal is limited to filter out noise, and then the processed signal is short-time Fourier transformed to obtain a two-dimensional time-frequency matrix of the CW signal.

[0024] For the end-to-end recognition model, in the present invention, the input is the time-frequency matrix obtained by preprocessing. The front end of the model adopts the FC-N structure, such as Figure 3 As shown in the figure, this network structure uses fully connected layers and Relu activation functions multiple times to perform feature extraction and dimension transformation while keeping the time dimension unchanged, and outputs the feature sequence of Morse code. Then, the feature sequence passes through the Bi-LSTM layer, and inputs the features of each frame in turn to learn the contextual association information of the Morse code. Next, a fully connected layer is used to obtain the unnormalized probability distribution of different characters output at each moment, that is, the serialized features output by Bi-LSTM are reduced in dimension, and the dimension of the frequency domain is reduced to the minimum number of modeling units (consisting of each label character and a character representing a blank) while keeping the time dimension unchanged. As shown in the figure Figure 4 As shown in the figure, after the Softmax layer, a probability distribution matrix is ​​output. In this matrix, each column represents the probability of a corresponding character on the frame. The first row of each column represents the probability of outputting a blank character (indicated by "—" in the figure), and the following rows represent the probability of outputting other corresponding characters. Then, CTC is used to take the difference between the predicted output and the true label at each moment as the loss function of the end-to-end decoding model to achieve automatic alignment of the input sequence and the output label. Finally, the probability distribution matrix is ​​used to obtain the output sequence corresponding to each frame, and the final predicted message is obtained by integration.

[0025] The amplitude limiting process uses the equal amplitude characteristic of the CW signal to limit the amplitude of the CW signal to be decoded, so as to filter out some noise in the signal. Assuming that the received CW signal is r(n) and the threshold is A, the signal r′(n) after amplitude limiting can be expressed as

[0026]

[0027] The short-time Fourier transform is to perform frame-by-frame windowing on the signal r′(n), and perform discrete Fourier transform on each frame to obtain the time-frequency matrix X(k,l) of the CW signal, which can be expressed as

[0028]

[0029] Where N is the length of the window function, K is the number of discrete Fourier transform points, and L is the length of each step of the window function.

[0030] The FC-N layer includes N fully connected layers. The fully connected layer has the function of dimensional transformation, which can record useful information while realizing the dimensional transformation of the feature matrix. At the same time, after being processed by the FC-N layer, the time dimension of the Morse code feature matrix remains unchanged, ensuring that the subsequent Bi-LSTM layer fully utilizes the time domain information.

[0031] Among them, the Bi-LSTM layer is an improvement of the bidirectional RNN structure. It solves the problem of gradient explosion or gradient disappearance of the RNN network by introducing a gating mechanism. It is composed of multiple layers of Bi-LSTM and can selectively forget the previous accumulated information or add new information, so that the model can fully learn the contextual association information of the Morse code.

[0032] Among them, after being processed by the Softmax layer and the CTC layer, the probability of outputting different characters in each frame is obtained. According to the probability relationship, the predicted characters output by each frame can be obtained, and then integrated to obtain the predicted message.

[0033] The mapping relationship between the complete input and output of the present invention is as follows Figure 5 shown.

[0034] The preprocessing process of the present invention adopts the method of limiting and short-time Fourier transform, and uses the amplitude characteristics of the CW signal itself to limit the signal to be identified, effectively filtering out some noise. At the same time, the time-frequency matrix of the Morse signal is obtained by short-time Fourier transform for learning and training the recognition model.

[0035] The end-to-end recognition model of the present invention adopts the FC-N-Bi-LSTM-CTC framework to directly transcribe the CW signal to be recognized into a message without the need for table lookup mapping operations. At the same time, the time dimension is kept unchanged during the feature extraction process, and the contextual association information of the Morse code can be fully utilized. In addition, CTC is used to automatically solve the alignment problem between the input sequence and the predicted label, and the output of each frame is determined by the probability distribution of the model output, and finally the output sequence corresponding to each frame is integrated into a predicted message.

[0036] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. An end-to-end hand-transmitted constant-amplitude telegraph decoding system based on deep learning, characterized in that: Includes preprocessing module and end-to-end recognition model; The preprocessing module includes signal limiting and short-time Fourier transform; after the input CW signal is processed by the limiting and short-time Fourier transform, a time-frequency matrix of the signal is obtained, and the time dimension of the time-frequency matrix is ​​the number of frames obtained by the short-time Fourier transform; The end-to-end recognition model includes an FC-N layer, a Bi-LSTM layer, a fully connected layer, a Softmax layer, and a CTC layer; after the time-frequency matrix of the Morse signal is obtained, it is processed by the FC-N layer, the Bi-LSTM layer, the fully connected layer, the Softmax layer, and the CTC layer to obtain a probability distribution matrix of each frame output sequence, and the predicted output of each frame is integrated into a predicted message according to the probability distribution matrix; For the end-to-end recognition model, the input is the time-frequency matrix obtained by preprocessing; the front end of the end-to-end recognition model uses the FC-N layer, which uses the fully connected layer and the Relu activation function multiple times to extract features and transform dimensions while keeping the time dimension unchanged, and outputs the feature sequence of the Morse code; then, the feature sequence passes through the Bi-LSTM layer, and inputs the features of each frame in turn to learn the contextual association information of the Morse code, and then, through a fully connected layer, the unnormalized probability distribution of the output of different characters at each moment is obtained, that is, the serialized features of the Bi-LSTM output are reduced in dimension, while keeping the time dimension unchanged. Under the condition of keeping the time dimension unchanged, the dimension of the frequency domain is reduced to the minimum number of modeling units; after passing the Softmax layer, a probability distribution matrix is ​​output, in which each column represents the probability of a certain character corresponding to the frame, the first row of each column represents the probability of outputting a blank character, and the following rows represent the probability of outputting other corresponding characters in turn; then, the CTC layer is used to take the difference between the predicted output and the true label at each moment as the loss function of the end-to-end decoding model to achieve automatic alignment of the input sequence and the output label; finally, the probability distribution matrix is ​​used to obtain the output sequence corresponding to each frame, and the final predicted message is obtained by integration; Amplitude limiting processing uses the equal amplitude characteristic of CW signals to limit the amplitude of the CW signal to be decoded, so as to filter out some noise in the signal. Short-time Fourier transform is to perform frame-by-frame windowing on the signal, and perform discrete Fourier transform on each frame to obtain the time-frequency matrix of the CW signal.

2. The end-to-end hand-transmitted constant-amplitude telegraph decoding system based on deep learning according to claim 1 is characterized in that: The FC-N layer includes a fully connected layer, and the output of each fully connected layer is activated by a Relu function. In the process of extracting features using the FC-N layer, the time domain feature dimension can be kept unchanged, and the features of the Morse code can be extracted in the frequency domain, so that the subsequent network can make full use of the time domain information of the Morse code.

3. The end-to-end hand-transmitted constant-amplitude telegraph decoding system based on deep learning according to claim 2 is characterized in that: The Bi-LSTM layer includes a multi-layer Bi-LSTM structure.

Images