Time series signal searching method based on two-stage deep learning model
By dividing the time series signal search process into two stages and using a deep learning model to predict the segmented images and binary matrices, the problem of high computational cost and low accuracy in existing technologies is solved, achieving more efficient and accurate signal recognition.
Patent Information
- Application Number
- CN202310271654.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-03-20
- Publication Date
- 2026-01-20
- Estimated Expiration
- 2043-03-20
AI Technical Summary
Existing time series signal search methods based on transform domain are computationally intensive and have low accuracy.
A two-stage deep learning model is adopted. First, the time series data is converted into image files and segmented into frame images. Then, the first deep learning model is used to predict the segmented images to generate a binary matrix. Finally, the second deep learning model is used to further predict the binary matrix to improve the search speed and accuracy.
It improves the efficiency and accuracy of time series signal search, while reducing computational load and memory requirements.
Smart Images
Figure CN116403027B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the field of high-energy signal intelligent recognition, and particularly relates to a time series signal searching method based on a two-stage deep learning model. BACKGROUND
[0002] Time series signals are generally random signals obtained by observation devices in a certain time period. Some signals randomly appear and last for a very short time on the time series. These signals have certain fixed characteristics and important physical meanings, and may represent fault events, notification events or the outbreak of other physical events. For example, a fast radio burst (FRB) is a high-energy radio pulse signal in the universe, which may last for only a few milliseconds, and the energy in the signal is very large. Abnormal voltage or current signals with a very short duration caused by occasional faults in a control system, etc. It is of great significance to quickly and accurately identify these target signals in time series signals.
[0003] Some time series signals have the characteristics of occasional occurrence and small sample size, and the characteristics of the subsequence are consistent with those of the original sequence. Traditional searching methods based on transform domain have large computational complexity and low accuracy. SUMMARY
[0004] The technical problem of the present application is that the existing time series signal searching method based on the transform domain has large computational complexity and low accuracy.
[0005] The purpose of the present application is to solve the above problems by providing a time series signal searching method based on two-stage deep learning. This method directly searches time series data, improves search speed and target signal recognition accuracy, reduces computational complexity, and saves memory.
[0006] The technical solution of the present application is a time series signal searching method based on a two-stage deep learning model, which includes the following steps:
[0007] Step 1: converting the time series data to be searched into an image file and cutting it into frame images according to the time axis;
[0008] Step 2: block processing the frame images obtained in step 1 to obtain block images;
[0009] Step 3: inputting the block images obtained in step 2 into a first deep learning model for first-stage prediction, i.e. using the first deep learning model to predict whether the block images contain target signals;
[0010] Step 4: reconstructing the prediction result of the first-stage segmented image, generating a binary matrix according to the position of the segmented image in the original frame image, if the prediction result of the segmented image is the target signal, the value of the binary matrix corresponding to the position of the segmented image is 1, otherwise 0;
[0011] Step 5: inputting the binary matrix of step 4 into the second deep learning model for the second-stage prediction, i.e. using the second deep learning model to predict whether there is a target signal in the corresponding frame image according to the binary matrix.
[0012] Preferably, the time series data to be searched is in FITS format.
[0013] Further, the segmentation rule of the image file of the time series data into frame images in step 1 includes:
[0014] 1) all frame images have the same width and height, i.e. the size of the frame images is the same;
[0015] 2) the frame image marked as containing the target signal should contain at least part of the target signal, so as to train the deep learning model for effective identification of the signal.
[0016] Preferably, in step 2, the size of the segmented image is 128*128 or 256*256 or 512*512.
[0017] Preferably, the first deep learning model uses DenseNet or Inception-Resnetv2 or ResNet-Conv17 neural network.
[0018] Preferably, the second deep learning model includes sequentially connected Flatten, Linear1, Linear2, Linear3 layers and Attention layer. The output of the Flatten layer is an MxX1 matrix, the output of the Linear1 layer is an MxX2 matrix, the output of the Linear2 layer is an MxX3 matrix, the output of the Linear3 layer is an MxX4 matrix, and the output of the Attention layer is an MxX4 matrix, wherein X1=(n+2)*(n+2), X2=n*n, X3=(n / 2)*(n / 2), M represents the row of the output matrix of the Flatten, Linear1, Linear2, Linear3 and Attention layers, X1, X2, X3, X4 represent the column of the output matrix of the Flatten, Linear1, Linear2, Linear3 and Attention layers respectively, and n is the order of the binary matrix generated in the first stage.
[0019] Preferably, X4=2.
[0020] The training process of the first deep learning model comprises:
[0021] 1) generating a training sample data set, the training sample data set containing sample images, the size of the sample images being consistent with the size of the block images of step 2, and the training sample data being real observation data or simulation data;
[0022] 2) labeling the training sample data set, and adding a class label to each sample image, the class label indicating that the sample image is a positive sample or a negative sample, the positive sample being a sample image containing a target signal, and the negative sample being a sample not containing a target signal; the class label being labeled manually or automatically labeled by a computer program, and if the training sample data is simulation data, the class label is automatically labeled by a computer program;
[0023] 3) the number of sample images in the data set is 500-10000, and the balance of positive and negative samples is ensured.
[0024] The training process of the second deep learning model comprises:
[0025] 1) generating a training sample data set, the training sample data set of the second deep learning model being automatically generated by a computer program, a binary matrix being generated for each frame image, and if the original frame image is a positive sample, the corresponding binary matrix is a positive sample, otherwise it is a negative sample;
[0026] 2) dividing each frame image into blocks, and the blocks correspond to the binary matrix. The block image containing the target signal has a corresponding value of 1 in the binary matrix, and the block not containing the target signal has a corresponding value of 0, and if the frame image is simulation data, the computer program automatically sets each value of the binary matrix according to whether the frame image has a target signal and the position of the signal.
[0027] Compared with the prior art, the beneficial effects of the present application include:
[0028] 1) The present application divides the search process of the time series signal into two stages, the first stage classifies the block images of the time series signal data, the classification results are reconstructed into a binary matrix according to the relative positions of the block images, and then the second stage classification is performed to obtain the final search result, the two stages use the first deep learning model and the second deep learning model respectively, compared with the traditional search method using a single classification model, the efficiency of image processing and target search is improved through concurrent processing of block images, and the accuracy of target search is improved.
[0029] 2) The second deep learning model of the present application simplifies the structure of the deep learning model, reduces the computational complexity, and reduces the dependence on the computer memory capacity under the premise of ensuring the accuracy of target search. BRIEF DESCRIPTION OF DRAWINGS
[0030] The application will be further described below in conjunction with the accompanying drawings and examples.
[0031] Figure 1 Flowchart of the time series signal searching method of the embodiment of the application.
[0032] Figure 2 Effect diagram of injecting a single pulse signal in the background of the FAST signal noise.
[0033] Figure 3 Diagram of using Presto software to perform single pulse screening on the fits file of the inserted analog signal.
[0034] Figure 4 Diagram of the search results of the fits file of the inserted analog signal in the embodiment.
[0035] Figure 5a Effect diagram of the FAST observation data containing only background noise and interference in the embodiment.
[0036] Figure 5b Effect diagram of adding the analog FRB signal in Figure 5a in the embodiment.
[0037] Figure 6 Diagram of dividing Figure 5b into 2048*2048 size block images in the embodiment.
[0038] Figure 7 Diagram of dividing the block images of Figure 6 into 256*256 size block images in the embodiment.
[0039] Figure 8 Diagram of the Dense121 model structure in the embodiment.
[0040] Figure 9a Diagram of the ResNet-Conv17 model structure in the embodiment.
[0041] Figure 9b Diagram of the residual unit in the ResNet-Conv17 model.
[0042] Figure 10 Diagram of the Inception-ResNetV2 model structure in the embodiment. DETAILED DESCRIPTION
[0043] As shown in Figure 1 , the time series signal searching method based on two-stage deep learning includes the following steps:
[0044] 1、The time series data to be searched is divided on the time axis to form a frame image. The time series data is two-dimensional data, with the horizontal coordinate being the time axis and the vertical coordinate being frequency, intensity, channel, voltage, current, etc. according to different data types. Part of the vertical coordinate data can be removed according to the characteristics of the target signal. When dividing, the original data and the characteristics of the signal are considered, and the length of each frame is not less than the duration of the target signal, so that a complete target signal can be accommodated. Once the division length is specified, it remains unchanged to ensure that all frames are of equal length. The frame data is converted into image data according to the sampling time-vertical coordinate physical quantity, and each sampling time corresponds to a column on the image width.
[0045] 2、The frame image is divided into blocks. The original frame image can be divided into 256, 64 or 16 small block images with different specifications such as 16*16, 8*8 and 4*4. The block image specification mainly considers the size of the original frame image and the input image specification of the later deep learning model. In order to use conveniently, the input image of the deep learning model generally needs to ensure that the height and width of the image are the same, and the block size and the width and height of the block image size need to be the same. The block image size is not greater than 256*256.
[0046] 3、First stage prediction, using the first deep learning model to predict the block image. The prediction of the block image is a binary classification, if a block image contains a target signal, the prediction result is 1, otherwise it is 0.
[0047] 4、Prediction result reconstruction. The first stage prediction result is reconstructed to generate a binary matrix according to the position of the block image in the original frame image. The size of the binary matrix is the different block specifications such as 16*16, 8*8 and 4*4 in step 2, when the corresponding block prediction result is 1, the corresponding position value of the binary matrix is 1, and when the corresponding block prediction result is 0, the corresponding position value of the binary matrix is 0.
[0048] 5、Second stage prediction, using the second deep learning model to predict the binary matrix. The training process of the second deep learning model is described in detail later, and the prediction result is whether the frame has a target signal.
[0049] The training process of the first deep learning model includes:
[0050] 1) Generate a training sample data set. The training sample data set is composed of sample images, and the size of the sample image is consistent with the size of the block image in step 2, such as 256*256. The training sample data can use real data or simulated data. If simulated data is used, technical means need to be used to ensure the effectiveness of the simulated data.
[0051] 2) The labeling of the dataset, each data sample of the generated dataset needs to be labeled with a class label, indicating that the sample belongs to a positive sample or a negative sample. The class label can be labeled by manual method, if the method of simulating the generated dataset is adopted, the class label can be automatically labeled by program.
[0052] 3) The size of the dataset, the size of the deep learning dataset is about 10000 samples, and the balance of positive and negative samples should be ensured.
[0053] 4) Select the mainstream deep learning algorithm, input the sample dataset for training, and get the classification model.
[0054] The training process of the second deep learning model includes:
[0055] 1) Generate the training sample set. The sample set is automatically generated by program. First, select the original frame image, each original frame image corresponds to a binary matrix. If the original frame image is a positive sample, the binary matrix is a positive sample, otherwise it is a negative sample.
[0056] 2) Divide the original frame image into blocks. If the original frame image is simulated, the program can automatically set the value of the corresponding position of the binary matrix according to whether the image contains the target signal and the position of the signal: if the corresponding block contains the target signal, the corresponding value is 1, otherwise it is 0.
[0057] The second deep learning model is used to classify binary matrices, which are usually small in size, so a self-designed deep learning model is used.
[0058] The following is an example of searching for fast radio bursts on the data of the Five-hundred-meter Aperture Spherical radio Telescope (FAST) multi-target drift scanning survey project.
[0059] The FAST multi-target drift scanning survey uses a 19-beam receiver drift scanning observation mode, with a data sampling time of 49.152us, a sampling accuracy of 8Bit, a frequency range of 1.05GHz-1.45GHZ, 4096 channels, and a FITS format. The observation data can be regarded as an 8Bit continuous sequence image with a height of 4096 and an infinite width. Usually FRB signals always appear in the middle channel. In order to improve efficiency, the channels are cropped, and the high and low 1024 channels are removed, and the middle 2048 channels are retained for experiment, which does not affect the correctness of the method result.
[0060] Similar to most electromagnetic signal sampling methods, FAST divides the 1.05GHz-1.45GHz frequency range into 4096 wave bands, and each wave band is weighted by the center frequency, signal strength and signal-to-noise ratio to obtain a sampling value, forming 4096 channels.
[0061] Due to the small amount of real FRB signal samples, in order to increase the sample size, a method of simulating FRB signals is used. First, a FRB simulation signal is generated, and then the simulation signal is injected into the FAST observation data to obtain FRB signal samples. The generation and injection method of the FRB simulation signal is as follows:
[0062] 1) Input the sampling time, start and end frequency, dispersion value, and channel number of the FRB simulation signal to be generated;
[0063] 2) Calculate the initial frequency interval according to the start and end frequency and the channel number;
[0064] 3) Calculate the channel interval value of the signal;
[0065] 4) Shape the signal and adjust the signal-to-noise ratio;
[0066] 5) Generate a FRB simulation signal that changes by channel;
[0067] 6) On the original FAST observation data, find the injection position according to the start time, and superimpose the FRB simulation signal channel by channel with the FAST observation data as the background.
[0068] In order to verify the effectiveness of the simulation method, a single pulse is injected into FAST without signal noise background, and Presto software based on the traditional dispersion principle is used for searching, and the signal injection effect is as shown in Figure 2 .
[0069] Figure 2 In the simulation, a FRB signal with a dispersion value (Dispersion Measure, DM) of 1000 pc cm -3 and a duration of 2ms is generated, and the signal is injected into the background file at 11 minutes and 26 seconds. Presto software is used to perform single pulse screening on the fits file inserted with the simulation signal, and the screening effect is as shown in Figure 3 .
[0070] Figure 3 In the table, RA represents declination, DEC represents right ascension, Num Channels represents the number of channels, Power represents signal strength, T_sample represents sampling time, and T_total represents the current total sampling duration. Figure 3From left to right, the four signal curves are the maximum, variance, mean of each channel in the total sampling time calculated by the fits file of the inserted analog signal, and the fourth signal curve is the signal curve obtained by superimposing the maximum, variance, mean of each channel in the total sampling time calculated by the fits file of the inserted analog signal. The frequency of the signal in the channel number is represented directly above the four signal curves. The calculation method is to calculate the average value of each channel in the total time. By reflecting the strength distribution of the signal segment in time slots and channels, it is analyzed whether there is a high-intensity signal in the time domain or frequency domain. The comprehensive frequency of the signal in the same time domain in different channels in the fits file is represented on the right side of the four signal curves. The calculation method is to average the signals of multiple channels in each sampling time. By reflecting the strength distribution of the signal segment in time slots and channels, it is analyzed whether there is a high-intensity signal in the time domain or frequency domain.
[0071] The single-pulse screening effect is as shown in Figure 4 It can be seen that at Time = 1s, DM = 1000pccm -3 The position has a strong single-pulse mark, where red is periodic RFI noise, and blue and green represent the visual changes of the mean and variance of the fits file in the time domain. It can be seen that due to the insertion of the single pulse, there is an obvious strong peak in the search time domain.
[0072] Figure 4 The upper three small graphs of the figure show the relationship between the number of pulses and the signal-to-noise ratio, the relationship between the number of pulses and the dispersion, and the relationship between the signal-to-noise ratio and the dispersion, which intuitively reflects the relationship between the signal-to-noise ratio change and the pulse. The three graphs show that Presto effectively finds the inserted analog signal and shows the distribution change of the analog signal data. The lower graph reflects the distribution change of the dispersion in the entire signal file, indicating that there is a significant dispersion change in the 1s time period before and after the signal segment, i.e. the inserted FRB analog signal. Figure 4 In the figure, Signal-to-Noise represents the signal-to-noise ratio, and Number of Pulses represents the number of pulses. Figure 4 The upper left part shows the relationship between the signal-to-noise ratio and the number of pulses. When the signal-to-noise ratio is low, i.e. the noise is less and the pulse signal is clear and the number is obvious. When the signal-to-noise ratio gradually rises, the pulse signal is blurred and the number decreases.
[0073] Through the above verification, the effectiveness of the analog signal and the injection method of the embodiment is shown. The batch generated DM range of the embodiment is 50-5050pccm -3, sample data files with pulse width range 0.25-10.25ms. About 1-25 FRB simulation signals are superimposed on each noiseless background noise fits file at equal intervals. FAST observation data containing only background noise and interference is shown in Fig. 1. Figure 5a The FITS image after injecting FRB simulation signals is shown in Fig. 2. Figure 5b
[0074] Convert the FITS file to an image file and cut it into pieces for processing using a two-stage deep learning model. Convert the FITS file to an image file according to time-frequency-intensity, and the image size generated by each FITS file is 2048*131072. Cut each FITS file into 64 2048*2048 images on the time axis every 2048 pixels, and the effect is shown in Fig. 3. Figure 6 Figure 7
[0075] Since the injection program records the exact position and duration of each FRB simulation signal injection, the program can determine whether each 2048*2048 sub-image is a positive sample or a negative sample based on the cutting position. For each 256*256 image, the program automatically labels the positive and negative samples based on the cutting position and duration. When each 2048*2048 image is cut into 64 256*256 images, an 8x8 matrix will be generated according to the position correspondence. If the 256*256 image at a certain position on the matrix is a positive sample, the value at that position on the matrix is set to 1, otherwise it is set to 0. If the 2048*2048 image is a positive sample, the 8*8 position matrix corresponding to the image is a positive sample.
[0076] Through the above steps, four sets of image data sets Dataset1, Dataset2, Dataset3, and Dataset4 are made.
[0077] The sample image size of Dataset1 is 2048*2048, a total of 7040. Among them, the positive samples are 4591, and the negative samples are 2449.
[0078] The sample image size of Dataset2 is 256*256, which is obtained by cutting the samples in Dataset1, a total of 42638, and the number of positive and negative samples is 21319 and 21319 respectively.
[0079] The sample image size of Dataset3 is 8*8, which is converted from the position matrix to a binary image, and part of the random noise negative samples are added, and the number of positive and negative samples is 5315 and 4025 respectively.
[0080] Dataset4 is specially used for testing and checking the generalization ability of the model, which is made of different FAST data backgrounds. The sample parameters of Dataset4 are similar to those of Dataset1, and the image size is 2048*2048, with a total of 7040 images. Among them, the number of positive samples is 4591, and the number of negative samples is 2449.
[0081] The process of searching the FITS file by using the two-stage deep learning model is as follows:
[0082] 1. Read the original FITS file to be searched, and cut it into multiple 2048*2048 size images;
[0083] 2. Cut each 2048*2048 size image into 8*8, i.e. 64 256*256 size images;
[0084] 3. Use the first deep learning model to classify the 64 256*256 images according to whether they have FRB signals;
[0085] According to the classification result of model I, reconstruct an 8*8 size binary image according to the position information. If a 256*256 image has an FRB signal, set the corresponding position value in the 8*8 image to 1, otherwise set the corresponding position to 0;
[0086] 4. Use the learning model II to classify the 8*8 image. If it is classified as having an FRB signal, the corresponding 2048*2048 image contains an FRB signal; if it is classified as not having an FRB signal, the corresponding 2048*2048 image does not contain an FRB signal.
[0087] The training process of the first deep learning model of the embodiment: the deep learning classification model is trained on the Dataset2 data set. The present application selects three deep learning algorithms, ResNet-conv17, Dense121 and Inception-Resnetv2. At the same time, the Efficient-B6 Net algorithm is selected as a comparison. The Efficient-B6 Net model is an optimal structure model obtained by searching for model parameters, which is proposed by Tan M et al. in 2019. The accuracy performance of this model per computing power on the public data set is better than that of ResNet, Dense-Net, Inception-ResNetv2, and the model is more general.
[0088] The structure of the Dense121 model of the embodiment is as shown in Figure 8 The structure of the ResNet-conv17 model is as shown in Figure 9a The residual block 1, the residual block 2, the residual block 3 and the residual block 4 in the ResNet-conv17 model each contain a plurality of residual units, and the structure of the residual unit is as shown in Figure 9b The structure of the Inception-Resnetv2 model is as shown in Figure 10
[0089] The learning rate of each model is uniformly set to 0.00001, the weight decay rate is set to 0.001, the loss function is the cross-entropy loss function, and the early stopping method is set, and the patience is set to 10. The training set and the test set are uniformly divided according to the ratio of 8:2, and each round of training is randomly divided.
[0090] The training process of the second deep learning model of the embodiment: the second deep learning model is a CAModel designed by the embodiment, which is composed of 4 fully connected layers and 1 Attention layer, and the network structure is as shown in Table 1. The 8*8 data set of Dataset3 is used for training.
[0091] Table 1 CAModel model parameters
[0092]
[0093]
[0094] Using the 2048*2048 image in the data set 1, the image is reduced to 256*256, and then trained and tested. The results are shown in Table 2. Table 2 Training and testing accuracy of the original model
[0095]
[0096] First, train each deep learning model with the 256*256 data in the data set 2, then train the CAModel with the Dataset3 data set, finally integrate the two models into a two-stage deep learning model, and test it on the Dataset4.
[0097] The accuracy of training each deep learning model with the data set Dataset2 is shown in Table 3.
[0098] Table 3 Training and testing accuracy of each deep learning model on the data set Dataset2
[0099] Model Training accuracy Test accuracy ResNet-Conv17 97.80％ 93.42％ Inception-Resnetv2 99.76％ 98.16％ Dense121 99.70％ 97.51％ Efficient-b6 Net 99.63％ 98.60％
[0100] The training and testing accuracy of the CAModel on the data set Dataset3 is shown in Table 4.
[0101] Table 4 Training and testing accuracy of CAModel model
[0102] Model Training accuracy Test accuracy CAModel 100.00％ 100.00％
[0103] Experimental results of two-stage deep learning model: given a 2048*2048 image, first cut it into 8*8 256*256 images, then use ResNet-conv17, Dense121, Inception-Resnetv2, Efficient-B4 net, Efficient-B6 net and other models for classification, and then reconstruct an 8*8 binary image according to the classification result, and then use CAModel for classification. When testing as a whole, the data in the original training set is not used for testing, but a Dataset4 dataset is tested on another FAST data background. The experimental results are shown in Table 5.
[0104] Table 5 Overall experimental results of two-stage model
[0105] Model Accuracy Precision Recall F1-Score ResNet-Conv17-CAModel 66.46％ 97.28％ 65.36％ 75.12％ Inception-Resnetv2-CAModel 96.14％ 95.74％ 97.46％ 95.96％ Dense121-CAModel 90.05％ 84.72％ 98.05％ 87.45％ Efficient-b6 Net-CAModel 91.12％ 98.89％ 86.56％ 90.97％
[0106] The implementation results show that the time series signal searching method based on the two-stage deep learning model has the following advantages:
[0107] 1) Compared with a single deep learning model, the two-stage model uses the consistency of the signal after the time series signal of a specific target is blocked and the original signal, has a more specific direction, and has sensitivity to image rotation and stronger representation ability for specific target signals through the binary matrix generated by reconstruction.
[0108] 2) Compared with a single deep learning model, the two-stage model occupies less memory resources. If the two-stage model is not used, the size of the input image after cropping of the original image is 2048*2048, and if it is directly used as the input of the deep learning model, the memory resources consumed will be doubled compared with the 256*256 size image input. If the original image is reduced and input, the image quality will be greatly reduced.
[0109] 3) Compared with a single deep learning model, the accuracy of the two-stage model is greatly improved.
[0110] 4) Compared with other transform domain search methods, the two-stage model has the advantages of small calculation amount and less storage resource occupation.
Claims
1. A time series signal search method based on a two-stage deep learning model, characterized in that, The time series signal searching method divides the searching process of the time series signal into two stages, the first stage classifies the block images of the time series signal data, the classification results are reconstructed into a binary matrix according to the relative positions of the block images, and then the second stage classification is performed to obtain the final searching result, and the two stages respectively adopt a first deep learning model and a second deep learning model; The first deep learning model adopts DenseNet or Inception-Resnetv2 or ResNet-Conv17; The second deep learning model comprises sequentially connected Flatten, Linear1, Linear2, Linear3 layers and an Attention layer; The time series signal searching method comprises the following steps: Step 1: converting the time series data to be searched into an image file, and cutting along the time axis to divide into frame images; Step 2: block processing the frame images obtained in step 1 to obtain block images; Step 3: inputting the block images obtained in step 2 into the first deep learning model for first-stage prediction, that is, using the first deep learning model to classify the block images according to whether they contain target signals; Step 4: reconstructing the prediction results of the block images in the first stage, generating a binary matrix from the prediction results of the block images in step 3 according to the positions of the block images in the original frame images, if the prediction result of the block image is that there is a target signal, then the value of the corresponding position of the binary matrix is 1, otherwise it is 0; Step 5: inputting the binary matrix in step 4 into the second deep learning model for second-stage prediction, that is, using the second deep learning model to predict whether the corresponding frame image has a target signal according to the binary matrix.
2. The time series signal search method according to claim 1, characterized by, The cutting rule of step 1 for cutting the image file of the time series data into frame images comprises: 1) all frame images have the same width and height, that is, the sizes of the frame images are the same; 2) the frame image marked as containing a target signal should contain at least part of the target signal, so as to facilitate the training of the deep learning model for effective identification of signals.
3. The time series signal search method according to claim 2, characterized by, In step 2, the size of the block image is 128*128 or 256*256 or 512*512.
4. The time series signal search method according to claim 3, characterized by, The output of the Flatten layer is M × X A matrix, the output of Linear1 layer is M × X 2 matrices, the output of the Linear2 layer is M × X 3 matrices, the output of the Linear3 layer is M × X 4 matrices, the output of the Attention layer is M × X 4 matrices, where X 1 = ( n +2)*( n +2), X 2= n*n , X 3 = ( n / 2)*( n / 2), M This represents the rows of the output matrices of the Flatten, Linear1, Linear2, Linear3, and Attention layers. X 1. X 2. X 3. X 4 represents the columns of the output matrices of the Flatten, Linear1, Linear2, Linear3 layers and the Attention layer, respectively. n The order of the binary matrix generated in the first stage.
5. The time series signal search method according to claim 4, characterized by, X 4=2。 6. The time series signal search method according to claim 5, characterized by, The training process of the first deep learning model comprises: 1) generating a training sample data set, the training sample data set contains sample images, the size of the sample images is consistent with the size of the block images in step 2, and the training sample data uses real observation data or simulation data; 2) labeling the training sample data set, adding a class label to each sample image, indicating that the sample image is a positive sample or a negative sample through the class label, the positive sample is a sample image containing a signal, and the negative sample does not contain a signal; the class label is labeled manually or automatically labeled by a computer program, if the training sample data is simulation data, the class label is automatically labeled by a computer program according to the generation rule of the simulation data; 3) the number of sample images in the data set is 500-10000, and the balance of positive and negative samples is ensured.
7. The time series signal search method according to claim 6, characterized by, The training process of the second deep learning model comprises: 1) generating a training sample data set, the training sample data set of the second deep learning model is automatically generated by a computer program, a binary matrix is generated for each frame of image, if the original frame of image is a positive sample, the corresponding binary matrix is a positive sample, otherwise it is a negative sample; 2) dividing each frame of image into blocks, the value of the binary matrix corresponding to the block image with target signal is 1, the value of the binary matrix without target signal is 0, if the frame of image is simulation data, the value of each binary matrix is set automatically by a computer program according to whether the frame of image has target signal and the position of the signal.