Improved method for fast decoding of morse signals using ctc

Abstract

The application discloses a method for quickly decoding a morse signal of a CTC, comprising the following steps: performing short-time Fourier transform on extracted morse signal data, obtaining a time-frequency matrix, extracting one-dimensional features of the morse signal in the time-frequency matrix, and constructing a training set by corresponding labels; constructing a deep learning framework, including a deep residual shrinkage module, a full connection layer and a softmax structure; outputting the deep residual shrinkage module through the full connection layer and the softmax structure, taking the output as a sub-model, calculating the CTC loss of the sub-model, weighting the CTC loss of the full model, taking the weighted CTC loss as a loss function, and training a network model; inputting the one-dimensional features of the extracted morse signal time-frequency matrix into the trained network model, performing CTC decoding-based post-processing on the output, and obtaining a decoding result. The application combines the CTC loss of the full model with the loss of the sub-model for training, has very small calculation overhead, and can effectively improve the decoding accuracy.

Description

Technical Field

[0001] This invention relates to an improved CTC method for fast Morse signal decoding, belonging to the field of Morse telegraph automatic decoding technology. Background Technology

[0002] Morse code is a communication method that uses a combination of dots (short signals) and dashes (long signals) to represent letters, numbers, and symbols. It requires very little bandwidth, making it suitable for low-bandwidth environments. Furthermore, Morse code rules are simple and can be mastered after a short training period. In the event of a failure in modern communication systems, Morse code serves as an important emergency communication method. Morse code maintains a high recognition rate even in noisy environments due to its distinct signal characteristics and strong anti-interference capabilities. Morse code can be transmitted over long distances via radio waves, making it particularly suitable for long-distance communication, such as in maritime and aviation fields. With the development of radio communication, Morse code has become an important encoding method in radio communication, widely used in military and civilian communications. However, traditional Morse code decoding relies heavily on the operator's auditory recognition ability, requiring specialized training to master the transmission and reception of Morse code. In practice, operator fatigue or inattention can lead to decoding errors, and the operator's decoding speed is limited, failing to meet the demands of high-speed communication. Therefore, research on end-to-end automatic Morse signal decoding methods is both valuable and necessary.

[0003] Traditional machine learning methods, such as Support Vector Machines (SVM) and clustering algorithms, can automatically decode Morse signals at high signal-to-noise ratios (SNR). However, these methods are sensitive to noise, and their decoding accuracy drops significantly at low SNRs. Currently, mainstream automatic decoding methods are based on deep learning. After preprocessing the extracted Morse signal data, short-time Fourier transform or wavelet transform is used to obtain the time-frequency feature map of the Morse signal, which is then used as network input. Features are extracted using deep learning models such as Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), or Long Short-Term Memory Networks (LSTM). During training, the Connectionist Temporal Classification (CTC) loss function is used, and its greedy search algorithm can solve the problem that the length of the input feature sequence is greater than the length of the output message during decoding.

[0004] It is worth noting that current deep learning-based decoding methods use time-frequency graphs derived from short messages as network inputs. While these graphs offer limited information and are easy to train and infer, they cannot separate or concatenate long messages into short ones. Furthermore, the fixed-size time-frequency graphs input to the network significantly limit the message length, making it impossible to decode variable-length Morse signals. Additionally, due to the extremely small bandwidth of Morse signals, frequency bands without Morse signal information occupy a large area in the time-frequency graph. This redundant information undoubtedly increases the computational complexity of the inference process, potentially limiting real-time processing capabilities and consuming unnecessary computational resources. In complex real-world applications, decoding may encounter problems such as low signal-to-noise ratio and large-scale fading of Morse signals, affecting the extraction of Morse signal features and leading to incorrect decoding results. Therefore, it is necessary to design an automatic decoding method with noise robustness and fading resistance. The deep learning architecture needs to supervise the features extracted by networks of different depths to ensure the reliability and accuracy of the decoding results. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide an improved CTC fast decoding method for Morse signals, which can effectively extract the deep features of the model signal and improve the decoding accuracy in scenarios such as low signal-to-noise ratio and large-scale fading of Morse signals.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] An improved method for fast Morse signal decoding of CTC includes the following steps:

[0008] Step 1: Obtain the Morse signal and perform a short-time Fourier transform to obtain the time-frequency matrix. Extract the one-dimensional features of the Morse signal from the time-frequency matrix and construct a training set using the one-dimensional features of the Morse signal and the corresponding labels of the Morse signal.

[0009] Step 2: Construct a deep learning network model, including a deep residual shrinking module, first and second fully connected layers, and first and second softmax functions. The deep residual shrinking module, the first fully connected layer, and the first softmax function are connected sequentially. The deep residual shrinking module includes a one-dimensional convolutional input layer and several basic modules connected sequentially. After the m-th basic module, the second fully connected layer and the second softmax function are connected sequentially, where m = 2, ..., n-1, and n is the number of basic modules.

[0010] Step 3: Take the output of the m-th basic module after passing through the second fully connected layer and the second softmax function as the sub-model, calculate the CTC loss of the sub-model, take the output of the last basic module after passing through the first fully connected layer and the first softmax function as the full model, and use the weighted sum of the CTC loss of the sub-model and the CTC loss of the full model as the loss function. Use the training set to train the deep learning network model to obtain the Morse signal decoding model.

[0011] Step 4: Extract the one-dimensional features of the time-frequency matrix of the Morse signal to be decoded, input them into the Morse signal decoding model, and perform CTC decoding post-processing on the output of the model to obtain the decoding result of the Morse signal to be decoded.

[0012] As a preferred embodiment of the present invention, in step 1, the one-dimensional features of the Morse signal in the time-frequency matrix are extracted as follows:

[0013] Taking the square of the short-time Fourier transform of the Morse signal x(n) yields the time-frequency matrix X(k,l) of x(n), where k and l represent the column and row indices of the matrix, i.e., the frequency axis and time axis, respectively. The calculation formula is as follows:

[0014]

[0015] Where w(n) is the window function, N is the length of the window function, L is the window step size, and K is the number of points in the discrete Fourier transform;

[0016] The one-dimensional feature of a Morse signal is the row vector corresponding to the carrier frequency coordinate in the time-frequency matrix, i.e., X(k c ,l), where k c Satisfying k c =N·f c / f s f s f is the sampling frequency of the Morse signal. c This is the Morse code carrier frequency.

[0017] As a preferred embodiment of the present invention, in step 2, the basic module includes a first convolutional layer, a second convolutional layer, an SE channel attention mechanism, and a soft thresholding processing unit connected in sequence. The SE channel attention mechanism includes an average pooling layer, a third fully connected layer, a fourth fully connected layer, and a sigmoid activation function.

[0018] The input x of the basic module is obtained after passing through the first and second convolutional layers. After the average pooling layer, the dimensionality is reduced to a one-dimensional vector X. This one-dimensional vector X then passes through the third and fourth fully connected layers, and is activated by the sigmoid activation function to obtain a one-dimensional vector Y. Multiplying X and Y yields the threshold T for each channel, where T is a 1*1*C vector, and C is... The number of channels; then use the threshold T to... The soft thresholding process is calculated using the following formula:

[0019]

[0020] Among them, y thresholded for The result is then soft-thresholded using a threshold T; finally, a cross-layer identity connection is used to select y. thresholded After being added to the input x, and then processed by the ReLU activation function, it becomes the output of the basic module.

[0021] In a preferred embodiment of the present invention, the loss function in step 3 is calculated using the following formula:

[0022]

[0023] in, This represents the loss function used during training, where w represents the weights. The loss function for the entire model. This is the loss function for the sub-model.

[0024] As a preferred embodiment of the present invention, step 4, the post-processing of CTC decoding, specifically includes:

[0025] The number of input channels of the first fully connected layer is the number of channels of the Morse signal depth feature sequence output by the deep residual shrinkage module, and the number of output channels is the number of categories of the label characters plus 1. Then, the probability distribution matrix of each character at each time step is obtained by the first softmax function.

[0026] Using a greedy decoding algorithm, the character with the highest probability value at each time step is selected from the probability distribution matrix output by the first softmax function. After removing empty characters and duplicate characters, the decoding result is obtained.

[0027] A computer device includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the steps of the improved CTC Morse signal fast decoding method.

[0028] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the improved CTC Morse signal fast decoding method.

[0029] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0030] 1. This invention reduces the use of redundant information by extracting a one-dimensional feature sequence of the Morse signal as the input of the network model, which can effectively save computing resources, improve inference speed, and enable Morse signal decoding of arbitrary length.

[0031] 2. This invention integrates soft thresholding processing through the SE (Squeeze-and-Excitation) channel attention mechanism, which can effectively reduce noise interference and make the model noise robust when facing extreme noise.

[0032] 3. This invention uses a deep residual shrinkage network with many basic modules stacked together to extract the deep features of the sequence, effectively reducing the impact of fading on the decoding results.

[0033] 4. This invention uses weighted intermediate CTC loss and final layer CTC loss to supervise the features extracted by networks of different depths, which not only improves training efficiency but also enables the model to focus on useful features earlier, thereby achieving higher decoding accuracy. Attached Figure Description

[0034] Figure 1 This is the overall flowchart of the improved CTC Morse signal fast decoding method of the present invention;

[0035] Figure 2 This is an example diagram of the time and frequency of a Morse signal;

[0036] Figure 3 This is an example of a visualization of a one-dimensional Morse code sequence;

[0037] Figure 4 This is a basic module structure diagram;

[0038] Figure 5 This is a diagram of the CTC loss calculation structure;

[0039] Figure 6 It is the overall process of reasoning and decoding. Detailed Implementation

[0040] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0041] like Figure 1 As shown, this invention proposes an improved method for fast decoding of Morse signals using CTC, with the following specific steps:

[0042] Step 1: Perform a short-time Fourier transform on the extracted Morse signal data to obtain the time-frequency matrix. Extract the one-dimensional features of the Morse signal in the time-frequency matrix as the input of the neural network, and use the corresponding labels to form the training set.

[0043] In step 1, the method for extracting the one-dimensional features of the Morse signal from the time-frequency matrix is ​​as follows:

[0044] The time-frequency matrix X(k,l) is obtained by squaring the short-time Fourier transform of the Morse signal x(n). The matrix can then be visualized and plotted as follows: Figure 2 As shown, k and l represent the column (frequency axis) and row (time axis) indices of the matrix, respectively, and are calculated using the following formula:

[0045]

[0046] Where w(n) is the added window function, N is the length of the window function, K is the number of points in the discrete Fourier transform, and L is the window step size. The one-dimensional feature of the Morse signal is the row vector corresponding to the carrier frequency coordinate in the time-frequency matrix, i.e., X(k... c ,l). Where, k c Satisfying k c =N·f c / f s f s f is the sampling frequency of the Morse signal. c This represents the Morse code carrier frequency. The visualization results are as follows: Figure 3 As shown, the Morse signal sequence after extracting one-dimensional features is a row of the time-frequency matrix.

[0047] Step 2: Construct a deep learning framework, which includes a deep residual shrinking module, first and second fully connected layers, and first and second softmax functions.

[0048] The deep learning framework in step 2 is as follows Figure 5 As shown, the description is as follows:

[0049] The deep residual shrinking module consists of a one-dimensional convolutional input layer and many basic modules. Each basic module is composed of three fused parts: a residual network module, a Squeeze-and-Excitation (SE) channel attention mechanism, and a soft thresholding unit. The one-dimensional sequence of the Morse signal is input into the deep residual shrinking network to obtain its high-channel depth features; the output is denoted as the Morse signal depth feature sequence.

[0050] In the first fully connected layer and the first softmax function, the number of input channels of the first fully connected layer is the number of channels of the Morse signal depth feature sequence output by the deep residual shrinkage module, and the number of output channels is the number of categories of the label characters + 1. Then, the probability distribution matrix of each character at each time step is obtained by the first softmax function.

[0051] Basic module structure as follows Figure 4 As shown, the description is as follows:

[0052] Let x be the input of the basic module. After x passes through two convolutional layers, it becomes... Then, based on the SE channel attention mechanism, The vector is reduced to a 1x1xC vector by global average pooling. This vector is then fed into two fully connected layers and activated by the sigmoid function. The resulting 1x1xC vector is multiplied by the previously pooled 1x1xC vector to obtain the threshold for each channel, denoted as T, where T is a 1x1xC vector and C is a 1x1xC vector. The number of channels. Then use the threshold T to... The soft thresholding process is calculated using the following formula:

[0053]

[0054] Among them, y thresholded for The result is soft-thresholded using a threshold T. Finally, a cross-layer identity connection is used to take y. thresholded After being added to the input x, and then processed by the ReLU activation function, it becomes the output of the basic module.

[0055] Step 3: The basic module in the middle of the deep residual shrinkage module is output through the second fully connected layer and the second softmax function as a sub-model. The CTC loss of the sub-model is calculated and weighted with the CTC loss of the whole model as the loss function to train the network model.

[0056] In the intermediate layer of the basic stacked modules of the deep residual shrinkage module, after passing through the second fully connected layer and the second softmax function, the output is denoted as the sub-model, and the loss function is... Let the loss function of the full model be... The total loss function for training is a weighted sum of the two losses, calculated as follows:

[0057]

[0058] in, Here, w represents the loss function used in the final training, and w represents the weights.

[0059] Step 4: Input the one-dimensional features of the extracted Morse signal time-frequency matrix into the trained network model, and perform post-processing based on CTC decoding on the network output to obtain the decoding result of the Morse signal.

[0060] CTC decoding post-processing uses a greedy decoding algorithm. From the probability distribution matrix output by the first softmax function, it selects the character with the highest probability value at each time step and outputs it. After removing empty and duplicate characters, the decoded result is obtained. The complete reasoning process from Morse signal audio to the decoded result is as follows: Figure 6 As shown.

[0061] Based on the same inventive concept, embodiments of this application provide a computer 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, it implements the steps of the aforementioned improved CTC Morse signal fast decoding method.

[0062] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned improved CTC Morse signal fast decoding method.

[0063] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0064] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0065] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0066] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0067] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. An improved CTC Morse signal fast decoding method, characterized in that, Includes the following steps: Step 1: Obtain the Morse signal and perform a short-time Fourier transform to obtain the time-frequency matrix. Extract the one-dimensional features of the Morse signal from the time-frequency matrix and construct a training set using the one-dimensional features of the Morse signal and the corresponding labels of the Morse signal. Step 2: Construct a deep learning network model, including a deep residual shrinking module, first and second fully connected layers, and first and second softmax functions. The deep residual shrinking module, the first fully connected layer, and the first softmax function are connected sequentially. The deep residual shrinking module includes a one-dimensional convolutional input layer and several basic modules connected sequentially. After the m-th basic module, the second fully connected layer and the second softmax function are connected sequentially, where m = 2, ..., n-1, and n is the number of basic modules. In step 2, the basic module includes a first convolutional layer, a second convolutional layer, an SE channel attention mechanism, and a soft thresholding processing unit connected in sequence. The SE channel attention mechanism includes an average pooling layer, a third fully connected layer, a fourth fully connected layer, and a sigmoid activation function. The input x of the basic module is obtained after passing through the first and second convolutional layers. After the average pooling layer, the dimensionality is reduced to a one-dimensional vector X. This one-dimensional vector X then passes through the third and fourth fully connected layers, and is activated by the sigmoid activation function to obtain a one-dimensional vector Y. Multiplying X and Y yields the threshold T for each channel, where T is a 1*1*C vector, and C is... The number of channels; then use the threshold T to... The soft thresholding process is calculated using the following formula: Among them, y thresholded for The result is then soft-thresholded using a threshold T; finally, a cross-layer identity connection is used to select y. thresholded After being added to the input x, the result is processed through the ReLU activation function and used as the output of the basic module. Step 3: Take the output of the m-th basic module after passing through the second fully connected layer and the second softmax function as the sub-model, calculate the CTC loss of the sub-model, take the output of the last basic module after passing through the first fully connected layer and the first softmax function as the full model, and use the weighted sum of the CTC loss of the sub-model and the CTC loss of the full model as the loss function. Use the training set to train the deep learning network model to obtain the Morse signal decoding model. Step 4: Extract the one-dimensional features of the time-frequency matrix of the Morse signal to be decoded, input them into the Morse signal decoding model, and perform CTC decoding post-processing on the output of the model to obtain the decoding result of the Morse signal to be decoded.

2. The improved CTC Morse signal fast decoding method according to claim 1, characterized in that, In step 1, the one-dimensional features of the Morse signal in the time-frequency matrix are extracted, as follows: Taking the square of the short-time Fourier transform of the Morse signal x(n) yields the time-frequency matrix X(k,l) of x(n), where k and l represent the column and row indices of the matrix, i.e., the frequency axis and time axis, respectively. The calculation formula is as follows: Where w(n) is the window function, N is the length of the window function, L is the window step size, and K is the number of points in the discrete Fourier transform; The one-dimensional feature of a Morse signal is the row vector corresponding to the carrier frequency coordinate in the time-frequency matrix, i.e., X(k c ,l), where k c Satisfying k c =N·f c / f s f s f is the sampling frequency of the Morse signal. c This is the Morse code carrier frequency.

3. The improved CTC Morse signal fast decoding method according to claim 1, characterized in that, In step 3, the loss function is calculated using the following formula: in, This represents the loss function used during training, where w represents the weights. The loss function for the entire model. This is the loss function for the sub-model.

4. The improved CTC Morse signal fast decoding method according to claim 1, characterized in that, In step 4, the post-processing of CTC decoding specifically includes: The number of input channels of the first fully connected layer is the number of channels of the Morse signal depth feature sequence output by the deep residual shrinkage module, and the number of output channels is the number of categories of the label characters plus 1. Then, the probability distribution matrix of each character at each time step is obtained by the first softmax function. Using a greedy decoding algorithm, the character with the highest probability value at each time step is selected from the probability distribution matrix output by the first softmax function. After removing empty characters and duplicate characters, the decoding result is obtained.

5. A computer device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the improved CTC Morse signal fast decoding method as described in any one of claims 1 to 4.

6. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the improved CTC Morse signal fast decoding method as described in any one of claims 1 to 4.

Images