A method, apparatus, device, and storage medium for recognizing Morse signals.

Abstract

This invention relates to the field of Morse communication and discloses a method, apparatus, device, and storage medium for recognizing Morse signals. These methods alleviate the pressure of manual decoding of Morse signals and further improve the accuracy of Morse signal recognition under various interference conditions. The method includes: converting the received Morse signal into a corresponding time-frequency image using wavelet transform; dividing the time-frequency image into corresponding multi-frame sub-images at equal intervals, and using a 3D convolutional layer to extract and fuse the spatiotemporal features of the continuous multi-frame sub-images; inputting the spatiotemporal features into a preset bidirectional convolutional long short-term memory network model for recognition and prediction training to obtain a batch of target feature sequences predicted from the model; and decoding and transcribing the target feature sequences predicted from the model using a connection-time classification method to obtain the decoded Morse signal.

Description

Technical Field

[0001] This invention relates to the field of Morse communication technology, and in particular to a method, apparatus, device, and storage medium for recognizing Morse signals. Background Technology

[0002] Morse code communication is a relatively traditional communication method. Due to its simple coding, strong anti-interference ability, and long communication distance, it is still widely used in aerospace, maritime, and military communications. In Morse code communication, manual decoding is a crucial task, but it still faces many challenges. For example, training experienced operators requires significant resources and time; operators are susceptible to physiological or psychological factors after performing large-scale, long-term work, leading to omissions or misinterpretations. Furthermore, Morse signals can experience low signal-to-noise ratios, frequency shifts, and code length deviations during the transmission and reception stages.

[0003] To address the aforementioned problems, this invention presents a Morse signal method that significantly improves decoding accuracy in complex environments and alleviates the burden of manual decoding. Summary of the Invention

[0004] The main objective of this invention is to ensure that the identification method and system can maintain a high identification accuracy even under various interference conditions, such as low signal-to-noise ratio, code length deviation, and frequency offset, either individually or in combination.

[0005] To achieve the above objectives, the first aspect of the present invention provides a method for recognizing Morse signals, comprising:

[0006] The received Morse signal is converted into a corresponding time-frequency image using wavelet transform;

[0007] The signal time-frequency image is divided into corresponding multi-frame sub-images at equal intervals, and a 3D convolutional layer is used to extract and fuse the spatiotemporal features of the continuous multi-frame sub-images, wherein the spatiotemporal features are multi-channel two-dimensional feature sequences;

[0008] The spatiotemporal features are input into a preset bidirectional convolutional long short-term memory network model for recognition and prediction training to obtain a batch of target feature sequences of prediction output.

[0009] The target feature sequence predicted by the model is decoded and transcribed by connecting temporal classification to obtain the decoded Morse signal.

[0010] Optionally, in another implementation of the first aspect of the present invention, before converting the received Morse signal into a corresponding time-frequency image of the signal via wavelet transform, the method further includes:

[0011] Randomly batch generate raw datasets of Morse signals under low signal-to-noise ratio and save the Morse signals as readable audio. The raw datasets include standard Morse signals, Morse signals with frequency offsets, Morse signals with code length deviations, and Morse signals with both frequency offsets and code length deviations.

[0012] For each Morse signal in the original Morse signal dataset, an audio label is given, where the label is the character corresponding to the audio after decoding according to the standard Morse code table;

[0013] Raw data is extracted from the original Morse signal dataset using a random sampling method to form a training set, a test set, and a validation set;

[0014] The processed raw data is packaged into batches and input into the bidirectional convolutional long short-term memory network training model, and trained under the set training conditions to obtain the preset bidirectional convolutional long short-term memory network model.

[0015] Optionally, in another implementation of the first aspect of the present invention, the wavelet transform employs a scalable and shiftable finite-length attenuating wavelet basis function, and the step of converting the received Morse signal into a corresponding time-frequency image through wavelet transform specifically includes:

[0016] When transforming the high-frequency portion of the Morse signal, the wavelet basis function is narrowed and the transformation is performed with a lower frequency resolution. When transforming the low-frequency portion of the Morse signal, the wavelet basis function is widened and the transformation is performed with a lower time resolution, thereby obtaining a time-frequency image of the wavelet-transformed signal.

[0017] Optionally, in another implementation of the first aspect of the present invention, the step of dividing the signal time-frequency image into corresponding multi-frame sub-images at equal intervals and using a 3D convolutional layer to extract and fuse the spatiotemporal features of the continuous multi-frame sub-images specifically includes:

[0018] Pre-set two layers of 3D convolution and pooling;

[0019] The size of the signal time-frequency image is uniformly set to a predetermined number of pixels and then grayscale normalized.

[0020] Each signal time-frequency image is divided into a predetermined number of time steps at equal intervals, and the sub-image corresponding to each time step is used as the input of the first 3D convolutional layer;

[0021] Feature extraction is performed through convolution of the first and second 3D convolutional layers respectively. After each convolution, the ReLU activation function is used to normalize the feature data obtained from each layer, and finally the spatiotemporal features of the fused multi-frame sub-images are obtained.

[0022] Optionally, in another implementation of the first aspect of the present invention, the step of inputting the spatiotemporal features into a preset bidirectional convolutional long short-term memory network model for recognition and prediction training to obtain a batch of target feature sequences for prediction output specifically includes:

[0023] The two-dimensional feature sequence of the multi-channel corresponding to each time step is used as the input of the first layer of the preset bidirectional convolutional long short-term memory network model;

[0024] Feature prediction is performed using the first and second layers of the bidirectional convolutional long short-term memory network model, and layer normalization is performed on the feature data obtained from each layer to obtain the output features.

[0025] A high-level feature extraction is performed on the output features using a convolutional and pooling layer to obtain the output features at each time step;

[0026] The output features at each time step are reorganized into a one-dimensional vector, which is then used as a fully connected layer at the corresponding time step and further classified into a predetermined number of categories.

[0027] The softmax function is used to convert the classified data into probabilities between 0 and 1, so that each time step forms a corresponding posterior probability matrix, which is the target feature sequence.

[0028] Optionally, in another implementation of the first aspect of the present invention, the row vector length of the posterior probability matrix represents the length of the time step, and the column vector length represents the number of characters in the predicted classification. The decoding and transcription of the target feature sequence predicted by the model through the connection of temporal classification to obtain the Morse signal specifically includes:

[0029] Path prediction is performed by connecting temporal classification, calculating the path probability of the corresponding label in each frame of the target feature sequence, and calculating the sum of the probabilities of all paths that meet the preset conditions.

[0030] Under the supervision of the target label, the maximum value of the sum of the path probabilities of the target label is taken as the sum of the probabilities of all paths. According to the preset loss function of the label, the parameters of the recognition network model are adjusted by gradient backpropagation to obtain the character content after the Morse signal is decoded.

[0031] Optionally, in another implementation of the first aspect of the present invention, the calculation of the path probability of the label corresponding to each frame in the target feature sequence specifically involves:

[0032] Calculate the conditional probability of the label for each frame. Assuming that the conditions are independent for each frame, the path probability of the corresponding label is:

[0033]

[0034] Where t represents the position of the frame. Indicates the output label π t The probability at frame t;

[0035] The sum of the probabilities of all paths is:

[0036]

[0037] in, This represents the set of paths that are all connected temporal classifications equal to the target label.

[0038] A second aspect of the present invention also provides a Morse signal identification device, the device comprising:

[0039] The signal conversion module is used to convert the received Morse signal into a corresponding time-frequency image through wavelet transform;

[0040] The feature extraction module is used to divide the signal time-frequency image into corresponding multi-frame sub-images at equal intervals, and use a 3D convolutional layer to extract and fuse the spatiotemporal features of the continuous multi-frame sub-images, wherein the spatiotemporal features are multi-channel two-dimensional feature sequences;

[0041] The prediction and recognition module is used to input the spatiotemporal features into a preset bidirectional convolutional long short-term memory network model for recognition and prediction training, and to obtain a batch of target feature sequences of prediction output.

[0042] The decoding module is used to decode and transcribe the target feature sequence predicted by the model through connection time-series classification to obtain the decoded Morse signal.

[0043] A third aspect of the present invention provides a Morse signal identification device, the device comprising: a memory and at least one processor, the memory storing instructions, the memory and the at least one processor being interconnected via a circuit; the at least one processor calling the instructions in the memory to cause the Morse signal identification device to perform the Morse signal identification method as described in any of the above descriptions.

[0044] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the Morse signal recognition method as described in any one of the preceding claims.

[0045] In the technical solution provided by this invention, the received Morse signal is converted into a corresponding time-frequency image by wavelet transform; the time-frequency image is divided into corresponding multi-frame sub-images at equal intervals, and the spatiotemporal features of the fused continuous multi-frame sub-images are extracted using a 3D convolutional layer; the spatiotemporal features are input into a preset bidirectional convolutional long short-term memory network model for recognition and prediction training to obtain a batch of target feature sequences predicted by the model; the target feature sequences predicted by the model are decoded and transcribed by connection temporal classification to obtain the decoded Morse signal. This invention utilizes a more flexible wavelet transform in the signal preprocessing stage to convert the Morse signal into a time-frequency image with clear features. The time-frequency image is then equidistantly segmented into multiple sub-images, and a 3D convolutional layer is used to extract and fuse the spatiotemporal features of several consecutive sub-images. Finally, a bidirectional convolutional long short-term memory neural network and a connection-based temporal classification method are employed to complete the recognition and decoding of the fixed-length Morse signal. This achieves high Morse signal recognition accuracy even under conditions of low signal-to-noise ratio, code length deviation, and frequency offset interference, whether individually or in combination, while significantly reducing the burden of manual Morse signal decoding. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a schematic diagram of an embodiment of the Morse signal recognition method according to the present invention;

[0048] Figure 2 The simulation results of comparing the short-time Fourier transform and wavelet transform in the embodiments of the present invention are shown.

[0049] Figure 3 This is a diagram of the recognition network model structure used in the embodiments of the present invention;

[0050] Figure 4 These are the experimental results of different recognition algorithms under three different interferences in the embodiments of the present invention;

[0051] Figure 5 These are experimental results of different recognition algorithms in the embodiments of the present invention under the simultaneous addition of code length deviation and frequency offset interference;

[0052] Figure 6 This is a schematic diagram of one embodiment of the Morse signal recognition device according to the present invention;

[0053] Figure 7 This is a schematic diagram of one embodiment of the Morse signal recognition device in this invention. Detailed Implementation

[0054] This invention provides a method, apparatus, device, and storage medium for recognizing Morse signals, which can still achieve high recognition accuracy for Morse signals under interference such as low signal-to-noise ratio, frequency offset, and code length deviation.

[0055] Because Morse signals can exhibit low signal-to-noise ratio, frequency shift, and code length shift during the transmission and reception stages, many scholars in various fields have proposed methods for automatic Morse identification. However, the identification effectiveness of these traditional methods is generally limited by signal quality and noise interference, making them difficult to apply to real-world environments with low signal-to-noise ratios.

[0056] In recent years, some scholars have proposed new methods combining machine learning and deep learning, such as using clustering algorithms and support vector machines, which have improved the recognition effect to some extent. However, these methods require high-level signal preprocessing, and the decoding accuracy drops significantly when affected by signal-to-noise ratio, frequency offset, and code length deviation. Furthermore, while using hidden Markov models and deep neural networks has achieved good results, it remains constrained by signal-to-noise ratio and does not consider other interference factors.

[0057] This invention utilizes wavelet transform based on CMOR to preprocess Morse signals, overcoming the limitation of short-time Fourier transform in maintaining accurate time-frequency resolution. Even when the signal has a low signal-to-noise ratio and experiences frequency shifts and code length deviations, the time-frequency image obtained through wavelet transform still clearly reflects the characteristics of the Morse signal. After obtaining the time-frequency image, this invention first segments it into corresponding time steps and uses 3D convolution to extract the spatiotemporal features of continuous time steps. Then, it feeds the data into a bidirectional convolutional long short-term memory network model (BiConvLSTM) for recognition and prediction. Finally, it decodes and transcribes the network output through connection-time classification (CTC), achieving character-level decoding. Experiments show that the method provided by this invention has high recognition accuracy for fixed-length Morse signals under interference such as low signal-to-noise ratio, frequency shifts, and code length deviations.

[0058] To enable those skilled in the art to better understand the present invention, the embodiments of the present invention will be described below with reference to the accompanying drawings. The decoding and recognition objects described below are limited to fixed-length Morse signals, and the method is not limited to the recognition of variable-length Morse signals or continuous long Morse signals.

[0059] See Figure 1 One embodiment of the Morse signal recognition method of the present invention includes:

[0060] Step 101: Convert the received Morse signal into a corresponding time-frequency image using wavelet transform.

[0061] Step 102: Divide the signal time-frequency image into corresponding multi-frame sub-images at equal intervals, and use a 3D convolutional layer to extract and fuse the spatiotemporal features of the continuous multi-frame sub-images, wherein the spatiotemporal features are multi-channel two-dimensional feature sequences.

[0062] Step 103: Input the spatiotemporal features into a preset bidirectional convolutional long short-term memory network model for recognition and prediction training to obtain a batch of target feature sequences for prediction output.

[0063] Step 104: Decode and transcribe the target feature sequence predicted by the model through connection time series classification to obtain the decoded Morse signal.

[0064] In Morse code communication, a standard Morse signal consists of equal-amplitude signals corresponding to dots and dashes, and silence signals corresponding to intervals. For ease of decoding, the duration of a dot signal is usually used as the unit of time, and Morse signals are divided into five types: dot, dash, code interval, character interval, and word interval. The duration relationship of these types generally follows a ratio of 1:3:1:3:5, referring to existing Morse standard code tables.

[0065] In actual communication, the transmission environment of signals is very complex. The Morse signal received by the receiver is often in a low signal-to-noise ratio state. At the same time, the Morse signal is prone to frequency shift or unstable code rate during the transmission stage. Therefore, obtaining the position information between dots and dashes before decoding is particularly important.

[0066] In implementing this invention, in order to accurately locate the point and the time-frequency relationship of the signal, the received Morse signal is converted into a corresponding time-frequency image by wavelet transform, so as to facilitate subsequent time-frequency joint analysis.

[0067] Specifically, in step 101, the wavelet transform employs a scalable and shiftable finite-length attenuated wavelet basis function, and the conversion of the received Morse signal into a corresponding time-frequency image via wavelet transform specifically includes:

[0068] When transforming the high-frequency portion of the Morse signal, the wavelet basis function is narrowed and the transformation is performed with a lower frequency resolution. When transforming the low-frequency portion of the Morse signal, the wavelet basis function is widened and the transformation is performed with a lower time resolution, thereby obtaining a time-frequency image of the wavelet-transformed signal.

[0069] However, currently, the common signal processing approach is to use the Short-Time Fourier Transform (SFT) to convert the Morse signal into a time-frequency image, followed by joint time-frequency analysis. The SFT is essentially a windowed Fourier Transform, and the window length directly affects both time and frequency resolution. A shorter window captures a shorter signal, resulting in higher time resolution but lower frequency resolution; conversely, a wider window captures a longer signal, resulting in lower time resolution but higher frequency resolution. However, the window length of the SFT is fixed, making it difficult to flexibly control the resolution accuracy of the signal at different locations.

[0070] This invention can compare the effects of short-time Fourier transform and wavelet transform through experimental simulation, with reference to... Figure 2 , Figure 2 To obtain the time-frequency graphs of short-time Fourier transform and wavelet transform under different signal-to-noise ratios, this simulation experiment first adds noise to the same Morse signal, then sets the signal-to-noise ratios, and finally performs short-time Fourier transform and wavelet transform to obtain the results. Figure 2 The time-frequency diagram.

[0071] Figure 2 The brightness of a pixel represents the frequency domain amplitude of the signal; higher brightness indicates greater signal power. In the image, brighter rectangular blocks correspond to Morse signals, while dimmer pixels correspond to Gaussian white noise. Simulation experiments show that when the signal-to-noise ratio (SNR) is 1 dB, both short-time Fourier transform and wavelet transform can yield a clear Morse signal time-frequency diagram. This is because at a high SNR, the power of background noise is low, and its impact on the Morse signal is weak. However, wavelet transform yields better results, resulting in a more concentrated signal frequency and almost invisible ghosting at the edges of signal features.

[0072] In practical communication, the Morse signal received by the receiver often has a low signal-to-noise ratio (SNR). In the simulation, when the SNR was set to a low -12dB, the results are shown in 2(c) and 2(d). As the SNR decreases, the adhesion and mixing of the Morse signal feature region and the background noise region in 2(c) becomes very obvious, exhibiting a severe "mosaic" phenomenon. Therefore, it is difficult to distinguish the edges of the signal features, which increases the difficulty of subsequent classification and recognition using time-frequency plot features. However, the time-frequency plot of the wavelet transform does not change significantly with the decrease in SNR; the high-brightness signal region is relatively concentrated and has a clear boundary with the background noise, facilitating better extraction of Morse signal features for recognition and classification by subsequent algorithms.

[0073] Therefore, compared to the short-time Fourier transform, the wavelet transform effectively solves the problem of the trade-off between time resolution and frequency resolution. A significant feature of the wavelet transform is that it replaces the infinitely long trigonometric basis functions used in the Fourier transform with finite-length decaying wavelet base functions.

[0074] This basis function can be scaled and shifted. When the wavelet basis function is narrowed, it corresponds to the high-frequency part of the signal; when it is widened, it corresponds to the low-frequency part. At a certain scale, multiplying the wavelet basis function by the corresponding signal yields the frequency components of the signal at that scale. For the low-frequency part of the signal, a lower time resolution is generally used to obtain accurate frequency localization; for the high-frequency part, a lower frequency resolution is used to obtain accurate time localization. This solves the problem that the short-time Fourier transform cannot simultaneously account for both time and frequency resolution, and more clearly reflects the characteristics of the signal.

[0075] Furthermore, to obtain a bidirectional convolutional long short-term memory network model for recognition and prediction, this invention is implemented through the following method. Specifically, before converting the received Morse signal into a corresponding time-frequency image using wavelet transform, the method further includes:

[0076] Randomly batch generate raw datasets of Morse signals under low signal-to-noise ratio and save the Morse signals as readable audio. The raw datasets include standard Morse signals, Morse signals with frequency offsets, Morse signals with code length deviations, and Morse signals with both frequency offsets and code length deviations.

[0077] For each Morse signal in the original Morse signal dataset, an audio label is given, where the label is the character corresponding to the audio after decoding according to the standard Morse code table;

[0078] Raw data is extracted from the original Morse signal dataset using a random sampling method to form a training set, a test set, and a validation set;

[0079] The processed raw data is packaged into batches and input into the bidirectional convolutional long short-term memory network training model, and trained under the set training conditions to obtain the preset bidirectional convolutional long short-term memory network model.

[0080] In practice, since there is currently no publicly available general dataset for Morse signal recognition, the Morse signal dataset used in this invention is automatically generated in batches by the program. The frequency, code length, and signal-to-noise ratio of the Morse signal can be customized and saved as a WAV audio file for reading.

[0081] To simulate real-world conditions as closely as possible, this invention generates standard Morse signals under low signal-to-noise ratio conditions. The original datasets include Morse signals with frequency offsets, Morse signals with code length deviations, and Morse signals with both frequency and code length deviations. Each dataset contains 5000 Morse signals, and each signal maps to 5 characters. That is, for each Morse signal's WAV audio, the given tag is the character corresponding to the audio after decoding according to the standard Morse code table.

[0082] The dataset was generated by randomly selecting combinations from the 26 lowercase English letters and Arabic numerals 0 to 9. The standard Morse code signal has a frequency of 800 Hz and a sampling rate of 8000 Hz. The durations of the dot, dash, code, character, and word intervals are 100 ms, 300 ms, 100 ms, 300 ms, and 500 ms, respectively, with the overall signal duration ranging from 4 to 7 seconds. When frequency offset exists, an offset of 0 to 200 Hz is randomly added or subtracted from the standard signal frequency for the dot and dash signals; when code length deviation exists, an offset coefficient of 0.8 to 1.2 is randomly multiplied from the standard signal time interval.

[0083] Before training, this invention partitioned the dataset, allocating approximately 70% of the original data as the training set, 20% as the test set, and the remaining 10% as the validation set. Random sampling was used to extract the original data into the training, test, and validation sets. After partitioning the dataset, the processed data was packaged into batches (32 data points per batch in this invention) and loaded into the training model. The Adam optimizer was used for training, employing L2 regularization, with an initial learning rate of 0.001, which decayed exponentially by 0.96 after each epoch. The graphics card used for training was an RTX 3080Ti. Through training, testing, and validation on the original dataset, a bidirectional convolutional long short-term memory network model meeting the requirements was obtained.

[0084] Furthermore, this invention employs feature extraction based on 3D convolution. In another implementation, the step of equidistantly segmenting the signal time-frequency image into corresponding multi-frame sub-images and using a 3D convolutional layer to extract and fuse the spatiotemporal features of the continuous multi-frame sub-images specifically includes:

[0085] Pre-set two layers of 3D convolution and pooling;

[0086] The size of the signal time-frequency image is uniformly set to a predetermined number of pixels and then grayscale normalized.

[0087] Each signal time-frequency image is divided into a predetermined number of time steps at equal intervals, and the sub-image corresponding to each time step is used as the input of the first 3D convolutional layer;

[0088] Feature extraction is performed through convolution of the first and second 3D convolutional layers respectively. After each convolution, the ReLU activation function is used to normalize the feature data obtained from each layer, and finally the spatiotemporal features of the fused multi-frame sub-images are obtained.

[0089] Compared to existing technologies that use the time-frequency map of a single Morse signal as input to the recognition network, employ two layers of 2D convolution for feature extraction, and then reassemble the feature map into a multi-frame feature sequence according to the divided time steps, where each frame of the time sequence contains high-level features from the original time-frequency map, this invention does not use the entire time-frequency map of the signal as input. Instead, it first uniformly divides the original time-frequency map into multiple sub-maps, and then uses 3D convolution to extract the spatiotemporal features of several consecutive sub-maps, simultaneously obtaining the spatial features of each sub-map itself and the local spatiotemporal dependencies between sub-maps.

[0090] In specific implementation, for example, the 3D convolution kernel used is (3, 3, 3), that is, the (3, 3) convolution kernel is used to extract features from three consecutive time steps. The spatiotemporal information between each time step is linked together by the continuous sliding of the convolution window, which also expands the receptive field.

[0091] In practice, the Morse code at the receiving end is usually input into the recognition system as a signal stream, making pre-segmentation of the signal stream crucial. However, inappropriate segmentation can alter the dot-dash features of adjacent sub-signals near the segmentation line. If 2D convolution is then used to operate independently on the time-frequency graph of each sub-signal, the spatiotemporal correlation features between the two segmented sub-signals cannot be obtained, easily leading to misjudgments at the tail of the preceding signal and the head of the following signal, thus degrading the overall decoding performance. This invention uses 3D convolution, which effectively integrates information from consecutive frames into the overall features, meaning it is not sensitive to the segmentation position and accuracy. Compared to 2D convolution, it has a greater advantage in processing long signals.

[0092] Furthermore, after feature extraction of the signal time-frequency image is completed in step 102, in order to enhance feature extraction of sub-images at each time step and to learn the global long-term spatiotemporal feature correlation of all time steps as much as possible, this invention uses a bidirectional convolutional long short-term memory network as the recognition network. In another implementation, the step of inputting the spatiotemporal features into a preset bidirectional convolutional long short-term memory network model for recognition and prediction training to obtain a batch of target feature sequences of prediction output specifically includes:

[0093] The two-dimensional feature sequence of the multi-channel corresponding to each time step is used as the input of the first layer of the preset bidirectional convolutional long short-term memory network model;

[0094] Feature prediction is performed using the first and second layers of the bidirectional convolutional long short-term memory network model, and layer normalization is performed on the feature data obtained from each layer to obtain the output features.

[0095] A high-level feature extraction is performed on the output features using a convolutional and pooling layer to obtain the output features at each time step;

[0096] The output features at each time step are reorganized into a one-dimensional vector, which is then used as a fully connected layer at the corresponding time step and further classified into a predetermined number of categories.

[0097] The softmax function is used to convert the classified data into probabilities between 0 and 1, so that each time step forms a corresponding posterior probability matrix, which is the target feature sequence.

[0098] In a specific implementation of this invention, for example, the input image size is uniformly set to 160×32 pixels and grayscale normalization is performed on it. Then, each image is divided into 40 time steps, and the sub-image (4×32 pixels in size) corresponding to each time step is used as the input of the first 3D convolutional layer.

[0099] In the feature extraction part, the network uses two layers of convolution and pooling, with a ReLU activation function applied after each convolution, followed by data normalization. In the prediction and recognition part, a two-layer BiConvLSTM network is used, and layer normalization is applied to the output of each ConvLSTM layer at all time steps. After obtaining the output of the BiConvLSTM, another convolution and pooling layer is used to further extract high-level features and remove redundant data. Finally, the output features of each time step are reorganized into a one-dimensional vector, which serves as the fully connected layer for each time step and is further divided into 37 categories (labels include 10 digits, 26 lowercase English letters, and whitespace placeholders introduced in CTC). The softmax function is used to convert the classified data into probabilities between 0 and 1, resulting in a 40×37 (time step × number of categories) posterior probability matrix for each batch. This facilitates subsequent transcription of the prediction results using connected temporal classification. (Reference) Figure 3 As shown, this is the identification network structure used in this invention. The figure clearly illustrates the specific connections between each layer and also shows the basic operation units of each layer.

[0100] Specifically, after the aforementioned 3D convolution, the features at each time step Figure 1Typically, LSTM inputs are multi-channel two-dimensional sequences, thus they can be well matched to the input of each time step in a ConvLSTM network without requiring additional dimensionality adjustments. In contrast, traditional LSTMs usually require a single-channel or multi-channel one-dimensional sequence as input at each time step, necessitating feature extraction and dimensionality reduction when processing two-dimensional images. ConvLSTM transforms weight calculation from Hadamard multiplication to convolution operations, preserving temporal information between data points and enabling each LSTM unit to extract spatial features from the data.

[0101] In step 103, the recognition model outputs a batch of feature sequences, i.e., the posterior probability matrix, after each round of training. The row vector length of the posterior probability matrix represents the length of the time step, and the column vector length represents the number of characters to be predicted and classified.

[0102] Furthermore, in the final decoding process, this invention employs CTC to decode and transcribe the network output to achieve character-level decoding. Specifically, the decoding and transcription of the target feature sequence predicted by the model through connection-time classification to obtain the Morse signal includes:

[0103] Path prediction is performed by connecting temporal classification, calculating the path probability of the corresponding label in each frame of the target feature sequence, and calculating the sum of the probabilities of all paths that meet the preset conditions.

[0104] Under the supervision of the target label, the maximum value of the sum of the path probabilities of the target label is taken as the sum of the probabilities of all paths. According to the preset loss function of the label, the parameters of the recognition network model are adjusted by gradient backpropagation to obtain the character content after the Morse signal is decoded.

[0105] Specifically, calculating the path probability of the label corresponding to each frame in the target feature sequence is as follows:

[0106] Calculate the conditional probability of the label for each frame. Assuming that the conditions are independent for each frame, the path probability of the corresponding label is:

[0107]

[0108] Where t represents the position of the frame. Indicates the output label π t The probability at frame t;

[0109] The sum of the probabilities of all paths is:

[0110]

[0111] in, This represents the set of paths that are all connected temporal classifications equal to the target label.

[0112] In practice, each frame of the feature sequence corresponds to a set of column vectors of the probability matrix. Usually, multiple frames in the feature sequence need to be combined sequentially to align a certain part of the label content. That is, after combining the features of multiple time steps of the i-frequency image, it may ultimately only be mapped to a single character. Therefore, the length of the string obtained after decoding is generally less than the length of the feature sequence.

[0113] Therefore, during the training process of this invention, it is essential to ensure that the decoding result of the Morse signal prediction sequence is equal in length to the label string before subsequent calculations of the loss function and recognition accuracy can be performed. CTC introduces frame-level prediction paths, aiming to find the sum of the probabilities of the paths closest to the label. After finding the prediction path, CTC introduces whitespace placeholders to separate two adjacent identical labels. Another function is to merge duplicate characters between two whitespace placeholders. Finally, simply deleting all whitespace characters yields the decoding result.

[0114] Specifically, determining the CTC path first requires calculating the conditional probability of the tag corresponding to each frame. Assuming that the conditions are independent in each frame, the probability of a CTC path is:

[0115]

[0116] Where t represents the position of the frame. Indicates the output label π t The probability at frame t;

[0117] The sum of the final path probabilities is equal to the sum of the probabilities of all CTC paths for the output label:

[0118] The sum of the probabilities of all paths is:

[0119]

[0120] in, This represents the set of paths that are all connected temporal classifications equal to the target label.

[0121] In this invention, since the number of classification labels is 37 and t is 10, it is impractical to calculate the probability of all paths that meet the conditions. The verification experiment of this invention uses a beam-search algorithm here, that is, for each frame the path passes through from the starting point, only the 10 most likely paths to reach that frame are selected. Although some CTC paths that meet the reachability conditions are sacrificed, the computational efficiency is significantly improved. By selecting the 10 most likely paths to reach that frame and calculating the sum of the probabilities of all paths reaching the preset conditions, after obtaining the sum of path probabilities, the loss function can be calculated based on the labels, and then the parameters of the recognition network can be adjusted using backpropagation of gradients.

[0122] Specifically, the following loss function can be used, i.e., the optimization objective is:

[0123]

[0124] Where F represents the sum of CTC path probabilities to be optimized, and label represents the true label. The ultimate goal of training is to maximize the sum of path probabilities that map the predicted result to that label under the supervision of a given target label, so that the predicted decoded character content is as close as possible to the character content of the label.

[0125] The Morse code identification method described above can be applied to situations with low signal-to-noise ratio, code length deviation, and frequency offset interference, either individually or in combination. This invention provides an implementation verification process and related data to verify that the provided method maintains high identification accuracy under various interference conditions.

[0126] This experiment uses character recognition accuracy and word recognition accuracy as performance indicators of the recognition algorithm. The performance indicators of each phase are verified by comparing the Convolutional Recurrent Neural Network (CRNN) with the preset bidirectional convolutional long short-term memory network model 3D CNN-BiConvLSTM provided in this invention.

[0127] Character accuracy refers to the proportion of correctly identified characters out of the total number of characters in all sample labels, while word accuracy refers to the proportion of correctly identified words out of the total number of words in all sample labels.

[0128] Since no filtering or data augmentation is performed on the noisy Morse signal during signal preprocessing, the recognition accuracy can also reflect the algorithm's ability to resist noise interference from a robustness perspective. (Reference) Figure 4 This paper presents experimental results and analyzes the performance of different recognition algorithms, namely loss curves, word recognition accuracy, and character recognition accuracy, under three different types of interference.

[0129] In practice, to simulate the actual low signal-to-noise ratio environment, all types of interference were trained under a signal-to-noise ratio (SNR) of -10dB.

[0130] As can be seen from 4(a) and 4(b), in the early stage of training, the loss function of 3DCNN-BiConvLSTM experienced a very slow decline under the interference of code length bias and frequency offset. However, as training progresses, compared with the performance of CRNN, the overall loss function of 3DCNN-BiConvLSTM converges to near 0 faster and with less fluctuation.

[0131] When Morse code length deviations exist, although CRNN achieves character recognition accuracy very close to 3DCNN-BiConvLSTM, it still lags behind in word recognition accuracy. Furthermore, even in the later stages of training, CRNN's performance remains volatile, which negatively impacts recognition stability.

[0132] When the Morse signal has a frequency shift, although the recognition accuracy of 3DCNN-BiConvLSTM fluctuated to varying degrees in the early stage of training, it became more and more stable as training progressed, especially in word recognition accuracy, which remained at around 98%, while CRNN never broke through 95%.

[0133] Finally, to verify the recognition performance of the automatic recognition network model under comprehensive interference, this invention simultaneously adds code length deviation and frequency shift interference to each original Morse signal, while still placing it under a signal-to-noise ratio of -10dB. The final experimental results are as follows: Figure 5 As shown.

[0134] In the initial training phase, both the loss functions of the Convolutional Recurrent Neural Network (CRNN) and the pre-defined bidirectional convolutional long short-term memory network model 3D CNN-BiConvLSTM provided in this invention exhibited a very gradual decrease. However, thereafter, the convergence speed of 3DCNN-BiConvLSTM was significantly faster. In word recognition, 3DCNN-BiConvLSTM generally outperformed CRNN in accuracy and was more stable. In character recognition, the difference between the two was small, but CRNN was more prone to fluctuations.

[0135] To address the interference introduced individually or in combination, 3DCNN-BiConvLSTM has achieved satisfactory results in both word and character recognition accuracy. To improve word recognition accuracy, it is essential to ensure high accuracy for each character, while the spatiotemporal relationship between any two characters is a key factor influencing character recognition accuracy.

[0136] While CRNN uses 2D convolutions to extract features from time-frequency images and reconstruct them into corresponding time-step sequences, it cannot capture the spatiotemporal feature information between each time step in advance and can only rely on the LSTM network itself to obtain it. In contrast, 3DCNN-BiConvLSTM, before performing time-series prediction, has already fused the local spatiotemporal information of several consecutive frames of time-frequency sub-images into a new time step through 3D convolutions, thereby establishing the local spatiotemporal dependencies between time steps. Then, ConvLSTM is used to further strengthen the long-term spatiotemporal dependencies between time steps.

[0137] Experimental results show that, under low signal-to-noise ratio conditions, using the pre-defined bidirectional convolutional long short-term memory network model 3D CNN-BiConvLSTM provided in this invention as the automatic recognition network for fixed-length Morse signals can effectively resist interference caused by code length deviation and frequency shift, and achieve a high recognition accuracy.

[0138] In summary, the Morse signal recognition method provided by this invention converts the received Morse signal into a corresponding time-frequency image using wavelet transform; the time-frequency image is equidistantly segmented into corresponding multi-frame sub-images, and a 3D convolutional layer is used to extract and fuse the spatiotemporal features of the continuous multi-frame sub-images; the spatiotemporal features are input into a preset bidirectional convolutional long short-term memory network model for recognition and prediction training to obtain a batch of target feature sequences predicted by the model; the target feature sequences predicted by the model are decoded and transcribed by connection temporal classification to obtain the decoded Morse signal. This invention utilizes a more flexible wavelet transform in the signal preprocessing stage to convert the Morse signal into a time-frequency image with clear features. Then, the time-frequency image is equidistantly segmented into multiple sub-images, and a 3D convolutional layer is used to extract and fuse the spatiotemporal features of several consecutive sub-images. Finally, a bidirectional convolutional long short-term memory neural network and a connection-based temporal classification method are employed to complete the recognition and decoding of the fixed-length Morse signal. This achieves high Morse signal recognition accuracy even under conditions of low signal-to-noise ratio, code length deviation, and frequency offset interference, whether individually or in combination, while significantly reducing the workload of manual Morse signal decoding.

[0139] The method for recognizing Morse signals in the embodiments of the present invention has been described above. The device for recognizing Morse signals in the embodiments of the present invention is described below. Please refer to [link / reference needed]. Figure 6 One embodiment of the Morse signal recognition device in this invention includes:

[0140] The signal conversion module 11 is used to convert the received Morse signal into a corresponding time-frequency image of the signal through wavelet transform;

[0141] The feature extraction module 12 is used to divide the signal time-frequency image into corresponding multi-frame sub-images at equal intervals, and use a 3D convolutional layer to extract and fuse the spatiotemporal features of the continuous multi-frame sub-images, wherein the spatiotemporal features are multi-channel two-dimensional feature sequences;

[0142] The prediction and recognition module 13 is used to input the spatiotemporal features into a preset bidirectional convolutional long short-term memory network model for recognition and prediction training, and obtain a batch of target feature sequences of prediction output.

[0143] The decoding module 14 is used to decode and transcribe the target feature sequence predicted by the model through connection time series classification to obtain the decoding of the Morse signal.

[0144] Optionally, the device further includes:

[0145] The raw dataset acquisition module is used to randomly generate raw Morse signal datasets under low signal-to-noise ratio and save the Morse signals as readable audio. The raw dataset includes standard Morse signals, Morse signals with frequency offsets, Morse signals with code length deviations, and Morse signals with both frequency offsets and code length deviations.

[0146] The labeling module is used to assign a label to the audio of each Morse signal in the original Morse signal dataset, wherein the label is the character corresponding to the audio after decoding according to the standard Morse code table;

[0147] The dataset partitioning module is used to extract raw data from the original Morse signal dataset using a random sampling method to form a training set, a test set, and a validation set.

[0148] The training module is used to package the processed raw data into a set batch and input it into the bidirectional convolutional long short-term memory network training model, and train it under set training conditions to obtain the preset bidirectional convolutional long short-term memory network model.

[0149] Optionally, the wavelet transform employs a scalable and shiftable finite-length attenuated wavelet basis function, and the signal conversion module specifically includes:

[0150] A high-frequency signal processing unit is used to narrow the wavelet basis function and perform the transformation with a lower frequency resolution when converting the high-frequency part of the Morse signal.

[0151] The low-frequency signal processing unit is used to expand the wavelet basis function more widely and perform the transformation with a lower time resolution when transforming the low-frequency part of the Morse signal, thereby obtaining a wavelet transform signal time-frequency image.

[0152] Optionally, the feature extraction module specifically includes:

[0153] Pre-configured cells are used to pre-configure two layers of 3D convolution and pooling;

[0154] A pixel and grayscale normalization processing unit is used to uniformly set the size of the signal time-frequency image to a predetermined number of pixels and perform grayscale normalization on it;

[0155] The segmentation unit is used to divide each signal time-frequency image into a predetermined number of time steps at equal intervals, and the sub-image corresponding to each time step is used as the input of the first 3D convolutional layer;

[0156] The feature extraction and normalization unit is used to extract features through the convolution of the first and second 3D convolutional layers respectively, and after each convolution, the ReLU activation function is used to normalize the feature data obtained by each layer, and finally obtain the spatiotemporal features of the fused continuous multi-frame sub-images.

[0157] Optionally, the prediction and recognition module specifically includes:

[0158] The input unit is used to take the two-dimensional feature sequence of the multi-channel network corresponding to each time step as the input of the first layer of the preset bidirectional convolutional long short-term memory network model;

[0159] The feature prediction and normalization unit is used to perform feature prediction through the first and second layers of the bidirectional convolutional long short-term memory network model, and to perform layer normalization on the feature data obtained from each layer's feature prediction to obtain the output features.

[0160] A high-level feature extraction unit is used to perform high-level feature extraction on the output features using a convolutional and pooling layer to obtain the output features at each time step;

[0161] The recombination and classification unit is used to recombine the output features of each time step into a one-dimensional vector, which serves as the fully connected layer of the corresponding time step and is further classified into a predetermined number of categories.

[0162] The probability transformation unit is used to convert the classified data into probabilities between 0 and 1 using the softmax function, so that each time step forms a corresponding posterior probability matrix, which is the target feature sequence.

[0163] Optionally, the row vector length of the posterior probability matrix represents the length of the time step, and the column vector length represents the number of characters to be predicted for classification. The decoding module specifically includes:

[0164] The path prediction and probability calculation unit is used to perform path prediction by connecting temporal classification, calculate the path probability of the corresponding label in each frame of the target feature sequence, and calculate the sum of the probabilities of all paths that meet the preset conditions.

[0165] The character content acquisition unit is used to, under the supervision of the target label, make the maximum value of the sum of the path probabilities of the target label take the sum of the probabilities of all paths, and adjust the parameters of the recognition network model through gradient backpropagation according to the preset loss function of the label to obtain the character content after Morse signal decoding.

[0166] Optionally, the calculation of the path probability of the label corresponding to each frame in the target feature sequence is specifically as follows:

[0167] Calculate the conditional probability of the label for each frame. Assuming that the conditions are independent for each frame, the path probability of the corresponding label is:

[0168]

[0169] Where t represents the position of the frame. Indicates the output label π t The probability at frame t;

[0170] The sum of the probabilities of all paths is:

[0171]

[0172] in, This represents the set of paths that are all connected temporal classifications equal to the target label.

[0173] It should be noted that the apparatus in the embodiments of the present invention can be used to implement all the technical solutions in the above method embodiments. The functions of each functional module can be specifically implemented according to the methods in the above method embodiments. The specific implementation process can be referred to the relevant descriptions in the above examples, which will not be repeated here.

[0174] above Figure 6 The Morse signal recognition device in this embodiment of the invention is described in detail from the perspective of modular functional entities. The Morse signal recognition device in this embodiment of the invention is described in detail from the perspective of hardware processing.

[0175] Figure 7This is a schematic diagram of the structure of a Morse signal recognition device 300 provided in an embodiment of the present invention. The Morse signal recognition device 300 can vary significantly due to different configurations or performance characteristics. It may include one or more central processing units (CPUs) 301 (e.g., one or more processors) and a memory 309, and one or more storage media 308 (e.g., one or more mass storage devices) storing application programs 307 or data 306. The memory 309 and storage media 308 can be temporary or persistent storage. The program stored in the storage media 308 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations stored in Boolean variables for graph computation. Furthermore, the processor 301 may be configured to communicate with the storage media 308 and execute the series of instruction operations in the storage media 308 on the Morse signal recognition device 300.

[0176] The Morse code identification device 300 may also include one or more power supplies 302, one or more wired or wireless network interfaces 303, one or more input / output interfaces 304, and / or one or more operating systems 305, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 7 The structure of the Morse signal identification device shown does not constitute a limitation on the Morse signal identification device, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0177] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0178] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0179] In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between devices or units through some interfaces, and may be electrical, mechanical, or other forms.

[0180] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0181] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0182] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium, which can be non-volatile or volatile. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0183] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for recognizing Morse signals, characterized in that, The method includes: The received Morse signal is converted into a corresponding time-frequency image by wavelet transform. The wavelet transform uses a finite-length attenuating wavelet basis function. When the wavelet basis function is narrowed, it corresponds to the conversion of the high-frequency part of the Morse signal. When the wavelet basis function is broadened, it corresponds to the conversion of the low-frequency part of the Morse signal. The signal time-frequency image is divided into corresponding multi-frame sub-images at equal intervals, and a 3D convolutional layer is used to extract and fuse the spatiotemporal features of the continuous multi-frame sub-images, wherein the spatiotemporal features are multi-channel two-dimensional feature sequences; The spatiotemporal features are input into a preset bidirectional convolutional long short-term memory network model for recognition and prediction training to obtain a batch of target feature sequences of prediction output. The target feature sequence predicted by the model is decoded and transcribed by connecting time-series classification to obtain the decoded Morse signal; Specifically, the process of converting the received Morse signal into a corresponding time-frequency image using wavelet transform includes: When converting the high-frequency part of the Morse signal, the wavelet basis function is narrowed and a lower frequency resolution is used for transformation to obtain accurate time positioning and obtain the corresponding signal time-frequency image; When converting the low-frequency portion of the Morse signal, the wavelet basis function is broadened and transformed using a lower time resolution to obtain accurate frequency localization, resulting in the corresponding time-frequency image of the signal. Specifically, the step of dividing the signal time-frequency image into corresponding multi-frame sub-images at equal intervals and using a 3D convolutional layer to extract and fuse the spatiotemporal features of the continuous multi-frame sub-images includes: Pre-set two layers of 3D convolution and pooling; The size of the signal time-frequency image is uniformly set to a predetermined number of pixels and then grayscale normalized. Each signal time-frequency image is divided into a predetermined number of time steps at equal intervals, and the sub-image corresponding to each time step is used as the input of the first 3D convolutional layer; Feature extraction is performed through convolution of the first and second 3D convolutional layers respectively. After each convolution, the ReLU activation function is used to normalize the feature data obtained from each layer, so as to obtain the spatiotemporal features of fused continuous multi-frame sub-images.

2. The method for recognizing Morse signals according to claim 1, characterized in that, Before converting the received Morse signal into a corresponding time-frequency image using wavelet transform, the method further includes: Randomly batch generate raw datasets of Morse signals under low signal-to-noise ratio and save the Morse signals as readable audio. The raw datasets include standard Morse signals, Morse signals with frequency offsets, Morse signals with code length deviations, and Morse signals with both frequency offsets and code length deviations. For each Morse signal in the original Morse signal dataset, an audio label is given, where the label is the character corresponding to the audio after decoding according to the standard Morse code table; Raw data is extracted from the original Morse signal dataset using a random sampling method to form a training set, a test set, and a validation set; The processed raw data is packaged into batches and input into the bidirectional convolutional long short-term memory network training model, and trained under the set training conditions to obtain the preset bidirectional convolutional long short-term memory network model.

3. The method for recognizing Morse signals according to claim 1, characterized in that, The step of inputting the spatiotemporal features into a preset bidirectional convolutional long short-term memory network model for recognition and prediction training to obtain a batch of target feature sequences for prediction output specifically includes: The two-dimensional feature sequence of the multi-channel corresponding to each time step is used as the input of the first layer of the preset bidirectional convolutional long short-term memory network model; Feature prediction is performed using the first and second layers of the bidirectional convolutional long short-term memory network model, and layer normalization is performed on the feature data obtained from each layer to obtain the output features. A high-level feature extraction is performed on the output features using a convolutional and pooling layer to obtain the output features at each time step; The output features at each time step are reorganized into a one-dimensional vector, which is then used as a fully connected layer at the corresponding time step and further classified into a predetermined number of categories. The softmax function is used to convert the classified data into probabilities between 0 and 1, so that each time step forms a corresponding posterior probability matrix, which is the target feature sequence.

4. The method for recognizing Morse signals according to claim 3, characterized in that, The row vector length of the posterior probability matrix represents the length of the time step, and the column vector length represents the number of characters in the predicted classification. The decoding and transcription of the target feature sequence predicted by the model through the connection of temporal classification to obtain the Morse signal specifically includes: Path prediction is performed by connecting temporal classification, calculating the path probability of the corresponding label in each frame of the target feature sequence, and calculating the sum of the probabilities of all paths that meet the preset conditions. Under the supervision of the target label, the maximum value of the sum of the path probabilities of the target label is taken as the sum of the probabilities of all paths. According to the preset loss function of the label, the parameters of the recognition network model are adjusted by gradient backpropagation to obtain the character content after the Morse signal is decoded.

5. The method for recognizing Morse signals according to claim 4, characterized in that, The specific steps for calculating the path probability of the label corresponding to each frame in the target feature sequence are as follows: Calculate the conditional probability of the label for each frame. Assuming that the conditions are independent for each frame, the path probability of the corresponding label is: Where t represents the position of the frame. Indicates the output label π t The probability at frame t; where the sum of the probabilities of all paths is: in, This represents the set of all connection time-series classifications that are equal to the target label.

6. A Morse signal recognition device, characterized in that, The device includes: The signal conversion module is used to convert the received Morse signal into a corresponding time-frequency image of the signal through wavelet transform. The wavelet transform uses a finite-length attenuating wavelet basis function. When the wavelet basis function is narrowed, it corresponds to the conversion of the high-frequency part of the Morse signal. When the wavelet basis function is broadened, it corresponds to the conversion of the low-frequency part of the Morse signal. The feature extraction module is used to segment the signal time-frequency image into corresponding multi-frame sub-images at equal intervals, and to extract and fuse the spatiotemporal features of the continuous multi-frame sub-images using a 3D convolutional layer, wherein the spatiotemporal features are multi-channel two-dimensional feature sequences; the prediction and recognition module is used to input the spatiotemporal features into a preset bidirectional convolutional long short-term memory network model for recognition. Instead of predictive training, a batch of target feature sequences is obtained from the predicted output; The decoding module is used to decode and transcribe the target feature sequence predicted by the model through connection time-series classification to obtain the decoded Morse signal; The signal conversion module further includes: A high-frequency signal conversion unit is used to narrow the wavelet basis function and use a lower frequency resolution to transform the high-frequency part of the Morse signal to obtain accurate time positioning and obtain the corresponding signal time-frequency image. The low-frequency signal conversion unit is used to broaden the wavelet basis function and perform transformation with a lower time resolution when converting the low-frequency part of the Morse signal to obtain accurate frequency positioning and obtain the corresponding signal time-frequency image. The feature extraction module further includes: 3D convolution preset unit, used to pre-set two layers of 3D convolution and pooling; The time-frequency image size and normalization processing unit is used to uniformly set the size of the signal time-frequency image to a predetermined number of pixels and perform grayscale normalization on it. The time-frequency image equidistant segmentation processing unit is used to equidistantly segment each signal time-frequency image into a predetermined number of time steps, and the sub-image corresponding to each time step is used as the input of the first 3D convolutional layer; The spatiotemporal feature acquisition unit is used to extract features through the convolution of the first and second 3D convolutional layers respectively, and after each convolution, the ReLU activation function is used to normalize the feature data obtained by each layer to obtain the spatiotemporal features of the fused continuous multi-frame sub-images.

7. A Morse signal recognition device, characterized in that, The Morse signal identification device includes: a memory and at least one processor, wherein the memory stores instructions, and the memory and the at least one processor are interconnected via a line; the at least one processor invokes the instructions in the memory to cause the Morse signal identification device to perform the Morse signal identification method as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the Morse signal recognition method as described in any one of claims 1-5.

Images