A signal data fusion method and identification method

Abstract

The present invention relates to a signal data fusion method and an identification method, wherein the fusion method comprises the following steps: obtaining data of m sensors at time t to t+h; decomposing the data of the m sensors at time t to t+h by using adaptive empirical mode decomposition to obtain the intrinsic mode function of all data and the corresponding residual value; extracting and fusing the intrinsic mode function of each sensor data at time t to t+h according to the time, obtaining the modal decomposition data matrix of each sensor data at time t to t+h; stacking the modal decomposition data matrix of each sensor data at time t to t+h, obtaining the modal decomposition data matrix of all data at time t to t+h. The present invention can display more data feature details while converting one-dimensional data into matrix data.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to a signal data fusion method and an identification method. Background Art

[0002] In wireless communication cognitive radio, spectrum feature recognition is an important part of dynamic spectrum management. In traditional spectrum feature recognition, Fourier transform, wavelet transform, Hilbert transform and other methods are usually used to convert time domain information into spectrum information, thereby showing the characteristics of the signal. However, the characteristics of the time dimension of the signal are lost during the conversion process of Fourier transform and wavelet transform. Although the Hilbert transform can retain the time domain information of the signal to the maximum extent, the signal is processed in a one-dimensional way. This one-dimensional sequence signal processing method loses the structural relationship information in time.

[0003] One of the development directions is to introduce deep learning into wireless communication equipment to empower the equipment with intelligence. However, in the classification problem of deep learning, time series-based deep learning networks such as long short-term memory networks (LSTM) and recurrent neural networks (RNN) are generally used to process vector data, and their functions are also focused on sequence prediction. Fully connected neural networks (FNN) can usually achieve a variety of tasks such as classification and regression, but when facing massive data, the scale of FNN network parameters is too large, the hardware resources are high, and the stability of this network is poor, and it is easy to have problems such as gradient explosion and gradient disappearance. Convolutional neural networks (CNN) are often used in visual image recognition, and CNN is usually used to solve classification problems. However, CNN is generally used to process matrix data and cannot process sequence data. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a signal data fusion method and a recognition method, which can display more data feature details while converting one-dimensional data into matrix data.

[0005] The technical solution adopted by the present invention to solve the technical problem is: to provide a signal data fusion method, comprising the following steps:

[0006] (1) Obtain data from m sensors at time t to t+h;

[0007] (2) Adaptive empirical mode decomposition is used to decompose the data of the m sensors at time t to t+h to obtain the eigenmode functions of all the data and the corresponding residual values;

[0008] (3) extracting and fusing the intrinsic mode functions of each sensor data at time t to t+h according to the time, and obtaining the modal decomposition data matrix of each sensor data at time t to t+h;

[0009] (4) The modal decomposition data matrices of the obtained sensor data at the time t to t+h are stacked to obtain the modal decomposition data matrices of all the data at the time t to t+h.

[0010] The step (2) specifically includes the following sub-steps:

[0011] (21) Select the data D of the i-th sensor at time t it , for the data D it Add n Gaussian white noises to get n data D it1 ~D itn ;

[0012] (22) After adding white noise, the data D it1 ~D itn The empirical mode decomposition algorithm is used to perform an iterative calculation to obtain the initial eigenmode functions IMF11~IMF1 n ;

[0013] (23) Take the average of the obtained n initial eigenmode functions and use the average as data D it The intrinsic mode function IMF1, using data D it Subtract the intrinsic mode function IMF1 to obtain the residual value r i1 ;

[0014] (24) For the residual value r i1 Continue K-1 decompositions according to the method of step (21) to step (23) until the residual value r is obtained. i1 When the number of extreme points of is less than 2, the decomposition is completed, and a total of K intrinsic mode functions IMF1~IMFK and residual value Res are obtained, and the intrinsic mode function IMF1 of the i-th sensor data at time t is recorded as it ~IMFK it and residual value Res it ;

[0015] (25) Take the data D of the i-th sensor at time t+1~t+h respectively it+1 ~D it+h , repeat the steps (21) to (24) to obtain the intrinsic mode function IMF1 of the i-th sensor data at the time t+1 to t+h it+1 ~IMFK it+1 ,……IMF1 it+h ~IMFK it+h And the corresponding residual value Res it+1 ~Res it+h ;

[0016] (26) Repeat steps (21) to (25) to obtain the intrinsic mode function IMF1 of all sensor data 1t ~IMFK 1t ,……IMF1 1t+h ~IMFK 1t+h ...IMF1 mt ~IMFK mt ,……IMF1 mt+h ~IMFK mt+h And the corresponding residual value Res 1t ~Res 1t+h ……Res mt ~Res mt+h .

[0017] The step (3) specifically includes the following sub-steps:

[0018] (31) The intrinsic mode function IMF1 of the first sensor data at time t~t+h 1t ~IMF1 1t+h Arrange them in rows and get the IMF1 matrix of the first sensor data at time t~t+h:

[0019] (32) The intrinsic mode function IMF2 of the first sensor data at time t~t+h 1t ~IMFK 1t+h and residual value Res 1t ~Res 1t+h According to the method in step (31), the IMF2 matrix to the IMFK matrix and the Res1 matrix of the first sensor data at time t to t+h are obtained in parallel;

[0020] (33) The IMF1 matrix, IMFK matrix and Res1 matrix of the first sensor data at time t to t+h are stacked to obtain the modal decomposition data matrix of the first sensor data at time t to t+h:

[0021] (34) Using the method of steps (31) to (33), the modal decomposition data matrix of the 2nd to mth sensor data at time t to t+h is obtained:

[0022] The technical solution adopted by the present invention to solve the technical problem is: a signal data recognition method is also provided, comprising the following steps:

[0023] (1) The above-mentioned signal data fusion method is used to obtain the modal decomposition data matrix of the data to be identified at time t to t+h;

[0024] (2) Input the modal decomposition data matrix into the trained recognition model to complete the recognition of the data.

[0025] The recognition model adopts a CNN deep learning network, including two convolutional layers, two pooling layers and a fully connected layer, wherein the first convolutional layer is provided with 64 convolutional kernels and the second convolutional layer is provided with 16 convolutional kernels.

[0026] Beneficial Effects

[0027] Due to the adoption of the above-mentioned technical scheme, the present invention has the following advantages and positive effects compared with the prior art: the present invention adopts the method of adaptive complete empirical mode decomposition to transform the one-dimensional sequence data from one-dimensional space to high-dimensional space by decomposition, fusion, etc., while retaining the characteristic details of the data such as time sequence and amplitude to the maximum extent, thereby improving the extraction and recognition of data features. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a flow chart of a first embodiment of the present invention;

[0029] Figure 2 is a flow chart of a single sensor data fusion process in a first embodiment of the present invention;

[0030] Figure 3 It is a schematic diagram of the training results of the recognition model in the second embodiment of the present invention. DETAILED DESCRIPTION

[0031] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0032] The first embodiment of the present invention relates to a signal data fusion method, such as Figure 1 As shown, the method comprises the following steps: obtaining data of m sensors at time t to t+h; decomposing the data of the m sensors at time t to t+h by using adaptive empirical mode decomposition to obtain the intrinsic mode functions of all data and the corresponding residual values; extracting and fusing the intrinsic mode functions of the data of each sensor at time t to t+h according to the time to obtain the modal decomposition data matrix of the data of each sensor at time t to t+h; stacking the modal decomposition data matrices of the obtained data of each sensor at time t to t+h to obtain the modal decomposition data matrix of all data at time t to t+h.

[0033] This method uses adaptive complete empirical mode decomposition to transform the one-dimensional sequence data from one-dimensional space to high-dimensional space by decomposing and merging the one-dimensional sequence data, while retaining the time sequence, amplitude and other characteristic details of the data to the maximum extent. It specifically includes the following steps:

[0034] Step 1: Get the data of sensors 1 to m at time t to t+h.

[0035] Step 2: If Figure 2 As shown, any sensor data D at time t is taken it , for data D it Add n Gaussian white noises respectively to get n data D it1 ~D itn .

[0036] Step 3: Add white noise to the data D it1 ~D itn The empirical mode decomposition algorithm is used to perform an iterative calculation; the first intrinsic mode function IMF11~IMF1 is obtained. n .

[0037] Step 4: Take the average of the n eigenmode functions obtained in step 3 IMF1 is data D it The first eigenmode function IMF1.

[0038] Step 5: Use data D it Subtract the first local oscillation mode function IMF1 to get the residual value r i1 .

[0039] Step 6: Continue to i1 According to the method from step 2 to step 5, the residual value r i1 Continue K-1 times of decomposition until the residual value r is obtained i1 When the number of extreme points of is less than 2, the decomposition is completed. After the decomposition is completed, a total of K intrinsic mode functions IMF1~IMFK and residual value Res are obtained, which is recorded as the intrinsic mode function IMF1 of the i-th sensor data at time t it ~IMFK it and residual value Res it .

[0040] Step 7: Take the data D of sensor No. i at time t+1~t+h respectively it+1 ~D it+h Repeat steps 2 to 6 to obtain the intrinsic mode function IMF1 of the i-th sensor data at time t+1~t+h. it+1 ~IMFK it+1 ,……IMF1 it+h ~IMFK it+hAnd the corresponding residual value Res it+1 ~Res it+h .

[0041] Step 8: Decompose the remaining data according to the methods of steps 2 to 8 to obtain the intrinsic mode function IMF1 of all data 1t ~IMFK 1t ,……IMF1 1t+h ~IMFK 1t+h ...IMF1 mt ~IMFK mt ,……IMF1 mt+h ~IMFK mt+h And the corresponding residual value Res 1t ~Res 1t+h ……Res mt ~Res mt+h .

[0042] Step 9: The intrinsic mode function IMF1 of the sensor data No. 1 obtained in step 8 at time t~t+h 1t ~IMF1 1t+h By arranging them in rows, we get the IMF1 matrix of sensor No. 1 at time t~t+h:

[0043]

[0044] Step 10: The local oscillator mode function IMF2 of the sensor data at time t~t+h 1t ~IMFK 1t+h and residual value Res 1t ~Res 1t+h The data are merged in the manner of step nine to obtain n-1 matrices IMF1 to IMFK of sensor No. 1 at time t to t+h.

[0045] Step 11: Stack the IMF matrices of the data obtained in step 9 and step 10 at time t to t+h to obtain the final modal decomposition data matrix of the sensor data No. 1 at time t to t+h:

[0046]

[0047] Step 12: Repeat the methods of steps 9, 10 and 11 for the modal decomposition data of the remaining data to obtain the modal decomposition data matrix of m-1 data at time t~t+h:

[0048]

[0049] Step 11: Stack the matrices obtained in step 9 and step 10 again to obtain the final modal decomposition data matrix of all data from 0 to i at time t to t+h:

[0050]

[0051] The second embodiment of the present invention relates to a signal data recognition method, comprising the following steps: using the signal data fusion method in the first embodiment to obtain the modal decomposition data matrix of the data to be recognized at time t to t+h; inputting the modal decomposition data matrix into a trained recognition model to complete the recognition of the data.

[0052] In this embodiment, the recognition model uses a CNN deep learning network, including two convolutional layers, two pooling layers and a fully connected layer, wherein the first convolutional layer is provided with 64 convolutional kernels and the second convolutional layer is provided with 16 convolutional kernels. The recognition model can achieve an accuracy rate of 98.05% in the training set and 99.33% in the validation set. Figure 3 Give the accuracy curves on the training set and validation set during the training process.

[0053] Judging from the verification results, the time series data processed and fused by adaptive complete empirical mode decomposition retains more of the data characteristics. Through different intrinsic mode functions, the characteristics of the data will be more prominent in different dimensions, which is more conducive to the extraction and identification of data features.

[0054] It is not difficult to find that the present invention aims to convert one-dimensional data into matrix data while, for the purpose of showing more data feature details, especially for the feature display of nonlinear and non-stationary data, an adaptive complete empirical mode decomposition method is adopted to decompose and fusion one-dimensional sequence data, thereby converting the data from one-dimensional space to high-dimensional space, while retaining the time sequence, amplitude and other feature details of the data to the maximum extent, thereby improving the extraction and recognition of data features.

Claims

1. A signal data fusion method, characterized in that: The following steps are involved: (1) Obtain data from m sensors at time t to t+h; (2) Adaptive empirical mode decomposition is used to decompose the data of the m sensors at time t to t+h to obtain the intrinsic mode functions of all the data and the corresponding residual values, which specifically includes the following sub-steps: (21) Select the data D of the i-th sensor at time t it , for the data D it Add n Gaussian white noises to get n data D it1 ~D itn ; (22) After adding white noise, the data D it1 ~D itn The empirical mode decomposition algorithm is used to perform an iterative calculation to obtain the initial eigenmode functions IMF11~IMF1 n ; (23) Take the average of the obtained n initial eigenmode functions and use the average as data D it The intrinsic mode function IMF1, using data D it Subtract the intrinsic mode function IMF1 to obtain the residual value r i1 ; (24) For the residual value r i1 Continue K-1 decompositions according to the method of step (21) to step (23) until the residual value r is obtained. i1 When the number of extreme points of is less than 2, the decomposition is completed, and a total of K intrinsic mode functions IMF1~IMFK and residual value Res are obtained, and the intrinsic mode function IMF1 of the i-th sensor data at time t is recorded as it ~IMFK it and residual value Res it ; (25) Take the data D of the i-th sensor at time t+1~t+h respectively it+1 ~D it+h , repeat the steps (21) to (24) to obtain the intrinsic mode function IMF1 of the i-th sensor data at the time t+1 to t+h it+1 ~IMFK it+1 ,……IMF1 it+h ~IMFK it+h And the corresponding residual value Res it+1 ~Res it+h ; (26) Repeat steps (21) to (25) to obtain the intrinsic mode function IMF1 of all sensor data 1t ~IMFK 1t ,……IMF1 1t+h ~IMFK 1t+h ...IMF1 mt ~IMFK mt ,……IMF1 mt+h ~IMFK mt+h And the corresponding residual value Res 1t ~Res 1t+h ……Res mt ~Res mt+h ; (3) The intrinsic mode functions of each sensor data at time t to t+h are extracted and fused according to the time to obtain the modal decomposition data matrix of each sensor data at time t to t+h, which specifically includes the following sub-steps: (31) The intrinsic mode function IMF1 of the first sensor data at time t~t+h 1t ~IMF1 1t+h Arrange them in rows and get the IMF1 matrix of the first sensor data at time t~t+h: (32) The intrinsic mode function IMF2 of the first sensor data at time t~t+h 1t ~IMFK 1t+h and residual value Res 1t ~Res 1t+h According to the method in step (31), the IMF2 matrix to the IMFK matrix and the Res1 matrix of the first sensor data at time t to t+h are obtained in parallel; (33) The IMF1 matrix, IMFK matrix and Res1 matrix of the first sensor data at time t to t+h are stacked to obtain the modal decomposition data matrix of the first sensor data at time t to t+h: (34) Using the method of steps (31) to (33), the modal decomposition data matrix of the 2nd to mth sensor data at time t to t+h is obtained: (4) The modal decomposition data matrices of the obtained sensor data at the time t to t+h are stacked to obtain the modal decomposition data matrices of all the data at the time t to t+h.

2. A signal data recognition method, characterized in that: The following steps are involved: (1) using the signal data fusion method as described in any one of claims 1 to obtain a modal decomposition data matrix of the data to be identified at time t to t+h; (2) Input the modal decomposition data matrix into the trained recognition model to complete the recognition of the data.

3. The signal data recognition method according to claim 2, characterized in that: The recognition model adopts a CNN deep learning network, including two convolutional layers, two pooling layers and a fully connected layer, wherein the first convolutional layer is provided with 64 convolutional kernels and the second convolutional layer is provided with 16 convolutional kernels.

Images