Method and system for traffic training based on attention mechanism and recurrent neural network

Abstract

This application relates to the field of communication technology, and in particular to a message training method and system based on attention mechanism and recurrent neural network. The method includes acquiring audio input data, processing the audio input data sequentially to obtain a high-dimensional time-frequency feature sequence; inputting the high-dimensional time-frequency feature sequence into a convolutional neural network to obtain a semantic feature map; performing parallel heterogeneous dual-branch processing on the semantic feature map, obtaining attention-weighted features through a convolutional attention module and context-aware features through a bidirectional long short-term memory network; using adaptive feature fusion to assign weights and fuse the attention-weighted features and context-aware features to generate a unified feature representation, and mapping it to the probability distribution of the corresponding character at each time step; using connectionist temporal classification to decode the probability sequence and output the decoded character sequence. The method can adapt to different firing habits, has strong anti-interference ability, and achieves high-precision and stable end-to-end decoding.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a message training method and system based on attention mechanism and recurrent neural network. Background Technology

[0002] With the deep integration of information technology in military, emergency response, and amateur radio fields, the intelligentization and automation of radio communication training and support have become key development trends. One of the core skills of radio communication is the accurate identification and decoding of audio signals, represented by Morse code, which are encoded by manual rhythmic striking.

[0003] In existing technologies for automatic Morse code recognition, mainstream solutions rely on threshold-based decision-making or static clustering algorithms (such as the Günther algorithm, relative comparison method, and probability partitioning method). These solutions generally suffer from the following problems: 1) Due to the different sending habits and speeds of different radio operators, the actual duration of dots, dashes, and intervals in the signal is dynamically changing. The traditional method of distinguishing and judging dots, dashes, and intervals by manually setting fixed thresholds cannot adapt to this change in real time, and is prone to misjudgment of dots and dashes, resulting in an increased misjudgment rate and failing to meet the needs of high-quality radio operator training.

[0004] 2) Actual key signals are susceptible to interference from hardware jitter and electromagnetic noise. Morse code automatic recognition schemes relying on threshold-based decisions or static clustering algorithms are highly sensitive to noise and exhibit unstable performance in complex environments. Specifically, the lack of a customized multi-layer convolutional pooling stacking design tailored to the time-frequency characteristics of Morse code makes it impossible to accurately extract the core local spatiotemporal features corresponding to dots, dashes, and intervals. Furthermore, the serial architecture's unidirectional progressive relationship between feature purification and temporal modeling means that the output errors of preceding modules are propagated and accumulated, leading to a mutual constraint between feature purification and temporal context modeling capabilities. Additionally, the comprehensive feature purification process struggles to effectively suppress noise interference in complex environments, resulting in weak anti-interference robustness and unstable recognition performance in complex conditions. Summary of the Invention Based on this, this application provides a message training method and system based on attention mechanism and recurrent neural network, which aims to solve the technical problems in related technologies where message training methods cannot adapt to the different shooting habits and shooting speeds of different operators, resulting in easy misjudgment of dots and strokes, low recognition accuracy, and weak anti-interference robustness.

[0005] In a first aspect, embodiments of this application provide a message service training method based on an attention mechanism and a recurrent neural network, comprising: Audio input data, including analog audio signals or digital audio sequences, is acquired. The audio input data is then subjected to bandpass filtering, frame windowing, Mel frequency cepstral coefficients, multi-order difference feature extraction, and standardization processing in sequence to obtain a high-dimensional time-frequency feature sequence. The high-dimensional time-frequency feature sequence is input into a convolutional neural network composed of multiple stacked two-dimensional convolutional layers and pooling layers to obtain a semantic feature map containing local spatiotemporal patterns of Morse code dots, dashes and intervals. The semantic feature map is processed in parallel with two branches; wherein, based on the first branch, channel and spatial weighting is performed to obtain attention-weighted features; and based on the second branch, bidirectional temporal modeling is performed to obtain context-aware features. An adaptive feature fusion module based on a gating mechanism is used to perform end-to-end learnable dynamic weight allocation and feature fusion on the attention-weighted features and the context-aware features to generate a unified feature representation. The unified feature representation is mapped to the probability distribution of the corresponding character at each time step. The probability sequence is decoded using connectionist temporal classification, and the decoded character sequence corresponding to the Morse code is output.

[0006] In some embodiments, the step of acquiring audio input data including analog audio signals or digital audio sequences, and sequentially performing bandpass filtering, frame windowing, Mel-frequency cepstral coefficients, multi-order difference feature extraction, and standardization on the audio input data to obtain a high-dimensional time-frequency feature sequence includes: The analog audio signal is converted from analog to digital using a sampling specification adapted to the effective frequency band of Morse code to obtain a digital audio sequence; The digital audio sequence is subjected to digital bandpass filtering that matches the Morse code carrier frequency band to obtain a filtered, clean audio sequence. Based on the pure audio sequence, the frame length and frame shift parameters of the minimum timing unit of the matching Morse code are used for frame division. A Hamming window function that adapts to the short-time stationary characteristics of the audio is applied to each frame of data after frame division to obtain the windowed and normalized frame sequence. The Mel frequency cepstral coefficients are extracted frame by frame from the frame sequence as static features. Each static feature is then concatenated with its corresponding first-order difference feature and second-order difference feature along the feature dimension to obtain an initial feature sequence that fuses the Morse code timing variation information. Global standardization is then performed based on the initial feature sequence to obtain the high-dimensional time-frequency feature sequence.

[0007] In some embodiments, the step of inputting the high-dimensional time-frequency feature sequence into a convolutional neural network composed of multiple stacked two-dimensional convolutional layers and pooling layers to obtain a semantic feature map containing local spatiotemporal patterns of Morse code dots, dashes, and intervals includes: The high-dimensional time-frequency feature sequence is subjected to dimensionality transformation to generate a two-dimensional time-frequency feature matrix; The two-dimensional time-frequency feature matrix is ​​input into a stacked multi-layer two-dimensional convolutional layer. By synchronously sliding the convolutional kernel in the time and frequency dimensions, the energy distribution features of Morse code dots and dashes and the background texture features of the intervals are extracted layer by layer to obtain multi-level local convolutional features. The multi-level local convolutional features are downsampled using pooling operations adapted to the temporal continuity of Morse code to obtain downsampled feature maps. The semantic feature map is obtained by performing nonlinear activation and feature regularization on the downsampled feature map.

[0008] In some embodiments, the step of inputting the two-dimensional time-frequency feature matrix into a stacked multi-layer two-dimensional convolutional layer, and extracting the energy distribution features of Morse code dots and dashes and the background texture features of intervals by synchronously sliding the convolutional kernel in the time and frequency dimensions to obtain multi-level local convolutional features includes: The two-dimensional time-frequency feature matrix is ​​input into a parallel convolution branch targeting two discriminative spatiotemporal patterns of Morse code dots and dashes. The parallel convolution branch includes a short-time local feature convolution branch and a long-time continuous feature convolution branch. Specifically, the short-time local feature convolution branch extracts the short-time local time-frequency features of the Morse code dot signal, and the long-time continuous feature convolution branch extracts the long-time continuous time-frequency features of the Morse code dash signal. Based on the inherent encoding mutual exclusion rules of Morse code dots and dashes, cross-branch feature complementarity verification is performed on the short-term local time-frequency features and the long-term continuous time-frequency features to filter out noise pseudo features extracted from single branches and obtain multi-dimensional convolutional features. The multi-dimensional convolutional features are fused between layers and their dimensions are normalized to obtain multi-level local convolutional features that include different firing rhythm characteristics.

[0009] In some embodiments, the step of performing parallel dual-branch processing on the semantic feature map, wherein the first branch performs channel and spatial weighting to obtain attention-weighted features, and the second branch performs bidirectional temporal modeling to obtain context-aware features, includes: The semantic feature map is transformed in dimension to obtain a shared feature matrix; The shared feature matrix is ​​input into the parallel feature purification branch and the temporal modeling branch. Based on the convolutional attention module in the feature purification branch, the shared feature matrix is ​​subjected to adaptive weighting of the channel dimension and spatial dimension in sequence to obtain attention-weighted features. The shared feature matrix is ​​flattened along the time axis into a temporal feature sequence. Based on the bidirectional long short-term memory network within the temporal modeling branch, forward and backward temporal modeling is performed on the temporal feature sequence to obtain context-aware features containing global temporal semantics.

[0010] In some embodiments, the step of performing adaptive weighting of the shared feature matrix in sequence based on the convolutional attention module within the feature purification branch to obtain attention-weighted features includes: Based on the energy distribution characteristics of Morse code dots and dashes, and using the channel attention units of the convolutional attention module to perform global context aggregation on the shared feature matrix along the channel dimension, channel weight coefficients are generated; the channel weight coefficients are then multiplied by the shared feature matrix to obtain channel-weighted features. Based on the temporal boundary characteristics of the Morse code interval, and using the spatial attention unit of the convolutional attention module to perform spatial dimension context aggregation on the channel weighted features, spatial weight coefficients are generated; the spatial weight coefficients are multiplied by the channel weighted features to obtain the attention weighted features.

[0011] In some embodiments, the step of employing a gating-based adaptive feature fusion module to perform end-to-end learnable dynamic weight allocation and feature fusion on the attention-weighted features and the context-aware features to generate a unified feature representation includes: The attention-weighted features and the context-aware features are respectively aligned in feature dimension and mapped in semantic space to obtain two sets of heterogeneous features to be fused. The two sets of heterogeneous features to be fused are input into an adaptive feature fusion module based on a gating mechanism to generate a contribution representation of the two sets of features to the Morse code decoding task. Based on the contribution characterization, dynamic weight allocation is performed on the two sets of heterogeneous features to be fused, and the fusion ratio of local detail features and global temporal features is adaptively adjusted to obtain two sets of weighted fused features; then, nonlinear fusion and feature regularization are performed on the two sets of weighted fused features to generate the unified feature representation.

[0012] In some embodiments, the step of inputting the two sets of heterogeneous features to be fused into an adaptive feature fusion module based on a gating mechanism to generate contribution representations of the two sets of features to the Morse code decoding task includes: Two sets of heterogeneous features to be fused are spliced ​​together to generate fused correlation features; the fused correlation features are then input into the fully connected layer and nonlinear activation layer of the gated unit to extract the implicit representations related to signal attributes. Based on the implicit representation, dynamic gating weights are generated for the corresponding attention-weighted features and context-aware features, respectively. The dynamic gating weights are used as the contribution of the two sets of heterogeneous features to be fused, and the dynamic gating weights are adaptively adjusted in real time according to the characteristics of the input signal.

[0013] In some embodiments, the step of mapping the unified feature representation to the probability distribution of characters at each time step, decoding the probability sequence using connectionist temporal classification, and outputting the decoded character sequence corresponding to the Morse code includes: The unified feature representation is dimensionally regularized and feature mapped to generate a sequence feature matrix that adapts to the input specifications of the fully connected layer. The sequence feature matrix is ​​input into a fully connected layer. Through linear transformation and nonlinear activation processing, the feature map of each time step is made into the predicted probability of all candidate characters in the corresponding Morse code character set, generating a probability distribution sequence of characters corresponding to each time step. Based on the inherent encoding rules and whitespace mechanism of Morse code, and using a connectionist temporal classification framework to filter invalid information and regularize the path of the probability distribution sequence, a set of candidate decoding paths is obtained. The candidate decoding path set is subjected to global probability optimal screening, and the decoded character sequence corresponding to the Morse code is output.

[0014] Compared with the prior art, the technical solution provided in the first aspect of this application includes at least the following beneficial effects or advantages: The telegraph training method provided in this application obtains a high-dimensional time-frequency feature sequence by preprocessing multi-source audio input data. Then, it extracts the spatiotemporal features of Morse code dot-dash intervals based on a network composed of multi-layer convolution and pooling structures, forming a semantic feature map with clear representation. This adapts to the time-frequency characteristics of Morse code and improves the feature representation of effective signals. A parallel heterogeneous dual-branch architecture is used to simultaneously complete the attention feature purification and the temporal modeling of the bidirectional long short-term memory network, so that the feature purification capability and the temporal context modeling capability reach the optimal state simultaneously. This can effectively suppress noise interference caused by hardware jitter and complex electromagnetic environment, obtain the long-distance temporal dependency pattern of Morse code sequences under different firing rhythms, and improve feature denoising and context understanding capabilities. Dynamic feature fusion based on gating mechanism can adaptively adjust the proportion of local details and global temporal features to form a unified feature with complete representation. It can effectively adapt to the different sending habits and speed differences of different telegraph operators. Then, through feature mapping and connectionist temporal classification decoding, it can achieve end-to-end decoding without fixed threshold constraints. It can still maintain stable and efficient recognition performance in complex electromagnetic interference and key jitter environments, and improve the accuracy and robustness of decoding in telegraph training scenarios.

[0015] Secondly, embodiments of this application provide a message service training system based on an attention mechanism and a recurrent neural network, comprising: The acquisition module is configured to acquire audio input data including analog audio signals or digital audio sequences, and sequentially perform bandpass filtering, frame windowing, Mel frequency cepstral coefficients, multi-order difference feature extraction and standardization on the audio input data to obtain a high-dimensional time-frequency feature sequence. The convolutional feature extraction module is configured to input the high-dimensional time-frequency feature sequence into a convolutional neural network composed of multiple stacked two-dimensional convolutional layers and pooling layers to obtain a semantic feature map containing local spatiotemporal patterns of Morse code dots, dashes and intervals. The dual-branch feature encoding module is configured to perform parallel dual-branch processing on the semantic feature map; wherein, based on the first branch, channel and spatial weighting is performed to obtain attention-weighted features; and based on the second branch, bidirectional temporal modeling is performed to obtain context-aware features. The feature fusion module is configured to adopt an adaptive feature fusion module based on a gating mechanism to perform end-to-end learnable dynamic weight allocation and feature fusion on the attention-weighted features and the context-aware features to generate a unified feature representation. The decoding output module is configured to map the unified feature representation to the probability distribution of the corresponding character at each time step, decode the probability sequence using connectionist temporal classification, and output the decoded character sequence corresponding to the Morse code.

[0016] It is understood that the beneficial effects of the technical solution provided in the second aspect above can be found in the relevant description in the first aspect above, and will not be repeated here.

[0017] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating the message training method based on attention mechanism and recurrent neural network provided in the embodiments of this application; Figure 2 A structural block diagram of a message service training system based on attention mechanism and recurrent neural network provided in an embodiment of this application; Figure 3A structural block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0020] To facilitate understanding of this application, a more complete description will be provided below with reference to the accompanying drawings, which illustrate several embodiments of the present application. However, the present application can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that the disclosure of this application will be thorough and complete.

[0021] It should be noted that, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0022] Please see Figure 1 , Figure 1 This illustration shows a flowchart of a message service training method based on an attention mechanism and a recurrent neural network, according to an embodiment of this application. This embodiment provides a message service training method based on an attention mechanism and a recurrent neural network, specifically including steps S100 to S500.

[0023] Step S100: Acquire audio input data including analog audio signals or digital audio sequences, and sequentially perform bandpass filtering, frame windowing, Mel frequency cepstral coefficients, multi-order differential feature extraction and standardization on the audio input data to obtain a high-dimensional time-frequency feature sequence; In this step, audio input data, including analog audio signals or digital audio sequences, can be acquired through an audio input module. This module may include a real-time signal input interface and a digital file input interface. The real-time signal input interface is used to connect to a physical keypad device; this interface captures the pulse signals generated by the keypad switch and modulates them in real time into a continuous wave analog audio signal of a specific frequency. The digital file input interface is used to process recorded audio files (such as WAV and MP3 formats); the system directly reads their digital audio sequences through a file decoder.

[0024] It should be noted that the bandpass filter is a dedicated filter adapted to the effective carrier frequency band of Morse code, used to filter out out-of-band power frequency interference, equipment noise floor, and clutter; frame segmentation and windowing are based on the short-time stationary signal characteristics of Morse code, dividing the long audio sequence into single-frame signals that conform to the short-time stationary characteristics, thus suppressing spectral leakage; Mel-Frequency Cepstral Coefficients (MFCC) are time-frequency features designed to simulate the characteristics of human hearing, which are highly consistent with the auditory logic of telegraph operators manually identifying Morse code; multi-order differential features are the differential features of MFCC in the time dimension, used to characterize the dynamic change law of the signal, corresponding to the time boundary information of Morse code dot-dash switching and interval start and end; standardization processing is used to eliminate the signal amplitude differences caused by different firing forces and equipment gains, and to unify the feature distribution.

[0025] In some embodiments, analog-to-digital conversion is performed on the analog audio signal using sampling specifications adapted to the effective frequency band of Morse code to obtain a digital audio sequence. Subsequently, the digital audio sequence is subjected to digital bandpass filtering matching the Morse code carrier frequency band to obtain a filtered, clean audio sequence. Based on the clean audio sequence, frame segmentation is performed using frame length and frame shift parameters matching the minimum temporal unit of Morse code. A Hamming window function adapted to the short-time stationary characteristics of audio is applied to each frame of data after segmentation to obtain a windowed and normalized frame sequence. Mel-frequency cepstral coefficients are extracted frame by frame from the windowed and normalized frame sequence as static features to construct a static frequency domain feature sequence. First-order and second-order difference features are calculated along the time axis of the static frequency domain feature sequence to obtain a dynamic time domain feature sequence. The static frequency domain feature sequence is concatenated with the first-order and second-order difference feature sequences along the feature dimension to obtain an initial time-frequency joint feature sequence. Global normalization is then performed based on the initial feature sequence to eliminate feature offsets caused by amplitude differences in different beat signals, resulting in a high-dimensional time-frequency feature sequence.

[0026] For example, for the input analog audio signal, an analog-to-digital converter (ADC) is first performed, digitizing it at a sampling rate of at least 8kHz; audio file selection is not required. Next, a digital bandpass filter with a passband of 400Hz–2000Hz is applied to suppress power line interference, equipment noise, and high-frequency clutter. Then, a short-time analysis is performed on the filtered audio signal: it is divided into frames with a fixed frame length (e.g., 25ms) and a frame shift (e.g., 10ms), and a Hamming window function is applied to each frame. Then, the Mel-frequency cepstral coefficients (MFCCs) of each frame are extracted as static features, and the first-order and second-order differential dynamic features of the MFCCs are concatenated to fuse dynamic temporal information. Finally, the feature sequences of all frames are standardized and used as input to the intelligent recognition module.

[0027] In this embodiment, analog-to-digital conversion is performed by adapting the sampling specifications to the effective frequency band of the Morse code, preserving the core effective information of the Morse code signal. Then, digital bandpass filtering, matched to the carrier frequency band, selectively filters out power frequency interference, equipment noise floor, and out-of-band clutter unrelated to the telegraph signal. This achieves complete preservation of effective information and precise removal of invalid interference from the signal source, laying a clean and stable signal foundation for subsequent processing. Framing is performed using the frame length and frame shift parameters of the smallest timing unit matched to the Morse code, ensuring that each frame of data completely covers the smallest signal unit, avoiding the omission of key information such as the start and end points of dots and dashes. Combined with a Hamming window function that matches the short-time stationary characteristics of the audio, spectral leakage caused by framing can be effectively suppressed, making the spectral characteristics of the single-frame signal smoother and more stable, further improving the accuracy of subsequent feature extraction.

[0028] It should be noted that using the Mel frequency cepstral coefficients as static features aligns with the characteristics of human hearing and the underlying logic of telegraph operators' manual listening. This allows for the accurate acquisition of the inherent energy distribution of a single-frame signal. By concatenating the first-order and second-order differential dynamic features of MFCC with the corresponding first-order and second-order differential dynamic features of the static features, the signal change trend between adjacent frames can be fully restored. This accurately corresponds to the temporal change information of Morse code dot-dash switching and interval start and end, enabling the generated initial feature sequence to simultaneously possess both single-frame detail representation capabilities and inter-frame temporal correlation capabilities, fully covering the static and dynamic core discrimination information of the Morse code signal. Finally, global standardization is used to distribute and normalize the features of the entire sequence, eliminating signal amplitude fluctuations caused by differences in the striking force of different operators and the gain of equipment. This ensures that all features are within a uniform numerical distribution range, maintaining feature consistency without disrupting the inherent temporal variation of the code signal. As a result, the generated high-dimensional time-frequency feature sequence has high discriminative power and strong robustness, and can adapt to different striking habits and the signal output of different acquisition devices. This provides reliable feature input for the stable convergence and accurate recognition of subsequent deep learning networks.

[0029] Step S200: Input the high-dimensional time-frequency feature sequence into a convolutional neural network composed of multiple stacked two-dimensional convolutional layers and pooling layers to obtain a semantic feature map containing local spatiotemporal patterns of Morse code dots, dashes and intervals. In this step, the two-dimensional convolutional layer performs convolution operations simultaneously in the time and frequency dimensions, unlike the conventional one-dimensional temporal convolution, and can simultaneously extract the temporal distribution of Morse code in the time dimension and the spectral characteristics in the frequency dimension; the pooling layer is used to downsample and compress the feature dimension and filter redundant information; the stacked structure is a cascaded structure of multiple convolutions and pooling layers, which can extract features from the bottom texture to the high semantic layer by layer; the local spatiotemporal pattern refers to the unique distribution pattern of Morse code dots, dashes, and intervals in the two-dimensional space of time and frequency, such as the short energy burst of "dots", the continuous energy band of "dashes", and the background texture of the silent segment, that is, dots are short-term narrow-band high-energy regions, dashes are long-term narrow-band high-energy regions, and intervals are low-energy regions.

[0030] In some embodiments, the high-dimensional time-frequency feature sequence is dimensionally adapted and normalized to generate a two-dimensional time-frequency feature matrix adapted to the input specifications of a two-dimensional convolutional neural network. The two-dimensional time-frequency feature matrix is ​​input into a stacked multi-layer two-dimensional convolutional layer. By synchronously sliding the convolutional kernel in the time and frequency dimensions, the energy distribution features of Morse code dots and dashes and the background texture features of the intervals are extracted layer by layer to obtain multi-level local convolutional features. For the multi-level local convolutional features, a pooling operation adapted to the temporal continuity of Morse code is used for downsampling processing to filter redundant features while retaining the temporal boundary information of dots, dashes, and intervals, resulting in a downsampled feature map. The downsampled feature map is then subjected to nonlinear activation and feature normalization to obtain a semantic feature map containing local spatiotemporal patterns of Morse code dots, dashes, and intervals.

[0031] It should be noted that this embodiment designs a multi-level convolutional feature extraction architecture for the time-frequency domain characteristics of Morse code signals. By converting the high-dimensional time-frequency feature sequence into a time-frequency feature matrix adapted to the two-dimensional convolution input, the convolution kernel can slide synchronously in the time and frequency dimensions. This not only accurately obtains the instantaneous energy accumulation and distribution features of Morse code dots and dashes, but also effectively distinguishes the background texture features of the interval regions. Downsampling is performed using pooling operations adapted to the temporal continuity of Morse code. This compresses the feature dimension, filters redundant information, and retains the temporal boundary information of the signal, avoiding the defect of losing the start and end boundaries of dots and dashes that is easily found in general pooling operations. This allows subsequent temporal modeling to accurately obtain the temporal variation law of the signal. Nonlinear activation introduces nonlinear expressive power into the features, enabling the network to fit the complex nonlinear mapping relationship of Morse code. Feature regularization makes the output feature map distribution more stable, avoiding the gradient vanishing problem. The final generated semantic feature map can completely represent the local spatiotemporal pattern of the core units of Morse code.

[0032] In some embodiments, the step of inputting a high-dimensional time-frequency feature sequence into a convolutional neural network composed of multiple stacked two-dimensional convolutional layers and pooling layers to obtain a semantic feature map containing local spatiotemporal patterns of Morse code dots, dashes, and intervals specifically includes: The two-dimensional time-frequency feature matrix is ​​input into parallel convolution branches customized for the two core discriminative spatiotemporal patterns of Morse code dots and dashes. These parallel convolution branches include short-term local feature convolution branches and long-term continuous feature convolution branches. The short-term local feature convolution branches capture the instantaneous energy bursts and interval boundaries of the Morse code dot signals. The long-term continuous feature convolution branches capture the sustained energy distribution and character temporal correlation of the Morse code dash signals. Based on the inherent encoding mutual exclusion rules of Morse code dots and dashes, cross-branch feature complementarity verification is performed on the short-term local and long-term continuous time-frequency features to filter out noise pseudo-features extracted from single branches, resulting in multi-dimensional convolutional features. Subsequently, inter-level feature fusion and dimensional regularization are performed on the multi-dimensional convolutional features to obtain multi-level local convolutional features including different beat rhythm characteristics.

[0033] It should be noted that the dot signal of Morse code is a short-term instantaneous energy burst, while the swipe signal is a long-term continuous energy distribution. The spatiotemporal patterns of the two signals are very different. A single convolutional structure cannot simultaneously and optimally adapt to the feature extraction requirements of the two signals. The customized design of dual parallel branches can be optimally optimized for the characteristics of the two types of core signals respectively, while covering both short-term and long-term feature patterns.

[0034] Furthermore, in Morse code, dots, dashes, and intervals are mutually exclusive at the same temporal position and will not appear simultaneously. This inherent rule allows for cross-validation of features extracted from two branches, filtering out false features extracted by a single branch due to noise interference, resulting in higher accuracy of the final output convolutional features. Feature fusion and dimensional regularization between layers enable multi-scale features to complement each other, adapting to signal characteristics under different firing rhythms. Whether firing quickly or slowly, effective local spatiotemporal features can be stably extracted.

[0035] In this embodiment, the two-dimensional time-frequency feature matrix is ​​input into parallel convolution branches of short-term local features and long-term continuous features customized for the two core discriminative spatiotemporal patterns of Morse code dots and dashes. These two branches are optimized specifically for the instantaneous energy bursts and interval boundary characteristics of dot signals, and the continuous energy distribution and character temporal correlation characteristics of dash signals, respectively. Simultaneously, it accurately captures the two vastly different spatiotemporal patterns of short-term instantaneous and long-term continuous signals, fully covering the feature representation of the core units of the Morse code under different firing rhythms. Based on the inherent encoding mutual exclusion rules of the Morse code dot-dash intervals, cross-branch feature complementarity verification is performed, utilizing the inherent encoding logic of the Morse code itself. The features output from the two branches are cross-validated to filter out pseudo-features that do not conform to the coding rules due to noise interference in a single branch, making the feature representation of the effective signal more accurate. At the same time, the complementarity of the features of the two branches is strengthened to avoid invalid feature redundancy. Then, through feature fusion and dimension regularization between layers, the two complementary spatiotemporal features can be deeply integrated to generate multi-level local convolutional features that take into account the characteristics of different beat rhythms. This makes the final output features contain both fine local boundary information of dot-stroke intervals and long-term temporal correlation information of corresponding characters, providing a highly discriminative and robust feature foundation for subsequent feature purification and temporal modeling.

[0036] Step S300: Perform parallel dual-branch processing on the semantic feature map; wherein, based on the first branch, channel and spatial weighting is performed to obtain attention-weighted features; based on the second branch, bidirectional temporal modeling is performed to obtain context-aware features; In this step, the semantic feature map undergoes dimensionality adaptation and feature normalization to generate a shared feature matrix that simultaneously adapts to the input specifications of the convolutional attention module and the bidirectional long short-term memory network. This shared feature matrix is ​​then input into the parallel feature purification branch and the temporal modeling branch. The parallel heterogeneous dual branches are two completely independent and synchronously executed processing branches, with no sequential processing relationship. Compared to the traditional serial architecture, this avoids the problem of error propagation and accumulation from preceding modules. Each branch can optimize for its own processing objective without mutual constraint. Dimensional adaptation and feature normalization generate a shared feature matrix that simultaneously adapts to the input requirements of both branches, ensuring that both branches process the exact same original features and avoiding feature bias caused by input differences.

[0037] In one example, the deep feature maps extracted by the CNN are fed in parallel into two structurally and functionally complementary neural network units. These units perform in-depth processing from two dimensions: "feature importance selection" and "sequence context understanding." The two complementary neural network units include a Convolutional Block Attention Module (CBAM) and a Bidirectional Long Short-Term Memory (Bi-LSTM) module. The attention module passes through a channel attention module and then a spatial attention module. The channel attention adaptively recalibrates the weights of each frequency channel, highlighting key frequency bands that contribute significantly to recognition. Spatial attention (primarily the time dimension in the time-frequency map) focuses on the time regions in the feature map that contain key information (such as character start points). The output of this branch is a refined attention-weighted feature. The temporal context module flattens and rearranges the feature map spatially before inputting it into the bidirectional long short-term memory network. The temporal context module scans the entire sequence bidirectionally, modeling the long-distance dynamic dependencies and combinatorial logic between points, dashes, and intervals. Its output is a context-aware feature containing rich global temporal semantics.

[0038] In some embodiments, the semantic feature map is processed in parallel with two branches. Specifically, the steps of performing channel and spatial weighting based on the first branch to obtain attention-weighted features, and performing bidirectional temporal modeling based on the second branch to obtain context-aware features, include: adapting the semantic feature map to dimensions and regularizing the features to generate a shared feature matrix that simultaneously adapts to the input specifications of the convolutional attention module and the bidirectional long short-term memory network; inputting the shared feature matrix into a parallel feature purification branch and a temporal modeling branch, wherein, based on the convolutional attention module within the feature purification branch, adaptive weighting of the channel and spatial dimensions is performed sequentially on the shared feature matrix to obtain purified attention-weighted features; flattening the shared feature matrix along the time axis into a temporal feature sequence, specifically, performing global average pooling of the features in the shared feature matrix along the time axis to generate a frame-by-frame temporal feature sequence; and using the bidirectional long short-term memory network within the temporal modeling branch to perform forward and backward temporal modeling on the frame-by-frame temporal feature sequence to capture the long-distance temporal dependence of Morse code dots, dashes, and intervals on global encoding rules, thereby obtaining context-aware features containing global temporal semantics.

[0039] In this embodiment, the semantic feature map undergoes dimensional adaptation and feature normalization to generate a shared feature matrix that simultaneously adapts to the input specifications of the convolutional attention module and the bidirectional long short-term memory network. This ensures that the two subsequent processing branches operate based on complete features from the same source, avoiding information bias and representation misalignment caused by differences in input features, and laying a unified input foundation for synchronous dual-branch processing. The shared feature matrix is ​​synchronously input into the parallel feature purification branch and temporal modeling branch. The two branches execute their processing flows independently and synchronously, allowing each branch to achieve optimal optimization for its own processing objectives without being constrained by the sequential processing logic of a serial architecture. This maximizes the synchronous performance of feature purification and temporal modeling capabilities.

[0040] In some embodiments, the step of performing adaptive weighting of the channel dimension and spatial dimension on the shared feature matrix sequentially based on the convolutional attention module within the feature purification branch to obtain attention-weighted features specifically includes: performing global context aggregation of the channel dimension on the shared feature matrix based on the energy distribution characteristics of Morse code dot and dash signals and the channel attention unit of the convolutional attention module to generate channel weight coefficients; multiplying the channel weight coefficients with the shared feature matrix to obtain channel-weighted features; performing context aggregation of the spatial dimension on the channel-weighted features based on the temporal boundary characteristics of the Morse code interval and the spatial attention unit of the convolutional attention module to generate spatial weight coefficients; and multiplying the spatial weight coefficients with the channel-weighted features to obtain attention-weighted features.

[0041] In this embodiment, based on the convolutional attention module within the feature purification branch, adaptive weighting of the channel and spatial dimensions is sequentially applied to the shared feature matrix. This selectively strengthens the feature weights of effective signal regions such as Morse code dots, dashes, and intervals, while simultaneously suppressing irrelevant noise and background interference, thus achieving full-dimensional feature purification. The resulting attention-weighted features can more accurately characterize the local key details and discriminative information of the Morse code signal. Based on the bidirectional long short-term memory network within the temporal modeling branch, a forward and backward bidirectional scan of the shared feature matrix is ​​performed, capturing the full-link contextual association information of the preceding and following sequences of the Morse code sequence. This models the long-distance temporal dependencies of Morse code dots, dashes, and intervals and extracts global encoding rules. The resulting context-aware features can completely reconstruct the global temporal semantic logic of the Morse code sequence. The two types of features output synchronously from the two branches form a completely complementary representation system, which takes into account both the accuracy of local details of the code signal and the integrity of the global timing. This improves the feature fusion and decoding output quality and robustness, adapts to the needs of telegraph signal processing under different shooting habits and different interference environments, and provides a solid guarantee for the stable operation and accurate output of the entire decoding process.

[0042] Step S400: An adaptive feature fusion module based on a gating mechanism is used to perform end-to-end learnable dynamic weight allocation and feature fusion on the attention-weighted features and the context-aware features to generate a unified feature representation; In this embodiment, feature dimension alignment and semantic space mapping are performed on attention-weighted features and context-aware features, respectively, to generate two sets of heterogeneous features to be fused that are dimensionally matched and can be collaboratively computed. The two sets of heterogeneous features to be fused are input into an adaptive feature fusion module based on a gating mechanism. Through an end-to-end learnable gating unit, the contribution representation of the Morse code decoding task corresponding to the two sets of features is learned. According to the contribution representation, dynamic weight allocation is performed on the two sets of heterogeneous features to be fused, and the fusion ratio of local detail features and global temporal features is adaptively adjusted to obtain two sets of weighted fused features. Nonlinear fusion and feature regularization are then performed on the two sets of weighted fused features to generate a unified feature representation that takes into account both local key details and global temporal context.

[0043] It should be understood that the unified feature representation is a weighted fusion feature, the attention-weighted feature is a highly discriminative feature output by the front-end feature purification branch that focuses on the local key details of Morse code dots, dashes, and intervals, and the context-aware feature is a semantic feature output by the front-end temporal modeling branch that represents the global temporal coding rules of the Morse code sequence. The two types of features are naturally heterogeneous in terms of dimensionality and semantic expression space. Feature dimensional alignment and semantic space mapping refer to mapping the two types of heterogeneous features to the same feature space through independent linear transformation layers, eliminating dimensional differences and semantic misalignment, and ensuring that the two types of features can be collaboratively fused for computation.

[0044] In this embodiment, the adaptive feature fusion module based on the gating mechanism is an end-to-end learnable fusion module customized for two types of heterogeneous features. Unlike the conventional fixed-weight fusion method, it can automatically adjust the fusion strategy according to the real-time characteristics of the input signal. The contribution characterization is a quantitative representation of the two sets of features to be fused, which is an indicator of the importance of the correct decoding of Morse code under the current input signal. According to the contribution characterization, dynamic weight allocation is performed on the two sets of heterogeneous features to be fused, and the fusion ratio of local detail features and global temporal features is adaptively adjusted to obtain two sets of weighted fused features. Then, nonlinear fusion and feature regularization are performed on the two sets of weighted fused features to generate a unified feature representation that takes into account both local key details and global temporal context.

[0045] In this way, by performing feature dimension alignment and semantic space mapping on the two types of heterogeneous features output by the parallel dual branches of the front end, local detail features and global temporal features belonging to different semantic spaces and with different dimension specifications are mapped to the same feature space, eliminating the dimensional differences and semantic misalignment of the two types of features, ensuring the synergy and effectiveness of subsequent fusion calculations, and avoiding feature conflicts and information redundancy caused by direct fusion of heterogeneous features.

[0046] In some embodiments, two sets of heterogeneous features to be fused are input into an adaptive feature fusion module based on a gating mechanism. Through an end-to-end learnable gating unit, the module learns and generates contribution representations of the two sets of features for the Morse code decoding task. Specifically, this includes: concatenating the two sets of heterogeneous features to generate fused correlation features containing both local and global information; inputting the fused correlation features into the fully connected layer and nonlinear activation layer of the gating unit to learn implicit representations of the firing rhythm and noise interference level corresponding to the current input signal; generating dynamic gating weights for the corresponding attention-weighted features and context-aware features based on the implicit representations, with the dynamic gating weights adaptively adjusted in real time according to the characteristics of the input signal; and using the dynamic gating weights as contribution representations of the two sets of heterogeneous features to be fused for subsequent weight allocation.

[0047] It should be explained that the fusion-related features combine the local detail information of attention-weighted features with the global temporal information of context-aware features, which can comprehensively reflect the complete characteristics of the current input signal. The implicit representation is a high-dimensional feature obtained by the gating unit through nonlinear learning, which can implicitly represent the intrinsic characteristics of the current input signal, such as the firing speed, the degree of fluctuation in firing rhythm, and the strength of background noise interference. For example, the dynamic gating weight is a learnable parameter with a value range between 0 and 1. The sum of the dynamic gating weights corresponding to the two sets of features is 1, which can change in real time with the intrinsic characteristics of the input signal without the need for manual setting of fixed values.

[0048] In this embodiment, two sets of features to be fused based on dimension matching are concatenated to generate fused associated features. This integrates the local details and global temporal information of the current input signal, providing a comprehensive and complete signal feature foundation for the weight learning of the gating unit. The gating unit learns the implicit representation of the firing rhythm and noise interference level of the input signal through fully connected layers and nonlinear activation layers, mining the intrinsic characteristics of the current input signal and providing accurate judgment criteria for the generation of fused weights. The dynamic gating weights generated based on the implicit representations can be adaptively adjusted in real time according to the characteristics of the input signal. In scenarios with strong signal and noise interference, it automatically increases the fusion ratio of local detail features, strengthens the discriminative expression of the effective signal, and enhances the firing rhythm. In scenarios with significant rhythm fluctuations and complex temporal logic, the system automatically increases the fusion ratio of global temporal features, strengthens the global constraints of Morse code encoding rules, and uses dynamic gating weights as a representation of the contribution of the two types of features to complete dynamic weight allocation. This achieves deep synergy between the two complementary features, rather than simple feature superposition. The final unified feature representation obtained through nonlinear fusion and feature regularization takes into account both the local key details of Morse code dot-dash intervals and the global temporal context of the entire sequence. It has both high discriminative power and strong robustness, adapting to the personalized sending habits of different telegraph operators and complex and ever-changing application environments. This provides a high-quality feature foundation for subsequent character probability mapping and decoding output, and improves the stability of the end-to-end decoding performance.

[0049] Step S500: Map the unified feature representation to the probability distribution of the corresponding character at each time step, use connectionist temporal classification to decode the probability sequence, and output the decoded character sequence corresponding to the Morse code.

[0050] Specifically, the unified feature representation undergoes dimensionality regularization and feature mapping to generate a sequence feature matrix adapted to the input specifications of the fully connected layer. The unified feature representation is a high-dimensional fusion feature sequence output by the front-end adaptive feature fusion module, which takes into account both the local key details of Morse code and the global temporal context. Dimensional regularization and feature mapping refer to the standardization and adjustment of the dimensional arrangement and numerical range of the unified feature representation through linear transformation operations, so that it fully matches the input dimensionality requirements of the subsequent fully connected layer, ensuring that the complete feature information after fusion enters the subsequent processing stage without deviation or omission.

[0051] Furthermore, the sequence feature matrix is ​​input into the fully connected layer. Through linear transformation and nonlinear activation processing, the feature at each time step is mapped to the predicted probability of all candidate characters in the corresponding Morse code character set, generating a probability distribution sequence of characters at each time step. Based on the inherent encoding rules and whitespace mechanism of Morse code, and using a connectionist temporal classification framework, invalid information is filtered and path regularization is performed on the probability distribution sequence to obtain a set of candidate decoding paths. Subsequently, the set of candidate decoding paths is subjected to global probability optimization screening to output the decoded character sequence corresponding to the Morse code.

[0052] It should be noted that the Morse code-specific character set is a character set customized for telegraph training and communication scenarios. It includes punctuation marks commonly used in telegraph communication, as well as whitespace characters required by the connectionist temporal classification framework, covering the decoding needs of the entire telegraph training scenario. Linear transformation and nonlinear activation processing can map high-dimensional temporal features to a character-dimensional prediction space, so that the output of each time step corresponds to the probability of occurrence of all candidate characters at that time. The generated probability distribution sequence completely represents the predicted probability distribution of each candidate character in the entire temporal range. The inherent encoding rules of Morse code refer to the dots, dashes, spaces, etc. in the Morse code system. The standard combinational logic and character mapping specifications of the separator, as well as the inherent characteristic of mutual exclusion of dots, dashes, and intervals at the same time position; the whitespace mechanism is the core mechanism of the connectionist temporal classification framework, used to handle the problem of mismatch between the length of the input feature sequence and the output character sequence, and can mark meaningless repeated predictions and interval regions without the need for manual character-level annotation of each frame of signal; invalid information filtering and path regularization refer to merging continuous repeated invalid predictions and filtering meaningless whitespace based on the whitespace mechanism, while combining the inherent encoding rules of Morse code to remove prediction paths that do not conform to the standard encoding logic, thus completing the initial screening and regularization of the decoding path.

[0053] The telegraph training method provided in the above embodiments obtains a high-dimensional time-frequency feature sequence by preprocessing multi-source audio input data. Then, based on a network composed of multi-layer convolutional and pooling structures, it accurately extracts the spatiotemporal features of Morse code dot-dash intervals, forming a semantic feature map with clear representation. This adapts to the unique time-frequency characteristics of Morse code and further enhances the feature expression of effective signals. A parallel heterogeneous dual-branch architecture is used to simultaneously complete attention feature purification and temporal modeling of the bidirectional long short-term memory network, so that feature purification capability and temporal context modeling capability reach the optimal state simultaneously. At the same time, it can effectively suppress noise interference caused by hardware jitter and complex electromagnetic environment, and fully capture the long-distance temporal dependency pattern of Morse code sequences under different firing rhythms, so that feature denoising and context understanding capabilities are synergistically improved. Dynamic feature fusion based on gating mechanism performs end-to-end learnable dynamic weight allocation and feature fusion on two types of heterogeneous features. It can adaptively adjust the proportion of local details and global temporal features to form a unified feature with complete representation, effectively adapting to the firing habits and speed differences of different telegraph operators. Furthermore, by using feature mapping and connectionist temporal classification decoding, end-to-end decoding without fixed threshold constraints is achieved. This ensures stable and efficient recognition performance even under complex electromagnetic interference and key jitter environments, thereby improving the accuracy and robustness of decoding in telegraph training scenarios.

[0054] Please see Figure 2 , Figure 2This embodiment illustrates a message service training system based on an attention mechanism and a recurrent neural network. In this embodiment, the message service training system based on an attention mechanism and a recurrent neural network is used to perform the above-described... Figure 1 The steps in the corresponding embodiments. Please refer to the details. Figure 1 as well as Figure 1 The relevant descriptions in the corresponding embodiments are shown below. For ease of explanation, only the parts relevant to this embodiment are shown. See also... Figure 2 The message training system 200 based on attention mechanisms and recurrent neural networks includes: The acquisition module 210 is configured to acquire audio input data including analog audio signals or digital audio sequences, and sequentially perform bandpass filtering, frame windowing, Mel frequency cepstral coefficient extraction, multi-order difference feature extraction and standardization on the audio input data to obtain a high-dimensional time-frequency feature sequence. The convolutional feature extraction module 220 is configured to input the high-dimensional time-frequency feature sequence into a convolutional neural network composed of multiple stacked two-dimensional convolutional layers and pooling layers to obtain a semantic feature map containing local spatiotemporal patterns of Morse code dots, dashes and intervals. The dual-branch feature encoding module 230 is configured to perform parallel dual-branch processing on the semantic feature map; wherein, based on the first branch, channel and spatial weighting is performed to obtain attention-weighted features; and based on the second branch, bidirectional temporal modeling is performed to obtain context-aware features. The feature fusion module 240 is configured to adopt an adaptive feature fusion module based on a gating mechanism to perform end-to-end learnable dynamic weight allocation and feature fusion on the attention-weighted features and the context-aware features to generate a unified feature representation. The decoding output module 250 is configured to map the unified feature representation to the probability distribution of the corresponding character at each time step, decode the probability sequence using connectionist temporal classification, and output the decoded character sequence corresponding to the Morse code.

[0055] The telegraph operator training system provided in this embodiment obtains a high-dimensional time-frequency feature sequence by preprocessing multi-source audio input data. Then, based on a network composed of multi-layer convolutional and pooling structures, it accurately extracts the spatiotemporal features of Morse code dot-dash intervals, forming a semantic feature map with clear representation. This adapts to the time-frequency characteristics of Morse code and further improves the feature representation of effective signals. A parallel heterogeneous dual-branch architecture is used to simultaneously complete attention feature purification and temporal modeling of a bidirectional long short-term memory network, so that feature purification capability and temporal context modeling capability reach the optimal state simultaneously. At the same time, it effectively suppresses noise interference caused by hardware jitter and complex electromagnetic environment, and completely extracts the long-distance temporal dependency rules of Morse code sequences under different firing rhythms, so that feature denoising and context understanding capabilities are synergistically improved. Dynamic feature fusion based on a gating mechanism performs end-to-end learnable dynamic weight allocation and feature fusion on two types of heterogeneous features. It can adaptively adjust the proportion of local details and global temporal features to form a unified feature with complete representation, adapting to the firing habits and speed differences of different telegraph operators. Furthermore, by using feature mapping and connectionist temporal classification decoding, end-to-end decoding without fixed threshold constraints is achieved. This ensures stable and efficient recognition performance even under complex electromagnetic interference and key jitter environments, thereby improving the accuracy and robustness of decoding in telegraph training scenarios.

[0056] It should be understood that, Figure 2 The block diagram of the reporting training system based on attention mechanism and recurrent neural network shown below illustrates that each module is used to perform... Figure 1 The steps in the corresponding embodiments, and for Figure 1 The steps in the corresponding embodiments have been explained in detail in the above embodiments. Please refer to them for details. Figure 1 as well as Figure 1 The relevant descriptions in the corresponding embodiments will not be repeated here.

[0057] In some embodiments, an electronic device is also provided, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the service training method based on attention mechanism and recurrent neural network described in the above embodiments.

[0058] Please see Figure 3 , Figure 3This diagram illustrates a structural block diagram of an electronic device provided in this embodiment. The server 500 of this electronic device includes a processor 501, a memory 502, and a computer program 503 stored in the memory 502 and executable on the processor 501, such as a program for a message service training method based on an attention mechanism and a recurrent neural network. When the processor 501 executes the computer program 503, it implements the steps of the message service training method based on an attention mechanism and a recurrent neural network in the above embodiments, for example... Figure 1 Steps S100 to S500 in the corresponding embodiment. Alternatively, the processor 501 executes the computer program 503 to implement the above. Figure 2 The functions of each module in the corresponding embodiments, for example, Figure 2 For details on the functions of the modules shown (e.g., module 210), please refer to [link / reference]. Figure 2 The relevant descriptions in the corresponding embodiments are not repeated here.

[0059] For example, computer program 503 can be divided into one or more units, one or more units are stored in memory 502 and executed by processor 501 to complete the technical solution provided in the above embodiments. One or more units can be a series of computer program instruction segments capable of performing a specific function, which are used to describe the execution process of computer program 503 in server 500.

[0060] The electronic device may include, but is not limited to, a processor 501 and a memory 502. Those skilled in the art will understand that... Figure 3 This is merely an example of server 500 in an electronic device and does not constitute a limitation on server 500. It may include more or fewer components than shown, or combine certain components, or different components. For example, a turntable terminal device may also include input / output terminal devices, network access terminal devices, buses, etc.

[0061] The processor 501 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0062] The memory 502 can be an internal storage unit of the server 500, such as the server 500's hard drive or memory. The memory 502 can also be an external storage terminal device of the server 500, such as a plug-in hard drive, SmartMediaCard (SMC), SecureDigital (SD) card, or FlashCard equipped on the server 500. Furthermore, the memory 502 can include both internal storage units and external storage terminal devices of the server 500. The memory 502 is used to store computer programs and other programs and data required by the turntable terminal device. The memory 502 can also be used to temporarily store data that has been output or will be output.

[0063] In one embodiment, a computer-readable storage medium is also provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the message training method based on attention mechanism and recurrent neural network as described in the above embodiments.

[0064] Furthermore, the functional units in the various embodiments of this application 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.

[0065] 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. The computer-readable storage medium can be non-volatile or volatile. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable storage medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0066] The terms "first," "second," "third," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects and not to describe a particular order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, it may include a series of steps or units, or optionally, steps or units not listed, or other steps or units inherent to these processes, methods, products, or devices.

[0067] The accompanying drawings show only the portions relevant to this application, not all of them. Before discussing exemplary embodiments in more detail, it should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe operations (or steps) as sequential processes, many of these operations may be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations may be rearranged. The process may be terminated when its operation is completed, but may also have additional steps not included in the drawings. The process may correspond to a method, function, procedure, subroutine, subprogram, etc.

[0068] The terms “component,” “module,” “system,” “unit,” etc., used in this specification are used to refer to computer-related entities, hardware, firmware, combinations of hardware and software, software, or software in execution. For example, a unit can be, but is not limited to, a process running on a processor, a processor, an object, an executable file, a thread of execution, a program, and / or distributed between two or more computers. Furthermore, these units can be executed from various computer-readable media on which various data structures are stored. Units can communicate, for example, via local and / or remote processes based on signals having one or more data packets (e.g., data from a second unit interacting with another unit between a local system, a distributed system, and / or a network; for example, the Internet interacting with other systems via signals).

[0069] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example.

[0070] Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The reference to "embodiment" herein means that a specific feature, structure, or characteristic described in connection with an embodiment can be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily indicate the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0071] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

Claims

1. A message training method based on attention mechanism and recurrent neural network, characterized in that, include: Audio input data, including analog audio signals or digital audio sequences, is acquired. The audio input data is then subjected to bandpass filtering, frame windowing, Mel frequency cepstral coefficients, multi-order difference feature extraction, and standardization processing in sequence to obtain a high-dimensional time-frequency feature sequence. The high-dimensional time-frequency feature sequence is input into a convolutional neural network composed of multiple stacked two-dimensional convolutional layers and pooling layers to obtain a semantic feature map containing local spatiotemporal patterns of Morse code dots, dashes and intervals. The semantic feature map is processed in parallel with two branches. The first branch performs channel and spatial weighting to obtain attention-weighted features, and the second branch performs bidirectional temporal modeling to obtain context-aware features. An adaptive feature fusion module based on a gating mechanism is used to perform end-to-end learnable dynamic weight allocation and feature fusion on the attention-weighted features and the context-aware features to generate a unified feature representation. The unified feature representation is mapped to the probability distribution of the corresponding character at each time step. The probability sequence is decoded using connectionist temporal classification, and the decoded character sequence corresponding to the Morse code is output.

2. The message training method based on attention mechanism and recurrent neural network according to claim 1, characterized in that, The steps of acquiring audio input data, including analog audio signals or digital audio sequences, and sequentially performing bandpass filtering, frame windowing, Mel-frequency cepstral coefficients, multi-order difference feature extraction, and standardization on the audio input data to obtain a high-dimensional time-frequency feature sequence include: The analog audio signal is converted from analog to digital using a sampling specification adapted to the effective frequency band of Morse code to obtain a digital audio sequence; The digital audio sequence is subjected to digital bandpass filtering that matches the Morse code carrier frequency band to obtain a filtered, clean audio sequence. Based on the pure audio sequence, the frame length and frame shift parameters of the minimum timing unit of the matching Morse code are used for frame division. A Hamming window function that adapts to the short-time stationary characteristics of the audio is applied to each frame of data after frame division to obtain the windowed and normalized frame sequence. The Mel frequency cepstral coefficients are extracted frame by frame from the frame sequence as static features. Each static feature is then concatenated with its corresponding first-order difference feature and second-order difference feature along the feature dimension to obtain an initial feature sequence that fuses the Morse code timing variation information. Global standardization is then performed based on the initial feature sequence to obtain the high-dimensional time-frequency feature sequence.

3. The message training method based on attention mechanism and recurrent neural network according to claim 1, characterized in that, The step of inputting the high-dimensional time-frequency feature sequence into a convolutional neural network composed of multiple stacked two-dimensional convolutional layers and pooling layers to obtain a semantic feature map containing local spatiotemporal patterns of Morse code dots, dashes, and intervals includes: The high-dimensional time-frequency feature sequence is subjected to dimensionality transformation to generate a two-dimensional time-frequency feature matrix; The two-dimensional time-frequency feature matrix is ​​input into a stacked multi-layer two-dimensional convolutional layer. By synchronously sliding the convolutional kernel in the time and frequency dimensions, the energy distribution features of Morse code dots and dashes and the background texture features of the intervals are extracted layer by layer to obtain multi-level local convolutional features. The multi-level local convolutional features are downsampled using pooling operations adapted to the temporal continuity of Morse code to obtain downsampled feature maps. The semantic feature map is obtained by performing nonlinear activation and feature regularization on the downsampled feature map.

4. The message training method based on attention mechanism and recurrent neural network according to claim 3, characterized in that, The step of inputting the two-dimensional time-frequency feature matrix into a stacked multi-layer two-dimensional convolutional layer, and extracting the energy distribution features of Morse code dots and dashes and the background texture features of the intervals by synchronously sliding the convolutional kernel in the time and frequency dimensions to obtain multi-level local convolutional features includes: The two-dimensional time-frequency feature matrix is ​​input into a parallel convolution branch targeting two discriminative spatiotemporal patterns of Morse code dots and dashes. The parallel convolution branch includes a short-time local feature convolution branch and a long-time continuous feature convolution branch. Specifically, the short-time local feature convolution branch extracts the short-time local time-frequency features of the Morse code dot signal, and the long-time continuous feature convolution branch extracts the long-time continuous time-frequency features of the Morse code dash signal. Based on the inherent encoding mutual exclusion rules of Morse code dots and dashes, cross-branch feature complementarity verification is performed on the short-term local time-frequency features and the long-term continuous time-frequency features to filter out noise pseudo features extracted from single branches and obtain multi-dimensional convolutional features. The multi-dimensional convolutional features are fused between layers and their dimensions are normalized to obtain multi-level local convolutional features that include different firing rhythm characteristics.

5. The message training method based on attention mechanism and recurrent neural network according to claim 1, characterized in that, The step of performing parallel dual-branch processing on the semantic feature map, wherein channel and spatial weighting is performed based on the first branch to obtain attention-weighted features, and bidirectional temporal modeling is performed based on the second branch to obtain context-aware features, includes: The semantic feature map is transformed in dimension to obtain a shared feature matrix; The shared feature matrix is ​​input into the parallel feature purification branch and the temporal modeling branch. Based on the convolutional attention module in the feature purification branch, the shared feature matrix is ​​subjected to adaptive weighting of the channel dimension and spatial dimension in sequence to obtain attention-weighted features. The shared feature matrix is ​​flattened along the time axis into a temporal feature sequence. Based on the bidirectional long short-term memory network within the temporal modeling branch, forward and backward temporal modeling is performed on the temporal feature sequence to obtain context-aware features containing global temporal semantics.

6. The message training method based on attention mechanism and recurrent neural network according to claim 5, characterized in that, The step of performing adaptive weighting of the shared feature matrix in sequence by the convolutional attention module within the feature purification branch to obtain attention-weighted features includes: Based on the energy distribution characteristics of Morse code dots and dashes, and using the channel attention units of the convolutional attention module to perform global context aggregation on the shared feature matrix along the channel dimension, channel weight coefficients are generated; the channel weight coefficients are then multiplied by the shared feature matrix to obtain channel-weighted features. Based on the temporal boundary characteristics of the Morse code interval, and using the spatial attention unit of the convolutional attention module to perform spatial dimension context aggregation on the channel weighted features, spatial weight coefficients are generated; the spatial weight coefficients are multiplied by the channel weighted features to obtain the attention weighted features.

7. The message training method based on attention mechanism and recurrent neural network according to claim 1, characterized in that, The step of employing a gating-based adaptive feature fusion module to perform end-to-end learnable dynamic weight allocation and feature fusion on the attention-weighted features and the context-aware features to generate a unified feature representation includes: The attention-weighted features and the context-aware features are respectively aligned in feature dimension and mapped in semantic space to obtain two sets of heterogeneous features to be fused. The two sets of heterogeneous features to be fused are input into an adaptive feature fusion module based on a gating mechanism to generate a contribution representation of the two sets of features to the Morse code decoding task. Based on the contribution characterization, dynamic weight allocation is performed on the two sets of heterogeneous features to be fused, and the fusion ratio of local detail features and global temporal features is adaptively adjusted to obtain two sets of weighted fused features; then, nonlinear fusion and feature regularization are performed on the two sets of weighted fused features to generate the unified feature representation.

8. The message training method based on attention mechanism and recurrent neural network according to claim 7, characterized in that, The step of inputting the two sets of heterogeneous features to be fused into an adaptive feature fusion module based on a gating mechanism to generate contribution representations of the two sets of features to the Morse code decoding task includes: Two sets of heterogeneous features to be fused are spliced ​​together to generate fused correlation features; the fused correlation features are then input into the fully connected layer and nonlinear activation layer of the gated unit to extract the implicit representations related to signal attributes. Based on the implicit representation, dynamic gating weights are generated for the corresponding attention-weighted features and context-aware features, respectively. The dynamic gating weights are used as the contribution of the two sets of heterogeneous features to be fused, and the dynamic gating weights are adaptively adjusted in real time according to the characteristics of the input signal.

9. The message training method based on attention mechanism and recurrent neural network according to claim 1, characterized in that, The steps of mapping the unified feature representation to the probability distribution of characters at each time step, decoding the probability sequence using connectionist temporal classification, and outputting the decoded character sequence corresponding to the Morse code include: The unified feature representation is dimensionally regularized and feature mapped to generate a sequence feature matrix that adapts to the input specifications of the fully connected layer. The sequence feature matrix is ​​input into a fully connected layer. Through linear transformation and nonlinear activation processing, the feature map of each time step is made into the predicted probability of all candidate characters in the corresponding Morse code character set, generating a probability distribution sequence of characters corresponding to each time step. Based on the inherent encoding rules and whitespace mechanism of Morse code, and using a connectionist temporal classification framework to filter invalid information and regularize the path of the probability distribution sequence, a set of candidate decoding paths is obtained. The candidate decoding path set is subjected to global probability optimal screening, and the decoded character sequence corresponding to the Morse code is output.

10. A telecom training system based on attention mechanism and recurrent neural network, characterized in that, include: The acquisition module is configured to acquire audio input data including analog audio signals or digital audio sequences, and sequentially perform bandpass filtering, frame windowing, Mel frequency cepstral coefficients, multi-order difference feature extraction and standardization on the audio input data to obtain a high-dimensional time-frequency feature sequence. The convolutional feature extraction module is configured to input the high-dimensional time-frequency feature sequence into a convolutional neural network composed of multiple stacked two-dimensional convolutional layers and pooling layers to obtain a semantic feature map containing local spatiotemporal patterns of Morse code dots, dashes and intervals. The dual-branch feature encoding module is configured to perform parallel dual-branch processing on the semantic feature map; wherein, based on the first branch, channel and spatial weighting is performed to obtain attention-weighted features; and based on the second branch, bidirectional temporal modeling is performed to obtain context-aware features. The feature fusion module is configured to adopt an adaptive feature fusion module based on a gating mechanism to perform end-to-end learnable dynamic weight allocation and feature fusion on the attention-weighted features and the context-aware features to generate a unified feature representation. The decoding output module is configured to map the unified feature representation to the probability distribution of the corresponding character at each time step, decode the probability sequence using connectionist temporal classification, and output the decoded character sequence corresponding to the Morse code.

Images