CN108564000B

CN108564000B - Fragment signal automatic identification method based on neural network

Info

Publication number: CN108564000B
Application number: CN201810228814.9A
Authority: CN
Inventors: 张斌; 李沅; 赵冬娥; 赵辉
Original assignee: North University of China
Current assignee: North University of China
Priority date: 2018-03-20
Filing date: 2018-03-20
Publication date: 2021-09-28
Anticipated expiration: 2038-03-20
Also published as: CN108564000A

Abstract

The invention discloses an automatic identification method of fragment signals based on neural network. The invention utilizes the extremely strong nonlinear mapping ability of BP neural network and the ability of associative memory for external stimuli and input information to improve the performance of the fragment speed test system. Correctly identify fragment over-target signals in large amounts of data.

Description

Fragment signal automatic identification method based on neural network

Technical Field

The invention belongs to the technical field of fragment signal automatic identification, and particularly relates to a fragment signal automatic identification method based on a neural network.

Background

In the process of acquiring the fragment over-target signal, a lot of noises interfere the accurate identification of the over-target signal, although the noises in the signal can be removed by a denoising method based on wavelet decomposition and reconstruction, the noises cannot be removed for the noise signals with frequency components similar to the fragment over-target signal, and a forward signal is formed instead after wavelet filtering, so that the subsequent identification of the fragment over-target signal is wrong.

Disclosure of Invention

In view of this, the present invention provides a method for automatically identifying a fragment signal based on a neural network, which can effectively remove the influence of various noises.

A fragment signal automatic identification method based on a neural network comprises the following steps:

step 1, collecting original data of a fragment speed testing process, and selecting a plurality of groups of typical fragment target-passing signals and a plurality of groups of noise signals with obvious peak values as sample data;

step 2, carrying out interference removal processing on the sample data obtained in the step 1 to obtain sample data for neural network training and testing;

step 3, carrying out differential processing on the sample data obtained in the step 2, then respectively obtaining 5 signal slopes at-6 sampling points, 0 sampling point, 3 sampling points and 6 sampling points, wherein the intervals between the sample data and the peak value are-3 sampling points, and simultaneously obtaining the signal pulse width at the peak value, thereby obtaining 6 characteristic parameter values of the fragment over-target signal;

step 4, carrying out classification marking on the sample data obtained in the step 2;

step 5, constructing a BP neural network comprising an input layer, two hidden layers and an output layer;

step 6, inputting the characteristic value of the sample data subjected to classification marking in the step 4 into the BP neural network established in the step 5, and training the BP neural network; obtaining a trained BP neural network model after the network error reaches a set condition;

and 7, acquiring 6 characteristic values of the input fragment signals by adopting the processing method in the step 3, and inputting the characteristic values into the BP neural network model trained in the step 6 for identification.

Preferably, the weight of the BP neural network is a random number between [ -1,1], and the bias is a random number between [0,1 ].

Preferably, the weight of the BP neural network is adjusted by adopting the principle of adjusting negative gradient descent.

Preferably, the setting error is 3 × 10^-7。

The invention has the following beneficial effects:

the invention utilizes the extremely strong nonlinear mapping capability of the BP neural network and the capability of associative memory of external stimulation and input information to improve the accurate recognition of the fragment over-target signals in a large amount of data in the fragment speed test system.

Drawings

FIG. 1 shows the fragment target-crossing signals detected by the system at different magnitudes and speeds;

FIG. 2 is a diagram illustrating the detection of a noise waveform by the system;

FIG. 3 is a process of extracting feature parameters of fragment over-target signals;

FIG. 4 is a diagram of a neural network training process;

FIG. 5 is a graph of the original waveform, peak and neural network filtered waveforms of noise and fragmentation signal characteristics.

Detailed Description

The invention is described in detail below by way of example with reference to the accompanying drawings.

Step 1, collecting an original oscillogram of a fragment signal, wherein fig. 1 is an effective over-target signal in the original oscillogram, and fig. 2 is noise which cannot be removed by a wavelet denoising method in the original signal.

And 2, filtering all effective over-target signals and all noises in the acquired original oscillogram of the fragment signals to remove high-frequency interference, and obtaining a sample for training the neural network.

And 3, observing and analyzing the difference between the over-target signal and the noise, and extracting characteristic parameters and classification rules according to the difference between the pulse width and the smoothness of the over-target signal and the noise.

And 4, constructing a BP neural network in MATLAB. The BP neural network is further described below. The constructed BP neural network is a four-layer neural network, the first layer of the network structure is an input layer, the second layer and the third layer of the network structure are hidden layers, and the last layer of the network structure is an output layer. Neurons in the same layer are not connected with each other, namely, connection weights are not arranged between the layers, and the connection mode between the layers is full connection, namely, all outputs of the previous layer can influence the layer.

The input of the BP neural network is a p-column vector in the dimension of R multiplied by 1. If the number of the nodes of the input layer of the BP neural network is N, the number of the nodes of the two hidden layers is R, Q respectively, and the number of the output nodes is M. w is a_ij(i 1, 2., N, j 1, 2., R) and b1j are the connection weights and offsets of the input layer and the first hidden layer, and v 1_jk(

j

1, 2.. said, R,

k

1, 2.. said, Q) and b2k are the connection weights and offsets of the first hidden layer and the second hidden layer, and m is the connection weight and offset of the first hidden layer and the second hidden layer_kl(k 1, 2., Q, l 1, 2., M) and b3l are connection weights and offset values of the second hidden layer and the output layer.

The function c between input and output can be obtained according to the weight and bias value of each layer network_kF (p) and defines an error function

Wherein y is_kIs the target vector.

And 5, calculating each parameter value in the neural network model according to the characteristic parameters of the known samples. For the smooth characteristic of the signal, the slope of the pulse is taken as a characteristic. Solving all peak values of the signal, wherein the obtained peak value information comprises the peak values of all fragment passing target signals and the peak value of noise, differentiating the signal with high-frequency interference removed to obtain the signal slope of-6 to-3, 0, 3 and 6 sampling points at the peak value interval, and simultaneously solving the signal pulse width at the peak value; and finally, obtaining 6 characteristic parameter values of the fragment target-passing signal, and then knowing that the value of R in the step 4 is 6. Fig. 3 is a flow of feature parameter extraction. Effective fragment signals all have certain peak values, in addition, the effective fragment signals are round and smooth, and different fragment signals have more consistent slope characteristics at-6, -3, 0, 3 and 6 sampling points away from the peak value. In addition, because the size and the speed value of the fragments are in a fixed interval, the pulse widths are consistent. The six features are different from the noise signal, so the six features of the fragment are selected as the feature parameters of the neural network.

And 6, adjusting the weight according to the principle that the BP neural network adjusts the negative gradient descent, namely initializing the weight and the offset value of the network, wherein the weight is a random number between [0 and 1], and the offset is a random number between [ -1 and 1 ]. And a group of weight values and bias values are arranged between every two layers.

And 7, inputting the sample filtered in the step 1 into the constructed neural network for training. The training of the neural network involves a number of iterations, each using all samples of the training set. And (3) using 6 parameter values to represent the fragment target-passing signal as an input, simultaneously defining 0 and 1 as output values of the neural network, representing the fragment by 1, representing noise by 0, and training the BP network by using known groups of data as training samples.

Step 8, setting the error to be 3 multiplied by 10^-7After 67 iterations, as shown in fig. 4, the network error reaches the set condition. Each iteration is performed, so that the weight matrix calculated last time is added with the change amount of the weight respectively. Thus, the weight matrix is continuously corrected and gradually tends to the target result.

And 9, identifying fragment over-target signals in the signals by using the obtained network model, reserving the fragment signals and setting all noise signals except the fragment signals to be zero. The multiple groups of signals are processed to obtain a comparison result of the original waveform of the noise and fragment signal characteristics and the neural network filtering, and fig. 5 shows the original waveform, the peak value and the neural network filtered waveform of the noise and fragment signal characteristics.

And step 10, calculating a new sample by using the trained network model, obtaining an output result through calculation, classifying the result according to the range of the result, and judging the classification attribution of the new sample.

And step 11, calculating the precision. The neural network is used for identifying the fragment speed measurement signal, the identification rate reaches 100%, and the error identification rate is 5.3%.

In summary, the above description is only a preferred embodiment of the present invention, and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. a method for automatic identification of fragment signals based on neural network, is characterized in that, comprises the steps:

Step 1. Collect the raw data of the fragment velocity test process, and select several groups of typical fragment passing target signals and several groups of noise signals with obvious peaks as sample data;

Step 2, performing interference removal processing on the sample data obtained in step 1, to obtain sample data for neural network training and testing;

Step 3. Differentiate the sample data obtained in step 2, and then obtain the peak interval of -6 sampling points, -3 sampling points, 0 sampling points, 3 sampling points and 6 sampling points respectively. At the same time, the signal pulse width at the peak value is obtained, and the 6 characteristic parameter values of the fragment passing target signal are obtained;

Step 4, classify and mark the sample data obtained in step 2;

Step 5. Construct a BP neural network including an input layer, two hidden layers and an output layer;

Step 6, input the eigenvalues of the sample data classified and marked in step 4 into the BP neural network established in step 5, and train it; after the network error reaches the set condition, a trained BP neural network model is obtained;

Step 7. For the input fragment signal, adopt the processing method of Step 3 to obtain 6 eigenvalues, and then input them into the BP neural network model trained in Step 6 for identification.

2. The method for automatic identification of fragment signals based on a neural network according to claim 1, wherein the weight of the BP neural network is a random number between [-1, 1], and the bias is [ A random number between 0,1].

3 . The method for automatic identification of fragment signals based on a neural network according to claim 2 , wherein the weights of the BP neural network are adjusted using the principle of adjusting negative gradient descent. 4 .

4 . The method for automatic identification of fragment signals based on neural network according to claim 2 , wherein the setting error is 3×10 ⁻⁷ .