CN115964670A - Frequency spectrum anomaly detection method - Google Patents
Frequency spectrum anomaly detection method Download PDFInfo
- Publication number
- CN115964670A CN115964670A CN202211628547.7A CN202211628547A CN115964670A CN 115964670 A CN115964670 A CN 115964670A CN 202211628547 A CN202211628547 A CN 202211628547A CN 115964670 A CN115964670 A CN 115964670A
- Authority
- CN
- China
- Prior art keywords
- data
- detection module
- parameter
- spectrum
- updating
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D30/00—Reducing energy consumption in communication networks
- Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a frequency spectrum anomaly detection method, which comprises the following steps: carrying out short-time Fourier transform on the acquired radio signal data to obtain short-time Fourier transform data; carrying out classification extraction processing on the short-time Fourier transform data to obtain channel data; initializing a first detection module and a second detection module to obtain initial parameter values of the first detection module and the second detection module; updating parameters of the first detection module and the second detection module to obtain a first optimized detection module and a second optimized detection module; and establishing an electromagnetic spectrum anomaly detection model by using a second optimization detection module, and realizing anomaly detection by using the electromagnetic spectrum anomaly detection model. The method of the invention fully utilizes the statistical information of the electromagnetic spectrum data, and solves the problems of limited detection precision and limitation in practical application caused by acquiring the spectrum abnormal sample by the traditional method.
Description
Technical Field
The invention relates to the field of electromagnetic spectrum monitoring, in particular to a spectrum anomaly detection method.
Background
At present, with the rapid development of electromagnetic technology and wireless communication technology, the forms of radio signals are in a diversified trend, the demand of human beings on radio spectrum resources is more and more intense, and radio spectrum is not inexhaustible. The increasing contradiction between the application requirements and the limited spectrum resources increases the difficulty for the supervision of the electromagnetic spectrum and the security guarantee of the electromagnetic space. In recent years, personal use conditions of amateur radio stations, unmanned aerial vehicles and wireless communication equipment are more and more common, and due to the lack of knowledge on electromagnetic space safety, cases of illegal invasion of wireless communication frequency bands sometimes occur, and even civil aviation radio is interfered, so that safety accidents occur.
The existing electromagnetic spectrum anomaly detection method mainly comprises two main types: one is to judge whether the spectrum state is abnormal or not by analyzing the change of the spectrum characteristic parameters by using a spectrum analysis method; the other type is that electromagnetic spectrum data is subjected to two-classification processing by using a supervised machine learning algorithm so as to judge whether the spectrum is abnormal or not, such as: support vector machine, naive Bayes classification method, etc.
The existing electromagnetic spectrum anomaly detection method mainly has two problems in the practical application process. On one hand, in an actual radio propagation frequency band, radio signals are in a normal working state most of time, and the probability of abnormity is relatively low; moreover, due to the fact that the composition and the working process of a radio system are complex, and various reasons such as internal faults of the system and external interference signals can cause the frequency spectrum signals of the detection end of the system to be abnormal, the frequency spectrum abnormal sample is difficult to obtain, and the supervised detection method is difficult to fully master experience knowledge, so that the detection precision is influenced; on the other hand, the existing method is mainly based on specific tasks and scenes, and the type of the electromagnetic spectrum abnormal signal is defined by artificial standards, so that the method has limitations in practical application.
Disclosure of Invention
The invention discloses a method and a device for detecting spectrum abnormity, aiming at the problems of limited detection precision and limitation in practical application caused by the fact that a spectrum abnormity sample is difficult to obtain in the existing electromagnetic spectrum abnormity monitoring method.
The invention discloses a frequency spectrum anomaly detection method, which comprises the following steps:
s1, acquiring radio signal data; using the acquired radio signal data as data to be processed; carrying out short-time Fourier transform on data to be processed to obtain short-time Fourier transform data;
s2, classifying and extracting the short-time Fourier transform data according to the class composition information and the parameter information of the short-time Fourier transform data to obtain channel data; constructing first picture data by using the channel data; setting a forward tag for the first picture data to obtain first tag data;
the category composition information of the short-time Fourier transform data comprises real data information and imaginary data information, and the parameter information of the short-time Fourier transform data comprises power data information and phase data information;
s3, establishing a first detection module and a second detection module; respectively initializing the first detection module and the second detection module to respectively obtain parameter initial values of the first detection module and the second detection module;
respectively updating parameters of the first detection module and the second detection module by using the first picture data and the first label data to respectively obtain the first detection module with the updated parameters and the second detection module with the updated parameters;
establishing an electromagnetic spectrum anomaly detection model by using the second detection module after the parameters are updated;
s4, processing the channel data by using the electromagnetic spectrum anomaly detection model to obtain output data of the electromagnetic spectrum anomaly detection model; and judging and counting the output data to obtain the occurrence probability of the electromagnetic spectrum abnormality.
The window function of the short-time Fourier transform is a hamming window with the length of W; the number of overlapped samples of the short-time Fourier transform is 2/3 xW, and the sampling frequency is Fs; the length of the data to be processed is L, and the sampling frequency is Fs.
The step S2 includes:
classifying, extracting and processing the short-time Fourier transform data according to the class composition information and the parameter information of the short-time Fourier transform data to obtain real part data, imaginary part data, power data and phase data of the short-time Fourier transform data; integrating real part data, imaginary part data and power data or phase data of the short-time Fourier transform data to obtain channel data; constructing first picture data by using the channel data; and setting the label of the first picture data as 1 to obtain first label data.
The step S3 includes:
s31, respectively establishing a first detection module and a second detection module by using a deep neural network model; generating first random matrix data, and respectively training a first detection module and a second detection module by using the first random matrix data to complete the initialization processing of the first detection module and the second detection module so as to obtain parameter initial values of the first detection module and the second detection module; setting the cycle number as an initial value;
s32, generating second random matrix data; inputting the second random matrix data into the first detection module to obtain an output matrix of the first detection module, and taking the output matrix of the first detection module as second picture data; negative label setting is carried out on the second picture data to obtain second label data;
s33, inputting the first picture data and the first label data into a second detection module to obtain first judgment result data; inputting the second picture data and the second label data into a second detection module to obtain second judgment result data; updating parameters of the second detection module by using a difference value between the first judgment result data and the second judgment result data to obtain a second detection module with updated parameters;
s34, setting a forward tag for the second picture data to obtain third tag data; inputting the second picture data and the third label data into the second detection module after the parameter is updated, and obtaining third judgment result data; updating parameters of the first detection module by using a difference value between the label data of the third discrimination result data and the third label data to obtain a parameter-updated first detection module; increasing the number of cycles by 1;
s35, judging whether the updating stop condition is met, if the updating stop condition is not met, taking the first detection module with the updated parameters as a first detection module and taking the second detection module with the updated parameters as a second detection module, returning to the step S32, and continuously updating the parameters of the first detection module and the second detection module; and if the update stop condition is met, using the second detection module with the updated parameters as an electromagnetic spectrum anomaly detection model.
The parameter updating is performed on the second detection module by using the difference value between the first judgment result data and the second judgment result data to obtain the second detection module after the parameter updating, and the method includes the following steps:
inputting the first discrimination result data and the second discrimination result data into a cost function, and processing the first discrimination result data and the second discrimination result data by using the cost function to obtain an error cost function; calculating a partial derivative of the error cost function to the parameter of the second detection module by using the error cost function; performing conversion processing on the partial derivative to obtain a parameter updating value; and updating the parameters of the second detection module by using the parameter updating value to obtain the second detection module with updated parameters.
Performing conversion processing on the partial derivative to obtain a parameter update value; updating the parameters of the second detection module by using the parameter update value to obtain the second detection module with updated parameters, comprising:
the partial derivative is expressed as
k i The ith parameter of the second detection module is represented, the E represents an error cost function, and a calculation formula for performing conversion processing on the partial derivative is as follows:
wherein, Δ k i Updating a parameter of the ith parameter of the second detection module, wherein pi is a circumferential rate constant value;
and multiplying the parameter updating value by the corresponding parameter of the second detection module obtained after the last round of parameter updating to obtain the updated parameter of the second detection module.
The judging whether the update stop condition is satisfied includes:
judging whether the cycle number exceeds a set threshold or not, or performing difference judgment on the tag data of the third judgment result data and the third tag data;
judging whether the cycle number exceeds a set threshold value, when the cycle number exceeds the set threshold value, determining that the updating stop condition is met, and when the cycle number does not exceed the set threshold value, determining that the updating stop condition is not met;
the performing difference judgment on the tag data of the third judgment result data and the third tag data includes:
performing autoregressive-moving average model modeling processing on the tag data of the third discrimination result data and the third tag data respectively to obtain a first autoregressive-moving average model and a second autoregressive-moving average model; extracting coefficient vectors of the two autoregressive-moving average models, and calculating to obtain a cross-correlation matrix of the coefficient vectors; performing eigenvalue decomposition operation on the cross-correlation matrix to obtain the maximum eigenvalue of the cross-correlation matrix; and judging the maximum characteristic value, and when the maximum characteristic value is greater than the difference judgment threshold value, determining that the updating stop condition is not met, and when the maximum characteristic value is less than or equal to the difference judgment threshold value, determining that the updating stop condition is met.
The step S4 includes:
s41, splitting channel data according to the dimensionality of the picture pixel matrix data to obtain a plurality of single-picture data;
s42, inputting each single picture data into the electromagnetic spectrum abnormality detection model to obtain output data of the electromagnetic spectrum abnormality detection model of the single picture data;
s43, judging and processing output data of the electromagnetic spectrum abnormality detection model of the single-sheet picture data to obtain a spectrum abnormality judgment result of the single-sheet picture data;
s44, counting the frequency spectrum abnormality judgment results of all the single picture data and the number of the single picture data to obtain the occurrence probability of the electromagnetic frequency spectrum abnormality.
The judging and processing of the output data of the electromagnetic spectrum abnormality detection model of the single picture data to obtain the spectrum abnormality judging result of the single picture data comprises the following steps:
the output data of the electromagnetic spectrum anomaly detection model of the single picture data is expressed as outputs, the spectrum anomaly judgment result of the single picture data is expressed as labels, and the judgment processing process is expressed as follows:
if labels is equal to 1, the frequency spectrum corresponding to the picture data is considered to be in a normal working state, and if labels is equal to 0, the frequency spectrum corresponding to the picture data is considered to be abnormal.
The method for counting the frequency spectrum abnormity discrimination results of all single picture data and the number of the single picture data to obtain the occurrence probability of the electromagnetic frequency spectrum abnormity comprises the following steps:
summing the frequency spectrum abnormality judgment results of all single picture data to obtain the occurrence frequency of the abnormality judgment result; and performing division operation by using the occurrence frequency of the abnormity discrimination result and the number of single picture data to obtain the occurrence probability of the electromagnetic spectrum abnormity.
The beneficial effects of the invention are as follows:
1. the method provided by the invention can input the picture formed after short-time Fourier transform into the first detection module to learn the statistical distribution rule when the frequency spectrum works normally, and then establish and judge the second detection module. The method provided by the invention fully utilizes the statistical information of the electromagnetic spectrum data, simplifies the discrimination process, and solves the problems of limited detection precision and limitation in practical application caused by acquiring a spectrum abnormal sample by a traditional method.
2. The invention realizes the discrimination of the spectrum abnormity through the change of the statistical rule of the received electromagnetic spectrum data without manual supervision, thereby avoiding the misdiscrimination caused by manual subjective assumption and improving the accuracy of the monitoring and identification of the spectrum abnormity.
3. The invention obtains more information beneficial to analyzing the frequency spectrum statistical characteristics, such as power information and phase information, through data processing, the information is in nonlinear relation with the original data and the real part and imaginary part data, the sensitivity and the information range of the network are increased, in addition, the selection process of interested and uninteresting regions of a user is increased, a template filtering form is provided, and the analysis is more efficient and flexible.
Drawings
Fig. 1 is a flowchart of an implementation of the spectrum anomaly detection method of the present invention.
Detailed Description
For a better understanding of the present disclosure, two examples are given herein.
Fig. 1 is a flowchart of an implementation of the spectrum anomaly detection method of the present invention.
The first embodiment is as follows:
the invention discloses a frequency spectrum anomaly detection method, which comprises the following steps:
s1, acquiring radio signal data; using the acquired radio signal data as data to be processed; carrying out short-time Fourier transform on data to be processed to obtain short-time Fourier transform data;
s2, classifying and extracting the short-time Fourier transform data according to the class composition information and the parameter information of the short-time Fourier transform data to obtain channel data; constructing first picture data by utilizing the channel data; setting a forward tag for the first picture data to obtain first tag data;
the class composition information of the short-time Fourier transform data comprises real data information and imaginary data information, and the parameter information of the short-time Fourier transform data comprises power data information and phase data information;
the step S2 includes:
respectively taking real part data, imaginary part data and power data or phase data information of the short-time Fourier transform data as channel data to form standard three-channel data, constructing N pictures with dimensions W3 by utilizing the three-channel data, wherein W is the height and the width of the pictures, and all picture labels are set to be 1;
s3, establishing a first detection module and a second detection module; respectively initializing the first detection module and the second detection module to respectively obtain parameter initial values of the first detection module and the second detection module;
respectively updating parameters of the first detection module and the second detection module by using the first picture data and the first label data to respectively obtain a first optimized detection module with updated parameters and a second optimized detection module with updated parameters;
establishing an electromagnetic spectrum anomaly detection model by using the second optimized detection module after the parameters are updated;
s4, processing the channel data by using the electromagnetic spectrum anomaly detection model to obtain output data of the electromagnetic spectrum anomaly detection model; and distinguishing and counting the output data to obtain the occurrence probability of the electromagnetic spectrum abnormality.
The window function of the short-time Fourier transform is a hamming window with the length of W; the number of overlapped samples of the short-time Fourier transform is 2/3 x W, and the sampling frequency is Fs; the length of the data to be processed is L, and the sampling frequency is Fs. And after short-time Fourier transform is carried out, first short-time Fourier transform data are obtained, and template filtering processing is carried out on the first short-time Fourier transform data to obtain finally output short-time Fourier transform data. The template filtering process is implemented by the following formula:
S＝T·C
wherein, S is a template filtering output data sequence (complex number form or real number form), i.e. the finally output short-time fourier transform data, T is a template filtering input data sequence, i.e. the first short-time fourier transform data, representing the sequence dot product, C is a template filtering function, which has the same output length as the short-time fourier transform, the value of each point is set to 1 or 0 by the user, the value of the corresponding point of the frequency set by the user is 1, otherwise, it is 0. After the template filtering processing, the length of the data sequence is kept unchanged. The template filtering process processes input data according to a continuous or discontinuous frequency domain set by a user.
The step S2 includes:
classifying, extracting and processing the short-time Fourier transform data according to the class composition information and the parameter information of the short-time Fourier transform data to obtain real part data, imaginary part data and power data or phase data of the short-time Fourier transform data; integrating real part data, imaginary part data and power data or phase data of the short-time Fourier transform data to obtain channel data; constructing first picture data by using the channel data; and setting the label of the first picture data as 1 to obtain first label data.
The step S3 includes:
s31, respectively establishing a first detection module and a second detection module by using a deep neural network model; generating first random matrix data, and training a first detection module and a second detection module respectively by using the first random matrix data to complete initialization processing of the first detection module and the second detection module so as to obtain parameter initial values of the first detection module and the second detection module; setting the cycle number as an initial value;
s32, generating second random matrix data; inputting the second random matrix data into the first detection module to obtain an output matrix of the first detection module, and taking the output matrix of the first detection module as second picture data; negative label setting is carried out on the second picture data to obtain second label data;
s33, inputting the first picture data and the first label data into a second detection module to obtain first judgment result data; inputting the second picture data and the second label data into a second detection module to obtain second judgment result data; updating parameters of the second detection module by using a difference value between the first judgment result data and the second judgment result data to obtain a second detection module with updated parameters;
s34, setting a forward tag for the second picture data to obtain third tag data; inputting the second picture data and the third label data into the second detection module after the parameter is updated, and obtaining third judgment result data; updating parameters of the first detection module by using a difference value between the label data of the third discrimination result data and the third label data to obtain a parameter-updated first detection module; increasing the number of cycles by 1;
s35, judging whether the updating stop condition is met, if the updating stop condition is not met, taking the first detection module with the updated parameters as a first detection module and taking the second detection module with the updated parameters as a second detection module, returning to the step S32, and continuously updating the parameters of the first detection module and the second detection module; and if the update stop condition is met, the second detection module with the updated parameters is used as an electromagnetic spectrum anomaly detection model.
The parameter updating is performed on the second detection module by using the difference value between the first judgment result data and the second judgment result data to obtain the second detection module after the parameter updating, and the method includes the following steps:
inputting the first discrimination result data and the second discrimination result data into a cost function, and processing the first discrimination result data and the second discrimination result data by using the cost function to obtain an error cost function; calculating a partial derivative of the error cost function to the parameter of the second detection module by using the error cost function; performing conversion processing on the partial derivative to obtain a parameter updating value; and updating the parameters of the second detection module by using the parameter updating value to obtain the second detection module with updated parameters.
Performing conversion processing on the partial derivative to obtain a parameter update value; updating the parameters of the second detection module by using the parameter update value to obtain the second detection module with updated parameters, and the method comprises the following steps:
the partial derivative is expressed as
k i The ith parameter of the second detection module is represented, E represents an error cost function, and a calculation formula for converting the partial derivative is as follows:
wherein, Δ k i Updating a parameter value for the ith parameter of the second detection module;
and multiplying the parameter updating value by the corresponding parameter of the second detection module obtained after the last round of parameter updating to obtain the updated parameter of the second detection module.
The judging whether the update stop condition is satisfied includes:
judging whether the cycle number exceeds a set threshold or not, or performing difference judgment on the tag data of the third judgment result data and the third tag data;
judging whether the cycle number exceeds a set threshold value, when the cycle number exceeds the set threshold value, considering that the updating stop condition is met, and when the cycle number does not exceed the set threshold value, considering that the updating stop condition is not met;
the performing difference judgment on the tag data of the third judgment result data and the third tag data includes:
performing autoregressive-moving average model modeling processing on the tag data of the third discrimination result data and the third tag data respectively to obtain a first autoregressive-moving average model and a second autoregressive-moving average model; extracting to obtain coefficient vectors of two autoregressive-moving average models, and calculating to obtain a cross-correlation matrix of the coefficient vectors; performing eigenvalue decomposition operation on the cross-correlation matrix to obtain the maximum eigenvalue of the cross-correlation matrix; and judging the maximum characteristic value, and when the maximum characteristic value is greater than the difference judgment threshold value, determining that the updating stop condition is not met, and when the maximum characteristic value is less than or equal to the difference judgment threshold value, determining that the updating stop condition is met.
By adopting a difference discrimination method, the overfitting problem in the training process of the detection module can be effectively solved, and the accuracy of the detection module in realizing the frequency spectrum anomaly detection is improved.
The step S4 includes:
s41, splitting channel data according to the dimensionality of the picture pixel matrix data to obtain a plurality of single-picture data;
s42, inputting each single picture data into the electromagnetic spectrum abnormality detection model to obtain output data of the electromagnetic spectrum abnormality detection model of the single picture data;
s43, judging and processing output data of the electromagnetic spectrum abnormality detection model of the single-sheet picture data to obtain a spectrum abnormality judgment result of the single-sheet picture data;
s44, counting the frequency spectrum abnormality judgment results of all the single picture data and the number of the single picture data to obtain the occurrence probability of the electromagnetic frequency spectrum abnormality.
The picture pixel matrix data dimension may be 64 x 3.
According to the dimension of the image pixel matrix data, the channel data is split to obtain a plurality of single image data, which can be: and splitting the channel data according to the dimension 64 × 3 to obtain a plurality of single picture data with the dimension 64 × 3.
The judging and processing of the output data of the electromagnetic spectrum abnormality detection model of the single picture data to obtain the spectrum abnormality judging result of the single picture data comprises the following steps:
the output data of the electromagnetic spectrum anomaly detection model of the single picture data is expressed as outputs, the spectrum anomaly judgment result of the single picture data is expressed as labels, and the judgment processing process is expressed as follows:
if labels is equal to 1, the frequency spectrum corresponding to the image data is considered to be in a normal working state, and if labels is equal to 0, the frequency spectrum corresponding to the image data is considered to be abnormal. Labels equals 0 when outputs = 0.5.
The method for counting the frequency spectrum abnormality judgment results of all the single picture data and the number of the single picture data to obtain the occurrence probability of the electromagnetic frequency spectrum abnormality comprises the following steps:
summing the frequency spectrum abnormality judgment results of all single picture data to obtain the occurrence frequency of the abnormality judgment result; and carrying out division operation by using the occurrence frequency of the abnormity discrimination result and the number of single picture data to obtain the occurrence probability of the electromagnetic spectrum abnormity.
The second embodiment:
the invention discloses a frequency spectrum anomaly detection method, which comprises the following steps:
s1, acquiring radio signal data; using the acquired radio signal data as data to be processed; carrying out short-time Fourier transform on data to be processed to obtain short-time Fourier transform data;
s2, classifying and extracting the short-time Fourier transform data according to the class composition information and the parameter information of the short-time Fourier transform data to obtain channel data; constructing first picture data by using the channel data; setting a forward tag for the first picture data to obtain first tag data;
the category composition information of the short-time Fourier transform data comprises real data information and imaginary data information, and the parameter information of the short-time Fourier transform data comprises power data information and phase data information;
the step S2 includes:
respectively taking real part data, imaginary part data and power data or phase data information of the short-time Fourier transform data as channel data to form standard three-channel data, constructing N pictures with dimensions W3 by utilizing the three-channel data, wherein W is the height and the width of the pictures, and all picture labels are set to be 1;
s3, establishing a first detection module and a second detection module; respectively initializing the first detection module and the second detection module to respectively obtain parameter initial values of the first detection module and the second detection module;
respectively updating parameters of the first detection module and the second detection module by using the first picture data and the first label data to respectively obtain a first optimized detection module with updated parameters and a second optimized detection module with updated parameters;
establishing an electromagnetic spectrum anomaly detection model by using the second optimized detection module after the parameters are updated;
s4, processing the channel data by using the electromagnetic spectrum anomaly detection model to obtain output data of the electromagnetic spectrum anomaly detection model; and distinguishing and counting the output data to obtain the occurrence probability of the electromagnetic spectrum abnormality.
The window function of the short-time Fourier transform is a hamming window with the length of W; the number of overlapped samples of the short-time Fourier transform is 2/3 xW, and the sampling frequency is Fs; the length of the data to be processed is L, and the sampling frequency is Fs.
And after short-time Fourier transform is carried out, first short-time Fourier transform data are obtained, and template filtering processing is carried out on the first short-time Fourier transform data to obtain finally output short-time Fourier transform data. The template filtering process is implemented by the following formula:
S＝T·C
wherein, S is a template filtering output data sequence (complex number form or real number form), i.e. the finally output short-time fourier transform data, T is a template filtering input data sequence, i.e. the first short-time fourier transform data, representing the sequence dot product, C is a template filtering function, which has the same output length as the short-time fourier transform, the value of each point is set to 1 or 0 by the user, the value of the corresponding point of the frequency set by the user is 1, otherwise, it is 0. After the template filtering processing, the length of the data sequence is kept unchanged. The template filtering process processes input data according to a continuous or discontinuous frequency domain set by a user.
The step S2 includes:
classifying, extracting and processing the short-time Fourier transform data according to the class composition information and the parameter information of the short-time Fourier transform data to obtain real part data, imaginary part data, power data and phase data of the short-time Fourier transform data; integrating real part data, imaginary part data and power data or phase data of the short-time Fourier transform data to obtain channel data; constructing first picture data by utilizing the channel data; and setting the label of the first picture data as 1 to obtain first label data.
The step S2 includes: and after the channel data are obtained, preprocessing the channel data, and constructing first picture data by using the preprocessed channel data.
The channel data is preprocessed, and a sampling vector formed by discrete sampling values of the channel data acquired within a period of time is represented as [ x ] 1 ,x 2 ,…,x N ]N is the number of discrete sampling values of channel data acquired in a period of timeCalculating to obtain a cross-correlation matrix C1 of the sampling vector, and performing eigenvalue decomposition on the cross-correlation matrix C1 to obtain:
C1＝VDV H ，
and V is an eigenvector matrix, D is an eigenvalue matrix, the diagonal elements of the matrix D are normalized and used as weight vectors, and the sampling vectors are subjected to weighted summation to obtain a smooth value of the channel data, which is used as the preprocessed channel data.
By carrying out the preprocessing operation on the channel data, the noise data and the false alarm data in the channel data can be effectively filtered, the accuracy of the training data and the training efficiency of the detection module are improved, and the establishment of a more accurate and efficient electromagnetic spectrum anomaly detection model in the subsequent steps is facilitated.
The step S2 includes: the phase data and the power data are calculated from real data and imaginary data outputted by short-time fourier transform (or outputted after windowing), the power data (linear value) is equal to the sum of the square of the real data value and the square of the imaginary data value, and the correlation processing can be operated in a point mode. The phase data is equal to arg (real/imag), where real represents the real data values and imag represents the imaginary data values, and the correlation process can be operated on as points.
The step S3 includes:
s31, respectively establishing a first detection module and a second detection module by using a deep neural network model; generating first random matrix data, and training a first detection module and a second detection module respectively by using the first random matrix data to complete initialization processing of the first detection module and the second detection module so as to obtain parameter initial values of the first detection module and the second detection module; setting the cycle number as an initial value;
s32, generating second random matrix data; inputting the second random matrix data into the first detection module to obtain an output matrix of the first detection module, and taking the output matrix of the first detection module as second picture data; performing negative label setting on the second picture data to obtain second label data;
s33, inputting the first picture data and the first label data into a second detection module to obtain first judgment result data; inputting the second picture data and the second label data into a second detection module to obtain second judgment result data; updating parameters of the second detection module by using a difference value between the first judgment result data and the second judgment result data to obtain a second detection module with updated parameters;
s34, setting a forward tag for the second picture data to obtain third tag data; inputting the second picture data and the third label data into the second detection module after the parameters are updated to obtain third judgment result data; updating parameters of the first detection module by using a difference value between the label data of the third discrimination result data and the third label data to obtain a parameter-updated first detection module; increasing the number of cycles by 1;
s35, judging whether the updating stop condition is met, if the updating stop condition is not met, taking the first detection module with the updated parameters as a first detection module and taking the second detection module with the updated parameters as a second detection module, returning to the step S32, and continuously updating the parameters of the first detection module and the second detection module; and if the update stop condition is met, the second detection module with the updated parameters is used as an electromagnetic spectrum anomaly detection model.
The first detection module comprises an input layer, a deconvolution kernel and an output layer, wherein each layer and the deconvolution kernel comprise a plurality of neural network layers;
the second detection module comprises an input layer, a convolution kernel and an output layer, wherein each layer and the convolution kernel comprise a plurality of neural network layers.
The step S3 specifically includes:
s301, initializing the first detection module and the second detection module, and cycling the number i =0.
S3011, initializing a first detection module, including:
a) Inputting a randomly generated Q matrix to generate a matrix L1, L1= ∑ Sigma i w i *Q i ，Q i Column i of the Q matrix, convert L1 to 4 x 512 dimensionsA matrix of degree, taking L1= max (L1, α × L1), α being a set constant;
b) Inputting the L1 matrix with the dimension of 4 × 512 into a first deconvolution kernel of a first detection module, wherein the first deconvolution kernel uses a 3 × 3 deconvolution matrix, the step length is moved to 2, deconvolution is performed on the L1 matrix to generate an output matrix L2 with the dimension of 8 × 256, and L2= max (L2, α × L2) is taken to obtain an output result L2;
c) Inputting the L2 high-dimensional matrix with the dimension size of 8 × 256 into a second deconvolution kernel of the first detection module, wherein the second deconvolution kernel uses a deconvolution matrix with the dimension size of 3 × 3, the step length is moved to 2, deconvolution is performed on the L2 matrix, output L3 with the dimension size of 16 × 128 is generated, L3= max (L3, α × L3) is taken, and the output result is L3;
d) Inputting the L3 high-dimensional matrix with the dimension size of 16 × 128 into a third deconvolution kernel of the first detection module, wherein the third deconvolution kernel uses a 3 × 3 deconvolution matrix, the step length is moved to 2, deconvolution is performed on the L3 matrix, an output L4 with the dimension size of 32 × 64 is generated, L4= max (L4, α × L4) is taken, and after 0.8 times of parameters are randomly selected, an output result is L4;
e) And inputting the L4 high-dimensional matrix with the dimension of 32 x 64 into a fourth deconvolution kernel of the first detection module, wherein the fourth deconvolution kernel uses a deconvolution matrix with the dimension of 3 x 3, the step length is moved to 2, deconvolution is carried out on the L4 high-dimensional matrix to generate output logits with the dimension of 64 x 3, and the output is output (the pixel points of the generated picture are in the range of [ -1,1 ]) after the logits pass through a tanh function.
S3012, initializing a second detection module, including:
a) The second detection module adjusts pixel points to be in a range of [ -1,1] for each input picture, a high-dimensional matrix of dimension 64 x 3 corresponding to the input picture is used for randomly generating a convolution matrix of 3 x 3 for the high-dimensional matrix, the step length is moved to be 2, convolution operation is carried out, an output layer1 of dimension 32 x 64 is generated, layer1= max (layer 1, α x layer 1) is taken, and the layer1 is output;
b) Inputting the layer1 high-dimensional matrix with the dimension of 32 × 64 into a second convolution kernel of a second detection module, using a convolution matrix of 3 × 3 for the layer1 high-dimensional matrix, moving the step by 2, performing convolution operation, generating an output layer2 with the dimension of 16 × 128, taking layer2= max (layer 2, α × layer 2), and outputting the layer2;
c) Inputting the layer2 high-dimensional matrix with the dimension of 16 × 128 into a third convolution kernel of the second detection module, using a convolution matrix of 3 × 3 for the layer2 high-dimensional matrix, moving the step length to 2, performing convolution operation to generate an output layer3 with the dimension of 8 × 256, taking layer3= max (layer 3, α × layer 3), and outputting layer3;
d) Inputting the layer3 high-dimensional matrix with the dimension size of 8 × 256 into a fourth convolution kernel of the second detection module, using a convolution matrix with 3 × 3 for the layer3 high-dimensional matrix, moving the step length to 2, performing convolution to generate an output layer4 with the dimension size of 4 × 512, taking layer4= max (layer 4, α × layer 4), and outputting the layer4;
e) The layer4 high-dimensional matrix with the dimension size of 4 x 512 is input and output into a layer, all coefficients are expanded to generate a one-dimensional vector of (4 x 512) x 1, and the vector is calculated with random parameters of the layer to generate 1-bit output logits = ∑ sigma i W i *X i Input logits into
And obtaining output of the second detection module.
S302, updating the parameters of the second detection module, including:
1) Inputting the uniformly distributed noise matrix Q matrix into a first detection module, setting the size of a convolution matrix to be 3 x 3, outputting a picture with the dimension of W x 3, and setting all corresponding labels to be 0;
a) Calculating the Q matrix of 100 × 1 by using each random parameter of the input layer of the first detection module, outputting a one-dimensional vector with the size of 4 × 512, converting the one-dimensional vector into a matrix L11 with the size of 4 × 512, taking L11= max (L1, α × L1), and outputting the result as L11;
b) Inputting the L11 high-dimensional matrix with the dimension size of 4 x 512 into a first deconvolution kernel, using the deconvolution matrix with the dimension size of 3 x 3, moving the step size to 2, performing deconvolution to generate an output L21 with the dimension size of 8 x 256, taking L22= max (L21, alpha L21), and outputting the result as L21;
c) Inputting the L21 high-dimensional matrix with the dimension size of 8 × 256 into a second deconvolution kernel, performing deconvolution by using the deconvolution matrix with the dimension size of 3 × 3 and the step size of 2, generating an output L31 with the dimension size of 16 × 128, taking L31= max (L31, α × L31), and outputting the output result as L31;
d) Inputting the L31 high-dimensional matrix with dimension size of 16 × 128 into a third deconvolution kernel, performing deconvolution using the deconvolution matrix with 3 × 3 and step size shift of 2, generating an output L41 with dimension size of 32 × 64, taking L41= max (L41, α × L41), and outputting the result as L41;
e) And inputting the L41 high-dimensional matrix with the dimension size of 32 x 64 into a fourth deconvolution kernel, using the deconvolution matrix with the dimension size of 3 x 3, moving the step length to 2, performing deconvolution to generate an output logits1 with the dimension size of 64 x 3, and outputting the output 1 through a tanh function (so that the range of pixel points of the generated picture is [ -1,1 ]).
2) Inputting all the images with labels generated in the step 1) in the steps 2 of S2 and S3 into a second detection module, wherein the size of the convolution matrix is 3 x 3, calculating a discrimination error loss1 by comparing an output discrimination result with the labels, and updating parameters of the second detection module;
a) For each picture, firstly adjusting the pixel points to be in a range of [ -1,1], inputting the high-dimensional matrix with the dimension of 64 × 3 into a first convolution kernel of a second detection module, randomly generating a convolution matrix with the dimension of 3 × 3, moving the step length to be 2, performing convolution to generate an output layer12 with the dimension of 32 × 64, taking layer12= max (layer 12, α layer 12), and outputting the layer12;
b) Inputting the layer12 high-dimensional matrix with the dimension size of 32 × 64 into a second convolution kernel of a second detection module, performing convolution by using a convolution matrix with 3 × 3 and step shift of 2 to generate an output layer22 with the dimension size of 16 × 128, taking layer22= max (layer 22, α × layer 22), and outputting the layer22;
c) Inputting the layer22 high-dimensional matrix with the dimension size of 16 × 128 into a third convolution kernel of the second detection module, performing convolution by using a convolution matrix with 3 × 3 and the step length being moved to 2 to generate an output layer32 with the size of 8 × 256, and taking layer32= max (layer 32, α × layer 32) and outputting the layer32;
d) Inputting the layer32 high-dimensional matrix with the dimension size of 8 × 256 into a fourth convolution kernel of the second detection module, performing convolution by using the convolution matrix with the dimension size of 3 × 3 and moving the step size to 2 to generate an output layer42 with the dimension size of 4 × 512, taking layer42= max (layer 42, α × layer 42), and outputting the layer42;
e) Inputting the layer42 high-dimensional matrix with the dimension size of 4 x 512 into the output layer of the second detection module, expanding the layer42 high-dimensional matrix, generating a one-dimensional vector of (4 x 512 x 1), calculating the vector and each parameter of the output layer to obtain 1-bit output logits2, inputting the logits2 into the output layer of the second detection module
Function, output outputs2,/>
Labelsl is compared to the label of each picture itself, and the cross entropy cost function is used to generate a trained loss1:
where lossl is the cost function, y represents the actual label, a represents the output value labels1, and n represents the total number of samples.
And optimizing the parameters of the second detection module by using an adam optimizer. For the second detection module, the output value of the j-th feature map of the 1 st layer
The calculation expression is as follows:
wherein, M j J-th input feature map combination, k, representing a selection ij Is the convolution matrix used for the connection between the input ith and jth feature maps, b j Is the bias corresponding to the jth feature map, and f is the feature map mapping function.
And (3) output values of the ith characteristic diagram of the l-1 layer are shown. * Representing a convolution operation.
The core formula for updating the parameters is as follows:
where up (-) represents the upsampling operation,
and represents the error transfer coefficient on the jth characteristic diagram of the 1+1 layer. Based on the formula (1.3), the error on the jth feature map of the l +1 th layer is determined>
An error on a feature map of type j of the layer l is calculated>
And realizing the back propagation calculation of the error. />
Error in j-th characteristic diagram, u, of l + 1-th layer l A calculated intermediate vector representing the neural network of layer1, expressed as:
and calculating the partial derivative of the error cost function E to the convolution matrix by using the error of the characteristic diagram of each layer, wherein the calculation formula is as follows:
here, the number of the first and second electrodes,
is/>
When making convolution, with k ij The central weight (u, v) of each patch for convolution is the central vector of the patch, and the value of the (u, v) position in the output feature map is determined by the (u, v) position in the input feature map and the convolution matrix k ij The resulting values are convolved. />
Represents->
The value under the center vector (u, v) of patch.
Calculating the partial derivative of the error cost function E to the bias b:
updating the formula with the parameters:
all parameters in the second detection module are updated, and α 1 represents an update constant.
S303, updating each parameter of the first detection module, where the cycle number i = i +1, and includes:
1) Setting all the picture labels generated in the step a) in the step S302 as 1, inputting the picture labels into the second detection module with the updated parameters in the step S302, wherein the size of the convolution matrix is 3 × 3, calculating a judgment error loss2 by comparing an output judgment result with the labels, and updating the parameters of the first detection module (the parameters of the second detection module are not changed in the process);
a) For each picture, firstly adjusting the pixel points to be in a range of [ -1,1], inputting a 64 × 3 high-dimensional matrix into a first deconvolution kernel of a first detection module, randomly generating a 3 × 3 convolution matrix, moving the step length to 2, performing deconvolution operation to generate an output layer13 with a dimension of 32 × 64, taking layer13= max (layer 13, α × layer 13), and outputting layer1;
b) Inputting the layer1 high-dimensional matrix with the dimension of 32 × 64 into a second deconvolution kernel of the first detection module, performing deconvolution operation by using the convolution matrix with 3 × 3 and the step length being shifted to 2, generating an output layer23 with the dimension of 16 × 128, taking layer23= max (layer 23, α × layer 23), and outputting layer23;
c) Inputting the layer23 high-dimensional matrix with the dimension size of 16 × 128 into a third deconvolution kernel of the first detection module, performing deconvolution operation by using a convolution matrix with 3 × 3 and step size shift of 2, generating an output layer33 with the size of 8 × 256, taking layer33= max (layer 33, α × layer 33), and outputting layer33;
d) Inputting the layer33 high-dimensional matrix with the dimension size of 8 × 256 into the fourth deconvolution kernel of the first detection module, performing deconvolution operation by using the convolution matrix with 3 × 3 and the step size shifted to 2, generating an output layer43 with the size of 4 × 512, taking layer43= max (layer 4, α × layer 4), and outputting layer4;
e) Inputting a layer43 high-dimensional matrix with the dimension size of 4 x 512 into an output layer of the first detection module, expanding the layer43 high-dimensional matrix into a one-dimensional vector of (4 x 512) x 1, and calculating the vector and each parameter of the layer to generate 1-bit output logits3= ∑ sigma i W i *X i Inputs logits3
Function, output outputs3, <' > is asserted>
The labels of labels3 and each picture are compared, and a cross entropy cost function is used for generatingLoss2 of live training:
where loss2 is the cost function, y represents the actual label, a represents the output value labels3, and n represents the total number of samples.
Optimization was performed using the adam optimizer, combining equations (1.2), (1.3), (1.4), (1.5) with the parameter update equations:
all parameters in the first detection module are updated.
S304, setting cycle times M, judging whether the current cycle times i are equal to the current cycle times M, if i is less than M, entering step 2 to continuously update parameters of a first detection module and a second detection module, and if i = M, ending the cycle to obtain the first detection module and the second detection module, wherein the second detection module is the established spectrum anomaly detection model.
The above description is only an example of the present application and is not intended to limit the present application. Various modifications and changes may occur to those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.
Claims (10)
1. A method for detecting spectrum anomalies, comprising:
s1, acquiring radio signal data; using the acquired radio signal data as data to be processed; carrying out short-time Fourier transform on data to be processed to obtain short-time Fourier transform data;
s2, classifying and extracting the short-time Fourier transform data according to the class composition information and the parameter information of the short-time Fourier transform data to obtain channel data; constructing first picture data by using the channel data; setting a forward tag for the first picture data to obtain first tag data;
the category composition information of the short-time Fourier transform data comprises real data information and imaginary data information, and the parameter information of the short-time Fourier transform data comprises power data information and phase data information;
s3, establishing a first detection module and a second detection module; respectively initializing the first detection module and the second detection module to respectively obtain parameter initial values of the first detection module and the second detection module;
respectively updating parameters of the first detection module and the second detection module by using the first picture data and the first label data to respectively obtain the first detection module with the updated parameters and the second detection module with the updated parameters;
establishing an electromagnetic spectrum anomaly detection model by using the second detection module after the parameters are updated;
s4, processing the channel data by using the electromagnetic spectrum anomaly detection model to obtain output data of the electromagnetic spectrum anomaly detection model; and judging and counting the output data to obtain the occurrence probability of the electromagnetic spectrum abnormality.
2. The spectrum anomaly detection method according to claim 1, wherein said step S2 comprises:
classifying, extracting and processing the short-time Fourier transform data according to the class composition information and the parameter information of the short-time Fourier transform data to obtain real part data, imaginary part data, power data and phase data of the short-time Fourier transform data; integrating real part data, imaginary part data and power data or phase data of the short-time Fourier transform data to obtain channel data; constructing first picture data by using the channel data; and setting the label of the first picture data as 1 to obtain first label data.
3. The spectrum anomaly detection method according to claim 1, wherein said step S3 comprises:
s31, respectively establishing a first detection module and a second detection module by using a deep neural network model; generating first random matrix data, and training a first detection module and a second detection module respectively by using the first random matrix data to complete initialization processing of the first detection module and the second detection module so as to obtain parameter initial values of the first detection module and the second detection module; setting the cycle number as an initial value;
s32, generating second random matrix data; inputting the second random matrix data into the first detection module to obtain an output matrix of the first detection module, and taking the output matrix of the first detection module as second picture data; performing negative label setting on the second picture data to obtain second label data;
s33, inputting the first picture data and the first label data into a second detection module to obtain first judgment result data; inputting the second picture data and the second label data into a second detection module to obtain second judgment result data; updating parameters of the second detection module by using a difference value between the first judgment result data and the second judgment result data to obtain a second detection module with updated parameters;
s34, setting a forward tag for the second picture data to obtain third tag data; inputting the second picture data and the third label data into the second detection module after the parameter is updated, and obtaining third judgment result data; updating parameters of the first detection module by using a difference value between the label data of the third discrimination result data and the third label data to obtain a parameter-updated first detection module; increasing the number of cycles by 1;
s35, judging whether the updating stop condition is met, if the updating stop condition is not met, taking the first detection module with the updated parameters as a first detection module and taking the second detection module with the updated parameters as a second detection module, returning to the step S32, and continuously updating the parameters of the first detection module and the second detection module; and if the update stop condition is met, using the second detection module with the updated parameters as an electromagnetic spectrum anomaly detection model.
4. The method for detecting spectrum abnormality according to claim 3, wherein said updating the parameter of the second detecting module by using the difference between the first discrimination result data and the second discrimination result data to obtain the updated parameter of the second detecting module includes:
inputting the first discrimination result data and the second discrimination result data into a cost function, and processing the first discrimination result data and the second discrimination result data by using the cost function to obtain an error cost function; calculating a partial derivative of the error cost function to the parameter of the second detection module by using the error cost function; performing conversion processing on the partial derivative to obtain a parameter updating value; and updating the parameters of the second detection module by using the parameter updating value to obtain the second detection module with updated parameters.
5. The method for detecting spectrum anomaly according to claim 4, wherein said partial derivatives are transformed to obtain updated values of parameters; updating the parameters of the second detection module by using the parameter update value to obtain the second detection module with updated parameters, comprising:
the partial derivative is expressed as
k i The ith parameter of the second detection module is represented, the E represents an error cost function, and a calculation formula for performing conversion processing on the partial derivative is as follows:
wherein, Δ k i Updating a parameter of the ith parameter of the second detection module;
and multiplying the parameter updating value by the corresponding parameter of the second detection module obtained after the last round of parameter updating to obtain the updated parameter of the second detection module.
6. The spectrum anomaly detection method of claim 3, wherein said determining whether an update-stop condition is satisfied comprises:
judging whether the cycle number exceeds a set threshold or not, or performing difference judgment on the tag data of the third judgment result data and the third tag data;
the judging whether the cycle number exceeds a set threshold value includes: when the number of cycles exceeds a set threshold, the update stop condition is considered to be satisfied, and when the number of cycles does not exceed the set threshold, the update stop condition is considered not to be satisfied;
the performing difference judgment on the tag data of the third judgment result data and the third tag data includes:
performing autoregressive-moving average model modeling processing on the tag data of the third discrimination result data and the third tag data respectively to obtain a first autoregressive-moving average model and a second autoregressive-moving average model; extracting to obtain coefficient vectors of two autoregressive-moving average models, and calculating to obtain a cross-correlation matrix of the coefficient vectors; performing eigenvalue decomposition operation on the cross-correlation matrix to obtain the maximum eigenvalue of the cross-correlation matrix; and judging the maximum characteristic value, and when the maximum characteristic value is greater than the difference judging threshold value, determining that the updating stop condition is not met, and when the maximum characteristic value is less than or equal to the difference judging threshold value, determining that the updating stop condition is met.
7. The method for spectrum anomaly detection according to claim 1, comprising:
the short-time Fourier transform adopts a hamming window with a window function of W length; the number of overlapped samples of the short-time Fourier transform is 2/3 x W, and the sampling frequency is Fs; the length of the data to be processed is L, and the sampling frequency is Fs.
8. The spectrum anomaly detection method of claim 1, wherein said step S4 comprises:
s41, splitting channel data according to the dimensionality of the picture pixel matrix data to obtain a plurality of single-picture data;
s42, inputting each single picture data into the electromagnetic spectrum abnormality detection model to obtain output data of the electromagnetic spectrum abnormality detection model of the single picture data;
s43, judging and processing output data of the electromagnetic spectrum abnormality detection model of the single picture data to obtain a spectrum abnormality judgment result of the single picture data;
s44, counting the frequency spectrum abnormity discrimination results of all single picture data and the number of the single picture data to obtain the occurrence probability of electromagnetic frequency spectrum abnormity.
9. The method for detecting spectrum abnormality according to claim 8, wherein said determining the output data of the electromagnetic spectrum abnormality detection model for single-sheet picture data to obtain the spectrum abnormality determination result for single-sheet picture data includes:
the output data of the electromagnetic spectrum abnormality detection model of the single picture data is expressed as outputs, the spectrum abnormality judgment result of the single picture data is expressed as labels, and the judgment processing process is expressed as follows:
if labels is equal to 1, the frequency spectrum corresponding to the image data is considered to be in a normal working state, and if labels is equal to 0, the frequency spectrum corresponding to the image data is considered to be abnormal.
10. The method for detecting spectrum abnormality according to claim 8, wherein the obtaining the occurrence probability of electromagnetic spectrum abnormality by statistically processing the spectrum abnormality discrimination results of all single-picture data and the number of the single-picture data includes:
summing the frequency spectrum abnormality judgment results of all single picture data to obtain the occurrence frequency of the abnormality judgment result; and carrying out division operation by using the occurrence frequency of the abnormity discrimination result and the number of single picture data to obtain the occurrence probability of the electromagnetic spectrum abnormity.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211628547.7A CN115964670B (en) | 2022-12-17 | 2022-12-17 | Spectrum anomaly detection method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211628547.7A CN115964670B (en) | 2022-12-17 | 2022-12-17 | Spectrum anomaly detection method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115964670A true CN115964670A (en) | 2023-04-14 |
| CN115964670B CN115964670B (en) | 2023-07-07 |
Family
ID=85895639
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211628547.7A Active CN115964670B (en) | 2022-12-17 | 2022-12-17 | Spectrum anomaly detection method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115964670B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117147966A (en) * | 2023-08-30 | 2023-12-01 | 中国人民解放军军事科学院系统工程研究院 | Electromagnetic spectrum signal energy anomaly detection method |
| CN117420527A (en) * | 2023-12-19 | 2024-01-19 | 南京航天工业科技有限公司 | Radar monitoring method based on broadband real-time spectrum analysis |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180012020A1 (en) * | 2016-01-14 | 2018-01-11 | Georgia Tech Research Corporation | Systems and methods for runtime program monitoring through analysis of side channel signals |
| KR101847638B1 (en) * | 2016-11-23 | 2018-04-10 | 주식회사 시그널웍스 | Method for analyzing acoustic and apparatus for testing non destructive of using the method |
| US20190072601A1 (en) * | 2017-01-23 | 2019-03-07 | DGS Global Systems, Inc. | Systems, methods, and devices for automatic signal detection based on power distribution by frequency over time within a spectrum |
| CN110705456A (en) * | 2019-09-29 | 2020-01-17 | 华南理工大学 | Micro motor abnormity detection method based on transfer learning |
| WO2020044495A1 (en) * | 2018-08-30 | 2020-03-05 | Nec Corporation | Anomaly detection apparatus, method and computer-readable medium |
| CN113296990A (en) * | 2020-09-30 | 2021-08-24 | 阿里云计算有限公司 | Method and device for recognizing abnormity of time sequence data |
| JP6942836B1 (en) * | 2020-04-08 | 2021-09-29 | 株式会社川本製作所 | Anomaly detection device |
-
2022
- 2022-12-17 CN CN202211628547.7A patent/CN115964670B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180012020A1 (en) * | 2016-01-14 | 2018-01-11 | Georgia Tech Research Corporation | Systems and methods for runtime program monitoring through analysis of side channel signals |
| KR101847638B1 (en) * | 2016-11-23 | 2018-04-10 | 주식회사 시그널웍스 | Method for analyzing acoustic and apparatus for testing non destructive of using the method |
| US20190072601A1 (en) * | 2017-01-23 | 2019-03-07 | DGS Global Systems, Inc. | Systems, methods, and devices for automatic signal detection based on power distribution by frequency over time within a spectrum |
| WO2020044495A1 (en) * | 2018-08-30 | 2020-03-05 | Nec Corporation | Anomaly detection apparatus, method and computer-readable medium |
| CN110705456A (en) * | 2019-09-29 | 2020-01-17 | 华南理工大学 | Micro motor abnormity detection method based on transfer learning |
| JP6942836B1 (en) * | 2020-04-08 | 2021-09-29 | 株式会社川本製作所 | Anomaly detection device |
| CN113296990A (en) * | 2020-09-30 | 2021-08-24 | 阿里云计算有限公司 | Method and device for recognizing abnormity of time sequence data |
Non-Patent Citations (4)
| Title |
|---|
| 于永强: "基于深度迁移学习的铣刀磨损量预测研究", 《中国优秀硕士学位论文全文数据库 工程科技Ⅰ辑》, no. 3 * |
| 夏庆升;: "基于特征频段法的舰船水压场信号检测", 水雷战与舰船防护, no. 03 * |
| 姜煜 等: "基于声矢量场处理的海洋内波预警监测技术", 《电子科技》 * |
| 朱军涛;廖红建;谢勇勇;昝月稳;: "采用短时傅里叶变换的铁路车载探地雷达数据解译方法", 西安交通大学学报, no. 07 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117147966A (en) * | 2023-08-30 | 2023-12-01 | 中国人民解放军军事科学院系统工程研究院 | Electromagnetic spectrum signal energy anomaly detection method |
| CN117147966B (en) * | 2023-08-30 | 2024-05-07 | 中国人民解放军军事科学院系统工程研究院 | Electromagnetic spectrum signal energy anomaly detection method |
| CN117420527A (en) * | 2023-12-19 | 2024-01-19 | 南京航天工业科技有限公司 | Radar monitoring method based on broadband real-time spectrum analysis |
| CN117420527B (en) * | 2023-12-19 | 2024-03-12 | 南京航天工业科技有限公司 | Radar monitoring method based on broadband real-time spectrum analysis |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115964670B (en) | 2023-07-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107529650B (en) | Closed loop detection method and device and computer equipment | |
| CN115964670B (en) | Spectrum anomaly detection method | |
| CN105740894B (en) | Semantic annotation method for hyperspectral remote sensing image | |
| CN106846729B (en) | Tumble detection method and system based on convolutional neural network | |
| CN112287816B (en) | Dangerous work area accident automatic detection and alarm method based on deep learning | |
| CN111582320B (en) | Dynamic individual identification method based on semi-supervised learning | |
| CN111611874B (en) | Face mask wearing detection method based on ResNet and Canny | |
| CN112673381B (en) | Method and related device for identifying confrontation sample | |
| CN111738054B (en) | Behavior anomaly detection method based on space-time self-encoder network and space-time CNN | |
| CN113283356B (en) | Multistage attention scale perception crowd counting method | |
| CN108171119B (en) | SAR image change detection method based on residual error network | |
| CN115761900B (en) | Internet of things cloud platform for practical training base management | |
| Mathias et al. | Deep Neural Network Driven Automated Underwater Object Detection. | |
| CN114154532A (en) | Unmanned aerial vehicle individual multi-dimensional domain electromagnetic signal feature deep learning identification method | |
| CN116894207A (en) | Intelligent radiation source identification method based on Swin transducer and transfer learning | |
| CN113537240B (en) | Deformation zone intelligent extraction method and system based on radar sequence image | |
| CN115358952A (en) | Image enhancement method, system, equipment and storage medium based on meta-learning | |
| CN115909144A (en) | Method and system for detecting abnormity of surveillance video based on counterstudy | |
| CN114972871A (en) | Image registration-based few-sample image anomaly detection method and system | |
| CN112560712B (en) | Behavior recognition method, device and medium based on time enhancement graph convolutional network | |
| CN114724245A (en) | CSI-based incremental learning human body action identification method | |
| CN114760104A (en) | Distributed abnormal flow detection method in Internet of things environment | |
| Tralie et al. | Multi-scale geometric summaries for similarity-based sensor fusion | |
| CN111382761A (en) | CNN-based detector, image detection method and terminal | |
| CN117421723B (en) | SERVER MESH-based microservice system |
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 |