CN108564000A - A kind of fragmentation signal automatic-identifying method based on neural network - Google Patents

Abstract

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

Description

A method for automatic identification of fragment signals based on neural network

technical field

The invention belongs to the technical field of fragment signal automatic recognition, and in particular relates to a neural network-based automatic fragment signal recognition method.

Background technique

In the process of fragment passing signal acquisition, there will be many noises that interfere with the accurate identification of passing signals. Although the noise in the signal can be removed by the denoising method based on wavelet decomposition and reconstruction, for some frequency components and fragment passing The noise signal with relatively similar signal cannot be removed, and instead forms a positive signal after wavelet filtering, which makes the subsequent recognition of the fragment passing the target signal wrong.

Contents of the invention

In view of this, the object of the present invention is to provide a method for automatic identification of fragment signals based on neural network, which can more effectively remove the influence of various noises.

A method for automatic recognition of fragment signals based on neural networks, comprising the steps of:

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

Step 2. De-interference processing is performed on the sample data obtained in step 1 to obtain sample data for neural network training and testing;

Step 3. Perform differential processing on the sample data obtained in step 2, and then obtain the peak intervals of -6 sampling points, -3 sampling points, 0 sampling points, 3 sampling points and 6 sampling points A total of 5 signal slopes at the point, and at the same time calculate the signal pulse width at the peak, and thus obtain 6 characteristic parameter values of the fragment passing the target signal;

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

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

Step 6, input the eigenvalues of the sample data that are classified and marked in step 4 into the BP neural network set up in step 5, and train it; the network error reaches the set condition to obtain the trained BP neural network model;

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

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 weights of the BP neural network are adjusted using the principle of adjusting negative gradient descent.

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

The present invention has following beneficial effects:

The invention utilizes the extremely strong non-linear mapping capability of the BP neural network and the capability of associative memory for external stimuli and input information to improve the correct recognition of fragment passing target signals in a large amount of data in a fragment velocity testing system.

Description of drawings

Figure 1 shows the signals of fragments passing the target at different sizes and speeds detected by the system;

Figure 2 is the noise waveform detected by the system;

Fig. 3 is the procedure for extracting the characteristic parameters of the fragment passing the target signal;

Fig. 4 is a neural network training process diagram;

Figure 5 shows the original waveform, peak value and neural network filtered waveform of noise and fragment signal characteristics.

Detailed ways

The present invention will be described in detail below with reference to the accompanying drawings and examples.

Step 1. Collect the original wave form of the fragment signal. Figure 1 shows the effective cross-target signal in the original picture, and Figure 2 shows the noise in the original signal that cannot be removed by the wavelet denoising method.

Step 2. Filter all effective target passing signals and all noises in the collected fragment signal original waveform diagram to remove high-frequency interference, and obtain training samples for the neural network.

Step 3. Observe and analyze the difference between the passing signal and noise, and extract the characteristic parameters and classification rules according to the difference in pulse width and smoothness of the passing signal and noise.

Step 4, construct BP neural network in MATLAB. The BP neural network is further described below. The BP neural network constructed is a four-layer neural network, the first layer of the network structure is the input layer, the second and third layers of the network structure are hidden layers, and the last layer of the network structure is the output layer. The neurons in the same layer are not connected to each other, that is, there is no connection weight between the same layers, and the connection mode between layers is full connection, that is, all the outputs of the previous layer will have an impact on this layer.

The input of the BP neural network is a p-column vector of R×1 dimension. If the number of nodes in the input layer of the BP neural network is N, the number of nodes in the two hidden layers is R and Q respectively, and the number of output nodes is M. w _ij (i=1,2,...,N, j=1,2,...,R) and b1j are the connection weights and bias values of the input layer and the first hidden layer, v _jk (j =1,2,...,R, k=1,2,...,Q) and b2k are the connection weights and bias values of the first hidden layer and the second hidden layer, m _kl (k=1 ,2,...,Q, l=1,2,...,M) and b3l are the connection weights and bias values of the second hidden layer and the output layer.

According to the weight and bias value of each layer network, the function c _k =f(p) between input and output can be obtained, and the error function is defined where y _k is the target vector.

Step 5. Calculate the value of each parameter in the neural network model according to the characteristic parameters of the known samples. The smoothness of the signal is characterized by the slope of the pulse. Calculate all its peak values, and the obtained peak information includes the peak values of all fragments passing the target signal and the peak value of noise. After differential processing of the signal that filters out high-frequency interference, the peak intervals obtained are -6, -3, 0, 3, The slope of the signal at the 6 sampling points and the pulse width of the signal at the peak value are calculated at the same time; finally, the 6 characteristic parameter values of the fragment passing the target signal are obtained, and the value of R in step 4 is known to be 6. Figure 3 is the feature parameter extraction process. The effective fragmentation signals all have a certain peak value, and the effective fragmentation signals are relatively round and smooth. Different fragmentation signals have relatively consistent slope characteristics at the sampling points of -6, -3, 0, 3, and 6 from the peak value. In addition, because the fragment size and velocity values are in a relatively fixed interval, the pulse width is also relatively consistent. These six features are different from the noise signal, so these six features of the fragments are selected as the characteristic parameters of the neural network.

Step 6. Adjust the weight according to the principle of BP neural network to adjust the negative gradient descent, that is, initialize the weight and bias value of the network, the weight is a random number between [0,1], and the bias is [-1 ,1] is a random number. Set a set of weights and bias values between every two layers.

Step 7. Input the sample filtered in step 1 to the constructed neural network for training. The training of the neural network includes multiple iterations, and each iteration (training) uses all the samples in the training set. Use 6 parameter values to characterize the fragment passing signal, take it as input, define 0 and 1 as the output value of the neural network at the same time, use 1 to represent fragments, 0 to represent noise, and use multiple sets of known data as training samples to train the BP network .

Step 8, the set error is 3×10 ^-7 , as shown in Fig. 4 , after 67 iterations, the network error reaches the set condition. Each iteration adds the change amount of the weight to the weight matrix calculated last time. In this way, the weight matrix is constantly revised and gradually tends to the target result.

Step 9. Use the obtained network model to identify the fragment passing signal in the signal, retain the fragment signal and set all noise signals other than the fragment signal to zero. Multiple groups of signals are processed to obtain the comparison results of the original waveforms of noise and fragment signal characteristics and neural network filtering. Figure 5 shows the original waveforms, peak values of noise and fragment signal characteristics, and waveforms after neural network filtering.

Step 10. Use the trained network model to calculate new samples, obtain output results through calculation, classify the results according to the range of the results, and determine the classification of the new samples.

Step 11. Calculate the accuracy. The neural network is used to identify the fragment speed measurement signal, the identification rate is 100%, and the false identification rate is 5.3%.

To sum up, the above are only preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims (4)

1. a kind of fragment signal automatic recognition method based on neural network, is characterized in that, comprises the steps: Step 1. Collect the original data of the fragment velocity test process, and select several groups of typical fragment passing signals and several groups of noise signals with obvious peaks as sample data; Step 2. De-interference processing is performed on the sample data obtained in step 1 to obtain sample data for neural network training and testing; Step 3. Perform differential processing on the sample data obtained in step 2, and then obtain the peak intervals of -6 sampling points, -3 sampling points, 0 sampling points, 3 sampling points and 6 sampling points A total of 5 signal slopes at the point, and at the same time calculate the signal pulse width at the peak, and thus obtain 6 characteristic parameter values of the fragment passing the target signal; Step 4, classify and mark the sample data obtained in step 2; Step 5, constructing a BP neural network comprising an input layer, two hidden layers and an output layer; Step 6, input the eigenvalues of the sample data that are classified and marked in step 4 into the BP neural network set up in step 5, and train it; the network error reaches the set condition to obtain the trained BP neural network model; Step 7. For the input fragment signal, adopt the processing method of step 3 to obtain 6 eigenvalues, and then input them into the trained BP neural network model in step 6 for recognition. 2. a kind of fragment signal automatic recognition method based on neural network as claimed in claim 1, is characterized in that, the weight of described BP neural network gets a random number between [-1,1], and offset gets [ A random number between 0,1]. 3. a kind of fragment signal automatic recognition method based on neural network as claimed in claim 2, is characterized in that, adopts the principle of adjusting negative gradient descent to adjust described BP neural network weight. 4. A neural network-based automatic fragment signal identification method according to claim 2, characterized in that the setting error is 3×10 ^-7 .