AI intelligent coal blending method based on deep learning

Abstract

The invention discloses an AI intelligent coal blending method based on deep learning, relates to the technical field of intelligent coal blending, and solves the technical problems of low efficiency in feature association recognition and screening and poor model architecture scene adaptability. The Pearson coefficient is used for linear features, the mutual information method is used for hierarchical redundancy elimination for nonlinear features, the limitation of traditional single statistic feature screening is avoided, hidden association is generated and mined through composite features, feature dimensions are more simplified, association with coal blending targets is more direct, the model training efficiency and prediction precision are effectively improved, and the method is suitable for large-scale popularization and application. Models are customized for different scenes, compared with an existing method that a single model adapts to a full scene, precise coverage of lightweight-dynamic-complex scenes is achieved, and the applicability and optimization effect of coal blending schemes under different scenes are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent coal blending technology, specifically to an AI-based intelligent coal blending method based on deep learning. Background Technology

[0002] Coal blending refers to the process of mixing and blending two or more types of coal with different properties in a certain proportion according to specific production needs, thereby obtaining a blended coal with better performance and that better meets usage standards. The coal blending scheme is central to coking production; its core function is to balance output, cost, and by-product revenue while ensuring coke quality and maintaining production stability by controlling the coal type ratio. The coal blending scheme has a decisive impact on subsequent industrial production, as well as the quantity and quality of the products. Therefore, AI-based coal blending has significant research value.

[0003] Current coal blending processes suffer from insufficient intelligence, rely heavily on human experience, and have a high probability of errors. Resource waste due to unreasonable coal blending schemes occurs frequently. Existing intelligent systems also have many problems, such as being unable to handle special situations and having poor anti-interference capabilities. These problems not only restrict the stability of coke quality and the competitiveness of enterprises, but also make optimizing coal blending processes through intelligent technology an important direction for upgrading the coking industry. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an AI-based intelligent coal blending method based on deep learning, which solves the problems of inefficient feature association recognition and screening, and poor model architecture adaptability to various scenarios.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an AI-based intelligent coal blending method based on deep learning, which specifically includes the following steps: Collect basic coal quality data, production process data, and target and constraint data. Preprocess the production process data and target and constraint data, including handling missing values ​​and verifying their reasonableness. The core input features are classified into relevant features and irrelevant features. The Pearson coefficient is used to remove redundant features of the relevant features, and the mutual information method is used to remove redundant features of the irrelevant features. The retained features are combined to generate composite features. Different model architectures are selected for different scenarios, the dataset is divided and a loss function is selected based on the number of targets. Hyperparameters are adjusted by optimizing the algorithm, and the model accuracy is verified using the test set. If the model does not meet the target, the process is returned to step one and optimized again. Collect real-time coal quality data and load constraints, input the trained model to generate candidate solutions, push them to the decision-making level for selection, and trigger model retraining or parameter adjustment based on feedback from the decision-making level.

[0006] As a further aspect of the present invention, the basic coal quality data includes industrial analysis indicators, elemental analysis indicators, and process characteristic indicators, which are collected using a combination of offline testing and online monitoring. Production process data includes equipment operating conditions, historical coal blending records, and actual production results, which are captured in real time through the manufacturing execution system and distributed control system. The target and constraint data includes target indicators and constraints. The target indicators cover cost, quality, and environmental dimensions, while the constraints include coal availability, equipment limitations, and process limitations.

[0007] As a further aspect of the present invention, linear interpolation is used to handle missing values: After sorting the target features by collection time to obtain an ordered sequence, the missing points are located. The missing values ​​are estimated using a linear formula based on the known values ​​before and after the missing points and the corresponding times. The rationality check adopts the 3σ principle. After verifying that the target feature is normally distributed, the mean μ and standard deviation σ are calculated, Lower = μ - 3σ, Upper = μ + 3σ. Outliers exceeding the threshold are marked and processed.

[0008] As a further aspect of the present invention, the classification of relevance features and non-relevance features is as follows: Plot a scatter plot with feature A as the x-axis and feature B as the y-axis. If the scatter points are distributed along a straight line, they tend to be linearly correlated; if they are distributed in a curve or a non-linear pattern, they tend to be non-linearly correlated. Next, the residual plot is used for verification. A scatter plot is drawn with feature A as the x-axis and the residual as the y-axis. If the residuals are randomly distributed on both sides of y=0, it is a linear correlation. If the residuals show a systematic trend, it is a non-linear correlation.

[0009] As a further aspect of the present invention, when performing redundant feature removal for the Pearson coefficient, all continuous features are obtained and standardized, a Pearson correlation coefficient matrix between features is constructed, the correlation coefficient is calculated and compared with a high correlation threshold, feature pairs with correlation coefficients exceeding the threshold are selected, and the features with the highest correlation to the target indicator are retained. The high correlation threshold is set by business requirements and is set to 0.8.

[0010] As a further aspect of the present invention, when performing redundant feature removal using the mutual information method, the continuous features and the target features are discretized, the mutual information value between the continuous features and the target features is calculated, the top N features are selected as screening features by sorting the mutual information values ​​from largest to smallest, the mutual information values ​​between each pair of screening features are calculated again, redundant features with mutual information values ​​exceeding the threshold are removed to obtain retained features, and the retained features are normalized to the [0,1] interval.

[0011] As a further aspect of the present invention, the scenario-based model architecture is specifically as follows: For single-objective optimization, a basic regression model based on multilayer perceptron is selected. The number of neurons in the input layer is consistent with the feature dimension. There are 3-5 hidden layers with the number of neurons in each layer decreasing by 2 / 3 of the number in the input layer. The output layer has 1 neuron. For scenarios with large fluctuations in coal quality, we select LSTM or GRU time series models. The input is a sliding window time series feature, with 2-3 hidden layers and 64-128 neurons in each layer. The output is the dynamic ratio for the next 1-3 hours. For high-value-added scenarios such as coking and coal chemical industry, the Transformer model or hybrid intelligent model is selected. The hybrid intelligent model includes models that combine CNN+MLP+linear programming, deep learning and traditional algorithms.

[0012] As a further aspect of the present invention, the dataset is divided into a training set, a validation set, and a test set in a 7:2:1 time order, and a loss function is selected based on the target number. In the case of a single objective, the mean squared error is selected to calculate the deviation; in the case of multiple objectives, a weighted loss function is selected and combined with the algorithm to generate the Pareto optimal solution.

[0013] As a further aspect of the present invention, the Pearson correlation coefficient matrix between features is constructed as follows: Obtain all continuous features and standardize them, denoted as X, where X = 1, 2, ..., Y, and Y represents the type of continuous feature. Construct the Pearson correlation coefficient matrix R between the features. , where r XY Let represent the Pearson correlation coefficient between feature X and feature Y, and its value ranges from [-1, 1], where b is the sample size and x is the number of samples. a Let represent the observed value of feature X in the a-th sample, and ya represent the observed value of feature Y in the a-th sample. r represents the sample mean of feature X. Y This represents the sample mean of feature Y.

[0014] As a further aspect of the present invention, the mutual information value between continuous features and target features is calculated as follows: Obtain continuous features and target features, and discretize the continuous features and target indicators to convert them into discrete variables. Then, calculate the mutual information value between the continuous feature X and the target feature Y according to the formula. The mutual information values ​​of continuous feature X and target feature Y are calculated, where P(X,Y) is the joint probability, P(X) is the marginal probability of X taking the value x, and P(Y) is the marginal probability of Y taking the value y.

[0015] This invention provides an AI-based intelligent coal blending method based on deep learning. Compared with existing technologies, it has the following advantages: This invention employs a multi-source data acquisition mode combining offline testing and online monitoring, along with refined preprocessing using linear interpolation and the 3σ principle. This approach preserves the high accuracy of offline data while leveraging the real-time performance of online data. By using time-series alignment and outlier-linked equipment status verification, it addresses the issues of traditional coal blending data being dimensionally limited and of poor quality, providing a more reliable input for the model.

[0016] This invention innovatively distinguishes between linear and nonlinear correlation features through a scatter plot and residual plot system. For linear features, Pearson coefficient is used, and for nonlinear features, mutual information is used to remove redundancy in a hierarchical manner. This not only avoids the limitations of traditional single statistical quantity screening of features, but also mines implicit correlations through composite feature generation, making the feature dimensions more concise and the correlation with coal blending targets more direct, effectively improving model training efficiency and prediction accuracy.

[0017] This invention customizes models for different scenarios. For single-objective scenarios, a lightweight MLP is used to meet the high-efficiency needs of small and medium-sized enterprises. For coal quality fluctuation scenarios, LSTM / GRU is used to capture temporal dynamics. For high-value-added scenarios, hybrid models such as Transformer or CNN+MLP+linear programming are used to handle multimodal data and multi-objective constraints. Compared to existing approaches that use a single model to adapt to all scenarios, this invention achieves accurate coverage of lightweight, dynamic, and complex scenarios, significantly improving the applicability and optimization effect of coal blending schemes under different scenarios. Attached Figure Description

[0018] Figure 1 This is a diagram illustrating the steps and methods of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] First embodiment: Please see Figure 1 This application provides an AI-based intelligent coal blending method based on deep learning, which specifically includes the following steps: Step 1: Collect basic coal quality data, specifically including industrial analysis indicators such as moisture (M), ash (A), volatile matter (V), and fixed carbon (FC); elemental analysis indicators such as carbon (C), hydrogen (H), oxygen (O), nitrogen (N), and sulfur (S); and process characteristic indicators such as caking index (G), plastic layer thickness (Y), ash fusion point (ST), and calorific value (Qnet, v, ar). Data collection is conducted using a combination of offline testing and online monitoring. Offline testing follows national standard methods, while online monitoring utilizes near-infrared spectrometers and laser particle size analyzers to achieve second-level data updates, using these as core input features. Simultaneously, production process data and target / constraint data are collected. Specific process data includes equipment operating conditions such as coking oven temperature and gasification... Furnace pressure; historical coal blending records and actual production results, such as coke strength M40 / CSR, boiler thermal efficiency, and gasifier carbon conversion rate, are captured in real time through Manufacturing Execution System (MES) and Distributed Control System (DCS). The data sampling frequency is usually 1-5 minutes / time. The target and constraint data include target indicators and constraints. The target indicators include cost, quality, and environmental dimensions, while the constraints include coal availability, equipment limitations, and process limitations. The sources of the target and constraint data include indicators issued by the production planning department, the supply chain inventory system, and equipment parameter manuals. At the same time, the production process data and target and constraint data are preprocessed, including handling missing values ​​and verifying rationality. Missing values ​​are handled using linear interpolation. The target features are sorted according to the acquisition time to obtain an ordered sequence X0=[x1, x2, ..., x... i-1 NaN, x i+1 , ..., x n Next, locate the missing point and the position i of the missing value, and find its preceding known value x. i-1 and the next known value x i+1 The two times correspond to t-1 and t+1, respectively. Missing values ​​are estimated using a linear formula. ; Reasonableness verification is performed using the 3σ principle. The target features are subjected to Shapiro-Wilk test or QQ plot analysis to verify whether they conform to a normal distribution. Here, it is assumed that they conform to a normal distribution. Then, the mean μ and standard deviation σ of the features are calculated, and anomaly thresholds are set: Lower = μ - 3σ, Upper = μ + 3σ. Data values ​​less than Lower or greater than Upper are marked as outliers and removed after confirmation with the original records. If it is a false alarm by the equipment, it can be replaced with normal data from the adjacent time.

[0021] Step 2: Obtain the core input features and classify them into relevant and irrelevant features, with the specific classification method as follows: Plot the scatter distribution of all samples with feature A as the x-axis and feature B as the y-axis, and observe the aggregation trend of the points. If the scatter points are roughly distributed along a straight line, they tend to be linearly correlated. If the scatter points show a curved trend or a non-linear regular distribution, such as U-shaped, S-shaped, or logarithmic, they tend to be non-linearly correlated. Next, the presence of nonlinear correlation is verified by residual plot. A scatter plot is drawn with feature A as the x-axis and residual as the y-axis. If the residual is randomly distributed on both sides of y=0, it is linearly correlated. If the residual shows a systematic trend, such as the residual increasing with A as a positive, negative and positive curve, it is nonlinearly correlated. For relevance features, the Pearson coefficient is used to remove redundant features, and for non-relevance features, the mutual information method is used to remove redundant features. For the analysis of the Pearson coefficient, all continuous features are obtained and standardized, denoted as X, where X = 1, 2, ..., Y, and Y represents the type of continuous feature, such as ash content A, volatile matter V, and calorific value Q. A Pearson correlation coefficient matrix R is then constructed between the features. , where r XY Let x represent the Pearson correlation coefficient between feature X and feature Y, with a value range of [-1, 1], where b is the sample size, such as the number of coal batches or historical coal blending schemes in coal blending data. a Let represent the observed value of feature X in the a-th sample, such as the ash content and volatile matter of the a-th batch of coal; let ya represent the observed value of feature Y in the a-th sample, such as the calorific value and coke strength M40 of the a-th batch of coal. The sample mean of feature X, such as the average ash content of all batches of coal, r Y This represents the sample mean of feature Y, such as the average calorific value of all batches of coal. Next, the calculated Pearson correlation coefficient is compared with a high correlation threshold. The specific value of the high correlation threshold is set according to business requirements, typically set to 0.8. Feature pairs with correlation coefficients exceeding the threshold are filtered out. This process is repeated to obtain all highly correlated feature pairs, and the feature with the highest correlation to the target indicator is retained. The analysis using the mutual information method involves obtaining continuous features and target features, then discretizing these features and targets into discrete variables. Next, the mutual information value between the continuous feature X and the target feature Y is calculated using the formula... The mutual information value between continuous feature X and target feature Y is calculated, where P(X,Y) is the joint probability, P(X) is the marginal probability of X taking x, and P(Y) is the marginal probability of Y taking y. Then, the mutual information values ​​MI are sorted from largest to smallest, and the top N features are selected as screening features. For the screening features, the mutual information values ​​between each pair are calculated again, and redundant features with mutual information values ​​MI exceeding the threshold are removed to obtain retained features. At the same time, the retained features are transformed into the interval [1,1]. Then, the retained features are combined to generate composite features.

[0022] Step 3: Select different model architectures based on different scenarios. For single-objective optimization, a basic regression model is selected, mainly using a multilayer perceptron. This model learns simple nonlinear relationships through an input layer, 3-5 hidden layers, and an output layer. Specifically, the number of neurons in the input layer corresponds to the feature dimension; for example, 8-15 coal quality features correspond to 8-15 input neurons. The hidden layer has 3-5 layers, with the number of neurons in each layer decreasing by 2 / 3 of the input layer's number. For example, with a 10-dimensional input, the first layer has 6-8 neurons, the second layer has 4-5 neurons, avoiding overfitting. The output layer has 1 neuron. For scenarios with large fluctuations in coal quality, a time-series model is selected, using LSTM / GRU to capture coal quality data. The algorithm adapts to time-series changes and outputs a dynamic blending scheme. The input is a sliding window of time-series features, such as coal quality data from the previous 6 hours plus current inventory. The window length is set according to the fluctuation cycle, typically 4-12 hours. There are 2-3 hidden layers, each with 64-128 neurons. The output layer is the dynamic blending scheme for the next 1-3 hours. For high-value-added scenarios such as coking and coal chemical industries, multimodal or complex models are selected, including the Transformer model and hybrid intelligent models, which combine CNN+MLP+linear programming and deep learning+traditional algorithms. Then, the dataset is divided into training, validation, and test sets in a 7:2:1 ratio according to the time sequence. Next, different loss functions are selected based on the number of objectives. For the single-objective case, mean squared error is selected to calculate the bias. For the multi-objective case, weighted loss function and other algorithms are selected to generate Pareto optimal solutions. The learning rate, number of hidden layer neurons, and number of iterations are adjusted through grid, random, or Bayesian optimization, with the learning rate being 0.001-0.01 and the number of iterations being 1000-5000 rounds to ensure model convergence. The accuracy is tested using a test set. If the accuracy is not met, the process returns to the data preprocessing stage for re-optimization.

[0023] Step 4: Real-time data is collected using an online near-infrared spectrometer (for detecting ash, volatile matter, and moisture, with an error ≤0.5%) and an X-ray fluorescence spectrometer (for detecting sulfur and phosphorus, with an error ≤0.02%). Instantaneous values ​​are generated every 10 seconds, and the data is aggregated into a 30-second moving average every 30 seconds. Constraints related to inventory, procurement, and environmental protection are loaded from the ERP system and converted into mathematical conditions that the model can recognize. When constraints conflict, a priority mechanism is triggered: Level 1 priority: Environmental and safety constraints (must be met; for example, if SO2 emissions exceed the limit, the scheme will be directly eliminated); Level 2 priority: Equipment constraints (e.g., the coal mill load cannot be exceeded, otherwise the mix ratio will be adjusted); Level 3 priority: Flexible constraints such as cost and inventory (optimization can be performed after Levels 1 and 2 are met). The real-time coal quality characteristics and constraints are input into the trained model. The model uses the learned nonlinear relationships to calculate the optimal proportion of each coal type and simultaneously predicts the target indicators. In response to the dynamic changes in target priority, the model generates 3-5 candidate schemes by adjusting the weights of the loss function. Each candidate scheme includes the proportion of different coal types and the corresponding predicted values ​​of target indicators, such as the lowest cost scheme, the best quality scheme, and the environmental compliance scheme. Then, the candidate schemes are pushed to the production decision-making level. The decision-making level selects or adjusts the final coal blending scheme from the candidate schemes based on actual production needs, market dynamics, and strategic objectives. If the decision-making level is not satisfied with any of the candidate schemes, a feedback mechanism can be triggered to require the model to be retrained or its parameters adjusted to generate new candidate schemes that better meet the requirements.

[0024] Second embodiment: As a second embodiment of the present invention, it is implemented based on the first embodiment, and the difference from the first embodiment is as follows: Missing values ​​were handled using box plots. The data was divided into distribution intervals based on the quartiles, and data exceeding 1.5 times the interquartile range were identified as outliers, unaffected by extreme values. The target features were sorted according to the collection time, resulting in an ordered sequence X0 = [x1, x2, ..., x...]. i-1 NaN, x i+1 , ..., x n Simultaneously calculate the lower quartile Q1 (25th quartile), median Q2 (50th quartile), and upper quartile Q3 (75th quartile), then calculate the interquartile range IQR = Q3 - Q1; Next, the anomaly threshold is determined based on the calculated interquartile range, with the lower boundary being Q1 - 1.5 × IQR and the upper boundary being Q3 + 1.5 × IQR. Data exceeding the boundaries are marked as outliers. If it is an isolated case, it is directly removed; if it is a continuous anomaly, it is necessary to check whether the equipment is faulty.

[0025] Third embodiment: As a third embodiment of the present invention, the focus is on implementing the first and second embodiments in combination.

[0026] The data in the above formulas are all calculated using numerical values, without substituting the units of the parameters. In addition, the contents not described in detail in this specification are all prior art known to those skilled in the art.

[0027] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A deep learning-based AI-powered intelligent coal blending method, characterized in that, The method specifically includes the following steps: Collect basic coal quality data, production process data, and target and constraint data. Preprocess the production process data and target and constraint data, including handling missing values ​​and verifying their reasonableness. The core input features are classified into relevant features and irrelevant features. The Pearson coefficient is used to remove redundant features of the relevant features, and the mutual information method is used to remove redundant features of the irrelevant features. The retained features are combined to generate composite features. Different model architectures are selected for different scenarios, the dataset is divided and a loss function is selected based on the number of targets. Hyperparameters are adjusted by optimizing the algorithm, and the model accuracy is verified using the test set. If the model does not meet the target, the process is returned to step one and optimized again. Collect real-time coal quality data and load constraints, input the trained model to generate candidate solutions, push them to the decision-making level for selection, and trigger model retraining or parameter adjustment based on feedback from the decision-making level.

2. The AI-based intelligent coal blending method based on deep learning according to claim 1, characterized in that, Basic coal quality data includes industrial analysis indicators, elemental analysis indicators, and process characteristic indicators, which are collected using a combination of offline testing and online monitoring. Production process data includes equipment operating conditions, historical coal blending records, and actual production results, which are captured in real time through the manufacturing execution system and distributed control system. The target and constraint data includes target indicators and constraints. The target indicators cover cost, quality, and environmental dimensions, while the constraints include coal availability, equipment limitations, and process limitations.

3. The AI-based intelligent coal blending method based on deep learning according to claim 1, characterized in that, Missing values ​​are handled using linear interpolation: After sorting the target features by collection time to obtain an ordered sequence, the missing points are located. The missing values ​​are estimated using a linear formula based on the known values ​​before and after the missing points and the corresponding times. The rationality check adopts the 3σ principle. After verifying that the target feature is normally distributed, the mean μ and standard deviation σ are calculated, Lower = μ - 3σ, Upper = μ + 3σ. Outliers exceeding the threshold are marked and processed.

4. The AI-based intelligent coal blending method based on deep learning according to claim 1, characterized in that, The classification method for features into relevant and irrelevant features is as follows: Plot a scatter plot with feature A as the x-axis and feature B as the y-axis. If the scatter points are distributed along a straight line, they tend to be linearly correlated; if they are distributed in a curve or a non-linear pattern, they tend to be non-linearly correlated. Next, the residual plot is used for verification. A scatter plot is drawn with feature A as the x-axis and the residual as the y-axis. If the residuals are randomly distributed on both sides of y=0, it is a linear correlation. If the residuals show a systematic trend, it is a non-linear correlation.

5. The AI-based intelligent coal blending method based on deep learning according to claim 1, characterized in that, When removing redundant features for the Pearson coefficient, all continuous features are obtained and standardized, a Pearson correlation coefficient matrix between features is constructed, the correlation coefficient is calculated and compared with a high correlation threshold, feature pairs with correlation coefficients exceeding the threshold are selected, and the features with the highest correlation to the target indicator are retained. The high correlation threshold is set by business requirements and is set to 0.

8.

6. The AI-powered intelligent coal blending method based on deep learning according to claim 1, characterized in that, When performing redundant feature removal using the mutual information method, the continuous features and target features are discretized, the mutual information value between the continuous features and the target features is calculated, the top N features are selected as screening features by sorting the mutual information values ​​from largest to smallest, the mutual information values ​​between each pair of screening features are calculated again, redundant features with mutual information values ​​exceeding the threshold are removed to obtain retained features, and the retained features are normalized to the [0,1] interval.

7. The AI-based intelligent coal blending method based on deep learning according to claim 1, characterized in that, The specific architecture of the scenario-based selection model is as follows: For single-objective optimization, a basic regression model based on multilayer perceptron is selected. The number of neurons in the input layer is consistent with the feature dimension. There are 3-5 hidden layers with the number of neurons in each layer decreasing by 2 / 3 of the number in the input layer. The output layer has 1 neuron. For scenarios with large fluctuations in coal quality, we select LSTM or GRU time series models. The input is a sliding window time series feature, with 2-3 hidden layers and 64-128 neurons in each layer. The output is the dynamic ratio for the next 1-3 hours. For high-value-added scenarios such as coking and coal chemical industry, the Transformer model or hybrid intelligent model is selected. The hybrid intelligent model includes models that combine CNN+MLP+linear programming, deep learning and traditional algorithms.

8. The AI-based intelligent coal blending method based on deep learning according to claim 1, characterized in that, The dataset was divided into training, validation, and test sets in a 7:2:1 chronological order, and a loss function was selected based on the number of targets. In the case of a single objective, the mean squared error is selected to calculate the deviation; in the case of multiple objectives, a weighted loss function is selected and combined with the algorithm to generate the Pareto optimal solution.

9. The AI-based intelligent coal blending method based on deep learning according to claim 5, characterized in that, The method for constructing the Pearson correlation coefficient matrix between features is as follows: Obtain all continuous features and standardize them, denoted as X, where X = 1, 2, ..., Y, and Y represents the type of continuous feature. Construct the Pearson correlation coefficient matrix R between the features. , where r XY Let represent the Pearson correlation coefficient between feature X and feature Y, and its value ranges from [-1, 1], where b is the sample size and x is the number of samples. a Let represent the observed value of feature X in the a-th sample, and ya represent the observed value of feature Y in the a-th sample. r represents the sample mean of feature X. Y This represents the sample mean of feature Y.

10. The AI-based intelligent coal blending method based on deep learning according to claim 6, characterized in that, The method for calculating the mutual information value between continuous features and target features is as follows: Obtain continuous features and target features, and discretize the continuous features and target indicators to convert them into discrete variables. Then, calculate the mutual information value between the continuous feature X and the target feature Y according to the formula. The mutual information values ​​of continuous feature X and target feature Y are calculated, where P(X,Y) is the joint probability, P(X) is the marginal probability of X taking the value x, and P(Y) is the marginal probability of Y taking the value y.

Images