Method of iron ore grinding granularity soft measurement model based on deep forest

Abstract

The invention relates to the technical field of iron ore grinding and separation processes, in particular to an iron ore grinding granularity soft measurement model method based on a depth forest, which comprises the steps of collecting iron ore grinding production data and preprocessing the data, extracting input features of an iron ore grinding granularity soft measurement model of the depth forest, and setting initial parameters for a depth forest model. Comprising the number n of decision trees in each forest, the size di of a sliding window, the step length stepi and the upper limit of the number of sliding times, multi-granularity window sliding and model generation, and generation of an iron ore grinding granularity prediction result; the deep learning method based on the random forest has higher prediction precision and stronger generalization ability when processing complex ore grinding process data, can monitor the granularity change in the ore grinding process in real time, and helps operators to carry out more accurate control, so that the production efficiency and the resource utilization rate are improved, and the energy consumption and the cost are reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of iron ore grinding and beneficiation processes, and in particular to a method for a soft measurement model of iron ore grinding particle size based on deep forest. Background Art

[0002] Under the background of globalization, the stable supply of mineral products is of great significance to the security and sustainable development of the national economy. Especially in the non-coal mine industry, the efficiency, energy consumption, and production quality of the beneficiation process are directly related to the economic benefits of mining enterprises and the efficient utilization of resources. The beneficiation process mainly includes multiple links such as crushing, screening, grinding and classification, separation, and tailings thickening. As a key link among them, grinding and classification plays an important role in connecting the previous and the following. Through reasonable grinding and classification, not only can the processing capacity of the grinding mill be improved, but also the full dissociation of useful minerals and gangue can be effectively promoted, thus creating ideal conditions for the subsequent separation operation.

[0003] Currently, grinding and classification mainly adopt a closed-circuit operation of a grinding mill and a hydrocyclone. In this process, through the real-time detection and control of the operating state, process parameters of the grinding mill, and the production process of the hydrocyclone, the stability of the grinding fineness can be achieved, and then the classification efficiency can be improved. The grinding fineness of the hydrocyclone is an important indicator to measure the effect of grinding and classification, and its stability directly affects the effect of the beneficiation process and the grade of the ore. However, there are still many deficiencies in the existing grinding fineness detection methods.

[0004] Currently, the monitoring of grinding fineness mainly adopts two methods: manual sampling and testing, and instrument detection. Due to its periodicity and lag, manual sampling cannot reflect the fluctuations in the production process in real time, resulting in the inability to adjust process parameters in a timely manner, affecting the stability of production and product quality. Although instrument detection can provide real-time data, it has high investment costs, large maintenance workloads, and problems such as frequent failures and insufficient detection accuracy. Therefore, traditional detection methods are difficult to meet the requirements of precise monitoring and real-time adjustment of grinding fineness in modern beneficiation processes.

[0005] To solve these problems, soft sensor technology based on machine learning has gradually been applied. In particular, the random forest method exhibits good performance in dealing with complex and non-linear problems. As a new ensemble learning method, the deep random forest requires fewer parameters, has strong robustness, and can automatically adjust complexity. For high-dimensional, large-scale, and complex data processing problems, it can improve the prediction ability and generalization ability of the model, thereby effectively improving data processing efficiency and quality. While ensuring prediction accuracy, it avoids the dependence on large-scale data and computing resources of traditional deep learning methods. The simple forest method has deficiencies in multi-source data mining due to the lack of cross-layer interaction, feature decay caused by a fixed architecture, weak time series capture, poor noise resistance of homogeneous base learners, and insufficient adaptive and dynamic weights. To address the above problems, an improved deep forest-based soft sensor prediction model for grinding particle size is proposed. Summary of the Invention

[0006] The present invention provides a method for a soft sensor model of iron ore grinding particle size based on deep forest, which realizes online real-time prediction of grinding fineness, improves the automation and intelligence level of the beneficiation process, thereby increasing production efficiency, reducing costs, and ensuring the stability and high efficiency of the beneficiation process.

[0007] To achieve the above object, the present invention is implemented by adopting the following technical solutions: A method for a soft sensor model of iron ore grinding particle size based on deep forest includes the following steps: S1. Collect iron ore grinding production data and perform data preprocessing; S2. Extract input features of the deep forest-based soft sensor model for iron ore grinding particle size; S3. Model training and parameter setting: Set initial parameters for the deep forest model, including the number n of decision trees in each forest, the sliding window size d i , the step size step i and the upper limit limit of the sliding times i ; S4. Multi-granularity window sliding and model generation, including the following steps: S4.1. Multi-granularity window sliding: Use a sliding window of size d i to divide the training set D1 to generate feature vectors g of different sizes vi ; S4.2. Train the random forest and the completely random forest: For each feature vector g i extracted from the sliding window d vi , respectively use them to train the random forest and the completely random forest models to obtain two different local feature vectors, and splice them together to finally form the feature vector g i ; S4.3. Training the cascade forest: Use the feature vector g generated in the previous step i to train the cascade forest. At the ni-th layer of the cascade forest, g i is converted into an augmented feature vector ag ni . Connect ag ni with the original feature vector to form the feature vector g at the n-th layer of the i-th sliding window ni . S4.4. Determining the number of layers of the cascade forest: Calculate the training error of the feature vector g ni . As the error gradually decreases, use the current feature vector as the input for the next layer of the combined forest. If the error does not further decrease in three consecutive layers, stop training and determine the number of layers of the cascade forest.

[0008] S5. Generating the prediction result of iron ore grinding particle size: The last layer of the number of layers of the cascade forest serves as the evaluation layer, and calculate the average value of all forest prediction results as the final deep forest regression prediction result.

[0009] Furthermore, in step S1, collecting the iron ore grinding production data is to read the production data set D of the concentrator through a Python program, and divide the data set into a training set D1 and a test set T1, where D1 ∈ R l×m , l represents the number of samples in the training set D1, and m represents the number of feature variables.

[0010] Furthermore, step S2 is specifically as follows: Randomly shuffle each sample in the training set D1 n times to establish a two-dimensional matrix set Matrix corresponding to all samples, and use this two-dimensional matrix set Matrix to train a one-dimensional convolutional neural network Conv and the subsequent connected linear layer Linear; during the training process, set the number of input channels of the Conv network to n, the number of output channels to 1, and adjust the size and stride of the convolutional kernel to ensure that the size of the output vector is m / 2; the output of the connected linear layer Linear is 1, representing the overflow particle size at the next moment; after completing the training of the Conv network and the Linear layer, use the pre-trained Conv network to generate an intermediate vector, which contains m / 2 features and serves as the input for the iron ore grinding particle size soft measurement model based on deep forest.

[0011] Furthermore, the iron ore grinding production data includes the rotational speed, current, load, grinding time, mill decibel, overflow concentration of the classifier, inlet pressure of the hydrocyclone, return sand ratio, ore density and hardness index, steel ball filling rate and ratio, pulp pH value, actual measurement data of the particle size analyzer, and ore feed particle size.

[0012] Furthermore, the data preprocessing includes: for the missing values in the collected iron ore grinding production data, linear interpolation method is used for filling; for the outliers, mean filtering method is used for replacement.

[0013] Furthermore, the evaluation layer evaluates the prediction effects of the sub-samples through different evaluation indexes, and selects the optimal 85% sub-samples for splicing as the output result.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention improves the feature parsing ability through multi-granularity dynamic window: through the multi-granularity window sliding mechanism, the window size and step length are dynamically adjusted, combined with the parallel training of random forest and completely random forest, effectively capturing the short-period fluctuations and long-range trends of the grinding data. The one-dimensional convolutional neural network is used to extract the dynamic time series patterns of the original features. By adjusting the convolutional kernel size and step length, the compressed intermediate feature vector is generated, significantly enhancing the modeling ability of the time-delay effect of parameters such as the current of the ball mill and the ore feed. By splicing the augmented features and the original features to form a new feature vector, the original feature space information is retained, avoiding information loss in the hierarchical transmission, and significantly improving in terms of stability and generalization compared with the traditional method. Verified by industry, the prediction error of the improved model is reduced by 2.7% in a strong noise environment, the adaptation speed to time-varying working conditions is increased by 1.5 times, the real-time prediction response time ≤ 5 seconds, and it is significantly improved in terms of stability and generalization compared with the traditional method. Through the training of a large amount of historical data, it can automatically learn the non-linear relationships in the data and accurately predict the grinding particle size. Compared with the traditional prediction models based on physical models or simple statistical methods, the deep learning method based on random forest has higher prediction accuracy and stronger generalization ability in dealing with the complex grinding process data. The present invention can real-time monitor the particle size change of the grinding process, help the operator to control more precisely, thereby improving the production efficiency and resource utilization rate, reducing the energy consumption and cost, and having good industrial application prospects. This technology is not only applicable to iron ore, but also can be extended to the grinding processes of other minerals. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a schematic diagram of the establishment process of the soft sensor model for iron ore grinding particle size based on deep forest of the present invention.

[0016] Figure 2 is a schematic diagram of the improved multi-granularity scanning process of the present invention.

[0017] Figure 3 is a schematic diagram of the random forest prediction result of the present invention.

[0018] Figure 4 is a schematic diagram of the improved deep forest prediction result of the present invention.

[0019] Figure 5 is the learning curve of the grinding fineness predicted by the deep forest of the present invention. Specific embodiments

[0020] The following further describes the specific embodiments of the present invention in conjunction with the accompanying drawings: See Figure 1 , which is the flowchart of the method of the present invention. A method for a soft sensor model of iron ore grinding particle size based on deep forest of the present invention includes the following steps: S1. Collect iron ore grinding production data and data preprocessing: The iron ore grinding production data includes the rotational speed, current, load, grinding time, mill decibels, overflow concentration of the classifier, inlet pressure of the cyclone, return sand ratio, ore density and hardness index, steel ball filling rate and ratio, pulp pH value, actual measurement data of the particle size analyzer, and ore feed particle size. Obtain the production dataset D of the concentrator. Use the Python programming language to load the dataset D and perform preliminary processing. The preprocessing includes filling the missing values in the dataset using the linear interpolation method, and replacing the outliers using the mean filtering method. By calculating the mean and standard deviation of the data, identify the outliers and replace them with the mean of the samples where they are located. After the processing is completed, divide the dataset D into a training set D1 and a test set T1 according to a ratio of 7:3, where D1 ∈ R l×m ; Training set D1: with a size of Rl×m, where l is the number of samples in the training set and m is the number of feature variables; Test set T1: used to verify the model performance subsequently.

[0021] S2. Input feature extraction of the soft sensor model of iron ore grinding particle size by deep forest, see Figure 2 , extract the dynamic time series pattern of the original input features through a one-dimensional convolutional neural network, set the convolution kernel size and sliding step parameters, dynamically calculate the output feature dimension according to the formula, map the original high-dimensional features to a low-dimensional space, retain the key time series evolution rules, and balance feature dimensionality reduction and dynamic pattern capture by adjusting the convolution kernel parameters, such as the combination of kernel size and step length, to solve the problem of feature coupling caused by the time delay effect of grinding parameters. After the output features are aligned and corrected, a compressed intermediate representation is formed as the input of the deep forest model. Specifically, it includes the following steps: S2.1. Randomly shuffle the training set data: Randomly shuffle each sample in the training set D1 n times to generate n different shuffled data samples, and save the shuffled data as a two-dimensional matrix. The set of two-dimensional matrices corresponding to all samples is called Matrix, and the size of Matrix is l×n×m; S2.2. Training the Convolutional Neural Network: Use the two-dimensional matrix set Matrix to train a one-dimensional convolutional neural network Conv and the subsequent connected linear layer Linear. During the training process, set the number of input channels of Conv to n, which is the number of shuffling times, the number of output channels to 1, and adjust the size and stride of the convolutional kernel to ensure that the size of the output vector is m / 2. Obtain the weights of the one-dimensional convolutional neural network through training and extract features through this network; S2.3. Training the Linear Layer: On the basis of the convolutional layer, connect a linear layer Linear. The role of the linear layer is to further process the output vector of the convolutional neural network and finally output a value representing the grinding fineness at the next moment. Specifically, the output of the connected linear layer Linear is 1, representing the grinding fineness at the next moment. Optimize the parameters of the linear layer through training so that it can effectively predict the particle size at the next moment; S2.4. Generating the Feature Vector: After completing the training of the Conv network and the Linear layer, use the intermediate vector generated by the pre-trained Conv network through the trained convolutional neural network Conv and the linear layer Linear. This vector contains m / 2 features, and this vector is used as the input of the deep forest model and enters the deep forest model for further training.

[0022] S3. Model Training and Parameter Setting: Set the initial parameters for the deep forest model to ensure that the model can operate efficiently and obtain good prediction results. The initial parameters include the number of decision trees n in each forest, the sliding window size d i , the stride step i and the upper limit of the sliding times limit i ; The number of decision trees n in each forest: Set the number of decision trees to ensure the diversity and robustness of the forest model; The sliding window size d i : Set for the training data to ensure the multi-granularity characteristics of effectively extracting data features; The stride step i : Set the stride of the sliding window, which determines the granularity of data segmentation; The upper limit of the sliding times limit i : Set the maximum number of times of the sliding window to prevent excessive consumption of computing resources during the training process; Different from the traditional method, artificially limit the sliding process of the multi-granularity window to reduce the spatio-temporal complexity generated by the deep forest and avoid the memory overflow problem caused by resource limitations.

[0023] S4. Multi-granularity Window Sliding and Model Generation, including the following steps: S4.1. Multi-granularity Window Sliding: Use a size of di The training set D1 is segmented by a sliding window, and feature vectors g of different sizes are extracted from each sliding window vi Specifically, the implementation method is as follows: Apply a sliding window to the training set D1, and the window size is d i Subsets are extracted from different positions, and each window slide generates a feature vector g vi representing the features under this window; S4.2. Train the random forest and completely random forest: For each sliding window d i The extracted feature vector g vi is used to train the random forest and completely random forest respectively. The specific steps are as follows: Based on g vi Train the random forest to obtain the local features corresponding to this feature vector. Based on g vi Train the completely random forest to obtain another local feature vector. Concatenate these two different feature vectors to obtain the final feature vector g i ; S4.3. Train the cascade forest: Use the feature vector g generated in the previous step i to train the cascade forest layer by layer; The specific process is that at the ni-th layer of the cascade forest, gi is converted into an augmented feature vector ag ni Connect ag ni with the original feature vector to form the feature vector g of the i-th sliding window at the n-th layer ni and perform training and optimization; S4.4. Determine the number of layers of the cascade forest: Calculate the training error of the feature vector g ni As the error gradually decreases, use the current feature vector as the input of the next layer of the combined forest; The specific steps are as follows: If the error does not further decrease in 3 consecutive layers, stop training and determine the number of layers of the cascade forest, and select the optimal number of layers to obtain the lowest error and improve the prediction accuracy.

[0024] See Figures 3-4 , The improved model adapts to the multi-granularity feature fusion mechanism through dynamic parameters, significantly improving the sample point fitting degree. Compared with the traditional method, the goodness of fit R² of the present invention on the training set and test set is increased by about 15% and 12% respectively. Especially in the area where the grinding working condition changes suddenly, such as the fluctuation of ore hardness and the jump of equipment load, it shows stronger adaptability. This improvement effectively solves the problem of insufficient fitting of the traditional method under complex working conditions and provides a more accurate model support for grinding particle size prediction.

[0025] S5. Generate the prediction result of iron ore grinding particle size: The last layer of the cascade forest serves as the evaluation layer, calculates the average value of all forest prediction results, and takes it as the final deep forest regression prediction result y. This method evaluates the prediction effect of the sub-samples through different evaluation indicators, including mean square error, mean absolute error, etc. Select the 85% sub-samples with the best performance for splicing as the improved output result to reduce the impact of low-quality sub-samples on the model result. This method can effectively improve the prediction accuracy of the model, thus enhancing the prediction accuracy and robustness of the model, strengthening the feature learning ability of the deep forest, and avoiding the negative impact of sub-samples with poor performance on the model result.

[0026] See Figure 5 , Experimental verification shows that the dynamic parameter optimization mechanism of the present invention significantly improves the model prediction accuracy. Under the condition of a fixed tree depth max_depth, as the number of base learners n_trees gradually increases from 0 to 300, the mean square error MSE of the model shows a continuous downward trend, decreasing from the initial 0.018 to the optimal 0.005, with a decrease of 72.2%. Especially when n_trees exceeds 150, the convergence rate of the MSE curve accelerates, indicating that the model effectively suppresses the overfitting risk by increasing the diversity of heterogeneous forests. This characteristic verifies the synergistic advantage of the multi-granularity window and the dynamic cascade architecture: by adaptively adjusting the scale of the base learner and the hierarchical depth, the model can still maintain high accuracy and strong generalization ability under complex working conditions. Compared with the traditional fixed-parameter model, the prediction stability of the improved method is increased by about 2.3 times under the same experimental conditions, providing a reliable technical guarantee for the real-time control of grinding particle size.

[0027] The above embodiments are implemented on the premise of the technical solution of the present invention, and the detailed implementation manners and specific operation processes are given. However, the protection scope of the present invention is not limited to the above embodiments. The methods used in the following embodiments are conventional methods unless otherwise specified.

Claims

1. A method for a soft sensor model of iron ore grinding particle size based on deep forest, characterized in that, It includes the following steps: S1. Collect iron ore grinding production data and perform data preprocessing; S2. Extract input features of the soft sensor model for iron ore grinding particle size by deep forest; S3. Model Training and Parameter Setting: Set the initial parameters for the deep forest model, including the number of decision trees n in each forest, the sliding window size d i , the step size step i , and the upper limit limit of the sliding times i ; S4. Multi-granularity window sliding and model generation, including the following steps: S4.

1. Multi - granularity window sliding: Use a sliding window of size d i to segment the training set D1 and generate feature vectors g vi of different sizes; S4.

2. Training Random Forest and Completely Random Forest: For each sliding window d i The extracted feature vector g vi , which are respectively used to train the random forest and completely random forest models, thereby obtaining two different local feature vectors, and splicing them together to finally form the feature vector g i ; S4.

3. Training the cascaded forest: Use the feature vector g generated in the previous step i to train the cascaded forest. At the ni-th layer of the cascaded forest, g i is converted into an augmented feature vector ag ni . Connect ag ni with the original feature vector to form the feature vector g of the n-th layer of the i-th sliding window ni ; S4.

4. Determine the number of layers of the cascade forest: Calculate the training error of the feature vector g ni and, as the error gradually decreases, use the current feature vector as the input to the next layer of the combined forest. If the error does not further decrease in three consecutive layers, stop training and determine the number of layers of the cascade forest; S5. Generate the prediction result of iron ore grinding particle size: The last layer of the cascade forest is used as the evaluation layer, and the average value of all forest prediction results is calculated as the final regression prediction result of the deep forest.

2. The method of a soft sensor model for iron ore grinding particle size based on deep forest according to claim 1, wherein, In the step S1, the iron ore grinding production data is collected by using a Python program to read the production data set D of the concentrator, and the data set is divided into a training set D1 and a test set T1, where D1 ∈ R l×m , l represents the number of samples in the training set D1, and m represents the number of feature variables.

3. The method of a soft sensor model for iron ore grinding particle size based on deep forest according to claim 2, characterized in that, The specific content of step S2 is as follows: Each sample in the training set D1 is randomly shuffled n times to establish a two-dimensional matrix set Matrix corresponding to all samples. Use this two-dimensional matrix set Matrix to train a one-dimensional convolutional neural network Conv and the subsequent connected linear layer Linear; during the training process, the number of input channels of the Conv network is set to n, the number of output channels is 1, and the size and stride of the convolutional kernel are adjusted to ensure that the size of the output vector is m / 2; the output of the connected linear layer Linear is 1, representing the overflow particle size at the next moment; after completing the training of the Conv network and the Linear layer, use the pre-trained Conv network to generate an intermediate vector, which contains m / 2 features and is used as the input of the soft sensor model for iron ore grinding particle size based on the deep forest.

4. A method for a soft measurement model of iron ore grinding particle size based on deep forest according to claim 1, characterized in that, The iron ore grinding production data includes the rotation speed, current, load, grinding time, mill decibel, overflow concentration of the classifier, inlet pressure of the cyclone, return sand ratio, ore density and hardness index, steel ball filling rate and ratio, pulp pH value, actual measurement data of the particle size analyzer, and ore feed particle size of the ball mill.

5. A method for a soft sensor model of iron ore grinding particle size based on deep forest according to claim 1, characterized in that, The data preprocessing includes: For the missing values in the collected iron ore grinding production data, linear interpolation method is used for filling; for the outliers, mean filtering method is used for replacement.

6. A method for a soft sensor model of iron ore grinding particle size based on deep forest according to claim 1, characterized in that, The evaluation layer evaluates the prediction effect of the sub-samples through different evaluation indexes, and selects the optimal 85% sub-samples for splicing as the output result.

Images