A universal carrier modulation method and system

Abstract

The present application provides a universal carrier modulation method and system, which relates to the technical field of carrier modulation. The method includes: generating a modulated signal with different modulation modes, namely a first signal, obtaining a signal to be modulated, processing the first signal, establishing a first prediction model, using the processed first signal and the modulation mode of the first signal to train the first prediction model, using the trained first prediction model to analyze the modulated signal to be processed, and obtaining the modulation mode of the modulated signal to be processed. The present application can combine the inherent characteristics of the modulated signal with deep learning to automatically identify the modulation mode of the modulated signal.

Description

Technical Field

[0001] The present application relates to the technical field of carrier modulation, and in particular to a general carrier modulation method and system. Background Art

[0002] The development of human society has always been accompanied by the progress of communication technology. From beacon fires to satellite reconnaissance, from wild geese to the era of the Internet, human communication technology has developed rapidly and made qualitative leaps, and communication means have become increasingly rich, efficient and complex. The purpose of communication is to transmit information quickly and accurately. Looking back at the brief history of the development of communication technology, although it is only more than a hundred years, the development of communication technology can be said to be changing with each passing day. After more than a hundred years of development, communication technology has undergone a transition from analog to digital. In the process of the development of communication technology, in order to improve the quality and efficiency of communication, more efficiently utilize channel capacity, and meet the needs of different users, communication signals adopt different modulation systems. The modulation of communication signals can be divided into analog modulation and digital modulation. Analog modulation can be divided into amplitude modulation and frequency phase modulation; the modulation methods of digital signals are more diverse, which can be divided into amplitude shift keying (ASK), frequency shift keying (FSK), phase shift keying (PSK), and quadrature amplitude modulation (QAM). Each modulation method can be further divided, for example, quadrature amplitude modulation can be divided into 8QAM, 16QAM, 32QAM, 64QAM, 128QAM, 256QAM, etc.

[0003] In recent years, it has become a trend to combine deep learning technology with radio signal detection. However, the signal characteristics of different modulation methods vary greatly, and multiple signal features need to be combined to complete the modulation identification of the signal. How to combine the inherent characteristics of the modulated signal with deep learning to automatically identify the modulation method of the modulated signal has attracted the attention of more and more researchers. Summary of the invention

[0004] In order to automatically and accurately identify the modulation mode of a modulated signal, the present application provides a universal carrier modulation method and system.

[0005] In a first aspect, the present application provides a general carrier modulation method, which adopts the following technical solution:

[0006] A general carrier modulation method, comprising:

[0007] Acquiring a signal: generating a modulated signal with different modulation modes, recorded as a first signal, and acquiring a modulated signal to be processed;

[0008] Signal processing: including extraction of IQ data, time domain conversion, integration and normalization;

[0009] Extracting IQ data: extracting IQ data of the first signal, obtaining a signal-to-noise ratio of the IQ data, recorded as first data, and obtaining a modulation mode of the IQ data, recorded as a first modulation mode;

[0010] Time domain conversion: Perform feature conversion on the IQ data in the time domain to obtain the conversion result, and extract the instantaneous amplitude and instantaneous phase from the conversion result;

[0011] Integration: integrating the instantaneous amplitude and the instantaneous phase according to the first data and the first modulation method to obtain a second signal;

[0012] Normalization: normalizing the amplitude of the second signal to obtain a third signal;

[0013] First modeling: Based on a deep learning algorithm, a first prediction model is established, wherein the first prediction model includes: a ResNet sub-model, a fully connected layer, and a SoftMax classifier;

[0014] First training: training the first prediction model using the third signal and the modulation method of the third signal;

[0015] Prediction: The modulation signal to be processed is used as the input of the trained first prediction model, and the predicted modulation mode is output.

[0016] By adopting the above technical solution, the present application extracts IQ data and calculates the signal-to-noise ratio (i.e., the first data), which helps to distinguish the signal quality in subsequent processing, and the identification of the modulation method (i.e., the first modulation method) provides an important basis for subsequent integration. Converting the IQ data from the frequency domain to the time domain, extracting the instantaneous amplitude and instantaneous phase, and then integrating the instantaneous amplitude and instantaneous phase, and adjusting them according to the first data and the first modulation method can enhance the representativeness of the signal and reduce redundant information. Normalization processing helps to eliminate the dimensional differences between different signals and improve the generalization ability of deep learning models. ResNet (residual network) is used as a sub-model, combined with a fully connected layer and a SoftMax classifier, to construct a complex deep learning model, namely the first prediction model. The ResNet sub-model is good at processing complex data, and its residual learning mechanism helps to alleviate the problem of gradient disappearance or explosion in deep networks and improve the learning ability of the first prediction model. The first prediction model is trained using the processed third signal and its modulation method, so that the first prediction model can learn the signal characteristics under different modulation methods. The modulated signal to be processed is input into the trained first prediction model, so that its modulation method can be predicted more accurately.

[0017] Optionally, after performing the first modeling step and before performing the first training step, the method further includes:

[0018] Sequencing: arranging the third signal in time sequence to obtain a third signal sequence;

[0019] Model reconstruction: Add a self-attention layer to the first prediction model to obtain the second prediction model, and use the second prediction model as the new first prediction model.

[0020] By adopting the above technical solution, by maintaining the original order of the data, the first prediction model can more effectively capture the time dependency and long-term dependency in the sequence, which helps to improve the accuracy of the first prediction model in prediction, classification or generation tasks. The noise in the modulated signal will increase due to the disorder of the signal order. By serialization, the noise introduced by improper signal arrangement can be reduced. Serialization makes it easier for the first prediction model to find the modulation mode of the third signal during the training process, thereby accelerating the training process. The main purpose of adding a self-attention layer to the first prediction model is to enable the first prediction model to consider the relationship between each third signal in the third signal sequence and all other signals at the same time when processing sequence signals, rather than just adjacent elements. The self-attention layer can significantly improve the ability of the first prediction model to process long sequences while maintaining efficient processing of short sequences, so that the first prediction model performs better when processing tasks that require capturing long-distance dependencies. Combining the two steps of serialization and model reconstruction can significantly improve the performance of the first prediction model before executing the first training step. Serialization arranges the signals in chronological order, providing high-quality training data for the first prediction model; and the model reconstruction step enhances the ability of the first prediction model to process complex sequence data by adding a self-attention layer to the first prediction model. This combination enables the first prediction model to learn patterns and relationships in the data more effectively during training, thereby achieving higher accuracy and efficiency in subsequent tasks.

[0021] Optionally, after executing the serialization step and before executing the model reconstruction step, the method further includes:

[0022] Extract key signals: Use K-means clustering algorithm to extract the key third signal in the third signal sequence, which is recorded as the fourth signal.

[0023] This application actually reduces the redundant information in the third signal sequence by extracting the key signal (recorded as the fourth signal) in the third signal sequence, which helps the first prediction model to focus more on learning the features that are truly important for the prediction or classification task, thereby improving the efficiency and accuracy of the model. Key signals often contain the most important information in the data, which is crucial to the decision-making process of the model. By extracting these key signals, their representation in the data set can be enhanced, making it easier for the first prediction model to capture these important features.

[0024] Optionally, the self-attention layer includes a local self-attention part and a global self-attention part,

[0025] The local self-attention part is provided with a sliding window of a fixed size, and the third signal sequence is traversed by using the sliding window, and the attention score of the third signal and the remaining third signals in the sliding window is calculated in each sliding window, which is recorded as a first attention score;

[0026] The global self-attention part calculates the attention score of each fourth signal and the remaining third signals in the third signal sequence, which is recorded as the second attention score;

[0027] The first attention score and the second attention score are weighted averaged as the output of the self-attention layer.

[0028] By adopting the above technical solution, the present application sets a self-attention layer, which includes a local self-attention part for calculating the local attention score, and a global self-attention part for calculating the key third signal (i.e., the fourth signal) and the remaining third signal attention score, so as to realize the function of extracting the local features and global features of the third signal sequence. This process does not require adding additional layers in the second prediction model, and simplifies the model structure. The local self-attention part can capture the dependencies between the signals in each window in detail by setting a sliding window of a fixed size. Since only the signals in each window are focused on, the local self-attention part is more efficient than the global self-attention part in calculation, which helps to reduce the amount of calculation, especially when processing long sequences. The global self-attention part can capture the dependencies between any two signals in the entire sequence without being restricted by distance, and is very suitable for tasks that require understanding the context of the entire sequence. By combining the local self-attention part and the global self-attention part, the second prediction model can simultaneously utilize the local details and global context of the sequence. This combination makes the second prediction model more flexible in processing complex tasks. Since it can capture local and global information at the same time, the self-attention layer can improve the accuracy of the second prediction model in tasks such as prediction and classification.

[0029] Optionally, after executing the step of extracting key signals and before executing the step of reconstructing the model, the method further includes:

[0030] Extract features: extract features of the fourth signal, and take all features of the fourth signal as a sample;

[0031] First calculation: calculate the similarity between the i-th sample and the j-th sample, recorded as the first similarity;

[0032] First judgment: judging whether the first similarity is less than a preset similarity threshold, if so, executing the retaining step; if not, executing the deleting step;

[0033] Retention: retain the i-th sample and the j-th sample;

[0034] Deletion: In the third signal sequence, delete the jth sample;

[0035] Iteration: Take the j+1th sample as the new jth sample and execute the first calculation step until the preset stop iteration condition is reached;

[0036] Update signal: take the fourth signal corresponding to the retained sample as the new fourth signal.

[0037] The present application converts the original data into a form that is easier for the first prediction model to process by extracting the features of the fourth signal. Feature extraction helps to reduce the redundancy and noise of the data while retaining key information useful for model training. By calculating the similarity between samples and retaining or deleting them according to a preset similarity threshold, samples that are too similar and do not contribute much to model training can be removed, which helps to reduce the redundancy of the data and improve the generalization ability of the first prediction model. The iterative process can continuously screen samples until the preset stop iteration condition is met. This process helps to clean up the data more thoroughly and improve the quality of the final retained samples. By explicitly calculating the similarity between samples and screening samples based on the similarity, it is easier to understand which samples are useful for model training and which samples are redundant, and then delete redundant samples, reducing the amount of data that needs to be processed during model training, thereby reducing the computational cost. Due to the removal of redundant and similar samples, the first prediction model can focus more on learning those truly useful features during the training process, thereby improving the accuracy and generalization ability of the model.

[0038] Optionally, after executing the deleting step and before executing the iterating step, the method further includes:

[0039] Establishing a sample set: establishing a deleted sample set, in which all deleted samples are stored;

[0040] Second calculation: calculate the similarity between the jth sample and the samples in the deleted sample set, recorded as the second similarity;

[0041] Second judgment: judging whether the second similarity is less than a preset similarity threshold, if so, executing the adding step; if not, adding the jth sample to the deleted sample set;

[0042] Adding: obtaining a sample corresponding to the second similarity, recording it as a first sample, obtaining a fourth signal corresponding to the first sample, recording it as a fifth signal, and re-adding the fifth signal to the third signal sequence.

[0043] By adopting the above technical solution, the deleted sample set of the present application is used to store samples that are determined to need to be deleted. By calculating the similarity between the sample deleted this time and each sample in the deleted sample set, it is determined whether the similarity is lower than the preset similarity threshold. If so, it means that the current sample and the sample in the deleted sample set have large differences in features, and the samples in the sample set need to be re-added to the third signal sequence, thereby reducing the possibility of accidental deletion.

[0044] Optionally, after performing the first training step and before performing the prediction step, the method further includes:

[0045] Obtaining parameter combinations: in the first training step, various parameters of the first prediction model are obtained, and various parameter combinations of the first prediction model are obtained based on the various parameters, which are recorded as first parameter combinations;

[0046] Second modeling: establish Gaussian process model;

[0047] Second training: the Gaussian process model is trained using the first parameter combination;

[0048] Generate parameter combinations: Use the trained Gaussian process model to generate new candidate parameter combinations, which are recorded as the second parameter combinations;

[0049] Deployment: deploying the second parameter combination in the first prediction model, evaluating the first prediction model, and obtaining an evaluation result;

[0050] Update model: Update the Gaussian process model based on the evaluation results and execute the steps of generating parameter combinations until the preset stop update conditions are met.

[0051] By adopting the above technical solution, after the first training step, by obtaining various parameters of the first prediction model and generating parameter combinations (i.e., first parameter combinations), it is helpful to explore different configurations of the parameters of the first prediction model, so that better model performance may be found. Based on the knowledge it has learned, the Gaussian process model generates new candidate parameter combinations (second parameter combinations). These new parameter combinations may have not been explored before and may bring better model performance. By continuously generating new parameter combinations and evaluating their effects, the optimal solution can be gradually approached. The newly generated parameter combination is deployed to the first prediction model, and the performance of the first prediction model is evaluated to obtain evaluation results to verify the effectiveness of the new parameter combination. By comparing the model performance under different parameter combinations, it can be determined which parameter combinations are better. Through continuous learning and updating, the Gaussian process model can gradually approach the optimal parameter combination, thereby improving the performance of the first prediction model.

[0052] Optionally, after executing the step of generating the parameter combination and before executing the step of deploying, the method further includes:

[0053] Third modeling: establishing a Bayesian network model, wherein the nodes of the Bayesian network model are the first parameter combinations, and the directed edges of the Bayesian network model are the correlations between the first parameter combinations;

[0054] Reasoning: Take the second parameter combination as the input of the Bayesian network model and output the probability of the predicted parameter combination, which is recorded as the first probability;

[0055] Probability judgment: judge whether the first probability is greater than the preset probability threshold, if so, execute the deployment step; if not, execute the step of generating parameter combination.

[0056] The present application establishes a Bayesian network model, which uses the first parameter combination as a node and the correlation between them as a directed edge to construct a model that can express the complex relationship between parameters. The Bayesian network model uses prior knowledge (i.e., the correlation between parameters) to optimize the subsequent reasoning process and improve the accuracy of the prediction. The reasoning step uses the established Bayesian network model to reason about the new parameter combination (i.e., the second parameter combination) and output the probability (first probability) corresponding to the parameter combination under the Bayesian network model. This probability reflects the confidence that the first prediction model may reach the optimal state under the given parameter combination. By setting a preset probability threshold, the predicted probability of the new parameter combination is judged. If the predicted probability is greater than the threshold, it is considered that the parameter combination has sufficient potential for successful deployment, so the deployment step is executed; otherwise, it is considered that the parameter combination may not be the best choice and a new parameter combination needs to be regenerated.

[0057] Optionally, after the step of performing reasoning and before the step of performing probability judgment, the method further includes:

[0058] Adding a node: taking the predicted parameter combination as a new node of the Bayesian network model, and establishing a directed edge with the first parameter combination.

[0059] By adopting the above technical solution, the predicted parameter combination (i.e., the second parameter combination) is used as a new node of the Bayesian network model, which not only expands the coverage of the model, but also enables the Bayesian network model to conduct a more in-depth analysis and evaluation of this new parameter combination. The operation of adding nodes immediately after the reasoning step is actually a further utilization and expansion of the reasoning results. The reasoning step outputs the probability of the predicted parameter combination (i.e., the first probability), and adding nodes connects the first probability with other parameter combinations in the model, forming a more complete and rich probability graph.

[0060] In a second aspect, the present application provides a universal carrier modulation system, which adopts the following technical solution:

[0061] A general carrier modulation system comprises the following steps:

[0062] A signal acquisition module is used to generate a modulation signal with different modulation modes, which is recorded as a first signal; and is also used to acquire a modulation signal to be processed;

[0063] A signal processing module, which is in communication with the signal acquisition module, comprises an IQ data extraction unit, a time domain conversion unit, an integration unit and a normalization unit;

[0064] An IQ data extraction unit is used to extract IQ data of the first signal, obtain a signal-to-noise ratio of the IQ data, recorded as the first data, and obtain a modulation mode of the IQ data, recorded as the first modulation mode;

[0065] A time domain conversion unit is used to perform feature conversion on the IQ data in the time domain to obtain a conversion result, and extract the instantaneous amplitude and instantaneous phase from the conversion result;

[0066] An integration unit, configured to integrate the instantaneous amplitude and the instantaneous phase according to the first data and the first modulation method to obtain a second signal;

[0067] a normalization unit, configured to perform normalization processing on the amplitude of the second signal to obtain a third signal;

[0068] A first modeling module, used to establish a first prediction model based on a deep learning algorithm, wherein the first prediction model includes: a ResNet sub-model, a fully connected layer, and a SoftMax classifier;

[0069] A first training module, which is in communication with the signal processing module and the first modeling module, and is used to train the first prediction model using the third signal and the modulation method of the third signal;

[0070] The prediction module is connected to the signal acquisition module and the first training module for taking the modulation signal to be processed as the input of the trained first prediction model and outputting the predicted modulation mode.

[0071] This application generates a variety of modulated signals by acquiring a signal module, and uses a signal processing module to perform fine IQ data extraction, time domain conversion, integration and normalization processing on the signal, effectively extracting the key features of the signal. The first modeling module builds a first prediction model based on a powerful deep learning algorithm (especially the ResNet sub-model), which performs feature mapping through a fully connected layer, and outputs the probability distribution of the modulation mode by a SoftMax classifier. The first training module fully trains the first prediction model using the processed signal and its modulation mode, thereby improving the accuracy and generalization ability of the first prediction model. The prediction module can efficiently and accurately predict the modulation mode of the modulated signal to be processed.

[0072] In summary, the present application includes at least one of the following beneficial technical effects:

[0073] 1. Extracting IQ data and calculating the signal-to-noise ratio helps distinguish signal quality in subsequent processing, while the identification of the modulation method (i.e., the first modulation method) provides an important basis for subsequent integration. Converting IQ data from the frequency domain to the time domain, extracting the instantaneous amplitude and instantaneous phase, and then integrating the instantaneous amplitude and instantaneous phase, and adjusting them according to the first data and the first modulation method can enhance the representativeness of the signal and reduce redundant information. Normalization processing helps to eliminate the dimensional differences between different signals and improve the generalization ability of the deep learning model.

[0074] 2. This application uses ResNet (residual network) as a sub-model, combined with a fully connected layer and a SoftMax classifier, to construct a complex deep learning model, namely the first prediction model. The ResNet sub-model is good at processing complex data, and its residual learning mechanism helps to alleviate the problem of gradient vanishing or exploding in deep networks and improve the learning ability of the first prediction model. The first prediction model is trained using the processed third signal and its modulation method, so that the first prediction model can learn the signal characteristics under different modulation methods. The modulated signal to be processed is input into the trained first prediction model, and the predicted modulation method is output. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 is a flow chart of Example 1 of the present application;

[0076] Figure 2 is a flow chart of Example 2 of the present application;

[0077] Figure 3 is a flow chart of Example 3 of the present application;

[0078] Figure 4 is a flow chart of Example 4 of the present application;

[0079] Figure 5 It is a circuit topology diagram of the hardware circuit part of the carrier modulation and demodulation device of this application. DETAILED DESCRIPTION

[0080] The following combination Figures 1 to 5 This application is described in further detail.

[0081] Embodiment 1: This embodiment discloses a general carrier modulation method, referring to Figure 1The modulation method includes: S1 acquiring a signal, S2 signal processing, S3 first modeling, S4 first training and S5 prediction. First, a modulation signal (i.e., a first signal) with different modulation modes is generated, then a signal to be modulated is acquired, then the first signal is processed, then a first prediction model is established, then the processed first signal and the modulation mode of the first signal are used to train the first prediction model, then the trained first prediction model is used to analyze the modulation signal to be processed, and the modulation mode of the modulation signal to be processed is obtained. This embodiment includes the following steps:

[0082] S1 obtains the signal: Use MATLAB to generate a series of modulated signals with different modulation modes (such as BPSK, QPSK, QAM, etc.), and mark these signals as the first signal.

[0083] First, set the parameters of the modulation signal to be generated, signal length N = 1024; sampling frequency fs = 1000; time vector t = (0:N-1) / fs. Then, based on the set parameters, generate modulation signals of different modulation modes, that is, the first signal. The parameters set in this step can be set according to requirements.

[0084] This step also captures the modulation signal to be processed, and the modulation mode of the modulation signal to be processed is unknown.

[0085] S2 signal processing, including S21 extracting IQ data, S22 time domain conversion, S23 integration and S24 normalization. The purpose of this step is to extract key features from the first signal and perform preprocessing.

[0086] S21 extracts IQ data: extracts IQ data (in-phase and quadrature components) of each first signal. First, each first signal needs to be sampled and quantized to convert it into a digital signal. The sampling frequency uses the same sampling frequency fs = 1000 as the signal acquired in S1, and then the sampled value is mapped to the corresponding digital level.

[0087] In quadrature modulation, the signal already exists in the form of IQ components or can be converted into this form.

[0088] For non-orthogonal modulation such as BPSK, although it does not directly contain orthogonal components, it can be mapped to the IQ plane. In practical applications, BPSK is usually regarded as a special QPSK (only using I component or Q component), so when extracting IQ data, only one of the components needs to be considered. However, in some analysis or processing, in order to be consistent with other orthogonal modulation methods, the BPSK signal can also be regarded as a QPSK signal with zero Q component.

[0089] This embodiment uses MATLAB software to extract IQ data of each first signal.

[0090] This step also obtains the signal-to-noise ratio of each IQ data, recorded as the first data, and is calculated as follows:

[0091] ;

[0092] in, It represents the signal-to-noise ratio in decibels; A value representing the signal power of the first signal corresponding to the current first data; Indicates the value of the noise power of the first signal corresponding to the current first data.

[0093] At the same time, this step also obtains the modulation mode of the IQ data (that is, the modulation mode of the first signal is used as the modulation mode of the IQ data), and records the modulation mode of the IQ data as the first modulation mode.

[0094] S22 Time domain conversion: This step uses short-time inverse Fourier transform to divide the IQ data into multiple short time periods, performs feature conversion in the time domain on the IQ data of each time period, and obtains the instantaneous amplitude and instantaneous phase based on the conversion results.

[0095] S23 integration: using MATLAB to integrate the instantaneous amplitude and the instantaneous phase according to the first data and the first modulation mode (ie, the signal-to-noise ratio and the modulation mode), to obtain a plurality of second signals.

[0096] S24 normalization: using a minimum-maximum normalization algorithm or a Z-score normalization algorithm to normalize the amplitude of the second signal to obtain a third signal (the normalized amplitude is used as the amplitude of the third signal, and the modulation mode of the second signal is used as the modulation mode of the third signal).

[0097] The amplitude of the i-th second signal is normalized using the minimum-maximum normalization algorithm as follows:

[0098] ;

[0099] in, is the amplitude of the third signal; is the amplitude of the i-th second signal; is the maximum amplitude of all second signals; is the minimum amplitude of all second signals.

[0100] The Z-score normalization algorithm is used to normalize the amplitude of the i-th second signal as follows:

[0101] ;

[0102] is the amplitude of the third signal; is the amplitude of the i-th second signal; is the average amplitude of all second signals; is the standard deviation of the amplitudes of all second signals.

[0103] S3 First Modeling: Based on the deep learning algorithm, the first prediction model is established. The first prediction model includes: ResNet sub-model, fully connected layer and SoftMax classifier. The ResNet sub-model is responsible for extracting deep features from the input signal, the fully connected layer is used to map the features to the output space of the modulation mode, and the SoftMax classifier provides the final modulation mode prediction result.

[0104] This step uses MATLAB's Deep Learning Toolbox or Machine Learning Toolbox to build the ResNet sub-model, fully connected layers, and SoftMax classifier.

[0105] S4 first training: using the third signal and the modulation mode of the third signal as training data to train the first prediction model.

[0106] S5 prediction: the modulation signal to be processed is used as the input of the trained first prediction model, and the predicted modulation mode is output.

[0107] This embodiment extracts IQ data, performs time domain feature conversion (such as instantaneous amplitude and phase), integration and normalization, and establishes and trains a first prediction model based on a deep learning algorithm (including a ResNet sub-model, a fully connected layer, and a SoftMax classifier), ultimately achieving automatic recognition of the modulation mode of an unknown modulated signal.

[0108] Example 2: Reference Figure 2 The difference between this embodiment and embodiment 1 is that after executing S3 first modeling and before executing S4 first training, it also includes:

[0109] S31 serialization: arranging the third signal in time sequence to obtain a third signal sequence, so that the timing characteristics of the third signal are preserved.

[0110] S32 extracting a key signal: extracting a key third signal in the third signal sequence, and recording it as a fourth signal.

[0111] The scheme for extracting key signals based on the K-means clustering algorithm is as follows:

[0112] Feature extraction: Extract useful signal features from the third signal. These signal features can reflect the key properties of the third signal, such as amplitude, frequency, phase, energy, etc.

[0113] The signal features are clustered using a K-means clustering algorithm.

[0114] Feature distribution analysis: Observe the distribution of signal features in each cluster, including statistics such as mean, median, and standard deviation. These statistics can help understand the overall characteristics of the signals within the cluster.

[0115] Feature correlation analysis: Analyzing the correlation between signal features within a cluster using tools such as correlation coefficient matrices or scatter plots can help identify which features play a key role in distinguishing different clusters.

[0116] A signal feature whose correlation is greater than a preset threshold is obtained, and a third signal corresponding to the signal feature is used as a key third signal.

[0117] S33 Model reconstruction: Add a self-attention layer to the first prediction model to obtain a second prediction model, and use the second prediction model as the new first prediction model.

[0118] The self-attention layer includes a local self-attention part and a global self-attention part:

[0119] The local self-attention part is provided with a sliding window of a fixed size for limiting the scope of attention calculation. The window slides on the third signal sequence, and only pays attention to the third signal in the window each time. By using the sliding window to traverse the third signal sequence, the attention score of the third signal and the remaining third signals in the sliding window is calculated in each sliding window, which is recorded as the first attention score. The first attention score reflects the correlation or importance between the current third signal and the remaining third signals.

[0120] The global self-attention part calculates the attention score of each fourth signal and the remaining third signals in the third signal sequence, which is recorded as the second attention score.

[0121] The first attention score and the second attention score are weighted averaged as the output of the self-attention layer. The calculation model is as follows:

[0122] ;

[0123] Among them, A is the output of the self-attention layer; B is the first attention score; C is the second attention score; is the weight of the first attention score; is the weight of the second attention score.

[0124] The structure of the second prediction model is:

[0125] The self-attention layer is used to generate a weight for each third signal in the third signal sequence.

[0126] The modulation mode of the original third signal is used as the true label of the new third signal. The ResNet sub-model performs deep feature extraction on the new third signal and the true label corresponding to the third signal through its convolutional layer, residual connection and other structures. These features can capture the complex patterns and structures in the signal and provide strong support for subsequent classification tasks. The ResNet sub-model outputs the feature representation of the new third signal, which is recorded as the first feature.

[0127] The fully connected layer maps the first feature to the modulation mode to obtain the predicted value of the modulation mode of the first feature. The function of the fully connected layer is to convert the extracted first feature into a vector corresponding to the number of modulation mode categories, and each element of the vector represents a predicted value of a modulation mode.

[0128] The SoftMax classifier converts the predicted value into a probability distribution. The function of the SoftMax classifier is to convert a real number vector into a probability distribution vector, in which the value of each element is between 0 and 1, and the sum of all elements is 1. After being processed by the SoftMax classifier, each modulation mode corresponds to a probability value, indicating the possibility that the input signal belongs to this modulation mode.

[0129] This embodiment adds a second prediction model including a local self-attention part and a global self-attention part to the original first prediction model to more accurately capture the correlation between signals. The self-attention layer calculates the local and global attention scores and outputs them in a weighted average manner to enhance the second prediction model's ability to understand the signal sequence. The reconstructed model extracts features through the ResNet sub-model, and the fully connected layer maps the features to the modulation mode prediction results, which are converted to probability distributions by the SoftMax classifier to support more accurate modulation mode predictions.

[0130] Example 3: Reference Figure 3 The difference between this embodiment and embodiment 2 is that after executing S32 to extract the key signal and before executing S33 to reconstruct the model, it also includes:

[0131] S601 extracts features: extracts features of the fourth signal, and takes all features of the fourth signal as a sample.

[0132] S602 First calculation: Calculate the similarity between the i-th sample and the j-th sample using cosine similarity, which is recorded as the first similarity.

[0133] Set the coordinates of the i-th sample in the feature space to , the coordinates of the jth sample in the feature space are , n is the number of features, and the calculation formula of the cosine similarity D of two samples is as follows:

[0134] .

[0135] S603 first judgment: judge whether the first similarity is less than a preset similarity threshold. If so, it means that the difference between the two samples is large and may represent different information, and execute S604 to retain; if not, it indicates that there is redundancy between the two samples, and execute S605 to delete.

[0136] S604 Retain: retain the i-th sample and the j-th sample.

[0137] S605 Delete: In the third signal sequence, delete the jth sample.

[0138] S606: Create a sample set: create a deleted sample set, in which all deleted samples are stored.

[0139] S607 Second calculation: Calculate the similarity between the jth sample and the samples in the deleted sample set, recorded as the second similarity, to check whether the deleted sample (jth sample) has a similarity value with a sample in the deleted sample set that is also less than a preset similarity threshold, thereby reducing the possibility of mistaken deletion.

[0140] S608 second judgment: determine whether the second similarity is less than the preset similarity threshold. If so, it means that the deleted sample (the jth sample) is not highly similar to the samples in the deleted sample set and may contain additional information or patterns. Therefore, it is necessary to consider restoring the samples in the deleted sample set and execute S609 to add; if not, add the jth sample to the deleted sample set.

[0141] S609 adding: obtaining a sample corresponding to the second similarity, recording it as a first sample, obtaining a fourth signal corresponding to the first sample, recording it as a fifth signal, and re-adding the fifth signal to the third signal sequence.

[0142] S610 iteration: take the j+1th sample as the new jth sample, and execute the first calculation of S602 to S609 addition, until the preset stop iteration condition is reached, and the stop iteration condition can be that the number of executions reaches a preset execution number threshold, or the number of samples deleted in the sample set reaches a preset sample upper limit.

[0143] S611 updates the signal: uses the fourth signal corresponding to the sample retained in S604 as a new fourth signal.

[0144] This embodiment optimizes the fourth signal through feature extraction, similarity calculation and sample screening to improve data quality and reduce redundancy, thereby providing better input for subsequent model reconstruction or analysis tasks.

[0145] Example 4: Reference Figure 4 The difference between this embodiment and embodiment 1 is that after executing S4 first training and before executing S5 prediction, it also includes:

[0146] S701 Obtain parameter combination: In S4 first training, various parameters of the first prediction model are obtained, and the various parameters include learning rate, regularization coefficient, number of network layers, number of nodes in each layer, etc.

[0147] Based on the different values ​​of the various parameters, a plurality of parameter combinations, namely first parameter combinations, are generated. These first parameter combinations represent different configurations of the first prediction model.

[0148] S702 Second modeling: Establishing a Gaussian process model. The Gaussian process model models the function space based on Gaussian distribution, can capture the correlation between various first parameter combinations, and predict new first parameter combinations.

[0149] S703 second training: the first parameter combination in the parameter combination obtained in S701 and the prediction accuracy of the first prediction model are used to train the Gaussian process model, so that the Gaussian process model can learn the relationship between the parameter combination and the model performance.

[0150] S704 Generate parameter combination: Use the Gaussian process model trained in the second training in S703 to generate a new candidate parameter combination, which is recorded as the second parameter combination. The second parameter combination is a parameter configuration found based on the exploration of the parameter space by the Gaussian process model.

[0151] S705 The third modeling: Establishing a Bayesian network model, wherein the nodes of the Bayesian network model are the first parameter combinations, the directed edges of the Bayesian network model are the correlations between the first parameter combinations, and the Bayesian network model is used to infer the mutual influence between different parameter combinations and predict the performance of new parameter combinations.

[0152] S706 Reasoning: The second parameter combination is used as the input of the Bayesian network model, and the probability of the predicted parameter combination is output, which is recorded as the first probability. The first probability reflects the possibility that the performance of the first prediction model reaches the expected level under the given second parameter combination.

[0153] S707 Adding a node: Using the predicted parameter combination as a new node of the Bayesian network model, and establishing a directed edge with the first parameter combination.

[0154] S708 probability judgment: judge whether the first probability is greater than a preset probability threshold, if so, execute S709 deployment; if not, execute S704 to generate a parameter combination.

[0155] S709 Deployment: deploy the second parameter combination in the first prediction model, and use the test set to evaluate the first prediction model, and obtain the evaluation result to understand the actual effect of the new parameter combination.

[0156] S710 updates the model: based on the evaluation results, the parameters of the Gaussian process model are updated, and S704 is executed to generate parameter combinations until a preset stop updating condition is met (such as reaching a certain number of iterations, the performance of the first prediction model reaches expectations, etc.).

[0157] In this embodiment, various parameter combinations of the first prediction model are obtained through the first training step, and the Gaussian process model is trained using these combinations to predict better parameters. A Bayesian network model is then established to infer the performance probability of the newly generated candidate parameter combination (second parameter combination). If the probability is higher than the preset threshold, the second parameter combination is deployed and evaluated; otherwise, new combinations are generated. Based on the evaluation results, the parameters of the Gaussian process model are continuously updated, and this process is repeated until the stop update condition is met to find the optimal configuration of the first prediction model and improve the prediction performance.

[0158] Embodiment 5: This embodiment discloses a universal carrier modulation system, the system comprising:

[0159] The signal acquisition module is used to generate a modulated signal with different modulation modes, which is recorded as the first signal. This module is also used to acquire the modulated signal to be processed.

[0160] The signal processing module is connected to the signal acquisition module in communication, and includes an IQ data extraction unit, a time domain conversion unit, an integration unit and a normalization unit.

[0161] The IQ data extraction unit is responsible for extracting IQ data (In-phase and Quadrature components) from the first signal and obtaining the signal-to-noise ratio (first data) and modulation mode (first modulation mode) of these data.

[0162] The time domain conversion unit is used to perform feature conversion on the IQ data in the time domain and obtain the instantaneous amplitude and instantaneous phase based on the conversion result.

[0163] The integration unit is used to integrate the instantaneous amplitude and the instantaneous phase according to the first data and the first modulation method to obtain a second signal.

[0164] The normalization unit is used to normalize the amplitude of the second signal (to eliminate the dimensional difference between different features and improve the training efficiency and prediction accuracy of the model) to obtain the third signal.

[0165] The first modeling module is used to establish a first prediction model using a deep learning algorithm, wherein the first prediction model includes: a ResNet sub-model, a fully connected layer and a SoftMax classifier.

[0166] The first training module is communicatively connected with the signal processing module and the first modeling module, and is used to train the first prediction model using the third signal and the modulation mode of the third signal. The training process includes forward propagation, loss calculation, back propagation and parameter update, etc., and the first prediction model can accurately predict the modulation mode of the input signal by continuously optimizing the parameters of the model.

[0167] The prediction module is connected to the signal acquisition module and the first training module for taking the modulation signal to be processed as the input of the trained first prediction model and outputting the predicted modulation mode.

[0168] The system generates a modulated signal and processes the signal to be tested by acquiring a signal module, extracts IQ data, performs time domain conversion, integration and normalization using the signal processing module, and then trains the system through the first prediction model based on ResNet, fully connected layer and SoftMax classifier. Finally, the prediction module outputs the predicted modulation mode of the signal to be processed, thereby realizing automatic recognition of the signal modulation mode.

[0169] The carrier modulation system is deployed in a carrier modulation and demodulation device, referring to Figure 5 The carrier modulation and demodulation device also includes a hardware circuit part, which includes: a strong current interface terminal, a zero-crossing detection circuit, a strong current carrier signal processing circuit, a signal transformer, an optical coupler, a weak current interface terminal (connected to a protocol conversion backplane), a main control system circuit, a weak current carrier signal processing circuit and an LED indication circuit. The connection method of each circuit part is as follows:

[0170] One end of the zero-crossing detection circuit is connected to one end of the strong-current interface terminal, and is used to process the signal transmitted from the strong-current interface terminal. The other end of the zero-crossing detection circuit is connected to one end of the optical coupler.

[0171] One end of the strong-electric carrier signal processing circuit is connected to the other end of the strong-electric interface terminal, and the other end of the strong-electric carrier signal processing circuit is connected to one end of the signal transformer.

[0172] The main control system circuit is connected with the other end of the optical coupler, one end of the weak-current carrier signal processing circuit, the LED indication circuit and the weak-current interface terminal.

[0173] The other end of the weak-current carrier signal processing circuit is connected to the other end of the signal transformer.

[0174] Its working principle is: the zero-crossing detection circuit collects the signal from the high-voltage interface terminal, and transmits the processed signal to the main control system circuit through the optical coupler to realize the zero-crossing detection of the AC line; after the carrier output signal enters the high-voltage interface terminal, it passes through the high-voltage carrier signal processing circuit and the signal transformer to be converted into a low-voltage signal.

[0175] The weak-current carrier signal processing circuit receives the weak-current signal, which enters the main control system circuit after processing, and obtains the output signal of the main control system circuit. The output signal is amplified by the weak-current carrier signal processing circuit, and finally output to the strong-current interface terminal through the signal transformer and the strong-current carrier signal processing circuit.

[0176] The LED indication circuit receives the control signal of the main control system circuit to realize LED display; the main control system circuit performs two-way information exchange with the external meter board through the weak current interface terminal.

[0177] The carrier modulation and demodulation device realizes HPLC (High-speed Power Line Carrier) communication interaction through an embedded program. The signal modulation part of the power carrier communication modulates the data signal to be transmitted onto the high-frequency carrier signal through the carrier modulation method, thereby obtaining the effect of converting the low-frequency signal into a high-frequency signal. After that, the modulated high-frequency signal is coupled with the power line through a coupler to realize signal transmission. The function of the coupler is to couple the signal to the power line and reduce the interference to the power line system. After that, at the receiving end, the high-frequency carrier signal is demodulated and restored to the original data signal through a demodulator. The demodulation process is opposite to the modulation process, and the high-frequency signal is restored to a low-frequency signal. Finally, the demodulated data signal can be transmitted and applied through various interface devices, such as computer networks, telephones, televisions, etc.

[0178] The above are all preferred embodiments of the present application, and the protection scope of the present application is not limited thereto. Therefore, any equivalent changes made according to the structure, shape, and principle of the present application should be included in the protection scope of the present application.

Claims

1. A general carrier modulation method, characterized in that: include: Acquire signal: generate a modulated signal with different modulation modes, recorded as a first signal; Obtaining a modulated signal to be processed; Signal processing: including extraction of IQ data, time domain conversion, integration and normalization; Extracting IQ data: extracting IQ data of the first signal, obtaining a signal-to-noise ratio of the IQ data, recorded as first data, and obtaining a modulation mode of the IQ data, recorded as a first modulation mode; Time domain conversion: Perform feature conversion on the IQ data in the time domain to obtain the conversion result, and extract the instantaneous amplitude and instantaneous phase from the conversion result; Integration: integrating the instantaneous amplitude and the instantaneous phase according to the first data and the first modulation method to obtain a second signal; Normalization: normalizing the amplitude of the second signal to obtain a third signal; First modeling: Based on a deep learning algorithm, a first prediction model is established, wherein the first prediction model includes: a ResNet sub-model, a fully connected layer, and a SoftMax classifier; Sequencing: arranging the third signal in time sequence to obtain a third signal sequence; Extracting key signals: Using K-means clustering algorithm to extract the key third signal in the third signal sequence, recorded as the fourth signal; Model reconstruction: add a self-attention layer to the first prediction model to obtain a second prediction model, and use the second prediction model as the new first prediction model; The self-attention layer includes a local self-attention part and a global self-attention part. The local self-attention part is provided with a sliding window of a fixed size, and the third signal sequence is traversed by using the sliding window, and the attention score of the third signal and the remaining third signals in the sliding window is calculated in each sliding window, which is recorded as a first attention score; The global self-attention part calculates the attention score of each fourth signal and the remaining third signals in the third signal sequence, which is recorded as the second attention score; The first attention score and the second attention score are weighted averaged as the output of the self-attention layer; First training: training the first prediction model using the third signal and the modulation method of the third signal; Prediction: The modulation signal to be processed is used as the input of the trained first prediction model, and the predicted modulation mode is output.

2. The universal carrier modulation method according to claim 1, characterized in that: After executing the step of extracting key signals and before executing the step of model reconstruction, the method further includes: Extract features: extract features of the fourth signal, and take all features of the fourth signal as a sample; First calculation: calculate the similarity between the i-th sample and the j-th sample, recorded as the first similarity; First judgment: judging whether the first similarity is less than a preset similarity threshold, if so, executing the retaining step; if not, executing the deleting step; Retention: retain the i-th sample and the j-th sample; Deletion: In the third signal sequence, delete the jth sample; Iteration: Take the j+1th sample as the new jth sample and execute the first calculation step until the preset stop iteration condition is reached; Update signal: take the fourth signal corresponding to the retained sample as the new fourth signal.

3. The universal carrier modulation method according to claim 2, characterized in that: After the deletion step and before the iteration step, it also includes: Establishing a sample set: establishing a deleted sample set, in which all deleted samples are stored; Second calculation: calculate the similarity between the jth sample and the samples in the deleted sample set, recorded as the second similarity; Second judgment: judging whether the second similarity is less than a preset similarity threshold, if so, executing the adding step; if not, adding the jth sample to the deleted sample set; Adding: obtaining a sample corresponding to the second similarity, recording it as a first sample, obtaining a fourth signal corresponding to the first sample, recording it as a fifth signal, and re-adding the fifth signal to the third signal sequence.

4. The universal carrier modulation method according to claim 1, characterized in that: After performing the first training step and before performing the prediction step, the method further includes: Obtaining parameter combinations: in the first training step, various parameters of the first prediction model are obtained, and various parameter combinations of the first prediction model are obtained based on the various parameters, which are recorded as first parameter combinations; Second modeling: establish Gaussian process model; Second training: using the first parameter combination to train the Gaussian process model; Generate parameter combinations: Use the trained Gaussian process model to generate new candidate parameter combinations, which are recorded as the second parameter combinations; Deployment: deploying the second parameter combination in the first prediction model, evaluating the first prediction model, and obtaining an evaluation result; Update model: Update the Gaussian process model based on the evaluation results and execute the steps of generating parameter combinations until the preset stop update conditions are met.

5. The universal carrier modulation method according to claim 4, characterized in that: After executing the step of generating parameter combinations and before executing the deployment step, it also includes: Third modeling: establishing a Bayesian network model, wherein the nodes of the Bayesian network model are the first parameter combinations, and the directed edges of the Bayesian network model are the correlations between the first parameter combinations; Reasoning: Take the second parameter combination as the input of the Bayesian network model and output the probability of the predicted parameter combination, which is recorded as the first probability; Probability judgment: judge whether the first probability is greater than the preset probability threshold, if so, execute the deployment step; if not, execute the step of generating parameter combination.

6. The universal carrier modulation method according to claim 5, characterized in that: After the step of executing reasoning and before the step of executing probability judgment, it also includes: Adding a node: taking the predicted parameter combination as a new node of the Bayesian network model, and establishing a directed edge with the first parameter combination.

7. A universal carrier modulation system, characterized in that: The system is applicable to the method according to any one of claims 1 to 6, and the system further comprises: A signal acquisition module is used to generate a modulation signal with different modulation modes, which is recorded as a first signal; and is also used to acquire a modulation signal to be processed; A signal processing module, which is in communication with the signal acquisition module, comprises an IQ data extraction unit, a time domain conversion unit, an integration unit and a normalization unit; An IQ data extraction unit is used to extract IQ data of the first signal, obtain a signal-to-noise ratio of the IQ data, recorded as the first data, and obtain a modulation mode of the IQ data, recorded as the first modulation mode; A time domain conversion unit is used to perform feature conversion on the IQ data in the time domain to obtain a conversion result, and extract the instantaneous amplitude and instantaneous phase from the conversion result; An integration unit, configured to integrate the instantaneous amplitude and the instantaneous phase according to the first data and the first modulation method to obtain a second signal; a normalization unit, configured to perform normalization processing on the amplitude of the second signal to obtain a third signal; A first modeling module, used to establish a first prediction model based on a deep learning algorithm, wherein the first prediction model includes: a ResNet sub-model, a fully connected layer, and a SoftMax classifier; A first training module, which is in communication with the signal processing module and the first modeling module, and is used to train the first prediction model using the third signal and the modulation method of the third signal; The prediction module is connected to the signal acquisition module and the first training module for taking the modulation signal to be processed as the input of the trained first prediction model and outputting the predicted modulation mode.

Images