Methods, apparatus, storage media and electronic equipment for predicting iron ore pixel-level grades

Abstract

This application discloses a method, apparatus, storage medium, and electronic device for predicting iron ore pixel-level grade, relating to the field of mineral resource prediction technology. The method includes: inputting hyperspectral data of the iron ore to be predicted into a convolutional autoencoder of a pixel-level grade prediction model; extracting low-dimensional features from the hyperspectral data using the convolutional autoencoder; inputting the low-dimensional features into a multilayer perceptron of the pixel-level grade prediction model; and mapping the low-dimensional features to a predicted iron ore grade value for each pixel in the iron ore to be predicted using the multilayer perceptron. This application can improve the accuracy of iron ore pixel-level grade prediction.

Description

Technical Field

[0001] This application relates to the field of mineral resource prediction technology, and in particular to a method, apparatus, storage medium and electronic equipment for predicting the pixel-level grade of iron ore. Background Technology

[0002] Iron ore, as an indispensable core raw material for the steel industry, directly impacts the stable operation of the national economy and the process of industrial upgrading through its resource reserves and development efficiency. Whether sedimentary metamorphic, magmatic, or contact metasomatic, most iron ore deposits are mined using open-pit mining. Efficient identification and precise zoning of ore grades in open-pit mines are crucial prerequisites for achieving refined mining, reducing resource waste, controlling mining dilution rates, and improving development efficiency. Imaging hyperspectral remote sensing technology, with its nanometer-level spectral resolution and refined spatial resolution, can simultaneously capture the spatial distribution information and spectral characteristic responses of ore, providing a new technical approach for the quantitative prediction of iron ore grades. This breaks through the spatiotemporal limitations of traditional methods and demonstrates enormous application potential in the field of iron ore resource exploration and development.

[0003] Currently, grade zoning in open-pit iron ore mines mainly relies on traditional technical solutions. The core process involves on-site manual sampling, laboratory chemical analysis to determine the grade, and drawing grade zoning maps based on geological maps. While this method provides basic grade data, it has significant drawbacks: First, the sampling and analysis process is cumbersome and time-consuming, resulting in a significant lag between the grade zoning results and the actual mining process, failing to provide real-time support for dynamic adjustments to the mining plan. Second, the coverage of manual sampling is limited, making it difficult to fully reflect the spatial heterogeneity of ore grades, thus leading to insufficient zoning accuracy, exacerbating resource waste, and increasing mining costs. Summary of the Invention

[0004] In view of this, this application provides a method, apparatus, storage medium and electronic equipment for predicting the grade of iron ore at the pixel level, which can improve the accuracy of iron ore pixel-level grade prediction.

[0005] According to a first aspect of this application, a method for predicting the pixel-level grade of iron ore is provided, comprising: The hyperspectral data of the iron ore to be predicted is input into the convolutional autoencoder of the pixel-level grade prediction model, and the low-dimensional features of the hyperspectral data are extracted using the convolutional autoencoder. The low-dimensional features are input into the multilayer perceptron of the pixel-level grade prediction model, and the multilayer perceptron maps the low-dimensional features to the iron ore grade prediction value of each pixel in the iron ore to be predicted. The training process of the pixel-level quality prediction model includes: Obtain the hyperspectral data and overall grade label of each of the multiple iron ore samples. Divide the overall grade of all iron ore samples into intervals and determine the distribution frequency of each grade interval. Based on the distribution frequency, use an inverse probability weighting strategy to allocate sampling weights. Based on the sampling weights and Monte Carlo random sampling rules, extract multiple pixel-level training samples from the hyperspectral data of each iron ore sample. Use the overall grade of the iron ore sample itself as the weak supervision label of all its corresponding pixel-level training samples to construct a weakly supervised training dataset. Based on the sample hyperspectral data, the convolutional autoencoder is pre-trained in an unsupervised manner to extract the sample low-dimensional features of the sample hyperspectral data and freeze the encoder parameters of the convolutional autoencoder. The low-dimensional features of the sample are associated with the overall grade labels of the samples in the weakly supervised training dataset, and the multilayer perceptron is trained under supervision to obtain a trained iron ore pixel-level grade prediction model.

[0006] According to a second aspect of this application, an iron ore pixel-level grade prediction device is provided, comprising: The extraction module is used to input the hyperspectral data of the iron ore to be predicted into the convolutional autoencoder of the pixel-level grade prediction model, and use the convolutional autoencoder to extract the low-dimensional features of the hyperspectral data. The mapping module is used to input the low-dimensional features into the multilayer perceptron of the pixel-level grade prediction model, and to map the low-dimensional features into the iron ore grade prediction value of each pixel in the iron ore to be predicted through the multilayer perceptron. The training module is used to train the pixel-level quality prediction model; The training module is specifically used for: acquiring the hyperspectral data and overall grade labels of multiple iron ore samples; dividing the overall grade of all iron ore samples into intervals and determining the distribution frequency of each grade interval; allocating sampling weights based on the distribution frequency using an inverse probability weighting strategy; extracting multiple pixel-level training samples from the hyperspectral data of each iron ore sample based on the sampling weights and Monte Carlo random sampling rules; constructing a weakly supervised training dataset by using the overall grade of the iron ore sample itself as the weakly supervised label of all its corresponding pixel-level training samples; pre-training the convolutional autoencoder using unsupervised training based on the sample hyperspectral data to extract the low-dimensional features of the sample hyperspectral data and freezing the encoder parameters of the convolutional autoencoder; associating the low-dimensional features of the samples with the overall grade labels of the samples in the weakly supervised training dataset; supervising the training of the multilayer perceptron to obtain the trained iron ore pixel-level grade prediction model.

[0007] According to a third aspect of this application, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described iron ore pixel-level grade prediction method.

[0008] According to a fourth aspect of this application, an electronic device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described iron ore pixel-level grade prediction method.

[0009] By employing the aforementioned technical solutions, the iron ore pixel-level grade prediction method, apparatus, storage medium, and electronic equipment provided in this application acquire sample hyperspectral data and overall grade labels of iron ore samples. They then construct a weakly supervised training dataset by combining grade interval division, inverse probability weighted sampling, and Monte Carlo random sampling. This eliminates the need for cumbersome pixel-level grade labeling, reducing data acquisition costs and ensuring data representativeness through sample balancing strategies, thus laying a high-quality data foundation for subsequent model training. Furthermore, by utilizing unsupervised pre-training of a convolutional autoencoder to extract low-dimensional robust features from the sample hyperspectral data and freezing encoder parameters, the core information related to iron ore grade in the sample hyperspectral data can be effectively mined, redundant noise interference can be eliminated, and the feature table can be improved. The effectiveness of the feature model is demonstrated; then, by training the association between low-dimensional features of the sample and weakly supervised labels through a multilayer perceptron, accurate mapping from features to grade can be achieved; finally, the hyperspectral data of the iron ore to be predicted is processed by the trained pixel-level grade prediction model to output pixel-level grade prediction values. This process eliminates the need for on-site manual sampling and laboratory chemical analysis, which can greatly simplify the grade acquisition process, shorten the analysis time, and completely solve the lag problem of traditional solutions. It can support the dynamic adjustment of mining plans in real time. At the same time, the pixel-level prediction accuracy can comprehensively characterize the spatial heterogeneity of grade within the ore, significantly improve the accuracy of grade zoning, reduce resource waste caused by inaccurate zoning, reduce mining costs, and provide reliable technical support for the refined and efficient development of iron ore resources.

[0010] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating a method for predicting the pixel-level grade of iron ore provided in an embodiment of this application is shown. Figure 2 A flowchart illustrating a pixel-level quality prediction model training method provided in an embodiment of this application is shown. Figure 3 A flowchart illustrating a method for predicting the pixel-level grade of iron ore according to another embodiment of this application is shown. Figure 4 This paper shows a schematic diagram of the structure of an iron ore pixel-level grade prediction device provided in an embodiment of this application; Figure 5 A schematic diagram of the structure of an iron ore pixel-level grade prediction device provided in another embodiment of this application is shown. Detailed Implementation

[0012] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0013] Currently, grade zoning in open-pit iron ore mines mainly relies on traditional technical solutions. The core process involves on-site manual sampling, laboratory chemical analysis to determine the grade, and drawing grade zoning maps based on geological maps. While this method provides basic grade data, it has significant drawbacks: First, the sampling and analysis process is cumbersome and time-consuming, resulting in a significant lag between the grade zoning results and the actual mining process, failing to provide real-time support for dynamic adjustments to the mining plan. Second, the coverage of manual sampling is limited, making it difficult to fully reflect the spatial heterogeneity of ore grades, thus leading to insufficient zoning accuracy, exacerbating resource waste, and increasing mining costs.

[0014] Accordingly, in order to solve the above-mentioned technical problems, embodiments of the present invention provide a method for predicting the grade of iron ore at the pixel level, such as... Figure 1 As shown, the method includes: Step 110: Input the hyperspectral data of the iron ore to be predicted into the convolutional autoencoder of the pixel-level grade prediction model, and use the convolutional autoencoder to extract the low-dimensional features of the hyperspectral data.

[0015] The iron ore to be predicted refers to the target iron ore region whose iron ore grade distribution information of each pixel needs to be obtained through hyperspectral data combined with a pixel-level grade prediction model. Hyperspectral data refers to integrated data of images and spectral information containing multiple continuous and subdivided spectral bands of the iron ore to be predicted, acquired through hyperspectral imaging equipment, which can simultaneously reflect the spatial distribution characteristics and spectral characteristics of the iron ore. The pixel-level grade prediction model is a pre-trained machine learning model used to map the hyperspectral data of iron ore to pixel-level grade prediction values. Its core consists of two functional modules: a convolutional autoencoder and a multilayer perceptron. The convolutional autoencoder is an unsupervised deep learning network containing an encoder and a decoder. Its core function is to extract and reconstruct features from the input hyperspectral data, achieving dimensionality reduction and effective feature representation of high-dimensional data. The multilayer perceptron is a feedforward neural network composed of multiple layers of neurons, including multiple hidden layers and one output layer. It performs nonlinear transformations and depth mapping on the low-dimensional features of the input, ultimately outputting a predicted iron ore grade. The encoder is the core submodule of the convolutional autoencoder, consisting of multiple one-dimensional convolutional layers, batch normalization layers, and pooling layers. It is used to extract features and compress dimensions layer by layer from the input hyperspectral data, outputting low-dimensional features. The one-dimensional convolutional layer is a network layer that performs convolution operations on the one-dimensional spectral dimension of the hyperspectral data, capturing local correlation features in the spectral dimension. The batch normalization layer is a network layer used to standardize the output features of the network layers, which can accelerate the model training convergence speed and improve model stability. The pooling layer is a network layer that performs downsampling operations on the feature maps output by the convolutional layers, which can compress the data dimension and reduce computational complexity while preserving core features. The decoder is a submodule of the convolutional autoencoder, consisting of multiple one-dimensional transpose layers. The system consists of convolutional layers, batch normalization layers, and upsampling layers. These layers expand and reconstruct the low-dimensional features output by the encoder, generating reconstructed spectral data with the same dimension as the input hyperspectral data. A one-dimensional transposed convolutional layer expands the feature map dimension through transposed convolution operations, corresponding to the encoder's convolutional compression process. An upsampling layer interpolates and amplifies the feature map, gradually restoring its size to match the input hyperspectral data. Hidden layers, located between the input and output layers in the multilayer perceptron, use activation functions to achieve non-linear mapping of input features, uncovering complex relationships between features and the prediction target. The output layer, the final layer of the multilayer perceptron, outputs the final prediction result, i.e., the iron ore pixel-level grade prediction value. The low-dimensional features are simplified feature vectors containing core spectral information, obtained by the encoder of the convolutional autoencoder after feature extraction and dimensional compression of the high-dimensional hyperspectral data.

[0016] In this embodiment of the present disclosure, the hyperspectral data of the iron ore to be predicted can be input into a preset pixel-level grade prediction model. This model is composed of a convolutional autoencoder and a multilayer perceptron. The convolutional autoencoder includes two major modules: an encoder and a decoder. The encoder integrates multiple one-dimensional convolutional layers, batch normalization layers, and pooling layers, while the decoder is configured with multiple one-dimensional transposed convolutional layers, batch normalization layers, and upsampling layers. The multilayer perceptron consists of multiple hidden layers and one output layer. After the data is input, the encoder of the convolutional autoencoder performs layer-by-layer feature extraction and dimensionality compression operations on the hyperspectral data, filters redundant information and retains the core spectral features related to the iron ore grade, and finally outputs the corresponding low-dimensional features, providing accurate feature support for subsequent grade prediction.

[0017] This technical solution constructs a hybrid model architecture combining convolutional autoencoders and multilayer perceptrons. By utilizing convolutional autoencoders to efficiently extract the core low-dimensional features of hyperspectral data, it can effectively filter redundant and noisy information in the data, improve the accuracy and robustness of feature representation, and lay a high-quality data foundation for subsequent grade prediction.

[0018] In specific application scenarios, such as Figure 2 As shown, the training process of the pixel-level grade prediction model includes the following steps: Step 210: Obtain the hyperspectral data and overall grade label of each of the multiple iron ore samples. Divide the overall grade of all iron ore samples into intervals and determine the distribution frequency of each grade interval. Based on the distribution frequency, use an inverse probability weighting strategy to allocate sampling weights. Based on the sampling weights and Monte Carlo random sampling rules, extract multiple pixel-level training samples from the hyperspectral data of each iron ore sample. Use the overall grade of the iron ore sample itself as the weak supervision label of all its corresponding pixel-level training samples to construct a weakly supervised training dataset.

[0019] Among them, iron ore samples are representative ore samples selected from different iron ore regions, serving as the basic data carrier for training the pixel-level grade prediction model; sample hyperspectral data are integrated data of the spectral and spatial information of iron ore samples collected by hyperspectral imaging equipment, including the continuous band spectral characteristics and spatial distribution characteristics of surface micro-regions of the samples; the overall grade label of the sample is the overall iron ore content value of a single iron ore sample determined by methods such as chemical analysis, serving as the basis for weak supervision labeling of the corresponding pixel-level training samples; the grade interval is a continuous numerical interval divided according to the overall grade value range of all iron ore samples, used to classify iron ore samples of different grade levels; the distribution frequency is the proportion of iron ore samples contained in each grade interval to the total number of samples, reflecting the distribution of samples at different grade levels; the inverse probability weighting strategy is a sample-based... The sampling weighting method assigns sampling weights based on distribution frequency, giving higher sampling weights to grade intervals with lower distribution frequencies to achieve balanced sample utilization. Sampling weights are weighting coefficients used to control the number of pixel-level training samples extracted from iron ore samples in different grade intervals; higher weights correspond to a greater number of samples extracted. The Monte Carlo random sampling rule is a sampling method based on random sampling, achieving unbiased sampling of the sample's hyperspectral data by randomly generating sampling locations. Pixel-level training samples are spectral feature samples corresponding to small spatial regions extracted from the hyperspectral data of iron ore samples, serving as direct input data for model training. Weakly supervised labels use the overall grade label of the iron ore sample as the annotation information for all its corresponding pixel-level training samples. The weakly supervised training dataset is a dataset composed of pixel-level training samples and their corresponding weakly supervised labels, used for training the pixel-level grade prediction model.

[0020] In this embodiment of the present disclosure, when dividing the overall grade of all iron ore samples into intervals and determining the distribution frequency of each grade interval, allocating sampling weights based on the distribution frequency using an inverse probability weighting strategy, and extracting multiple pixel-level training samples from the hyperspectral data of each iron ore sample based on the sampling weights and Monte Carlo random sampling rules, step 210 of the embodiment may include the following steps: Step 210-1: Based on the overall grade of all iron ore samples, use the equidistant division method to divide it into multiple continuous grade intervals to ensure coverage of the lowest to highest grade of all iron ore samples.

[0021] Among them, the equal interval division method is a kind of interval division method. Based on the numerical range of the overall grade of all samples, the numerical range is divided into multiple continuous intervals at fixed intervals; the lowest grade is the minimum value among the overall grade values ​​of all iron ore samples; the highest grade is the maximum value among the overall grade values ​​of all iron ore samples.

[0022] In this embodiment of the disclosure, the overall grade values ​​corresponding to all iron ore samples used for model training can be summarized first, and the lowest and highest values ​​among these values ​​can be determined. Based on the range of values ​​formed by the lowest and highest values, the range can be divided into multiple continuous grade intervals using the equidistant division method. This ensures that all intervals can completely cover the entire range of values ​​from the lowest to the highest grade, thereby achieving a systematic classification of samples with different grade levels.

[0023] For example, a batch of iron ore samples was collected for training. After summarizing the overall grade of all samples, it was found that the lowest grade was 25% and the highest grade was 55%. Using an equidistant division method, the grade range of 25% to 55% was divided, resulting in multiple consecutive grade intervals. These intervals completely covered the entire range from 25% to 55%, and samples of different grade levels could be assigned to their corresponding intervals. Subsequently, the proportion of samples in each interval can be statistically analyzed based on these intervals, providing data support for subsequent sampling weight allocation.

[0024] This technical solution divides the overall grade of all iron ore samples into intervals using the equidistant partitioning method. This allows for the rapid and standardized classification of samples at different grade levels, with consistent and repeatable partitioning processes. It also ensures that the intervals completely cover the entire grade range, avoiding the omission of samples at any grade level. This provides accurate and comprehensive interval partitioning data for subsequent operations such as statistically analyzing the distribution frequency of each grade interval and implementing balanced sampling. This ensures the scientific rigor and rationality of subsequent dataset construction, thereby laying the foundation for improving model training performance.

[0025] Step 210-2: Calculate the distribution frequency of each grade interval based on the number of iron ore samples in each grade interval.

[0026] In this embodiment of the disclosure, after completing the interval division of the overall grade of all iron ore samples, the specific number of iron ore samples contained in each grade interval can be counted, and then the ratio of the number of samples in each grade interval to the total number of iron ore samples can be calculated to finally obtain the distribution frequency corresponding to each grade interval.

[0027] For example, 100 iron ore samples were collected and grade intervals were divided into three intervals. The first interval contained 20 samples, the second interval contained 50 samples, and the third interval contained 30 samples. The distribution frequencies of the three intervals were calculated to be 20%, 50%, and 30%, respectively.

[0028] This technical step, by statistically analyzing the number of samples in each grade range and calculating the distribution frequency, can accurately quantify the proportion of samples of different grade levels in the overall sample set, clearly presenting the balance or difference in sample distribution. This provides crucial data support for subsequent implementation of inverse probability weighting strategies based on distribution frequency and allocation of sampling weights, ensuring the scientific rigor and relevance of the subsequent sampling process, thereby improving the representativeness of the constructed training dataset and laying a solid foundation for the model training effect.

[0029] Step 210-3: Calculate the sampling weights based on the distribution frequency using an inverse probability weighting strategy, so that the sampling weights are inversely proportional to the distribution frequency.

[0030] In this embodiment of the disclosure, the distribution frequency of each grade interval can be used as the core basis, and an inverse probability weighting strategy is adopted to calculate the sampling weight, so as to ensure that the final sampling weight is inversely proportional to the distribution frequency of the corresponding grade interval. That is, the higher the distribution frequency of the grade interval, the lower its corresponding sampling weight, and the lower the distribution frequency of the grade interval, the higher its corresponding sampling weight.

[0031] Following the example in step 210-2 of the embodiment, the 100 iron ore samples are divided into three grade intervals with distribution frequencies of 20%, 50%, and 30%, respectively. Sampling weights are calculated based on an inverse probability weighting strategy, ensuring that the weights are inversely proportional to the distribution frequencies. The final sampling weights for the three intervals are 0.5, 0.2, and 0.33, respectively. This means that the first interval with the lowest distribution frequency receives the highest weight, resulting in the largest number of pixel-level training samples extracted from each sample in that interval. Conversely, the second interval with the highest distribution frequency has the lowest weight, resulting in the fewest extracted samples, thus balancing the sampling contributions of each interval.

[0032] This technical solution employs an inverse probability weighting strategy, making the sampling weights inversely proportional to the distribution frequency. This effectively balances the sampling scale across different grade ranges, compensating for the uneven distribution where low-grade ranges have fewer samples and high-grade ranges have more. This allows samples from grade ranges that initially have a low proportion to receive more attention in subsequent sampling, ensuring that sample information from each grade range is fully integrated into the training dataset. This enhances the representativeness and balance of the dataset, providing reliable data support for the generalization ability and prediction accuracy of subsequent model training.

[0033] Step 210-4: For the hyperspectral data of each iron ore sample, determine the number of pixel-level training samples to be extracted from the iron ore sample based on the sampling weight, and randomly determine multiple sampling positions matching the number of pixel-level training samples within the effective area of ​​the sample hyperspectral data based on the Monte Carlo random sampling algorithm.

[0034] The number of pixel-level training samples is the specific number of training samples to be extracted for a single iron ore sample, determined by the sampling weight of the grade range to which the sample belongs; the Monte Carlo random sampling algorithm is a sampling method based on the principle of randomness, which achieves unbiased selection of data by randomly generating sampling positions; the effective region is the core region in the sample hyperspectral data that can truly reflect the attributes of the iron ore sample after removing invalid edge information; the sampling position is the specific spatial coordinate point determined within the effective region of the sample hyperspectral data for extracting pixel-level training samples.

[0035] In this embodiment of the disclosure, for each iron ore sample's hyperspectral data, the number of pixel-level training samples to be extracted can be determined first based on the sampling weight of the grade range to which the sample belongs. Then, the Monte Carlo random sampling algorithm is used to randomly generate multiple sampling locations that match the determined number of samples within the effective area of ​​the sample's hyperspectral data, thus clarifying the spatial location for subsequent extraction of pixel-level training samples.

[0036] Following the example in step 210-3 of the embodiment, the sampling weights for the three grade intervals are 0.5, 0.2, and 0.33, respectively. If the overall sampling baseline is set to 10 samples per iron ore sample, then each iron ore sample in the first interval (weight 0.5) requires 5 pixel-level training samples, each iron ore sample in the second interval (weight 0.2) requires 2 pixel-level training samples, and each iron ore sample in the third interval (weight 0.33) requires 3 pixel-level training samples. For a specific iron ore sample in the first interval, the Monte Carlo random sampling algorithm is used to randomly generate 5 non-repeating sampling locations within the effective area of ​​the hyperspectral data of that sample, which are then used to extract the corresponding pixel-level training samples.

[0037] This technical solution precisely controls the sampling quantity of a single iron ore sample by using sampling weights and determines the sampling location by combining the Monte Carlo random sampling algorithm. This not only continues the balance objective of inverse probability weighting, ensuring that a sufficient number of training samples are extracted from low-frequency grade ranges, but also ensures the unbiasedness of the sampling location through random sampling. This can fully cover the spatial features of the effective sample area and avoid feature omissions caused by sampling bias. The extracted pixel-level training samples can more comprehensively reflect the spatial heterogeneity of iron ore samples, further improving the quality of the training dataset and providing strong support for the model training effect.

[0038] Step 210-5: Generate a sampling pixel window of a preset size with each sampling position as the center, calculate the spectral mean of all pixels in the sampling pixel window, and form a single pixel-level training sample.

[0039] The sampling pixel window is a rectangular area centered on the sampling position and covering a certain number of adjacent pixels. It is the basic unit for extracting local spectral features. The preset size is the size specification of the sampling pixel window, which determines the number of pixels contained in the window. The spectral mean is the value obtained by arithmetically averaging the spectral data of all pixels in the sampling pixel window, which can comprehensively reflect the spectral characteristics of the area within the window.

[0040] In this embodiment of the present disclosure, for each determined sampling location, a sampling pixel window of a preset size is defined with the location as the center, covering a certain number of adjacent pixels around the location. The mean of the spectral data of all pixels within the sampling pixel window is calculated, and the obtained spectral mean is used as the core data to form a single pixel-level training sample for model training.

[0041] Following the example in step 210-4 of the embodiment, five sampling locations have been determined for a certain iron ore sample in the first grade range. The preset size of the sampling pixel window is set to 3×3 pixels. For each sampling location, a 3×3 sampling pixel window centered on it is defined. Each window covers 9 pixels. The mean of the spectral data of the 9 pixels in each window is calculated. Finally, five sets of spectral mean data are obtained. Each set of data forms a pixel-level training sample, which can be used for subsequent model training.

[0042] This technical solution forms pixel-level training samples by defining sampling pixel windows and calculating the spectral mean within the window. This effectively smooths random noise in hyperspectral data, improves the stability and reliability of sample spectral features, and integrates spectral features of local areas by using windows of preset size, avoiding random interference from single pixel data. Combined with the sampling positions determined in the early stage based on sampling weights and random sampling, it ensures that the extracted training samples are both balanced and representative, providing high-quality data input for subsequent model training, thereby improving the model's generalization ability and prediction accuracy.

[0043] Step 220: Based on the sample hyperspectral data, pre-train the convolutional autoencoder through unsupervised training to extract the low-dimensional features of the sample hyperspectral data and freeze the encoder parameters of the convolutional autoencoder.

[0044] For embodiments of this disclosure, step 220 may include the following steps: Step 220-1: Input the preprocessed sample hyperspectral data into the encoder, extract convolutional features from the sample hyperspectral data through a one-dimensional convolutional layer, standardize the convolutional features through a batch normalization layer to obtain a standardized feature map, and downsample the standardized feature map through a pooling layer to compress the dimensionality and obtain the initial low-dimensional features of the sample.

[0045] Among them, the preprocessed sample hyperspectral data is the regular data that meets the model input requirements after denoising and normalization of the original iron ore sample hyperspectral data; the standardized feature map is a feature matrix with a more regular distribution of feature values ​​after batch normalization layer processing, which can reduce the interference of data distribution differences on model training; the initial sample low-dimensional features are simplified feature vectors with a dimension much lower than the input data and containing core spectral information, obtained by the encoder layer by layer processing during the pre-training of the convolutional autoencoder.

[0046] In this embodiment of the disclosure, the regular preprocessed sample hyperspectral data can be input into the encoder of the convolutional autoencoder, and the convolutional features can be extracted sequentially through a one-dimensional convolutional layer. The batch normalization layer standardizes the extracted convolutional features to obtain a standardized feature map, and the pooling layer downsamples the standardized feature map to compress the data dimension. Finally, the initial low-dimensional features of the sample containing the core spectral information of the iron ore sample are output.

[0047] This technical solution utilizes the collaborative processing of a one-dimensional convolutional layer, a batch normalization layer, and a pooling layer within the encoder to efficiently mine core features related to iron ore grade in sample hyperspectral data. The standardization processing of the batch normalization layer effectively unifies feature distribution, accelerates model training convergence, and improves model stability. The downsampling operation of the pooling layer significantly compresses data dimensionality, reducing the computational cost of subsequent model training. The resulting low-dimensional features of the initial samples eliminate redundant noise information while retaining key spectral features, providing high-quality feature support for accurate grade mapping of subsequent multilayer perceptrons, thereby improving the overall model's prediction accuracy and generalization ability.

[0048] Step 220-2: Input the low-dimensional features of the initial sample into the decoder, expand the dimensions of the low-dimensional features of the initial sample through a one-dimensional transposed convolutional layer, standardize the expanded low-dimensional features of the initial sample through a batch normalization layer to obtain an expanded feature map, and restore the size of the expanded feature map to match the preprocessed sample hyperspectral data through an upsampling layer to obtain the reconstructed spectral data.

[0049] Among them, the expanded feature map is the feature matrix obtained after one-dimensional transposed convolution dimensional expansion and batch normalization, with a dimension higher than the low-dimensional features of the initial input sample; the reconstructed spectral data is the data that has been fully processed by the decoder, whose size matches the preprocessed sample hyperspectral data and contains reconstructed spectral information, and is used to compare with the original input data to calculate the reconstruction loss.

[0050] In this embodiment of the present disclosure, the low-dimensional features of the initial sample output by the encoder can be input into the decoder of the convolutional autoencoder. The low-dimensional features of the initial sample are then expanded sequentially through a one-dimensional transposed convolutional layer. A batch normalization layer is used to standardize the expanded low-dimensional features of the initial sample to obtain an expanded feature map. An upsampling layer is used to enlarge the size of the expanded feature map. Finally, the size of the feature map is restored to be consistent with the preprocessed hyperspectral data of the sample to obtain reconstructed spectral data.

[0051] This technical solution achieves accurate reconstruction of low-dimensional features from initial samples to hyperspectral data through the synergistic effect of a one-dimensional transposed convolutional layer, a batch normalization layer, and an upsampling layer within the decoder. The combination of one-dimensional transposed convolution and upsampling effectively restores the original dimensionality and core spectral distribution features of the hyperspectral data, while batch normalization ensures the stability of the feature distribution, facilitating efficient model training. Simultaneously, the comparison between the reconstructed spectral data and the preprocessed initial sample hyperspectral data generates a reconstruction loss, providing an optimization basis for the encoder's unsupervised training. This encourages the encoder to mine more representative core low-dimensional features, thereby improving the feature extraction capability of the entire convolutional autoencoder and laying a solid foundation for accurate mapping of subsequent pixel-level quality prediction models.

[0052] Step 220-3: Using the difference between the initial sample hyperspectral data and the reconstructed spectral data as the reconstruction loss, apply L2 regularization based on the first preset training parameters to suppress overfitting, and perform unsupervised iterative training on the convolutional autoencoder until the reconstruction loss converges to below the first preset threshold, thus completing the pre-training of the convolutional autoencoder and extracting the final low-dimensional features of the samples.

[0053] The first preset training parameters are pre-set parameters used to guide the training process of the convolutional autoencoder, including core configurations such as learning rate and number of iterations; L2 regularization is a regularization method used to suppress model overfitting, which constrains the model parameter size by adding the squared term of the parameters to the loss function; unsupervised iterative training is a cyclic training process that does not require manual labeling and relies solely on the features of the data itself to drive the model to continuously optimize parameters; the first preset threshold is a pre-set critical value for judging whether the reconstruction loss has converged, used to determine the timing of training termination; the final low-dimensional features of the samples are obtained after the convolutional autoencoder training is completed, by using the trained convolutional autoencoder to sequentially perform convolutional feature extraction, standardization, and downsampling on the preprocessed sample hyperspectral data.

[0054] In this embodiment of the disclosure, the reconstruction loss can be constructed from the difference between the sample hyperspectral data and the reconstructed spectral data. Specifically, the mean squared error (MSE) can be used as the reconstruction loss function, and the formula is as follows: In the formula, For sample hyperspectral data, To reconstruct the spectral data, N represents the batch size. Then, combining the pre-set first training parameters, L2 regularization is applied to constrain the model parameters to suppress overfitting. Unsupervised iterative training is performed on the convolutional autoencoder, continuously optimizing the model parameters to reduce reconstruction loss until the reconstruction loss decreases and stabilizes below the first preset threshold. At this point, the pre-training of the convolutional autoencoder is complete, and low-dimensional features containing core spectral information are extracted from the encoder.

[0055] This technical solution guides the unsupervised iterative training of the convolutional autoencoder through reconstruction loss, which drives the model to continuously optimize its feature extraction and reconstruction capabilities, enabling the extracted low-dimensional features to more accurately retain the core information related to iron ore grade; the application of L2 regularization can effectively suppress model overfitting and improve the model's generalization ability and stability.

[0056] Step 230: Associate the low-dimensional features of the samples with the overall grade labels of the samples in the weakly supervised training dataset, and conduct supervised training on the multilayer perceptron to obtain the trained iron ore pixel-level grade prediction model.

[0057] For embodiments of this disclosure, step 230 may include the following steps: Step 230-1: Associate the low-dimensional features of the samples extracted by the convolutional autoencoder with the corresponding overall quality labels of the samples in the weakly supervised training dataset to form training data pairs.

[0058] Among them, the training data pair is a data unit consisting of a set of low-dimensional features of samples and the corresponding overall quality label of the samples, which is the basic input unit for supervised training of multilayer perceptrons.

[0059] In this embodiment of the present disclosure, the low-dimensional features of each iron ore sample extracted by the convolutional autoencoder can be matched one-to-one with the overall grade labels of the same iron ore sample in the weakly supervised training dataset. Each set of corresponding features and labels is integrated into a data unit, and finally training data pairs for supervised training of multilayer perceptrons are formed.

[0060] This technical solution forms training data pairs by associating low-dimensional features of samples with overall grade labels, thus establishing a clear correspondence between features and targets. This provides standardized input data for the subsequent supervised training of multilayer perceptrons. It can fully utilize the high-value spectral features extracted by convolutional autoencoders and avoid the high cost of obtaining pixel-level real labels by using weak supervised labels, ensuring the effectiveness and relevance of the multilayer perceptron training process. This, in turn, enables the model to learn the accurate mapping law from spectral features to iron ore grade, improving the accuracy and reliability of the final pixel-level grade prediction.

[0061] Step 230-2: Input the training data into the multilayer perceptron, perform the first nonlinear transformation on the low-dimensional features of the sample through the first hidden layer, perform depth mapping on the low-dimensional features of the sample after the first nonlinear transformation through subsequent hidden layers, and convert the mapping result into the standardized prediction value of the training stage through the activation function of the output layer.

[0062] Among them, the standardized prediction value during the training phase is the intermediate result of the grade prediction with a regular numerical range obtained after processing by the activation function of the output layer during the model training process. It is used to compare with the label to calculate the training loss.

[0063] In this embodiment of the present disclosure, the constructed training data can be input into the multilayer perceptron. First, the first hidden layer performs a first nonlinear transformation on the low-dimensional features of the samples to mine the basic correlation information of the features. Then, the subsequent multilayer hidden layers perform progressive depth mapping on the features after the initial transformation to continuously strengthen the correlation between the features and the iron ore grade. Finally, the activation function of the output layer transforms the result after depth mapping to obtain the standardized prediction value of the training stage.

[0064] This technical solution utilizes the collaborative operation of multiple hidden layers in a multilayer perceptron to perform nonlinear transformation and depth mapping of features. This fully leverages the complex information related to iron ore grade contained in the low-dimensional features of the samples, overcoming the limitations of linear models and enhancing the mapping capability from features to grade. The standardization of the results by the output layer activation function can regulate the range of predicted values, facilitating the effective calculation of training loss and the accurate optimization of model parameters. The overall process ensures that the multilayer perceptron can efficiently learn the correspondence between spectral features and iron ore grade, laying a solid foundation for achieving accurate pixel-level grade prediction in the future.

[0065] Step 230-3: Substitute the standardized predicted values ​​from the training phase and the overall grade label of the sample into the preset loss function to calculate the loss value. With minimizing the loss value as the training objective, perform supervised iterative training of the multilayer perceptron based on the second preset training parameters with cross-validation until the loss value converges to below the second preset threshold. This determines that the multilayer perceptron training is complete, and the trained iron ore pixel-level grade prediction model is obtained.

[0066] The preset loss function is a pre-defined mathematical function used to quantify the difference between the predicted value and the overall quality label of the sample. It is a core tool for measuring the training effect of the model. Specifically, the L1 norm loss function can be used, and the formula is as follows: In the formula, For the first The grade prediction value of each spectral pixel, Y is the overall grade label of the sample; the loss value is a quantitative value calculated by a preset loss function, representing the difference between the standardized prediction value and the overall grade label of the sample during the training phase; the second preset training parameters are the core configurations that guide the training process of the multilayer perceptron, including parameters such as learning rate and iteration batch; cross-validation is a training and validation method to improve the generalization ability of the model, which avoids model overfitting by dividing the training data into multiple subsets and alternating between training and validation; supervised iterative training is a training process that uses labels as supervision signals to iteratively optimize model parameters to reduce the loss value, so that the model gradually learns the mapping law between features and the target; the second preset threshold is a preset critical value for judging whether the loss value has converged, which is used to determine the timing of termination of multilayer perceptron training.

[0067] In this embodiment of the disclosure, the standardized predicted value during the training phase and the corresponding overall grade label of the sample can be substituted into a pre-set loss function to calculate the loss value. Minimizing the loss value is the core training objective. Combined with the second preset training parameters, the multilayer perceptron is subjected to supervised iterative training through cross-validation to continuously optimize the model parameters to reduce the loss value until the loss value decreases and stabilizes below the second preset threshold. At this point, the training of the multilayer perceptron is determined to be complete. The multilayer perceptron that has been trained and the previously pre-trained convolutional autoencoder together constitute the trained iron ore pixel-level grade prediction model.

[0068] This technical solution conducts supervised iterative training with the goal of minimizing the loss value, which can drive the multilayer perceptron to accurately learn the mapping law between the low-dimensional features of the samples and the iron ore grade, thereby improving the prediction accuracy of the model. The introduction of cross-validation can effectively suppress model overfitting and enhance the model's ability to generalize to new data.

[0069] Step 120: Input the low-dimensional features into the multilayer perceptron of the pixel-level grade prediction model, and map the low-dimensional features into the predicted iron ore grade value of each pixel in the iron ore to be predicted through the multilayer perceptron.

[0070] Among them, a pixel is the basic unit of a hyperspectral image, corresponding to a tiny area on the surface of the iron ore to be predicted. The information of each pixel can reflect the spectral and spatial characteristics of the corresponding location. The iron ore grade prediction value is the estimated value of iron ore content at each pixel location of the iron ore to be predicted, calculated by the prediction model. It is the core data to guide iron ore mining and grade zoning.

[0071] In this embodiment of the present disclosure, low-dimensional features containing core spectral information of iron ore extracted by a convolutional autoencoder can be input into a multilayer perceptron of a pixel-level grade prediction model. The multilayer perceptron performs layer-by-layer deep nonlinear transformation on the low-dimensional features through its internal multilayer hidden layers to fully explore the potential correlation between the features and the iron ore grade. Finally, the output layer maps the transformed features into numerical values ​​that correspond one-to-one with the input low-dimensional features. These numerical values ​​are the predicted iron ore grade values ​​for each pixel in the iron ore to be predicted.

[0072] This technical step involves inputting high-value low-dimensional features extracted by a convolutional autoencoder into a multilayer perceptron for depth mapping. This fully leverages the multilayer perceptron's ability to fit complex features, achieving accurate conversion from spectral features to pixel-level grade. Compared to traditional grade analysis methods that rely on manual sampling, this significantly improves the spatial resolution and efficiency of grade prediction. It can comprehensively capture the spatial heterogeneity of grades within iron ore deposits, providing reliable data support for real-time grade zoning of iron ore mines and dynamic adjustment of mining plans, thereby reducing resource waste and lowering mining costs.

[0073] In summary, the iron ore pixel-level grade prediction method provided in this application acquires the sample hyperspectral data and overall grade labels of iron ore samples. It then constructs a weakly supervised training dataset by combining grade interval division, inverse probability weighted sampling, and Monte Carlo random sampling. This eliminates the need for cumbersome pixel-level grade labeling, reducing data acquisition costs and ensuring data representativeness through sample balancing strategies, thus laying a high-quality data foundation for subsequent model training. Secondly, by utilizing unsupervised pre-training of a convolutional autoencoder to extract low-dimensional robust features from the sample hyperspectral data and freezing the encoder parameters, the core information related to iron ore grade in the sample hyperspectral data can be effectively mined, redundant noise interference can be eliminated, and the effectiveness of feature representation can be improved. Furthermore, through multiple... Layer perceptrons are trained to correlate low-dimensional features of samples with weakly supervised labels, enabling precise mapping from features to grades. Finally, the hyperspectral data of the iron ore to be predicted is processed by the trained pixel-level grade prediction model, outputting pixel-level grade prediction values. This process eliminates the need for on-site manual sampling and laboratory chemical analysis, significantly simplifying the grade acquisition process, shortening analysis time, and completely solving the lag problem of traditional methods. It can support real-time dynamic adjustments to mining plans. Furthermore, the pixel-level prediction accuracy can comprehensively characterize the spatial heterogeneity of grades within the ore, significantly improving grade zoning accuracy, reducing resource waste caused by inaccurate zoning, lowering mining costs, and providing reliable technical support for the refined and efficient development of iron ore resources.

[0074] Furthermore, as a refinement and extension of the specific implementation of the above embodiments, and to fully illustrate the implementation of this embodiment, this embodiment also provides another method for predicting iron ore pixel-level grade, such as... Figure 3 As shown, the method includes: Step 310: Input the hyperspectral data of the iron ore to be predicted into the convolutional autoencoder of the pixel-level grade prediction model, and use the convolutional autoencoder to extract the low-dimensional features of the hyperspectral data.

[0075] For embodiments of this disclosure, step 310 may include the following steps: Step 310-1: Input the preprocessed hyperspectral data into the convolutional autoencoder that has been trained for the pixel-level grade prediction model.

[0076] Among them, the preprocessed hyperspectral data is the standardized data that meets the model input requirements after the original hyperspectral data of the iron ore to be predicted has been denoised, normalized and other regularized. The pixel-level grade prediction model is composed of a pre-trained convolutional autoencoder and a trained multilayer perceptron, which can realize a complete mapping model from hyperspectral data to pixel-level iron ore grade prediction values. The trained convolutional autoencoder is a convolutional autoencoder with stable feature extraction ability after unsupervised iterative training, whose reconstruction loss converges to below the preset threshold.

[0077] In this embodiment of the present disclosure, the collected hyperspectral data of iron ore to be predicted can be preprocessed to obtain regularized data. Then, the preprocessed hyperspectral data can be input into the corresponding convolutional autoencoder in the trained pixel-level grade prediction model to prepare data input for the subsequent extraction of core spectral features.

[0078] This technical solution inputs preprocessed hyperspectral data into a trained convolutional autoencoder, ensuring the standardization and adaptability of the input data and avoiding noise or formatting issues in the original data from affecting feature extraction. Simultaneously, the trained convolutional autoencoder possesses mature feature extraction capabilities, efficiently mining core information related to iron ore grade from the hyperspectral data. This provides high-quality feature input for subsequent multilayer perceptrons to achieve accurate grade mapping, ensuring the efficiency and accuracy of the entire prediction model.

[0079] Step 310-2: By sequentially processing the one-dimensional convolutional layer, batch normalization layer and pooling layer of the encoder in the convolutional autoencoder, feature extraction and dimensionality compression of the hyperspectral data are achieved, and low-dimensional features of the hyperspectral data are obtained.

[0080] In this embodiment of the disclosure, the input hyperspectral data can be processed sequentially by means of the one-dimensional convolutional layer, batch normalization layer and pooling layer contained in the encoder of the convolutional autoencoder. First, the core features related to iron ore grade are extracted by the one-dimensional convolutional layer. Then, the extracted features are standardized and normalized by the batch normalization layer. Finally, the feature dimensionality is compressed by the pooling layer, and low-dimensional features that retain the core spectral information are obtained.

[0081] This technical solution employs a collaborative and progressive processing approach involving a one-dimensional convolutional layer, a batch normalization layer, and a pooling layer within the encoder. This approach not only accurately extracts core information closely related to iron ore grade from hyperspectral data but also ensures the stability of feature distribution through the batch normalization layer, preventing data differences from interfering with subsequent processing. Simultaneously, the dimensionality compression operation of the pooling layer effectively eliminates redundant information, reducing the computational cost of data processing. The resulting low-dimensional features are both representative and concise, providing high-quality feature support for subsequent multilayer perceptrons to achieve accurate pixel-level grade mapping, thus ensuring the efficiency and accuracy of the entire prediction model.

[0082] Step 320: Input the low-dimensional features into the multilayer perceptron of the pixel-level grade prediction model, and map the low-dimensional features into the predicted iron ore grade value of each pixel in the iron ore to be predicted through the multilayer perceptron.

[0083] For embodiments of this disclosure, step 320 may include the following steps: Step 320-1: Input the low-dimensional features extracted by the convolutional autoencoder into the trained multilayer perceptron.

[0084] Among them, the trained multilayer perceptron is a multilayer perceptron that has undergone supervised iterative training, whose loss value has converged to below a preset threshold, and is able to stably learn the mapping rules between features and quality.

[0085] In this embodiment of the present disclosure, the low-dimensional features extracted by the convolutional autoencoder after processing the hyperspectral data can be directly input into the trained multilayer perceptron, so as to prepare the data input for the subsequent generation of iron ore grade prediction values ​​through feature transformation and mapping of the multilayer perceptron.

[0086] This technical solution precisely connects the high-value low-dimensional features extracted by the convolutional autoencoder with the trained multilayer perceptron. This not only fully utilizes the advantages of low-dimensional features being concise yet rich in core information, but also relies on the mature mapping capabilities of the trained multilayer perceptron to ensure efficient conversion from features to grade predictions. This seamless connection between modules avoids information loss during feature transfer, ensuring that the multilayer perceptron can quickly and accurately uncover the correlation between low-dimensional features and iron ore grade, thereby improving the overall efficiency of the prediction model and the accuracy of pixel-level grade prediction.

[0087] Step 320-2: Perform deep mapping on low-dimensional features through nonlinear transformation of multiple hidden layers in the multilayer perceptron.

[0088] Among them, depth mapping is a progressive feature processing process completed collaboratively by multiple hidden layers, which gradually strengthens the correlation between low-dimensional features and iron ore grade, and achieves accurate correspondence from features to targets.

[0089] In the embodiments of this disclosure, after the multilayer perceptron is trained by low-dimensional feature input, nonlinear transformations can be carried out sequentially by using multiple hidden layers inside the perceptron. Through progressive processing, the low-dimensional features are deeply mapped, and key information related to iron ore grade contained in the features is continuously mined, gradually building a precise correlation between features and grade targets.

[0090] This technical solution achieves deep mapping of low-dimensional features through nonlinear transformation of multiple hidden layers, which can effectively overcome the limitations of linear models and fully explore the complex potential information related to iron ore grade in low-dimensional features. The multi-layer progressive processing method can further strengthen the correlation between features and grade, improve the accuracy of mapping, and lay a solid foundation for generating accurate iron ore grade prediction values. At the same time, relying on the mature multilayer perceptron structure, it ensures the high efficiency and stability of the mapping process, helping the entire prediction model to achieve high-precision pixel-level grade prediction.

[0091] Step 320-3: Use the activation function of the output layer in the multilayer perceptron to convert the mapping result into a standardized prediction value.

[0092] The activation function is a function embedded in the output layer, used to perform nonlinear transformation on the depth mapping result and standardize the numerical range of the predicted value. The mapping result is the intermediate feature data obtained after the multiple hidden layers of the multilayer perceptron perform depth mapping on the low-dimensional features, carrying the correlation information between the features and the iron ore grade. The standardized predicted value is the intermediate grade prediction result after being transformed by the activation function of the output layer. It has a regular numerical range and meets the model prediction requirements, which facilitates further processing to obtain the final grade value.

[0093] In the embodiments of this disclosure, after the deep mapping of low-dimensional features is completed in the multiple hidden layers of the multilayer perceptron, the activation function configured in the output layer of the perceptron can be used to transform the obtained mapping result and normalize it into a standardized prediction value with a uniform numerical range.

[0094] This technical solution converts the depth mapping results into standardized predicted values ​​through the output layer activation function. This effectively standardizes the numerical range of the predicted results and avoids excessive fluctuations in the mapping results that could interfere with subsequent processing. At the same time, the nonlinear transformation capability of the activation function can further optimize the accuracy of the predicted results, making the standardized predicted values ​​more closely match the numerical distribution of the actual grade. This provides high-quality basic data for subsequent inverse normalization to obtain the final iron ore grade predicted value, ensuring the reliability and accuracy of the entire prediction model's output results.

[0095] Step 320-4: Through inverse normalization, the standardized predicted values ​​are converted into the iron ore grade predicted values ​​of the corresponding pixels.

[0096] Among them, denormalization is a data restoration operation used to restore standardized data that has been previously normalized to the numerical range of the original data, thus restoring the actual physical meaning of the data.

[0097] In this embodiment, the standardized predicted values ​​output by the multilayer sensor can be denormalized. Through inverse transformation, the standardized predicted values ​​within the normalized range are restored to the numerical range of the actual iron ore grade, ultimately obtaining iron ore grade prediction values ​​that directly characterize the iron ore content at each pixel location. Then, using the spatial location of each pixel in the model's hyperspectral data as a reference, the iron ore grade prediction value corresponding to each pixel is matched one-to-one according to its original spatial location. All matched pixel locations and grade prediction values ​​are integrated to ultimately form a pixel-level grade prediction result that fully reflects the spatial distribution of the iron ore grade. Here, the pixel location in the hyperspectral data refers to the spatial coordinates corresponding to each pixel in the hyperspectral image. These coordinates correspond one-to-one with the actual micro-regions on the surface of the iron ore to be predicted, representing the spatial positioning information of the predicted value.

[0098] This technical solution uses inverse normalization to convert standardized predicted values ​​into predicted iron ore grades, effectively restoring the actual physical meaning of the prediction results. This allows the predicted values ​​to directly correspond to the actual iron ore grade levels, meeting the actual needs for grade data in engineering applications. Simultaneously, inverse normalization ensures the accuracy of the measured values, avoiding numerical deviations caused by the standardization process. This makes the final output predicted iron ore grade more reliable, providing precise data support for subsequent iron ore mining area grade zoning and mining plan development.

[0099] Step 330: Based on the pixel position of the hyperspectral data, associate the predicted iron ore grade of each pixel in the iron ore to be predicted with the spatial coordinates of the hyperspectral data one by one, and generate an iron ore grade heat map reflecting the spatial distribution of grade based on the preset grade color scale mapping rules.

[0100] Among them, the preset grade color gradation mapping rule is a pre-set rule that corresponds the predicted iron ore grade values ​​of different ranges with specific color gradients, and the grade value and the color depth or hue form a fixed relationship; the grade spatial distribution is the distribution pattern and difference characteristics of iron ore grade in different locations within the iron ore to be predicted; the iron ore grade heat map is an image that intuitively presents the spatial distribution of the iron ore grade to be predicted through color gradients, with different colors corresponding to different grade levels, which can clearly show the spatial distribution pattern of high and low grades.

[0101] In this embodiment of the present disclosure, the predicted iron ore grade of each pixel in the iron ore to be predicted can be associated with the spatial coordinates corresponding to the hyperspectral data based on the pixel location of the hyperspectral data, thereby determining the actual spatial location of each grade prediction value. Then, according to the pre-set grade color scale mapping rules, the grade prediction values ​​of different numerical ranges are converted into corresponding colors, and finally, an iron ore grade heat map that can intuitively reflect the spatial distribution characteristics of the iron ore grade to be predicted is generated.

[0102] This technical solution generates a heat map by linking the predicted grade value with spatial coordinates and combining it with color-gradient mapping, which enables the visualization of grade distribution information. This makes the originally abstract grade values ​​intuitive and easy to understand, facilitating the rapid identification of the spatial distribution pattern of high-grade and low-grade areas. This visualization method not only improves the efficiency of interpreting grade distribution information, but also provides intuitive and accurate spatial references for decisions such as optimizing mining plans and rationally utilizing resources, effectively reducing the difficulty of decision-making and improving the scientific and economic aspects of iron ore mining.

[0103] Step 340: Using the gradient-weighted class activation mapping method, calculate the gradient weight of the feature map output by the last one-dimensional convolutional layer in the encoder relative to the iron ore grade prediction value. The gradient weight represents the degree of influence of each feature point in the feature map on the iron ore grade prediction value.

[0104] Among them, the gradient-weighted activation mapping method is an analytical method for quantifying the influence of features on prediction results. It determines the feature weights by calculating the gradient relationship between features and predicted values, thereby enabling the location and evaluation of key features.

[0105] In this embodiment of the disclosure, the Gradient Weighted Class Activation Mapping (Grad-CAM) method can be used. With the predicted iron ore grade as the target variable, the gradient of the feature map output by the last one-dimensional convolutional layer in the encoder is calculated. This gradient is then converted into a gradient weight for each feature point in the corresponding feature map. This gradient weight clarifies the degree of influence of each feature point in the feature map on the predicted iron ore grade. The calculation formula is as follows: In the formula, The gradient weight is the overall influence weight of the k-th feature point on the predicted iron ore grade in the feature map output by the last one-dimensional convolutional layer of the encoder. It is the feature point of the k-th feature at the i-th position in the feature map; This is the predicted value for iron ore grade; Iron ore grade prediction value For this feature point The partial derivative of H directly reflects the degree of influence of the change of this feature point on the grade prediction value; H is the size of the feature map in a certain dimension (such as the length of a one-dimensional convolutional feature map).

[0106] This technical solution calculates gradient weights using the gradient-weighted activation mapping method, which can accurately quantify the influence of each feature point in low-dimensional features on the grade prediction results and clearly locate the core feature points that play a key role in the prediction results.

[0107] Step 350: Perform global average pooling on the gradient weights, sum the obtained average weights with the low-dimensional features, and then filter out the features with negative weights using the ReLU non-linear activation function to obtain the activated features.

[0108] Among them, global average pooling is a global average calculation of gradient weights, which integrates the scattered gradient weights into an average weight that comprehensively reflects the overall importance of the features; the activation features are features obtained after weighted summation and ReLU function processing, which only contain core feature information that has a positive contribution to iron ore grade prediction.

[0109] In this embodiment of the disclosure, the calculated gradient weights can first be subjected to global average pooling to obtain an average weight that comprehensively reflects the overall importance of the features. Then, the average weight is weighted and summed with the corresponding low-dimensional features to adjust the importance ratio of the features. Finally, the ReLU nonlinear activation function is used to filter out the feature information with negative weights after weighting, and finally, the activated features that retain only positive and valid information are obtained.

[0110] Step 360: Map the activation features back to the original spectral band dimensions of the hyperspectral data through deconvolution operation, and establish a one-to-one correspondence between the activation features and the original spectral bands.

[0111] Among them, the deconvolution operation, as an operation to expand the feature dimension and map the space, can reverse the simplified features to restore the original feature space with a higher dimension; the original spectral band dimension of hyperspectral data is the dimension corresponding to the number of spectral bands initially included in the collected hyperspectral data, which is the original feature dimension of hyperspectral data.

[0112] In this embodiment of the disclosure, a deconvolution operation can be performed on the obtained activation features. This operation is used to reverse-map the dimension-simplified activation features to the original spectral band dimension of the hyperspectral data, so that each information unit of the activation feature establishes a precise one-to-one correspondence with the corresponding original spectral band.

[0113] Step 370: Based on the mapping relationship, calculate the weighted sum of each original spectral band as the contribution value of the original spectral band to the iron ore grade prediction value. Sort the contribution values ​​from high to low, and select and output the key feature bands that play a core role in the iron ore grade prediction from the original spectral bands.

[0114] Among them, the key characteristic bands are those spectral bands that have a high contribution value in the original spectral bands and play a major role in predicting iron ore grade.

[0115] In this embodiment of the disclosure, the contribution value of each original spectral band can be calculated by summing the weights corresponding to the activation features and the original spectral bands, and then all the original spectral bands can be sorted from high to low according to the contribution value. One or more key feature bands with higher contribution values ​​can be selected from the sorting results, and finally these key feature bands that play a core role in predicting iron ore grade are output.

[0116] This technical solution calculates the contribution value of the original spectral bands and filters key feature bands, which can clearly identify the original spectral information that plays a core role in iron ore grade prediction. This not only reduces the dimensionality and computational cost of subsequent hyperspectral data processing, but also improves the targeting of data applications. At the same time, the output of key feature bands can provide accurate direction for the acquisition and analysis of hyperspectral data, enhance the efficiency and interpretability of the iron ore grade prediction process, and promote the practical application of hyperspectral technology in iron ore resource assessment.

[0117] In summary, the iron ore pixel-level grade prediction method provided in this application inputs preprocessed hyperspectral data into a trained convolutional autoencoder. Through sequential processing of one-dimensional convolutional layers, batch normalization layers, and pooling layers within the encoder, it can accurately extract core grade-related information from the hyperspectral data while eliminating redundant noise through dimensionality compression, ensuring the robustness and effectiveness of low-dimensional features and laying a solid foundation for high-quality features in subsequent grade mapping. Subsequently, the low-dimensional features are input into a multilayer perceptron, undergoing nonlinear transformation through multiple hidden layers to achieve deep feature mapping. The output layer activation function converts the data into standardized predicted values, which are then inversely normalized to obtain the actual grade prediction value. Finally, the complete prediction result is output according to pixel location. This process eliminates the cumbersome manual sampling and laboratory chemical analysis steps of traditional methods, significantly simplifying the operation process, shortening analysis time, and completely solving the lag problem of traditional grade zoning. It can provide real-time data support for dynamic adjustments to mining plans, and the pixel-level prediction accuracy can comprehensively capture the spatial heterogeneity of grades within the ore, significantly improving zoning accuracy. Based on this... This method associates grade predictions with the spatial coordinates of hyperspectral data, generating iron ore grade heatmaps based on predefined grade color-gradient mapping rules. This visually presents the spatial distribution characteristics of grade, reducing the difficulty of interpreting and applying subsequent mining plans. Furthermore, a gradient-weighted activation mapping method is used to calculate the gradient weights of low-dimensional features relative to grade predictions. Activation features are obtained by filtering negative contribution features through global average pooling, weighted summation, and the ReLU nonlinear activation function. These activation features are then deconvolutionally mapped back to the original spectral band dimensions of the hyperspectral data, establishing a one-to-one correspondence. Finally, based on this mapping, the contribution of each original spectral band is calculated, and key feature bands are selected. This not only clarifies the core spectral information affecting grade prediction, providing a clear direction for targeted optimization and lightweight model upgrades in subsequent hyperspectral data acquisition, but also further enhances the model's generalization ability and prediction reliability. Ultimately, this achieves refined management of iron ore resources, reducing resource waste and increased mining costs caused by inaccurate zoning, and providing comprehensive technical support for the efficient development of iron ore resources.

[0118] Furthermore, as Figure 1 and Figure 3 The specific implementation of the method shown in this embodiment provides an iron ore pixel-level grade prediction device, such as... Figure 4 As shown, the device includes: an extraction module 41, a mapping module 42, and a training module 43; Extraction module 41 can be used to input the hyperspectral data of the iron ore to be predicted into the convolutional autoencoder of the pixel-level grade prediction model, and use the convolutional autoencoder to extract the low-dimensional features of the hyperspectral data. The mapping module 42 can be used to input low-dimensional features into the multilayer perceptron of the pixel-level grade prediction model, and map the low-dimensional features to the iron ore grade prediction value of each pixel in the iron ore to be predicted through the multilayer perceptron. Training module 43 can be used to train a pixel-level quality prediction model; The training module 43 can be used to: acquire the hyperspectral data and overall grade labels of multiple iron ore samples; divide the overall grade of all iron ore samples into intervals and determine the distribution frequency of each grade interval; allocate sampling weights based on the distribution frequency using an inverse probability weighting strategy; extract multiple pixel-level training samples from the hyperspectral data of each iron ore sample based on the sampling weights and Monte Carlo random sampling rules; construct a weakly supervised training dataset by using the overall grade of the iron ore sample itself as the weakly supervised label of all its corresponding pixel-level training samples; pre-train a convolutional autoencoder using unsupervised training based on the sample hyperspectral data to extract the low-dimensional features of the sample hyperspectral data and freeze the encoder parameters of the convolutional autoencoder; associate the low-dimensional features of the samples with the overall grade labels of the samples in the weakly supervised training dataset to supervise the training of the multilayer perceptron, and obtain the trained iron ore pixel-level grade prediction model.

[0119] In some embodiments of this application, the training module 43 can be specifically used to divide the overall grade of all iron ore samples into multiple continuous grade intervals using an equidistant division method, ensuring coverage of the lowest to highest grade of all iron ore samples; calculate the distribution frequency of each grade interval based on the number of iron ore samples in each grade interval; calculate the sampling weight using an inverse probability weighting strategy based on the distribution frequency, making the sampling weight inversely proportional to the distribution frequency; for the sample hyperspectral data of each iron ore sample, determine the number of pixel-level training samples to be extracted from the iron ore sample based on the sampling weight, and randomly determine multiple sampling positions matching the number of pixel-level training samples within the effective area of ​​the sample hyperspectral data based on the Monte Carlo random sampling algorithm; generate a sampling pixel window of a preset size with each sampling position as the center, calculate the spectral mean of all pixels within the sampling pixel window, and form a single pixel-level training sample.

[0120] In some embodiments of this application, the convolutional autoencoder includes an encoder and a decoder. The encoder includes multiple one-dimensional convolutional layers, multiple batch normalization layers, and multiple pooling layers. The decoder includes multiple one-dimensional transposed convolutional layers, multiple batch normalization layers, and multiple upsampling layers. The training module 43 is specifically used to input preprocessed sample hyperspectral data into the encoder, extract convolutional features from the sample hyperspectral data through one-dimensional convolutional layers, standardize the convolutional features through batch normalization layers to obtain a standardized feature map, and downsample the standardized feature map through pooling layers to compress the dimensionality, obtaining initial low-dimensional features of the sample. The initial low-dimensional features of the sample are then input into the decoder. The device expands the low-dimensional features of the initial samples through a one-dimensional transposed convolutional layer, and then standardizes the expanded low-dimensional features through a batch normalization layer to obtain an expanded feature map. An upsampling layer restores the size of the expanded feature map to match the size of the preprocessed hyperspectral data of the samples, resulting in reconstructed spectral data. The difference between the initial hyperspectral data and the reconstructed spectral data is used as the reconstruction loss. L2 regularization is applied based on the first preset training parameters to suppress overfitting. The convolutional autoencoder is then subjected to unsupervised iterative training until the reconstruction loss converges to below the first preset threshold, thus completing the pre-training of the convolutional autoencoder and extracting the final low-dimensional features of the samples.

[0121] In some embodiments of this application, the multilayer perceptron includes multiple hidden layers and an output layer; the training module 43 is specifically used to associate the low-dimensional features of the samples extracted by the convolutional autoencoder with the corresponding overall grade labels of the samples in the weakly supervised training dataset to form training data pairs; the training data pairs are input into the multilayer perceptron, the low-dimensional features of the samples are subjected to a first nonlinear transformation through the first hidden layer, the low-dimensional features of the samples after the first nonlinear transformation are subjected to depth mapping through subsequent hidden layers, and the mapping result is converted into a standardized prediction value for the training stage through the activation function of the output layer; the standardized prediction value for the training stage and the overall grade label of the samples are substituted into a preset loss function to calculate the loss value, with the minimization of the loss value as the training objective, and the multilayer perceptron is subjected to supervised iterative training with cross-validation based on the second preset training parameters until the loss value converges to below the second preset threshold, and the multilayer perceptron training is determined to be completed, thus obtaining a trained iron ore pixel-level grade prediction model.

[0122] In some embodiments of this application, the extraction module 41 can be specifically used to input the preprocessed hyperspectral data into the convolutional autoencoder corresponding to the pixel-level grade prediction model after training; through the one-dimensional convolutional layer, batch normalization layer and pooling layer of the encoder in the convolutional autoencoder, the feature extraction and dimensionality compression of the hyperspectral data are realized, and the low-dimensional features of the hyperspectral data are obtained.

[0123] In some embodiments of this application, the mapping module 42 can be specifically used to input the low-dimensional features extracted by the convolutional autoencoder into the trained multilayer perceptron; perform deep mapping on the low-dimensional features through nonlinear transformation of the multiple hidden layers in the multilayer perceptron; convert the mapping result into a standardized prediction value using the activation function of the output layer in the multilayer perceptron; and convert the standardized prediction value into the iron ore grade prediction value of the corresponding pixel through inverse normalization processing.

[0124] In some embodiments of this application, such as Figure 5 As shown, the device also includes: an association module 44, a calculation module 45, a processing module 46, and a filtering module 47; The association module 44 can be used to associate the predicted iron ore grade of each pixel in the iron ore to be predicted with the spatial coordinates of the hyperspectral data according to the pixel position of the hyperspectral data, and generate an iron ore grade heat map reflecting the spatial distribution of grade based on the preset grade color scale mapping rules. Calculation module 45 can be used to calculate the gradient weight of low-dimensional features relative to the predicted iron ore grade using the gradient weighted class activation mapping method. The gradient weight represents the degree of influence of each feature point in the low-dimensional features on the predicted iron ore grade. Processing module 46 can be used to perform global average pooling on gradient weights, sum the obtained average weights with low-dimensional features, and then filter out features with negative weights through the ReLU nonlinear activation function to obtain activated features. The mapping module 42 can also be used to map the activation features back to the original spectral band dimension of the hyperspectral data through deconvolution operation, and establish a one-to-one correspondence between the activation features and the original spectral bands. The filtering module 47 can be used to calculate the weighted sum of each original spectral band based on the mapping relationship, which is used as the contribution value of the original spectral band to the iron ore grade prediction value. The contribution values ​​are sorted from high to low, and the key feature bands that play a core role in the iron ore grade prediction are filtered and output from the original spectral bands.

[0125] It should be noted that other corresponding descriptions of the functional units involved in the iron ore pixel-level grade prediction device provided in this embodiment can be found in [reference needed]. Figure 1 and Figure 3 The corresponding description in [the document] will not be repeated here.

[0126] Based on the above, Figure 1 and Figure 3 Accordingly, this embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the above-described method. Figure 1 and Figure 3 The method for predicting the grade of iron ore at the pixel level is shown.

[0127] Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause an electronic device (such as personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of this application.

[0128] Based on the above, Figure 1 and Figure 3 The method shown, and Figure 4 , 5 To achieve the above objectives, the present application also provides an electronic device, specifically a personal computer, tablet computer, server, or other network device, as shown in the virtual device embodiment. This device includes a storage medium and a processor; the storage medium stores a computer program; the processor executes the computer program to achieve the above-described objectives. Figure 1 and Figure 3 The method for predicting the grade of iron ore at the pixel level is shown.

[0129] Optionally, the aforementioned physical devices may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB interfaces, card reader interfaces, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Wi-Fi interfaces), etc.

[0130] Those skilled in the art will understand that the physical device structure provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0131] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the aforementioned physical device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software in the information processing physical device.

[0132] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platform, or it can be implemented by hardware.

[0133] This invention acquires hyperspectral data and overall grade labels of iron ore samples, and constructs a weakly supervised training dataset by combining grade interval division, inverse probability weighted sampling, and Monte Carlo random sampling. This eliminates the need for cumbersome pixel-level grade labeling, reducing data acquisition costs and ensuring data representativeness through sample balancing strategies, thus laying a high-quality data foundation for subsequent model training. Secondly, unsupervised pre-training using a convolutional autoencoder extracts low-dimensional robust features from the sample hyperspectral data and freezes the encoder parameters, effectively mining core information related to iron ore grade from the sample hyperspectral data, eliminating redundant noise interference, and improving the effectiveness of feature representation. Finally, a multilayer perceptron is used to process the samples... The association training between low-dimensional features and weakly supervised labels enables accurate mapping from features to grades. Finally, the hyperspectral data of the iron ore to be predicted is processed by the trained pixel-level grade prediction model to output pixel-level grade prediction values. This process eliminates the need for on-site manual sampling and laboratory chemical analysis, significantly simplifying the grade acquisition process, shortening analysis time, and completely solving the lag problem of traditional solutions. It can support the dynamic adjustment of mining plans in real time. At the same time, the pixel-level prediction accuracy can comprehensively characterize the spatial heterogeneity of grades within the ore, significantly improve the accuracy of grade zoning, reduce resource waste caused by inaccurate zoning, lower mining costs, and provide reliable technical support for the refined and efficient development of iron ore resources.

[0134] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application. Those skilled in the art will understand that the modules in the apparatus of the embodiment can be distributed within the apparatus of the embodiment as described, or can be modified to be located in one or more apparatuses different from this embodiment. The modules of the above-described embodiment can be combined into one module, or further divided into multiple sub-modules.

[0135] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of any particular implementation scenario. The above disclosures are merely a few specific implementation scenarios of this application; however, this application is not limited thereto, and any variations conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. A method for predicting the pixel-level grade of iron ore, characterized in that, include: The hyperspectral data of the iron ore to be predicted is input into the convolutional autoencoder of the pixel-level grade prediction model, and the low-dimensional features of the hyperspectral data are extracted using the convolutional autoencoder. The low-dimensional features are input into the multilayer perceptron of the pixel-level grade prediction model, and the multilayer perceptron maps the low-dimensional features to the iron ore grade prediction value of each pixel in the iron ore to be predicted. The training process of the pixel-level quality prediction model includes: Obtain the hyperspectral data and overall grade label of each of the multiple iron ore samples. Divide the overall grade of all iron ore samples into intervals and determine the distribution frequency of each grade interval. Based on the distribution frequency, use an inverse probability weighting strategy to allocate sampling weights. Based on the sampling weights and Monte Carlo random sampling rules, extract multiple pixel-level training samples from the hyperspectral data of each iron ore sample. Use the overall grade of the iron ore sample itself as the weak supervision label of all its corresponding pixel-level training samples to construct a weakly supervised training dataset. Based on the sample hyperspectral data, the convolutional autoencoder is pre-trained in an unsupervised manner to extract the sample low-dimensional features of the sample hyperspectral data and freeze the encoder parameters of the convolutional autoencoder. The low-dimensional features of the sample are associated with the overall grade labels of the samples in the weakly supervised training dataset, and the multilayer perceptron is trained under supervision to obtain a trained iron ore pixel-level grade prediction model.

2. The method according to claim 1, characterized in that, The overall grade of all iron ore samples is divided into intervals, and the distribution frequency of each grade interval is determined. Based on the distribution frequency, an inverse probability weighting strategy is used to allocate sampling weights. Based on the sampling weights and Monte Carlo random sampling rules, multiple pixel-level training samples are extracted from the hyperspectral data of each iron ore sample, including: For the overall grade of all iron ore samples, the equidistant division method is used to divide it into multiple continuous grade intervals to ensure coverage of the lowest to highest grades of all iron ore samples. Based on the number of iron ore samples in each grade range, the distribution frequency of each grade range is calculated. Based on the distribution frequency, an inverse probability weighting strategy is used to calculate the sampling weight, so that the sampling weight is inversely proportional to the distribution frequency; For each iron ore sample's hyperspectral data, the number of pixel-level training samples to be extracted from the iron ore sample is determined based on the sampling weight. Multiple sampling positions matching the number of pixel-level training samples are randomly determined within the effective area of ​​the sample's hyperspectral data based on the Monte Carlo random sampling algorithm. A sampling pixel window of a preset size is generated with each sampling location as the center, and the spectral mean of all pixels within the sampling pixel window is calculated to form a single pixel-level training sample.

3. The method according to claim 1, characterized in that, The convolutional autoencoder includes an encoder and a decoder. The encoder includes multiple one-dimensional convolutional layers, multiple batch normalization layers, and multiple pooling layers. The decoder includes multiple one-dimensional transposed convolutional layers, multiple batch normalization layers, and multiple upsampling layers. Based on the sample hyperspectral data, the convolutional autoencoder is pre-trained using an unsupervised training method to extract low-dimensional features from the sample hyperspectral data, including: The preprocessed sample hyperspectral data is input into the encoder. The one-dimensional convolutional layer extracts convolutional features from the sample hyperspectral data. The batch normalization layer standardizes the convolutional features to obtain a standardized feature map. The pooling layer downsamples the standardized feature map to compress the dimensionality, resulting in the initial low-dimensional features of the sample. The initial low-dimensional features of the sample are input into the decoder. The initial low-dimensional features of the sample are expanded in dimension by the one-dimensional transposed convolutional layer. The expanded initial low-dimensional features of the sample are standardized by the batch normalization layer to obtain an expanded feature map. The size of the expanded feature map is restored to match the preprocessed sample hyperspectral data by the upsampling layer to obtain reconstructed spectral data. The difference between the initial sample hyperspectral data and the reconstructed spectral data is used as the reconstruction loss. Based on the first preset training parameters, L2 regularization is applied to suppress overfitting. The convolutional autoencoder is then subjected to unsupervised iterative training until the reconstruction loss converges to below the first preset threshold. This completes the pre-training of the convolutional autoencoder and extracts the final low-dimensional features of the samples.

4. The method according to claim 1, characterized in that, The multilayer perceptron includes multiple hidden layers and an output layer; Associating the low-dimensional features of the samples with the overall grade labels of the samples in the weakly supervised training dataset, and performing supervised training on the multilayer perceptron, yields a trained iron ore pixel-level grade prediction model, including: The low-dimensional features of the samples extracted by the convolutional autoencoder are associated one by one with the overall quality labels of the corresponding samples in the weakly supervised training dataset to form training data pairs. The training data is input into the multilayer perceptron. The first nonlinear transformation is performed on the low-dimensional features of the sample through the first hidden layer. The low-dimensional features of the sample after the first nonlinear transformation are then deeply mapped through the subsequent hidden layers. The mapping result is converted into a standardized prediction value during the training phase through the activation function of the output layer. The standardized predicted values ​​from the training phase and the overall grade label of the sample are substituted into a preset loss function to calculate the loss value. The loss value is minimized as the training objective. The multilayer perceptron is subjected to supervised iterative training with cross-validation based on the second preset training parameters until the loss value converges to below the second preset threshold. The training of the multilayer perceptron is then determined to be complete, and the trained iron ore pixel-level grade prediction model is obtained.

5. The method according to claim 3, characterized in that, The process of inputting hyperspectral data of the iron ore to be predicted into a pixel-level grade prediction model using a convolutional autoencoder, and extracting low-dimensional features from the hyperspectral data using the convolutional autoencoder, includes: The preprocessed hyperspectral data is input into the pixel-level grade prediction model corresponding to the trained convolutional autoencoder. The hyperspectral data is processed sequentially by the one-dimensional convolutional layer, batch normalization layer, and pooling layer of the encoder in the convolutional autoencoder to achieve feature extraction and dimensionality compression, thereby obtaining the low-dimensional features of the hyperspectral data.

6. The method according to claim 1, characterized in that, The low-dimensional features are input into the multilayer perceptron of the pixel-level grade prediction model. The multilayer perceptron maps the low-dimensional features to the predicted iron ore grade for each pixel in the iron ore to be predicted, including: The low-dimensional features extracted by the convolutional autoencoder are input into the trained multilayer perceptron. The low-dimensional features are deeply mapped by nonlinear transformation of the multiple hidden layers in the multilayer perceptron. The mapping result is converted into a standardized prediction value using the activation function of the output layer in the multilayer perceptron. By performing inverse normalization, the standardized predicted values ​​are converted into predicted iron ore grades for the corresponding pixels.

7. The method according to claim 1, characterized in that, The method further includes: Based on the pixel position of the hyperspectral data, the predicted iron ore grade of each pixel in the iron ore to be predicted is associated with the spatial coordinates of the hyperspectral data one by one, and an iron ore grade heat map reflecting the spatial distribution of grade is generated based on the preset grade color scale mapping rules. The gradient weighted class activation mapping method is used to calculate the gradient weight of the low-dimensional feature relative to the iron ore grade prediction value. The gradient weight represents the degree of influence of each feature point in the low-dimensional feature on the iron ore grade prediction value. The gradient weights are subjected to global average pooling, the average weights are weighted and summed with the low-dimensional features, and then the features with negative weights are filtered out by the ReLU non-linear activation function to obtain the activated features. The activation features are mapped back to the original spectral band dimension of the hyperspectral data through deconvolution operation, thus establishing a one-to-one correspondence between the activation features and the original spectral bands. Based on the mapping relationship, the weighted summation value corresponding to each of the original spectral bands is calculated as the contribution value of the original spectral bands to the iron ore grade prediction value. The contribution values ​​are sorted from high to low, and the key feature bands that play a core role in the iron ore grade prediction are selected and output from the original spectral bands.

8. A device for predicting the grade of iron ore at the pixel level, characterized in that, include: The extraction module is used to input the hyperspectral data of the iron ore to be predicted into the convolutional autoencoder of the pixel-level grade prediction model, and use the convolutional autoencoder to extract the low-dimensional features of the hyperspectral data. The mapping module is used to input the low-dimensional features into the multilayer perceptron of the pixel-level grade prediction model, and to map the low-dimensional features into the iron ore grade prediction value of each pixel in the iron ore to be predicted through the multilayer perceptron. The training module is used to train the pixel-level quality prediction model; The training module is specifically used to: acquire the hyperspectral data and overall grade label of each of the multiple iron ore samples; divide the overall grade of all iron ore samples into intervals and determine the distribution frequency of each grade interval; allocate sampling weights based on the distribution frequency using an inverse probability weighting strategy; extract multiple pixel-level training samples from the hyperspectral data of each iron ore sample based on the sampling weights and Monte Carlo random sampling rules; and construct a weakly supervised training dataset by using the overall grade of the iron ore sample itself as the weakly supervised label of all its corresponding pixel-level training samples. Based on the sample hyperspectral data, the convolutional autoencoder is pre-trained in an unsupervised manner to extract the sample low-dimensional features of the sample hyperspectral data and freeze the encoder parameters of the convolutional autoencoder; the sample low-dimensional features are associated with the overall grade labels of the samples in the weakly supervised training dataset, and the multilayer perceptron is trained in a supervised manner to obtain the trained iron ore pixel-level grade prediction model.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

10. An electronic device comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

Images