Ore grade rapid detection method and system, medium and product

Abstract

The invention discloses an ore grade rapid detection method and system, a medium and a product. The ore grade rapid detection method comprises the following steps: extracting ore features from an input XRT image of ore by using a neural network model, and predicting an implicit grade value; and converting the ore grade prediction value into an ore grade prediction value in an ordered binary form through a pre-trained ordered regression neural network with ordered consistency, and performing inverse uniform distribution mapping after real numeralization to obtain the ore grade prediction value. The method aims at directly learning the regression model for predicting the continuous value of the ore grade so as to effectively avoid the problem of a dichotomy separation model, the problems of'skewed distribution 'and'unsteady state' in regression prediction of the ore grade are specifically optimized and solved, the separation performance of the ore separation model is improved, and meanwhile, the adjustability and adaptability of the ore separation model are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of ore sorting, and in particular to a method, system, medium and product for rapid detection of ore grade. Background Art

[0002] The core of an ore sorter is to accurately identify the ore to be sorted. Common data-driven ore identification algorithms, such as classic statistical models and machine learning models, are based on the idea of ​​binary classification and directly output binary classification results of the ore (waste ore or concentrate ore). However, the binary classification model only roughly classifies the ore and does not fully learn the data differences between continuous ores of different grades. It is easy to introduce errors, resulting in poor classification performance and thus reduced sorting performance. Specifically, the binary classification model requires the ore to be classified in advance, which is generally done by artificially setting a grade threshold. For example, ore with a copper grade of more than 1% is considered to be concentrate ore, and vice versa. However, the setting of this grade threshold is full of experience and uncertainty. (2) The true value of the ore grade is generally the test result, which is a precise continuous value. However, in the binary classification model, the continuous grade true value is converted into a discrete binary true value. This conversion is irreversible, resulting in information loss, so that the test grade value is not fully utilized. (3) The adjustability and adaptability of the binary classification model are poor. When we need to adjust the sorting results, such as discarding more waste rock than before, retaining lower-grade ore, etc., the binary classification model often needs to be retrained and deployed to meet such changes in demand. Summary of the invention

[0003] Technical problem to be solved by the present invention: In view of the above-mentioned problems of the prior art, a method, system, medium and product for rapid detection of ore grade are provided. The present invention aims to directly learn the regression model for predicting the continuous value of ore grade to effectively avoid the problems of the binary classification sorting model, and to optimize and solve the "skewed distribution" and "non-steady state" problems in the regression prediction of ore grade in a targeted manner, thereby improving the sorting performance of the ore sorting model and enhancing its adjustability and adaptability.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is: A method for rapid detection of ore grade comprises the following steps: S1, extracting ore features from the XRT image of the input ore using a neural network model; S2, predicting the ore characteristics through a pre-trained ordered regression neural network with sorting consistency to generate an ordered binary form of ore grade prediction value; S3, convert the predicted ore grade value in the ordered binary form into real numbers, and obtain the predicted ore grade value by inverse uniform distribution mapping of the predicted ore grade value in the ordered binary form after real number conversion: , , In the above formula, is the predicted value of the ore grade of the i-th ore, is the predicted ore grade of the i-th ore in the real-numbered ordered binary form, is the quantile conversion function The inverse function of is the quantile conversion function The true value of the ore grade for the i-th ore The output, is the total number of ore samples; is an indicator function, which takes the value 1 when the condition is met, otherwise it is 0; is the true value of the ore grade of the jth piece of ore.

[0005] Optionally, step S1 includes: standardizing the XRT image of the ore into a single-channel real tensor of a specified size after padding and normalization operations, and converting the single-channel real tensor into a three-channel real tensor of a specified size by repeated expansion and inputting it into a trained feature extraction network to obtain a multi-dimensional real feature vector as the extracted ore feature.

[0006] Optionally, the ordered regression neural network for sorting consistency in step S2 includes a multi-perceptron model MLP and a sorting consistency module, and step S2 includes: S2.1, inputting the real feature vector representing the ore characteristics into the multi-perceptron model MLP to convert the ore characteristics into the corresponding implicit grade value, the function expression of the multi-perceptron model MLP is: , In the above formula, is the implicit grade value of the i-th ore sample, is a linear rectification function, and is the weight parameter, and is the bias parameter, is the ore characteristics of the i-th ore sample; S2.2, using the sort consistency module to convert the implicit grade value into an ore grade prediction value in ordered binary form: , In the above formula, is the predicted ore grade in ordered binary form, is the sigmoid activation function, is the bias parameter of the sorting consistency module.

[0007] Optionally, before step S2, the method further includes training an ordered regression neural network with sorting consistency, wherein the ordered regression neural network with sorting consistency includes using a quantile conversion function to convert the true value of the ore grade of the ore sample into a true value of the ore grade of the ore sample. The uniform distribution mapping, the ordered binary true value vector obtained by ordered binarization, and the binary classification loss calculated by the ore grade prediction value of the ordered binary form of the ore sample are used as the loss function to optimize the network parameters of the ordered regression neural network with sorting consistency, wherein the ore grade true value of the ore sample is transformed by the quantile conversion function The function expression of uniform distribution mapping is: , In the above formula, is the true value of the ore grade of the i-th ore sample Mapping value after uniform distribution mapping; The function expression of ordered binarization is: , In the above formula, is the first ordered binary truth vector of the i-th ore sample values, For the k Threshold value, The dimension of the ore grade prediction value in ordered binary form.

[0008] Optionally, the calculation function expression of the binary classification loss is: , In the above formula, represents the binary classification loss function, is the total number of ore samples, is the dimension of the ore grade prediction value in ordered binary form, is the first ordered binary truth vector of the i-th ore sample values, is the kth value of the predicted ore grade in ordered binary form for the i-th ore sample.

[0009] Optionally, in step S3, the function expression for converting the ordered binary ore grade prediction value into a real number is: , In the above formula, is the predicted ore grade of the i-th ore in the real-numbered ordered binary form, is the dimension of the ore grade prediction value in ordered binary form, is the kth value of the predicted ore grade in ordered binary form for the i-th ore, If Returns 1 if true, otherwise returns 0.

[0010] Optionally, step S3 further includes classifying the ore according to the following formula: , In the above formula, is the classification result of the i-th ore, is the predicted value of the ore grade of the i-th ore, is the grade sorting threshold.

[0011] In addition, the present invention also provides a system for rapid detection of ore grade, comprising a microprocessor and a memory connected to each other, wherein the microprocessor is programmed or configured to execute the method for rapid detection of ore grade.

[0012] In addition, the present invention also provides a computer-readable storage medium, in which a computer program or instruction is stored. The computer program or instruction is programmed or configured to execute the method for rapid ore grade detection through a processor.

[0013] In addition, the present invention also provides a computer program product, including a computer program or instructions, which are programmed or configured to execute the method for rapid ore grade detection through a processor.

[0014] Compared with the prior art, the present invention mainly has the following advantages: the method of the present invention comprises extracting ore features from the input XRT image of the ore using a neural network model and predicting implicit grade values; converting the ore grade prediction value into an ordered binary form through a pre-trained ordered regression neural network with sorting consistency, and obtaining the ore grade prediction value by performing inverse uniform distribution mapping after real number conversion. The present invention effectively avoids the problems of the binary classification sorting model by directly learning the regression model for predicting the continuous value of ore grade, and solves the "skewed distribution" and "non-steady state" problems in the ore grade regression prediction through targeted optimization, thereby improving the sorting performance of the ore sorting model and enhancing its adjustability and adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Schematic diagram of the basic flow of the method of the embodiment of the present invention.

[0016] Figure 2 Schematic diagram of the principle of the method of the embodiment of the present invention. DETAILED DESCRIPTION

[0017] The embodiments of the present invention will be described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention. In the description of the present invention, if there is a description of the first, the second, etc., it is only for the purpose of distinguishing the technical features, and cannot be understood as indicating or implying the relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features. If it involves the description of the orientation, such as the orientation or position relationship indicated by the upper and lower indications, the above-mentioned orientation description is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. Unless otherwise clearly defined, the terms such as setting, installing, connecting, etc. in the description of the present invention should be understood in a broad sense, and the technicians in the relevant technical field can reasonably determine the specific meaning of the above terms in the present invention in combination with the specific content of the technical solution.

[0018] like Figure 1 and Figure 2 As shown, the ore grade rapid detection method of this embodiment includes the following steps: S1, extracting ore features from the XRT image of the input ore using a neural network model; S2, predicting the ore characteristics through a pre-trained ordered regression neural network with sorting consistency to generate an ordered binary form of ore grade prediction value; S3, convert the predicted ore grade value in the ordered binary form into real numbers, and obtain the predicted ore grade value by inverse uniform distribution mapping of the predicted ore grade value in the ordered binary form after real number conversion: , , In the above formula, is the predicted value of the ore grade of the i-th ore, is the predicted ore grade of the i-th ore in the real-numbered ordered binary form, is the quantile conversion function The inverse function of is the quantile conversion function The true value of the ore grade for the i-th ore The output, is the total number of ore samples; is an indicator function, which takes the value 1 when the condition is met, otherwise it is 0; is the true value of the ore grade of the jth ore. Through the quantile conversion function , such as the Uniform Quantile Transformation, which maps the true grade value to make the distribution of the true grade value more uniform.

[0019] In order to predict the grade of the ore, the ore sorting model needs to obtain the key feature information of the ore. The innovation of the present invention does not lie in the feature extraction method. Therefore, the feature extraction here can be a model that outputs ore features in various technical solutions. In this embodiment, ore feature information is extracted by machine vision technology, such as processing the XRT image of the ore with a convolutional neural network (CNN). Specifically, step S1 of this embodiment includes: the XRT image of the ore is standardized into a single-channel real tensor of a specified size after filling and normalization operations, and the single-channel real tensor is converted into a three-channel real tensor of a specified size by repeated expansion and input into the trained feature extraction network to obtain a multi-dimensional real feature vector as the extracted ore feature. Specifically, in this embodiment, the XRT image of the ore is standardized into a single-channel real tensor of 224×224×1 after padding and normalization operations, and the single-channel real tensor is converted into a three-channel real tensor of 224×224×3 by repeated expansion and input into the trained feature extraction network to obtain a real feature vector with a dimension of 512 as the extracted ore feature. The feature extraction network is a common network model in the convolutional neural network CNN, and the required network type can be selected as needed. For example, as an optional implementation, the feature extraction network in this embodiment specifically adopts the backbone network of the ResNet-18 neural network model (with the prediction head removed). The ResNet-18 model has been pre-trained on the ImageNet dataset and can provide excellent image feature extraction capabilities.

[0020] The ordered regression neural network for sorting consistency in step S2 of this embodiment includes a multi-perceptron model MLP and a sorting consistency module. Step S2 includes: S2.1, inputting the real feature vector representing the ore characteristics into the multi-perceptron model MLP to convert the ore characteristics into the corresponding implicit grade value, the function expression of the multi-perceptron model MLP is: , In the above formula, is the implicit grade value of the i-th ore sample, is a linear rectification function, and is the weight parameter, and is the bias parameter, is the ore characteristics of the i-th ore sample; S2.2, using the sort consistency module to convert the implicit grade value into an ore grade prediction value in ordered binary form: , In the above formula, is the predicted ore grade in ordered binary form, is the sigmoid activation function, is the bias parameter of the sorting consistency module.

[0021] In this embodiment, the ore characteristics will be represented The real feature vector with a dimension of 512 is input into the multi-perceptron model MLP to convert the ore features into the corresponding implicit grade values , the multi-perceptron model MLP includes two linear layers and a linear rectification function ReLU, the first linear layer contains K=1024 nodes, the second linear layer has 1 node, the first linear layer is used to convert the real feature vector with a dimension of 512 representing the ore characteristics into a 1024-dimensional vector, and the 1024-dimensional vector is then converted by a linear rectification function ReLU and input into the second linear layer again to be converted into a scalar as an implicit grade value. Due to the characteristics of the neural network model (multi-perceptron model MLP), the information contained in the implicit grade value expressed in the form of this feature vector is implicit, and it implicitly encodes the key visual features of the input ore XRT image. When training the prediction model, the grade information of the ore can be obtained by scientific measurement methods such as assays as a label for the implicit grade value. In addition, the true grade of each ore, that is, the true value of the grade, is obtained through precise inspection technology. This ore grade detection technology is superior to the detection technology used in ore sorting machines, but it is often carried out in an offline and cumbersome manner, and cannot be quickly detected online, so it is used as a true value guide for the sorting model in the ore sorting machine. For example, the ore is first crushed and ground into powder particles of about 200 meshes, and then fully mixed and extracted through a reduction process. About 5g of sample is then obtained by inductively coupled plasma mass spectrometry (ICP-MS).

[0022] Before step S2 of this embodiment, the step further includes training an ordered regression neural network with sorting consistency. The ordered regression neural network with sorting consistency includes using the ore grade true value of the ore sample through a quantile conversion function during training. The uniform distribution mapping, the ordered binary true value vector obtained by ordered binarization, and the binary classification loss calculated by the ore grade prediction value of the ordered binary form of the ore sample are used as the loss function to optimize the network parameters of the ordered regression neural network with sorting consistency, wherein the ore grade true value of the ore sample is transformed by the quantile conversion function The function expression of uniform distribution mapping is: , In the above formula, is the true value of the ore grade of the i-th ore sample Mapping value after uniform distribution mapping; The function expression of ordered binarization is: , In the above formula, is the first ordered binary truth vector of the i-th ore sample values, For the k Threshold value, is the dimension of the ore grade prediction value in ordered binary form. For the ore stockpile to be processed by the ore sorter, its grade distribution will inevitably be a skewed distribution. Since one of the innovative points proposed in this embodiment is to build a regression model for directly predicting the ore grade, and this regression model is more sensitive to the specific size of the output target value, if the original grade value is directly used as the supervisory signal for model training, it is easy to cause the learning process of the model to be affected by the skewed distribution of the ore grade, resulting in the sorting performance being unable to meet the actual needs. In one embodiment, the original ore grade distribution analysis of a mine mining site shows that most of the ores have a low grade, and only a few ores have a high grade, as shown in the figure. The expected grade sorting threshold of the ore is about 1%. When we directly use the original grade value to train the model, we find that because the grade value of the high-grade ore is particularly large, the trained model tends to predict low-grade ores, such as 0.01%, as a higher value (5%). This is because although the difference between 5% and 0.01% is large, it is still much smaller than the difference between 5% and 50%. Although this prediction strategy can ensure that the model has a lower training loss, it is very disadvantageous for the actual sorting requirements, that is, the grade sorting threshold is 1%, because almost all ores are classified as concentrates. Therefore, this embodiment proposes to use a distribution conversion function to map the true value distribution of ore grade to a uniformly distributed space, which can effectively alleviate the skewed distribution problem of the true value of ore grade. The true value of the ore grade of the ore sample is converted into a uniformly distributed space by the quantile conversion function. The true value of the grade obtained after uniform distribution mapping can be expressed as: , In the above formula, is the true value of the ore grade of the i-th ore sample The mapping value after uniform distribution mapping, is the quantile transformation function. In one embodiment, the quantile transformation function, also called the empirical cumulative distribution function (ECDF), is used to process the true value of the ore grade. After the transformation, the ore grade distribution is effectively converted to a more uniform distribution, making the true value of the grade more continuous, so as to better train the regression model, then: , In the above formula, is the quantile conversion function The true value of the ore grade for the i-th ore The output, is the total number of ore samples; is an indicator function, which takes the value 1 when the condition is met, otherwise it is 0; is the true value of the ore grade of the jth piece of ore.

[0023] In this embodiment, a neural network model is constructed which takes the ore characteristics as input and the converted grade truth value as output guidance signal, and the neural network model is trained using binary classification loss. Furthermore, in order to avoid the problem of "rank inconsistency" existing in conventional ordered regression models, a rank consistent logits module is used in the ordered regression neural network. Specifically, the ordered regression neural network with rank consistency in step S2 includes a multi-perceptron model MLP and a rank consistency module. The implicit grade value is input into a linear layer with K nodes of a feature, and the linear layer has only learnable bias parameters but no weight parameters to convert the implicit grade value in scalar form into a K-dimensional vector. The K-dimensional vector is activated by a Sigmoid function and converted into a K-dimensional vector with a value between 0 and 1 as an ordered binary form of ore grade prediction value. The ordered regression neural network with rank consistency includes the use of the ore grade truth value of the ore sample through a quantile conversion function during training. The uniform distribution mapping, the ordered binary truth vector obtained by ordered binarization (Binary Decomposition), and the binary classification loss calculated by the ore grade prediction value of the ordered binary form of the ore sample are used as the loss function to optimize the network parameters of the ordered regression neural network with sorting consistency. Through the ordered binarization processing (Binary Decomposition), the grade truth value is converted into a cumulative binary vector to avoid the "non-stationary problem" in regression prediction. Although the skewness problem of the ore grade truth value has been effectively alleviated after the previous processing, since the ore grade prediction is a regression task, it still has the "non-stationary" problem. This is because the ore grade truth value is a real value, and different grade truth values ​​are homogeneous. However, due to the natural complex composition mechanism of natural ores, the characteristic information of the ore may be presented in different key forms in different grade intervals, that is, the change of the characteristic information of the ore may be non-homogeneous with the change of grade. This contradiction leads to the fact that real numbers cannot be used as a good supervision signal form to characterize ore grade. In one embodiment, low-grade copper ore on a mine is mainly impregnated, so at low grade, the grade change of this copper ore has a high correlation with the overall light and dark brightness characteristics of the XRT image of the ore. The high-grade copper ore on the mine is mainly vein-like, so at high grade, the grade relationship of this copper ore has a high correlation with the number and brightness of veins in the XRT image of the ore. For the ore of this mine, the grade true values ​​of low-grade and high-grade ores are homogeneous real numbers and have a linear relationship, but their characteristics are presented in different ways, are non-homogeneous, and do not have a linear relationship. Therefore, using real numbers to encode the grade true value of this ore is not the optimal solution. Therefore, this embodiment proposes to use an ordered binarization method to further convert the grade true value from a real number into an ordered binary vector true value, and the function expression of the ordered binarization is: , In the above formula, is the first ordered binary truth vector of the i-th ore sample values, For the k Threshold value, is the dimension of the ore grade prediction value in ordered binary form. Since this vector belongs to ordered coding, it still retains the high-precision coding of the grade. At the same time, by independently learning different binary true values ​​in this vector, the non-homogeneity problem can be effectively alleviated. Since the grade true value has been converted to a uniform distribution, the thresholds here are can be easily set by a linear function, just control The numerical value can be: , In the above formula, The value determines the accuracy of the conversion to the true value of the taste. The larger it is, the smaller the information loss caused by the transformation will be, but at the same time the average number of samples within the threshold interval will also decrease, resulting in more ore samples being required to build a fitting model. The value cannot be greater than the total number of ore samples , whose specific value can be obtained by parameter search and cross-validation. In one embodiment, the total number of ores is , found during parameter search The sorting performance of cross validation is optimal. On this basis, in this embodiment, a multi-perceptron model MLP is constructed and trained to learn to predict the true value of ore grade (after conversion) from ore characteristics. In order to further solve the "sort inconsistency problem" in ordered binary coding, the neural network model of this embodiment also ensures that the sorting result output by the model is consistent through the sorting consistency module. The specific method is to first output an implicit grade value through a linear regression layer, and then obtain K biased grade values ​​by adding K independent learnable biases (Bias) to the implicit grade value, and finally calculate the binary classification error between the K biased grade values ​​and the ordered binary grade true value as a supervisory signal on the K biased grade values, ensuring that the ordered binary predicted grade value output by the model after training is completed must be sorted consistently. This is because in this case, the K biases learned by the model must be monotonically increasing or decreasing. In one embodiment, a multi-perceptron model MLP (Multilayerperceptron, MLP) is used to predict grade information from ore feature information. The 512-dimensional feature vector of the ore output by the ResNet-18 neural network model is directly input into the MLP model, and after being calculated by the linear layer of 1024 nodes in the first layer, it is converted into a 1024-dimensional vector, and then converted by a rectified linear unit (ReLU), and then input into a linear layer with only 1 node again, and converted into a scalar. This scalar is the specific manifestation of the implicit grade value described in the present invention in this example. Thereafter, the scalar is input into a linear layer with K nodes of a feature, which has only learnable bias parameters but no weight parameters (or is understood as the weight parameters are fixed to 1), and is converted into a K-dimensional vector. Finally, the K-dimensional vector is activated by a sigmoid function and converted into a K-dimensional vector with a value between 0 and 1. This K-dimensional vector is the grade prediction value of the ore (ordered binary form). Finally, the error between the grade prediction value and the grade true value is calculated as the binary classification loss, and the calculation function expression of the binary classification loss is: , In the above formula, represents the binary classification loss function, is the total number of ore samples, is the dimension of the ore grade prediction value in ordered binary form, is the first ordered binary truth vector of the i-th ore sample values, is the kth value of the predicted ore grade in ordered binary form for the i-th ore sample. The weights of the model are updated using the loss function until the model fits the data set. The training of the neural network model involves common general techniques, which will not be elaborated here. It is worth mentioning that in some embodiments, the weights of the ResNet-18 model used as a feature extraction model are fine-tuned and trained together with the MLP model. This training paradigm of transfer learning is also a common technique in the field of machine learning and will not be elaborated on.

[0024] Since the neural network model learns the converted grade value, the grade prediction value output by the model needs to be decoded before it can be understood and used. At this time, it is only necessary to use the inverse function of the quantile conversion function on the grade prediction value in the reverse order to obtain the ore grade prediction real value in the same distribution range as the actual ore grade. Specifically, the function expression for converting the ordered binary form of the ore grade prediction value into a real number in step S3 is: , In the above formula, is the predicted ore grade of the i-th ore in the real-numbered ordered binary form, is the dimension of the ore grade prediction value in ordered binary form, is the kth value of the predicted ore grade in ordered binary form for the i-th ore, If If true, it returns 1, otherwise it returns 0. Then according to: , You can get , is the conversion function described in step 2 The inverse function of The converted grade prediction value is the final grade prediction value (after decoding).

[0025] In one embodiment, the ore is fed into a belt-type XRT ore sorter. After passing through the XRT detection equipment, the XRT image of the ore on the belt is first obtained, and the XRT image of a single ore is obtained through the ore positioning algorithm. The XRT image is pre-processed and sent to the neural network model to obtain the feature vector, and the feature is further sent to the neural network model to obtain the grade prediction value (before decoding). Since Quantile Transformation is used as the grade true value conversion in this embodiment, the ore grade prediction value is first converted to a real value through the inverse operation of ordered binarization, and then further decoded to the final grade prediction value through the inverse function of the quantile transformation function. The specific formula is: , In the above formula, Indicates from Find the smallest one in a set , so that .

[0026] Unlike the traditional ore sorting model that outputs a rough binary classification prediction result (high-grade ore or low-grade ore), the ore sorting model proposed in this embodiment outputs an accurate grade prediction value. To confirm whether an ore is classified as high-grade or low-grade ore, it is only necessary to manually set a grade threshold to obtain the same binary classification prediction result as the traditional binary classification model. Specifically, after step S3 of this embodiment, the ore classification is also performed according to the following formula: , In the above formula, is the classification result of the i-th ore, is the predicted value of the ore grade of the i-th ore, is the grade sorting threshold. However, the difference from the binary classification model is that the grade threshold can be flexibly adjusted according to the needs of the site. For example, when the grade requirement of the ore on site becomes lower and it is required to retain ore with a lower grade than before, it is only necessary to lower the grade threshold, while the traditional binary classification ore sorting model requires complex debugging, or even retraining and deployment. In addition, the accurate grade value of the ore prediction provides more detailed grade information, which helps the mine site understand the grade distribution of the ore and is also beneficial to the digital management of the mine.

[0027] In one embodiment, by comparing the performance of the traditional binary classification model with the model of the method of this embodiment, the significant advantages of this embodiment are proved: keeping other conditions unchanged (such as using the same ore data set, the same backbone network ResNet-18, etc.), the traditional binary classification model uses a single-node MLP to output the probability of discarding, and uses binary cross-entropy training, and the final discard rate is 37.11% and the recovery rate is 95.03%. Using the model proposed in this embodiment (without converting the true value of the ore grade), the discard rate is increased to 42.10% while the recovery rate remains basically unchanged. When the complete regression model proposed in this embodiment (including grade true value conversion and ordered binary processing) is used, the discard rate is further increased to 47.3%. The above data clearly show the improvement of the ore sorting performance of this embodiment. Compared with the traditional binary classification model, this embodiment not only maintains a high level in recovery rate, but also significantly improves the discard rate, indicating its high efficiency and practicality in the ore sorting process. It can be seen that this embodiment has significant technical advantages in the field of ore sorting, which not only improves the sorting performance, but also enhances the flexibility and adaptability of the system.

[0028] In summary, this embodiment significantly improves the performance of ore sorting through innovative designs for ore grade prediction and sorting. Its innovations are mainly reflected in the following aspects. These innovative designs are closely combined with the specific needs of ore sorting and effectively improve the deficiencies in the prior art: Regression model design for direct prediction of ore grade: This embodiment proposes a regression model that can directly predict the continuous value of ore grade. Compared with the existing sorting method based on the binary classification idea, this model makes full use of the continuity information of ore grade and avoids the information loss caused by the discretization of continuous grade values ​​in the binary classification model, thereby significantly improving the accuracy of ore grade prediction. This direct regression model design can better adapt to the natural distribution of ore grade and meet the requirements of ore sorting for fine particle size prediction. Uniform distribution mapping of grade true value: In view of the skewed distribution problem of ore grade true value in actual operation, this embodiment proposes a solution for uniform distribution mapping of ore grade true value using quantile conversion function. This processing method effectively alleviates the impact of skewed distribution on model training, avoids the imbalanced learning effect of the model between high-grade and low-grade ores, improves the prediction accuracy of the model in different grade intervals, and thus improves the overall performance of ore sorting. Ordered binary processing and its inverse conversion: In order to overcome the "non-steady state" problem in ore grade prediction, this embodiment performs an ordered binary processing on the grade true value after uniform distribution mapping, and converts the continuous grade value into an ordered binary vector. This ordered binary encoding not only maintains the fine characterization of ore grade, but also effectively solves the problem of inconsistent ore characteristics in different grade intervals. By designing the corresponding inverse conversion function, the output of the model can be losslessly restored to a continuous grade prediction value, meeting the needs of ore sorting for high-precision prediction. Construction of an ordered regression neural network with ranking consistency: In view of the "ranking inconsistency" problem that may occur in the traditional ordered regression model, this embodiment introduces a ranking consistency module (Rank Consistent Logits) in the neural network to ensure the consistency of the ordered binary prediction results output by the model in the grade order. This design not only improves the prediction accuracy of the model, but also ensures the stability and interpretability of the output results in ore grade prediction, and is suitable for the continuous division of concentrate and waste ore in ore sorting. The adjustability and adaptability of the model are enhanced: Since the model proposed in this embodiment directly outputs continuous ore grade prediction values, users can flexibly adjust the grade sorting threshold according to actual on-site needs without having to retrain and deploy the model like traditional binary classification models. This feature greatly enhances the adaptability of the ore sorting model, and is especially suitable for on-site applications in different mines or with different grade requirements, which helps to improve the efficiency and flexibility of the sorting process. In summary, this embodiment proposes a series of innovative methods from grade prediction model design to grade truth processing, neural network structure optimization, etc., in response to the actual needs in the ore sorting process.Targeted improvements have been made in the specific scenario of ore sorting, so that this embodiment can effectively improve the prediction accuracy, stability and operational flexibility of the ore sorting model, thereby improving the sorting performance.

[0029] In addition, this embodiment also provides a system for rapid detection of ore grade, comprising a microprocessor and a memory connected to each other, wherein the microprocessor is programmed or configured to execute the method for rapid detection of ore grade.

[0030] In addition, this embodiment also provides a computer-readable storage medium, in which a computer program or instruction is stored. The computer program or instruction is programmed or configured to execute the method for rapid ore grade detection through a processor.

[0031] In addition, this embodiment also provides a computer program product, including a computer program or instructions, which are programmed or configured to execute the method for rapid ore grade detection through a processor.

[0032] Those skilled in the art should understand that the technical solutions provided by the embodiments of the present application may be in the form of methods, systems, or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the process Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which implements the functions specified in the process. Figure 1 A process or multiple processes and / or boxes Figure 1These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0033] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A method for rapid detection of ore grade, characterized in that: The steps include: S1, extracting ore features from the XRT image of the input ore using a neural network model; S2, predicting the ore characteristics through a pre-trained ordered regression neural network with sorting consistency to generate an ordered binary form of ore grade prediction value; S3, convert the predicted ore grade value in the ordered binary form into real numbers, and obtain the predicted ore grade value by inverse uniform distribution mapping of the predicted ore grade value in the ordered binary form after real number conversion: ， ， In the above formula, is the predicted value of the ore grade of the i-th ore, is the predicted ore grade of the i-th ore in the real-numbered ordered binary form, is the quantile conversion function The inverse function of is the quantile conversion function The true value of the ore grade for the i-th ore The output, is the total number of ore samples; is an indicator function, which takes the value 1 when the condition is met, otherwise it is 0; is the true value of the ore grade of the jth piece of ore.

2. The method for rapid detection of ore grade according to claim 1, characterized in that: Step S1 includes: the XRT image of the ore is normalized into a single-channel real tensor of a specified size after padding and normalization operations, and the single-channel real tensor is converted into a three-channel real tensor of a specified size by repeated expansion and input into a trained feature extraction network to obtain a multi-dimensional real feature vector as the extracted ore feature.

3. The method for rapid detection of ore grade according to claim 2, characterized in that: The ordered regression neural network for sorting consistency in step S2 includes a multi-perceptron model MLP and a sorting consistency module. Step S2 includes: S2.1, inputting the real feature vector representing the ore characteristics into the multi-perceptron model MLP to convert the ore characteristics into the corresponding implicit grade value, the function expression of the multi-perceptron model MLP is: ， In the above formula, is the implicit grade value of the i-th ore sample, is a linear rectification function, and is the weight parameter, and is the bias parameter, is the ore characteristics of the i-th ore sample; S2.2, using the sort consistency module to convert the implicit grade value into an ore grade prediction value in ordered binary form: ， In the above formula, is the kth value of the predicted ore grade in ordered binary form for the i-th ore sample, is the sigmoid activation function, is the bias parameter of the sorting consistency module.

4. The method for rapid detection of ore grade according to claim 1, characterized in that: Before step S2, the step further includes training a sorting consistency ordered regression neural network, wherein the sorting consistency ordered regression neural network includes using the ore grade true value of the ore sample through a quantile conversion function during training. The uniform distribution mapping, the ordered binary true value vector obtained by ordered binarization, and the binary classification loss calculated by the ore grade prediction value of the ordered binary form of the ore sample are used as the loss function to optimize the network parameters of the ordered regression neural network with sorting consistency, wherein the ore grade true value of the ore sample is transformed by the quantile conversion function The function expression of uniform distribution mapping is: ， In the above formula, is the true value of the ore grade of the i-th ore sample Mapping value after uniform distribution mapping; The function expression of ordered binarization is: ， In the above formula, is the first ordered binary truth vector of the i-th ore sample values, For the k Threshold value, The dimension of the ore grade prediction value in ordered binary form.

5. The method for rapid detection of ore grade according to claim 4, characterized in that: The calculation function expression of the two-class loss is: ， In the above formula, represents the binary classification loss function, is the total number of ore samples, is the dimension of the ore grade prediction value in ordered binary form, is the first ordered binary truth vector of the i-th ore sample values, is the kth value of the predicted ore grade in ordered binary form for the i-th ore sample.

6. The method for rapid detection of ore grade according to claim 1, characterized in that: The function expression for converting the predicted ore grade value in ordered binary form into real number in step S3 is: ， In the above formula, is the predicted ore grade of the i-th ore in the real-numbered ordered binary form, is the dimension of the ore grade prediction value in ordered binary form, is the kth value of the predicted ore grade in ordered binary form for the i-th ore, If Returns 1 if true, otherwise returns 0.

7. The method for rapid detection of ore grade according to claim 1, characterized in that: Step S3 also includes ore classification according to the following formula: ， In the above formula, is the classification result of the i-th ore, is the predicted value of the ore grade of the i-th ore, is the grade sorting threshold.

8. A system for rapid detection of ore grade, comprising a microprocessor and a memory connected to each other, characterized in that: The microprocessor is programmed or configured to execute the method for rapid ore grade detection according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program or instruction stored therein, characterized in that: The computer program or instruction is programmed or configured to execute the method for rapid ore grade detection according to any one of claims 1 to 7 through a processor.

10. A computer program product comprising a computer program or instructions, characterized in that The computer program or instruction is programmed or configured to execute the method for rapid ore grade detection according to any one of claims 1 to 7 through a processor.

Images