A residual learning lightweight convolutional neural network quantification method for LIBS

Abstract

The application discloses a residual learning lightweight convolutional neural network quantitative method for LIBS. The method comprises the following steps: collecting laser-induced breakdown spectroscopy (LIBS) original spectrum data of training samples; performing spectrum selection on each original spectrum and taking the mean value of the selected spectrum; establishing and training a residual learning lightweight convolutional neural network model by using the selected original spectrum; and performing quantitative analysis on test samples by using the trained residual learning lightweight convolutional neural network model. The application performs quantitative analysis of LIBS by using the residual learning convolutional neural network model. The residual module in the model improves the signal-to-noise ratio and retains the details of the spectrum signal, combines data preprocessing and quantitative analysis organically, and makes the preprocessing more targeted. Attention is concentrated on different effective information to different degrees, and feature selection is completed. The expression ability of the complex nonlinear relationship of the model is improved, and the precision of the LIBS quantitative analysis is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of spectral analysis, and particularly relates to a residual learning light convolutional neural network quantification method for LIBS. BACKGROUND

[0002] The laser-induced breakdown spectroscopy (LIBS) technology is an atomic emission spectroscopy technology taking laser-induced plasma as a light source. The LIBS is very suitable for in-situ analysis in the fields of geological and soil exploration, deep sea and deep space exploration, and online analysis in the industrial fields of mineral extraction, metallurgy, etc. due to its simple structure, little or no sample preparation, simultaneous online analysis of multiple elements, remote analysis, etc. However, it is difficult to obtain a reliable quantitative analysis model due to the following two points, which restricts the application and development of the LIBS technology in the field of online quantitative analysis. (1) The physical parameters of the plasma formed by the strike are different due to the noise such as photon scattering and photodetector, and therefore the noise is regarded as interference of the effective LIBS signal, which hinders the repeatability and accuracy of the measurement. (2) The concentration of the element to be measured and the spectral line intensity present a complex multivariate nonlinear relationship in the actual application of the LIBS technology due to the influence of the matrix effect and self-absorption effect, and it is difficult to express the complex nonlinear relationship by simple machine learning or neural network in the case of a small sample. SUMMARY

[0003] The present application proposes a residual learning light convolutional neural network quantification method for LIBS aiming at the above two problems, which aims to improve the signal-to-noise ratio of the spectral signal while retaining useful detail signals, and to organically integrate the pretreatment and quantitative analysis together. The pretreatment part improves the signal-to-noise ratio of the signal, and the quantitative analysis part completes the expression of the complex nonlinear relationship of the small sample, so that the pretreatment is more targeted, and the pretreatment and quantitative analysis can be completed by the same model, improving the efficiency of the model. In this way, the interference of the noise can be effectively solved, and the complex nonlinear relationship caused by the self-absorption effect and the matrix effect can be effectively expressed.

[0004] To this end, the present application is implemented by adopting the following technical scheme: a residual learning light convolutional neural network quantification method for LIBS, comprising the following steps:

[0005] Step 1: collecting laser-induced breakdown spectroscopy original spectral data of training samples;

[0006] Step 2: selecting spectral segments for the original spectrum;

[0007] Step 3: establishing and training a residual learning light convolutional neural network model by using the selected original spectral segments;

[0008] Step 4: using the trained residual learning lightweight convolutional neural network model to quantitatively analyze the test sample, and output the predicted element concentration.

[0009] is to take the average spectrum of multiple spectra after laser hitting multiple times for each sample as the original spectrum data.

[0010] The spectral segment where the characteristic spectral line related to the element to be measured is located is selected as the original spectral segment.

[0011] The specific network structure of the residual learning lightweight convolutional neural network model is as follows:

[0012] The residual module layer Residual Module includes: a plurality of residual modules connected in sequence, and the difference between the input and output of the current residual module is taken as the input of the next residual module, and each residual module includes a plurality of residual units connected in sequence; for processing the selected original spectrum segment, output the extracted spectrum segment information s;

[0013] The lightweight convolutional neural module layer includes: a plurality of convolutional layers, Inception Module layers, flattening layers Flatten, and fully connected layers FC connected in sequence; for further feature extraction of the spectrum segment information s output by the residual module layer, output the predicted element concentration y.

[0014] The residual unit includes a convolution structure, batch normalization BN, and a Relu activation function.

[0015] The formula of the Relu activation function is: f(x)=max(0,x).

[0016] The convolutional layer of the lightweight convolutional neural module layer deletes the nonlinear activation and pooling operation.

[0017] The Inception Module layer is composed of four branches, each branch includes a 1×1 convolution structure, and the four branches are nonlinearized by a tanh activation function before merging; the fully connected layer FC adopts a tanh activation function for nonlinearization.

[0018] The training residual learning lightweight convolutional neural network model comprises:

[0019] The target function is calculated by the following formula, and the back propagation is performed under the condition of minimizing the target function, and the residual module output is the preprocessed spectrum s, and the lightweight convolutional neural module output is the element concentration y to be measured after continuous iteration when the training reaches the iteration number;

[0020]

[0021] Wherein, M is the number of training samples, yn is the label concentration value of the nth training sample, is the model predicted concentration value, and is a penalty factor, and ||w|| is the weight of the network layer. 2 is the network layer weight, and in the above formula, an L2 regularization constraint term is added to the hidden layer to limit the weight value w, so that w is as close to 0 as possible.

[0022] The root mean square error RMSE and the determination coefficient R between the element content predicted by the quantitative analysis model and the reference label are quantitatively analyzed. 2 The index is used as an evaluation function to further optimize the network parameters to obtain the final quantitative analysis model.

[0023] Advantages and technical features of the present application:

[0024] The present application performs LIBS spectrum preprocessing and quantitative analysis by using a residual learning lightweight convolutional neural network model. The residual module in the model improves the signal-to-noise ratio and also retains the details of the spectral signal. The data preprocessing and quantitative analysis are organically combined together, so that the preprocessing is more targeted. Attention is concentrated on different effective information to different degrees to complete feature selection. The expression ability of the complex nonlinear relationship of the model is improved, and the precision of the ore pulp grade analysis is improved. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 is the implementation flowchart of the present application.

[0026] Figure 2 is the residual learning lightweight convolutional neural network structure diagram.

[0027] Figure 3 is the Inception module network structure diagram.

[0028] Figure 4 is the original spectrum and residual module output spectrum time domain comparison diagram.

[0029] Figure 5 is the comparison diagram of the true value and the predicted value of the phosphorus ore pulp P205 training set and the prediction set. DETAILED DESCRIPTION

[0030] The technical solutions of the present application will be further described below in combination with the LIBS grade analysis example of the phosphorus ore pulp.

[0031] Embodiment: A residual learning lightweight convolutional neural network quantitative method for LIBS, the implementation flowchart is as shown in Figure 1 , and specifically includes the following steps:

[0032] (1) Spectral band selection and data averaging. A total of 100,000×6116 spectral matrices were obtained from 100 samples, with 1,000 spectral data points for each sample. The 1,000 spectral data points collected for each sample were then averaged, resulting in a 100×5958 spectral data matrix for all 100 samples. Spectral bands rich in P-type lines and containing Mg, Si, Fe, and C were selected as input data, and the final input spectral data was 100*900.

[0033] (2) Divide the sample into training and test sets. The total number of samples is 100. 21 samples are randomly selected as the test set, and the remaining 79 samples are used as the training set. When dividing the training and test samples, the concentration range of the test set is evenly distributed within the concentration range of the training set, and the samples with the highest and lowest concentrations are both in the training set.

[0034] (3) A residual learning lightweight convolutional neural network model is established and trained using the selected original spectral segments. The weights of each layer are randomly initialized. The input is the original spectrum, the output is the elemental concentration, and the label is the elemental concentration. The original spectral signal is processed gradually through the layers of the residual module in the network using a feedforward method. Similarly, the P2O5 concentration is predicted through the layers of the lightweight convolutional neural module using a feedforward method. The loss function of the output layer (mean squared error and L2 norm regularization to minimize the sum of squares loss) is calculated, and the weights are updated by gradient descent until the number of iterations is completed.

[0035] like Figure 2 The diagram shows the architecture of a lightweight convolutional neural network for residual learning. The residual module consists of two residual units, each with identical internal structure and parameter settings. Residual unit 1 consists of three convolutional layers. The first layer uses a 1×1 convolutional kernel with a stride of 1 and one output channel, which is batch normalized before activation by the ReLU function. The second layer uses a 1×3 convolutional kernel with a stride of 1 and one output channel, which is batch normalized before activation by the ReLU function. The third layer uses a 1×1 convolutional kernel with a stride of 1 and one output channel, which is batch normalized before activation by the ReLU function. The original spectral segment, after identity transformation, is subtracted from the output of residual unit 1 after three convolutional layers; this difference is used as the input to residual unit 2. The output of Residual Unit 2 after three convolutional layers is subtracted from its input, and this difference is used as the input to the lightweight convolutional neural module. Specific network parameters for the Residual Unit are shown in Table 1.

[0036] Table 1

[0037]

[0038] The Lightweight Convolutional Neural Module (L-CNN module) consists of a convolutional neural network nested with an Inception module. Specifically, it includes three sequentially connected convolutional neural layers, an Inception layer, a flattened layer, and a fully connected (FC) layer, with the activation functions and pooling operations of the convolutional layers (Conv_1, Conv_2, Conv_3) removed. For example... Figure 3 As shown, an InceptionModule layer consists of four branches, each with a 1×1 convolutional structure. The four branches are nonlinearized using a tanh activation function before merging. Fully connected (FC) layers are nonlinearized using a tanh activation function, and the output of the fully connected layer is the predicted concentration. The parameters of the lightweight convolutional neural module are shown in Table 2.

[0039] The specific network structure of the residual learning lightweight convolutional neural network model is as follows:

[0040] Table 2

[0041]

[0042]

[0043] The parameter optimization of the established residual learning lightweight convolutional neural network model includes:

[0044] a. Spectral preprocessing process, the input vector x is processed by parameter W r b r Obtain the output s of the residual module, where x is the original spectral segment information and s is the preprocessed spectral segment vector;

[0045] b. In the quantitative spectral analysis process, the input vector s is passed through parameter W. t b t Obtain the output of a lightweight convolutional neural module s represents the preprocessed spectral information. Element concentrations after quantitative analysis;

[0046] The formula for model feedforward propagation is as follows:

[0047]

[0048] In the formula: X is the original spectral information, W r b r These are the weights of the residual module process matrix, S = XR(X, W) r ,b r ) represents the preprocessed spectral information, W t b t where represents the weights of the process matrix of the lightweight convolutional neural module, and R and T are the activation functions of the residual module and the lightweight convolutional neural module, respectively. Predicted concentration of the element to be measured.

[0049] c. By setting the residual unit structure, the number of residual units, the convolutional layer structure, the activation function, the learning rate, the number of iterations, the L2 regularization coefficient, repeatedly iterating, optimizing the weight of the residual learning lightweight convolutional neural network model, and obtaining the optimized model;

[0050] d. The target function is calculated using the following formula, and the back propagation is performed under the condition of minimizing the target function, and the residual module outputs the preprocessed spectrum s, and the lightweight convolutional neural module outputs the concentration of the element to be measured

[0051]

[0052] Wherein, M is the number of training samples, y n is the label concentration value of the nth training sample, is the model predicted concentration value, λ is the penalty factor, ||w|| 2 is the network layer weight, and the above formula is to limit the weight w by adding an L2 regularization constraint term in the hidden layer, so that w is as close to 0 as possible.

[0053] (4) The trained model is used to predict the element concentration.

[0054] The method is realized by using the sklearn function package in python.

[0055] Results verification: Figure 4 The figure shows the time domain comparison of the original spectrum and the residual module output spectrum obtained by using the method, Figure 5 The figure shows the comparison of the true value and the predicted value of the P205 training set and the validation set of the phosphate ore slurry obtained by using the method.

[0056] Table 3

[0057]

[0058] Table 3 compares the results of the method of the application and the original spectrum sent to the lightweight convolutional neural network model and the spectrum after wavelet threshold denoising sent to the lightweight convolutional neural network model, and the method of the application achieves the lowest predicted root mean square error (RMSEP) and the highest determination coefficient R2 R 2 . It shows that the method of the application has obvious improvement effect on feature extraction of high-dimensional spectral data, reduction of self-absorption and matrix effect, and quantitative analysis.

[0059] The phosphate ore slurry is used in the example, which is only an optimal embodiment, and the neural network weight parameters can be adjusted according to different application objects during implementation.

[0060] The above detailed description is merely exemplary in nature and is not intended to limit the application as described herein. Any modification or equivalent arrangement within the spirit or scope of the application should be considered to fall within the scope of the application.

Claims

1. A quantitative method for residual learning lightweight convolutional neural networks for LIBS, characterized in that, Includes the following steps: Step 1: Collect raw spectral data of laser-induced breakdown spectrum of training samples; Step 2: Select spectral bands from the original spectrum; Step 3: Build and train a residual learning lightweight convolutional neural network model using the selected original spectral bands; the specific network structure of the residual learning lightweight convolutional neural network model is as follows: The ResidualModule layer includes: multiple residual modules connected in sequence, with the difference between the input and output of the current residual module serving as the input of the next level residual module; each residual module contains multiple residual units connected in sequence; it is used to process the selected original spectral segment of the input and output the extracted spectral segment information s; the residual unit includes a convolution structure, batch normalization (BN), and ReLU activation function; The lightweight convolutional neural module layer includes: several convolutional layers, an Inception Module layer, a Flatten layer, and a fully connected layer FC, connected sequentially. This layer is used for further feature extraction of the spectral information s output by the residual module layer, outputting the predicted elemental concentration y. The Inception Module layer consists of four branches, each including a 1×1 convolutional structure, and the four branches are non-linearized using a tanh activation function before merging. The fully connected layer FC is non-linearized using a tanh activation function. Step 4: Use the trained residual learning lightweight convolutional neural network model to perform quantitative analysis on the test samples and output the predicted element concentrations.

2. The quantitative method for LIBS-oriented residual learning lightweight convolutional neural networks according to claim 1, characterized in that, The original spectral data is obtained by averaging the spectra of multiple laser strikes on each sample.

3. The quantitative method for LIBS-oriented residual learning lightweight convolutional neural networks according to claim 1, characterized in that, Select the spectral segment containing the characteristic spectral line that is strongly correlated with the element to be measured as the original spectral segment.

4. The quantitative method for LIBS-oriented residual learning lightweight convolutional neural networks according to claim 1, characterized in that, The ReLU activation function is expressed as: f(x) = max(0,x).

5. The quantitative method for LIBS-oriented residual learning lightweight convolutional neural networks according to claim 1, characterized in that, Nonlinear activation and pooling operations are removed from the convolutional layers of the lightweight convolutional neural module layer.

6. The quantitative method for LIBS-oriented residual learning lightweight convolutional neural networks according to claim 1, characterized in that, The training residual learning lightweight convolutional neural network model includes: The objective function is calculated using the following formula. Backpropagation is performed with minimizing the objective function as the condition. The process is iterated continuously. When the training reaches the number of iterations, the residual module outputs the preprocessed spectrum s, and the lightweight convolutional neural module outputs the concentration y of the element to be tested. Where M is the number of training samples, y n The label concentration value of the nth training sample. The model predicts the concentration value, λ is the penalty factor, and ||w|| 2 For the network layer weights, the above formula represents the L2 regularization constraint added to the hidden layer to restrict the weights w, making w as close to 0 as possible.

7. The quantitative method for LIBS-oriented residual learning lightweight convolutional neural networks according to claim 1, characterized in that, The root mean square error (RMSE) and coefficient of determination (R²) between the elemental content predicted by the quantitative analysis model and the reference label are calculated. 2 The indicators are used as evaluation functions to further optimize network parameters and obtain the final quantitative analysis model.

Images