Rapid detection method and system of aluminum content in bauxite based on near infrared spectroscopy

Abstract

The present invention discloses a method and system for rapid detection of aluminum content in bauxite based on near infrared spectroscopy, aiming to meet the needs of cross-fields such as earth science, resource exploration, and artificial intelligence. 2 O 3 The detection method is expensive, the instruments are expensive, and the efficiency is low, which limits its application in scientific research, industry, and management. 2 O 3 A sample database was established based on standard values; a deep learning model based on DenseNet and Roformer was constructed and trained to make it suitable for capturing the spectrum of bauxite and its Al 2 O 3 The nonlinear relationship between the content of bauxite Al 2 O 3 Rapid detection and analysis.

Description

Technical Field

[0001] The present invention relates to the fields of bauxite quality parameter analysis, near infrared spectrum analysis and artificial intelligence, and more specifically, to a method and system for rapid detection of aluminum content in bauxite based on near infrared spectrum. Background Art

[0002] Bauxite, as the main raw material of the aluminum industry, has important economic and industrial value worldwide. This ore is a sedimentary ore containing a high proportion of aluminum oxide (Al2O3), usually composed of aluminum hydroxide minerals. These minerals have gradually formed through chemical sedimentation over the long history of the earth. Bauxite is distributed all over the world, and China, as a powerhouse in the aluminum industry, occupies a pivotal position in the vigorous development of the national economy and is an important pillar for promoting the development of the national economy. Rapid detection of the Al2O3 content in bauxite can effectively improve the utilization efficiency of resources, is the key to promoting the sustainable and high-quality development of China's aluminum industry, and is also a manifestation of new quality productivity. However, traditional Al2O3 chemical analysis methods have many limitations, such as time-consuming and labor-intensive, possible sample damage, and extremely high professional and technical requirements for analysts.

[0003] With the rapid development of computer technology, a series of innovative instrumental analysis methods have emerged, which can realize the rapid detection of material components. For example, X-ray fluorescence and laser induced breakdown spectroscopy have the advantages of rapid and multi-element analysis compared with traditional chemical analysis. However, these methods have strict requirements on sample pretreatment and high instrument costs, which to a certain extent limit their widespread application. In this context, near-infrared spectroscopy technology, with its significant advantages such as relatively low cost and no need for complex sample preparation, provides a new solution for the detection of bauxite quality parameters. The application of this technology is expected to bring revolutionary changes to the effective utilization of bauxite resources and the sustainable development of the aluminum industry.

[0004] Chinese patent application publication number CN112179930A records a method for determining the Al2O3 content in high-sulfur bauxite by X-ray fluorescence spectrometry. The patent prepares a standard sample glass sheet by stirring a sulfur-containing mixed flux, a bauxite standard sample, etc. evenly and pre-oxidizing them. The fluorescence intensity of the glass sheet is measured using an X-ray fluorescence spectrometer, and a working curve is established with the fluorescence intensity as the ordinate and the corresponding Al2O3 mass concentration as the abscissa, thereby calculating the Al2O3 content in the glass sheet to be tested. Although the patent uses X-ray fluorescence spectrometry to measure the Al2O3 content in bauxite, it has advantages in accuracy, but requires complex sample preparation and high instrument costs. Summary of the invention

[0005] Technical issues to be solved

[0006] Aiming at the problems existing in the prior art that the detection process of Al2O3 content in bauxite requires complicated sample preparation and high instrument cost, the present invention provides a method for rapid detection of aluminum content in bauxite based on near-infrared spectroscopy, which realizes rapid detection of Al2O3 content by collecting near-infrared spectrum of bauxite and inputting it into a deep learning network based on DenseNet and Roformer.

[0007] Technical Solution

[0008] The purpose of the present invention is achieved through the following technical solutions.

[0009] The first aspect of the present invention provides a method for rapid detection of aluminum content in bauxite based on near infrared spectroscopy, comprising the following steps:

[0010] S1: Collect near-infrared spectral data of bauxite samples and construct a bauxite dataset;

[0011] S2: Determination of Al2O3 content of bauxite samples using X-ray fluorescence spectrometry as a label;

[0012] S3: Use the SNV method to preprocess the bauxite data set to obtain the bauxite sample set to be tested;

[0013] Furthermore, the calculation formula of the SNV data preprocessing method is:

[0014]

[0015] in, represents a single NIR spectrum from the bauxite dataset; represents the average absorbance of the spectrum, calculated according to formula (2); Represents the spectrum a The absorbance value corresponding to each wave number point; Represents the total number of wave number points of the spectrum;

[0016] S4: Using five-fold cross validation, the bauxite sample set to be tested and the corresponding Al2O3 content labels are divided into a training set, a validation set and a test set;

[0017] Furthermore, in each fold, the dataset is divided into three parts: 20% of the data is assigned as a test set for the final evaluation of the model; of the remaining 80% of the data, 10% is further divided as a validation set for adjusting and selecting the hyperparameters of the model; and the remaining data is used as a training set for the actual training process of the model;

[0018] S5: Input the training set into the deep learning model based on DenseNet and Roformer for training, and input the validation set into the model for model validation to obtain the optimal model;

[0019] Furthermore, the DenseNet model includes at least three densely connected modules, at least two Roformer modules, at least two transition modules and a regression head. Every two densely connected modules are connected in series using a Roformer module and a transition module. The layers of the densely connected modules use jump connections, that is, the input of each layer in each densely connected module comes from the output of all previous layers; the Roformer module includes a rotated position encoding block, a self-attention block, a layer normalization, and a forward propagation block; the transition module includes a convolutional layer, a batch normalization layer, a Mish activation function layer, and an average pooling layer; the backbone network of the regression head includes no less than one linear layer;

[0020] S6: Input the test set into the optimal model to obtain the Al2O3 content corresponding to the bauxite sample set to be tested.

[0021] Another aspect of the present invention further provides a system for rapid detection of aluminum content in bauxite based on near infrared spectroscopy, using a method for rapid detection of aluminum content in bauxite based on near infrared spectroscopy provided in the first aspect of the present invention. The software system that can be developed based on the above algorithm integrates near infrared spectroscopy technology and deep learning technology, and the user can obtain the content analysis results of Al2O3 in bauxite by directly connecting to a near infrared spectrometer to measure the crushed bauxite sample or submitting the near infrared spectrum of the bauxite.

[0022] Beneficial Effects

[0023] Compared with the prior art, the advantages of the present invention are:

[0024] (1) The present invention uses near-infrared spectroscopy combined with deep learning technology to detect the content of Al2O3 in bauxite, which can overcome the problems of cumbersome preparation work, expensive analytical instruments, and high requirements for the test environment of traditional identification methods while maintaining high accuracy;

[0025] (2) Aiming at the characteristics of bauxite near-infrared spectral data, the present invention designs a deep learning model based on DenseNet and Roformer for regression, which realizes the rapid detection of Al2O3 content in bauxite. The dense connection module uses skip connections to enhance the feature propagation and feature reuse of the model, helping it to learn richer bauxite feature representations. The combination of the rotation position encoding block and the self-attention block significantly enhances the model's ability to capture long-distance dependencies of the spectrum. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a principle block diagram of a method for rapid detection of aluminum content in bauxite based on near infrared spectroscopy;

[0027] Figure 2 Schematic diagram of near infrared spectrum of some bauxite;

[0028] Figure 3 Schematic diagram of a deep learning model based on DenseNet and Roformer;

[0029] Figure 4 This is a schematic diagram of the Roformer network structure;

[0030] Figure 5 Schematic diagram of the self-attention block structure;

[0031] Figure 6 Schematic diagram of the feedforward propagation layer structure;

[0032] Figure 7 Schematic diagram for realizing rotational position encoding;

[0033] Figure 8 This is the relationship between the actual value and predicted value of Al2O3 content in bauxite. DETAILED DESCRIPTION

[0034] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0035] Example 1

[0036] Figure 1 This is a schematic diagram of the principle of a rapid detection method for aluminum content in bauxite based on near-infrared spectroscopy.

[0037] This embodiment provides a method for rapid detection of aluminum content in bauxite based on near infrared spectroscopy, which can quickly analyze and predict Al2O3 in bauxite, and specifically includes the following steps:

[0038] S1: Collect near-infrared spectral data of bauxite samples and construct a bauxite dataset.

[0039] In this example, 420 bauxite samples with a particle size of 0.15 mm after crushing, grinding and sieving were collected, and the Al2O3 content was between 31.1% and 54.62%. The MicroNIR Pro handheld near-infrared spectrometer produced by VIAVI was used to collect spectra of the bauxite samples, with a wavelength range of 900-1700 nm, a resolution of 6.24 nm, and a wavelength point number of 125. Figure 2The figure shows the schematic diagram of the near infrared spectrum of some bauxite. The near infrared spectrum is a typical one-dimensional data, the horizontal axis represents the wave number range of the near infrared spectrum, and the vertical axis represents the absorbance value of the sample at the corresponding wave number point.

[0040] S2: Determination of Al2O3 content of bauxite samples using X-ray fluorescence spectrometry as a label.

[0041] S3: The SNV method is used to preprocess the bauxite data set to obtain the bauxite sample set to be tested.

[0042] Sample scattering media and random noise can interfere with spectral data, mask the true signal, and thus reduce the reliability of the analysis. Standard normal variable (SNV) is a spectral data preprocessing technique that standardizes the data by subtracting the mean of each wavelength and dividing by its standard deviation. Applying SNV to preprocess near-infrared spectral data can significantly improve data quality and reduce noise interference. The SNV calculation formula is as follows:

[0043]

[0044] in, represents a single NIR spectrum from the bauxite dataset; represents the average absorbance of the spectrum, calculated according to formula (2); Represents the spectrum a The absorbance value corresponding to each wave number point; Represents the total number of wavenumber points in the spectrum.

[0045] S4: Five-fold cross validation is used to divide the bauxite sample set to be tested and the corresponding Al2O3 content labels into training set, validation set and test set. In each fold, the data set is divided into three parts: 20% of the data is assigned as the test set for the final evaluation of the model; in the remaining 80% of the data, 10% is further divided as the validation set for adjusting and selecting the hyperparameters of the model; and the remaining data is used as the training set for the actual training process of the model.

[0046] S5: Input the training set into the deep learning model based on DenseNet and Roformer for training, and use the validation set to input into the model for model validation to obtain the optimal model.

[0047] Preferably, Figure 3As shown, the deep learning model based on DenseNet and Roformer used in this embodiment includes three dense connection modules, two Roformer blocks, two transition modules and one regression head. In other embodiments, other numbers of dense connection modules, Roformer blocks and transition modules may also be used. A Roformer module and a transition module need to be connected in series between every two dense connection modules.

[0048] The layers of the densely connected module use skip connections, that is, the input of each densely connected layer in each densely connected module comes from the output of all the densely connected layers before it. This design not only promotes the sharing and reuse of features, but also strengthens the flow of information in the network, which helps the network learn more complex feature representations. Assuming that a densely connected module consists of B densely connected layers, the entire module will contain approximately B*(B+1) / 2 skip connections. Each densely connected layer contains two convolutional sub-blocks, and each convolutional sub-block consists of a batch normalization layer, a Mish activation function layer, and a convolutional layer.

[0049] like Figure 4 The diagram of the Roformer network structure is shown in Figure 1. The Roformer module includes a rotated position encoding block, a self-attention block, a layer normalization block, and a forward propagation block. The rotated position encoding block and the self-attention block are the core of Roformer, which can capture the long-distance dependencies between points of different wavelengths, thereby better understanding the global information.

[0050] Furthermore, if Figure 4 As shown, the output matrix of the self-attention block is added to the elements of the corresponding positions in the input matrix of the Roformer model, and the matrix after layer normalization is added to the elements of the corresponding positions in the input matrix of the feedforward propagation layer, using the symbol Represents addition of matrix elements.

[0051] In detail, Figure 5 As shown, the rotation position encoding is performed by the feature matrix of the self-attention block Perform matrix multiplication to embed the position information, using the symbol Represents matrix element-wise multiplication.

[0052] like Figure 5 The following is a schematic diagram of the self-attention block. The feature matrix of the self-attention block From your own input matrix After linear layer transformation, L represents the sequence length, d represents the feature dimension, that is, the number of channels. Q and K The transpose multiplication makes the vector corresponding to each sequence point and Then, after the Softmax function, the original learning result is converted into a probability distribution, which can be used to V Assign attention weights to all sequence points. This attention weight can suppress features such as noise in network learning and give higher weights to features that can help predict Al2O3 content. Assume that the output of the self-attention block is , then the mathematical expression of the self-attention block is as follows:

[0053]

[0054] in, Represents the input matrix K The feature dimension size is used to scale the matrix multiplication result. After scaling, the calculation is not affected by the input dimension. impact. represents matrix multiplication. Correspondingly, The attention score vector obtained by formula (3) The mathematical expression is as follows:

[0055]

[0056] Due to the characteristics of matrix multiplication, the above self-attention calculation obviously cannot distinguish the order between sequence points, and can only consider the relationship between different sequence points. Therefore, it is necessary to add position encoding to the self-attention calculation to help the model capture this position information, thereby improving the performance and generalization ability of the model. Rotational Position Encoding (RoPE) has better extrapolation. It helps the model better understand the relative position relationship between sequence points by adding the absolute rotation angle to the feature vector corresponding to each sequence point. This relative position relationship is achieved by converting the absolute rotation angle into the relative rotation angle through matrix multiplication in the self-attention calculation. The position conversion needs to satisfy the following mathematical expression:

[0057]

[0058] in, and Respectively represent m and n The sequence point vector of the input matrix at positions, , , are different rotational attention encoding functions, , and Respectively represent the embedded m and nThe sequence point vector representation after the position information.

[0059] The length of the corresponding vector at the sequence point d When is 2, the solution of formula (6) can be calculated according to the geometric properties of vectors on the two-dimensional plane and the complex form of vectors, as shown below:

[0060]

[0061] in, is the real part of the complex number, i represents the imaginary unit, , represents the weight matrix, express The complex conjugate of . is a preset non-zero constant, which here represents the rotation angle. Further,

[0062]

[0063] in, is the representation in a two-dimensional coordinate system, is the rotation matrix. Similarly, The solution can also be obtained by similar derivation.

[0064] like Figure 7 The figure shows the schematic diagram of the rotation position encoding implementation. The length of the vector corresponding to the sequence point is extended to d In this case, you can d The dimensional space is divided into d / 2 subspaces, according to the linear properties in the inner product space, formula (13) is extended to d The mathematical expression in dimensional space is as follows:

[0065]

[0066] in, represents the rotation matrix, d is an even number, .

[0067] Substituting the rotation attention encoding function (14) into the matrix multiplication calculation of the attention formula (5), we can obtain formula (15):

[0068]

[0069] As mentioned above, the calculation of the entire rotation position encoding block and the self-attention block is completed. Through the joint action of these two blocks, the model can capture the long-range dependencies of the spectrum.

[0070] Layer normalization is achieved by normalizing the input of each layer in the neural network, which helps to reduce the internal covariate shift problem and improve the stability of the model.

[0071] like Figure 6 Schematic diagram of the forward propagation block. The forward propagation block extracts features from the feature dimension level of the feature matrix, making up for the shortcoming of not aggregating the feature dimension level in the self-attention calculation.

[0072] The transition module includes batch normalization layer, Mish activation function, convolution layer and average pooling layer. The average pooling layer is used to halve the channel dimension, reduce the amount of model calculation, and control model overfitting.

[0073] The regression head is the tail of the entire network architecture, using a fully connected layer to process the output of the last transition module to generate the final Al2O3 content prediction.

[0074] S6: Input the test set into the optimal model to obtain the Al2O3 content corresponding to the bauxite sample set to be tested.

[0075] Preferably, the deep learning network used in steps S5 to S6 in this embodiment is based on the PyTorch framework, and uses the Adam optimizer with a regularization weight of 0.001. The initial learning rate is set to 0.0001, and the ReduceLROnPlateau learning rate decay strategy is used to dynamically adjust the learning rate according to the loss value of the validation set. The size of each batch is 64, and the maximum number of iterations is set to 150.

[0076] Preferably, the model of step S5 to step S6 uses root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R 2 ) is used as the evaluation index of the embodiment of the present invention. 2 The closer it is to 1 and the smaller the values ​​of MAE and RMSE are, the better the performance of the model is.

[0077]

[0078] in, f Indicates the number of samples; and Indicates g The true and detected values ​​for each bauxite sample; Indicates g The mean of the true values ​​of the bauxite samples.

[0079] The model proposed in this invention has been tested, and the RMSE is 0.8657, the MAE is 0.5873, and the R 2 is 0.9239, reaching a certain performance index. Figure 8 The figure shows the relationship between the true value and predicted value of Al2O3 content in bauxite. As can be seen from the figure, the predicted value and the true value of Al2O3 content show a strong correlation, indicating that the detection accuracy of the model is high, which further proves the effectiveness and necessity of the proposed method.

[0080] Finally, based on the above algorithm, a rapid detection and analysis software for bauxite Al2O3 based on DenseNet can be developed. The software integrates near-infrared spectroscopy technology and deep learning technology. Users can obtain the content analysis results of bauxite Al2O3 by directly connecting to a near-infrared spectrometer to measure bauxite samples or submitting the near-infrared spectrum of bauxite.

[0081] Another aspect of the present invention also provides a system for rapid detection of aluminum content in bauxite based on near-infrared spectroscopy, and a method for rapid detection of aluminum content in bauxite based on near-infrared spectroscopy provided by the present invention.

[0082] The above schematically describes the invention and its implementation methods, which is not restrictive. Without departing from the spirit or basic features of the invention, the invention can be implemented in other specific forms. What is shown in the accompanying drawings is only one of the implementation methods of the invention. The actual structure is not limited thereto, and any figure mark in the claims should not limit the claims involved. Therefore, if a person of ordinary skill in the art is inspired by it, without departing from the purpose of the invention, a structural method and an embodiment similar to the technical solution are designed without creativity, which should all belong to the scope of protection of the present application. In addition, the word "including" does not exclude other elements or steps, and the word "one" before the element does not exclude the inclusion of "multiple" elements. The multiple elements stated in the product claim can also be implemented by one element through software or hardware. The words first, second, etc. are used to indicate names, and do not indicate any specific order.

Claims

1. A method for rapid detection of aluminum content in bauxite based on near infrared spectroscopy, characterized in that: The steps include: S1: Collect near-infrared spectral data of bauxite samples and construct a bauxite dataset; S2: Determination of Al2O3 content of bauxite samples using X-ray fluorescence spectrometry as a label; S3: Use the SNV method to preprocess the bauxite data set to obtain the bauxite sample set to be tested; S4: Five-fold cross validation was used to divide the bauxite sample set to be tested and the corresponding Al2O3 content labels into training set, validation set and test set. In each fold, 20% of the data was assigned as the test set for the final evaluation of the model. In the remaining 80% of the data, 10% was further divided as the validation set for adjusting and selecting the hyperparameters of the model. The remaining data was used as the training set for the actual training process of the model. S5: Input the training set into the deep learning model based on DenseNet and Roformer for training, and use the validation set to input into the model for model validation to obtain the optimal model. The Roformer module includes: a rotated position encoding block, a self-attention block, a layer normalization and a forward propagation block. The rotated position encoding block and the self-attention block capture the long-distance dependency between points of different wavelengths. S6: Input the test set into the optimal model to obtain the Al2O3 content corresponding to the bauxite sample set to be tested.

2. The method for rapid detection of aluminum content in bauxite based on near infrared spectroscopy according to claim 1 is characterized in that: The calculation formula of the SNV data preprocessing method described in S3 is: , in, represents a single NIR spectrum from the bauxite dataset; represents the average absorbance of the spectrum, calculated according to formula (2); Represents the spectrum a The absorbance value corresponding to each wave number point; Represents the total number of wavenumber points in the spectrum.

3. The method for rapid detection of aluminum content in bauxite based on near infrared spectroscopy according to claim 1 is characterized in that: The DenseNet model includes at least three densely connected modules, at least two Roformer modules, at least two transition modules and a regression head; every two densely connected modules are connected in series using a Roformer module and a transition module.

4. The method for rapid detection of aluminum content in bauxite based on near infrared spectroscopy according to claim 3 is characterized in that: The layers of the densely connected module use skip connections, that is, the input of each layer in each densely connected module comes from the output of all previous layers.

5. The method for rapid detection of aluminum content in bauxite based on near infrared spectroscopy according to claim 3 is characterized in that: The transition module includes a convolutional layer, a batch normalization layer, a Mish activation function layer, and an average pooling layer.

6. The method for rapid detection of aluminum content in bauxite based on near infrared spectroscopy according to claim 3 is characterized in that: The backbone network of the regression head consists of no less than one linear layer.

7. A rapid detection system for aluminum content in bauxite based on near infrared spectroscopy, characterized in that: A method for rapid detection of aluminum content in bauxite based on near infrared spectroscopy as described in any one of claims 1 to 6.

Images