CN112566174A - Abnormal I/Q signal identification method and system based on deep learning - Google Patents
Abnormal I/Q signal identification method and system based on deep learning Download PDFInfo
- Publication number
- CN112566174A CN112566174A CN202011399508.5A CN202011399508A CN112566174A CN 112566174 A CN112566174 A CN 112566174A CN 202011399508 A CN202011399508 A CN 202011399508A CN 112566174 A CN112566174 A CN 112566174A
- Authority
- CN
- China
- Prior art keywords
- signal
- data
- neural network
- network structure
- abnormal
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W76/00—Connection management
- H04W76/20—Manipulation of established connections
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W24/00—Supervisory, monitoring or testing arrangements
- H04W24/08—Testing, supervising or monitoring using real traffic
Abstract
The invention discloses an abnormal I/Q signal identification method and system based on deep learning, wherein the method comprises the steps of obtaining an original I/Q signal to obtain I (t) and Q (t) signal components, wherein t is a time variable; preprocessing an original I/Q signal to form a training data set; constructing a signal identification model, wherein the signal identification model comprises a convolutional neural network structure, a cyclic neural network structure and a full-connection neural network structure; training the signal recognition model based on a training data set; and outputting an I/Q signal identification result by utilizing the I/Q signal to be identified which is trained to the optimal signal identification model, wherein the I/Q signal identification result comprises a category and confidence thereof, and the category is a normal signal or an abnormal signal. The invention aims to solve the problems of high cost and low accuracy of abnormal I/Q signal identification.
Description
Technical Field
The invention belongs to the technical field of wireless signal identification, and particularly relates to an abnormal I/Q signal identification method and system based on deep learning.
Background
In order to expand signal processing bandwidth in the field of wireless communication, frequency conversion technology based on I/Q quadrature modulation and demodulation is becoming increasingly popular. The frequency conversion technology based on the I/Q quadrature modulation and demodulation can simplify the frequency conversion times, has the advantages of high integration, low power consumption and the like, and provides a flexible I/Q channel for digital processing.
In electronic warfare applications, when a competitor works in an unauthorized frequency band bandwidth or an unauthorized device by adopting a certain transmission mode, a frequency band access enhancement phenomenon occurs, and detection through an abnormal signal detection technology is urgently needed; in addition, in radio communications, abnormal signal detection also provides the ability to quickly identify interfering transmitters, malfunctioning equipment, or malicious attacks within its licensed band to facilitate timely action.
Most of such applications focus on detecting changes in the fingerprint characteristics or expert characteristics of the sensor without paying attention to anomalies in the high-rate raw physical layer radio signal itself, which are prone to identification errors. Moreover, the expert system corresponding to the feature extraction usually has higher requirements on the professional performance of personnel and the computing capability of equipment, and the implementation cost is very high.
Disclosure of Invention
The invention aims to provide an abnormal I/Q signal identification method and system based on deep learning, and aims to solve the problems of high cost and low accuracy of abnormal I/Q signal identification.
In order to achieve the purpose, the technical scheme adopted by the invention is as follows:
an abnormal I/Q signal identification method based on deep learning, comprising the following steps:
step 2.1, respectively carrying out filtering operation on the I (t) signal component and the Q (t) signal component by adopting a one-dimensional low-pass filter;
step 2.2, respectively adopting sliding windows to segment the filtered I (t) signal components and the filtered Q (t) signal components, and obtaining signal samples with the same lengths of the I (t) signal components and the Q (t) signal components by segmentation;
step 2.3, correspondingly splicing the I (t) signal components and the Q (t) signal components section by section to form a data sequence comprising a plurality of sections of complex signal samples, carrying out time-frequency analysis on each section of complex signal sample to obtain a two-dimensional time-frequency graph, and converting the two-dimensional time-frequency graph into one-dimensional data;
step 2.4, splitting the complex structure in the one-dimensional data obtained after conversion to obtain two paths of frequency domain characterization data, respectively calculating the original data sequence containing multiple sections of complex signal samples and the mode of each complex number in the one-dimensional data obtained after conversion, and then obtaining two paths of mode data;
step 2.5, correspondingly splicing the two paths of signal components I (t) and Q (t), the two paths of frequency domain representation data and the two paths of analog data which are obtained by segmentation section by section according to the division of a sliding window to obtain six paths of signal data so as to construct a training data set;
step 3, constructing a signal identification model, wherein the signal identification model comprises a convolutional neural network structure, a cyclic neural network structure and a fully-connected neural network structure, the convolutional neural network structure is used for extracting the spatial characteristics of input data, the cyclic neural network structure extracts time sequence characteristics on the basis of the spatial characteristics to obtain time-space domain characteristics, and the fully-connected neural network structure outputs an I/Q signal identification result on the basis of the time-space domain characteristics;
and 5, outputting an I/Q signal identification result to be identified by utilizing the I/Q signal to be identified trained to the optimal signal identification model, wherein the I/Q signal identification result comprises a category and a confidence coefficient thereof, and the category is a normal signal or an abnormal signal.
Several alternatives are provided below, but not as an additional limitation to the above general solution, but merely as a further addition or preference, each alternative being combinable individually for the above general solution or among several alternatives without technical or logical contradictions.
Preferably, the convolutional neural network structure comprises a plurality of composite network layers connected in sequence, and the composite network layer comprises a convolutional layer and a max pooling layer which are connected in sequence from a data input side to a data output layer.
Preferably, the convolutional neural network structure further comprises an inclusion structure, the inclusion structure comprises four branches connected to the last composite network layer, and the outputs of all the branches are superposed to serve as the spatial features output by the convolutional neural network structure;
the first branch uses a layer of 1 × 1 convolution; the second branch comprises two convolution layers, the sizes of the corresponding convolution kernels are 1 multiplied by 1 and 2 multiplied by 1 respectively; the third branch adopts two convolution layers with convolution kernel sizes of 1 multiplied by 1 and 4 multiplied by 1 respectively; the fourth branch is preceded by an average pooling operation with a window size of 2, followed by a convolution layer with a convolution kernel size of 1 x 1.
The invention also provides an abnormal I/Q signal identification system based on deep learning, which comprises the following components:
the signal acquisition module is used for acquiring an original I/Q signal to obtain I (t) and Q (t) signal components, wherein t is a time variable;
a signal processing module, configured to pre-process the original I/Q signal to form a training data set, and specifically perform the following operations:
respectively carrying out filtering operation on the I (t) signal component and the Q (t) signal component by adopting a one-dimensional low-pass filter;
respectively adopting sliding windows to divide the filtered I (t) signal components and the filtered Q (t) signal components to obtain signal samples with the same lengths of the I (t) signal components and the Q (t) signal components;
correspondingly splicing the I (t) signal components and the Q (t) signal components section by section to form a data sequence comprising a plurality of sections of complex signal samples, performing time-frequency analysis on each section of complex signal sample to obtain a two-dimensional time-frequency diagram, and converting the two-dimensional time-frequency diagram into one-dimensional data;
splitting a complex structure in the one-dimensional data obtained after conversion to obtain two paths of frequency domain representation data, respectively calculating a data sequence comprising a plurality of sections of complex signal samples and a mode of each complex in the one-dimensional data obtained after conversion, and then obtaining two paths of mode data;
correspondingly splicing the two paths of signal components I (t) and Q (t), the two paths of frequency domain representation data and the two paths of analog data obtained by segmentation section by section according to the division of a sliding window to obtain six paths of signal data so as to construct a training data set;
the model construction module is used for constructing a signal identification model, the signal identification model comprises a convolutional neural network structure, a cyclic neural network structure and a fully-connected neural network structure, the convolutional neural network structure is used for extracting the spatial characteristics of input data, the cyclic neural network structure extracts time sequence characteristics on the basis of the spatial characteristics to obtain time-space domain characteristics, and the fully-connected neural network structure outputs an I/Q signal identification result on the basis of the time-space domain characteristics;
a model training module for training the signal recognition model based on the training data set;
and the signal identification module is used for outputting an I/Q signal identification result to be identified by utilizing the I/Q signal trained to the optimal signal identification model, wherein the I/Q signal identification result comprises a category and a confidence coefficient thereof, and the category is a normal signal or an abnormal signal.
Preferably, the convolutional neural network structure comprises a plurality of composite network layers connected in sequence, and the composite network layer comprises a convolutional layer and a max pooling layer which are connected in sequence from a data input side to a data output layer.
Preferably, the convolutional neural network structure further comprises an inclusion structure, the inclusion structure comprises four branches connected to the last composite network layer, and the outputs of all the branches are superposed to serve as the spatial features output by the convolutional neural network structure;
the first branch uses a layer of 1 × 1 convolution; the second branch comprises two convolution layers, the sizes of the corresponding convolution kernels are 1 multiplied by 1 and 2 multiplied by 1 respectively; the third branch adopts two convolution layers with convolution kernel sizes of 1 multiplied by 1 and 4 multiplied by 1 respectively; the fourth branch is preceded by an average pooling operation with a window size of 2, followed by a convolution layer with a convolution kernel size of 1 x 1.
The invention also provides an abnormal I/Q signal identification system based on deep learning, which comprises a memory and a processor, wherein the memory stores a computer program, and the processor realizes the steps of the abnormal I/Q signal identification method based on deep learning when executing the computer program.
The abnormal I/Q signal identification method based on deep learning provided by the invention is different from the existing method based on artificial design for representing characteristics. The method can automatically extract the characteristics with more representation capability in the original I/Q signals and has strong classification modeling capability; the method does not depend on feature analysis and expert knowledge in a specific field or a specific scene, and can realize more accurate abnormality identification at lower cost.
Drawings
FIG. 1 is a flow chart of an abnormal I/Q signal identification method based on deep learning according to the present invention;
FIG. 2 is a flow chart of a method for deep learning based abnormal I/Q signal identification in an embodiment of the present invention;
FIG. 3 is a flow chart of the off-line training phase of the present invention;
FIG. 4 is a schematic diagram of a convolutional neural network structure according to the present invention;
FIG. 5 is a schematic diagram of a recurrent neural network architecture of the present invention;
FIG. 6 is a flow chart of the online identification phase of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The terminology used in the description of the invention herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.
In one embodiment, the abnormal I/Q signal identification method based on deep learning is provided, the abnormal I/Q signal can be accurately identified on the premise of low implementation cost, and the method has great application significance in the fields of electronic warfare, communication, broadcasting and the like.
The specific implementation steps of the abnormal I/Q signal identification method based on deep learning of the embodiment are shown in fig. 1, and include the following steps:
The original I/Q signal includes two signal components, I (t) ═ a (t) cos (Φ (t) - θ) and Q (t) ═ a (t) sin (Φ (t) - θ). Therefore, in the embodiment, the original I/Q signal is analyzed as two independent signals, so as to obtain more effective characteristics.
and 2.1, respectively carrying out filtering operation on the I (t) signal component and the Q (t) signal component by adopting a one-dimensional low-pass filter.
In this embodiment, a commonly used one-dimensional low-pass filter, a butterworth filter, is preferably used to perform filtering operations on the i (t) and q (t) signal components, respectively, so as to remove high-frequency noise in the signals. However, the present invention is not limited to a specific filter type, and other common one-dimensional signal filters may be used, for example, a gaussian filter may be used to eliminate white noise generated by a signal during communication, or a chebyshev filter may be used. The filter cutoff frequency may be set empirically for different signal types or based on statistical information of the signal data.
And 2.2, respectively adopting sliding windows to segment the filtered I (t) signal components and the filtered Q (t) signal components, and obtaining signal samples with the same lengths of the I (t) signal components and the Q (t) signal components by segmentation.
Since the signal is continuously present over time, and the training and recognition of the neural network are more precise to process fixed-length data, the present embodiment uses a sliding window for the segmentation process. The length of the window for sliding division is selected according to the signal frequency bandwidth and the sampling frequency, and the sliding step width generally takes a certain proportion of the window length, for example, 1/2 time window is selected. This embodiment preferably takes 128 sample points as an example of the window length, and selects the step value as 64 sample points.
And 2.3, correspondingly splicing the I (t) signal components and the Q (t) signal components section by section to form a data sequence comprising a plurality of sections of complex signal samples, carrying out time-frequency analysis on each section of complex signal sample to obtain a two-dimensional time-frequency graph, and converting the two-dimensional time-frequency graph into one-dimensional data.
After splicing the signal samples of the I (t) and Q (t) signal components, a data sequence IQ including multiple sections of complex signal samples is formedcomplexI.e. { Ik+QkJ | k ∈ (1,128) }, j being the corresponding imaginary unit. Then the data sequence IQ of the paircomplexAnd performing basic time-frequency analysis to clearly characterize the relation of the frequency of the signal changing along with time.
Time-frequency analysis can extract frequency domain information of signals in each sliding time window, and representation of radar signals is expanded. In this embodiment, preferably, short-time fourier transform is used to sequentially perform spectrum analysis on signals in a plurality of continuously overlapped small time windows, specifically, STFT is used to convert a time domain signal in a short time window into a representation of a frequency domain, and a spectral function is used to perform solution, so as to obtain a two-dimensional time-frequency diagram.
In the original I/Q signal, the I (t) and Q (t) signal components are all one-dimensional data, and in order to facilitate the uniform application of various types of data, the two-dimensional time-frequency diagram after time-frequency analysis needs to be converted into one-dimensional data STFcomplex. And taking time as a variable during conversion, and extracting frequency data as converted one-dimensional data according to time distribution.
And 2.4, splitting the complex structure in the one-dimensional data obtained after conversion to obtain two paths of frequency domain characterization data, respectively calculating the original data sequence containing multiple sections of complex signal samples and the mode of each complex number in the one-dimensional data obtained after conversion, and then obtaining two paths of mode data.
Because the data sequence after time-frequency analysis is a complex signal sample with a complex structure, the one-dimensional data can be split with the complex structure to obtain two paths of data representing the signal in the frequency domain.
For data in a complex form, analysis needs to be performed from two aspects, namely a real part and an imaginary part, and the change of a signal can be relatively closely reflected, but the size of the data cannot be directly obtained.
The introduction of the module data not only increases the multi-directional utilization of the original I/Q signal, but also avoids the problem that identification errors are easy to occur to samples which are too thin; moreover, for two signal components and two paths of data which are characterized in a frequency domain, the two signal components and the two paths of data are both single data of an I (t) signal or a Q (t) signal, and the introduction of the complex modulus increases the data of the fusion I (t) signal and the Q (t) signal, so that the defect that the analysis of the single signal component is easily influenced by noise or a fluctuation point is overcome, and the identification accuracy of the I/Q signal in the embodiment is effectively improved.
And 2.5, correspondingly splicing the two paths of signal components I (t) and Q (t), the two paths of frequency domain representation data and the two paths of analog data which are obtained by segmentation section by section according to the division of a sliding window to obtain six paths of signal data so as to construct a training data set.
The invention obtains 6-channel signal data after processing based on the original I/Q signals, and splices the samples to obtain the training data of the neural network, which can be expressed as
Where n _ samples represents the number of samples, n _ channels represents the number of data channels, here 6, and n _ win represents the sliding window length, here 128.
It should be noted that, in the present embodiment, automatic characterization learning is performed based on 6-channel signal data, so that it is ensured that there is comprehensive and effective feature utilization, so as to improve the accuracy of signal identification. In other embodiments, however, on the premise of low requirement for identification accuracy, two signal components of the original I/Q signal may be directly input to the neural network for automatic characterization learning. That is, on the basis of the 6-channel signal data of the present invention, the form of performing simple channel signal data change or combination is within the scope of the present invention.
And 3, constructing a signal identification model, wherein the signal identification model comprises a convolutional neural network structure, a cyclic neural network structure and a fully-connected neural network structure, the convolutional neural network structure is used for extracting the spatial characteristics of input data, the cyclic neural network structure is used for extracting time sequence characteristics on the basis of the spatial characteristics to obtain time-space domain characteristics, and the fully-connected neural network structure outputs an I/Q signal identification result on the basis of the time-space domain characteristics.
In the embodiment, a convolutional neural network structure is adopted to extract the spatial characteristics of the I/Q signals, a cyclic neural network structure is adopted to extract the time sequence characteristics of the I/Q signals, the I/Q signals are automatically characterized and learned from two layers of a time-space domain, the learning effect is good, the I/Q signals are comprehensively analyzed, the characteristics of the I/Q signals are modeled through a fully-connected neural network structure, and abnormal I/Q signals are identified.
Wherein, the input of the convolutional neural network layer structure is IQtrainThe output is a spatial feature set XS. Since I (t) and Q (t) are all one-dimensional signals, the two-way signal data and the four-way data obtained after preprocessing are alternately subjected to one-dimensional convolution operation and maximum pooling operation. The choice of the convolution kernel and the pooling window size can be set empirically for different signal types or based on statistical information of the signal data.
Specifically, in one embodiment, the convolutional neural network structure comprises a plurality of sequentially connected composite network layers, including convolutional layers and max-pooling layers sequentially connected from a data input side to a data output layer.
In order to further increase the width of the network structure, the convolutional neural network structure further comprises an inclusion structure, the inclusion structure comprises four branches behind the current composite network layer, namely the inclusion structure comprises four branches connected to the last composite network layer, and the outputs of all the branches are superposed to serve as the spatial features output by the convolutional neural network structure.
Wherein the first branch employs a layer of 1 × 1 convolution; the second branch comprises two convolution layers, the sizes of the corresponding convolution kernels are 1 multiplied by 1 and 2 multiplied by 1 respectively; the third branch adopts two convolution layers with convolution kernel sizes of 1 multiplied by 1 and 4 multiplied by 1 respectively; the fourth branch is preceded by an average pooling operation with a window size of 2, followed by a convolution layer with a convolution kernel size of 1 x 1. The embedded inclusion network structure can increase the network width, reduce the calculated parameter quantity and efficiently and accurately identify abnormal I/Q signals.
The recurrent neural network structure is realized based on the existing network structure, for example, an RNN network, an LSTM network, a GRU network and the like are adopted; the fully-connected neural network structure is also realized based on the existing network structure and is formed by utilizing a plurality of fully-connected layers, and the details are not repeated here.
And 4, training the signal recognition model based on the training data set. It is readily understood that training neural network models is a well-established technique in the field of deep learning neural networks and will not be described in detail here.
And 5, outputting an I/Q signal identification result to be identified by utilizing the I/Q signal to be identified trained to the optimal signal identification model, wherein the I/Q signal identification result comprises a category and a confidence coefficient thereof, and the category is a normal signal or an abnormal signal.
Corresponding to the training signal identification model, if the filtered I (t) and Q (t) signal components are directly adopted for training during training, I (t) and Q (t) signal components obtained after filtering the I/Q signal to be identified are also required to be input into the signal identification model for identification during identification; if the 6-channel signal data preferred in this embodiment is used for training during training, the I/Q signal to be recognized also needs to be converted to obtain 6-channel signal data during recognition, and then input to the signal recognition model for recognition.
The embodiment can automatically extract the characteristics with more representation capability in the original I/Q signal and has strong classification modeling capability; the method does not depend on feature analysis and expert knowledge in a specific field or a specific scene, and can realize more accurate abnormality identification at lower cost.
To further illustrate the abnormal I/Q signal identification method based on deep learning of the present invention, the following description is made by way of a specific example.
As shown in fig. 2, the abnormal I/Q signal recognition method based on deep learning in this embodiment mainly includes an offline training stage and an online recognition stage, where the offline training stage is mainly implemented on a background server with GPU computing resources, and the acquired I/Q signals are input into a signal recognition model as training data for training, and a network model is derived for online recognition after training; the online identification stage is mainly executed on front-end real-time abnormity identification equipment (such as an airborne radar analyzer) of the embedded computing board card, and real-time signals after identification are output to a feedback and alarm mechanism, so that abnormal I/Q signals are identified and processed in time.
As shown in fig. 3, the off-line training phase is implemented as follows:
firstly, collecting original I/Q signals to obtain two paths of signals of I (t) and Q (t).
Secondly, preprocessing the original I/Q signal, which specifically comprises the following steps:
1) and respectively filtering the signals I (t) and Q (t) by using a common one-dimensional low-pass filter Butterworth filter to remove high-frequency noise.
2) And respectively adopting sliding windows to divide the signals I (t) and Q (t) into signal samples with the same length. The length here is taken to be 128 sample points as an example.
3) Splicing the signals I (t) and Q (t) to form a complex signal sample data sequence IQcomplexI.e. { Ik+QkJ | k ∈ (1,128) }, performing basic time-frequency analysis, solving by selecting a spectrum function to obtain a two-dimensional time-frequency graph, and converting the two-dimensional time-frequency graph into one-dimensional data STFcomplexSo as to facilitate the parallel connection and splicing with the original I/Q signals.
4) Analyzing original signals of I (t) and Q (t) in time frequency and STFcomplexSplitting to obtain two paths of frequency domain characterization data, and respectively calculating IQcomplexAnd STFcomplexThe two paths of module data obtained by the module are spliced by 6 paths of signal data to obtain training data of the neural network, which can be expressed as
Where n _ samples represents the number of samples, n _ channels represents the number of data channels, here 6, and n _ win represents the sliding window length, here 128.
Thirdly, designing a convolutional neural network structure, where the convolutional neural network structure designed in this embodiment is a structure with inclusion, and is specifically shown in fig. 4:
1) to facilitate subsequent convolution operations, the training data IQ is first appliedtrainIs converted into
And moreover, a mini-batch optimization method is adopted during training, and a batch of data is input every time. The batch size is represented as n _ batch, then the input to the convolutional neural network is(n _ batch, n _ win, n _ channels), for example, 1024, then
2) The first four layers of the convolutional network are four operations of "convolution + pooling", where the convolution kernel is one-dimensional, and the largest pooling operation is selected, specifically, const (kernel _ size ═ 2, streams ═ 1, streams ═ 16), max (pool _ size ═ 2, streams ═ 2), const (kernel _ size ═ 3, streams ═ 1, streams ═ 16), max (pool _ size ═ 4, streams ═ 2), const (kernel _ size ═ 4, streams ═ 1, streams ═ 32), max (pool _ size ═ 2, streams ═ 2), const (kernel _ size ═ 5, streams ═ 1, streams ═ 32), max (pool _ size ═ 2, streams ═ 2), const (kernel _ size ═ 4, streams ═ 1, streams ═ 32), and max (pool _ size ═ 2, streams ═ 8, and output a size of 8;
3) on the basis of the output of the previous step, four different operations are respectively carried out, which mainly comprise 3 convolution operations of 1 × 1 and an average pooling operation with a window size of 2, and the corresponding filter numbers are respectively 32, 16 and 16. Then, the convolution operation is further performed for the last three of the four outputs, the number of filters is set to 32 in a unified manner, and the sizes of convolution kernels are set to 2 × 1, 4 × 1, and 1 × 1, respectively. Through these convolution and pooling calculations of different spatial scales, four parallel outputs of (n _ batch,8,32) size can be obtained.
4) The four outputs are connected in parallel in the channel dimension to obtain the convolutional layer neural network output X with the size of (n _ batch,8,128)S. Through the form, the information of the signal receptive fields with different sizes is described, the fusion of different spatial scale characteristics is realized, the signal has stronger representation capability, and the quantity of parameters needing to be calculated is reduced.
Fourthly, designing a recurrent neural network structure for extracting the time sequence characteristic X of the signalS_T. The input of the network layer is the output X of the convolutional layerSThe output is a time sequence characteristic set XS_T. The time sequence characteristics of input data are extracted by adopting a long-time memory recurrent neural network (LSTM) to solve the problems of gradient loss and gradient explosion in the training process.
As shown in figure 5 of the drawings,XSbefore inputting into the LSTM network, data dimension is converted, so that time sequence modeling on the LSTM corresponds to sampling of samples on a time sequence, and the direction of sequence learning converted from the data dimension is consistent with the direction of a time axis; and let X before entering LSTM network layerSAnd fusing the output of the convolutional neural network through a full-connection network layer. In addition, a Dropout layer is used in the LSTM network to deactivate some nodes in the network to prevent overfitting.
The number of LSTM layers is generally selected according to the actual situation, and 2 layers are selected here, but different network layer numbers may be selected in a specific application scenario. The number of hidden nodes of the LSTM cell can be according to XSIs determined by the dimension and empirical values, where 64 is selected. The number of sequential steps of the timing inference of the LSTM unit is determined from the data dimension obtained in the previous step, here 8. Obtaining time-space domain characteristic X of original I/Q signal through two layers of LSTM networksS_T。
Designing a full-connection neural network structure, comprising a plurality of full-connection neural network layers, and obtaining the time-space domain characteristics XS_TBinary classification is performed. The input of the time-space domain feature X is the extracted time-space domain feature XS_TThe output is a determination that the signal is a "normal signal" or an "abnormal signal", and their respective confidence levels. Due to the binary classification, the activation function may select Sigmoid, or may select Softmax, which is more general. The optimization function may select Adam or RMSProp, and is not particularly limited herein.
Sixthly, performing iterative training on the signal identification model designed in the third step to the fifth step on a server with GPU computing resources to obtain a signal identification model anomallyIQ with an optimal effect; and optimizing the signals through thinning and quantification operations, solidifying and deriving the signals for subsequent online abnormal signal identification. The optimization and derivation of the model can be selected by means of the currently popular deep learning framework, such as Tensorflow, Caffe, Pyorch and the like, as the case may be.
As shown in fig. 6, the specific implementation steps of the online identification phase are as follows:
1) and leading the trained anomallyiq model into front-end equipment with a computing board card, such as an airborne radar analyzer.
2) And (3) analyzing the radio data acquired by the radar receiver in real time, and performing online preprocessing on the I/Q dual-channel signals obtained by analysis, wherein the online preprocessing comprises noise reduction and filtering, windowing and segmentation, interpolation and completion data, short-time and long-time frequency analysis and the like, and is the same as the step two in an offline training stage.
3) Performing online real-time reasoning and identification on the data sample obtained in the step 2) through anomalyIQ to obtain a classification result and confidence of the signal to be identified.
4) If the detected result is 'abnormal signal', the relevant personnel is informed to process through a feedback and alarm mechanism on the front-end equipment. The hierarchical alarm handling mechanism may be set according to the specific application. For example, when the confidence of the abnormal signal is between 50% and 75%, the professional can be reminded to pay more attention, when the confidence of the abnormal signal is between 75% and 90%, the professional is required to perform manual review, and when the confidence of the abnormal signal is greater than 90%, a corresponding emergency plan mechanism is immediately triggered.
5) And repeating the steps circularly to provide continuous abnormal signal detection and identification service for the user.
It should be understood that, although the steps in the flowchart of fig. 1 are shown in order as indicated by the arrows, the steps are not necessarily performed in order as indicated by the arrows. The steps are not performed in the exact order shown and described, and may be performed in other orders, unless explicitly stated otherwise. Moreover, at least a portion of the steps in fig. 1 may include multiple sub-steps or multiple stages that are not necessarily performed at the same time, but may be performed at different times, and the order of performance of the sub-steps or stages is not necessarily sequential, but may be performed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.
In another embodiment, an abnormal I/Q signal recognition system based on deep learning is characterized in that the abnormal I/Q signal recognition system based on deep learning comprises:
the signal acquisition module is used for acquiring an original I/Q signal to obtain I (t) and Q (t) signal components, wherein t is a time variable;
a signal processing module, configured to pre-process the original I/Q signal to form a training data set, and specifically perform the following operations:
respectively carrying out filtering operation on the I (t) signal component and the Q (t) signal component by adopting a one-dimensional low-pass filter;
respectively adopting sliding windows to divide the filtered I (t) signal components and the filtered Q (t) signal components to obtain signal samples with the same lengths of the I (t) signal components and the Q (t) signal components;
correspondingly splicing the I (t) signal components and the Q (t) signal components section by section to form a data sequence comprising a plurality of sections of complex signal samples, performing time-frequency analysis on each section of complex signal sample to obtain a two-dimensional time-frequency diagram, and converting the two-dimensional time-frequency diagram into one-dimensional data;
splitting a complex structure in the one-dimensional data obtained after conversion to obtain two paths of frequency domain representation data, respectively calculating a data sequence comprising a plurality of sections of complex signal samples and a mode of each complex in the one-dimensional data obtained after conversion, and then obtaining two paths of mode data;
correspondingly splicing the two paths of signal components I (t) and Q (t), the two paths of frequency domain representation data and the two paths of analog data obtained by segmentation section by section according to the division of a sliding window to obtain six paths of signal data so as to construct a training data set;
the model construction module is used for constructing a signal identification model, the signal identification model comprises a convolutional neural network structure, a cyclic neural network structure and a fully-connected neural network structure, the convolutional neural network structure is used for extracting the spatial characteristics of input data, the cyclic neural network structure extracts time sequence characteristics on the basis of the spatial characteristics to obtain time-space domain characteristics, and the fully-connected neural network structure outputs an I/Q signal identification result on the basis of the time-space domain characteristics;
a model training module for training the signal recognition model based on the training data set;
and the signal identification module is used for outputting an I/Q signal identification result to be identified by utilizing the I/Q signal trained to the optimal signal identification model, wherein the I/Q signal identification result comprises a category and a confidence coefficient thereof, and the category is a normal signal or an abnormal signal.
For specific definition of the abnormal I/Q signal identification system based on deep learning, reference may be made to the above definition of the abnormal I/Q signal identification method based on deep learning, and details are not repeated here.
In another embodiment, the convolutional neural network structure comprises a plurality of composite network layers connected in sequence, the composite network layers comprising convolutional layers and max pooling layers connected in sequence from a data input side to a data output layer.
In another embodiment, the convolutional neural network structure further comprises an inclusion structure comprising four branches connected on the last composite network layer, and the outputs of all the branches are superimposed as the spatial features of the convolutional neural network structure output;
the first branch uses a layer of 1 × 1 convolution; the second branch comprises two convolution layers, the sizes of the corresponding convolution kernels are 1 multiplied by 1 and 2 multiplied by 1 respectively; the third branch adopts two convolution layers with convolution kernel sizes of 1 multiplied by 1 and 4 multiplied by 1 respectively; the fourth branch is preceded by an average pooling operation with a window size of 2, followed by a convolution layer with a convolution kernel size of 1 x 1.
In another embodiment, the present application further provides a computer device, namely, an abnormal I/Q signal recognition system based on deep learning, including a memory and a processor, where the memory stores a computer program, and the processor implements the steps of the abnormal I/Q signal recognition method based on deep learning when executing the computer program.
The computer device may be a terminal whose internal structure may include a processor, a memory, a network interface, a display screen, and an input device connected through a system bus. Wherein the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises a nonvolatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of an operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used for communicating with an external terminal through a network connection. The computer program is executed by a processor to implement the above-described abnormal I/Q signal identification method based on deep learning. The display screen of the computer equipment can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer equipment can be a touch layer covered on the display screen, a key, a track ball or a touch pad arranged on the shell of the computer equipment, an external keyboard, a touch pad or a mouse and the like.
The technical features of the embodiments described above may be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the embodiments described above are not described, but should be considered as being within the scope of the present specification as long as there is no contradiction between the combinations of the technical features.
The above-mentioned embodiments only express several embodiments of the present invention, and the description thereof is more specific and detailed, but not construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the inventive concept, which falls within the scope of the present invention. Therefore, the protection scope of the present patent shall be subject to the appended claims.
Claims (7)
1. An abnormal I/Q signal identification method based on deep learning is characterized in that the abnormal I/Q signal identification method based on deep learning comprises the following steps:
step 1, obtaining an original I/Q signal to obtain I (t) and Q (t) signal components, wherein t is a time variable;
step 2, preprocessing the original I/Q signal to form a training data set, comprising:
step 2.1, respectively carrying out filtering operation on the I (t) signal component and the Q (t) signal component by adopting a one-dimensional low-pass filter;
step 2.2, respectively adopting sliding windows to segment the filtered I (t) signal components and the filtered Q (t) signal components, and obtaining signal samples with the same lengths of the I (t) signal components and the Q (t) signal components by segmentation;
step 2.3, correspondingly splicing the I (t) signal components and the Q (t) signal components section by section to form a data sequence comprising a plurality of sections of complex signal samples, carrying out time-frequency analysis on each section of complex signal sample to obtain a two-dimensional time-frequency graph, and converting the two-dimensional time-frequency graph into one-dimensional data;
step 2.4, splitting the complex structure in the one-dimensional data obtained after conversion to obtain two paths of frequency domain characterization data, respectively calculating the original data sequence containing multiple sections of complex signal samples and the mode of each complex number in the one-dimensional data obtained after conversion, and then obtaining two paths of mode data;
step 2.5, correspondingly splicing the two paths of signal components I (t) and Q (t), the two paths of frequency domain representation data and the two paths of analog data which are obtained by segmentation section by section according to the division of a sliding window to obtain six paths of signal data so as to construct a training data set;
step 3, constructing a signal identification model, wherein the signal identification model comprises a convolutional neural network structure, a cyclic neural network structure and a fully-connected neural network structure, the convolutional neural network structure is used for extracting the spatial characteristics of input data, the cyclic neural network structure extracts time sequence characteristics on the basis of the spatial characteristics to obtain time-space domain characteristics, and the fully-connected neural network structure outputs an I/Q signal identification result on the basis of the time-space domain characteristics;
step 4, training the signal recognition model based on the training data set;
and 5, outputting an I/Q signal identification result to be identified by utilizing the I/Q signal to be identified trained to the optimal signal identification model, wherein the I/Q signal identification result comprises a category and a confidence coefficient thereof, and the category is a normal signal or an abnormal signal.
2. The deep learning-based abnormal I/Q signal identification method according to claim 1, wherein the convolutional neural network structure comprises a plurality of layers of composite network layers connected in sequence, and the composite network layers comprise a convolutional layer and a max pooling layer connected in sequence from a data input side to a data output layer.
3. The deep learning-based abnormal I/Q signal identification method according to claim 2, wherein the convolutional neural network structure further comprises an inclusion structure, the inclusion structure comprises four branches connected to the last composite network layer, and the outputs of all the branches are superposed to be the spatial features output by the convolutional neural network structure;
the first branch uses a layer of 1 × 1 convolution; the second branch comprises two convolution layers, the sizes of the corresponding convolution kernels are 1 multiplied by 1 and 2 multiplied by 1 respectively; the third branch adopts two convolution layers with convolution kernel sizes of 1 multiplied by 1 and 4 multiplied by 1 respectively; the fourth branch is preceded by an average pooling operation with a window size of 2, followed by a convolution layer with a convolution kernel size of 1 x 1.
4. An abnormal I/Q signal identification system based on deep learning, which is characterized by comprising the following components:
the signal acquisition module is used for acquiring an original I/Q signal to obtain I (t) and Q (t) signal components, wherein t is a time variable;
a signal processing module, configured to pre-process the original I/Q signal to form a training data set, and specifically perform the following operations:
respectively carrying out filtering operation on the I (t) signal component and the Q (t) signal component by adopting a one-dimensional low-pass filter;
respectively adopting sliding windows to divide the filtered I (t) signal components and the filtered Q (t) signal components to obtain signal samples with the same lengths of the I (t) signal components and the Q (t) signal components;
correspondingly splicing the I (t) signal components and the Q (t) signal components section by section to form a data sequence comprising a plurality of sections of complex signal samples, performing time-frequency analysis on each section of complex signal sample to obtain a two-dimensional time-frequency diagram, and converting the two-dimensional time-frequency diagram into one-dimensional data;
splitting a complex structure in the one-dimensional data obtained after conversion to obtain two paths of frequency domain representation data, respectively calculating a data sequence comprising a plurality of sections of complex signal samples and a mode of each complex in the one-dimensional data obtained after conversion, and then obtaining two paths of mode data;
correspondingly splicing the two paths of signal components I (t) and Q (t), the two paths of frequency domain representation data and the two paths of analog data obtained by segmentation section by section according to the division of a sliding window to obtain six paths of signal data so as to construct a training data set;
the model construction module is used for constructing a signal identification model, the signal identification model comprises a convolutional neural network structure, a cyclic neural network structure and a fully-connected neural network structure, the convolutional neural network structure is used for extracting the spatial characteristics of input data, the cyclic neural network structure extracts time sequence characteristics on the basis of the spatial characteristics to obtain time-space domain characteristics, and the fully-connected neural network structure outputs an I/Q signal identification result on the basis of the time-space domain characteristics;
a model training module for training the signal recognition model based on the training data set;
and the signal identification module is used for outputting an I/Q signal identification result to be identified by utilizing the I/Q signal trained to the optimal signal identification model, wherein the I/Q signal identification result comprises a category and a confidence coefficient thereof, and the category is a normal signal or an abnormal signal.
5. The deep learning based abnormal I/Q signal identification system according to claim 4, wherein the convolutional neural network structure comprises a plurality of layers of composite network layers connected in sequence, the composite network layers comprising convolutional layers and maximum pooling layers connected in sequence from a data input side to a data output side.
6. The deep learning based anomaly I/Q signal identification system according to claim 5, wherein said convolutional neural network structure further comprises an inclusion structure comprising four branches connected on the last composite network layer, the outputs of all branches being superimposed as the spatial features of the convolutional neural network structure output;
the first branch uses a layer of 1 × 1 convolution; the second branch comprises two convolution layers, the sizes of the corresponding convolution kernels are 1 multiplied by 1 and 2 multiplied by 1 respectively; the third branch adopts two convolution layers with convolution kernel sizes of 1 multiplied by 1 and 4 multiplied by 1 respectively; the fourth branch is preceded by an average pooling operation with a window size of 2, followed by a convolution layer with a convolution kernel size of 1 x 1.
7. An abnormal I/Q signal recognition system based on deep learning, comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method according to any one of claims 1 to 3 when executing the computer program.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011399508.5A CN112566174B (en) | 2020-12-02 | 2020-12-02 | Abnormal I/Q signal identification method and system based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011399508.5A CN112566174B (en) | 2020-12-02 | 2020-12-02 | Abnormal I/Q signal identification method and system based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112566174A true CN112566174A (en) | 2021-03-26 |
| CN112566174B CN112566174B (en) | 2022-05-03 |
Family
ID=75047932
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011399508.5A Active CN112566174B (en) | 2020-12-02 | 2020-12-02 | Abnormal I/Q signal identification method and system based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112566174B (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113409213A (en) * | 2021-06-22 | 2021-09-17 | 中铁工程装备集团有限公司 | Plunger pump fault signal time-frequency graph noise reduction enhancement method and system |
| CN113466852A (en) * | 2021-06-08 | 2021-10-01 | 江苏科技大学 | Millimeter wave radar dynamic gesture recognition method applied to random interference scene |
| CN113780106A (en) * | 2021-08-24 | 2021-12-10 | 电信科学技术第五研究所有限公司 | Deep learning signal detection method based on radio waveform data input |
| CN115277324A (en) * | 2022-07-25 | 2022-11-01 | 电信科学技术第五研究所有限公司 | FSK signal identification method based on convolutional neural network |
| CN115296758A (en) * | 2022-07-07 | 2022-11-04 | 中山大学 | Method, system, computer device and storage medium for identifying interference signal |
| CN116633809A (en) * | 2023-06-26 | 2023-08-22 | 中国信息通信研究院 | Detection method and system based on artificial intelligence |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110166387A (en) * | 2019-05-22 | 2019-08-23 | 中国电力科学研究院有限公司 | A kind of method and system based on convolutional neural networks identification signal modulation system |
| CN110856178A (en) * | 2019-11-05 | 2020-02-28 | 天津大学 | Behavior identification method based on wireless network physical layer IQ signal |
| CN111428817A (en) * | 2020-04-22 | 2020-07-17 | 浙江工业大学 | Defense method for resisting attack by radio signal identification |
| US20200327397A1 (en) * | 2019-04-12 | 2020-10-15 | Motorola Solutions, Inc. | Systems and methods for modulation classification of baseband signals using multiple data representations of signal samples |
| CN111950423A (en) * | 2020-08-06 | 2020-11-17 | 中国电子科技集团公司第五十二研究所 | Real-time multi-scale dense target detection method based on deep learning |
-
2020
- 2020-12-02 CN CN202011399508.5A patent/CN112566174B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20200327397A1 (en) * | 2019-04-12 | 2020-10-15 | Motorola Solutions, Inc. | Systems and methods for modulation classification of baseband signals using multiple data representations of signal samples |
| CN110166387A (en) * | 2019-05-22 | 2019-08-23 | 中国电力科学研究院有限公司 | A kind of method and system based on convolutional neural networks identification signal modulation system |
| CN110856178A (en) * | 2019-11-05 | 2020-02-28 | 天津大学 | Behavior identification method based on wireless network physical layer IQ signal |
| CN111428817A (en) * | 2020-04-22 | 2020-07-17 | 浙江工业大学 | Defense method for resisting attack by radio signal identification |
| CN111950423A (en) * | 2020-08-06 | 2020-11-17 | 中国电子科技集团公司第五十二研究所 | Real-time multi-scale dense target detection method based on deep learning |
Non-Patent Citations (3)
| Title |
|---|
| FAISAL NADEEM KHAN ET.AL: "Modulation Format Identification in Coherent Receivers Using Deep Machine Learning", 《IEEE PHOTONICS TECHNOLOGY LETTERS》 * |
| 孟磊等: "基于机器学习的雷达辐射源识别方法综述", 《兵器装备工程学报》 * |
| 田昊旻: "基于深度学习的无线信号的识别技术研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113466852A (en) * | 2021-06-08 | 2021-10-01 | 江苏科技大学 | Millimeter wave radar dynamic gesture recognition method applied to random interference scene |
| CN113466852B (en) * | 2021-06-08 | 2023-11-24 | 江苏科技大学 | Millimeter wave radar dynamic gesture recognition method applied to random interference scene |
| CN113409213A (en) * | 2021-06-22 | 2021-09-17 | 中铁工程装备集团有限公司 | Plunger pump fault signal time-frequency graph noise reduction enhancement method and system |
| CN113409213B (en) * | 2021-06-22 | 2023-11-14 | 中铁工程装备集团有限公司 | Method and system for enhancing noise reduction of time-frequency diagram of fault signal of plunger pump |
| CN113780106A (en) * | 2021-08-24 | 2021-12-10 | 电信科学技术第五研究所有限公司 | Deep learning signal detection method based on radio waveform data input |
| CN113780106B (en) * | 2021-08-24 | 2024-02-27 | 电信科学技术第五研究所有限公司 | Deep learning signal detection method based on radio waveform data input |
| CN115296758A (en) * | 2022-07-07 | 2022-11-04 | 中山大学 | Method, system, computer device and storage medium for identifying interference signal |
| CN115296758B (en) * | 2022-07-07 | 2023-08-11 | 中山大学 | Method, system, computer equipment and storage medium for identifying interference signals |
| CN115277324A (en) * | 2022-07-25 | 2022-11-01 | 电信科学技术第五研究所有限公司 | FSK signal identification method based on convolutional neural network |
| CN115277324B (en) * | 2022-07-25 | 2023-11-10 | 电信科学技术第五研究所有限公司 | FSK signal identification method based on convolutional neural network |
| CN116633809A (en) * | 2023-06-26 | 2023-08-22 | 中国信息通信研究院 | Detection method and system based on artificial intelligence |
| CN116633809B (en) * | 2023-06-26 | 2024-01-23 | 中国信息通信研究院 | Detection method and system based on artificial intelligence |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112566174B (en) | 2022-05-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112566174B (en) | Abnormal I/Q signal identification method and system based on deep learning | |
| Ebert-Uphoff et al. | Evaluation, tuning, and interpretation of neural networks for working with images in meteorological applications | |
| CN108509911B (en) | Interference signal identification method based on convolutional neural network | |
| CN110926782B (en) | Circuit breaker fault type judgment method and device, electronic equipment and storage medium | |
| Lin et al. | Learning of time-frequency attention mechanism for automatic modulation recognition | |
| CN109766992B (en) | Industrial control abnormity detection and attack classification method based on deep learning | |
| CN114553648B (en) | Wireless communication modulation mode identification method based on space-time diagram convolutional neural network | |
| CN110569752A (en) | convolutional neural network-based radar signal category determination method | |
| CN111461037B (en) | End-to-end gesture recognition method based on FMCW radar | |
| Ke et al. | Blind detection techniques for non-cooperative communication signals based on deep learning | |
| CN112087442B (en) | Time sequence related network intrusion detection method based on attention mechanism | |
| CN114998695B (en) | Method and system for improving image recognition speed | |
| CN114881092A (en) | Signal modulation identification method based on feature fusion | |
| CN112598303A (en) | Non-invasive load decomposition method based on combination of 1D convolutional neural network and LSTM | |
| Kamal et al. | Generative adversarial learning for improved data efficiency in underwater target classification | |
| CN117079005A (en) | Optical cable fault monitoring method, system, device and readable storage medium | |
| CN110908365A (en) | Unmanned aerial vehicle sensor fault diagnosis method and system and readable storage medium | |
| Hou et al. | The recognition of multi-components signals based on semantic segmentation | |
| CN114912545A (en) | Power load classification method based on optimized VMD algorithm and DBN network | |
| Kezunovic et al. | Neural network applications to real-time and off-line fault analysis | |
| CN110516566B (en) | Filtering method and device based on convolutional layer | |
| Liu et al. | Automatic modulation classification with deep learning‐based frequency selection filters | |
| CN112766465A (en) | Training method of neural network for intelligent rotation performance detection | |
| Wang et al. | MIANet: Multi-level temporal information aggregation in mixed-periodicity time series forecasting tasks | |
| Junpeng et al. | MTCNN: a deep neural network for recognizing stochastic phase shifted time series data |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |