A method and system for classifying low signal-to-noise ratio ultra-short time signals

Abstract

The application discloses a kind of low signal-to-noise ratio ultra-short time signal classification method and system, and classification method includes the following process: the input sub-sample obtained is extracted, and the feature vector of multiple double-point pairs is obtained, the feature vector is composed of the feature amplitude of multiple double-point pairs;Utilize the feature vector of multiple double-point pairs to classify the low signal-to-noise ratio ultra-short time signal;The acquisition process of the input sub-sample includes: the double-point pair of the narrowband signal of low signal-to-noise ratio ultra-short time signal is sampled, and input sub-sample is obtained;The double-point pair is any two sampling points of low signal-to-noise ratio ultra-short time signal.The application can classify low signal-to-noise ratio ultra-short time signal in real time.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of time series signal analysis, and particularly relates to a classification method and system for low signal-to-noise ratio ultra-short time signals. BACKGROUND

[0002] The monitoring and control system based on time series signal classification has a wide range of application scenarios in life and production, for example, the health status of a patient is determined by real-time monitoring of electrocardiogram; an external device is controlled by classifying electroencephalogram signals and converting them into instructions; the working state of a gear bearing is determined by the vibration time series of the gear bearing, so as to realize fault warning; and a power grid fault is determined by a voltage and current sensing array.

[0003] In recent years, deep learning networks have been applied to the prediction and classification of various time series signals, showing excellent performance and being deployed in practical scenarios. However, time series signals collected in a high-noise environment have extremely low signal-to-noise ratios, which are difficult to be quickly classified by various classification models based on deep learning networks. Specifically, a classification model needs to extract useful information from a low signal-to-noise ratio signal of sufficient length to obtain satisfactory classification accuracy, which leads to significant lag in monitoring and control. This thorny problem is particularly prominent in a brain-computer interface (BCI) based on motor imagery electroencephalogram (EEG) signals.

[0004] EEG is a typical low signal-to-noise ratio signal. Motor imagery EEG signals are collected when a subject imagines different movements, and are converted into different control instructions after classification and identification. At present, most studies focus on the classification of samples of about 3 seconds in length. This means that when a user of a BCI wants to control an external device, he or she needs to imagine a certain action for 3 seconds. Such lagged control cannot meet the needs of most practical scenarios, and the user experience is extremely poor. A few researchers have tried to classify samples of about 1.5 seconds, but have not obtained satisfactory results. SUMMARY

[0005] In view of the problems existing in the existing monitoring and control system based on time series signal classification, the purpose of the present application is to provide a classification method and system for low signal-to-noise ratio ultra-short time signals, which can classify low signal-to-noise ratio ultra-short time signals in real time.

[0006] The technical solution adopted by the present application is as follows:

[0007] A classification method for low signal-to-noise ratio ultra-short time signals, comprising the following processes:

[0008] The double-point pairs of the narrowband signals of the low signal-to-noise ratio ultra-short time signals are sampled, and a feature vector is calculated according to the sampling values; the double-point pairs are any two sampling points of the low signal-to-noise ratio ultra-short time signals, and the feature vector is composed of feature amplitudes of multiple double-point pairs;

[0009] The feature vector is used to classify the low signal-to-noise ratio ultra-short time signal.

[0010] Preferably, the method for classifying the low signal-to-noise ratio ultra-short time signal further comprises the following process: decomposing the low signal-to-noise ratio ultra-short time signal into a plurality of narrow-band components.

[0011] When the two-point pair of the narrow-band signal of the low signal-to-noise ratio ultra-short time signal is sampled, the two-point pair of the narrow-band component obtained by the decomposition is used for sampling.

[0012] Preferably, the method for classifying the low signal-to-noise ratio ultra-short time signal further comprises a pre-processing process of the low signal-to-noise ratio ultra-short time signal, and when the low signal-to-noise ratio ultra-short time signal is pre-processed, the direct current bias and the fixed frequency noise of the interference in the low signal-to-noise ratio ultra-short time signal are filtered out.

[0013] When the low signal-to-noise ratio ultra-short time signal is a wide-band signal, the low signal-to-noise ratio ultra-short time signal after the pre-processing is decomposed into a plurality of narrow-band components, and then the two-point pair of the narrow-band component obtained by the decomposition is used for sampling.

[0014] When the low signal-to-noise ratio ultra-short time signal is a narrow-band signal, the low signal-to-noise ratio ultra-short time signal after the pre-processing is decomposed into a plurality of narrow-band components, and then the two-point pair of the narrow-band component obtained by the decomposition is used for sampling.

[0015] Or, when the low signal-to-noise ratio ultra-short time signal is a narrow-band signal, the two-point pair of the low signal-to-noise ratio ultra-short time signal after the pre-processing is directly sampled.

[0016] In the above description, the narrow-band component is the narrow-band signal obtained by the decomposition.

[0017] Preferably, the feature amplitude A of the two-point pair is calculated by the following formula:

[0018]

[0019] Wherein, p1 and p2 respectively represent the sampling values of the two sampling points in the two-point pair, ω represents the central angular frequency of the narrow-band signal, and Δt represents the time interval of the two sampling points in the two-point pair.

[0020] The application further provides a classification system of low signal-to-noise ratio ultra-short time signal, comprising:

[0021] The feature extraction module samples the two-point pair of the narrow-band signal of the low signal-to-noise ratio ultra-short time signal, and calculates the feature vector according to the sampling value; the two-point pair is any two sampling points of the low signal-to-noise ratio ultra-short time signal, and the feature vector is composed of the feature amplitudes of a plurality of two-point pairs;

[0022] The classification module is used to classify the low signal-to-noise ratio ultra-short time signal by using the feature vector.

[0023] Preferably, the feature extraction module comprises a grouped dilated convolution layer, a square activation function and a square root activation function.

[0024] The grouped dilated convolution layer is used to extract a plurality of continuous double-point pairs from a single narrowband signal, each group in the grouped dilated convolution layer has two convolution kernels, and the convolution kernel hole rate is set as Δn=f s ×Δt, the time interval Δt of all double-point pairs is equal, and the feature amplitude calculation process of any narrowband signal is as follows:

[0025]

[0026] Wherein, v 11 and v 12 respectively represent two trainable parameters of the first convolution kernel corresponding to each narrowband signal, which can be initialized as or randomly initialized; v 21 and v 22 respectively represent two trainable parameters of the second convolution kernel corresponding to each narrowband signal, which can be initialized as and or randomly initialized, and ω represents the center angular frequency of the narrowband signal.

[0027] Preferably, the low signal-to-noise ratio ultra-short time signal classification system of the application further comprises:

[0028] A decomposition module is used to decompose the low signal-to-noise ratio ultra-short time signal into a plurality of narrowband components; when sampling the double-point pairs of the narrowband signals of the low signal-to-noise ratio ultra-short time signal, the double-point pairs of the narrowband components obtained by the above decomposition are used for sampling.

[0029] Preferably, the decomposition module comprises a frequency domain decomposition convolution layer, and the frequency domain decomposition convolution layer decomposes the low signal-to-noise ratio ultra-short time signal into a plurality of narrowband signals through a group of time domain bandpass filters.

[0030] The convolution kernel of the frequency domain decomposition convolution layer is initialized as a filter bank, and the filter bank contains a plurality of finite impulse response bandpass filters with a preset length, and the calculation formula of the finite impulse response bandpass filter h[n] is as follows:

[0031]

[0032] Wherein, f h represents the upper limit cutoff frequency of the narrowband signal, f l represents the lower limit cutoff frequency of the narrowband signal, f s represents the sampling frequency, n is the index of the array h[n], n=0, 1, 2, …, N-1, and N is the length of the finite impulse response bandpass filter.

[0033] Preferably, the low signal-to-noise ratio ultra-short time signal classification system of the application further comprises:

[0034] The preprocessing module is used to filter the fixed frequency noise and the DC bias in the low signal-to-noise ratio ultra-short time signal;

[0035] When the low signal-to-noise ratio ultra-short time signal is a wideband signal, the preprocessed low signal-to-noise ratio ultra-short time signal is decomposed into multiple narrowband components by the decomposition module, and the two-point pairs of the narrowband components obtained by decomposition are sampled by the feature extraction module;

[0036] When the low signal-to-noise ratio ultra-short time signal is a narrowband signal, the preprocessed low signal-to-noise ratio ultra-short time signal is decomposed into multiple narrowband components by the decomposition module, and the two-point pairs of the narrowband components obtained by decomposition are sampled by the feature extraction module;

[0037] Alternatively, when the low signal-to-noise ratio ultra-short time signal is a narrowband signal, the preprocessed low signal-to-noise ratio ultra-short time signal is transmitted to the feature extraction module by the preprocessing module, and the two-point pairs of the preprocessed low signal-to-noise ratio ultra-short time signal are sampled by the feature extraction module.

[0038] Preferably, the classification module predicts a sub-label for the feature vector through a fully connected layer and a Softmax activation function, and classifies the low signal-to-noise ratio ultra-short time signal through all sub-labels.

[0039] The application has the following beneficial effects:

[0040] In the low signal-to-noise ratio signal classification problem, the signal to be classified is cropped into multiple segments to extract features and then combined, which is a common technical means. However, when the length of the low signal-to-noise ratio signal itself is too short, it is not possible to obtain enough and long enough segments through cropping. The segments obtained by cropping the low signal-to-noise ratio short time signal are too short, so that high-quality features cannot be obtained through existing feature extraction methods, significantly reducing the classification accuracy. To solve the above problems, the two-point pair composed of any two sampling values of the low signal-to-noise ratio short time signal is used to replace the segments obtained by cropping. The application can fully utilize the information of the two sampling values to extract features that can represent the low signal-to-noise ratio ultra-short time signal, and then obtain a classification result with high enough confidence. Specifically, for the low signal-to-noise ratio problem of the signal, the two-point pairs sampled at different times contain the same useful information for classification and different noise interference information. The features extracted by the algorithm combined with multiple two-point pairs can highlight the same useful information while suppressing the noise interference information. At the same time, even for ultra-short time signals, enough two-point pairs can be sampled to form a large data set, which is more conducive to the training of the classification system, and then the optimal parameters are obtained. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 is a flow chart of the low signal-to-noise ratio ultra-short time signal classification method of the present application;

[0042] Figure 2 is a structural block diagram of the low signal-to-noise ratio ultra-short time signal classification system of the present application;

[0043] Figure 3 is an example diagram of the electroencephalogram signal collection in the embodiment of the present application;

[0044] Figure 4 is a structural diagram of the low signal-to-noise ratio ultra-short time signal classification system of the motor imagery electroencephalogram signal provided by the embodiment of the present application. DETAILED DESCRIPTION

[0045] In order to make the object, technical scheme and advantages of the present application clearer, the present application is further described in detail below with reference to the embodiments and drawings. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0046] Referring to Figure 1 The low signal-to-noise ratio ultra-short time signal classification method of the present application includes the following processes:

[0047] The two-point pair of the narrowband signal of the low signal-to-noise ratio ultra-short time signal is sampled, and the feature vector is calculated according to the sampling value, and the process is as follows:

[0048] Two points p1 and p2 (two-point pair) are collected from the arbitrary narrowband signal of the low signal-to-noise ratio ultra-short time signal, and the feature amplitude A of the two-point pair is solved based on the following formula (1):

[0049]

[0050] Wherein, ω represents the center frequency of the narrowband signal, and Δt represents the time interval of the two sampling points in the two-point pair. Affected by noise interference, the amplitude of the narrowband signal changes over time, so different feature amplitudes are solved based on different two-point pairs. These solved feature amplitudes collectively represent the characteristics of the narrowband signal and are collectively used to identify the narrowband signal.

[0051] When the low signal-to-noise ratio ultra-short time signal is a wideband signal, the low signal-to-noise ratio ultra-short time signal needs to be decomposed into multiple narrowband signals. When decomposing, a set of bandpass filters is used to decompose the low signal-to-noise ratio ultra-short time signal into multiple narrowband components, and each narrowband component is a narrowband signal. The bandpass filter can be realized based on a filter circuit before signal analog-to-digital conversion, or realized based on a digital filter after signal analog-to-digital conversion. When extracting features, the feature amplitudes of each narrowband signal are extracted respectively, and the amplitudes of all narrowband signals collectively constitute the characteristics of the narrowband signal.

[0052] In addition, for the case that the low signal-to-noise ratio ultra-short time signal is a narrowband signal, the narrowband signal can be directly subjected to feature extraction, or the low signal-to-noise ratio ultra-short time signal is further decomposed into multiple narrowband signals, and the narrowband signals obtained by the decomposition are subjected to feature extraction again.

[0053] In the present application, for a signal with extremely short duration, a large number of feature vectors can be obtained by increasing the sampling rate; for a low signal-to-noise ratio signal, the large number of feature vectors are combined to represent the low signal-to-noise ratio signal, which can highlight the information beneficial to classification and effectively suppress noise.

[0054] For some low signal-to-noise ratio ultra-short time signals (such as motor imagery electroencephalogram signals), preprocessing of the low signal-to-noise ratio ultra-short time signal is required before sampling by a two-point pair, and the preprocessing generally includes filtering and noise reduction processing, that is, removing the interference of direct current bias and some fixed frequency noise in the original signal by a time domain filter. For some low signal-to-noise ratio ultra-short time signals (such as mechanical fault signals), filtering and noise reduction preprocessing is not required, and whether to perform filtering and noise reduction preprocessing can be selected by a person skilled in the art according to the low signal-to-noise ratio ultra-short time signal, and the present application is not limited in this regard.

[0055] The present application also provides a classification system for low signal-to-noise ratio ultra-short time signals, which is an end-to-end classification model for low signal-to-noise ratio ultra-short time signals. The low signal-to-noise ratio ultra-short time signal is sampled to serve as an input sub-sample of the end-to-end classification model. The input sub-sample has a size of CxT, where C is the number of channels of the low signal-to-noise ratio ultra-short time signal, and T is the total number of sampling points of the low signal-to-noise ratio ultra-short time signal; a sufficient number of input sub-samples of multiple categories with labels are divided into a training set and a test set for supervised training; and the trained end-to-end classification model can quickly identify the label of an unknown signal.

[0056] Referring to Figure 2 The classification system for low signal-to-noise ratio ultra-short time signals of the present application is a convolutional neural network that can learn autonomously, and includes:

[0057] a feature extraction module: sampling a two-point pair of a narrowband signal of a low signal-to-noise ratio ultra-short time signal, and calculating a feature vector according to the sampling value; the two-point pair is any two sampling points of the low signal-to-noise ratio ultra-short time signal, and the feature vector is composed of feature amplitudes of multiple two-point pairs;

[0058] a classification module: used for classifying the low signal-to-noise ratio ultra-short time signal by using the feature vector.

[0059] Based on the above-mentioned low signal-to-noise ratio ultra-short time signal classification method of the application, according to requirements, the low signal-to-noise ratio ultra-short time signal classification system of the application can also be provided with a preprocessing module and / or a decomposition module, wherein the preprocessing module is used for preprocessing the low signal-to-noise ratio ultra-short time signal, and the decomposition module is used for decomposing the low signal-to-noise ratio ultra-short time signal into a plurality of narrowband signals; if both the preprocessing module and the decomposition module are required, then the low signal-to-noise ratio ultra-short time signal is first preprocessed by the preprocessing module, and then decomposed into a plurality of narrowband signals by the decomposition module.

[0060] Hereinafter, the low signal-to-noise ratio ultra-short time signal classification system of the application is further described in detail with the low signal-to-noise ratio ultra-short time signal classification system comprising a preprocessing module, a decomposition module, a feature extraction module and a classification module as an example.

[0061] The low signal-to-noise ratio ultra-short time signal classification system of the application is specifically composed of a preprocessing module, a frequency domain decomposition convolution layer, a feature extraction module and a classification module in cascade. The preprocessing module comprises a time domain band-pass filtering operation and a spatial convolution layer; the frequency domain decomposition convolution layer is responsible for decomposing the input sample into a plurality of narrowband components through a set of time domain band-pass filters; the feature extraction module is responsible for extracting a two-point pair from each narrowband signal of the input sample and calculating the amplitude based on the above-mentioned formula (1); and the classification module classifies the input sample based on the feature vector composed of the feature amplitudes calculated by the feature extraction module.

[0062] Specifically, the time domain band-pass filtering operation is used to filter out the direct current bias and the fixed frequency noise of interference in the low signal-to-noise ratio ultra-short time signal.

[0063] Specifically, the spatial convolution layer further improves the signal-to-noise ratio of the low signal-to-noise ratio ultra-short time signal by performing a spatial filtering operation, and contains F1 convolution kernels with a size of Cx1.

[0064] The convolution kernel of the frequency domain decomposition convolution layer is initialized as a filter bank. The filter bank contains F2 finite impulse response (FIR) band-pass filters with a length of N. Each FIR band-pass filter h[n] is obtained by the following formula (2):

[0065]

[0066] wherein f h represents the upper limit cutoff frequency of the narrowband signal, f l represents the lower limit cutoff frequency of the narrowband signal, and f swhere f represents the sampling frequency, n is the index of the array h[n], n = 0, 1, 2, …, N-1, and N is the length of the finite impulse response band-pass filter. Based on prior knowledge or spectrum analysis of the low signal-to-noise ratio ultra-short time signal, a frequency band of useful information is obtained, and the frequency band of useful information is divided into F2 narrower frequency bands. The upper and lower cutoff frequencies of each of the narrow bands are substituted into formula (2) to obtain F2 FIR band-pass filters.

[0067] The feature extraction module includes a grouped dilated convolution layer, a square activation function, and a square root activation function. The size of the convolution kernel of the grouped dilated convolution is 1x2, and the convolution kernel of the grouped dilated convolution is divided into F2 groups, each group having two convolution kernels. The two convolution kernels of each group are used to extract a plurality of consecutive two-point pairs from a single narrowband component output by the frequency domain decomposition layer. The convolution kernel hole rate Δn = f s x Δt, then the time interval Δt of all two-point pairs is equal, and the hole rate Δn = f s x Δt of the dilated convolution is determined through multiple trials based on the nature of the actual signal. The feature amplitude feature calculation process of any narrowband signal is as shown in the following formula (3):

[0068]

[0069] where v 11 and v 12 respectively represent two trainable parameters of the first convolution kernel in the two convolution kernels corresponding to each narrowband signal, which can be initialized as the coefficients or randomly initialized in formula (1); v 21 and v 22 respectively represent two trainable parameters of the second convolution kernel in the two convolution kernels corresponding to each narrowband signal, which can be initialized as the coefficients and or randomly initialized in formula (1).

[0070] Therefore, the dilated convolution kernel slides one step along the time axis direction to extract a feature vector for the input sample, and multiple sliding can extract a plurality of feature amplitudes for each narrowband signal. The feature vector is composed of an amplitude feature of each narrowband component from all signal channels.

[0071] The classification module predicts a sub-label for the feature vector through a full connection layer and a Softmax activation function, and finally classifies the low signal-to-noise ratio ultra-short time signal through all the sub-labels. In the training process of the end-to-end classification model, the cross-entropy of the sub-labels and the true label of the input sample is summed as the training loss; in the testing process of the end-to-end classification model, the final prediction label of the input sample is obtained by majority voting from the prediction labels of all the feature vectors extracted from the sample.

[0072] Embodiment

[0073] The invented end-to-end classification model is implemented on the 2a dataset of the fourth brain-computer interface competition (BCIC IV2a). The dataset is collected through 22 electrode channels and a sampling rate of 250 Hz, containing four types of tasks such as left-hand motor imagery, right-hand motor imagery, both feet motor imagery, and tongue motor imagery. Figure 3 The positions of the 22 electrodes on the scalp and a typical motor imagery electroencephalogram signal collected through these electrodes are shown, which contains a large amount of noise and presents a disordered state. The dataset contains 9 subjects, and each subject is collected for two days. Each day collects 288 4s signals, of which there are 72 under each type of label. The first second part of the 4s signal is intercepted as the input sample. In order to adapt to the two-dimensional convolution operation, the input sample is expanded to three dimensions by adding a channel dimension, so the final size of each input sample is 1×22×250.

[0074] The workflow of the classification system is shown in Figure 4 The input sample with a size of 1×22×250 first passes through the preprocessing module composed of 1-48 Hz filtering operation and spatial convolution layer. The data matrix output by the spatial convolution layer has a size of 16×1×250, which is transposed to 1×16×250 and sent to the frequency domain decomposition convolution layer. Based on the prior knowledge of motor imagery electroencephalogram signal, 24 continuous FIR narrowband bandpass filters with a length of 51 are constructed as the convolution kernel of the frequency domain decomposition convolution layer, and their center frequencies are uniformly distributed in the range of 2-32 Hz. The data matrix S output by the frequency domain convolution layer has a size of 24×16×200, containing 24×16 narrowband components.

[0075] The grouping hollow convolution layer independently extracts a pair of double points from each narrowband component and solves its feature amplitude. The data matrix output by the grouping hollow convolution layer has a size of 48×16×100. Channel separation refers to separating the odd and even indexes in the first dimension (containing 48 channels) of the data matrix output by the grouping hollow convolution layer, obtaining two data matrices with a size of 24×16×100 and performing matrix addition operation. As Figure 4As shown, square and square root activation is added again. Through the above process, formula (3) can be executed in parallel in each frequency band without interference.

[0076] The data matrix finally output by the feature extraction module The size is 24x16x100, and 100 groups of amplitude features are obtained, each group of amplitude features being composed of feature amplitudes extracted from 24x16 narrowband components. Into the classification convolution, 100 sub-labels are predicted.

[0077] The classification system of the application is a convolutional neural network, which needs to be trained before application. In this embodiment, 4-fold cross-validation is used to train and evaluate the performance of the classification system, that is, the 288 samples obtained by each person on the first day are divided into 4 equal parts, and each time one of them is selected as the test set, and the remaining 3 are used as the training set. In the training stage, the cross-entropy of each sub-label and the true label of the sample is calculated as the training loss to update the parameters of the spatial convolution layer, the feature extraction module and the classification convolution layer. In the test stage, the majority vote of the 100 sub-labels is given to the final prediction label of the input sample. If the predicted label is consistent with the true label, the classification is correct. The number of correctly classified samples in the training set is counted, and then divided by the total number to obtain the classification accuracy.

[0078] The training and performance evaluation experiment is performed 10 times on each subject, and each training is performed for 800 generations. The maximum classification accuracy is shown in Table 1, and the unit is percentage.

[0079] Table 1

[0080]

[0081] As can be seen from Table 1, compared with the ATCNet, C-LSTM, KFCNet and other classification networks proposed in the existing literature, the end-to-end classification model proposed in the application obtains the highest classification accuracy.

Claims

1. A method of classifying low signal-to-noise ratio ultra-short time signals, characterized by, The method comprises the following steps: Preprocessing of low signal-to-noise ratio ultra-short time signal: filtering out the direct current bias and fixed frequency noise in the low signal-to-noise ratio ultra-short time signal; when the low signal-to-noise ratio ultra-short time signal is a wideband signal, the preprocessed low signal-to-noise ratio ultra-short time signal is decomposed into multiple narrowband components, and then the two-point pairs of the narrowband components obtained by decomposition are sampled; When the low signal-to-noise ratio ultra-short time signal is a narrowband signal, the preprocessed low signal-to-noise ratio ultra-short time signal is decomposed into multiple narrowband components, and then the two-point pairs of the narrowband components obtained by decomposition are sampled; Or, when the low signal-to-noise ratio ultra-short time signal is a narrowband signal, the two-point pairs of the preprocessed low signal-to-noise ratio ultra-short time signal are directly sampled; Decomposing the preprocessed low signal-to-noise ratio ultra-short time signal into multiple narrowband components; Sampling the two-point pairs of the narrowband signal of the low signal-to-noise ratio ultra-short time signal by using the two-point pairs of the narrowband components obtained by decomposition, and calculating the feature vector according to the sampling values; the two-point pairs are any two sampling points of the low signal-to-noise ratio ultra-short time signal, and the feature vector is composed of feature amplitudes of multiple two-point pairs; the feature amplitude A of the two-point pair is calculated by the following formula: wherein and denote the sample values of the two sampling points in the doublet pair, denotes the center angular frequency of the narrowband signal, denotes the time interval of the two sampling points in the doublet pair; Classifying the low signal-to-noise ratio ultra-short time signal by using the feature vector.

2. A low signal-to-noise ratio ultra-short time signal classification system, characterized by, It comprises: A preprocessing module for filtering out the direct current bias and fixed frequency noise in the low signal-to-noise ratio ultra-short time signal; When the low signal-to-noise ratio ultra-short time signal is a wideband signal, the preprocessed low signal-to-noise ratio ultra-short time signal is decomposed into multiple narrowband components by a decomposition module, and the two-point pairs of the narrowband components obtained by decomposition are sampled by a feature extraction module; when the low signal-to-noise ratio ultra-short time signal is a narrowband signal, the preprocessed low signal-to-noise ratio ultra-short time signal is decomposed into multiple narrowband components by the decomposition module, and the two-point pairs of the narrowband components obtained by decomposition are sampled by the feature extraction module; or, when the low signal-to-noise ratio ultra-short time signal is a narrowband signal, the preprocessed low signal-to-noise ratio ultra-short time signal is transmitted to the feature extraction module by the preprocessing module, and the feature extraction module samples the two-point pairs of the preprocessed low signal-to-noise ratio ultra-short time signal again; A decomposition module for decomposing the preprocessed low signal-to-noise ratio ultra-short time signal into multiple narrowband components; A feature extraction module for sampling the two-point pairs of the narrowband signal of the low signal-to-noise ratio ultra-short time signal by using the two-point pairs of the narrowband components obtained by decomposition, and calculating the feature vector according to the sampling values; the two-point pairs are any two sampling points of the low signal-to-noise ratio ultra-short time signal, and the feature vector is composed of feature amplitudes of multiple two-point pairs; the feature extraction module comprises a grouping hole convolution layer, a square activation function and a square root activation function; The grouping hollow convolutional layer is used to extract multiple continuous double-point pairs from a single narrowband signal, each group in the grouping hollow convolutional layer has two convolution kernels, and the hollow rate of the convolution kernel is set to , the time interval of all double-point pairs is equal, and the feature amplitude calculation process of any one narrowband signal is as follows: wherein, and respectively represent two trainable parameters of a first one of the two convolutional kernels corresponding to each narrowband signal, which can be initialized to or be randomly initialized; and respectively represent two trainable parameters of a second one of the two convolutional kernels corresponding to each narrowband signal, which can be initialized to and or be randomly initialized, denotes the center angular frequency of the narrowband signal. A classification module for classifying the low signal-to-noise ratio ultra-short time signal by using the feature vector.

3. A low signal-to-noise ratio ultra-short time signal classification system according to claim 2, characterized in that, The decomposition module comprises a frequency domain decomposition convolution layer, which decomposes the low signal-to-noise ratio ultra-short time signal into multiple narrowband signals through a set of time domain bandpass filters; The convolution kernel of the frequency domain decomposition convolution layer is initialized as a filter bank, the filter bank contains a plurality of finite impulse response band-pass filters of a preset length, and the finite impulse response band-pass filter The calculation formula is as follows: wherein represents the upper cut-off frequency of the narrowband signal, represents the lower cut-off frequency of the narrowband signal, represents the sampling frequency, n is the index of the array h[n], n = 0, 1, 2,..., N -1, N is the length of the finite impulse response bandpass filter.

4. A low signal-to-noise ratio ultra-short time signal classification system according to claim 2, characterized in that, The classification module predicts a sub-label for the feature vector through a fully connected layer and a Softmax activation function, and classifies the low signal-to-noise ratio ultra-short time signal through all sub-labels.

Images