A Deep Learning-Based Morse Signal Arrival Detection Method
By combining deep learning and narrowband filtering, the problems of automation and real-time performance in Morse signal arrival detection were solved, achieving high-accuracy detection in noisy environments and improving the efficiency and accuracy of Morse signal reception.
Patent Information
- Application Number
- CN202411049616.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-01
- Publication Date
- 2025-10-31
- Estimated Expiration
- 2044-08-01
AI Technical Summary
Existing Morse signal arrival detection methods rely on manual monitoring, which is inefficient and inaccurate in noisy environments, making it difficult to achieve automatic decoding and real-time detection.
A deep learning-based approach is adopted to detect Morse signals through frame-by-frame decision-making and narrowband filtering. A deep neural network is used to determine the arrival and end points of the signal. Feature extraction and classification are performed by combining a one-dimensional convolutional module, a Bi-GRU module, and a fully connected structure.
It achieves high accuracy and real-time Morse signal arrival detection, effectively filters noise in noisy environments, and improves the robustness and decision accuracy of the system.
Smart Images
Figure CN118887974B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a Morse signal arrival detection method based on deep learning, belonging to the field of shortwave communication technology. Background Technology
[0002] Morse communication continues to play a vital role as an emergency communication method. Currently, the reception of Morse signals primarily relies on manual monitoring, meaning that operators must constantly monitor shortwave radios for incoming telegram signals. Once a Morse signal is detected, the operator needs to perform highly focused manual decoding. Therefore, Morse signal arrival detection methods are not only a key prerequisite for achieving automatic decoding, but they can also significantly improve reception efficiency and accuracy while reducing the workload of manual monitoring. By automatically detecting the arrival of Morse signals, the automatic decoding system can be triggered promptly, thereby enabling rapid and accurate transmission of telegram information.
[0003] Currently, research on arrival detection technology for morse signals is limited. Voice Activity Detection (VAD) can be used to distinguish speech from silence in audio signals. VAD methods mainly include those based on energy, frequency domain features, and statistical models. Energy-based methods detect speech activity by comparing signal energy with a threshold; they are simple to implement but perform poorly in noisy environments. Frequency domain feature-based methods utilize Fourier transform and Mel-frequency cepstral coefficients (MFCCs) to improve detection performance in noisy environments. Statistical model-based methods, such as Gaussian Mixture Models (GMMs) and Hidden Markov Models (HMMs), achieve more accurate detection by modeling the statistical properties of speech and noise. In recent years, deep learning methods have been widely used in VAD. Deep learning methods automatically extract and learn complex features in speech signals through neural network models (such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Long Short-Term Memory Networks (LSTMs). These methods can handle large amounts of training data, adapt to different speech and noise environments, and improve detection accuracy and robustness. Deep learning methods have performed exceptionally well in VAD, especially in high-noise environments, significantly reducing false positive and false negative rates, making them one of the most advanced speech activity detection technologies available today. Summary of the Invention
[0004] To address the technical issue, this invention proposes a Morse signal arrival detection method based on deep learning, aiming to achieve arrival detection and termination point determination of Morse signals.
[0005] To address the aforementioned problems, this invention proposes a deep learning-based Morse signal arrival detection method. This method detects Morse signals by making decisions frame-by-frame (1 second) in a manner that provides excellent real-time performance. Narrowband filtering is used for signal preprocessing to effectively filter out noise outside the bandwidth, improving the system's noise immunity. A deep neural network is used to determine whether a frame contains a Morse signal; this deep learning decision model exhibits excellent noise and fading resistance, achieving a high decision accuracy. To achieve the above objectives, the technical solution adopted by this invention is as follows:
[0006] Using sampled data within 1 second as a frame, each frame is checked to determine whether it contains a Morse signal, with 1 representing the presence of a Morse signal and 0 representing its absence. When N consecutive... s When the decision result of the frame signal is 1, it is considered that a Morse message has been detected. After a Morse message is detected, the search for the end node of this Morse message begins based on the decision result of each frame signal. When N consecutive frames are detected... e When the decision result of the frame signal is 0, the Morse telegram is considered to have ended.
[0007] Furthermore, the method for determining whether a frame of signal contains a Morse signal is divided into two modules: a data preprocessing module and a deep learning decision module. The data preprocessing module filters and reduces noise on the sampled data within a frame; the deep learning decision module determines whether the frame of signal contains a Morse signal based on the output of a deep neural network. The deep neural network is trained using preprocessed features as input and the determination of whether a Morse signal is present as output.
[0008] Furthermore, the data preprocessing module performs narrowband filtering on the time-domain sampled data within 1 second. The filtering process uses a fourth-order bandpass Butterworth filter, and the passband frequency range of the filter is f. c -100 to f c +100, where f c The signal carrier frequency is denoted by . Morse signals occupy a narrow bandwidth; narrowband filtering of signals containing additive noise can effectively remove noise outside the bandwidth range, improving the robustness of the system. The filtered sampled data will serve as the input features for subsequent neural networks.
[0009] Furthermore, the deep learning decision module is implemented based on the designed neural network. The neural network structure includes a one-dimensional convolutional module, which extracts higher-dimensional information from the input features and generates sequence features; a Bi-GRU module, which inputs the sequence features into the Bi-GRU structure to fully extract contextual information; a fully connected structure, which inputs the feature vector corresponding to the last time step of the Bi-GRU output into the fully connected network to obtain a 1×2 prediction vector; and a softmax layer, which converts the prediction vector into a probability result and determines whether the signal in this frame contains a Morse signal based on this probability.
[0010] Furthermore, the one-dimensional convolution module consists of three stacked one-dimensional convolution layers, each followed by a normalization layer, a ReLU activation layer, and a max pooling layer.
[0011] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:
[0012] This invention uses data within 1 second as the smallest decision unit, balancing real-time performance and decision accuracy, and possesses significant practical value. It employs narrowband filtering for data preprocessing, effectively filtering out noise outside the bandwidth and improving system robustness. The neural network in this invention uses a Bi-GRU structure, which can fully extract the contextual features of the Morse signal. Attached Figure Description
[0013] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments are briefly described below.
[0014] Figure 1 This is the overall flowchart for Morse signal arrival detection and endpoint detection.
[0015] Figure 2 The time-domain waveform and spectrum of the Morse signal are shown.
[0016] Figure 3 This is a diagram of a deep neural network structure.
[0017] Figure 4 This is the test confusion matrix for the neural network model.
[0018] Figure 5 The ROC curve for testing the neural network model. Detailed Implementation
[0019] To more clearly illustrate the structure and function of the present invention, the invention will be further described below in conjunction with the accompanying drawings. The following embodiments will help those skilled in the art to further understand the invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the invention. These all fall within the protection scope of the present invention.
[0020] Please see Figure 1 , Figure 1 This is a flowchart illustrating the overall process of a deep learning-based Morse signal arrival detection method provided by the present invention. Figure 1 As shown, this invention uses time-domain sampled data within 1 second as a frame, and cyclically performs signal state judgment on each frame, determining the arrival and end times of the Morse signal based on the judgment results. When in the phase where no Morse telegram is detected, as... Figure 1 In the diagram, 'a' represents a loop that checks the arrival time of Morse telegrams. If the condition is met for N consecutive... s If the frame signal's decision result is 1, then a Morse message is considered to have been detected; when in the stage where a Morse message has already been detected, such as Figure 1 In the diagram, b represents a loop that checks the end time of the Morse telegram. If the condition is met for N consecutive... e If the decision result of the frame signal is 0, then the Morse telegram is considered to have ended.
[0021] Figure 1 The process of narrowband filtering on a single frame of signal, as shown, can effectively suppress noise. Figure 2 The time-domain waveform and spectrum of the Morse signal are shown. It can be observed that the Morse signal occupies a narrow bandwidth. Narrow-band filtering of signals containing additive noise can effectively filter out noise outside the bandwidth range, improving the robustness of the system. A fourth-order bandpass Butterworth filter is used in the filtering process, and the passband frequency range of the filter is f. c -100 to f c +100, where f c The signal carrier frequency is f, and the sampling frequency is f. s The filter design formula is as follows:
[0022]
[0023] Where b and a are the filter coefficients, 4 represents the filter order, and butter(N,W) n (,type) represents the design function for the Butterworth filter, where N is the filter order and W is the filter type. nThis is the normalized cutoff frequency, `type` is the filter type, and `bandpass` represents a bandpass filter. The sampled data obtained after filtering becomes the input features for the subsequent neural network. The deep neural network is trained by using the preprocessed features as input and determining whether a Morse signal is present as output.
[0024] The structural design of deep neural networks, such as Figure 3 As shown, taking a sampling rate of 8000 as an example, the temporal sampling data dimension of one frame (1s) is 8000×1. Overall, the neural network structure is divided into a convolutional module, a Bi-GRU module, and a fully connected classification module. The convolutional module consists of three stacked convolutional layers. Each convolutional layer is followed by a normalization layer, a ReLU activation layer, and a max-pooling layer. Selecting larger convolutional and pooling kernels can directly increase the receptive field, thereby capturing a wider range of contextual information in the input features, while significantly reducing the temporal dimension of the feature map. This helps reduce the computational load and memory consumption of subsequent layers. The Bi-GRU module consists of two Bi-GRU layers, with each layer having a hidden layer dimension of 128. GRU is an improved RNN that solves the gradient vanishing problem in traditional RNNs through a gating mechanism. Bidirectional RNNs improve the model's ability to capture contextual information by processing sequence information in both forward and backward directions. Taking the last time step of the output sequence features of the second Bi-GRU structure as the output of the Bi-GRU module, the output feature dimension is 1×256. Finally, the output of the Bi-GRU module is fed into a fully connected + softmax structure to calculate the binary classification probability, resulting in a 1×2 probability matrix. The index of the larger probability value is taken as the final classification result 0 or 1.
[0025] A Morse signal under a fading channel is generated through simulation, with the signal-to-noise ratio (SNR) of the simulated signal randomly ranging from -2dB to 8dB. Temporal sampling data from each second of the simulated signal is used as input to the deep learning model, with the presence or absence of a Morse signal (0 or 1) within that second as the model output. A weighted cross-entropy loss function is used for model training. In our binary classification task, the network output is a 1×2 probability matrix, representing the predicted probability of a sample belonging to one of the two classes. Let the network output be p = [p0, p1], where p0 represents the predicted probability of a sample belonging to class 0, and p1 represents the predicted probability of a sample belonging to class 1. For each sample, its true label y is either 0 or 1. The cross-entropy loss function is calculated using the following formula:
[0026] L = -[ylog(p1) + (1-y)log(p0)]
[0027] The cross-entropy loss function optimizes model parameters by minimizing the difference between the true label and the predicted probability, thereby improving the model's classification performance. The model with the minimum loss is used for testing, and the test confusion matrix is as follows: Figure 4 As shown, the model achieves a decision accuracy of 99.9% and a recall of 98.4%. For this binary classification task, the ROC curve of the model is plotted as follows. Figure 5 As shown, the AUC is 0.996, which means that the model is very robust to different samples and noise and can make the correct classification decision in most cases.
Claims
1. A Morse signal arrival detection method based on deep learning, characterized in that, The method includes the following steps: (1) Take the sampled data within 1 second as a frame, and judge whether each frame of signal contains a Morse signal in the order of time. Here, 1 represents the presence of a Morse signal and 0 represents the absence of a Morse signal. (2) Preset a starting frame number threshold N s When N consecutive s When the decision result of the frame signal is 1, it is considered that a Morse telegram has been detected; (3) After a Morse message is detected, the search begins based on the decision results of each frame of signal to find the end node of this Morse message, with a preset end frame number threshold N. e When N consecutive e When the decision result of the frame signal is 0, the Morse telegram is considered to have ended; In step (1), the method for determining whether a frame of signal contains a Morse signal is as follows: the sampled data within a frame is filtered and denoised, the preprocessed features are used as the input of a deep neural network, and the output of the deep neural network is used to determine whether the frame of signal contains a Morse signal; wherein, the deep neural network is trained by using the preprocessed features as input and determining whether a Morse signal is contained as output. The deep neural network structure includes a one-dimensional convolutional module, which extracts higher-dimensional information from the input features and generates sequence features; a Bi-GRU module, which inputs the sequence features into the Bi-GRU structure to fully extract contextual information; a fully connected structure, which inputs the feature vector corresponding to the last time step of the Bi-GRU output into the fully connected network to obtain a 1×2 prediction vector; and a softmax layer, which transforms the prediction vector into a probability result and determines whether the signal in this frame contains a Morse signal based on this probability. The decision result is represented by 0 and 1, with 1 representing the presence of a Morse signal and 0 representing the absence of a Morse signal.
2. The Morse signal arrival detection method based on deep learning according to claim 1, characterized in that, Narrowband filtering is used to remove signal noise from time-domain sampled data within 1 second. The filtering process uses a fourth-order bandpass Butterworth filter with a passband frequency range of f. c -100 to f c +100, where f c The signal carrier frequency is used, and the filtered sampled data is the input feature of the deep neural network.
3. The Morse signal arrival detection method based on deep learning according to claim 1, characterized in that, The one-dimensional convolution module consists of three stacked one-dimensional convolution layers. Each one-dimensional convolution layer is followed by a normalization layer, a ReLU activation layer, and a max pooling layer.