CN115856425A - Spectrum anomaly detection method and device based on hidden space probability prediction - Google Patents
Spectrum anomaly detection method and device based on hidden space probability prediction Download PDFInfo
- Publication number
- CN115856425A CN115856425A CN202211461149.0A CN202211461149A CN115856425A CN 115856425 A CN115856425 A CN 115856425A CN 202211461149 A CN202211461149 A CN 202211461149A CN 115856425 A CN115856425 A CN 115856425A
- Authority
- CN
- China
- Prior art keywords
- training
- complex value
- network
- time
- sequence signal
- 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
Abstract
The invention discloses a spectrum anomaly detection method and a device based on hidden space probability prediction, wherein the method comprises the following steps: acquiring a training complex value time sequence signal, a testing complex value time sequence signal and a complex value time sequence signal to be detected; preprocessing the training complex value time sequence signal to obtain time-frequency information of the training complex value time sequence signal; constructing an automatic coding network, and training the automatic coding network by using the time-frequency information of the training complex value time sequence signal to obtain a training automatic coding network; the training automatic coding network comprises a training convolution coding network and a training deconvolution decoding network; processing the test complex value time sequence signal by utilizing a training convolutional coding network to obtain hidden space vector information; training the Gaussian mixture long-time and short-time memory network by using the hidden space vector information to obtain a training Gaussian mixture long-time and short-time memory network; and processing the complex value time sequence signal to be detected by utilizing a training Gaussian mixture long-time and short-time memory network to obtain a frequency spectrum abnormity detection result. The method has high processing efficiency and strong robustness.
Description
Technical Field
The invention relates to the technical field of artificial intelligence, in particular to a spectrum anomaly detection method and device based on hidden space probability prediction.
Background
In thousands of information brought by the electromagnetic spectrum, timely capturing of the occurrence of anomalies in the spectrum can effectively target factors threatening the safety and reliability of the spectrum, such as detection of illegally built private radio stations in radio, electromagnetic interference weapons put into use in battlefields, and the like. However, the spectrum use behaviors in the complex electromagnetic environment of modern cities are often very random and sudden, so that the radio monitoring spectrum states show remarkable variability and fluctuation. The conventional spectrum anomaly detection method has several problems: (1) Relying on long-term monitoring data accumulation and a priori probability distribution knowledge of various frequency-using parameters; (2) The spectrum data volume is larger and wider, so that the calculation complexity is large and the detection delay is high; (3) The robustness to factors such as slow change of frequency spectrum and interference of noise is poor.
Two methods similar to the present invention are Automatic Encoding (AE) analysis and long-term memory (LSTM) networks. The AE analysis aims at encoding a signal with a wider frequency band into more lightweight implicit vectors, which are intended as a way of extracting features, and the distance between the implicit vectors is taken as the basis of spectrum anomaly classification, but the method is often lack of intuition in processing a timing problem and is not outstanding in performance. The LSTM network determines whether an abnormality occurs by predicting a spectrum vector at the next time and detecting a difference between the predicted spectrum vector and a real spectrum, and it is very intuitive to use this method to process a time-series signal. In addition, even if the two methods are combined with each other before the LSTM method, the AE method is used for feature extraction, and various interference items such as noise and environmental changes are still bottlenecks in improving the detection accuracy.
Disclosure of Invention
The technical problem to be solved by the invention is to provide a spectrum anomaly detection method and device based on hidden space probability prediction, which can be used for firstly coding time-frequency information (spectrum signals) into hidden space vectors of a low-dimensional space by combining an AE method through training AE, then passing through a special Gaussian mixture long-short time memory (GM-LSTM) network, predicting the probability distribution of each frequency point of the spectrum vectors at the next moment, and detecting the confidence probability of the distribution of the real spectrum at the next moment to judge the anomaly. The method skillfully combines the respective advantages of the AE method and the LSTM method, and has good robustness to noise and environmental changes due to modeling of distribution.
In order to solve the above technical problem, a first aspect of the embodiments of the present invention discloses a spectrum anomaly detection method based on implicit spatial probability prediction, where the method includes:
s1, acquiring a complex value time sequence signal; the complex value time sequence signal comprises a training complex value time sequence signal, a testing complex value time sequence signal and a complex value time sequence signal to be detected;
s2, preprocessing the training complex value time sequence signal to obtain time frequency information of the training complex value time sequence signal;
s3, constructing an automatic coding network; the automatic coding network comprises a convolution coding network and a deconvolution decoding network;
s4, training the automatic coding network by using the time-frequency information of the training complex value time sequence signal to obtain a training automatic coding network; the training automatic coding network comprises a training convolutional coding network and a training deconvolution decoding network;
s5, processing the test complex value time sequence signal by using the training convolutional coding network to obtain hidden space vector information;
s6, training the Gaussian mixture long-time and short-time memory network by using the hidden space vector information to obtain a training Gaussian mixture long-time and short-time memory network;
and S7, processing the complex value time sequence signal to be detected by using the training Gaussian mixture long-time and short-time memory network to obtain a frequency spectrum abnormity detection result.
As an optional implementation manner, in the first aspect of the embodiment of the present invention, the preprocessing the training complex value timing signal to obtain time-frequency information of the training complex value timing signal includes:
s21, performing windowing operation on the training complex value time sequence signal to obtain a window selection training complex value time sequence signal;
and S22, carrying out time-frequency transformation on the window selection training complex value time sequence signal to obtain time-frequency information of the training complex value time sequence signal.
As an optional implementation manner, in the first aspect of the embodiment of the present invention, the training the automatic coding network by using the time-frequency information of the training complex value timing signal to obtain a training automatic coding network includes:
s41, training the automatic coding network by using the time-frequency information of the training complex value time sequence signal to obtain reconstructed frequency spectrum information;
s42, constructing a loss function by using the reconstructed spectrum information;
s43, calculating the gradient of the parameter to be optimized in the automatic coding network by using the loss function, and obtaining an updated parameter to be optimized by using a gradient descent method;
and S44, repeating S41-S43 until the automatic coding network is converged to obtain a training automatic coding network.
As an optional implementation manner, in the first aspect of the embodiment of the present invention, the processing the test complex-valued timing signal by using the training convolutional coding network to obtain hidden space vector information includes:
s51, acquiring a test complex value time sequence signal, and preprocessing the test complex value time sequence signal to obtain time frequency information of the test complex value time sequence signal;
and S52, processing the time-frequency information of the test complex value time sequence signal by using the training convolutional coding network to obtain hidden space vector information.
As an optional implementation manner, in the first aspect of the embodiment of the present invention, the training the gaussian mixture long-short term memory network by using the implicit space vector information to obtain a trained gaussian mixture long-short term memory network includes:
s61, constructing a Gaussian mixture long-time and short-time memory network;
s62, processing the Gaussian mixture long-time memory network by using the hidden space vector information to obtain a mixture ratio, a mean value and a standard deviation of Gaussian distribution;
s63, constructing a Gaussian mixture long-time memory network loss function by using the mixing proportion, the mean value and the standard deviation of the Gaussian distribution;
s64, calculating the gradient of the parameter to be optimized in the Gaussian mixture long-short time memory network by using the loss function of the Gaussian mixture long-short time memory network, and updating the parameter to be optimized in the Gaussian mixture long-short time memory network by using a gradient descent method;
and S65, repeating S61-S64 until the Gaussian mixture long-time and short-time memory network converges to obtain the training Gaussian mixture long-time and short-time memory network.
As an optional implementation manner, in the first aspect of the embodiment of the present invention, the processing the complex-valued timing signal to be detected by using the training gaussian mixture long-time and short-time memory network to obtain a spectrum anomaly detection result includes:
s71, acquiring a complex value time sequence signal to be detected;
s72, preprocessing the complex value time sequence signal to be detected to obtain time frequency information of the complex value time sequence signal to be detected;
s73, processing the complex value time sequence signal time frequency information to be detected by using the training convolutional coding network to obtain testing hidden space vector information;
and S74, processing the test implicit space vector information by using the training Gaussian mixture long-time and short-time memory network to obtain a spectrum anomaly detection result.
As an alternative implementation manner, in the first aspect of the embodiment of the present invention, the automatic coding network includes a convolutional coding network and a deconvolution decoding network;
the automatic coding network comprises the following components:
wherein, x represents any time-frequency information input, E represents a convolution coding network, z represents a hidden vector obtained by the convolution coding network, D represents a deconvolution decoding network,
representing the output of the automatically encoded network.
The second aspect of the embodiment of the invention discloses a spectrum anomaly detection device based on hidden space probability prediction, which comprises:
the signal acquisition module is used for acquiring a complex value time sequence signal; the complex value time sequence signal comprises a training complex value time sequence signal, a testing complex value time sequence signal and a complex value time sequence signal to be detected;
the preprocessing module is used for preprocessing the training complex value time sequence signal to obtain time-frequency information of the training complex value time sequence signal;
the building module is used for building an automatic coding network; the automatic coding network comprises a convolution coding network and a deconvolution decoding network;
the first training module is used for training the automatic coding network by using the time-frequency information of the training complex value time sequence signal to obtain a training automatic coding network; the training automatic coding network comprises a training convolutional coding network and a training deconvolution decoding network;
the automatic coding module is used for processing the test complex value time sequence signal by utilizing the training convolutional coding network to obtain hidden space vector information;
the second training module is used for training the Gaussian mixture long-time and short-time memory network by using the implicit space vector information to obtain a training Gaussian mixture long-time and short-time memory network;
and the detection module is used for processing the complex value time sequence signal to be detected by utilizing the training Gaussian mixture long-time and short-time memory network to obtain a frequency spectrum abnormity detection result.
As an optional implementation manner, in a second aspect of the embodiment of the present invention, the preprocessing the training complex value timing signal to obtain time-frequency information of the training complex value timing signal includes:
s21, performing windowing operation on the training complex value time sequence signal to obtain a window selection training complex value time sequence signal;
and S22, carrying out time-frequency transformation on the window selection training complex value time sequence signal to obtain time-frequency information of the training complex value time sequence signal.
As an optional implementation manner, in a second aspect of the embodiment of the present invention, the training the automatic coding network by using the time-frequency information of the training complex value timing signal to obtain a training automatic coding network includes:
s41, training the automatic coding network by using the time-frequency information of the training complex value time sequence signal to obtain reconstructed frequency spectrum information;
s42, constructing a loss function by using the reconstructed spectrum information;
s43, calculating the gradient of the parameter to be optimized in the automatic coding network by using the loss function, and obtaining an updated parameter to be optimized by using a gradient descent method;
and S44, repeating S41-S43 until the automatic coding network is converged to obtain the training automatic coding network.
As an optional implementation manner, in a second aspect of the embodiment of the present invention, the processing the test complex-valued timing signal by using the training convolutional coding network to obtain hidden space vector information includes:
s51, acquiring a test complex value time sequence signal, and preprocessing the test complex value time sequence signal to obtain time frequency information of the test complex value time sequence signal;
and S52, processing the time-frequency information of the test complex value time sequence signal by using the training convolutional coding network to obtain hidden space vector information.
As an optional implementation manner, in a second aspect of the embodiments of the present invention, the training a gaussian mixture long-short term memory network by using the implicit spatial vector information to obtain a trained gaussian mixture long-short term memory network includes:
s61, constructing a Gaussian mixture long-time and short-time memory network;
s62, processing the Gaussian mixture long-time memory network by using the hidden space vector information to obtain a mixture ratio, a mean value and a standard deviation of Gaussian distribution;
s63, constructing a Gaussian mixture long-time memory network loss function by using the mixing proportion, the mean value and the standard deviation of the Gaussian distribution;
s64, calculating the gradient of the parameter to be optimized in the Gaussian mixture long-short time memory network by using the loss function of the Gaussian mixture long-short time memory network, and updating the parameter to be optimized in the Gaussian mixture long-short time memory network by using a gradient descent method;
and S65, repeating S61-S64 until the Gaussian mixture long-short term memory network converges to obtain the training Gaussian mixture long-short term memory network.
As an optional implementation manner, in a second aspect of the embodiment of the present invention, the processing the complex-valued timing signal to be detected by using the training gaussian mixture long-time and short-time memory network to obtain a spectrum anomaly detection result includes:
s71, acquiring a complex value time sequence signal to be detected;
s72, preprocessing the complex value time sequence signal to be detected to obtain time frequency information of the complex value time sequence signal to be detected;
s73, processing the time-frequency information of the complex value time sequence signal to be detected by using the training convolutional coding network to obtain testing hidden space vector information;
and S74, processing the test hidden space vector information by using the training Gaussian mixture long-time and short-time memory network to obtain a frequency spectrum abnormity detection result.
As an alternative implementation manner, in the second aspect of the embodiment of the present invention, the automatic coding network includes a convolutional coding network and a deconvolution decoding network;
the automatic coding network comprises the following steps:
wherein, x represents any time-frequency information input, E represents a convolutional coding network, z represents a hidden vector obtained by the convolutional coding network, D represents a deconvolution decoding network,
representing the output of the automatically encoded network. />
The third aspect of the invention discloses another spectrum anomaly detection device based on implicit spatial probability prediction, which comprises:
a memory storing executable program code;
a processor coupled with the memory;
the processor calls the executable program code stored in the memory to execute part or all of the steps of the spectrum anomaly detection method based on the hidden space probability prediction disclosed by the first aspect of the embodiment of the invention.
In a fourth aspect of the present invention, a computer storage medium is disclosed, where the computer storage medium stores computer instructions, and when the computer instructions are called, the computer instructions are configured to perform some or all of the steps in the spectrum anomaly detection method based on implicit spatial probability prediction disclosed in the first aspect of the embodiments of the present invention.
Compared with the prior art, the embodiment of the invention has the following beneficial effects:
in the embodiment of the invention, the processed time-frequency information is input into the automatic coding network for coding by constructing the automatic coding network; constructing a Gaussian mixture long-time and short-time memory network; and inputting the coded time sequence signal into a Gaussian mixture long-short time memory network, calculating the confidence probability that the code of the time-frequency information at the next moment belongs to the distribution output by the memory network, and outputting the abnormity when the probability is lower. The method of the invention can not only encode the high-dimensional frequency spectrum signal into a one-dimensional vector to accelerate the data processing efficiency, but also directly predict the signal distribution at the next moment by using the characteristics of the time sequence signal without relying on prior probability distribution knowledge, and simultaneously has strong robust capability to noise, environmental change and other interference factors.
Drawings
In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed to be used in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.
Fig. 1 is a schematic flowchart of a spectrum anomaly detection method based on implicit spatial probability prediction according to an embodiment of the present invention;
fig. 2 is a schematic diagram of an implementation framework for handling a spectrum anomaly detection problem of a spectrum anomaly detection method based on implicit spatial probability prediction according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of an output of a training convolutional encoding network disclosed in an embodiment of the present invention;
fig. 4 is a schematic structural diagram of a spectrum anomaly detection apparatus based on implicit spatial probability prediction according to an embodiment of the present invention;
fig. 5 is a schematic structural diagram of another spectrum anomaly detection apparatus based on implicit spatial probability prediction according to an embodiment of the present disclosure.
Detailed Description
In order to make those skilled in the art better understand the technical solutions of the present invention, 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.
The terms "first," "second," and the like in the description and claims of the present invention and in the above-described drawings are used for distinguishing between different objects and not necessarily for describing a particular sequential or chronological order. Furthermore, the terms "include" and "have," as well as any variations thereof, are intended to cover a non-exclusive inclusion. For example, a process, method, apparatus, product, or apparatus that comprises a list of steps or elements is not limited to those listed but may alternatively include other steps or elements not listed or inherent to such process, method, product, or apparatus.
Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the invention. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments. It is explicitly and implicitly understood by one skilled in the art that the embodiments described herein can be combined with other embodiments.
The invention discloses a spectrum anomaly detection method and a spectrum anomaly detection device based on implicit spatial probability prediction, which can solve the problem of spectrum anomaly detection of signals with time sequence dependency. The anomaly detection problem can be described as: given a set of time series in which all of the normal spectrum signals are in the training phase, a pattern of normal spectrum is required to be learned. Then in the time sequence of the test stage, the sequence is fixed as a normal signal at the early stage, and then abnormal signals generated by communication or interference of other frequency devices randomly appear, and the currently received frequency spectrum is required to be output as the abnormal signals when the abnormality occurs through the analysis of the time sequence frequency spectrum signals.
Based on the hidden space probability prediction, on the basis of the basic framework of an automatic encoder network and a long-time and short-time memory network, the Gaussian mixture distribution of the signal codes at the next moment which are output as the prediction is adjusted, then the probability density function value of the distribution to which the real signal codes at the next moment belong is calculated, and when the value exceeds a certain threshold value, the current spectrum signals are output as abnormal signals. The method has strong robustness to factors such as environment change, noise and the like. The following are detailed descriptions.
Example one
Referring to fig. 1, fig. 1 is a schematic flowchart illustrating a spectrum anomaly detection method based on implicit spatial probability prediction according to an embodiment of the present invention. The spectrum anomaly detection method based on implicit spatial probability prediction described in fig. 1 may be applied to the technical field of artificial intelligence, such as network security, fault analysis, artificial intelligence, spectrum coordination, radio security management and control, regional electromagnetic detection and protection, and the embodiment of the present invention is not limited thereto. As shown in fig. 1, the spectrum anomaly detection method based on implicit spatial probability prediction may include the following operations:
s1, acquiring a complex value time sequence signal; the complex value time sequence signal comprises a training complex value time sequence signal, a testing complex value time sequence signal and a complex value time sequence signal to be detected;
s2, preprocessing the training complex value time sequence signal to obtain time-frequency information of the training complex value time sequence signal;
s3, constructing an automatic coding network; the automatic coding network comprises a convolution coding network and a deconvolution decoding network;
s4, training the automatic coding network by using the time-frequency information of the training complex value time sequence signal to obtain a training automatic coding network; the training automatic coding network comprises a training convolution coding network and a training deconvolution decoding network;
s5, processing the test complex value time sequence signal by using the training convolutional coding network to obtain hidden space vector information;
s6, training the Gaussian mixture long-time and short-time memory network by using the hidden space vector information to obtain a training Gaussian mixture long-time and short-time memory network;
and S7, processing the complex value time sequence signal to be detected by using the training Gaussian mixture long-time and short-time memory network to obtain a frequency spectrum anomaly detection result.
Fig. 2 is a schematic diagram of an implementation framework for processing a spectrum anomaly detection problem of a spectrum anomaly detection method based on implicit spatial probability prediction according to an embodiment of the present invention.
Optionally, the long-time and short-time memory network of the gaussian mixture takes connection of the long-time and short-time memory network and three linear networks as main features, the long-time and short-time memory network processes a time sequence relation of the coding vectors, and the three linear networks respectively output a predicted mixture ratio, a predicted average value and a predicted standard deviation of the gaussian mixture distribution at the next time.
The Gaussian mixture long-time memory network is described as follows:
i t ＝σ(W xi z t +W hi h t-1 )
f t ＝σ(W xf z t +W hf h t-1 )
o t ＝σ(W xo z t +W ho h t-1 )
c t ＝f t ⊙c t-1 +i t ⊙tanh(W xc z t +W hc h t-1 )
h t ＝o t ⊙tanh(c t )
μ t ＝W μ h t
σ t ＝W σ h t
π t ＝W π h t
wherein z is t Code vector input representing time t, c t Cell variable, h, representing LSTM t Hidden layer output, i, representing LSTM t ,f t ,o t Respectively representing intermediate process variables, mu, of the input gate, the forgetting gate and the output gate t ,σ t ,π t Which represents the final output of the network,mean, standard deviation and mixture ratio of the Gaussian distributions, W xi ，W hi ，W xf ，W hf ，W xo ，W ho ，W xc ，W hc ，W μ ，W σ ，W π Denotes the weight of the network, σ denotes the sigmoid activation function, and a indicates the Hadamard product operation of the matrix.
Optionally, the probability density function of the mixture distribution output by the gaussian mixture length time memory network is described as follows:
where p (y | x) represents the probability density function value for a given input x whose output is y, the subscript k represents the kth component of the corresponding vector,
represents the specified mean value μ k (x) And variance->
The value of the probability density function corresponding to the y point in the normal distribution of (1).
Optionally, the loss function L used by the gaussian mixture long-and-short term memory network is described as follows:
optionally, the preprocessing the training complex value timing signal to obtain time-frequency information of the training complex value timing signal includes:
s21, performing windowing operation on the training complex value time sequence signal to obtain a window selection training complex value time sequence signal;
and S22, carrying out time-frequency transformation on the window selection training complex value time sequence signal to obtain time-frequency information of the training complex value time sequence signal.
The windowing, which is the multiplication of the signal and a window function, is a common method in signal processing.
Optionally, the time-frequency transform method includes:
given a signal x (t) e L 2 (R), time-frequency conversion STFT thereof x (t, Ω) is defined as
In the formula g t,Ω (τ)＝g(τ-t)e jΩτ G (τ) is a window function, t is a time variable, Ω is a frequency domain variable,<>representing an inner product operation;
and g (τ) | =1, | | g t,Ω (τ)||＝1
Optionally, a WVD time-frequency transform method may be adopted:
x (t) is the input signal, G (t, Ω) is the time-frequency distribution of a window function, let G (t, Ω) and W x The convolution of (t, Ω) in both the t and Ω directions is called smooth WVD, denoted SW x (t, Ω), namely:
optionally, the time-frequency information STFT x (t, Ω) or SW x (t, omega) carrying out bilateral filtering to obtain a time-frequency spectrogram after noise reduction and edge preservation
(i.e., audio information):
wherein W () represents an original time spectrogram,
representing the time-frequency spectrogram after bilateral filtering processing, m representing output pixel points, n representing input pixel points, s representing a set rectangular frame area for traversing an image, W being a two-dimensional template, and/or>
And
is two Gaussian functions respectively representing a spatial domain kernel and a pixel domain kernel, M m The specific expression of (A) is as follows:
in the formula, a and b represent horizontal and vertical coordinates of an input pixel, i and j represent coordinates of a box center, and σ represents s 、σ r The standard deviation of the gaussian function is indicated.
Therefore, a time-frequency diagram (time-frequency information) matrix arranged in a time sequence can be obtained, and the time-frequency diagram matrix is marked as X and has two dimensionalities:
X:＝[Time,Freq]
time and Freq respectively represent the Time dimension, namely the frame number of the whole spectrogram, and the frequency dimension, namely the frequency point number of each frame of the frequency spectrum.
The automatic coding network is composed of a convolution coding network and a deconvolution decoding network.
Optionally, the input data of the convolutional coding network is denoted as ConvIn, where ConvIn is a two-dimensional tensor expressed as:
ConvIn:＝[Batch,Freq]
where Batch represents the number of samples in each Batch, and Freq represents the number of frequency points in each spectrogram (i.e., time frequency graph).
Defining an optimizable parameter w, the three-dimensional tensor can be expressed as:
w:＝[FeatureMaps_ConvOut,FeatureMaps_ConvIn,KernelSize]
where, kernelSize, featureMaps _ ConvIn, featureMaps _ ConvOut represent the convolution kernel size, the number of input feature maps, and the number of output feature maps, respectively, of the convolutional layer. This step can be implemented using a torch. The above parameters can all be initialized using the Kaiming normal initialization method.
Calculating convolution output:
ConvOut＝F.conv1d(ConvIn,weight＝self.weight,bias＝None)
in the formula, f.conv1d is a one-dimensional convolution operation of torch.nn.functional, weight and bias are two parameters of the function f.conv1d, which respectively represent the weight and the offset of the convolution layer, and self.weight and None are values of the two parameters. Weight is w after initialization; bias = None indicates that the f.conv1d operation is unbiased.
The complete convolutional coding network is composed of a plurality of convolutional layers, the final output is a coded hidden space vector which is marked as LatentVector, and the dimensionality of the vector is as follows:
LatentVector:＝[Batch,LatentDim]
where LatentDim represents the length of the hidden space vector, usually a small value relative to the original input dimension.
The input data of the deconvolution decoding network is the output data of the convolution coding network, i.e. LatentVector, which is still denoted as DeconvIn for the sake of brevity, and is represented as:
DeconvIn:＝[Batch,LatentDim]
where the values and meanings of Batch and LatentDim are consistent with the description in the convolutional coding network.
Defining an optimizable parameter w, the three-dimensional tensor can be expressed as:
w:＝[FeatureMaps_ConvOut,FeatureMaps_ConvIn,KernelSize]
where, kernelSize, featureMaps _ ConvIn, featureMaps _ ConvOut represent the convolution kernel size, the number of input feature maps, and the number of output feature maps, respectively, of the convolutional layer. This step can be implemented using a torch. The above parameters can all be initialized using the Kaiming normal initialization method.
Calculating the deconvolution output:
in deconvo = f.convransposese 1d (DeconvIn, weight = self.weight, bias = None), f.convransposese 1d is a one-dimensional deconvolution operation of torch.nn.functional, and weight and bias are two parameters of the function f.convransposese 1d, respectively representing the weight and offset of the convolution layer, and self.weight and None are values of the two parameters, respectively. Weight is w after initialization; bias = None indicates that f.convtranspose1d operation is unbiased.
The complete convolutional coding network is composed of a plurality of deconvolution layers, the final output is a reconstructed spectrum signal obtained by coding and restoring, and the reconstructed spectrum signal is marked as ConvInRecon, and the dimensionality of the reconstructed spectrum signal is as follows:
ConvInRecon:＝[Batch,Freq]
where the values and meanings of Batch and Freq are consistent with the description in the convolutional coding network, i.e. the dimensions of ConvInRecon are consistent with ConvIn.
Optionally, the training the automatic coding network by using the time-frequency information of the training complex value timing sequence signal to obtain a training automatic coding network includes:
s41, training the automatic coding network by using the time-frequency information of the training complex value time sequence signal to obtain reconstructed frequency spectrum information;
s42, constructing a loss function by using the reconstructed spectrum information;
s43, calculating the gradient of the parameter to be optimized in the automatic coding and decoding network by using the loss function, and obtaining an updated parameter to be optimized by using a gradient descent method;
and S44, repeating S41-S43 until the automatic coding network is converged to obtain the training automatic coding network.
Optionally, the time-frequency information of the training complex value timing sequence signal is input, and the time-frequency information passes through the convolutional coding network and the deconvolution decoding network in sequence to obtain an output, that is, reconstructed frequency spectrum information. From its output, a loss function of the network output relative to the samples can then be calculated. As the mean squared error loss function:
in the above formula, L represents the value of the loss function, N is the number of samples per batch, x i Representing the ith input sample in a batch,
and the output of the coding and decoding network corresponding to the ith input sample in the batch is represented, and | | · | | represents the Euclidean distance between two one-dimensional vectors.
Then, using an automatic derivation mechanism in the torch, the gradient of the loss function with respect to each parameter to be optimized in the automatically encoded network is calculated, and the respective parameters are updated by a gradient descent method:
wherein, theta is any parameter to be optimized in the automatic coding network, theta' is the value of the parameter to be optimized theta after gradient descent updating,
is the gradient of the loss function relative to each parameter to be optimized in the automatic coding network, and eta is the learning rate, belongs to a hyper-parameter, and can be adjusted, for example, to 0.001.
And finally, repeating and iterating the loss function calculation and gradient descending process for a plurality of times until convergence, and obtaining the training automatic coding network.
Optionally, the processing the test complex value time sequence signal by using the training convolutional coding network to obtain hidden space vector information includes:
s51, acquiring a test complex value time sequence signal, and preprocessing the test complex value time sequence signal to obtain time frequency information of the test complex value time sequence signal;
and S52, processing the time-frequency information of the test complex value time sequence signal by using the training convolutional coding network to obtain hidden space vector information.
Optionally, the obtained time-frequency information of the test complex value time sequence signal is X, and the calculated X is output only by training a convolutional coding network, that is, the dimension of the Z is:
Z:＝[Time,LatentDim]
where Time represents the frame number of the test data spectrogram, and LatentDim represents the length of the hidden vector.
Optionally, the training of the gaussian mixture long-short term memory network by using the implicit spatial vector information to obtain a trained gaussian mixture long-short term memory network includes:
s61, constructing a Gaussian mixture long-time and short-time memory network;
s62, processing the Gaussian mixture long-time memory network by using the hidden space vector information to obtain a mixture proportion, a mean value and a standard deviation of Gaussian distribution;
s63, constructing a Gaussian mixture long-time memory network loss function by using the mixing proportion, the mean value and the standard deviation of the Gaussian distribution;
s64, calculating the gradient of the parameter to be optimized in the Gaussian mixture long-short time memory network by using the loss function of the Gaussian mixture long-short time memory network, and updating the parameter to be optimized in the Gaussian mixture long-short time memory network by using a gradient descent method;
and S65, repeating S61-S64 until the Gaussian mixture long-time and short-time memory network converges to obtain the training Gaussian mixture long-time and short-time memory network.
Optionally, input data of the gaussian mixture long-and-short term memory network is denoted as GMLSTMIn, which is a three-dimensional tensor expressed as:
GMLSTMIn:＝[Batch,Sequence,LatentDim]
where Batch represents the number of samples in each Batch, sequence represents the length of a time series in a sample, representing how long the number of past spectra was used each time the information of the spectra at the next time is predicted, latentDim represents the length of a hidden vector, consistent with the previous definition.
The LSTM layer LSTM0 and the three fully connected layers Fc1, fc2, fc3 are defined, which steps can be implemented using the torch.
lstm0 has input GMLSTMIn and output with three terms, LSTMOut, hidden, and Cell, expressed as:
LSTMOut,(Hidden,Cell)＝lstm0(GMLSTMIn)
this step can be done by the torch.nn.lstm function, where the only items to be utilized are Hidden, whose dimensions are:
Hidden:＝[Batch,HiddenDim]
where HiddenDim is a parameter defined when defining the LSTM network, it may be an appropriate value, such as equal to LatentDim.
The input of Fc1, fc2, fc3 is Hidden, the output is Fc1Out, fc2Out, fc3Out respectively, the dimensionality of all three is [ Batch, numGaussian LatensDim ], wherein NumGaussian represents the Gaussian distribution quantity that will form the final mixed distribution. The process is represented as:
Fc1Out＝F.softmax(Fc1(Hidden))
Fc2Out＝Fc2(Hidden)
Fc3Out＝Fc3(Hidden)
where f.softmax is a function in the mouth. Nn. Functional, the values of the first dimension of a tensor can be normalized (i.e., the first dimension sum is one).
Finally, dimension conversion is carried Out on the Fc1Out, the Fc2Out and the Fc3Out, the process can be completed through a function torch, a mean value mu and a standard deviation sigma are recorded as converted outputs, the dimensions of the converted outputs are [ Batch, numGaussian and Latensdim ], and the mixed proportion of the NumGaussian Gaussian distributions, the mean value of the NumGaussian Gaussian distributions and the standard deviation of the NumGaussian Gaussian distributions are respectively represented for each component of the coding hidden space vector.
Optionally, the method for training the gaussian mixture long-and-short term memory network includes:
firstly, a time Sequence of code vector data with a time length of Sequence and a code vector corresponding to a frequency spectrum at the next moment of the data are input by a network as labels, and three outputs pi, mu and sigma are obtained through a Gaussian mixture long-time memory network. Then, based on the three parameters pi, μ, σ, a loss function with respect to the next time instance code vector label y can be calculated, which is more specific and is described as:
in the above formula, L represents the value of the loss function, N is the number of samples per batch, π k ， i A k component, mu, representing the mixture ratio of the i-th sample in a batch through the Gaussian distribution of the network output k，i A k component, O, representing the mean of the Gaussian distribution of the ith sample over the network output in the batch k，i A k-th component representing the standard deviation of the gaussian distribution of the i-th sample through the network output in a batch,
represents the specified mean value μ k (x) And variance->
The value of the probability density function corresponding to the y point in the normal distribution of (1).
Then, calculating the gradient of each parameter to be optimized in the loss function relative to the Gaussian mixture time memory network by using an automatic derivation mechanism in the torch, and updating each parameter by using a gradient descent method:
wherein theta is any parameter to be optimized in the Gaussian mixture long-time memory network, theta' is a value of the parameter to be optimized theta after gradient descent updating,
the gradient of each parameter to be optimized in the memory network is memorized when the loss function is long relative to the Gaussian mixture time, eta is the learning rate, belongs to a hyper-parameter, and can be adjusted, for example, 0.001 is taken.
And finally, repeating and iterating the loss function calculation and gradient descent process for a plurality of times until convergence, and obtaining the training Gaussian mixture duration memory network.
Optionally, the processing the complex value time sequence signal to be detected by using the training gaussian mixture long-time and short-time memory network to obtain a spectrum anomaly detection result, including:
s71, acquiring a complex value time sequence signal to be detected;
s72, preprocessing the complex value time sequence signal to be detected to obtain time frequency information of the complex value time sequence signal to be detected;
s73, processing the complex value time sequence signal time frequency information to be detected by using the training convolutional coding network to obtain testing hidden space vector information;
and S74, processing the testing hidden space vector information by using the training Gaussian mixture long-time and short-time memory network to obtain a frequency spectrum abnormity detection result.
Optionally, a complex value time Sequence signal to be detected is obtained, the complex value time Sequence signal to be detected is preprocessed, time-frequency information of the complex value time Sequence signal to be detected is obtained, the training convolutional coding network is used for processing the time-frequency information of the complex value time Sequence signal to be detected, testing hidden space vector information is obtained, for each time, the time and previous Sequence testing hidden space vector information are input into a trained gaussian mixed long-short time memory network, and prediction of three parameters pi, mu and sigma of mixed distribution of coding vectors at the next time is obtained.
Traversing the test hidden space vector information, and calculating a loss function value obtained by the hidden space vector of the actual signal at each moment aiming at the hidden space vector mixed distribution of the network prediction in the last step. At the moment, a threshold value is determined, whether the calculated loss function value exceeds the threshold value or not is judged, and finally, the spectrum signal exceeding the threshold value is output as an abnormal signal, otherwise, the spectrum signal is output as a normal signal. As shown in fig. 3, the loss function value and the actual occurrence of an abnormality are shown, and the position with a higher value is an abnormal position.
Optionally, the automatic coding network includes a convolutional coding network and a deconvolution decoding network;
the automatic coding network comprises the following components:
wherein, x represents any time-frequency information input, E represents a convolution coding network, z represents a hidden vector obtained by the convolution coding network, D represents a deconvolution decoding network,
representing the output of the automatically encoded network.
Optionally, the loss function L of the automatic coding network is:
where | · | | represents the euclidean distance between two one-dimensional vectors.
Optionally, the graph convolution network may be trained by using the hidden space vector information to obtain a training graph convolution network; and processing the complex value time sequence signal to be detected by using a training graph convolution network to obtain a frequency spectrum anomaly detection result. The graph convolution neural network is a neural network model acting on graph data, and the graph convolution network based on a spectrum method firstly utilizes Fourier transform to convert convolution on a space domain into multiplication on a frequency domain, and finally uses inverse Fourier transform to convert characteristics back to the space domain.
The normalized laplacian matrix is real-symmetric and positive, and can be obtained by decomposing the laplacian matrix according to the matrix decomposition principle:
L＝UΛU T
wherein Λ is a diagonal matrix composed of eigenvalues of the Laplace matrix, and Λ is ii ＝λ i ，λ i Is a characteristic value; u is a matrix of eigenvectors arranged according to eigenvalues, and is also the basis of a Fourier transform, which is an orthogonal matrix, UU T = I, I is identity matrix, u 0 ,u 1 ,...,u n-1 For the elements in the matrix U:
U＝[u 0 ,u 1 ,...,u n-1 ]∈R N×N
ith node x in the graph i Is marked as x ∈ R N The fourier transform of the plot of signal x can then be represented in the form:
F(x)＝U T x
the inverse fourier transform is represented as:
wherein the content of the first and second substances,
is the output of the fourier transform of the plot of signal x.
Gg ∈ R of input signal x N The filtered graph convolution expression is:
y＝x*Gg＝F -1 (F(x)F(Gg))＝U(U T xU T Gg)
wherein Gg is a filter.
Define a form as Gg θ ＝diag(U T Gg), the form of graph convolution can be further simplified as:
y＝x*Gg θ ＝UGg θ U T x
this is a graph convolution network based on a spectral method.
Therefore, the method for detecting the spectrum anomaly, which is described by the embodiment of the invention, can be implemented by constructing an automatic encoder network and inputting the processed spectrum signal into the automatic encoder network for encoding; constructing a Gaussian mixture long-time and short-time memory network; inputting the coded time sequence signal into a Gaussian mixture long-time and short-time memory network; and calculating the confidence probability that the code of the spectrum signal at the next moment belongs to the distribution output by the memory network, and outputting an abnormity when the probability is lower. The method of the invention can not only code the high-dimensional frequency spectrum signal into a one-dimensional vector to accelerate the data processing efficiency, but also directly predict the signal distribution at the next moment by using the characteristics of the time sequence signal without relying on prior probability distribution knowledge, and simultaneously has strong robust capability to noise, environmental change and other interference factors.
Example two
Referring to fig. 4, fig. 4 is a schematic flowchart of a spectrum anomaly detection apparatus based on implicit spatial probability prediction according to an embodiment of the present invention. The spectrum anomaly detection device based on implicit spatial probability prediction described in fig. 4 may be applied to the technical field of artificial intelligence, such as network security, fault analysis, artificial intelligence, spectrum coordination, radio security management and control, regional electromagnetic detection and protection, and the embodiment of the present invention is not limited thereto. As shown in fig. 4, the spectrum anomaly detection apparatus based on implicit spatial probability prediction may include the following operations:
s301, a signal acquisition module is used for acquiring a training complex value time sequence signal; the training complex value time sequence signal comprises a plurality of pieces of training complex value time sequence signal data;
s302, a preprocessing module, configured to preprocess the training complex value timing signal to obtain time-frequency information of the training complex value timing signal;
s303, an automatic coding network construction module is used for constructing an automatic coding network; the automatic coding network comprises a convolution coding network and a deconvolution decoding network;
s304, an automatic coding network training module is used for training the automatic coding network by using the time-frequency information of the training complex value time sequence signal to obtain a training automatic coding network; the training automatic coding network comprises a training convolutional coding network and a training deconvolution decoding network;
s305, an automatic coding module is used for obtaining a test complex value time sequence signal, and processing the test complex value time sequence signal by using the training convolutional coding network to obtain hidden space vector information;
s306, a Gaussian mixture long-short time memory network training module used for training the Gaussian mixture long-short time memory network by using the implicit space vector information to obtain a training Gaussian mixture long-short time memory network;
and S307, a frequency spectrum abnormity detection module is used for acquiring a complex value time sequence signal to be detected, and processing the complex value time sequence signal to be detected by using the training Gaussian mixture long-short time memory network to obtain a frequency spectrum abnormity detection result.
Optionally, the preprocessing is performed on the training complex value timing signal to obtain time-frequency information of the training complex value timing signal, and the method includes:
s21, performing windowing operation on the training complex value time sequence signal to obtain a window selection training complex value time sequence signal;
and S22, carrying out time-frequency transformation on the window selection training complex value time sequence signal to obtain time-frequency information of the training complex value time sequence signal.
Optionally, training the automatic coding network by using the time-frequency information of the training complex value timing sequence signal to obtain a training automatic coding network, including:
s41, training the automatic coding network by using the time-frequency information of the training complex value time sequence signal to obtain reconstructed frequency spectrum information;
s42, constructing a loss function by using the reconstructed spectrum information;
s43, calculating the gradient of the parameter to be optimized in the automatic coding network by using the loss function, and obtaining an updated parameter to be optimized by using a gradient descent method;
and S44, repeating S41-S43 until the automatic coding network is converged to obtain the training automatic coding network.
Optionally, the processing the test complex value time sequence signal by using the training convolutional coding network to obtain hidden space vector information, including:
s51, acquiring a test complex value time sequence signal, and preprocessing the test complex value time sequence signal to obtain time frequency information of the test complex value time sequence signal;
and S52, processing the time-frequency information of the test complex value time sequence signal by using the training convolutional coding network to obtain hidden space vector information.
Optionally, training the gaussian mixture long-time and short-time memory network by using the implicit space vector information to obtain a trained gaussian mixture long-time and short-time memory network, including:
s61, constructing a Gaussian mixture long-time and short-time memory network;
s62, processing the Gaussian mixture long-time memory network by using the hidden space vector information to obtain a mixture ratio, a mean value and a standard deviation of Gaussian distribution;
s63, constructing a Gaussian mixture long-time memory network loss function by using the mixing proportion, the mean value and the standard deviation of the Gaussian distribution;
s64, calculating the gradient of the parameter to be optimized in the Gaussian mixture long-short time memory network by using the loss function of the Gaussian mixture long-short time memory network, and updating the parameter to be optimized in the Gaussian mixture long-short time memory network by using a gradient descent method;
and S65, repeating S61-S64 until the Gaussian mixture long-time and short-time memory network converges to obtain the training Gaussian mixture long-time and short-time memory network.
Optionally, the training gaussian mixture long-time and short-time memory network is used to process the complex value time sequence signal to be detected, so as to obtain a spectrum anomaly detection result, including:
s71, acquiring a complex value time sequence signal to be detected;
s72, preprocessing the complex value time sequence signal to be detected to obtain time frequency information of the complex value time sequence signal to be detected;
s73, processing the complex value time sequence signal time frequency information to be detected by using the training convolutional coding network to obtain testing hidden space vector information;
and S74, processing the test hidden space vector information by using the training Gaussian mixture long-time and short-time memory network to obtain a frequency spectrum abnormity detection result.
Optionally, the automatic coding network includes a convolutional coding network and a deconvolution decoding network;
the automatic coding network comprises the following steps:
wherein, x represents any time-frequency information input, E represents a convolutional coding network, z represents a hidden vector obtained by the convolutional coding network, D represents a deconvolution decoding network,
representing the output of the automatically encoded network.
EXAMPLE III
Referring to fig. 5, fig. 5 is a schematic flowchart illustrating another spectrum anomaly detection apparatus based on implicit spatial probability prediction according to an embodiment of the present invention. The spectrum anomaly detection device based on implicit spatial probability prediction described in fig. 5 may be applied to the technical field of artificial intelligence, such as network security, fault analysis, artificial intelligence, spectrum coordination, radio security management and control, regional electromagnetic detection and protection, and the embodiment of the present invention is not limited thereto. As shown in fig. 5, the spectrum anomaly detection apparatus based on implicit spatial probability prediction may include the following operations:
a memory 401 storing executable program code;
a processor 402 coupled with the memory 401;
the processor 402 calls the executable program code stored in the memory 401 for executing the steps of the spectrum abnormality detection method based on implicit spatial probability prediction described in the first embodiment.
Example four
The embodiment of the invention discloses a computer-readable storage medium which stores a computer program for electronic data exchange, wherein the computer program enables a computer to execute the steps of the spectrum anomaly detection method based on hidden space probability prediction.
The above-described embodiments of the apparatus are merely illustrative, and the modules described as separate parts may or may not be physically separate, and the parts displayed as modules may or may not be physical modules, may be located in one place, or may be distributed on a plurality of network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the present embodiment. One of ordinary skill in the art can understand and implement it without inventive effort.
Through the above detailed description of the embodiments, those skilled in the art will clearly understand that the embodiments may be implemented by software plus a necessary general hardware platform, and may also be implemented by hardware. Based on such understanding, the above technical solutions may be embodied in the form of a software product, which may be stored in a computer-readable storage medium, wherein the storage medium includes a Read-Only Memory (ROM), a Random Access Memory (RAM), a Programmable Read-Only Memory (PROM), an Erasable Programmable Read-Only Memory (EPROM), a One-time Programmable Read-Only Memory (OTPROM), an Electrically Erasable Programmable Read-Only Memory (EEPROM), a Compact Disc-Read-Only Memory (CD-ROM) or other Memory capable of storing data, a magnetic tape, or any other computer-readable medium capable of storing data.
Finally, it should be noted that: the method and the device for detecting spectrum anomaly based on implicit spatial probability prediction disclosed in the embodiments of the present invention are only preferred embodiments of the present invention, and are only used for illustrating the technical solution of the present invention, not for limiting the same; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art; the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and the modifications or the substitutions do not make the essence of the corresponding technical solutions depart from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims (10)
1. A spectrum anomaly detection method based on hidden space probability prediction is characterized by comprising the following steps:
s1, acquiring a complex value time sequence signal; the complex value time sequence signal comprises a training complex value time sequence signal, a testing complex value time sequence signal and a complex value time sequence signal to be detected;
s2, preprocessing the training complex value time sequence signal to obtain time-frequency information of the training complex value time sequence signal;
s3, constructing an automatic coding network; the automatic coding network comprises a convolution coding network and a deconvolution decoding network;
s4, training the automatic coding network by using the time-frequency information of the training complex value time sequence signal to obtain a training automatic coding network; the training automatic coding network comprises a training convolution coding network and a training deconvolution decoding network;
s5, processing the test complex value time sequence signal by using the training convolutional coding network to obtain hidden space vector information;
s6, training the Gaussian mixture long-time and short-time memory network by using the hidden space vector information to obtain a training Gaussian mixture long-time and short-time memory network;
and S7, processing the complex value time sequence signal to be detected by using the training Gaussian mixture long-time and short-time memory network to obtain a frequency spectrum abnormity detection result.
2. The spectrum anomaly detection method based on implicit spatial probability prediction according to claim 1, wherein the preprocessing the training complex value timing signal to obtain the time-frequency information of the training complex value timing signal includes:
s21, performing windowing operation on the training complex value time sequence signal to obtain a window selection training complex value time sequence signal;
and S22, carrying out time-frequency transformation on the window selection training complex value time sequence signal to obtain time-frequency information of the training complex value time sequence signal.
3. The spectrum anomaly detection method based on implicit spatial probability prediction according to claim 1, wherein the training of the automatic coding network by using the time-frequency information of the training complex value timing signal to obtain a training automatic coding network comprises:
s41, training the automatic coding network by using the time-frequency information of the training complex value time sequence signal to obtain reconstructed frequency spectrum information;
s42, constructing a loss function by using the reconstructed spectrum information;
s43, calculating the gradient of the parameter to be optimized in the automatic coding network by using the loss function, and obtaining an updated parameter to be optimized by using a gradient descent method;
and S44, repeating S41-S43 until the automatic coding network is converged to obtain the training automatic coding network.
4. The spectrum anomaly detection method based on implicit spatial probability prediction according to claim 1, wherein the processing the test complex value timing signal by using the training convolutional coding network to obtain implicit spatial vector information comprises:
s51, acquiring a test complex value time sequence signal, and preprocessing the test complex value time sequence signal to obtain time frequency information of the test complex value time sequence signal;
and S52, processing the time-frequency information of the test complex value time sequence signal by using the training convolutional coding network to obtain hidden space vector information.
5. The spectrum anomaly detection method based on implicit spatial probability prediction according to claim 1, wherein the training of the gaussian mixture long-short term memory network by using the implicit spatial vector information to obtain the trained gaussian mixture long-short term memory network comprises:
s61, constructing a Gaussian mixture long-time and short-time memory network;
s62, processing the Gaussian mixture long-time memory network by using the hidden space vector information to obtain a mixture ratio, a mean value and a standard deviation of Gaussian distribution;
s63, constructing a Gaussian mixture long-time memory network loss function by using the mixing proportion, the mean value and the standard deviation of the Gaussian distribution;
s64, calculating the gradient of the parameter to be optimized in the Gaussian mixture long-short time memory network by using the loss function of the Gaussian mixture long-short time memory network, and updating the parameter to be optimized in the Gaussian mixture long-short time memory network by using a gradient descent method;
and S65, repeating S61-S64 until the Gaussian mixture long-time and short-time memory network converges to obtain the training Gaussian mixture long-time and short-time memory network.
6. The spectrum anomaly detection method based on implicit spatial probability prediction according to claim 1, wherein the processing of the complex value time sequence signal to be detected by using the training gaussian mixture long-and-short-term memory network to obtain a spectrum anomaly detection result comprises:
s71, acquiring a complex value time sequence signal to be detected;
s72, preprocessing the complex value time sequence signal to be detected to obtain time frequency information of the complex value time sequence signal to be detected;
s73, processing the complex value time sequence signal time frequency information to be detected by using the training convolutional coding network to obtain testing hidden space vector information;
and S74, processing the test implicit space vector information by using the training Gaussian mixture long-time and short-time memory network to obtain a spectrum anomaly detection result.
7. The spectrum anomaly detection method based on implicit spatial probability prediction according to claim 1, characterized in that the automatic coding network comprises a convolutional coding network and a deconvolution decoding network;
the automatic coding network comprises the following steps:
wherein, x represents any time-frequency information input, E represents a convolutional coding network, z represents a hidden vector obtained by the convolutional coding network, D represents a deconvolution decoding network,
representing the output of the automatically encoded network.
8. An apparatus for detecting spectrum abnormality based on implicit spatial probability prediction, the apparatus comprising:
the signal acquisition module is used for acquiring a complex value time sequence signal; the complex value time sequence signal comprises a training complex value time sequence signal, a testing complex value time sequence signal and a complex value time sequence signal to be detected;
the preprocessing module is used for preprocessing the training complex value time sequence signal to obtain time-frequency information of the training complex value time sequence signal;
the building module is used for building an automatic coding network; the automatic coding network comprises a convolution coding network and a deconvolution decoding network;
the first training module is used for training the automatic coding network by using the time-frequency information of the training complex value time sequence signal to obtain a training automatic coding network; the training automatic coding network comprises a training convolutional coding network and a training deconvolution decoding network;
the automatic coding module is used for processing the test complex value time sequence signal by utilizing the training convolutional coding network to obtain hidden space vector information;
the second training module is used for training the Gaussian mixture long-time and short-time memory network by using the implicit space vector information to obtain a training Gaussian mixture long-time and short-time memory network;
and the detection module is used for processing the complex value time sequence signal to be detected by utilizing the training Gaussian mixture long-time and short-time memory network to obtain a frequency spectrum abnormity detection result.
9. An apparatus for detecting spectrum abnormality based on implicit spatial probability prediction, the apparatus comprising:
a memory storing executable program code;
a processor coupled with the memory;
the processor calls the executable program code stored in the memory to execute the spectrum abnormality detection method based on the implicit spatial probability prediction according to any one of claims 1 to 7.
10. A computer-readable medium storing computer instructions which, when invoked, perform the spectral anomaly detection method based on implicit spatial probability prediction according to any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211461149.0A CN115856425B (en) | 2022-11-21 | 2022-11-21 | Spectrum anomaly detection method and device based on hidden space probability prediction |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211461149.0A CN115856425B (en) | 2022-11-21 | 2022-11-21 | Spectrum anomaly detection method and device based on hidden space probability prediction |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115856425A true CN115856425A (en) | 2023-03-28 |
| CN115856425B CN115856425B (en) | 2023-10-17 |
Family
ID=85664587
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211461149.0A Active CN115856425B (en) | 2022-11-21 | 2022-11-21 | Spectrum anomaly detection method and device based on hidden space probability prediction |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115856425B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116976684A (en) * | 2023-09-25 | 2023-10-31 | 尚古智造(山东)智能装备有限公司 | Risk model predictive control method and system for logistics conveyor |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140254392A1 (en) * | 2013-03-05 | 2014-09-11 | Comcast Cable Communications, Llc | Network implementation of spectrum analysis |
| CN109034127A (en) * | 2018-08-31 | 2018-12-18 | 中国电子科技集团公司第三十六研究所 | A kind of frequency spectrum method for detecting abnormality, device and electronic equipment |
| CN109933741A (en) * | 2019-02-27 | 2019-06-25 | 京东数字科技控股有限公司 | User network behaviors feature extracting method, device and storage medium |
| US20190377979A1 (en) * | 2017-08-30 | 2019-12-12 | Tencent Technology (Shenzhen) Company Limited | Image description generation method, model training method, device and storage medium |
| CN110632386A (en) * | 2019-08-16 | 2019-12-31 | 山东大学 | Solar radio interference filtering method, readable storage medium and electronic equipment |
| CN113658583A (en) * | 2021-08-17 | 2021-11-16 | 安徽大学 | Method, system and device for converting ear voice based on generation countermeasure network |
| CN114254673A (en) * | 2021-12-16 | 2022-03-29 | 嘉兴深智科技有限公司 | Denoising countermeasure self-encoder-based spectrum anomaly detection method |
| CN114567398A (en) * | 2022-04-11 | 2022-05-31 | 浙江工业大学 | Frequency spectrum sensing method based on convolution long-time and short-time memory neural network |
| CN114897144A (en) * | 2022-05-13 | 2022-08-12 | 苏州大学 | Complex value time sequence signal prediction method based on complex value neural network |
-
2022
- 2022-11-21 CN CN202211461149.0A patent/CN115856425B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140254392A1 (en) * | 2013-03-05 | 2014-09-11 | Comcast Cable Communications, Llc | Network implementation of spectrum analysis |
| US20190377979A1 (en) * | 2017-08-30 | 2019-12-12 | Tencent Technology (Shenzhen) Company Limited | Image description generation method, model training method, device and storage medium |
| CN109034127A (en) * | 2018-08-31 | 2018-12-18 | 中国电子科技集团公司第三十六研究所 | A kind of frequency spectrum method for detecting abnormality, device and electronic equipment |
| CN109933741A (en) * | 2019-02-27 | 2019-06-25 | 京东数字科技控股有限公司 | User network behaviors feature extracting method, device and storage medium |
| CN110632386A (en) * | 2019-08-16 | 2019-12-31 | 山东大学 | Solar radio interference filtering method, readable storage medium and electronic equipment |
| CN113658583A (en) * | 2021-08-17 | 2021-11-16 | 安徽大学 | Method, system and device for converting ear voice based on generation countermeasure network |
| CN114254673A (en) * | 2021-12-16 | 2022-03-29 | 嘉兴深智科技有限公司 | Denoising countermeasure self-encoder-based spectrum anomaly detection method |
| CN114567398A (en) * | 2022-04-11 | 2022-05-31 | 浙江工业大学 | Frequency spectrum sensing method based on convolution long-time and short-time memory neural network |
| CN114897144A (en) * | 2022-05-13 | 2022-08-12 | 苏州大学 | Complex value time sequence signal prediction method based on complex value neural network |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116976684A (en) * | 2023-09-25 | 2023-10-31 | 尚古智造(山东)智能装备有限公司 | Risk model predictive control method and system for logistics conveyor |
| CN116976684B (en) * | 2023-09-25 | 2024-01-02 | 尚古智造(山东)智能装备有限公司 | Risk model predictive control method and system for logistics conveyor |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115856425B (en) | 2023-10-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Murakami et al. | Eigenvector spatial filtering for large data sets: fixed and random effects approaches | |
| Zhang et al. | A novel decomposition‐ensemble model for forecasting short‐term load‐time series with multiple seasonal patterns | |
| CN109841226A (en) | A kind of single channel real-time noise-reducing method based on convolution recurrent neural network | |
| Kohonen | Generalizations of the self-organizing map | |
| Huang | Mechanisms of dimensionality reduction and decorrelation in deep neural networks | |
| Xing et al. | Traffic flow decomposition and prediction based on robust principal component analysis | |
| CN115856425B (en) | Spectrum anomaly detection method and device based on hidden space probability prediction | |
| CN115964670B (en) | Spectrum anomaly detection method | |
| CN109471049B (en) | Satellite power supply system anomaly detection method based on improved stacked self-encoder | |
| CN114550223B (en) | Person interaction detection method and device and electronic equipment | |
| CN110851654A (en) | Industrial equipment fault detection and classification method based on tensor data dimension reduction | |
| CN113591728A (en) | Electric energy quality disturbance classification method based on integrated deep learning | |
| CN115600101A (en) | Unmanned aerial vehicle signal intelligent detection method and device based on priori knowledge | |
| CN115426671A (en) | Method, system and equipment for graph neural network training and wireless cell fault prediction | |
| EP3816805A1 (en) | Analysis of anomalies in a facility | |
| Deng et al. | Power system stability assessment method based on GAN and GRU‐Attention using incomplete voltage data | |
| Ren | The advance of generative model and variational autoencoder | |
| Sun et al. | Caae: A novel wireless spectrum anomaly detection method with multiple scoring criterion | |
| Deco et al. | Statistical-ensemble theory of redundancy reduction and the duality between unsupervised and supervised neural learning | |
| Mirmomeni et al. | Long-term prediction of solar and geomagnetic activity daily time series using singular spectrum analysis and fuzzy descriptor models | |
| EP4156036A1 (en) | Variational auto-encoder for multivariate time series generation | |
| Menges et al. | Computationally and Memory-Efficient Robust Predictive Analytics Using Big Data | |
| Cardoso | Financial Time Series Forecasting Using Artificial Neural Networks | |
| Guerra et al. | A 3D Dynamical Hybrid Model: Coupling Statistical And Machine Learning Techniques To SAR Time Series Prediction | |
| CN115100599A (en) | Mask transform-based semi-supervised crowd scene abnormality detection method |
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 |