Classification and recognition method of particle light scattering signals based on machine learning

Abstract

The present invention discloses a particle light scattering signal classification and recognition method based on machine learning, specifically: collecting light scattering signals of water droplets, coal powder, and polystyrene particles and performing noise reduction preprocessing, extracting 22 characteristic indicators, calculating the weight proportion of characteristic indicators, selecting 4 indicators with the highest weight proportion as the optimal characteristic vector, and arranging and combining the optimal characteristic variables to form a sample feature set. The label set is composed of the numbers of three kinds of particles, namely water droplets, coal powder, and polystyrene. The label sets correspond to the sample feature sets respectively to form a neural network data set, and are divided into a training set and a test set, which are respectively input into the GRNN and PNN neural networks for training and testing, so as to realize the classification and recognition of light scattering signals of three kinds of particles, namely water droplets, coal powder, and polystyrene. The method is suitable for accurate detection of atmospheric particle properties, determining the types and contributions of emission sources, and is an important decision-making basis for tracing the source.

Description

Technical Field

[0001] The present invention belongs to the technical field of digital signal processing, and in particular relates to a particle light scattering signal classification and recognition method based on machine learning. Background Art

[0002] Particles are widely present in nature, scientific research and engineering applications. Theoretical and technical issues related to particles are increasingly appearing in various fields such as medicine, atmospheric environment monitoring, and materials science. In the field of atmospheric environment monitoring, the study of atmospheric particles has become the focus of atmospheric monitoring. In addition, particles are closely related to human health. Harmful particles in the air, such as acidic particles, heavy metal particles, and volatile organic compounds, enter the human body and settle in the respiratory tract and alveoli, causing serious harm to human health. Therefore, accurate detection of the properties of atmospheric particles can determine the type and contribution of emission sources, so as to effectively control emission sources, thereby effectively controlling air pollution and improving air quality. Summary of the invention

[0003] The purpose of the present invention is to provide a particle light scattering signal classification and recognition method based on machine learning, which improves the recognition efficiency and accuracy of particles.

[0004] The technical solution adopted by the present invention is a particle light scattering signal classification and recognition method based on machine learning, which is specifically implemented according to the following steps:

[0005] Step 1, select three particles with different refractive indices, namely water droplets, coal powder and polystyrene particles, for experiment, and use a data acquisition card to complete the acquisition of light scattering signals of each particle through a photoelectric detector;

[0006] Step 2: Using the EMD-ICA algorithm to perform noise reduction preprocessing on the particle light scattering signal collected in step 1;

[0007] Step 3, extracting 22 characteristic indicators from the preprocessed particle light scattering signal;

[0008] Step 4: Use the ReliefF algorithm to calculate the weight ratios of the 22 characteristic indicators extracted in step 3, and select the 4 indicators with the highest weight ratios as the optimal characteristic vectors;

[0009] Step 5: Arrange and combine the four optimal feature variables according to their numbers to form 10 sample feature sets. The label set is composed of the numbers of three particles with different properties: water droplets, coal powder, and polystyrene. The label sets correspond to the 10 sample feature sets respectively to form 10 neural network data sets.

[0010] Step 6: Divide the 10 neural network data sets obtained in step 5 into training sets and test sets, and input them into the GRNN and PNN neural networks for training and testing, respectively, to achieve classification and recognition of particle light scattering signals of three particles with different properties: water droplets, coal powder, and polystyrene.

[0011] The present invention is also characterized in that:

[0012] In step 2, specifically:

[0013] Step 2.1: Perform EMD empirical mode decomposition on the collected noisy particle light scattering signal to obtain the particle light scattering signal intrinsic mode function IMFs components imf1, imf2, imf3,…, imf n and the residual component r n ;

[0014] Step 2.2, calculate and obtain the correlation coefficients and variance contribution rates of several IMFs components, use the modal correlation sorting criterion to determine the dividing point between the signal-dominant component and the noise-dominant component, and construct a virtual noise channel by linearly superposing the divided noise-dominant components;

[0015] Step 2.3, the virtual noise channel and the noisy particle light scattering signal are mixed, and the ICA algorithm is used to separate the particle light scattering signal from the noise, thereby completing the denoising preprocessing of the particle light scattering signal.

[0016] In step 3, the 22 characteristic indicators are maximum, mean, peak-to-peak value, rectified mean, variance, standard deviation, kurtosis, skewness, root mean square, waveform factor, peak factor, pulse factor, margin factor, centroid frequency, mean square frequency, root mean square frequency, frequency variance, frequency standard deviation, power spectrum entropy, singular spectrum entropy, and energy entropy.

[0017] In step 6, specifically: the 10 neural network data sets obtained in step 5 are randomly divided into training sets and test sets, among which 120 groups of data are selected as training sets and 30 groups of data are selected as test sets; the 10 divided neural network data sets are respectively input into the GRNN and PNN neural networks for training and testing, and the classification and recognition accuracy of the GRNN and PNN neural networks for the 10 test set data is compared, and the feature vector model corresponding to the neural network data set with the highest accuracy is selected as the optimal classification result to realize the classification and recognition of light scattering signals of three particles with different properties: water droplets, coal powder, and polystyrene.

[0018] The beneficial effect of the present invention is that the particle light scattering signal obtained by simulation and experimental verification is taken as the research object, the noise reduction preprocessing of the signal is completed by the EMD-ICA algorithm, and then the feature extraction and attribute recognition of the particle light scattering signal can be effectively realized by means of machine learning and data analysis; this method is suitable for the accurate detection of atmospheric particle properties, determining the type and contribution of emission sources, and is an important decision-making basis for tracing the source. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flowchart of the particle light scattering signal classification and recognition method based on machine learning of the present invention;

[0020] Figure 2 It is a graph of classification and recognition accuracy of GRNN and PNN neural networks in the particle light scattering signal classification and recognition method based on machine learning of the present invention;

[0021] Figure 3 It is a classification result diagram corresponding to the feature vector model with the highest classification and recognition accuracy of the GRNN and PNN neural networks in the particle light scattering signal classification and recognition method based on machine learning of the present invention. DETAILED DESCRIPTION

[0022] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] The particle light scattering signal classification and recognition method of the present invention is based on machine learning, such as Figure 1 As shown, the specific implementation steps are as follows:

[0024] Step 1, select water droplets, coal powder, and polystyrene particles with different refractive indices for the experiment, use a data acquisition card to complete the acquisition of light scattering signals of each particle through a photoelectric detector, and collect 50 groups of light scattering signals of water droplets, coal powder, and polystyrene particles, for a total of 150 groups;

[0025] Step 2: Using the EMD-ICA algorithm, the particle light scattering signal collected in step 1 is subjected to noise reduction preprocessing;

[0026] Firstly, several IMFs modal components are obtained by performing EMD empirical mode decomposition on the collected noisy particle light scattering signal, and the correlation coefficients and variance contribution rates of several IMFs modal components are calculated. The demarcation point between the signal-dominant component and the noise-dominant component is determined by the modal correlation sorting criterion, and a virtual noise channel is constructed by linear superposition of the noise-dominant component. The virtual noise channel and the noisy particle light scattering signal are used as the input of the ICA algorithm, and finally the particle light scattering signal and noise are separated, and the data denoising preprocessing is completed; specifically:

[0027] Step 2.1: Perform EMD empirical mode decomposition on the collected noisy particle light scattering signal to obtain the particle light scattering signal intrinsic mode function IMFs components imf1, imf2, imf3,…, imf n and the residual component r n ;

[0028] Step 2.2, calculate and obtain the correlation coefficients and variance contribution rates of several IMFs components, use the modal correlation sorting criterion to determine the dividing point between the signal-dominant component and the noise-dominant component, and construct a virtual noise channel by linearly superposing the divided noise-dominant components;

[0029] Step 2.3, mixing the virtual noise channel and the noisy particle light scattering signal, using the ICA algorithm to separate the particle light scattering signal from the noise, and completing the denoising preprocessing of the particle light scattering signal;

[0030] Step 3, extracting the maximum value, mean value, peak-to-peak value, rectified mean value, variance, standard deviation, kurtosis, skewness, root mean square, waveform factor, peak factor, pulse factor, margin factor, centroid frequency, mean square frequency, root mean square frequency, frequency variance, frequency standard deviation, power spectrum entropy, singular spectrum entropy, and energy entropy characteristic indicators from the preprocessed particle light scattering signal to form a sample data set;

[0031] There are 50 sets of sample feature data sets corresponding to the light scattering signals of water droplets, coal powder, and polystyrene particles, for a total of 150 sets;

[0032] The maximum value calculation formula is shown in formula (1) and formula (2):

[0033] F1=max(x(i)) (1);

[0034] F2 = min(x(i)) (2);

[0035] Among them, F1 is the maximum value, F2 is the minimum value; x(i) represents the sampled particle light scattering signal value;

[0036] The calculation formula of mean F3 is shown in formula (3):

[0037]

[0038] Among them, N s represents the number of sampled particle light scattering signals; the calculation formula of the peak-to-peak value F4 is shown in formula (4):

[0039] F4=max(x(i))-min(x(i)) (4);

[0040] The calculation formula of the rectified average value F5 is shown in formula (5):

[0041]

[0042] The calculation formula of variance F6 is shown in formula (6):

[0043]

[0044] The calculation formula of standard deviation F7 is shown in formula (7):

[0045]

[0046] The calculation formula of kurtosis F8 is shown in formula (8):

[0047]

[0048] The calculation formula of skewness F9 is shown in formula (9):

[0049]

[0050] RMS F 10 The calculation formula is shown in formula (10):

[0051]

[0052] Form Factor F 11 The calculation formula is shown in formula (11):

[0053]

[0054] Peak Factor F 12 The calculation formula is shown in formula (12):

[0055]

[0056] Pulse Factor F 13 The calculation formula is shown in formula (13):

[0057]

[0058] Margin factor F 14 The calculation formula is shown in formula (14):

[0059]

[0060] Center of gravity frequency F 15 The calculation formula is shown in formula (15):

[0061]

[0062] Among them, f i Indicates the frequency value corresponding to the spectrum of the signal at time i, p iRepresents the amplitude of the signal in the frequency domain at time i;

[0063] Mean square frequency F 16 The calculation formula is shown in formula (16):

[0064]

[0065] RMS frequency F 17 The calculation formula is shown in formula (17):

[0066]

[0067] Frequency variance F 18 The calculation formula is shown in formula (18):

[0068]

[0069] Frequency standard deviation F 19 The calculation formula is shown in formula (19):

[0070]

[0071] Power spectrum entropy F 20 The calculation steps are as follows:

[0072] First, the power spectrum of the signal is obtained using discrete Fourier transform, as shown in formula (20):

[0073]

[0074] In the formula, x(w i ) represents the discrete Fourier transform of each sample in the signal x(n);

[0075] Normalizing the power spectrum according to the total spectral power yields equation (21), and calculating the corresponding power spectrum entropy F 20 , as shown in formula (22):

[0076]

[0077] Singular spectral entropy F 21 The calculation steps are as follows:

[0078] Process x(n) into matrix A: Perform singular value decomposition on matrix A: A = USV T ;

[0079] Where U and V are both characteristic matrices, S is the eigenvalue matrix; N eigenvalues ​​s from large to small i (i=1,2,…,N) constitutes a singular spectrum;

[0080] Find the proportion of each singular value to the sum of all singular values ​​and get the probability density function P i , as shown in formula (23):

[0081]

[0082] Singular spectral entropy F 21 The calculation formula is shown in formula (24):

[0083]

[0084] Energy entropy F 22 The calculation formula is shown in formula (25):

[0085]

[0086] Among them, E i Indicates the proportion of each component energy to the total signal energy:

[0087] Step 4: Use the ReliefF algorithm in machine learning feature selection to calculate the weight ratios of the 22 feature indicators extracted in step 3, and select the 4 indicators with the highest weight ratios as the optimal feature vectors;

[0088] Step 5. Arrange and combine the four optimal feature variables selected in step 4 according to the numbers a, b, c, and d to obtain 10 feature vector models, namely {a}, {a,b}, {a,b,c}, {a,b,c,d}, {b}, {b,c}, {b,c,d}, {c}, {c,d}, and {d}, which constitute the required 10 sample feature sets. Each sample feature set corresponds to 50 groups of water droplets, coal powder, and polystyrene particles, for a total of 150 groups of data; the label set is composed of the numbers of three different attribute particles of water droplets, coal powder, and polystyrene, respectively recorded as {1}, {2}, and {3}, each corresponding to 50 groups, for a total of 150 groups. The label sets correspond to the 10 sample feature sets to form 10 neural network data sets;

[0089] Step 6: Randomly divide the 10 neural network data sets obtained in step 5 into training sets and test sets, among which 120 groups of data are selected as training sets and 30 groups of data are selected as test sets. The 10 divided neural network data sets are respectively input into the GRNN and PNN neural networks for training and testing, and the accuracy of classification and recognition of the 10 test set data by the GRNN and PNN neural networks is compared. The feature vector model corresponding to the neural network data set with the highest accuracy is selected as the optimal classification result to realize the classification and recognition of particle light scattering signals of three different attribute particles: water droplets, coal powder, and polystyrene.

[0090] Example

[0091] The particle light scattering signal classification and recognition method based on machine learning of the present invention is specifically implemented according to the following steps:

[0092] Step 1: Using a data acquisition card to complete the collection of light scattering signals of particles with different refractive indices, namely, water droplets, coal powder, and polystyrene particles, through a photoelectric detector, 50 groups of light scattering signals of water droplets, coal powder, and polystyrene particles are collected, for a total of 150 groups;

[0093] Step 2: Using the EMD-ICA algorithm, the particle light scattering signal collected in step 1 is subjected to noise reduction preprocessing;

[0094] Step 2.1: Perform EMD empirical mode decomposition on the light scattering signals of water droplets, coal powder, and polystyrene noisy particles to obtain the intrinsic mode function IMFs components imf1, imf2, imf3, …, imf n and the residual component r n ;

[0095] Step 2.2, calculate and obtain the correlation coefficients and variance contribution rates of several IMFs components, use the modal correlation sorting criterion to determine the dividing point between the signal-dominant component and the noise-dominant component, and construct a virtual noise channel by linearly superposing the divided noise-dominant components;

[0096] Step 2.3, mixing the virtual noise channel and the noisy particle light scattering signal, using the ICA algorithm to separate the particle light scattering signal from the noise, and completing the denoising preprocessing of the particle light scattering signal;

[0097] Step 3, extracting the maximum value, mean value, peak-to-peak value, rectified mean value, variance, standard deviation, kurtosis, skewness, root mean square, waveform factor, peak factor, pulse factor, margin factor, centroid frequency, mean square frequency, root mean square frequency, frequency variance, frequency standard deviation, power spectrum entropy, singular spectrum entropy, and energy entropy characteristic indicators from the preprocessed particle light scattering signal to form a sample data set;

[0098] Step 4: Use the ReliefF algorithm in machine learning feature selection to calculate the weight ratios of the 22 feature indicators extracted in step 3, and select the 4 indicators with the highest weight ratios as the optimal feature vectors;

[0099] Step 5: Construct a neural network dataset. The neural network dataset includes a label set and a sample feature set. The specific steps are as follows:

[0100] Step 5.1, mark the four optimal characteristic variables waveform factor, pulse factor, margin factor and singular value entropy selected in step 4 as a, b, c, d respectively and arrange and combine them;

[0101] Step 5.2: Get 10 sample feature sets, {a}, {a,b}, {a,b,c}, {a,b,c,d}, {b}, {b,c}, {b,c,d}, {c}, {c,d}, and {d}. Each sample feature set corresponds to 50 groups of water droplets, 50 groups of coal powder, and 50 groups of polystyrene particles, for a total of 150 groups of data.

[0102] Step 5.3: The label set is composed of the numbers of three particles with different properties: water droplets, coal powder, and polystyrene, which are recorded as {1}, {2}, and {3}, respectively, corresponding to 50 groups each, for a total of 150 groups.

[0103] Step 5.4, correspond the label sets to 10 sample feature sets respectively, form 10 neural network data sets, and complete the construction of the neural network data set;

[0104] Step 6: Build GRNN and PNN neural network models, complete the training and testing of the neural network data set, and realize the classification and recognition of particle light scattering signals of three different particle properties: water droplets, coal powder, and polystyrene. The specific steps are as follows:

[0105] Step 6.1: Randomly divide the 10 neural network data sets obtained in step 5 into training sets and test sets, where 120 sets of data are selected as training sets and 30 sets of data are selected as test sets.

[0106] Step 6.2: Build a GRNN neural network with four layers of network structure, including input layer, pattern layer, summation layer and output layer. Select 120 groups of 150 data sets as training sets, and the remaining 30 groups as test sets. The feature dimension of the data set is 4, the label dimension is 1, and the constructed GRNN neural network structure is (4:120:2:1).

[0107] Step 6.3, build a PNN neural network including input layer, pattern layer, summation layer and output layer. Select 120 of the 150 data sets as training sets and the remaining 30 as test sets. The feature dimension of the data set is 4, the category is 3 different substances, the label dimension is 1, and the constructed PNN neural network structure is (4:120:3:1).

[0108] Step 6.4: Input the 10 divided neural network data sets into the GRNN and PNN neural networks for training and testing respectively, compare the classification and recognition accuracy of the GRNN and PNN neural networks for the 10 test set data, select the feature vector model corresponding to the neural network data set with the highest accuracy as the optimal classification result, and realize the classification and recognition of particle light scattering signals of three particles with different properties: water droplets, coal powder, and polystyrene.

[0109] Figure 2This is a graph of the classification and recognition accuracy of GRNN and PNN neural networks in the particle light scattering signal classification and recognition method based on machine learning of the present invention. The figure shows 10 models generated corresponding to the 10 neural network data sets in step 5, and the test set is classified and recognized by GRNN and PNN neural networks, among which model 4 (combination of waveform factor, pulse factor, margin factor and singular value entropy) and model 7 (combination of pulse factor, margin factor and singular value entropy) have the highest accuracy through the PNN neural network, reaching 96.7%, and model 1 (combination of waveform factor) and model 10 (combination of singular value entropy) have the lowest accuracy through the GRNN neural network, which is 33.3%.

[0110] Figure 3 This is the classification result diagram corresponding to the feature vector model with the highest classification and recognition accuracy of the GRNN and PNN neural networks in the particle light scattering signal classification and recognition method based on machine learning of the present invention, that is, the classification result diagram of the test set corresponding to model 4 (a combination of waveform factor, pulse factor, margin factor and singular value entropy) through the GRNN and PNN neural networks. In the figure, for the classification of category 1 (water droplets): the GRNN neural network has 3 misidentifications, and the PNN neural network has 1 misidentification; for the classification of category 2 (polystyrene) in the figure: the GRNN neural network has 3 misidentifications, and the PNN neural network has all correct identifications; for the classification of category 3 (coal powder) in the figure: the GRNN neural network has all correct identifications, and the PNN neural network has all correct identifications. In total, the GRNN neural network has 6 misidentifications, with an accuracy rate of 80%, and the PNN neural network has 1 misidentification, with an accuracy rate of 96.7%.

[0111] The present invention takes the particle light scattering signals of water droplets, coal powder and polystyrene obtained by experimental collection as the research objects, and uses the EMD-ICA algorithm to reconstruct the virtual noise channel to complete the denoising preprocessing of the particle light scattering signals; then extracts 22 characteristic indicators such as waveform factor, peak factor time-frequency domain characteristics and sample entropy from the preprocessed particle light scattering signals, uses the ReliefF algorithm in machine learning feature selection to complete the selection of the optimal characteristic variables, constructs a sample data set, and inputs it into the GRNN and PNN neural networks to complete training and classification recognition. The present invention can simply and effectively distinguish the particle light scattering signals of the three particles with different properties of water droplets, coal powder and polystyrene, and the method has a recognition accuracy rate of 96.7% for the particle light scattering signals of the three particles with different properties of water droplets, coal powder and polystyrene.

Claims

1. A particle light scattering signal classification and recognition method based on machine learning, characterized in that: Follow the steps below to implement it: Step 1, select three particles with different refractive indices, namely water droplets, coal powder and polystyrene particles, for experiment, and use a data acquisition card to complete the acquisition of light scattering signals of each particle through a photoelectric detector; Step 2: Use the EMD-ICA algorithm to perform noise reduction preprocessing on the particle light scattering signal collected in step 1; specifically: Step 2.1: Perform EMD empirical mode decomposition on the collected noisy particle light scattering signal to obtain the particle light scattering signal intrinsic mode function IMFs components imf1, imf2, imf3,…,imf n and the residual component r n ; Step 2.2, calculate and obtain the correlation coefficients and variance contribution rates of several IMFs components, use the modal correlation sorting criterion to determine the dividing point between the signal-dominant component and the noise-dominant component, and construct a virtual noise channel by linearly superposing the divided noise-dominant components; Step 2.3, mixing the virtual noise channel and the noisy particle light scattering signal, using the ICA algorithm to separate the particle light scattering signal from the noise, and completing the denoising preprocessing of the particle light scattering signal; Step 3, extracting 22 characteristic indicators from the preprocessed particle light scattering signal; Step 4: Use the ReliefF algorithm to calculate the weight ratios of the 22 characteristic indicators extracted in step 3, and select the 4 indicators with the highest weight ratios as the optimal characteristic vectors; Step 5: Arrange and combine the four optimal feature variables according to their numbers to form 10 sample feature sets. The label set is composed of the numbers of three particles with different properties: water droplets, coal powder, and polystyrene. The label sets correspond to the 10 sample feature sets respectively to form 10 neural network data sets. Step 6: Divide the 10 neural network data sets obtained in step 5 into training sets and test sets, and input them into the GRNN and PNN neural networks for training and testing, respectively, to achieve classification and recognition of particle light scattering signals of three particles with different properties: water droplets, coal powder, and polystyrene.

2. The particle light scattering signal classification and recognition method based on machine learning according to claim 1 is characterized in that: In step 3, the 22 characteristic indicators are respectively maximum, mean, peak-to-peak value, rectified mean, variance, standard deviation, kurtosis, skewness, root mean square, waveform factor, peak factor, pulse factor, margin factor, center of gravity frequency, mean square frequency, root mean square frequency, frequency variance, frequency standard deviation, power spectrum entropy, singular spectrum entropy, and energy entropy.

3. The particle light scattering signal classification and recognition method based on machine learning according to claim 1 is characterized in that: In the step 6, specifically: the 10 neural network data sets obtained in step 5 are randomly divided into training sets and test sets, among which 120 groups of data are selected as training sets, and 30 groups of data are selected as test sets; the 10 divided neural network data sets are respectively input into the GRNN and PNN neural networks for training and testing, and the accuracy of classification and recognition of the 10 test set data by the GRNN and PNN neural networks is compared, and the feature vector model corresponding to the neural network data set with the highest accuracy is selected as the optimal classification result, so as to realize the classification and recognition of light scattering signals of three particles with different properties: water droplets, coal powder, and polystyrene.

Images