CN113138178B

CN113138178B - Method for identifying imported iron ore brands

Info

Publication number: CN113138178B
Application number: CN202110405872.6A
Authority: CN
Inventors: 刘曙; 李晨; 赵文雅; 严承琳; 闵红
Original assignee: Shanghai Customs Industrial Products And Raw Material Testing Technology Center
Current assignee: Shanghai Customs Industrial Products And Raw Material Testing Technology Center
Priority date: 2021-04-15
Filing date: 2021-04-15
Publication date: 2023-07-07
Anticipated expiration: 2041-04-15
Also published as: CN113138178A

Abstract

The invention discloses a method for identifying imported iron ore brands. The method comprises the following steps: s1, taking at least 16 brands, and detecting at least 10 batches of iron ores of each brand by using a laser-induced breakdown spectroscopy to obtain modeling spectrum data; building a convolutional neural network model by using modeling spectrum data; the convolutional neural network model comprises an input layer, a convolutional layer, a pooling layer, a full-connection layer and an output layer which are sequentially carried out; the full-connection layer further comprises a flattening layer and a hiding layer which are sequentially carried out; s2, detecting the iron ore of the sample to be detected by using a laser-induced breakdown spectroscopy to obtain spectral data to be detected, and substituting the spectral data to be detected into a convolutional neural network model. The brand recognition method has high accuracy and simple model establishment operation.

Description

Method for identifying imported iron ore brands

Technical Field

The invention relates to a method for identifying imported iron ore brands.

Background

Iron ore is an important raw material for the iron and steel industry. When entering the import iron ore, the import iron ore enters the customs declaration and reports information such as names, places of origin and the like, and a traceable analysis for rapidly identifying the places of origin and quality conditions of the iron ore on site is established, so that phenomena such as doping, adulteration and the like can be effectively screened, and convenience of trade is ensured.

Patent document CN111239103a discloses a method for identifying iron ore production countries and brands, which adopts Laser Induced Breakdown Spectroscopy (LIBS) in combination with an Artificial Neural Network (ANN) model to realize classification of iron ore production countries and brands. The model requires three preprocessing methods (Savitzky-Golay polynomial filtering, multiple scatter correction and quadratic fitting) to remove spectral background, dispersion and noise, and uses Principal Component Analysis (PCA) to reduce the dimension of the LIBS data. These approaches effectively overcome the problem of low LIBS spectral quality in iron ore brand classification. However, complex spectral preprocessing and feature selection can extend analysis time and can lead to spectral distortion and reduced recognition accuracy.

Therefore, there is a need to establish a technology of imported iron ore brands, which has high recognition accuracy.

Disclosure of Invention

The invention provides a method for identifying imported iron ore brands, which aims to solve the defect of low identification accuracy in the method for identifying imported iron ore brands in the prior art. The brand recognition method has high accuracy and simple model establishment operation.

The invention solves the technical problems through the following technical proposal.

The invention provides a method for identifying imported iron ore brands, which comprises the following steps:

s1, taking at least 16 brands, detecting at least 10 batches of iron ores of each brand by using a laser-induced breakdown spectroscopy to obtain modeling spectrum data, and establishing a convolutional neural network model;

the convolutional neural network model comprises an input layer, a convolutional layer, a pooling layer, a full-connection layer and an output layer which are sequentially carried out; the full-connection layer further comprises a flattening layer and a hiding layer which are sequentially carried out;

the input layer is the modeling spectrum data;

the size of the convolution kernel of the convolution layer is 10-90; the number of convolution kernels of the convolution layer is 5-25; the activation function of the convolution layer is an S-shaped function, a hyperbolic tangent function or a modified linear unit function;

the activation function of the hidden layer is an S-shaped function, a hyperbolic tangent function or a modified linear unit function; the number of neurons of the hidden layer is 100-160;

s2, detecting the iron ore of the sample to be detected by using a laser-induced breakdown spectroscopy to obtain spectral data to be detected, substituting the spectral data to be detected into the convolutional neural network model, and determining the brand of the iron ore of the sample to be detected.

In the present invention, those skilled in the art know that the more the data amount of brands and lots used for modeling is, the better, and therefore the upper limit of the data amount for brands and lots is not particularly limited, and preferably the number of batches of iron ore per brand is 10 to 27.

In the present invention, preferably, the modeling spectrum data is modeling spectrum raw data, and the spectrum data to be measured is spectrum raw data to be measured; or the modeling spectrum data is data obtained by preprocessing the modeling spectrum raw data, and the spectrum data to be detected is data obtained by preprocessing the modeling spectrum raw data.

Preferably, the modeling spectrum raw data or the spectrum raw data to be measured are 12814 data points with the wavelength range of 187-972 nm.

Wherein, preferably, the preprocessing is feature line selection or wavelet transformation, or is a combination of Savitzky-Golay polynomial filter, multi-element scattering correction, quadratic fitting and principal component analysis.

Preferably, the characteristic line is selected to be 5 consecutive data points around the elemental signature emission line of the selected Fe, mg, mn, si, al, ca, na, K and Ti.

Preferably, a bin 1.3 wavelet is used in the wavelet transform.

In the present invention, the modeling spectrum data in the step S1 may be divided into three groups, which are a training set, a verification set, and a prediction set, respectively. Wherein the training set, the validation set, and the prediction set are at 70%:15%: the 15% ratio is randomly selected, and the case number of each selection is different.

In the present invention, the convolutional layer functions to extract spectral features from the input modeling data.

Wherein, preferably, the size of the convolution kernel is 10, 30, 50, 70 or 90.

Wherein, preferably, the number of convolution kernels of the convolution layer is 5, 10, 15, 20, 25 or 30.

Wherein, preferably, the activation function of the convolution layer is a hyperbolic tangent function.

Wherein, the number of layers of the convolution layer is preferably 1-2.

Wherein, preferably, the step size of the convolution layer is 4.

In the invention, the pooling layer has the functions of reducing the characteristic size and compressing network parameters so as to reduce the overfitting and improve the fault tolerance of the model. For the modeling data, the pooling layer may maintain translational invariance to the extracted features and increase robustness of the displacement.

Wherein, preferably, the type of the pooling layer is maximum pooling or average pooling, such as maximum pooling.

Wherein, preferably, the number of pooling windows of the pooling layer is 20.

Wherein, preferably, the layer number of the pooling layer is 1.

Wherein, preferably, the pooling window size of the pooling layer is 4.

Wherein, preferably, the pooling window step length of the pooling layer is 4.

In the invention, the flattening layer is used for converting the output high-dimensional data into one-dimensional data after the spectrum characteristics are extracted from the convolution area, and the one-dimensional data are used as input data of a subsequent classifier.

Wherein, preferably, a dropout method is adopted in the flattening layer.

Preferably, the dropout technology inactivation probability in the flattened layer is 0.5, so that the model can be effectively prevented from being excessively fitted during training data.

In the invention, the hidden layer functions as a hidden layer of the traditional artificial neural network.

Preferably, the number of neurons of the hidden layer is 100, 110, 120, 130, 140, 150 or 160.

Preferably, a dropout method is adopted in the hidden layer.

Preferably, the dropout technology inactivation probability in the hidden layer is 0.5, so that the model can be effectively prevented from being excessively fitted during training data.

Preferably, the activation function of the hidden layer is an S-type function.

In the present invention, preferably, the activation function of the convolution layer is a modified linear unit function, and the activation function of the hidden layer is an S-type function; or the activation function of the convolution layer is a hyperbolic tangent function, and the activation function of the hidden layer is an S-shaped function; alternatively, the activation functions of the convolution layer and the hidden layer are modified linear unit functions; or the activation function of the convolution layer is a hyperbolic tangent function, and the activation function of the hidden layer is a modified linear unit function; alternatively, the activation functions of the convolution layer and the hidden layer are hyperbolic tangent functions; alternatively, the activation function of the convolution layer is a modified linear unit function, and the activation function of the hidden layer is a hyperbolic tangent function.

In the present invention, preferably, the output layer is the brand.

Preferably, the activation function of the output layer is a Softmax function.

In the present invention, preferably, the learning rate of the convolutional neural network model is 0.001.

In the invention, preferably, the optimizer of the convolutional neural network model is an Adam optimization algorithm, which requires less memory and has high calculation efficiency.

In the invention, preferably, the process of establishing the convolutional neural network model performs data dimension reduction through a t-distribution random neighborhood embedding algorithm so as to visualize the optimizing effect of the convolutional neural network model.

On the basis of conforming to the common knowledge in the field, the above preferred conditions can be arbitrarily combined to obtain the preferred examples of the invention.

The reagents and materials used in the present invention are commercially available.

The invention has the positive progress effects that:

the brand recognition method has high accuracy and simple model establishment operation.

In a preferred embodiment of the invention, the spectrum data preprocessing process is not needed before the convolutional neural network model is established, and the processing time is short.

Drawings

Fig. 1 is a schematic flow chart of a convolutional neural network model used in example 1.

Fig. 2 is a confusion matrix obtained by CNN model calculation in example 1.

Fig. 3 is a confusion matrix calculated by kNN model in comparative example 1.

Fig. 4 is a confusion matrix calculated by LDA model in comparative example 2.

Fig. 5 is a confusion matrix calculated by RF modeling in comparative example 3.

Fig. 6 is a confusion matrix calculated by the SVM model in comparative example 4.

Fig. 7 is a confusion matrix calculated by the BPANN model in comparative example 5.

Fig. 8 is a t-SNE two-dimensional scattergram of output data of each layer of the CNN model in example 2. (a) an input layer; (b) a P2 layer; (c) an F4 layer; (d) an output layer.

FIG. 9 is a graph showing the effect of different convolution kernel sizes on brand recognition accuracy and time in example 3.

FIG. 10 is a graph showing the effect of varying numbers of convolution kernels on brand recognition accuracy and time in example 3.

FIG. 11 is a graph showing the effect of the combination of different activation functions and pooling types on brand recognition accuracy in example 4.

FIG. 12 is a graph showing the effect of varying numbers of neurons in the full-link layer on brand recognition accuracy in example 5.

FIG. 13 is a classification accuracy curve visualized by a 3000 iteration process in example 6.

Fig. 14 is a graph of the loss function visualized for the 3000 iterations of example 6.

Detailed Description

The invention is further illustrated by means of the following examples, which are not intended to limit the scope of the invention. The experimental methods, in which specific conditions are not noted in the following examples, were selected according to conventional methods and conditions, or according to the commercial specifications.

In the examples and comparative examples of the present invention, the sample collection and sample detection process of iron ore is as follows:

1. sample collection

In the unloading process of ports in China Shanghai, ningbo, xinjiang and other cities, 266 batches of 16 types of brands of iron ores from three producing places of Australia, south Africa and Brazil are collected. According to the preparation standards of relevant samples of GB/T10322.1-2014 iron ore sampling and sample preparation method, imported iron ore samples are ground into iron ore powder with the particle diameter of about 100 mu m. All iron ore fines were kept in a dry environment prior to collection of the spectra using Laser Induced Breakdown Spectroscopy (LIBS). Table 1 lists sample detailed information.

TABLE 1 iron ore sample information Table

2. Sample detection

Samples in each of the examples and comparative examples of the present invention were prepared using the model number chem real, TSI company, usa ^TM -3764 commercial LIBS device. YAG laser equipped with Q-switch with emission wavelength of 1064nm, and peak energy of 200MJ. The optical fiber of the collimating optical element in the system is about 20mm according to the surface of the sample, and the spectrum detection range of the CCD detector is 190-950 nm. The spectrum collection process is in the atmospheric environment, the sample is directly placed on an X-Y-Z manual fine adjustment sample stage, and the shooting is utilizedMicro-area adjustment and focusing are carried out on a small window of an image system, the optimal diameter of an ablation light spot is set to be 200 mu m, the delay time between a laser pulse and a CCD detector is set to be 2 mu s, the laser repetition frequency is 5Hz, and the emission energy of the optimized laser is 30MJ for the purpose of maximizing the signal-to-noise ratio (SNR).

Under the condition of no adhesive, firstly placing a polyethylene ring on the surface of a stainless steel gasket, placing about 8g of iron ore powder in the stainless steel gasket, keeping the pressure for 60 seconds under the pressure of a table press model machine (ZHY-401B model, beijing Zhonghe Innovative science and technology development Co., ltd.) for withdrawing the die, finally preparing a sample sheet with the thickness of about 5mm, if the surface of the sample is uniform and has no crack and falling off, sealing the sample sheet by using a plastic package bag, and then placing the sample sheet in a dryer for spectrum acquisition. During measurement, the surface of the tablet is blown off by an ear-washing ball, and in order to reduce spectrum fluctuation caused by sample heterogeneity, spectrum measurement positions are randomly selected on each sample surface, and each position is formed by a 5×5 measurement lattice. At each measurement point, the surface ash is removed by 5 laser pulses, then 5 laser pulses are accumulated and emitted, and 5 LIBS spectra are collected. Finally, all spectra collected on the measurement lattice (5X 5) are averaged, an original analytical spectrum was obtained. For each spectrum, a total of 12814 data points were collected over the entire wavelength range of 187-972 nm. For each batch of samples, different positions are selected, and the above operation is repeated 6 times to obtain 6 original analysis spectrums respectively. In this study, the analytical spectrum at each position in each batch of samples was taken as an independent spectral sample, resulting in 1596 (266 samples×6) analytical spectral samples.

3. Brand identification accuracy calculation

The accuracy represents whether the predicted category is consistent with the real category, and the higher the accuracy is, the better the classification performance of the model is. The formula is as follows:

wherein N is _corr For correctly predicted samples, N is all predicted samples.

4. Confusion matrix (fusion matrix)

The confusion matrix is a visual method for presenting model performance in the field of artificial intelligence, and can intuitively evaluate the prediction capability of the model in supervised learning. In general, each column of the confusion matrix represents a prediction result, and each row represents a true class. The diagonal lines represent correctly distinguished samples, with the other positions being the misclassified samples.

Examples 1.1 to 1.266

The convolutional neural network model (CNN) parameters and schemes employed in example 1 are shown in fig. 1 and table 2 below.

TABLE 2 parameters of CNN model employed in example 1

The K-fold cross-validation (K-CV) method can estimate the generalization ability of CNN and eliminate correlations between samples. The K value is typically chosen to be 5 or 10 to achieve balanced deviations and variances. In this example, a five-fold cross-validation (5-CV) method was used to divide the original spectrum sample into 5 equal parts (4 of which were 319 each and 320 each), 4 of which were used for model validation at a time and the remaining 1 for model prediction, for a total of five complete CNN calculations. The classification accuracy and loss function values for the calibration and prediction sets were recorded until five sets of data were predicted as the prediction set, and the result statistics were performed as shown in table 3.

TABLE 3 CNN model Brand identification results

5-CV	Verification set accuracy (%)	Prediction set accuracy (%)	Loss function
				1	99.92(1276/1277)	100.00(319/319)	0.0329
2	99.77(1274/1277)	100.00(319/319)	0.0351
				3	99.92(1276/1277)	100.00(319/319)	0.0246
4	99.84(1275/1277)	99.69(318/319)	0.0452
				5	99.84(1275/1276)	99.69(319/320)	0.0401
Average value of	99.86	99.88	0.0356
				Standard deviation of	0.06	0.17	0.00785

As shown in Table 3, the classification accuracy and the prediction accuracy of the verification set are respectively in the interval range of 99.77-99.92% and 99.69-100.00%, which proves that the CNN model has stronger brand classification capability. The average value of the loss function is 0.0356, which is kept at a low level, indicating that CNN has a high prediction accuracy and a clear classification boundary. The standard deviation (standard deviation, SD) of the neural network 5-CV results was mainly used to evaluate the repeatability of the CNN model. The SD of the validation set, the prediction set and the loss function are 0.06, 0.17 and 0.00785 respectively, which proves that the model has relatively stable learning effect.

Comparative examples 1.1 to 1.236 to comparative examples 5.1 to 5.236

In comparison with example 1, in order to evaluate the difference in classification performance of CNN from other conventional machine learning methods on branded iron ore, a model comparison experiment was performed. The raw spectral samples were used as input variables, and the average classification correctness of the corresponding calibration and prediction sets of 5-CV were recorded in Table 4, using K-nearest neighbor (kNN), linear Discriminant Analysis (LDA), random Forest (RF), support Vector Machine (SVM), and Back Propagation Artificial Neural Network (BPANN) models. Wherein the BPANN model parameters are shown in table 5.

TABLE 4 Brand identification accuracy of different models

	Model	Verification set accuracy (%)	Prediction set accuracy(％)
				Comparative examples 1.1 to 1.236	kNN	91.48	90.35
Comparative examples 2.1 to 2.236	LDA	97.12	96.86
				Comparative examples 3.1 to 3.236	RF	100.00	94.11
Comparative examples 4.1 to 4.236	SVM	98.18	97.36
				Comparative examples 5.1 to 5.236	BPANN	99.25	97.93
Examples 1.1 to 1.236	CNN	99.86	99.88

TABLE 5 parameters of the BPANN model used in comparative example 5

Under the condition that the input data are modeling spectrum raw data, the established CNN model has the highest calibration set and prediction set classification accuracy, and shows good brand classification capability. The detailed misclassification of each model is as shown in fig. 2-7. For predictions of 16 brands of iron ore, the CNN model misjudged only two Yang Difen iron ores as Ha Yangfen iron ores, while the other models showed a large number of misclassified samples. Analyzing the reasons for the occurrence of the large-area misclassification phenomenon, the LIBS spectrum dimension of the iron ore is high (12814 pixel points), and the composition of the emission lines is complex due to the strong influence of the iron matrix, so that the traditional mode identification method needs complex pretreatment and feature selection engineering to improve the classification performance. In the CNN model, the convolution layer and the pooling layer are equivalent to a feature extractor, so that a plurality of local features can be densely mined; and the full connection layer is an optimized and improved BPANN classifier. Therefore, the CNN model can adaptively learn the characteristics of LIBS spectrum while simplifying the experimental flow so as to obtain better classification results. Compared with machine learning, the designed CNN model has stronger anti-interference capability and feature extraction capability, is beneficial to overcoming the spectrum distortion defect caused by traditional complex pretreatment, simplifies the analysis process and realizes higher prediction precision.

Examples 2.2 to 2.236

The present embodiment uses a t-distributed random neighborhood embedding (t-distributed symmetric neighbor embedding, t-SNE) algorithm to reduce the size of the output data of each feature layer in the CNN, and is intuitively shown in a 2D scatter diagram (fig. 8) to explain the effectiveness of the CNN model of embodiment 1. the t-SNE algorithm mainly comprises two steps: 1. a probability distribution between high-dimensional objects is constructed such that similar objects have a higher probability of being selected and different objects have a lower probability of being selected. 2. The probability distributions of these objects are established in a low dimensional space so that the two probability distributions are as similar as possible (Kullback-Leibler divergence is used in this embodiment), thereby measuring the similarity between the observed objects more intuitively. The embodiment adopts t distribution in a low-dimensional space, can effectively reduce the congestion and the difficult optimization problem possibly occurring in an SNE algorithm, maintains a local structure and captures the integral characteristics of data.

As a result, as shown in fig. 8a, a large number of cross-over phenomena occur in the input layer, representing the point set of raw LIBS data for different brands of iron ore, due to close correlation with the main chemical composition of the iron ore. According to the reduction method of titanium trichloride of GB/T6730.5-2007 and the X-ray fluorescence spectrometry of GB/T6730.62-2005, the Fe element and other main chemical components (SiO ₂ 、Al ₂ O ₃ 、TiO ₂ CaO and MgO), the contents of other chemical components, except Fe element, are very similar in all brands of iron ores. Meanwhile, as can be seen from fig. 8a, only the point clusters of two brands of south african iron concentrate and australian super-extract are far away from the clusters, and can be clearly distinguished. From the quantitative data of the element (table 6, the quantitative data in the table are all mean.+ -. Standard deviation): the super powder is the iron ore with the lowest iron grade in all brands, and is easy to identify; in addition, tiO in the fine powder of south Africa iron ₂ The CaO and MgO contents are the highest, siO ₂ Is clearly different from other brands. However, achieving classification of all brands still requires the CNN to further extract useful information. Based on the t-SNE two-dimensional scatter plot and quantitative data, progressive and rational extraction of spectral features from the CNN model layer by layer can be explained from the pooling layer (P2), the hiding layer (F4) to the output layer.

TABLE 6 quantitative elemental data for iron ore brands

The t-SNE two-dimensional scatter diagram of the convolution region of the CNN model after feature extraction of the original data is shown in FIG. 8b, and the point sets of the same category show obvious effectsAggregation, the cross-over phenomenon is significantly reduced, indicating that the convolution region plays a critical role in mining more nonlinear features of the spectral data. The convolution layer and pooling layer may quickly capture similar features within a brand and differences features between brands. However, there are still clearly seen four confusing areas for further analysis based on the elemental quantitative data. First, yang Difen iron ore and Hamersley Yang Difen are both from the lydices of the Piira region of West Australia, and it can be seen from FIG. 8b that the other components are almost identical and the spectral characteristics are relatively close, despite the slight difference in CaO content between the two brands. This is consistent with the confusion matrix misclassification situation of fig. 2, where only the two types of iron ore are misjudged in the practical application of the CNN model of example 1. Then, the kunba standard powder and the kunba standard block have only a small difference in the contents of Fe and CaO, both being products of kunba iron ore company in south africa, the matrices thereof are very similar, and thus the convolution area cannot effectively extract the difference information. Other overlapped areas are respectively the Pelbara mixed powder and the Newman mixed powder, the Pelbara mixed block and the Newman mixed block, and the Newman mixed powder and the Newman mixed block in the two groups are in Fe and SiO ₂ And Al ₂ O ₃ The content is slightly higher than the former. However, tiO ₂ The CaO and MgO contents show irregular distribution, resulting in complex spectral characteristics of both groups. Therefore, the degree of recognition of these iron ores in the convolution region still needs to be further improved.

And then, the BP classifier of the full-connection layer continuously learns the extracted spectral features, and fits the feature data with the category to which the feature data belongs, so that the sample brand is accurately identified. As shown in fig. 8c, the same color dot sets are almost all clustered together, except for a small number of intersections of Yang Difen and Ha Yangfen, through the learning of the fully connected layers. Notably, there were a small number of discrete points in the calix powder, pi Erba pull mixing block and brazil mixing powder scatter plots, which may be related to errors in the sample itself.

When the CNN operation is finished, as shown in fig. 8d, the scattered points of the samples of the same class are very converged and gathered into a heap. In particular, the south Africa iron concentrate powder gathers to a point, proving depthThe characteristics of the spectrum of the sample can be easily learned by learning. In addition, the Fe content in the Australian iron concentrate powder is relatively high, siO ₂ And the highest MgO content, al ₂ O ₃ And TiO ₂ The content is the lowest, and the products are easy to distinguish. Thus, in fig. 8d, the scattered points of such brands also converge into several points.

After the spectral characteristic information extracted by each layer in the deep learning is output and visualized, each layer can be intuitively observed to present better and better recognition effect, and the similarity and the difference of various iron ores are gradually obvious. In addition, by combining quantitative data of main chemical components of iron ore, misjudgment and misjudgment reasons occurring in the learning of LIBS spectral characteristics of deep learning can be fully explained, and parameters of a CNN model can be optimized in the mode.

Examples 3.1 to 3.236

In this embodiment, only the influence of different data preprocessing modes on the accuracy of recognizing iron ore brands is compared. Among the following 4 modes, there are: the method comprises the steps of (1) modeling spectrum raw data, (2) modeling spectrum raw data through manually selected characteristic lines (5 continuous data points around characteristic emission lines of each element of iron ore are selected as shown in table 7, intensity of the characteristic line of the element is directly extracted for calculation), (3) denoising the modeling spectrum raw data through wavelet transformation (4 layers of LIBS spectrum are decomposed through biorthogonal 1.3 wavelet and signals are filtered through a fixed soft threshold method), and (4) principal component analysis processing is carried out on the modeling data raw data through a Savitzky-Golay polynomial filter, multi-component scattering correction and secondary fitting. The original spectrum samples were first randomly divided into three groups at a ratio of 70% (training set), 15% (validation set), 15% (prediction set) for experimental study of this example. The data after different data preprocessing is used as an input variable to be input into a CNN model, the average accuracy is obtained after training is performed for 5 times in parallel, and the prediction results are compared as shown in the following table 8.

TABLE 7 principal element characteristic emission lines of iron ore spectra

TABLE 8 influence of four data preprocessing modes on iron ore brand recognition accuracy

As can be seen from Table 8, the method (2) effectively reduces the dimension (12814 to 320) of the spectrum data, shortens the training time, and ensures that the average classification accuracy of the CNN model reaches more than 99%, which indicates that the selected characteristic spectral line has strong representativeness, but the manually selected characteristic line has the defect of being incapable of fully reflecting the spectrum characteristics. Through verification, the signal to noise ratio can be effectively improved by the mode (3), the prediction accuracy of CNN is 98.75%, but the method has the defect that the parameter adjustment process of a wavelet transformation method is complex and time-consuming. After the comprehensive pretreatment in the mode (4), the classification accuracy of the CNN model on the training set, the verification set and the prediction set is 98.57%, 97.74% and 98.08% respectively, and the prediction effect is worse than that after single pretreatment. The phenomenon proves that the excessively complex preprocessing method can cause spectrum distortion, the data volume after dimension reduction is too small, and the effect brought by letting CNN learn spectrum characteristics is reduced.

In general, LIBS data has many complicated redundant information, so an appropriate preprocessing method is required, and a process of reducing the data size and noise reduction is generally unavoidable. However, in this embodiment, compared with the spectrum data after different preprocessing, the CNN learning efficiency of the original analysis spectrum is highest, the accuracy of the prediction set reaches 99.58%, and the most preferred mode is to directly use the original spectrum as the input variable, so that the experimental process is effectively simplified, and the spectrum distortion caused by human factors is avoided.

Comparative examples 6.1 to 6.236

Comparative example 6 is to compare brand recognition accuracy obtained after model calculation after the same data preprocessing method, except that the Convolutional Neural Network (CNN) and the back propagation artificial neural network (BPANN, model parameters are shown in table 5) are compared with example 3.

In comparative example 6, (1) modeling spectrum raw data, (2) modeling spectrum raw data are subjected to manually selected characteristic lines (5 continuous data points around each element characteristic emission line of iron ore are selected as shown in table 7, and the intensity of the element characteristic lines is directly extracted for calculation), and (3) modeling spectrum raw data are subjected to wavelet transformation and denoising (after 4 layers of decomposition of LIBS spectrum by using biorthogonal 1.3 wavelet, signals are filtered by using a fixed soft threshold method). The original spectrum samples were first randomly divided into three groups at a ratio of 70% (training set), 15% (validation set), 15% (prediction set) for experimental study of this series of comparative examples. The data after different data preprocessing is used as an input variable to be input into a CNN model, the average accuracy is obtained after training is performed for 5 times in parallel, and the prediction results are compared as shown in the following table 9.

TABLE 9 influence of different data pretreatment modes in CNN and BPANN models on iron ore brand recognition accuracy

As can be seen from table 9, regardless of the LIBS spectral data obtained by using any data preprocessing method, the brand recognition capability of the CNN model is better than that of BPANN.

Mode (2) effectively reduces the dimension (12814 to 320) of the spectral data, shortening the training time; compared with BPANN with the accuracy rate of 95.91%, the average accuracy of the CNN model reaches more than 99%, which shows the representativeness of the selected characteristic spectral lines. Moreover, CNNs exhibit strong data mining capabilities and high analytical accuracy by virtue of small amounts of spectral data. Noise is inevitably generated during data acquisition, and it is verified that wavelet transformation can effectively process noise containing signals and improve signal-to-noise ratio SNR. The data processed in the mode (3) are respectively input into a BPANN model and a CNN model, and the average accuracy is respectively 96.43% and 99.55%; the data processed in the mode (1) are respectively input into a BPANN model and a CNN model, and the average accuracy is respectively 96.43% and 99.55%; this demonstrates that deep learning performs better at higher spectral quality.

Examples 4.1 to 4.236

In this embodiment, only the influence of the size and number of convolution kernels in the convolution layer on the accuracy of identifying iron ore brands is analyzed. The original spectrum samples were first randomly divided into three groups at a ratio of 70% (training set), 15% (validation set), 15% (prediction set) for experimental study of this example.

As shown in fig. 9 and table 10, the convolution kernel sizes were set to 10, 30, 50, 70, 90, respectively, and the prediction correctness and the run time of the different schemes were recorded. When the convolution kernel size is 50, the accuracy of the training set, the verification set and the prediction set reaches the maximum value (99.10%, 99.58%, 97.50%), and the average prediction accuracy of the three sets is 98.73%; the line graph under this condition shows that the calculation time of the model is also a minimum value (59 minutes).

TABLE 10 influence of different convolution kernel sizes on brand recognition accuracy and time

As shown in fig. 10 and table 11, when the size of the convolution kernel is 50, the influence of the number of convolution kernels on the brand recognition performance was further studied. The number of convolution kernels is set to 5, 10, 15, 20, 25, 30, respectively. When the volume product number is 5, the calculation time is shortest, mainly because the number of convolution kernels is small, so that the overall variable number of the model is small, the running time of the model is fast, and the optimal classification accuracy cannot be achieved under the condition. When the convolution kernel increases to 20, the accuracy of classification is higher and the run time is shorter.

TABLE 11 influence of different numbers of convolution kernels on brand recognition accuracy and time

Examples 5.1 to 5.236

In this embodiment, the influence of the activation function type and the pooling type on the accuracy of identifying the iron ore brands is analyzed. The original spectrum samples were first randomly divided into three groups at a ratio of 70% (training set), 15% (validation set), 15% (prediction set) for experimental study of this example.

The pooling layer type is selected as average pooling, and for the nonlinear mapping relation related to the two layers of the convolution layer (C1 layer) and the hidden layer (F4 layer), three functions S-shaped functions (Sigmoid), hyperbolic tangent functions (Tanh) and modified linear unit functions (Rectified Linear Unit, reLU) are selected for 6 combination optimization, wherein in FIG. 11, A corresponds to the combination of ReLU and Sigmoid, B corresponds to the combination of the Tanh and Sigmoid functions, C corresponds to the combination of the ReLU and the ReLU, D corresponds to the combination of the Tanh and the ReLU, E corresponds to the combination of the Tanh and the Tanh, and F corresponds to the combination of the ReLU and the Tanh. As shown in fig. 11, when the combination (B combination) of the Tanh and Sigmoid functions is selected, the prediction set classification accuracy and the average classification accuracy of the three sets (training set, verification set, and prediction set) reach the maximum value. In addition, under the combination of the activation functions of the combination B, the classification accuracy of the prediction set when the pooling type is average pooling is 98.75%, and the average accuracy is 99.52% (at the beginning of the arrow). When the pooling type at this time is changed to the maximum pooling, the classification accuracy is improved to 99.17% and 99.60% (at the arrow end points), respectively.

At the time of average pooling, only the brand recognition accuracy of combination B, E, F at the time of maximum pooling was further compared because the performance of combination B, E, F was better, and the results are shown in table 12 below. The maximum pooling algorithm can effectively reserve local key information of LIBS spectrum, reduce the number of parameters and improve the generalization capability of the model.

Table 12 effects of combinations of different activation functions and pooling types on brand recognition accuracy and time

Examples 6.1 to 6.236

In this example, the effect of the number of neurons in the fully connected layer on the accuracy of identifying iron ore brands was analyzed. The original spectrum samples were first randomly divided into three groups at a ratio of 70% (training set), 15% (validation set), 15% (prediction set) for experimental study of this example.

As shown in fig. 12 and table 13, at the fully connected layer, too few neurons are set up resulting in poor fitting of the classification boundaries, while too many increase the model parameters, with the risk of overfitting. When the number of neurons in the fully connected layer is in the range of 100, 110, 120, 130, 140, 150 and 160, the classification accuracy of the CNN model fluctuates greatly, which indicates that the number of neurons affects the stability of the model. When the number of the neurons is 120, the prediction set and the average accuracy reach more than 99.5% of the highest value.

TABLE 13 influence of different numbers of neurons in full-connected layers on Brand recognition accuracy and time

In order to further confirm the comprehensive effect of all parameters of the CNN model after comprehensive optimization, in the training process of the training set and the verification set, the embodiment derives the classification accuracy and the loss function value at the end of each calculation (iteration), and draws the classification accuracy and the loss function value in a two-dimensional graph for visualization of the iteration process. The classification accuracy of the CNN model in 3000 iterative processes is shown in FIG. 13, the model converges at a relatively high speed, and the classification accuracy reaches 95% in 500 iterations; when the iteration is performed 1500 times, the classification accuracy exceeds 99%, which indicates that the model has good learning ability and stability. In addition, from the view of the loss function value (fig. 14), the loss function value gradually goes to 0 after 500 iterations, meaning that the predicted value and the actual value of the model gradually approach each other. Although approximately 1.5-2 hours are required for 3000 iterations of the entire training process, only 300-400 μs are required to predict hundreds of spectral samples. Experiments show that the dropout technology can be applied to the CNN classifier to effectively prevent the overfitting phenomenon; in addition, the parameter adjustment is proper, so that the model can quickly realize higher classification accuracy.

Claims

1. A method for identifying a brand of imported iron ore, comprising the steps of:

the input layer is the modeling spectrum data;

the modeling spectrum data are modeling spectrum raw data;

the convolution kernel of the convolution layer has a size of 50; the number of convolution kernels of the convolution layer is 20;

the activation function of the convolution layer is a hyperbolic tangent function, and the activation function of the hidden layer is an S-shaped function;

alternatively, the activation functions of the convolution layer and the hidden layer are hyperbolic tangent functions;

or the activation function of the convolution layer is a modified linear unit function, and the activation function of the hidden layer is a hyperbolic tangent function;

the type of the pooling layer is maximum pooling;

the number of neurons of the hidden layer is 100-160;

s2, detecting the iron ore of the sample to be detected by using a laser-induced breakdown spectroscopy to obtain spectral data to be detected, substituting the spectral data to be detected into the convolutional neural network model, and determining the brand of the iron ore of the sample to be detected;

and the process of establishing the convolutional neural network model carries out data dimension reduction through a t-distribution random neighborhood embedding algorithm so as to visualize the optimizing effect of the convolutional neural network model.

2. The method of identifying brands of imported iron ore of claim 1, wherein the number of batches of iron ore of each brand is from 10 to 27;

and/or the spectrum data to be detected is spectrum original data to be detected.

3. The method of identifying an imported iron ore brand of claim 2, wherein the modeled raw spectral data or the raw spectral data to be measured is 12814 data points within a wavelength range of 187-972 nm.

4. The method of identifying an imported iron ore brand of claim 1, wherein the activation function of the convolution layer is a hyperbolic tangent function;

and/or the number of the convolution layers is 1-2;

and/or, the step size of the convolution layer is 4.

5. The method for identifying brands of imported iron ore of claim 1,

the number of pooling windows of the pooling layer is 20;

and/or the number of layers of the pooling layer is 1;

and/or, the pooling window size of the pooling layer is 4;

and/or, the pooling window step length of the pooling layer is 4.

6. The method for identifying an imported iron ore brand as set forth in claim 1, wherein a dropout method is employed in the flattened layer;

and/or the number of neurons of the hidden layer is 100, 110, 120, 130, 140, 150 or 160;

and/or, a dropout method is adopted in the hidden layer;

and/or the activation function of the hidden layer is an S-shaped function.

7. The method of identifying an imported iron ore brand as set forth in claim 6, wherein the dropout technology deactivation probability in the flattened layer is 0.5.

8. The method of identifying an imported iron ore brand as set forth in claim 6, wherein the dropout technology deactivation probability in the hidden layer is 0.5.

9. The method of identifying a brand of imported iron ore of claim 1, wherein the output layer is the brand;

and/or the activation function of the output layer is a Softmax function.

10. The method of identifying an imported iron ore brand of claim 1, wherein the convolutional neural network model has a learning rate of 0.001;

and/or, the optimizer of the convolutional neural network model is an Adam optimization algorithm.