Alkali ore proportioning method and system based on big data

Abstract

The application provides a big data-based alkali ore proportioning method and system, and relates to the technical field of data processing.The method comprises the following steps: constructing original sample data containing historical alkali ore proportioning data and corresponding performance index data, performing data cleaning and normalization processing on the original sample data to obtain training data; constructing a neural network model, training the neural network model by using the training data, optimizing network weights by using a back propagation algorithm, and obtaining the trained neural network model until a preset stop condition is met; obtaining a target performance index, inputting the target performance index into the trained neural network model, obtaining alkali ore proportioning data, and performing proportioning production according to the alkali ore proportioning data.The application can improve the accuracy of alkali ore proportioning.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for alkali ore proportioning based on big data. Background Technology

[0002] Alkali mineral materials, as a novel type of green cementitious material, have broad application prospects in the field of civil engineering due to their excellent durability and environmentally friendly properties. Their performance is highly dependent on the ratio of alkali activator to mineral raw materials, traditionally relying on empirical formulas and extensive trial mixing experiments, resulting in long development cycles and high costs. With the development of materials informatics and big data technology, collecting massive amounts of historical mixing and performance data, and using data mining and machine learning methods to establish a complex nonlinear mapping relationship between mixing ratio and performance, has become an important direction for improving the design efficiency of alkali mineral materials and achieving precise performance control.

[0003] Related technologies employ statistical regression analysis to predict and optimize mix proportions by establishing linear or polynomial regression equations between proportioning parameters and key performance characteristics (such as strength). However, regression models have limited ability to fit complex relationships and cannot efficiently and accurately handle the complex relationships of multiple components, strong coupling, and nonlinearity in alkali-ore mix proportions. Consequently, the recommended mix proportions lack accuracy and adaptability when dealing with practical engineering projects that require multiple objectives and high performance, making it difficult to achieve accurate reverse design from performance targets to the optimal mix proportion. Summary of the Invention

[0004] In order to provide reliable proportioning guidance for alkali ore production, this application provides an alkali ore proportioning method and system based on big data.

[0005] Firstly, this application provides a method for alkali-ore proportioning based on big data, employing the following technical solution: A big data-based method for alkali ore proportioning includes the following steps: Original sample data containing historical alkali ore ratio data and corresponding performance index data was constructed. The original sample data was cleaned and normalized to obtain training data. Construct a neural network model, train the neural network model using training data, optimize the network weights using the backpropagation algorithm until the preset stopping condition is met, and obtain the trained neural network model. The target performance index is obtained and input into the trained neural network model to obtain alkali-ore ratio data. Production is carried out according to the alkali-ore ratio data.

[0006] This application first constructs original sample data containing historical alkali ore ratio data and corresponding performance index data, and then performs data cleaning and normalization processing on it, which helps the neural network model learn real and effective rules, avoids model bias caused by erroneous data as much as possible, and improves the accuracy and generalization ability of the model.

[0007] Neural network models possess powerful nonlinear mapping capabilities, enabling them to learn complex nonlinear relationships between alkali ore ratios and performance indicators, thereby improving prediction accuracy. These models can automatically adjust network weights during training to adapt to different data distributions and feature relationships. The models can continuously optimize themselves based on actual production data; as data accumulates and is updated, the model's performance continuously improves, better enabling it to cope with changes and uncertainties in the production process. This application uses a backpropagation algorithm to continuously optimize network weights, allowing the neural network model to gradually fit the patterns in the training data. Through multiple iterations of training, the model can continuously adjust its parameters, gradually reducing the error between the predicted results and actual performance indicators, thus improving the model's prediction accuracy.

[0008] Based on the target performance indicators, the trained neural network model can quickly and accurately obtain the corresponding alkali-ore ratio data, providing production personnel with scientific and precise ratio guidance. This avoids the blindness and uncertainty of relying on experience and trial-and-error methods in traditional production, thereby improving production efficiency and product quality.

[0009] Optionally, the neural network model is a hybrid prediction model, including a shared feature extraction layer, a long short-term memory network branch based on an attention mechanism, and a fully connected feedforward network branch; the outputs of the long short-term memory network branch and the fully connected feedforward network branch are concatenated in the fusion layer to obtain a concatenated vector, and the concatenated vector is passed through the output layer to obtain the alkali ore ratio data.

[0010] Sharing a feature extraction layer allows the Long Short-Term Memory (LSTM) network branch and the fully connected feedforward network branch to share some parameters to extract common features from the input data. This avoids designing separate feature extraction modules for each branch, reduces the overall number of parameters in the model, lowers computational complexity, and accelerates training and inference. At the same time, reducing the number of parameters also helps prevent overfitting, improves the model's generalization ability, and enables it to perform well on new data.

[0011] Long Short-Term Memory (LSTM) networks inherently possess the ability to process sequential data, capturing long-term dependencies within the data. In alkali-ore ratio prediction, historical data often exhibits time-series characteristics. LSTM branches can effectively learn patterns within this time-series data, understanding the impact of past data on the current alkali-ore ratio, thus providing more accurate historical information for prediction. The introduction of an attention mechanism allows LSTM branches to automatically focus on the most relevant parts of the input sequence to the prediction target, thereby improving the model's efficiency in utilizing important information and further enhancing prediction accuracy.

[0012] Fully connected feedforward networks excel at handling non-sequential feature data. In alkali ore proportion prediction, in addition to time series data, there may also be some static features, such as the chemical composition of raw materials and the initial parameters of equipment. The branches of the fully connected feedforward network can independently analyze and process these non-sequential features, extracting their effective information. This complements the time series features processed by the LSTM branch, enabling the model to comprehensively consider multiple types of data and improve the comprehensiveness of the prediction.

[0013] The fusion layer concatenates the outputs of the LSTM branch and the fully connected feedforward network branch to obtain a concatenated vector, enabling the model to comprehensively utilize time series features and non-sequence features to predict the alkali ore ratio from multiple dimensions.

[0014] This application leverages the time-series processing capabilities of the LSTM branch, the key information focusing capabilities of the attention mechanism, and the non-sequence feature processing capabilities of the fully connected feedforward network branch. This enables the model to more comprehensively and accurately capture the relationship between input data and alkali ore ratios. The fusion layer design allows for the full integration of different types of information, thereby improving the accuracy of model predictions and providing more reliable ratio guidance for alkali ore production.

[0015] Optionally, after obtaining the trained neural network model, the method further includes: The trained neural network model is used as a proxy model for the target performance. An optimization objective function including cost and resource consumption is defined. A Bayesian optimization algorithm is used to iteratively search in the continuous ratio parameter space until a Pareto optimal alkali ore ratio data set is found.

[0016] Optionally, the Bayesian optimization algorithm uses expected improvement as the acquisition function, and during the iteration process, it dynamically balances the exploration and utilization of the proportioning space based on the prediction uncertainty provided by the Gaussian process model.

[0017] In actual production, cost and resource consumption are key factors affecting a company's economic benefits and sustainable development. This application, by comprehensively considering these two objectives, can find a balanced ratio scheme between cost control and rational resource utilization, providing a more comprehensive and scientific basis for production decisions. Finding a Pareto optimal set of alkali-ore ratio data means that without improving any one objective (such as reducing costs or resource consumption), other objectives cannot be weakened. This provides decision-makers with a set of ratio schemes that achieve the best trade-off between different objectives. Decision-makers can select the most suitable ratio from this set of schemes based on actual production needs, market environment, and corporate strategy, thereby making better decisions.

[0018] Bayesian optimization is a probabilistic model-based optimization algorithm. It describes the uncertainty of the objective function by constructing a probabilistic model of the objective function (such as a Gaussian process model) and uses a sampling function (such as desired improvement) to guide the search process. Compared with traditional grid search or random search methods, Bayesian optimization can more effectively search within the matching parameter space, reducing the exploration of a large number of invalid regions, thereby improving search efficiency and reducing the consumption of computational resources and time.

[0019] This application employs expected improvement as the acquisition function and dynamically balances the exploration and utilization of the matching space based on the predictive uncertainty provided by the Gaussian process model during the iteration process. In the initial search phase, the algorithm explores unknown regions more extensively to discover potentially better solutions; as the search progresses and a certain understanding of the objective function is gained, the algorithm gradually shifts to utilizing known information to conduct a refined search in potentially optimal regions. This dynamic balancing mechanism enables the algorithm to achieve a good balance between global search and local optimization, further improving search efficiency.

[0020] Bayesian optimization algorithms can find the Pareto optimal alkali-ore ratio dataset, i.e., the Pareto front, in a relatively small number of iterations. The Pareto front contains all ratio schemes that cannot be further improved under a given objective. By quickly locating the Pareto front, decision-makers can understand the trade-offs between different objectives in a timely manner, providing important reference information for production decisions, while also saving a significant amount of experimentation and computation time.

[0021] Optionally, during the data cleaning and normalization process of the original sample data, an unsupervised learning method based on isolated forest is used to detect anomalies in the original sample data, automatically identifying and removing historical abnormal batch data that deviates from the main process mode.

[0022] As an unsupervised learning algorithm, Isolation Forest does not require pre-labeling of anomalous data. It can automatically learn the distribution characteristics of data from the original sample data, identify historical anomalous batch data that deviates from the main process mode, and improve the purity of the data.

[0023] Optionally, the neural network model is trained and inferred using Monte Carlo Dropout technology, and when outputting the alkali ore ratio data, the predicted mean and confidence interval of each performance index data are output simultaneously.

[0024] Monte Carlo Dropout randomly discards some neurons during training and inference, essentially introducing randomness into the model. Through multiple sampling and computations, multiple predicted values ​​of performance metrics can be obtained, and then the predicted mean and confidence interval can be calculated. This allows the model to simulate different possible scenarios, revealing the inherent uncertainty and variability of the model's predictions. This enables decision-makers to better understand how the model arrives at its predictions, rather than just a single predicted value.

[0025] In actual production, due to various factors such as fluctuations in raw material quality and unstable equipment operation, the production process itself inherently involves a certain degree of uncertainty. The confidence interval of the model output can reflect this uncertainty in actual production, making the model's predictions more consistent with reality and enhancing the model's interpretability and credibility.

[0026] Optionally, the method further includes: Clustering algorithms are used to process the training data to obtain clustering results. The clustering results are mapped to gray values ​​and gray-level histograms are drawn. The optimal threshold is obtained based on the gray-level histogram using the big rule method. The gray-level histogram is then binarized according to the optimal threshold, and the training data is divided into data subsets of different process feature intervals.

[0027] Optionally, the method further includes: formulating a differentiated alkali-ore ratio production process control strategy based on the significant interval of process characteristics obtained by the optimal threshold division, and deploying the control strategy to the DCS system.

[0028] This application employs a clustering algorithm to process the training data, grouping data points with similar characteristics together to form different clusters. This helps reveal the hidden internal structure and patterns within the training data, and uncover potential relationships between data points. Subsequently, the clustering results are mapped to grayscale values ​​and a grayscale histogram is plotted, providing an intuitive graphical representation of the data distribution. The horizontal axis of the grayscale histogram represents the grayscale value (corresponding to different process characteristics), and the vertical axis represents the frequency of that grayscale value. By observing the grayscale histogram, the distribution density of the data across different process characteristic ranges can be quickly understood, providing an intuitive basis for further analysis.

[0029] Subsequently, this application employs the Otsu's algorithm based on grayscale histograms to obtain the optimal threshold. The Otsu's algorithm determines the optimal threshold by maximizing the inter-class variance, thus dividing the data into two categories with the greatest difference. In alkali ore production data, the Otsu's algorithm can accurately find the optimal threshold for dividing different process characteristic intervals, reducing the subjectivity and uncertainty of manual threshold selection and improving the accuracy and reliability of data segmentation.

[0030] This application performs binarization processing on the grayscale histogram based on an optimal threshold, converting continuous grayscale values ​​into binary 0s and 1s, reducing data complexity. Through binarization, the training data can be divided into subsets of data with different process characteristic intervals. Subsequently, based on the significant process characteristic intervals obtained from the optimal threshold division, this application formulates differentiated alkali-ore ratio production process control strategies. Different process characteristic intervals may correspond to different production conditions and requirements; by formulating differentiated strategies, these different conditions can be better adapted.

[0031] Differentiated control strategies can be tailored to different process characteristic ranges to respond promptly to changes and fluctuations in the production process, thereby improving production stability.

[0032] Optionally, the method further includes: Determine whether the training samples in the training data are balanced. If so, generate adversarial samples and add them to the training data. If not, use a generative adversarial network to generate new training samples and add them to the training data.

[0033] In many real-world scenarios, training data often suffers from imbalance, meaning that some classes have significantly more samples than others. This imbalance can cause the model to overemphasize the majority class during learning, neglecting the minority class, thus reducing the model's ability to identify the minority class and impacting its generalization performance. This application addresses this issue by assessing the balance of training samples and taking appropriate measures to increase the number of samples. This effectively balances the data distribution, allowing the model to more comprehensively consider the characteristics of each class during learning, improving the model's prediction accuracy for different classes, and ultimately enhancing its generalization ability.

[0034] In the process of handling imbalanced data, this application utilizes existing training data for both generating adversarial examples and using generative adversarial networks to generate new samples. This means it can fully tap the potential of existing data, generating more valuable samples from limited data by adding small perturbations or leveraging the learning capabilities of generative adversarial networks. This avoids the problem of model performance degradation due to insufficient data and improves data utilization efficiency.

[0035] Secondly, this application provides an alkali ore proportioning system based on big data, employing the following technical solution: A big data-based alkali ore proportioning system includes: a processor and a memory communicatively connected to the processor; The memory is provided with a computer-readable storage medium, and a computer program is stored on the computer-readable storage medium. When the processor processes a computer program stored on the computer-readable storage medium, it implements the method as described in the first aspect.

[0036] In summary, this application includes at least one of the following beneficial technical effects: 1. This application first constructs original sample data containing historical alkali ore ratio data and corresponding performance index data, and performs data cleaning and normalization on the data, which helps the neural network model learn real and effective rules, avoids model bias caused by erroneous data as much as possible, and improves the accuracy and generalization ability of the model.

[0037] 2. Based on the target performance indicators, the trained neural network model can quickly and accurately obtain the corresponding alkali-ore ratio data, providing scientific and precise ratio guidance for production personnel. This avoids the blindness and uncertainty of relying on experience and trial-and-error methods in traditional production, thereby improving production efficiency and product quality. Attached Figure Description

[0038] Figure 1 This is a flowchart of an alkali-ore proportioning method based on big data, as described in this application. Detailed Implementation

[0039] The following combination Figure 1 This application will be described in further detail.

[0040] Example 1: This example discloses a method for alkali ore proportioning based on big data, referring to... Figure 1 The method includes: S11 acquiring training data, S12 modeling and training, and S13 acquiring the matching scheme. The execution process of each step in this embodiment is as follows: S11 acquires training data by extracting alkali-ore ratio data from the enterprise's production database over the past five years. This includes the ratio of various alkali-ore raw materials (such as soda ash, caustic soda, and limestone) and corresponding performance indicators, such as product strength, purity, and reactivity, to construct the original sample dataset. The original sample data is stored in tabular form, with each row representing a data record for a production batch, including ratio parameters, performance indicator parameters, cost, and resource consumption.

[0041] An unsupervised learning method based on isolated forests is used to detect anomalies in the original sample data. The isolated forest algorithm constructs isolated trees by randomly selecting features and then randomly selecting split values ​​on these features. Anomalous data can usually be isolated with fewer splits. In this embodiment, by setting the number of isolated trees (e.g., 100 trees) and the anomaly score threshold (e.g., 0.6), historical anomalous batch data that deviates from the main process mode is automatically identified and removed.

[0042] For example, if the proportion of a certain type of alkali ore in a batch of data is much higher than the normal range, and the corresponding performance indicators are also abnormal, the isolated forest algorithm can identify it as abnormal data and remove it.

[0043] The cleaned raw sample data is processed using a min-max normalization method to map the data to the [0,1] interval. After normalization, the training data is obtained to eliminate the influence of different feature parameters on their dimensions, thereby improving the training efficiency and accuracy of the neural network model.

[0044] In other embodiments, this step further includes the following steps: The number of samples in different performance index ranges in the training data is counted, and the proportion of each type of sample is calculated. If the proportions of each type of sample differ significantly (e.g., the number of samples in one type accounts for more than 65% of the total number of samples), then the training samples are considered imbalanced.

[0045] If the training samples are balanced, then the Fast Gradient Sign Method (FGSM) is used to generate adversarial examples. The computational model for adversarial examples is as follows:

[0046] in, For adversarial examples, For training data; It is a symbolic function; This represents the gradient of the loss function of the neural network model with respect to the input data.

[0047] FGSM adds small perturbations to the original samples, causing the model to make incorrect predictions about adversarial examples. The generated adversarial examples are then added to the training data at a rate of 8% to enhance the model's robustness.

[0048] If the training samples are imbalanced, a Generative Adversarial Network (GAN) is used to generate new training samples. A GAN consists of a generator and a discriminator. The generator produces fake samples that resemble real samples, while the discriminator determines whether a sample is real or fake. By iteratively training the generator and discriminator, the generator can produce high-quality new samples. These new samples are then added to the training data, adjusting the proportions of different types of samples to achieve a balanced training data.

[0049] S12 Modeling and Training: Constructing a neural network model. In this embodiment, the neural network model adopts a hybrid prediction model, which includes a shared feature extraction layer, a Long Short-Term Memory (LSTM) branch based on an attention mechanism, and a fully connected feedforward (MLP) branch.

[0050] The shared feature extraction layer employs a convolutional neural network (CNN) structure, comprising two convolutional layers and one pooling layer. The convolutional layers use 3×3 kernels with a stride of 1 and the ReLU activation function. The pooling layer uses max pooling with a 2×2 pooling window and a stride of 2. This shared feature extraction layer is used to extract general feature information from the input training data.

[0051] In the attention-based LSTM branch, the number of LSTM units is set to 64 to process data with temporal characteristics. This embodiment also introduces an attention mechanism after the LSTM layer. This mechanism calculates the weights for each time step and performs a weighted summation of the features output by the LSTM to highlight the features of important time steps.

[0052] The fully connected feedforward network branch contains two fully connected layers. The first fully connected layer has 128 neurons, and the second fully connected layer has 64 neurons. The activation function for both layers is the ReLU function.

[0053] The outputs of the Long Short-Term Memory (LSTM) network branch and the fully connected feedforward network branch are concatenated in the fusion layer to obtain the concatenated vector.

[0054] The concatenated vector is passed through the output layer and used to predict the alkali ore ratio data.

[0055] The training data was divided into training and validation sets in a 7:3 ratio, and the neural network model was trained using the stochastic gradient descent (SGD) optimization algorithm, with a learning rate of 0.0015 and a batch size of 32.

[0056] During training, this embodiment uses the backpropagation algorithm to optimize network weights. The model parameters are updated by calculating the gradient of the loss function (mean squared error loss function in this embodiment) with respect to the model parameters and then moving in the opposite direction of the gradient.

[0057] In this embodiment, the neural network model uses Monte Carlo Dropout technology for training and inference. During training, some neurons are randomly dropped with a probability of 0.4 to prevent overfitting. During inference, Dropout is kept on, and through 80 independent forward propagations, the mean and 95% confidence interval of each prediction index are output.

[0058] After each training iteration, the model's prediction error is evaluated on the validation set. Training stops when the model's prediction error on the validation set no longer decreases or falls below a set threshold (e.g., 0.06) for five consecutive iterations, resulting in the trained neural network model.

[0059] S13 obtains the formulation plan, acquires the target performance indicators set by the enterprise, and clarifies the performance constraints (such as product strength ≥30MPa, purity ≥98%, reactivity ≥85%). The target performance indicators are then input into the trained neural network model. The trained neural network model calculates through forward propagation and outputs alkali ore formulation data and corresponding cost and resource consumption predictions.

[0060] Using the trained neural network model as a surrogate model for the target performance, an optimization objective function is defined that includes cost, resource consumption, and performance constraints. The calculation model is as follows:

[0061]

[0062] in, The objective function is... Calculate to obtain the minimum value; and These are the weighting coefficients for cost and the consumption of the i-th resource, respectively. For cost; Let i be the consumption of the i-th type of resource.

[0063] A Bayesian optimization algorithm is used to iteratively search for the optimal solution in a continuous matching parameter space. The Bayesian optimization algorithm uses the expected improvement as the acquisition function, and during the iteration process, it dynamically balances the exploration and utilization of the matching space based on the prediction uncertainty provided by the Gaussian process model. The specific steps are as follows: Initialization: Within the ratio parameter space, 10 sets of alkali-ore ratio parameters that meet the constraints are randomly generated as initial sample points. The objective function value corresponding to each sample point is calculated through the surrogate model, and sample points that meet the performance constraints are selected, such as product strength ≥30MPa, purity ≥98%, and reactivity ≥85%.

[0064] Constructing the Gaussian process model: Based on the initial sample points and their corresponding objective function values, a Gaussian process model is constructed. The covariance function of the Gaussian process model adopts the squared exponential kernel function. :

[0065] in, It is the signal variance, used to control the amplitude of the function value; This is the length scale parameter.

[0066] For the new proportion parameter vector Its corresponding objective function value The predicted distribution is shown below:

[0067] Where m is the initial sample point mean vector; It is the mean of the new sample points; It is the covariance matrix between the initial sample points, and its elements are... ; It is the covariance matrix between the initial sample points and the new sample points, and its elements are... ;Through the properties of Gaussian processes, we can obtain The predicted mean and prediction variance .

[0068]

[0069]

[0070] The noise variance represents the noise in the observations.

[0071] Selecting the next sample point: Based on the expected improvement of the acquisition function and the prediction of the Gaussian process model, select the next sample point with the greatest potential, i.e., the combination of parameters expected to bring the greatest improvement, for the currently known optimal objective function value y. best The desired improvement in EI(x) is defined as:

[0072] The probability density function is the standard normal distribution. It is the cumulative distribution function of the standard normal distribution.

[0073] Evaluation of sample points: Substitute the selected sample points into the actual production environment or evaluate them through a more accurate simulation model to obtain the corresponding optimization objective function value.

[0074] Update the Gaussian process model: Add the newly evaluated sample points and their optimized objective function values ​​to the training data to update the Gaussian process model.

[0075] Iterative search: Repeat the above steps until the preset number of iterations (e.g., 80 times) is reached or the stopping condition is met (e.g., the objective function value is improved by less than 0.8 in 5 consecutive iterations), and output the set of alkali-ore ratio data that satisfies Pareto optimality.

[0076] Example 2: The difference between this example and Example 1 is that the method further includes: Processing the training data using the K-means clustering algorithm. First, determine the number of clusters K as 4 through the elbow method, and then divide the training data into 4 cluster clusters to obtain the clustering result.

[0077] Map the clustering result to a grayscale value. For example, map the eigenvalue corresponding to the center point of each cluster to a different grayscale value, and draw a grayscale histogram. In this example, the eigenvalues of the 4 cluster centers are C1 < C2 < C3 < C4, and they are mapped to grayscale values G1 = 50, G2 = 100, G3 = 150, G4 = 200.

[0078] Obtain the optimal threshold based on the grayscale histogram using the Otsu algorithm. The Otsu algorithm determines the optimal threshold by maximizing the between-class variance, divides the grayscale histogram into two parts to maximize the difference between the two parts, and perform binary processing on the grayscale histogram according to the optimal threshold to divide the training data into data subsets in different process feature intervals. For example, divide the data subset with a grayscale value greater than the optimal threshold into a high purity process feature interval, and divide the data subset with a grayscale value less than the optimal threshold into a low purity process feature interval.

[0079] Based on the process feature significant interval obtained by dividing according to the optimal threshold, formulate a differential control strategy for the production process of alkali ore ratio.

[0080] For example, for the high purity process feature interval, appropriately increase the mixing ratio of a certain key alkali ore raw material, and at the same time adjust process parameters such as reaction temperature and pressure; for the low purity process feature interval, adopt the opposite adjustment strategy.

[0081] Deploy the formulated control strategy to the DCS (Distributed Control System) system, and through the DCS system, realize the real-time monitoring and automatic control of the alkali ore production process, automatically adjust the production parameters according to different process feature intervals, and improve the stability of the production process and the consistency of product quality.

[0082] Example 3: This example discloses an alkali ore ratio system based on big data. The system includes: a processor, and a memory communicatively connected to the processor; A computer-readable storage medium is provided in the memory, and a computer program is stored on the computer-readable storage medium; When the processor processes the computer program stored on the computer-readable storage medium, the method is implemented.

[0083] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for alkali-ore proportioning based on big data, characterized in that, include: Original sample data containing historical alkali ore ratio data and corresponding performance index data was constructed. The original sample data was cleaned and normalized to obtain training data. Construct a neural network model, train the neural network model using training data, optimize the network weights using the backpropagation algorithm until the preset stopping condition is met, and obtain the trained neural network model. The target performance index is obtained and input into the trained neural network model to obtain alkali-ore ratio data. Production is carried out according to the alkali-ore ratio data.

2. The alkali-ore proportioning method based on big data according to claim 1, characterized in that, The neural network model is a hybrid prediction model, including a shared feature extraction layer, a long short-term memory network branch based on an attention mechanism, and a fully connected feedforward network branch. The outputs of the long short-term memory network branch and the fully connected feedforward network branch are concatenated in the fusion layer to obtain a concatenated vector. The concatenated vector passes through the output layer to obtain the alkali ore ratio data.

3. The alkali-ore proportioning method based on big data according to claim 1, characterized in that, After obtaining the trained neural network model, the method further includes: The trained neural network model is used as a proxy model for the target performance. An optimization objective function including cost and resource consumption is defined. A Bayesian optimization algorithm is used to iteratively search in the continuous ratio parameter space until a Pareto optimal alkali ore ratio data set is found.

4. The alkali-ore proportioning method based on big data according to claim 3, characterized in that, The Bayesian optimization algorithm uses expected improvement as the acquisition function and dynamically balances the exploration and utilization of the proportioning space based on the prediction uncertainty provided by the Gaussian process model during the iteration process.

5. The alkali-ore proportioning method based on big data according to any one of claims 1-4, characterized in that, During the data cleaning and normalization process of the original sample data, an unsupervised learning method based on isolated forest is used to detect anomalies in the original sample data, automatically identifying and removing historical abnormal batch data that deviates from the main process mode.

6. The alkali-ore proportioning method based on big data according to any one of claims 1-4, characterized in that, The neural network model is trained and inferred using Monte Carlo Dropout technology. When outputting alkali ore ratio data, it simultaneously outputs the predicted mean and confidence interval of each performance index data.

7. The alkali-ore proportioning method based on big data according to any one of claims 1-4, characterized in that, The method further includes: Clustering algorithms are used to process the training data to obtain clustering results. The clustering results are mapped to gray values ​​and gray-level histograms are drawn. The optimal threshold is obtained based on the gray-level histogram using the big rule method. The gray-level histogram is then binarized according to the optimal threshold, and the training data is divided into data subsets of different process feature intervals.

8. The alkali-ore proportioning method based on big data according to claim 7, characterized in that, The method further includes: formulating a differentiated alkali-ore ratio production process control strategy based on the significant interval of process characteristics obtained by the optimal threshold division, and deploying the control strategy to the DCS system.

9. The alkali-ore proportioning method based on big data according to any one of claims 1-4, characterized in that, The method further includes: Determine whether the training samples in the training data are balanced. If so, generate adversarial samples and add them to the training data. If not, use a generative adversarial network to generate new training samples and add them to the training data.

10. A big data-based alkali ore proportioning system, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory is provided with a computer-readable storage medium, and a computer program is stored on the computer-readable storage medium. When the processor processes a computer program stored on the computer-readable storage medium, it implements the method as described in any one of claims 1-9.

Images