CN117056680A - Data noise reduction and signal detection method, device and system and storage medium - Google Patents

Abstract

The application discloses a method, a device, a system and a storage medium for data noise reduction and signal detection, which relate to the technical field of data and signal processing and comprise the following steps: generating analog data with low signal-to-noise ratio; preprocessing the mixed signal in the analog data to obtain training data; fusing the first convolutional neural network, the second convolutional neural network, the self-attention mechanism and the multi-layer fully-connected neural network to obtain a fused deep neural network; step-by-step training is carried out on the fusion type deep neural network through training data, and a target model with a noise reduction function is obtained; and inputting the real data with low signal-to-noise ratio into a target model so that the target model can perform noise reduction and signal detection on the real data to obtain a processing result of the noise reduction and the signal detection. The application can accurately reduce noise and detect the low signal-to-noise ratio data.

Description

Data noise reduction and signal detection methods, devices, systems and storage media

Technical field

The present application relates to the field of data and signal processing technologies, and in particular to a data noise reduction and signal detection method, device, system and storage medium.

Background technique

In some application scenarios, such as deep space communications, gravitational wave detection, and seismic exploration, we often face the problem of low signal-to-noise ratio data processing. In the received signal, since the noise intensity is much higher than the signal corresponding to the valid information, effective methods are needed to denoise the data. General data signal processing methods often use matched filtering and nonlinear filters. When using this type of method to process low signal-to-noise ratio data, there is a problem of poor noise reduction effect. It is difficult to accurately denoise the data and eliminate potential noise. signals are detected. Therefore, how to process low signal-to-noise ratio data has become an urgent technical problem that needs to be solved.

Contents of the invention

This application aims to solve at least one of the technical problems existing in the prior art. To this end, this application proposes a data denoising and signal detection method, device, system and storage medium, which can accurately denoise and detect low signal-to-noise ratio data.

The data noise reduction and signal detection method according to the first embodiment of the present application includes:

Generating simulation data with a low signal-to-noise ratio, wherein the simulation data includes a plurality of mixed signals, the mixed signals are obtained by mixing noise and pure signals, and the pure signals represent noise-free signals;

Preprocess the mixed signals in the simulation data to obtain training data;

The first convolutional neural network, the second convolutional neural network, the self-attention mechanism and the multi-layer fully connected neural network are fused to obtain a fused deep neural network;

Perform step-by-step training on the fused deep neural network using the training data to obtain a target model with noise reduction function;

Real data with a low signal-to-noise ratio is input into the target model, so that the target model performs noise reduction and signal detection on the real data, and obtains processing results of noise reduction and signal detection.

The data noise reduction and signal detection method according to the embodiment of the present application has at least the following beneficial effects: first, generate simulation data with a low signal-to-noise ratio; second, preprocess the mixed signals in the simulation data to obtain training data; then, The first convolutional neural network, the second convolutional neural network, the self-attention mechanism and the multi-layer fully connected neural network are fused to obtain a fused deep neural network; then, the fused deep neural network is step-by-step through the training data After training, a target model with noise reduction function is obtained; finally, the real data with low signal-to-noise ratio is input into the target model, so that the target model can perform noise reduction and signal detection on the real data, and obtain the processing results of noise reduction and signal detection. The data noise reduction and signal detection method of this application, on the one hand, through sufficient training based on simulated data, can enable the target model to improve noise reduction and detection accuracy when processing real data with low signal-to-noise ratio, and does not require frequent calculation, improving the speed of noise reduction and detection; on the other hand, by constructing a fused deep neural network, it can process ultra-long data and extract signal features of real data with low signal-to-noise ratio from different dimensions, thereby achieving Higher noise reduction and detection accuracy. Therefore, the data denoising and signal detection method of this application can accurately denoise and detect low signal-to-noise ratio data.

According to some embodiments of the present application, generating simulation data with low signal-to-noise ratio includes:

Simulate the generation of noise based on the sensitivity curve of the gravitational wave detector;

Select corresponding parameter ranges for different types of gravitational wave sources;

According to the parameter range, generate multiple pure signals corresponding to multiple gravitational wave sources, where the gravitational wave sources and the pure signals correspond one to one;

Each of the pure signals is mixed with the noise respectively to obtain analog data including a plurality of mixed signals.

According to some embodiments of the present application, preprocessing the mixed signals in the analog data includes:

Perform whitening processing on the mixed signal in the simulated data so that the noise spectrum of the mixed signal is evenly distributed;

The whitened mixed signal is normalized so that all data in the simulated data have the same scale.

According to some embodiments of the present application, the first convolutional neural network, the second convolutional neural network, the self-attention mechanism and the multi-layer fully connected neural network are fused to obtain a fused deep neural network, including:

Build the first convolutional neural network to extract feature information from the mixed signal;

Build a fusion network of the self-attention mechanism and the second convolutional neural network to perform noise reduction processing on the mixed signal;

The multi-layer fully connected neural network is built to detect the target signal.

According to some embodiments of the present application, the step-by-step training of the fused deep neural network using the training data to obtain a target model with noise reduction function includes:

The mean square error between the mixed signal and the pure signal is used as a loss function, and the fused deep neural network is initially trained through the loss function;

Set the number of alternating cycles, wherein the number of alternating cycles includes a first training cycle number and a second training cycle number, and both the first training cycle number and the second training cycle number include multiple training cycles, each The training cycle represents the process of training all samples in the training data once;

Perform noise reduction training on the fused deep neural network within the first number of training cycles, and perform detection training on the fused deep neural network within the second number of training cycles;

Based on the number of alternation cycles, noise reduction training and detection training are alternately performed;

An optimizer is used to adjust the parameters of the fused deep neural network, and a learning rate adjustment strategy is used to adjust the learning rate of the fused deep neural network.

According to some embodiments of the present application, the fused deep neural network is trained for noise reduction within the first number of training cycles, and the fused deep neural network is trained for the second number of training cycles. Detection training, including:

Perform noise reduction training on the fused deep neural network within the first number of training cycles to minimize the mean square error between the mixed signal and the pure signal;

Perform detection training on the fused deep neural network within the second number of training cycles to minimize the cross-entropy loss of binary classification, wherein the number of cycles of the first number of training cycles is equal to the number of second training cycles number of cycles.

According to some embodiments of the present application, the data noise reduction and signal detection method further includes:

The processing results are evaluated to determine the noise reduction accuracy and detection accuracy of the target model.

The data noise reduction and signal detection device according to the second embodiment of the present application includes:

A generation module, configured to generate analog data with a low signal-to-noise ratio, wherein the analog data includes a plurality of mixed signals, the mixed signals are obtained by mixing noise and pure signals, and the pure signals represent noise-free signals;

A processing module, used to preprocess the mixed signal in the simulation data to obtain training data;

The fusion module is used to fuse the first convolutional neural network, the second convolutional neural network, the self-attention mechanism and the multi-layer fully connected neural network to obtain a fused deep neural network;

A training module, configured to perform step-by-step training on the fused deep neural network through the training data to obtain a target model with a noise reduction function;

The noise reduction and detection module is used to input real data with a low signal-to-noise ratio into the target model, so that the target model performs noise reduction and signal detection on the real data, and obtains processing results of noise reduction and signal detection.

A data noise reduction and signal detection system according to the third embodiment of the present application includes:

at least one memory;

at least one processor;

at least one program;

The program is stored in the memory, and the processor executes at least one of the programs to implement the data noise reduction and signal detection method as described in the embodiment of the first aspect.

According to the computer-readable storage medium according to the fourth embodiment of the present application, the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to cause the computer to execute the method described in the first embodiment. Data noise reduction and signal detection methods.

Additional aspects and advantages of the 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 the application.

Description of the drawings

The present application will be further described below in conjunction with the accompanying drawings and examples, wherein:

Figure 1 is a schematic flow chart of the data noise reduction and signal detection method provided by the embodiment of the present application;

Figure 2 is a waveform diagram of low signal-to-noise ratio data provided by the embodiment of the present application;

Figure 3 is a schematic diagram of a model of a fused deep neural network provided by an embodiment of the present application;

Figure 4 is a schematic structural diagram of the data noise reduction and signal detection device provided by the embodiment of the present application;

Figure 5 is a schematic structural diagram of the data noise reduction and signal detection system provided by the embodiment of the present application.

Reference signs:

Generation module 100, processing module 110, fusion module 120, training module 130, noise reduction and detection module 140, memory 200, processor 300.

Detailed ways

The embodiments of the present application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below with reference to the drawings are exemplary and are only used to explain the present application and cannot be understood as limiting the present application.

It should be noted that although the functional modules are divided in the system schematic diagram and the logical sequence is shown in the flow chart, in some cases, the modules can be divided into different modules in the system or the order in the flow chart can be executed. The steps shown or described. The terms used in the description, claims and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

In the description of this application, several means one or more, plural means two or more, greater than, less than, exceeding, etc. are understood to exclude the original number, and above, below, within, etc. are understood to include the original number. If there is a description of first and second, it is only for the purpose of distinguishing technical features, and cannot be understood as indicating or implying the relative importance or implicitly indicating the number of indicated technical features or implicitly indicating the order of indicated technical features. relation.

In the description of this application, unless otherwise explicitly limited, words such as setting, installation, and connection should be understood in a broad sense. Those skilled in the art can reasonably determine the specific meaning of the above words in this application in conjunction with the specific content of the technical solution.

In the description of this application, reference to the description of the terms "one embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples" is intended to be in conjunction with the description of the embodiment. or examples describe specific features, structures, materials, or characteristics that are included in at least one embodiment or example of the present application. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

Glossary:

Low signal-to-noise ratio: refers to the signal strength being lower than 0dB.

Data with low signal-to-noise ratio: usually refers to data whose signal strength is lower than 0dB, and the signal is completely submerged by noise;

Noise reduction: Suppress the noise in the data.

Signal detection: Determine whether there are signals of interest in the data, such as gravitational wave signals, communication signals, seismic wave signals, etc.

Deep space communication: refers to the communication between communication entities on the earth and aircraft that leave the earth's satellite orbit and enter the solar system. The distance can reach millions of kilometers, tens of millions of kilometers, or even hundreds of millions of kilometers;

Gravitational waves: Gravitational waves are ripples in space-time caused by accelerating mass.

Gravitational-wave observatory: It is a device used to detect gravitational waves in gravitational wave astronomy. By detecting gravitational waves, general relativity can be experimentally verified. Commonly used detectors include rod detectors and laser interferometers. The main operating principle of these detectors is to measure the impact of the passage of gravitational waves on the distance between two distant locations. Since the 1960s, multiple gravitational wave detectors have been built and commissioned one after another, with continuous improvements in detector sensitivity. Today, these detectors have the function of detecting gravitational wave sources within and outside the Milky Way, and are the main detection tools for gravitational wave astronomy.

Seismic exploration: refers to a geophysical exploration method that uses elastic waves caused by artificial excitation to use the difference in elasticity and density of underground media to infer the nature and shape of underground rock formations by observing and analyzing the propagation rules of seismic waves generated by artificial earthquakes in the underground. Seismic exploration is the most important method in geophysical exploration and the most effective method to solve oil and gas exploration problems. It is an important means of exploring oil and natural gas resources before drilling. It is also widely used in coalfield and engineering geological exploration, regional geological research and crustal research.

In scientific scenarios, such as deep space communications, gravitational wave detection, seismic exploration and other fields, we often face the problem of low signal-to-noise ratio data processing. Since signals are usually submerged by noise, traditional signal processing methods including matched filtering and nonlinear filters have problems such as poor noise reduction and low computational efficiency when processing low signal-to-noise ratio data, and cannot accurately reduce data. noise and detect potential signals within it.

Specifically, processing data with low signal-to-noise ratio has always been a research difficulty in the scientific field, especially in the fields of physics and communications. Since the noise intensity is much higher than the signal, effective methods are needed to de-noise the data to detect the data. Are there any useful or interesting signals? The current methods mainly have the following shortcomings:

Methods such as nonlinear filters have slow calculation speed, low noise reduction and detection accuracy, and cannot obtain accurate results;

The current neural network algorithm cannot handle ultra-long data, and the noise reduction and detection speed and accuracy need to be improved.

Based on the above, this application proposes a data noise reduction and signal detection method, which can effectively improve the noise reduction and detection accuracy, and can achieve data noise reduction and signal detection at the same time.

Next, the data noise reduction and signal detection method according to the embodiment of the present application will be described based on Figures 1-3.

It can be understood that, as shown in Figure 1, a data noise reduction and signal detection method is provided, including:

Step S100, generate simulation data with a low signal-to-noise ratio, where the simulation data includes multiple mixed signals, the mixed signals are obtained by mixing noise and pure signals, and the pure signals represent noise-free signals;

Step S110, preprocess the mixed signals in the simulation data to obtain training data;

Step S120, fuse the first convolutional neural network, the second convolutional neural network, the self-attention mechanism and the multi-layer fully connected neural network to obtain a fused deep neural network;

Step S130, perform step-by-step training on the fused deep neural network through training data to obtain a target model with noise reduction function;

Step S140, input the real data with low signal-to-noise ratio into the target model, so that the target model can perform noise reduction and signal detection on the real data, and obtain the processing results of noise reduction and signal detection.

First, simulated data with a low signal-to-noise ratio is generated. The simulated data includes multiple mixed signals. The mixed signals are obtained by mixing noise and pure signals. The pure signals represent noise-free signals. Secondly, the mixed signals in the simulated data are preprocessed. , obtain the training data; then, fuse the first convolutional neural network, the second convolutional neural network, the self-attention mechanism and the multi-layer fully connected neural network to obtain a fused deep neural network; then, use the training data to fuse the The deep neural network is trained step by step to obtain a target model with noise reduction function; finally, the real data with low signal-to-noise ratio is input into the target model, so that the target model can perform noise reduction and signal detection on the real data, and obtain the noise reduction and The processing result of signal detection. The data noise reduction and signal detection method of this application, on the one hand, through sufficient training based on simulated data, can enable the target model to improve noise reduction and detection accuracy when processing real data with low signal-to-noise ratio, and does not require frequent calculation, improving the speed of noise reduction and detection; on the other hand, by constructing a fused deep neural network, it can process ultra-long data and extract signal features of real data with low signal-to-noise ratio from different dimensions, thereby achieving Higher noise reduction and detection accuracy. Therefore, the data denoising and signal detection method of this application can accurately denoise and detect low signal-to-noise ratio data.

Understandably, simulation data that generates low signal-to-noise ratio include:

Simulate the generation of noise based on the sensitivity curve of the gravitational wave detector;

Select corresponding parameter ranges for different types of gravitational wave sources;

According to the parameter range, generate multiple pure signals corresponding to multiple gravitational wave sources, where the gravitational wave sources and the pure signals correspond one to one;

Each pure signal is mixed with noise separately to obtain analog data including multiple mixed signals.

It should be noted that in the gravitational wave detection scenario, simulation data with low signal-to-noise ratio can be generated through the following steps:

Step S101: Simulate and generate noise according to the sensitivity curve of the gravitational wave detector;

Step S102: Select appropriate parameter ranges for different types of gravitational wave sources to generate corresponding signals; among them, the gravitational wave sources can be extreme mass ratio precession systems, supermassive double black holes, double white dwarfs, or random gravitational wave backgrounds;

Step S103: Mix the data generated in steps S101 and S102 to obtain data with different signal-to-noise ratios; in this step, it is crucial to set a specific signal-to-noise ratio (SNR) according to the following formula.

Among them, s|s is the inner product, which can be a|b, and is calculated by the following formula:

The * sign in the formula represents complex conjugation, and S _n is the sensitivity curve.

It should be noted that in the field of deep space communication, noise can also be simulated and generated based on the signal curve of the communication between the communication entity and the earth satellite; in the field of seismic exploration, noise can be simulated and generated based on the curve corresponding to the seismic wave.

It is worth emphasizing that this application uses the matched filter signal-to-noise ratio, and when simulating the actual extremely low signal-to-noise ratio environment, the signal-to-noise ratio of the training data is deliberately set to 50, which is approximately equal to -40dB. After the above steps, a large amount of low signal-to-noise ratio simulation data is obtained. Among them, an example of a simulated data sample is shown in Figure 2. In Figure 2, orange is the pure signal and blue is the signal + noise.

Understandably, preprocessing of mixed signals in analog data includes:

Whiten the mixed signal in the analog data to make the noise spectrum of the mixed signal evenly distributed;

The whitened mixed signal is normalized so that all data in the simulated data have the same scale.

It should be noted that the data preprocessing part mainly consists of two steps: whitening and normalization, specifically including the following steps:

Step S111: Data whitening. The purpose of whitening is to remove the redundancy of the input data and improve its statistical characteristics; the specific calculation method is as follows, converting the noise spectrum of the original data into a uniformly distributed form, so that all frequency components are in There is a uniform distribution on the power spectrum, which can effectively reduce the complexity of noise and provide more homogeneous data for subsequent fused deep neural networks;

Step S112: Normalize the whitened data; normalization can ensure that the data has the same scale in the entire data set, making the training of the fused deep neural network more stable; in this step, adjust the simulation data to [ -1,1] to avoid numerical overflow, inaccurate calculation or instability in the learning process due to too large or too small values.

It can be understood that the first convolutional neural network, the second convolutional neural network, the self-attention mechanism and the multi-layer fully connected neural network are fused to obtain a fused deep neural network, including:

Build the first convolutional neural network to extract feature information from mixed signals;

Build a fusion network of the self-attention mechanism and the second convolutional neural network to reduce noise on mixed signals;

Build a multi-layer fully connected neural network to detect target signals.

It should be noted that in the model construction, please provide a fusion deep neural network algorithm, which is mainly composed of a convolutional neural network (CNN), a self-attention mechanism and a multi-layer fully connected neural network (MLP) ; The fused deep neural network model constructed in this application is an end-to-end design, which can directly perform denoising and detection on the input low signal-to-noise ratio data at the same time; this design avoids manual feature selection and threshold setting, and realizes data automatic processing and analysis.

Among them, Figure 3 gives an example of the overall structure of a model.

The specific steps to build the fused deep neural network model are as follows:

Step S121: Build the first convolutional neural network; use the convolutional neural network to perform the task of feature extraction. The data obtained in step S100 first passes through the CNN network; CNN has superior performance when processing data such as images, speech, and time series; In the model of this application, the role of CNN is to extract useful feature information from input data with low signal-to-noise ratio; more importantly, through the stacking of convolutional layers, the model has a larger receptive field and can process Ultra-long sequence data, thereby enhancing the accuracy of noise reduction and signal detection;

Step S122: Build the self-attention mechanism and the second convolutional neural network: This application designs an encoder module that integrates the self-attention mechanism and the convolutional neural network, specifically responsible for noise reduction processing. The principle of the self-attention mechanism is shown in the following formula, where Q, K, V are trainable parameters. This mechanism helps the neural network capture long-distance dependencies in the sequence, while CNN focuses on extracting local features; this unique The combination enables the fused deep neural network to consider both global and local information when dealing with noise reduction tasks.

Step S123: Build a multi-layer fully connected neural network; the multi-layer fully connected neural network is used as a classifier in the fused deep neural network model to detect target signals; MLP has the ability to learn nonlinear mapping relationships and can accurately reduce the Detect potential target signals in noisy data. The classifier formula is as follows:

Among them, ω _ij is the training parameter of the classifier, x _j is the output of step S123, and logistic is the logistic regression function.

It can be understood that the fused deep neural network is trained step by step through the training data to obtain a target model with noise reduction function, including:

The mean square error between the mixed signal and the pure signal is used as the loss function, and the fused deep neural network is initially trained through the loss function;

Set the number of alternating cycles, where the number of alternating cycles includes the number of first training cycles and the number of second training cycles. The number of first training cycles and the number of second training cycles each include multiple training cycles. Each training cycle represents the number of training cycles in the training data. The process of training all samples once;

Perform noise reduction training on the fused deep neural network within the first number of training cycles, and perform detection training on the fused deep neural network within the second number of training cycles;

Based on the number of alternation cycles, noise reduction training and detection training are performed alternately;

Use the optimizer to adjust the parameters of the fused deep neural network, and use the learning rate adjustment strategy to adjust the learning rate of the fused deep neural network.

It can be understood that the fused deep neural network is trained for noise reduction within the first number of training cycles, and the fused deep neural network is trained for detection within the second number of training cycles, including:

Perform noise reduction training on the fused deep neural network within the first number of training epochs to minimize the mean square error between the mixed signal and the pure signal;

The fused deep neural network is probe-trained within a second number of training cycles to minimize the cross-entropy loss of the binary classification, wherein the number of cycles of the first number of training cycles is equal to the number of cycles of the second number of training cycles.

It should be noted that in the model training stage, a unique step-by-step training strategy is introduced. This strategy involves alternating training of different loss functions to simultaneously achieve signal denoising and detection; the specific steps are as follows:

Step S131: Initial stage; first, in the initial stage, the mean square error between the mixed signal containing noise and the pure signal without noise is used as the loss function. The main goal is to let the network learn to reduce noise; the training in this stage helps The model initially understands and extracts valuable signal features, while also preventing the classifier from capturing excessive noise features;

Step S132: Alternate training; after the preliminary training is completed, a phased alternating training strategy is adopted, which is carried out every 6 epochs (an epoch represents the process of training once using all samples in the training set) (below) (two steps ab); in this training strategy, the model will alternate between noise reduction training and detection training, so that the two tasks can complement each other and improve the overall performance of the model. Specific steps are as follows:

a) Noise reduction training: In the first three epochs of each alternating cycle, training is mainly conducted on noise reduction tasks; the training goal of the model at this stage is to minimize the mean square error between the mixed signal and the pure signal, so that the model can be more accurate Good to learn how to denoise signals;

b) Detection training: In the last three epochs of each alternating cycle, the model is switched to training on detection tasks; the training goal of the model at this stage is to minimize the cross-entropy loss of the two classifications and improve the model's signal detection accuracy;

Step S133: Optimizer and learning rate adjustment; during the model training process, use the Adam optimizer to adjust network parameters, with an initial learning rate of 0.01, and a total training round of 100 epochs; at the same time, a learning rate adjustment strategy is introduced, when the training loss When stagnant, appropriately reduce the learning rate to ensure that the model can continue to learn effectively; through the above steps, the entire model training process not only improves the learning efficiency of the network, but also accelerates the convergence speed of the network, allowing the model to reach the ideal goal faster performance.

It can be understood that data noise reduction and signal detection methods also include:

The processing results are evaluated to determine the noise reduction accuracy and detection accuracy of the target model.

In this model, model inference is a step to determine whether the model meets the requirements; in the model inference stage, noise reduction and detection operations are the core steps. The specific process is as follows:

Step S141: Noise reduction and detection: Input the low signal-to-noise ratio data to be processed into the model trained in step S130, and the model will automatically perform noise reduction and signal detection tasks; first, the denoising network part passes through the convolutional neural network and self-attention mechanism for feature extraction and noise reduction; then, signal detection is performed through a pre-trained fully connected network.

Step S142: Result evaluation: Evaluate the denoised data and signal detection results output in step S141. The evaluation method mainly depends on the specific application scenario;

For example, in gravitational wave applications, the noise reduction accuracy is evaluated by calculating the overlap between the output noise reduction signal and the real signal, and the detection accuracy is evaluated by the ROC curve;

in:

In the above two formulas, h is the network output, s is the real signal, and the value of overlap is between [0,1]. The larger the value, the higher the degree of matching and the higher the noise reduction accuracy.

The data noise reduction and signal detection method of the present application will be further elaborated below in conjunction with the above embodiments.

It should be noted that traditional signal denoising methods include wavelet denoising methods, modular maximum denoising methods, etc. These methods have low generalization capabilities and require certain prior knowledge. For example, the determination of wavelet transform threshold and the selection of wavelet. Based on the learning of data representation, deep learning can better extract the key features hidden in the data and automatically filter out the influence of interference noise.

In deep learning technology, typical neural networks include convolutional neural networks, recurrent neural networks, fully connected neural networks, etc. Recurrent neural network is a neural network in which nodes are directionally connected into loops. It can better demonstrate dynamic sequential behavior. It can use the internal memory unit to process input sequences of any sequence and handle some timing problems, such as natural language processing, speech recognition, etc. Task.

The long short-term memory network is a special type of recurrent neural network. On the basis of the recurrent neural network, the input threshold, forgetting threshold and output threshold are added to make the weight change during the self-loop process, and the integration scale at different time nodes can be dynamically changed, cleverly avoiding the gradient generated during the loop process. vanishing or gradient expansion problems.

Specifically, wavelet transform is a frequency analysis method of radio signals. Compared with traditional Fourier transform analysis, wavelet transform has the advantages of time domain localization and frequency domain localization. It can decompose signals at different scales and retain the characteristics of signals under different modulation types. The signal has certain continuity in space or time, and its effective signal has a larger wavelet coefficient in the wavelet domain, while the noise signal generally presents a discrete state in space or time, and its wavelet coefficient in the wavelet domain is smaller. Utilizing this property, the noise reduction of the signal can be achieved through the wavelet transform method. Perform wavelet decomposition processing on the input original signal and calculate different wavelet coefficients. Assuming that the noise signal obeys Gaussian distribution, then most of the noise coefficients will be within a certain interval. Set the coefficients in this interval to zero to achieve Noise signals are suppressed to the greatest extent. The wavelet coefficients after threshold processing are used to reconstruct the radio signal and obtain the denoised signal.

Specifically, an autoencoder based on a convolutional neural network generally consists of a three-layer network, including an input layer, a hidden layer and an output layer, where the number of neurons in the input layer and the output layer is equal. During the training process, for each input sample, the autoencoder will produce an output sample of the same size. The optimization goal of autoencoder training is to make the output sample as close as possible to the input sample. On the basis of the autoencoder, this application proposes a De-noising Auto-encoder (DAE), which adds noise signals to the input data so that the trained autoencoder has a noise reduction function and is more efficient. Robustness improves the generalization ability of the model.

Existing radio signal recognition methods based on deep learning have good recognition effects on a large number of radio signal data, but in low signal-to-noise ratio areas, the recognition accuracy of these recognition methods is still low. To address this problem, this paper proposes a data denoising and signal detection method based on deep learning.

The data denoising and signal detection method of this application is implemented by setting up a low signal-to-noise ratio classifier, a denoising autoencoder and an identification network. The low signal-to-noise ratio classifier is essentially a two-classifier. By setting different The signal-to-noise ratio threshold can identify high signal-to-noise ratio signals and low signal-to-noise ratio signals in the signal-to-noise ratio. The denoising autoencoder can achieve denoising processing of low signal-to-noise ratio radio signals. The modulation type identification network is composed of a long short-term memory network (LSTM network) and can identify the modulation type of the input radio signal.

Specifically, inspired by the long short-term memory neural network, this application designed a radio signal modulation type identification model based on the long short-term memory network, including three levels: LSTM (128), LSTM (32), and FC (11), where , the input original signal size is len*2, len represents the number of sampling nodes, and 2 represents the time dimension of a certain sampling node. LSTM(128) represents mapping the time dimension of input data into a feature space of size len*128. FC(11) represents a fully connected network, which maps the input data to 11 distribution areas, where 11 is determined by the type of modulation type of the training data set. The LSTM layer uses tanh as the activation function, and the dropout is 0.8. The fully connected layer uses softmax as the activation function, cross entropy is used as the loss function for training, Adam with a learning rate of 0.001 is selected as the optimizer, batchsize is 64, and epoch is 20. Select top_one as the evaluation index of the model, that is, the model identifies correctly only when the class label corresponding to the highest confidence level is the correct class label.

The long short-term memory network is used to extract the characteristic information of the data from the radio signal data, and the classification of different types of signals is achieved based on the characteristic information. The noise reduction effect of the radio signal recognition model based on the long short-term memory network is judged by comparing the accuracy of type recognition before and after radio signal noise reduction.

Specifically, for low signal-to-noise ratio classifiers, the recognition model has good classification effects in high signal-to-noise ratio areas, but the recognition accuracy in low signal-to-noise ratio intervals is extremely low. Therefore, it is necessary to perform noise reduction processing on radio signals in the low signal-to-noise ratio interval to improve the accuracy of the identification interval of the recognition model. To perform noise reduction processing on low signal-to-noise ratio signals, we first need to extract the low signal-to-noise ratio signals from the signal data. In order to more accurately screen out low signal-to-noise ratio signals, this application designs a radio signal classifier based on LSTM to denoise low signal-to-noise ratio signals.

For example, the classifier can be such that the input signal size is len*2, where len represents the number of sampling nodes, and 2 represents the time dimension of a certain sampling node. LSTM(32) represents mapping the time dimension of the input data into a feature space of size len*32. The LSTM layer uses tanh as the activation function, and the dropout is 0.8. The FC layer is a fully connected layer, using softmax as the activation function. Cross entropy is used as the loss function for training, Adam with a learning rate of 0.008 is selected as the optimizer, batchsize is 64, and epoch is 20.

The low signal-to-noise ratio classification network can be used to map radio signals into two intervals, high signal-to-noise ratio and low signal-to-noise ratio, to achieve a two-classification task of radio signals based on LSTM.

The radio signal noise reduction and reconstruction model proposed in this application based on auto-encoding technology adds white noise obeying Gaussian distribution to the high signal-to-noise ratio signal, and inputs these signals containing additive noise and the original high signal-to-noise ratio signal into the reduction model. Noise model is trained. Use the trained noise reduction model to achieve noise reduction and reconstruction processing of low signal-to-noise ratio signals.

Generally speaking, this application proposes a data denoising and signal detection method based on autoencoding technology to solve the problem of low recognition accuracy of existing models in low signal-to-noise ratio areas, combined with autoencoder technology, which can to a certain extent. Improving the recognition accuracy of the modulation type recognition model based on deep learning in the low signal-to-noise ratio range can make full use of limited signal resources.

This application first uses LSTM to build a radio signal type recognition model to identify the types of radio signals with different signal-to-noise ratios. It uses a low signal-to-noise ratio classifier to achieve the second classification task of radio signals based on LSTM. Then it builds radio signal noise reduction based on auto-encoding technology. Reconstruct the model to achieve noise reduction and reconstruction of low signal-to-noise ratio radio signals. Finally, use the data after noise reduction and reconstruction to retrain the type recognition model and test it.

It can be understood that, as shown in Figure 4, this application also provides a data noise reduction and signal detection device, including:

The generation module 100 is used to generate analog data with a low signal-to-noise ratio, where the analog data includes multiple mixed signals. The mixed signals are obtained by mixing noise and pure signals, and the pure signals represent noise-free signals;

The processing module 110 is used to preprocess the mixed signals in the simulation data to obtain training data;

The fusion module 120 is used to fuse the first convolutional neural network, the second convolutional neural network, the self-attention mechanism and the multi-layer fully connected neural network to obtain a fused deep neural network;

The training module 130 is used to perform step-by-step training on the fused deep neural network through training data to obtain a target model with noise reduction function;

The noise reduction and detection module 140 is used to input real data with low signal-to-noise ratio into the target model, so that the target model can perform noise reduction and signal detection on the real data, and obtain processing results of noise reduction and signal detection.

It should be noted that the generation module 100 includes:

A simulation module used to simulate noise generation based on the sensitivity curve of the gravitational wave detector;

The selection module is used to select corresponding parameter ranges for different types of gravitational wave sources;

A signal module, used to generate multiple pure signals corresponding to multiple gravitational wave sources according to the parameter range, where the gravitational wave sources and the pure signals correspond one to one;

The mixing module is used to mix each pure signal with noise respectively to obtain analog data including multiple mixed signals.

It should be noted that the processing module 110 includes:

The whitening processing module is used to whiten the mixed signals in the analog data so that the noise spectrum of the mixed signals is evenly distributed;

The normalization processing module is used to normalize the whitened mixed signal so that all data in the simulation data have the same scale.

It should be noted that the fusion module 120 includes:

The first building module is used to build the first convolutional neural network to extract feature information from the mixed signal;

The second building module is used to build a fusion network of the self-attention mechanism and the second convolutional neural network to denoise the mixed signal;

The third building module is used to build a multi-layer fully connected neural network to detect target signals.

It should be noted that the training module 130 includes:

The initial training module is used to use the mean square error between the mixed signal and the pure signal as the loss function, and perform initial training of the fused deep neural network through the loss function;

The setting module is used to set the number of alternating cycles, wherein the number of alternating cycles includes a first training cycle number and a second training cycle number, and both the first training cycle number and the second training cycle number include multiple training cycles, each of which The training cycle represents the process of training all samples in the training data once;

The noise reduction and detection training module is used to perform noise reduction training on the fused deep neural network within the first number of training cycles, and to perform detection training on the fused deep neural network within the second number of training cycles;

The alternating training module is used to alternately perform noise reduction training and detection training based on the number of alternating cycles;

The adjustment module is used to use the optimizer to adjust the parameters of the fused deep neural network, and use the learning rate adjustment strategy to adjust the learning rate of the fused deep neural network.

It should be noted that the noise reduction and detection training module includes:

The first minimization module is used to perform noise reduction training on the fused deep neural network within the first number of training cycles to minimize the mean square error between the mixed signal and the pure signal;

The second minimization module is used to perform detection training on the fused deep neural network within the second number of training cycles to minimize the cross-entropy loss of the two classifications, where the number of cycles in the first number of training cycles is equal to the number of second training cycles number of cycles.

It should be noted that the data noise reduction and signal detection device also includes:

The evaluation module is used to evaluate the processing results to determine the noise reduction accuracy and detection accuracy of the target model.

The following describes a data noise reduction and signal detection system according to an embodiment of the present application with reference to FIG. 5 .

It can be understood that, as shown in Figure 5, the data noise reduction and signal detection system includes:

at least one memory 200;

At least one processor 300;

at least one program;

The program is stored in the memory 200, and the processor 300 executes at least one program to implement the above-mentioned data noise reduction and signal detection method. Figure 5 takes a processor 300 as an example.

The processor 300 and the memory 200 may be connected through a bus or other means. Figure 5 takes the connection through a bus as an example.

As a non-transitory computer-readable storage medium, the memory 200 can be used to store non-transitory software programs, non-transitory computer executable programs and signals, such as those corresponding to the data noise reduction and signal detection system in the embodiment of the present application. Program instructions/signals. The processor 300 executes various functional applications and data processing by running non-transient software programs, instructions and signals stored in the memory 200, that is, implementing the data noise reduction and signal detection methods of the above method embodiments.

The memory 200 may include a program storage area and a data storage area, where the program storage area may store an operating system and an application program required for at least one function; the storage data area may store data related to the above-mentioned data noise reduction and signal detection methods, etc. In addition, the memory 200 may include high-speed random access memory, and may also include non-transitory memory, such as at least one magnetic disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory 200 optionally includes memories remotely located relative to the processor 300 , and these remote memories can be connected to the data noise reduction and signal detection system through a network. Examples of the above-mentioned networks include, but are not limited to, the Internet of Things, software-defined networks, sensor networks, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

One or more signals are stored in the memory 200, and when executed by one or more processors 300, the data noise reduction and signal detection methods in any of the above method embodiments are performed. For example, the method in Figure 1 described above is performed.

The following describes a computer-readable storage medium according to an embodiment of the present application with reference to FIG. 5 .

As shown in Figure 5, the computer-readable storage medium stores computer-executable instructions that are executed by one or more processors 300, for example, by one processor 300 in Figure 5, which can cause the above-mentioned one Or multiple processors 300 execute the data noise reduction and signal detection methods in the above method embodiments. For example, the method in Figure 1 described above is performed.

The system embodiments described above are only illustrative, and the units illustrated as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in one place. , or it can be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

Through the description of the above embodiments, those of ordinary skill in the art can understand that all or some steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and appropriate combinations thereof. Some or all of the physical components may be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or as hardware, or as an integrated circuit, such as an application specific integrated circuit . Such software may be distributed on computer-readable media, which may include computer storage media and communication media. As is known to those of ordinary skill in the art, the term computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, Digital Versatile Disk or other optical disk storage, magnetic cassettes, magnetic tape, disk storage or other magnetic storage devices, or may be used for storage Any other medium on which the information is desired and can be accessed by the computer. Additionally, it is known to those of ordinary skill in the art that communication media typically embodies a computer-readable signal, data structure, program module or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media .

The embodiments of the present application have been described in detail above in conjunction with the accompanying drawings. However, the present application is not limited to the above-mentioned embodiments. Within the scope of knowledge possessed by those of ordinary skill in the art, various embodiments can be made without departing from the purpose of the present application. Variety. In addition, the embodiments of the present application and the features in the embodiments may be combined with each other without conflict.

Claims (10)

1. Data noise reduction and signal detection method, characterized by: Generating simulation data with a low signal-to-noise ratio, wherein the simulation data includes a plurality of mixed signals, the mixed signals are obtained by mixing noise and pure signals, and the pure signals represent noise-free signals; Preprocess the mixed signals in the simulation data to obtain training data; The first convolutional neural network, the second convolutional neural network, the self-attention mechanism and the multi-layer fully connected neural network are fused to obtain a fused deep neural network; Perform step-by-step training on the fused deep neural network using the training data to obtain a target model with noise reduction function; Real data with a low signal-to-noise ratio is input into the target model, so that the target model performs noise reduction and signal detection on the real data, and obtains processing results of noise reduction and signal detection. 2. The data noise reduction and signal detection method according to claim 1, characterized in that said generating simulation data with low signal-to-noise ratio includes: Simulate the generation of noise based on the sensitivity curve of the gravitational wave detector; Select corresponding parameter ranges for different types of gravitational wave sources; According to the parameter range, generate multiple pure signals corresponding to multiple gravitational wave sources, where the gravitational wave sources and the pure signals correspond one to one; Each of the pure signals is mixed with the noise respectively to obtain analog data including a plurality of mixed signals. 3. The data noise reduction and signal detection method according to claim 1, characterized in that the preprocessing of the mixed signals in the analog data includes: Perform whitening processing on the mixed signal in the simulated data so that the noise spectrum of the mixed signal is evenly distributed; The whitened mixed signal is normalized so that all data in the simulated data have the same scale. 4. The data noise reduction and signal detection method according to claim 1, characterized in that the first convolutional neural network, the second convolutional neural network, a self-attention mechanism and a multi-layer fully connected neural network are performed. Fusion to obtain a fused deep neural network, including: Build the first convolutional neural network to extract feature information from the mixed signal; Build a fusion network of the self-attention mechanism and the second convolutional neural network to perform noise reduction processing on the mixed signal; The multi-layer fully connected neural network is built to detect the target signal. 5. The data noise reduction and signal detection method according to claim 1, characterized in that, the fused deep neural network is trained step by step through the training data to obtain a target model with a noise reduction function, include: The mean square error between the mixed signal and the pure signal is used as a loss function, and the fused deep neural network is initially trained through the loss function; Set the number of alternating cycles, wherein the number of alternating cycles includes a first training cycle number and a second training cycle number, and both the first training cycle number and the second training cycle number include multiple training cycles, each The training cycle represents the process of training all samples in the training data once; Perform noise reduction training on the fused deep neural network within the first number of training cycles, and perform detection training on the fused deep neural network within the second number of training cycles; Based on the number of alternation cycles, noise reduction training and detection training are alternately performed; An optimizer is used to adjust the parameters of the fused deep neural network, and a learning rate adjustment strategy is used to adjust the learning rate of the fused deep neural network. 6. The data noise reduction and signal detection method according to claim 5, characterized in that, the fused deep neural network is trained for noise reduction within the first number of training cycles, and the noise reduction training is performed for the second number of training cycles. Detection training of the fused deep neural network within the number of training cycles includes: Perform noise reduction training on the fused deep neural network within the first number of training cycles to minimize the mean square error between the mixed signal and the pure signal; Perform detection training on the fused deep neural network within the second number of training cycles to minimize the cross-entropy loss of binary classification, wherein the number of cycles of the first number of training cycles is equal to the number of second training cycles number of cycles. 7. The data noise reduction and signal detection method according to claim 1, characterized in that the data noise reduction and signal detection method further includes: The processing results are evaluated to determine the noise reduction accuracy and detection accuracy of the target model. 8. Data noise reduction and signal detection device, characterized by including: A generation module, configured to generate analog data with a low signal-to-noise ratio, wherein the analog data includes a plurality of mixed signals, the mixed signals are obtained by mixing noise and pure signals, and the pure signals represent noise-free signals; A processing module, used to preprocess the mixed signal in the simulation data to obtain training data; The fusion module is used to fuse the first convolutional neural network, the second convolutional neural network, the self-attention mechanism and the multi-layer fully connected neural network to obtain a fused deep neural network; A training module, configured to perform step-by-step training on the fused deep neural network through the training data to obtain a target model with a noise reduction function; The noise reduction and detection module is used to input real data with a low signal-to-noise ratio into the target model, so that the target model performs noise reduction and signal detection on the real data, and obtains processing results of noise reduction and signal detection. 9. Data noise reduction and signal detection system, which is characterized by: at least one memory; at least one processor; at least one program; The program is stored in the memory, and the processor executes at least one of the programs to implement the data noise reduction and signal detection method according to any one of claims 1 to 7. 10. Computer-readable storage medium, characterized in that the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to cause the computer to execute the data according to any one of claims 1 to 7 Noise reduction and signal detection methods.