High-purity vein quartz raw material rapid intelligent discrimination method

Abstract

The invention relates to the crossing field of computers and geology, in particular to a rapid intelligent distinguishing method for a high-purity vein quartz raw material, and aims to solve the practical problems of high cost, low efficiency and the like of distinguishing the purification capability of the vein quartz raw material through a traditional purification experiment method. A multi-type microscopic image feature vector and a physical and chemical parameter feature vector of a vein quartz raw material are fused to form a comprehensive feature vector, the comprehensive feature vector is input into a final classifier and mapped into a single output value, and the output value represents the purification potential of a vein quartz sample. According to the method, multi-type microscopic images and physical and chemical parameter data can be comprehensively utilized, rapid and intelligent discrimination of the purification potential of the vein quartz raw material is realized through the deep learning model, the analysis efficiency is improved, the waste of funds and manpower is reduced, and the method is suitable for popularization and application. The method has important practical significance and economic value for screening and distinguishing high-purity vein quartz raw materials.

Description

Technical Field

[0001] The invention relates to the intersection of computer and geology, and in particular to a method for rapid intelligent identification of high-purity vein quartz raw materials. Background Art

[0002] High-purity quartz raw materials are natural quartz mineral resources and are also important raw materials in the fields of optical fiber communication, aerospace, and solar photovoltaic industries. With the increasing industrial demand, crystal as the main high-purity quartz raw material resource has gradually been exhausted. Using vein quartz and other materials to prepare high-purity quartz sand to replace crystal raw materials is a practical problem that needs to be solved urgently.

[0003] There are certain commonalities and differences between vein quartz raw materials from different origins, and not all vein quartz raw materials can be processed into high-purity quartz products. At present, whether vein quartz raw materials can be processed into high-purity vein quartz products is mainly determined through purification experiments in the laboratory, but this method is complex, time-consuming, and costly. Blind purification experiments will cause huge waste of funds and manpower; and most of them extract the characteristics of vein quartz raw materials through single modal data such as structured data of physical and chemical parameters. This method is not accurate enough for judging vein quartz raw materials; therefore, a method is needed to scientifically, quickly and accurately judge the high-purity quartz purification potential of vein quartz raw materials based on low-cost data information.

[0004] In recent years, artificial intelligence technology, especially deep learning, has made great progress in the industrial field. However, in the existing technology, there is no method to use deep learning models to fuse the multimodal data of vein quartz raw materials and to identify and screen their high-purity quartz purification potential. Summary of the invention

[0005] The purpose of the present invention is to provide a method for rapid and intelligent discrimination of high-purity vein quartz raw materials. By combining the unstructured data of multiple types of microscopic images of vein quartz samples with the structured data of physical and chemical parameters, the effective fusion of multimodal data is achieved. Compared with single-modal data, the method can more comprehensively extract the internal characteristics and external microstructural features of vein quartz, thereby improving the accuracy and reliability of discrimination. In addition, the present invention constructs a rapid and intelligent discrimination framework based on a deep learning model, which can effectively and quickly screen out vein quartz raw materials with high purification potential.

[0006] To achieve the above object, the present invention proposes the following technical solution: a method for rapid intelligent identification of high-purity vein quartz raw materials, comprising the following steps: Step S1: Perform a purification experiment on the vein quartz sample and make a label according to the purification result; assign a label of 1 to the sample whose purification result meets the high-purity vein quartz product standard of Sibelco (North America) Corporation, and assign a label of 0 to the sample whose purification result does not meet the standard; Step S2: constructing a vein quartz sample dataset, dividing the vein quartz sample dataset into a training set, a validation set, and a test set. Each vein quartz sample should include multiple types of microscopic images, such as inclusion images; and multiple physical and chemical parameters, such as the number of particles; Step S3: After conversion, the microscopic image data of the training set and the validation set are input into the deep learning model for model training and validation. After the conversion, the microscopic image data is flattened into an image feature vector in the fully connected layer after passing through the image feature processing module. The corresponding physicochemical parameters are input into the parameter feature processing module to obtain the physicochemical parameter feature vector, which is then spliced ​​with the image feature vector to form a comprehensive feature vector. Finally, intelligent discrimination of high-purity vein quartz raw materials is performed in the fully connected layer. Step S4: inputting the microscopic image data and physicochemical parameter data of the test set into a pre-trained deep learning model to perform model testing; for the test set, the deep learning model comprehensively analyzes the obtained microscopic image features and physicochemical features, and fuses the microscopic image features and physicochemical features through a fully connected layer; based on the fused features, obtaining a predicted value of the vein quartz sample through an output layer; judging whether the vein quartz sample has purification potential based on the obtained predicted value, and comparing it with the actual purification experimental results to judge the discrimination ability of the deep learning model; Step S5: Input the microscopic image and physicochemical parameters of the vein quartz sample to be identified into the trained deep learning model, so as to perform rapid and intelligent identification of the vein quartz sample.

[0007] Preferably, the deep learning model is a convolutional neural network (CNN) based on the AlexNet architecture, and the deep learning model can receive multiple types of microscope images and multiple physical and chemical parameters of vein quartz samples, and integrate multimodal data for analysis.

[0008] Preferably, the microscopic images and physicochemical parameters are further derived from vein quartz samples from different origins, different sampling depths and different microscopic observation conditions.

[0009] Preferably, the method for constructing the vein quartz sample data set in step S2 comprises the following steps: Step S21: traverse the root directory of the original data and load the sample file according to the category label directory; Step S22: for each vein quartz sample, read the microscopic image file and the physicochemical parameter file corresponding to the microscopic image file; Step S23: in the process of constructing the vein quartz sample data set, dynamically matching the microscopic image file with the physical and chemical parameter file of the same name, and only selecting the high-purity vein quartz samples that exist at the same time as valid data; Step S24: storing the paths of the microscopic image file and the physicochemical parameter file in the form of sample tuples when loading, each of the sample tuples containing a microscopic image path, a physicochemical parameter path and label information; Step S25: for each vein quartz sample, the microscopic image file is loaded and image preprocessing is performed, and at the same time, the physicochemical parameter file is read and converted into a floating point tensor; Step S26: When returning data, output the microscopic image tensor, physicochemical parameter tensor and the path of the microscopic image file of the vein quartz sample.

[0010] Preferably, the physicochemical parameters are subjected to characteristic normalization.

[0011] Preferably, the vein quartz sample is input into the deep learning model as a multimodal feature by loading the microscopic image tensor and the physicochemical parameter tensor in the form of a multidimensional tensor.

[0012] Preferably, the image preprocessing process in step S25 includes: Step S251: Divide the image into 4 blocks evenly; Step S252: rotating each segmented image by 90°, 180°, and 270° respectively, and saving the rotated image; Step S253: cropping a smaller area from the central portion of each segmented image; Step S254: performing horizontal and vertical flip operations on the image; Step S255: shift the image rightward and downward by 10 pixels respectively to generate a new image; Step S256: Standardize the enhanced image data, normalize the image pixel values ​​to the range of [0, 1], and scale the image so that the input size is fixed to 227×227 pixels.

[0013] Preferably, in step S3, the image feature processing module is composed of a convolutional neural network based on the AlexNet architecture, and the microscopic image data processing includes the following steps: Step S311: input is a three-channel microscopic image; Step S312: performing a convolution operation in the convolution layer; Step S313: After completing each convolution operation, a ReLU activation function is introduced; Step S314: introducing a maximum pooling layer at every certain convolution layer level; Step S315: Finally, the image features are normalized to a fixed size of 6×6 through an adaptive average pooling layer.

[0014] Preferably, in step S3, the parameter feature processing module is composed of a fully connected neural network, and the processing of the physicochemical parameters includes the following steps: Step S321: The input is a physical and chemical parameter data set, which contains multiple numerical features; Step S322: extracting high-level features of the parameters through a fully connected layer containing 64 neurons, and performing nonlinear mapping in combination with a ReLU activation function; Step S323: using the Dropout technique to randomly discard some neurons; Step S324: A fully connected layer containing 64 neurons is used to further extract the deep features of the parameters, and the ReLU activation function is applied again to improve the feature expression capability.

[0015] Preferably, in step S3, the image feature vector and the physicochemical parameter feature vector are fused by feature splicing, which specifically includes: Step S331: concatenate the image feature vector and the physicochemical parameter feature vector along the feature dimension to form a comprehensive feature vector containing comprehensive information; Step S332: Mapping the comprehensive feature vector through a final classifier, where the final classifier is a linear fully connected layer, mapping the comprehensive feature vector to a single output value; Step S333: The output value is used to determine the purification potential of the vein quartz sample.

[0016] Preferably, the model training method in step S3 comprises the following steps: Step S341: adopting a K-fold cross-validation method including multiple rounds of training, and re-dividing the training set and the validation set in each round; Step S342: Use the optimizer Adam to optimize the deep learning model parameters; Step S343: The mean square error (MSE) is selected as the loss function, and the loss is calculated and back-propagated according to the difference between the output value of the deep learning model and the true label to update the model weight; Step S344: After each round of training, save the weight file of the current round of deep learning model.

[0017] Preferably, in step S3, the method for verifying the model includes the following steps: Step S351: Use the validation set to perform a performance test on the deep learning model and calculate the mean square error (MSE) on the validation set; Step S352: After each round of training is completed, the MSE values ​​of the deep learning model on the training set and the validation set are recorded and visualized; Step S353: Dynamically adjust the hyperparameter settings of the deep learning model, including learning rate decay and training rounds, according to changes in the verification MSE.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention realizes the effective fusion of multimodal data by combining the microscopic image (unstructured data) and physicochemical parameters (structured data) of vein quartz samples. Compared with single-modal data, the present invention can more comprehensively extract the internal characteristics and external microstructural features of vein quartz, thereby improving the accuracy and reliability of discrimination.

[0019] 2. The present invention constructs a fast and intelligent discrimination framework through a deep learning model. The present invention realizes the training, verification and testing of the deep learning model. After the test is completed, the vein quartz sample to be identified only needs to be input into the deep learning model to identify the purification potential of the vein quartz sample. This method can simply, effectively and quickly screen out vein quartz raw materials with high purification potential, avoid waste of resources, improve raw material utilization efficiency, and save manpower and costs.

[0020] 3. In the image preprocessing stage, the present invention adopts a variety of data enhancement methods to significantly increase the diversity of training samples. This method effectively improves the robustness of the model, making it more generalizable to vein quartz samples from different origins or with different characteristics.

[0021] 4. The deep learning model of the present invention uses different convolution kernels to extract features in the image, uses the ReLU activation function to introduce nonlinear processing, and finally reduces the dimension of the feature map through the maximum pooling layer and the adaptive average pooling layer to extract more abstract features; the structured physicochemical parameters are processed by the fully connected layer and combined with the ReLU activation function to extract the key features of various parameters; the deep learning model has a wide range of applicability and helps to promote the development of high-purity quartz materials in the fields of optical fiber communication, photovoltaic industry, aerospace, etc., and has important industrial application value and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a flow chart of the steps of the vein quartz sample identification method of the present invention; Figure 2 Standard drawing of high-purity quartz products of Sibelco (North America) Co., Ltd. (reference: Wang Ling. Separation and purification of quartz mineral resources and material application [J]. Mineral Protection and Utilization, 2022 (5): 57); Figure 3 is a structural diagram of the deep learning model of the present invention; Figure 4 This is a system screenshot of the model training process of the vein quartz sample identification method of the present invention; Figure 5 It is a performance index diagram of the model training of the vein quartz sample identification method of the present invention; Figure 6 This is a diagram showing the model test results of the vein quartz sample identification method of the present invention; Figure 7 This is a comparison chart between the model test results and the purification experiment of the vein quartz raw material identification method of the present invention. DETAILED DESCRIPTION

[0023] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0024] The present invention provides a method for rapid and intelligent identification of high-purity vein quartz raw materials, which is based on unstructured data of multiple types of microscopic images of vein quartz samples and structured data of multiple physical and chemical parameters; microscopic images are used as unstructured modalities to capture the microstructural characteristics of vein quartz samples; physical and chemical parameters are used as structured modalities to extract the intrinsic properties of vein quartz raw materials; the present invention fully explores the synergistic characteristics of different modal data and constructs a rapid and intelligent identification framework based on a deep learning model.

[0025] like Figures 1 to 7 As shown, a method for rapid intelligent identification of high-purity vein quartz raw materials comprises the following steps: Step S1: Label construction. Perform purification experiments on each vein quartz sample, create labels based on the purification results, mark samples that meet the high-purity vein quartz product standards of Sibelco (North America) as "1", and mark samples that do not meet the standards as "0", and store them in different folders. For the high-purity vein quartz product standards of Sibelco (North America) Company, please refer to Figure 2 ; Step S2: data preparation, constructing a vein quartz sample dataset containing multiple types of microscopic images and multiple physical and chemical parameters, and dividing it into a training set, a validation set, and a test set; Data collection: Collect microscopic images of vein quartz samples such as inclusion images, and collect physical and chemical parameters of vein quartz samples such as particle number.

[0026] Dataset construction: The dataset includes training set, validation set and test set. Some vein quartz samples are randomly selected from the vein quartz samples with labels of "1" and "0" as the test set, and the remaining vein quartz samples are used to construct the training set and validation set. The dataset construction method is as follows: Step S21: traverse the root directory of the original data, and load the vein quartz sample file according to the category label directory, the vein quartz sample file includes a microscopic image file and a physical and chemical parameter file; Step S22: for each vein quartz sample, read the microscopic image file and the physicochemical parameter file corresponding to the microscopic image file; the microscopic image file format includes JPG, PNG and JPEG, etc., and the physicochemical parameter file is in CSV format, etc.; Step S23: in the process of constructing the vein quartz sample data set, dynamically matching the microscopic image file with the physical and chemical parameter file of the same name, and only selecting the high-purity vein quartz samples that exist at the same time as valid data; Step S24: storing the paths of the microscopic image file and the physicochemical parameter file in the form of sample tuples when loading, each of the sample tuples containing a microscopic image path, a physicochemical parameter path and label information; Step S25: for each vein quartz sample, the microscopic image file is loaded and image preprocessing is performed, and at the same time, the physicochemical parameter data in the physicochemical parameter file is read for feature standardization, that is, normalization and conversion into floating point tensors; the image preprocessing process includes: Step S251: Block operation: divide the image into 4 blocks evenly; Step S252: Rotation operation: rotate each segmented image by 90°, 180°, and 270° respectively, and save the rotated image; Step S253: Cropping: Cropping a smaller area from the central part of each segmented image, where the smaller area retains 90% of the size of the image; Step S254: Offset: perform horizontal and vertical flip operations on the image; Step S255: Flip: shift the image rightward and downward by 10 pixels each to generate a new image; Step S256: Data standardization: Standardize the enhanced image data, normalize the image pixel values ​​to the range of [0,1], and scale the image so that the input size is fixed to 227×227 pixels.

[0027] Step S26: When returning data, the microscopic image tensor, physicochemical parameter tensor and path of the microscopic image file of the vein quartz sample are output for the association and tracing of subsequent prediction results. The microscopic image tensor is used to capture the microstructural features, and the physicochemical parameter tensor is used to extract the physical and chemical properties of the sample.

[0028] The shape of the microscopic image tensor is (batch_size, 3, 227, 227), which means (batch size, three channels (RGB), image length, image width), representing a batch of three-channel (RGB) images; the shape of the physicochemical parameter tensor is (batch_size, num_chem_features), which means (batch size, number of parameters), representing a batch of physicochemical parameters associated with the image.

[0029] When reading and verifying the data set, traverse the root directory of the data set, identify the microscopic image file and the physicochemical parameter file according to the file type, check whether the files match, and ensure that the physicochemical parameter file has the same name as the microscopic image file and exists.

[0030] Step S3, the microscopic images of the training set and the validation set are converted and input into the deep learning model for model training and validation, the image feature vector is extracted by the image feature processing module, and the physical and chemical parameters are obtained by the parameter feature processing module. The physical and chemical parameter feature vectors are merged into a comprehensive feature vector, and the high-purity vein quartz raw material is intelligently identified through the fully connected layer.

[0031] The deep learning model is used for rapid and intelligent identification of high-purity vein quartz raw materials. It is a convolutional neural network (CNN) based on the AlexNet architecture. The deep learning model consists of convolutional layers, pooling layers, and fully connected layers. The convolutional layers of the deep learning model include five layers. The first layer includes 11×11 convolution kernels, ReLU activation functions, and 3×3 maximum pooling layers; the second layer consists of 5×5 convolution kernels, ReLU activation functions, and 3×3 maximum pooling layers; the third and fourth layers are both composed of 3×3 convolution kernels and ReLU activation functions; the fifth layer is composed of 3×3 convolution kernels, ReLU activation functions, and 3×3 maximum pooling layers; the fifth convolution layer is followed by an adaptive average pooling layer and two fully connected layers, and the two fully connected layers contain ReLU activation functions and Dropout layers. The deep learning model includes a multimodal feature input structure, an image feature processing module, a parameter feature processing module, and a feature fusion and classification module.

[0032] Multimodal feature input structure: Multimodal feature input includes microscopic image tensors and physical and chemical parameter tensors, which are fused after being processed by the image feature processing module and the parameter feature processing module respectively, and then loaded into the deep learning model in the form of multidimensional tensors.

[0033] Image feature processing module: It consists of a convolutional neural network based on the AlexNet architecture. The microscopic image tensor is extracted with high-level features through convolutional layers and pooling layers, including the following steps: Step S311: The input is a three-channel microscopic image, and the microscopic image features are extracted through a convolutional layer; Step S312: the first convolution layer uses a convolution kernel of size 11×11, a step size of 4, and a padding of 2. The second convolution layer uses a convolution kernel of size 5×5. The third, fourth, and fifth layers all use a convolution kernel of size 3×3. High-level microscopic image features are gradually extracted through convolution operations. Step S313: introducing a ReLU activation function after each convolution operation to enhance the nonlinear expression capability; Step S314: introducing a maximum pooling layer at regular intervals of convolutional layers to reduce the dimension of the feature map and extract more abstract features; Step S315: Finally, the feature map is normalized to a fixed size of 6×6 through an adaptive average pooling layer and flattened into a one-dimensional image feature vector of 4096 dimensions.

[0034] Parameter feature processing module: composed of a fully connected neural network, the processing of the physical and chemical parameters includes the following steps: Step S321: The input is a physical and chemical parameter data set, which contains multiple numerical features; Step S322: The physical and chemical parameter tensor is passed through two fully connected layers to extract features: a fully connected layer containing 64 neurons is used to extract high-level features of the parameters. The first fully connected layer outputs 64-dimensional features and is combined with a ReLU activation function for nonlinear mapping; Step S323: Use Dropout technology to randomly discard some neurons to prevent overfitting; Step S324: The 64-dimensional features of the first layer are then input again into the second fully connected layer containing 64 neurons. The weight matrix shape of the second layer is (64, 64), keeping the feature dimension unchanged. Then, a ReLU activation function is used to further extract higher-level nonlinear features to form a physical and chemical parameter feature vector.

[0035] Feature fusion and classification module: The image feature vector and the physical and chemical parameter feature vector are fused through feature splicing. The specific method is as follows: Step S331: concatenate the image feature vector and the physicochemical parameter feature vector into a comprehensive feature vector through torch.cat; Step S332: the comprehensive feature vector is input into the final classifier, and the final classifier is a linear fully connected layer, which maps the comprehensive feature vector into a single output value; Step S333: The output value represents the purification potential of the vein quartz sample; according to the label setting, with 0.5 as the threshold, if the model output result is greater than 0.5, it can be judged as having purification potential, otherwise it is judged as not having purification potential.

[0036] The training method of the deep learning model includes the following steps: Step S341: Use Xavier to initialize the weights of the convolutional layer and the fully connected layer of the deep model, set the bias value to 0, and use the mean square error (MSE) as the loss function; in each round of training, randomly select a part of each type of vein quartz sample (0 label, 1 label) as the validation set, and the remaining vein quartz samples are used as the training set. The vein quartz sample division can be dynamically adjusted, and a K-fold cross-validation method including multiple rounds of training is adopted. In each round of training, vein quartz samples are randomly re-selected to construct a new validation set, and the remaining vein quartz samples are used as new training sets to improve the robustness and generalization ability of the model; Step S342: Use the optimizer Adam to optimize the model parameters, and the initial learning rate is 0.001; Step S343, training process: In each round of training: the processed image feature tensor and physical and chemical parameter tensor are input into the deep learning model to calculate the predicted value; the MSE between the predicted value and the true label is calculated, and the model weight is updated by back propagation; the dynamic learning rate adjustment strategy is combined in the training process, and the learning rate is decayed to 10% of the current value every 10 rounds of training through StepLR; the model is trained for 50 rounds to ensure that the performance on the training set and the validation set is consistent. The final results are as follows Figure 4 After each round of training, the training loss and validation loss are recorded, and curves are drawn to monitor the convergence of the deep learning model; Step S344: After each round of training, save the model weight file of the current round to evaluate and compare the validation set performance of different rounds for subsequent testing and deployment.

[0037] The validation method for deep learning models includes the following steps: Step S351: Use the validation set to calculate the MSE value of the model on unseen data to evaluate the prediction and generalization capabilities of the model; Step S352: After each round of training is completed, the MSE value of each round of verification is recorded and visualized, and compared with the loss curve of the training set to determine whether there is overfitting or underfitting, so as to facilitate the convergence of the training process; Step S353: dynamically adjust the hyperparameter settings of the model according to the change of the verification MSE, including learning rate decay and training rounds; Figure 5 This is a model training performance index diagram of the vein quartz sample discrimination method of the present invention. The MSE values ​​recorded on the training set and the validation set indicate that the model has high accuracy and stability in discriminating the purification potential of vein quartz samples. The curve diagram shows that the training loss and the validation loss gradually converge without obvious overfitting.

[0038] Step S4, input the microscopic image data and physical and chemical parameter data of the test set into the trained deep learning model, and fuse the obtained microscopic image features and physical and chemical parameter features through the fully connected layer; obtain the predicted value of the vein quartz sample through the output layer; compare with the actual purification experimental results to judge the discrimination ability of the deep learning model.

[0039] The model is verified through the test set. The test set data, including microscopic images and physical and chemical parameters, are input into the trained model to obtain the intelligent discrimination results of the test samples. Figure 6 This is a model test result diagram of the vein quartz sample discrimination method of the present invention. The discrimination results are compared with the actual purification experimental results of the test set. Please refer to Figure 7 The results show that the model has high accuracy. When the discrimination ability meets the requirements, it can be used to judge the purification potential of other vein quartz samples.

[0040] Step S5: Input the microscopic image and physicochemical parameters of the vein quartz sample to be identified into the trained deep learning model to achieve rapid intelligent identification.

[0041] The present invention combines the microscopic image (unstructured data) and physicochemical parameters (structured data) of vein quartz samples to achieve effective fusion of multimodal data, and more comprehensively extract the internal characteristics and external microstructure characteristics of vein quartz, thereby improving the accuracy and reliability of discrimination. At the same time, a fast and intelligent discrimination framework is constructed based on the deep learning model. After model training, verification and testing, the vein quartz sample to be discriminated only needs to be input into the deep learning model to discriminate the purification potential of the vein quartz sample. This method can simply, effectively and quickly screen out vein quartz raw materials with high purification potential, avoid waste of resources, improve raw material utilization efficiency, and save manpower and costs.

[0042] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0043] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A rapid and intelligent identification method for high-purity vein quartz raw materials, characterized in that: The following steps are involved: Step S1: Perform a purification experiment on the vein quartz sample and make a label according to the purification result; mark the sample whose purification result meets the high-purity vein quartz product standard of Sibelco (North America) as 1, and mark the sample whose purification result does not meet the standard as 0; Step S2: constructing a vein quartz sample dataset containing multiple types of microscopic images and multiple physical and chemical parameters, and dividing it into a training set, a validation set, and a test set; Step S3: After conversion, the microscopic images of the training set and the validation set are input into the deep learning model for model training and validation. The image feature vector is extracted by the image feature processing module, and the physical and chemical parameters are processed by the parameter feature processing module to obtain the physical and chemical parameter feature vector. The two are fused into a comprehensive feature vector, and the intelligent discrimination of high-purity vein quartz raw materials is performed through the fully connected layer; Step S4: Input the microscopic image data and physicochemical parameter data of the test set into the trained deep learning model, and fuse the obtained microscopic image features and physicochemical parameter features through the fully connected layer; obtain the predicted value of the vein quartz sample through the output layer; compare it with the actual purification experiment results to determine the discrimination ability of the deep learning model; Step S5: Input the microscopic image and physicochemical parameters of the vein quartz sample to be identified into the trained deep learning model to achieve rapid intelligent identification.

2. The method for rapid intelligent identification of high-purity vein quartz raw materials according to claim 1, characterized in that: The deep learning model is a convolutional neural network (CNN) based on the AlexNet architecture, and the deep learning model can receive multiple types of microscope images and multiple physical and chemical parameters of vein quartz samples, and integrate multimodal data for analysis.

3. The method for rapid intelligent identification of high-purity vein quartz raw materials according to claim 1, characterized in that: The microscopic images and physicochemical parameters are further derived from vein quartz samples from different origins, different sampling depths and different microscopic observation conditions.

4. The method for rapid intelligent identification of high-purity vein quartz raw materials according to claim 1, characterized in that: The method for constructing the vein quartz sample data set in step S2 comprises the following steps: Step S21: traverse the root directory of the original data and load the sample file according to the category label directory; Step S22: for each vein quartz sample, read the microscopic image file and the physicochemical parameter file corresponding to the microscopic image file; Step S23: in the process of constructing the vein quartz sample data set, dynamically matching the microscopic image file with the physical and chemical parameter file of the same name, and only selecting the high-purity vein quartz samples that exist at the same time as valid data; Step S24: storing the paths of the microscopic image file and the physicochemical parameter file in the form of sample tuples when loading, each of the sample tuples containing a microscopic image path, a physicochemical parameter path and label information; Step S25: for each vein quartz sample, the microscopic image file is loaded and image preprocessing is performed, and at the same time, the physicochemical parameter file is read and converted into a floating point tensor; Step S26: When returning data, output the microscopic image tensor, physicochemical parameter tensor and the path of the microscopic image file of the vein quartz sample.

5. The method for rapid intelligent identification of high-purity vein quartz raw materials according to claim 1, characterized in that: The physicochemical parameters are characteristically normalized.

6. A method for rapid intelligent identification of high-purity vein quartz raw materials according to claim 4, characterized in that: The vein quartz sample is loaded with a microscopic image tensor and a physicochemical parameter tensor in the form of a multidimensional tensor and input into a deep learning model as a multimodal feature.

7. The method for rapid intelligent identification of high-purity vein quartz raw materials according to claim 4, characterized in that: The image preprocessing process in step S25 includes: Step S251: Divide the image into 4 blocks evenly; Step S252: rotating each segmented image by 90°, 180°, and 270° respectively, and saving the rotated image; Step S253: cropping a smaller area from the central part of each segmented image; Step S254: performing horizontal and vertical flip operations on the image; Step S255: shift the image rightward and downward by 10 pixels respectively to generate a new image; Step S256: Standardize the enhanced image data, normalize the image pixel values ​​to the range of [0, 1], and scale the image so that the input size is fixed to 227×227 pixels.

8. The method for rapid intelligent identification of high-purity vein quartz raw materials according to claim 1, characterized in that: In step S3, the image feature processing module is composed of a convolutional neural network based on the AlexNet architecture, and the microscopic image data processing includes the following steps: Step S311: input is a three-channel microscopic image; Step S312: performing a convolution operation in the convolution layer; Step S313: After completing each convolution operation, a ReLU activation function is introduced; Step S314: introducing a maximum pooling layer at every certain convolution layer level; Step S315: Finally, the image features are normalized to a fixed size of 6×6 through an adaptive average pooling layer.

9. The method for rapid intelligent identification of high-purity vein quartz raw materials according to claim 1, characterized in that: In step S3, the parameter feature processing module is composed of a fully connected neural network, and the processing of the physicochemical parameters includes the following steps: Step S321: The input is a physical and chemical parameter data set, which contains multiple numerical features; Step S322: extracting high-level features of the parameters through a fully connected layer containing 64 neurons, and performing nonlinear mapping in combination with a ReLU activation function; Step S323: using the Dropout technique to randomly discard some neurons; Step S324: A fully connected layer containing 64 neurons is used to further extract the deep features of the parameters, and the ReLU activation function is applied again to improve the feature expression capability.

10. The method for rapid intelligent identification of high-purity vein quartz raw materials according to claim 1, characterized in that: In the step S3, the image feature vector and the physicochemical parameter feature vector are fused by feature splicing, which specifically includes: Step S331: concatenate the image feature vector and the physicochemical parameter feature vector along the feature dimension to form a comprehensive feature vector containing comprehensive information; Step S332: Mapping the comprehensive feature vector through a final classifier, where the final classifier is a linear fully connected layer, mapping the comprehensive feature vector to a single output value; Step S333: The output value is used to determine the purification potential of the vein quartz sample.

11. The method for rapid intelligent identification of high-purity vein quartz raw materials according to claim 1, characterized in that: The method for model training in step S3 comprises the following steps: Step S341: adopting a K-fold cross-validation method including multiple rounds of training, and re-dividing the training set and the validation set in each round; Step S342: Use the optimizer Adam to optimize the deep learning model parameters; Step S343: The mean square error (MSE) is selected as the loss function, and the loss is calculated and back-propagated according to the difference between the output value of the deep learning model and the true label to update the model weight; Step S344: After each round of training, save the weight file of the current round of deep learning model.

12. The method for rapid intelligent identification of high-purity vein quartz raw materials according to claim 1, characterized in that: In step S3, the method for verifying the model includes the following steps: Step S351: Use the validation set to perform a performance test on the deep learning model and calculate the mean square error (MSE) on the validation set; Step S352: After each round of training is completed, the MSE values ​​of the deep learning model on the training set and the validation set are recorded and visualized; Step S353: Dynamically adjust the hyperparameter settings of the deep learning model, including learning rate decay and training rounds, according to changes in the verification MSE.

Images