A multi-working-condition prediction model for a grinding classification process and a construction method thereof

Abstract

The application discloses a kind of multi-working condition prediction models for grinding ore classification process and its construction method, belong to mineral processing technical field;The application includes the following steps: first, time delay correlation analysis algorithm is used to extract dynamic characteristics, to construct the dynamic characteristic data set for overflow concentration fineness prediction. Then, the model is initially pre-trained by full data, and the initial data set is divided into different working condition data sets using unsupervised clustering method. Subsequently, the pre-trained model is fine-tuned and retrained using these working condition data sets, so that the model can adapt to multiple working condition characteristics. In the prediction stage, the distance between the sample point to be predicted and each cluster center point is calculated to determine the working condition to which the sample belongs, and the corresponding model is called for accurate prediction. Compared with the prior art, the application improves the accuracy of grinding ore classification prediction, and promotes the intelligentization and stable operation of mineral processing production.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ore dressing, and particularly relates to a grinding and grading process multi-working condition prediction model based on time lag analysis and clustering fine tuning and a construction method thereof. BACKGROUND

[0002] The grinding and grading process is a key link in the ore dressing process, and has significant dynamic characteristics and time lag. In the grinding link, the material quantity in the mill, the movement state of the grinding medium, and the factors such as the feed size and hardness are in dynamic change at all times. For example, when the feed quantity suddenly increases, the material in the mill increases, the load of the ball mill increases, and the motor current also rapidly rises, which reflects the rapid response of the grinding process to the change of the input condition, that is, the dynamic characteristic. However, after the grinding product enters the grading equipment, the grading effect does not appear immediately. From the material entering the cyclone classifier to the accurate separation and output of the qualified product and the unqualified product after grading, there is a certain time delay, that is, the time lag. This lag is mainly due to the fact that the material needs to go through a series of complex physical processes such as sedimentation and separation in the classifier, resulting in a time difference from the change of the operation variable to the obvious change of the output variable. Moreover, various disturbance factors in the grinding and grading system, such as the fluctuation of the ore properties and the gradual aggravation of the equipment wear, also continuously affect the dynamic characteristics and time lag characteristics of the system, making the prediction effect of the grinding and grading overflow concentration worse.

[0003] The traditional prediction model usually takes simple linear hypothesis and static working condition as the premise, so that it has an insurmountable technical bottleneck in dealing with the complex characteristics of the grinding and grading process. For example, the method based on a simple mathematical model sets a fixed proportional coefficient to associate the production parameters, and cannot capture the influence of the complex and changeable factors such as the material quantity in the mill, the movement state of the grinding medium, the feed size and hardness on the grinding and grading effect in real time. When the mill load changes suddenly due to the sudden change of the feed quantity, the prediction deviation of the grinding product quality is easy to occur. The method based on time series analysis relies on the time series characteristics of the historical data, but the data is non-stationary due to the influence of factors such as the fluctuation of the ore properties and the equipment wear, and it is difficult to effectively adjust the prediction in the face of dynamic changes, affecting the accuracy. The static neural network model only considers the static relationship between the input and output data, and does not fully mine the time series characteristics and dynamic change law of the grinding and grading process. When the key parameters in the mill change rapidly, the dynamic information cannot be effectively captured, resulting in large prediction deviation. In addition, the traditional method for dealing with time lag is also seriously insufficient. The method based on mathematical model and experience often ignores the time delay of the material from entering to output in the classifier based on immediate feedback, and the method based on time series analysis and static neural network model also lacks effective processing mechanism, resulting in that the prediction result lags behind the actual working condition change, affecting the production efficiency, causing the product quality to be unstable, and increasing the production cost.

[0004] In view of the defects of the above-mentioned traditional prediction model, the patent proposes a grinding classification process multi-working condition prediction model based on time delay analysis and clustering fine-tuning and a construction method thereof. SUMMARY

[0005] The purpose of the present application is to provide a grinding classification process multi-working condition prediction model and a construction method thereof to solve the problems mentioned in the background art; the present application effectively improves the adaptability and prediction accuracy of the model to the time delay dynamics and multi-working conditions of the grinding classification process, and can ensure the efficient and stable operation of the beneficiation production.

[0006] In order to achieve the above-mentioned purpose of the application, the present application provides the following technical solutions:

[0007] A grinding classification process multi-working condition prediction model construction method, comprising the following steps:

[0008] S1, dynamic feature dataset construction: based on the historical operation parameter database of the grinding process, using time delay correlation analysis algorithm for feature time delay compensation, constructing dynamic feature dataset for overflow concentration fineness regression prediction, to solve the problem of poor effect caused by time delay;

[0009] S2, model construction and pre-training: selecting a pre-training model suitable for sequence data processing, pre-training the model using the dynamic feature dataset obtained in S1, to obtain an initial prediction model;

[0010] S3, multi-working condition dataset construction: using unsupervised clustering method to divide the initial dynamic feature dataset, to construct different working condition datasets, to identify various working condition characteristics in the grinding classification process;

[0011] S4, model training: designing a pre-training-clustering fine-tuning multi-working condition model training method, using the multi-working condition dataset obtained in S3 to retrain the initial prediction model pre-trained in S2, to obtain a multi-working condition prediction model, and to adapt the model to the characteristics of various works;

[0012] S5, prediction execution: based on the multi-working condition prediction model obtained in S4, designing a working condition recognition method based on distance calculation, determining the working condition to which the sample belongs by calculating the distance from the prediction sample point to each cluster center point, and then calling the corresponding model for accurate prediction.

[0013] Preferably, S1 specifically includes the following contents:

[0014] Collecting grinding classification process related parameters, after pre-processing the parameters, constructing a beneficiation process original dataset (X, Y), wherein the input parameter set X={x i}, i = 1, 2, …, m, containing m parameters affecting the overflow concentration; the output parameter set Y = {y j}, j = 1, 2, …, n, containing n target values to be predicted; N is the total number of sample points;

[0015] Subsequently, the time delay relationship between variables is accurately determined by the time delay correlation analysis algorithm, the dynamic characteristics of the grinding process are mined, and the dynamic characteristic data set (T, Y) for the overflow concentration regression prediction is constructed by extending and serializing the dynamic characteristics, wherein T represents the input characteristics after dynamic characteristic mining;

[0016] The construction process of the dynamic characteristic data set is specifically as follows:

[0017] The correlation coefficient of the i-th feature input time sequence of the grinding classification process to the j-th output variable time sequence is represented as follows:

[0018]

[0019] In the formula, x i (t) is the i-th feature of the input time sequence X at the t-th moment; y j (t+τ) is the j-th feature of the output time sequence Y to be predicted at the t+τ moment; and are the mean values of the i-th feature input time sequence and the j-th output variable time sequence, respectively; N is the total number of sample points; N0 is the number of observation points of the time sequence, which is greater than the maximum number of delay points, so that the time delay of the two time sequences is the number of time delays corresponding to the maximum correlation moment, which is represented as follows:

[0020]

[0021] For the i-th input feature, the maximum delay in all output sequences is taken, which is represented as:

[0022]

[0023] Through the above time delay information, the dynamic characteristic extension is performed on the data set (X, Y), and the serialization process of the i-th sample of the input variable is represented as:

[0024] x i (t)→[x i (t), x i (t-1), …, x i (t-τ i )]

[0025] Based on the above process, the dataset will combine each variable with its previous variable values at each time point according to the time lag relationship, forming a feature vector containing dynamic information.

[0026] Preferably, the pre-trained model in S2 is a neural network model, a decision tree model, or a combination thereof. The dynamic feature dataset (T, Y) is used as the training dataset of the pre-trained model, so that the model learns general features and patterns in large-scale data and masters the relationship between various parameters in the grinding process.

[0027] Preferably, the unsupervised clustering method in S3 is a K-means clustering algorithm, a density-based spatial clustering algorithm, or a Gaussian mixture model clustering algorithm. The dynamic feature dataset (T, Y) is used as the input feature of the clustering algorithm, so that the clustering adapts to the dynamic characteristics, and finally the clustering result (T, Y, C) of each sample is obtained; wherein C represents the clustering result.

[0028] Preferably, the clustering fine-tuning in S4 is a targeted adjustment of the pre-trained model according to the clustering result, specifically: for each working condition represented by the cluster, the performance of the pre-trained model on the working condition data is evaluated respectively, and the error pattern between the model prediction result and the actual value is analyzed; according to the obtained error analysis result, the connection weight and bias between the neurons of the neural network model are adaptively adjusted by means of the back propagation algorithm.

[0029] Preferably, S5 specifically includes the following contents:

[0030] Determine the working condition to which the to-be-predicted sample point belongs, and calculate the distance from the sample point to each cluster center point; the distance measurement is consistent with the clustering algorithm, and the Euclidean distance measurement is used, and the expression is:

[0031]

[0032] In the formula, μ k represents the cluster center of the kth cluster; D new represents the processed to-be-predicted sample point;

[0033] According to the nearest distance principle, the sample point is determined as the working condition represented by the nearest cluster center point, which is expressed as:

[0034]

[0035] In the formula, c k represents the kth cluster;

[0036] After determining the working condition, the corresponding prediction model is used for accurate prediction.

[0037] The application further protects a multi-working condition prediction model for the grinding classification process, which is constructed by the above method.

[0038] The application further protects a computer device, which comprises a processor and a memory, and the memory stores at least one instruction, at least one program, a code set or an instruction set, which are loaded and executed by the processor to realize the above grinding classification process multi-working condition prediction model construction method.

[0039] The application further protects a computer readable storage medium, which stores at least one instruction, at least one program, a code set or an instruction set, which are loaded and executed by the processor to realize the above grinding classification process multi-working condition prediction model construction method.

[0040] Compared with the prior art, the application has the following beneficial effects:

[0041] (1) Compared with the prior art, which mostly ignores the time lag characteristics in the grinding classification process and only considers the current static data without reflecting the dynamic changes in the process, the application uses a time lag correlation analysis algorithm to mine the time delay relationship between variables, thereby effectively dealing with the time lag dynamics in the grinding classification process. Based on the time delay relationship between variables, the application constructs a dynamic feature data set for overflow concentration fineness regression prediction. This data set can fully reflect the change law of the grinding classification process over time by incorporating rich dynamic information, thereby enabling the model to fully learn and adapt to the time lag dynamics of the grinding classification process, thereby significantly improving the timeliness and accuracy of the prediction.

[0042] (2) The application uses a large amount of grinding process historical data to pre-train the model, so that the model learns general features and patterns in large-scale data and has a certain knowledge reserve. Then, the model is fine-tuned under different working conditions to construct a multi-working condition prediction model. The training process of the existing prediction model is mostly to directly train and model the full amount of data, only once. However, the pre-training-fine-tuning mode of the application can pre-train a large amount of data, avoid the blindness of the model in the initial learning, speed up the learning progress, and reduce the dependence of the model on some specific working condition data.

[0043] (3) The existing prediction methods are mostly based on simple linear assumptions and single working conditions to build models, which are difficult to adapt to the complex and variable working conditions in the grinding and classification process. The present application uses an unsupervised clustering method, takes a dynamic feature data set as input, divides the grinding and classification process into different working conditions, and adjusts the pre-trained model according to different working condition data sets. When making predictions, first identify the working condition, and then use the corresponding working condition model to make predictions. In this way, the model can better capture the unique features of each working condition, make up for the shortcomings of the single pre-trained model in dealing with complex working conditions, and significantly improve the prediction accuracy and generalization ability of the model under different working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0044] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, a brief introduction is made to the drawings involved in the embodiments. Obviously, the drawings in the following description are only schematic illustrations of some embodiments of the present application, and other forms of drawings can be constructed by those skilled in the art without creative labor based on these drawings.

[0045] Figure 1 A flow chart of a multi-working condition prediction model construction method for a grinding and classification process is proposed in the present application;

[0046] Figure 2 A flow chart of a dynamic feature data set for constructing an overflow concentration fineness regression prediction is proposed in the embodiments of the present application;

[0047] Figure 3 A schematic diagram of dynamic feature expansion and serialization in the construction of a dynamic feature data set is proposed in the embodiments of the present application;

[0048] Figure 4 An elbow method curve diagram is proposed in the embodiments of the present application;

[0049] Figure 5 A comparison diagram of overflow concentration fineness prediction before and after TDMI is proposed in the embodiments of the present application. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all.

[0051] The application provides a multi-working-condition prediction model for a grinding and grading process and a construction method thereof, focuses on the prediction field, and effectively overcomes the deficiencies of traditional methods in dealing with time lag through accurate time lag analysis technology and in-depth analysis of time lag relationships in the grinding and grading process. Meanwhile, the model is optimized according to different working conditions by using clustering fine-tuning technology, successfully solves the problem that traditional methods are difficult to cope with dynamic characteristics, improves the accuracy and stability of prediction, better adapts to the complex and changeable actual production requirements of the grinding and grading process, and provides strong technical support for efficient and stable operation of the beneficiation industry. The multi-working-condition prediction model for the grinding and grading process and the construction method thereof are described below in combination with relevant drawings and specific examples, and the specific content is as follows.

[0052] Embodiment 1

[0053] As shown in Figure 1 , the application provides a multi-working-condition prediction model construction method for a grinding and grading process, which comprises the following steps.

[0054] S1, a large amount of historical parameter data of the grinding process is used to construct a dynamic feature data set for overflow concentration fineness regression prediction by means of an advanced time lag correlation analysis algorithm; the specific contents include the following.

[0055] S1.1, data collection and preprocessing: all kinds of parameters in the field decentralized control system of the grinding and grading process of the beneficiation plant are collected, including but not limited to feed quantity, mill power, classification machine overflow concentration, particle size and other data. The collected data are cleaned to remove abnormal values and noise data, so as to ensure the accuracy and reliability of the data. At the same time, the data are normalized to make the data of different parameters comparable. Accordingly, a grinding and grading process overflow concentration fineness prediction original data set (X, Y) with a dimension of (m+n) x N is constructed. Wherein, the input parameter set X = {x i}, i = 1, 2, …, m, contains m parameters affecting the overflow concentration fineness, the output parameter set Y = {y j}, j = 1, 2, …, n, contains n target values to be predicted, and N is the total number of sample points.

[0056] S1.2, time lag correlation analysis: the (X, Y) constructed in S11 is analyzed by using an advanced time lag correlation analysis algorithm, such as a time lag analysis method based on correlation coefficient, to comprehensively analyze all variables in the grinding and grading process. The correlation coefficients between different variables are calculated to accurately determine the time delay relationship between the variables. The dynamic characteristics of the grinding process are mined, and a dynamic feature data set (T, Y) for overflow concentration fineness regression prediction is constructed by extending and serializing the dynamic characteristics, as shown in Figure 2 .

[0057] The time lag analysis algorithm is as follows: by measuring the correlation of two time series at different time delays and accurately calculating the correlation coefficients at different time lags, the time lag that makes the correlation coefficient reach the maximum value is found, and this specific time lag is the optimal time delay between the two time series. The correlation coefficient of the ith characteristic input time series of the grinding classification process to the jth output variable time series is represented as follows:

[0058]

[0059] In the formula, x i (t) is the ith characteristic of the input time series X at time t, y j (t+τ) is the jth characteristic of the output time series Y to be predicted at time t+τ, and are the mean values of the ith characteristic input time series and the jth output variable time series, respectively, and N0 is the number of observation points of the time series, which should be greater than the number of maximum delay points. The time lag of the two time series is the number of time delays corresponding to the moment when the correlation is maximum, which is represented as follows:

[0060]

[0061] wherein

[0062]

[0063] The time lag information mined through dynamic characteristics is subsequently expanded to the dynamic characteristic of the data set (X, Y), so that each data sample not only contains the variable value at the current time, but also integrates the historical time variable value with strong correlation. By adding the historical time variable with strong correlation, the model can better learn and adapt to the time lag dynamics of the grinding classification process during training. Specifically, for each input variable, according to the calculated maximum delay τ i , the variable values of the previous τ i time points are added as new feature dimensions to the original data samples. For example, if the maximum delay of the feed rate to the overflow concentration is 3 time points, then in the construction of the dynamic characteristic data set, in addition to the current time value of the feed rate, the previous 3 time points of the feed rate value will also be included in the data sample, and finally a dynamic characteristic data set (T, Y) for overflow concentration regression prediction is constructed through serialization. For the sample corresponding to the t time of the ith characteristic of the input variable, the process is represented as follows: ​

[0064] x i (t)→[x i (t),x i (t-1),...,x i (t-τ i )]

[0065] This dataset combines the variable value at each time point with its values ​​at several preceding time points according to time lag relationships, forming a feature vector containing dynamic information. This unique modeling approach enables the model to fully learn and adapt to the time-lag dynamics of the grinding and classification process, better capture the mutual influences at different time scales, and effectively overcome the shortcomings of traditional methods that only consider static data at the current moment and cannot reflect the dynamic changes of the process. The specific implementation process is as follows: Figure 3 As shown. For ease of plotting, x is... i y(t) and y(t) are defined as follows: and y t .

[0066] S2. Using the complete historical data of the grinding process, the model is initially pre-trained to obtain a pre-trained model. The model used for initial pre-training includes, but is not limited to, neural network models or combinations thereof. The dynamic feature dataset obtained from S1 is used as the training dataset for the pre-trained model. Pre-training allows the model to learn common features and patterns from large-scale data, thus acquiring a certain knowledge reserve and laying the foundation for subsequent refined training for different working conditions. Through pre-training, the model initially grasps the relationships between various parameters in the grinding process, such as the potential connections between parameters like ore properties, mill speed, and feed rate and the grinding effect, which helps improve the efficiency and accuracy of subsequent model training, enabling it to adapt more quickly to data characteristics under different working conditions; specifically including the following:

[0067] Long Short-Term Memory (LSTM) networks, a neural network model suitable for processing sequential data, were used as the pre-trained model. This model has significant advantages in learning complex patterns and long-term trends in sequential data. Its unique structural design effectively captures long-term dependencies in the data and handles information inflow, outflow, and forgetting through gating mechanisms, overcoming the limitations of traditional neural networks when processing data with long time spans. In the pre-training phase, the entire historical data of the grinding process was further cleaned and organized to ensure data quality and consistency. The dynamic feature dataset (T, Y) constructed by S1 was divided into training and validation sets according to a certain ratio before initial pre-training of the selected model. The loss function chosen for the pre-training process is expressed as follows:

[0068]

[0069] wherein y represents the model predicted output value, y i (t) represents the actual value in the data set, in each round of training, the model predicts the grinding effect related indicators such as overflow concentration fineness according to the input dynamic characteristic data. By calculating the loss function between the predicted value and the true value, the parameters of the model are updated by using the back propagation algorithm. After multiple iterations of training, the model gradually learns the general characteristics and patterns between various parameters in the grinding process, and preliminarily masters the potential relationship between ore properties, mill speed, feed amount and grinding effect, laying a solid foundation for subsequent fine training for different working conditions. At the same time, the performance of the model is evaluated by using the validation set, and when the performance of the model on the validation set no longer improves, the training is stopped, and the model at this time is saved as a pre-trained model.

[0070] S3, using unsupervised clustering methods such as K-means clustering algorithm, density-based spatial clustering algorithm or Gaussian mixture model clustering algorithm, etc. to divide the initial data set. The dynamic characteristic data set is used as the input feature of the clustering algorithm, so that the clustering can better adapt to the dynamic characteristics of the grinding classification process. Through clustering analysis, the grinding classification process is divided into different working conditions, and different working condition data sets are obtained, each of which represents a specific working condition; specific contents include the following:

[0071] S3.1 Clustering algorithm based on dynamic characteristic data set is implemented as follows:

[0072] In view of the characteristics of the grinding classification process data and the clustering demand, the K-means unsupervised clustering method is used to cluster the above dynamic characteristic data set. This algorithm can quickly converge, and can distribute data by calculating the distance of data points to each cluster center. The following formula is used to determine the number of clusters K of clustering, and the sum of squared errors within classes (SSE) is calculated under different K values, as follows:

[0073]

[0074] In the formula, K represents the number of clusters, C k represents the kth cluster, D = {T, Y} represents the input sample set of the clustering algorithm, and its size is (m+n) × N d , D i ∈ C k represents the ith sample point in the kth cluster in D, N d is the number of sample points of the dynamic data set constructed after time delay analysis, μ k represents the clustering center of the kth cluster, which can be obtained by the following formula:

[0075]

[0076] Draw the curve of SSE with K, and select the K value corresponding to the point where the slope of the curve suddenly becomes smaller as the number of clustering clusters of the algorithm. At the same time, set the maximum number of iterations to avoid the algorithm from falling into an infinite loop.

[0077] S3.2 Unsupervised clustering of different working conditions of the grinding classification process, as follows:

[0078] After determining the clustering algorithm and K value, the dynamic feature data set (T, Y) obtained from S2 is used as the input feature of the clustering algorithm. These dynamic feature data sets contain time delay information and dynamic change characteristics in the grinding classification process, which can make the clustering process better adapt to the dynamic characteristics of the grinding classification process.

[0079] By minimizing the sum of the squares of the distances of all data points to their cluster centers, the objective function J is defined as the sum of squared intra-class errors, that is:

[0080]

[0081] Finally, the clustering result (D, C) of each sample is obtained, where C k is the kth cluster in the clustering result C.

[0082] S4, retrain the pre-trained model using the different working condition data sets to make the model adapt to the characteristics of various working conditions; specifically including the following contents:

[0083] S4.1, evaluate the performance of the pre-trained model under each working condition, as follows:

[0084] Further divide the different working condition data sets into training subsets and validation subsets, for example, divide them according to the ratio of 8:2. The training subset is used to retrain the model, and the validation subset is used to evaluate the performance of the model during the retraining process to prevent the model from overfitting. The mean squared error is used to measure the performance of the pre-trained model on each working condition data set. For each working condition data set, the MSE of the pre-trained model is calculated, and the error pattern between the model prediction result and the actual value is analyzed. For example, whether the error has a systematic bias, how the error fluctuates, etc. Through the analysis of the error pattern, the shortcomings of the model under different working conditions can be understood in depth, providing a basis for subsequent parameter adjustment.

[0085] S4.2, adjust the model parameters based on the error analysis, as follows:

[0086] For the parameters of the neural network model, the connection weights and biases between neurons can be adjusted. Specifically, when the predicted values of the model are generally too high under a certain working condition, the connection weights of the relevant neurons can be appropriately reduced; when the predicted values fluctuate greatly, the learning rate can be adjusted and the regularization term (such as L1 or L2 regularization) can be increased to stabilize the training process of the model. In addition, the structure of the network can also be adjusted, such as increasing or decreasing the number of neurons in the hidden layer, changing the activation function, etc.

[0087] S4.3, retraining process and monitoring, specifically as follows:

[0088] Training process: using the training subset of each working condition to retrain the model after adjusting the parameters. During the training process, appropriate optimization algorithms are used, and appropriate training rounds and batch sizes are set to update the parameters of the model to ensure that the model can fully learn the characteristics of the working condition dataset.

[0089] Monitoring and adjustment: during the retraining process, the performance of the model is monitored in real time using the validation subset. If it is found that the performance of the model on the validation subset is declining, such as the validation error starting to rise, it indicates that the model may have overfitting phenomenon, at which time the training strategy needs to be adjusted in time, such as stopping training in advance, increasing the regularization strength, etc.

[0090] S5, when making predictions, the distance of the sample point to be predicted to each cluster center point is calculated to determine the working condition to which the sample belongs, and then the corresponding model is called for accurate prediction, which includes the following contents:

[0091] S5.1, distance calculation and working condition determination, specifically as follows:

[0092] Determine the working condition to which the sample point to be predicted belongs, and calculate the distance of the sample point to each cluster center point. The distance measurement is consistent with the previous clustering algorithm, using the Euclidean distance measurement, the expression is:

[0093]

[0094] In the formula, D new represents the processed sample point to be predicted. According to the principle of the nearest distance, the sample point is determined as the working condition represented by the nearest cluster center point, which is expressed as follows:

[0095]

[0096] In the formula, C k represents the kth cluster of the clustering result.

[0097] S5.2, call the corresponding model for prediction, specifically as follows:

[0098] After determining the working condition to which the to-be-predicted sample belongs, a model retrained for the working condition is called to perform prediction. Since the model has been finely adjusted for different working conditions in S4, each model can well adapt to the characteristics of the respective working condition, and thus the prediction results for indicators such as overflow concentration are more accurate and robust, providing an important basis for decision-making in the beneficiation production process.

[0099] Embodiment 2:

[0100] Based on Embodiment 1 but with differences, a method for constructing a multi-working-condition prediction model for the grinding classification process based on time delay analysis and clustering fine-tuning is proposed below in combination with a specific example, including the following steps:

[0101] Step 1: Collect relevant parameters of the grinding classification process in the beneficiation plant, including 14 variables such as feed quantity, water quantity, and overflow concentration, with a total of 34,000 sample data, as shown in Table 1. Preprocess the collected data, including data cleaning (removing outliers and missing values), data normalization (unifying the data to the same scale range), and other operations, to construct an original data set (X, Y) of the beneficiation process. In this example, X = [x1, x2, …, x 11 ], Y = [y1, y2, y3], and the time delay characteristics of the grinding classification process are mined through a time delay analysis algorithm based on correlation coefficients, as shown in Table 1.

[0102] Table 1: Modeling variable information and related delay information

[0103] Variable Variable name Unit Number of delay points (τ) Delay time (min) x1 Feed rate t / d 0 0 x2 Water rate t / d 0 0 x3 Pump tank level cm 60 5 x4 Sizing concentration ％ 96 8 x5 Ball mill power kwh 96 8 x6 Oil pressure at feed end kpa 18 1.5 x7 Sizing flow rate m 3 / h]]> 204 17 x8 Cyclone pressure mPa 180 15 x9 Overflow flow rate m 3 / h]]> 288 24 x 10 ]]> Slurry pump frequency hz 252 21 x 11 ]]> Make-up water rate m 3 / h]]> 180 15 [ y1 ] Overflow concentration ％ 288 24 [ y2 ] Overflow fineness +80 mesh ％ 300 25 [cdta] Overflow fineness -200 mesh ％ 312 26

[0104] Then, the dynamic feature data set (T, Y) for overflow concentration regression prediction is constructed by dynamically expanding and serializing the original data set. Figure 3

[0105] Step 2: Use an LSTM model as the initial pre-training model, with the following model structure design: the number of input layer neurons corresponds to the 14 feature dimensions of the dynamic feature data set, the middle LSTM layer has 128 hidden units, the output end is connected to a fully connected layer and an output layer, and the output layer directly outputs the overflow concentration prediction result, with 3 feature dimensions set. After dividing the dynamic feature data set (T, Y) constructed in S1 into a training set and a validation set in a ratio of 8:2, the selected model is initially pre-trained. The loss function uses mean squared error, the learning rate of the optimizer is set to 1e-5, the weight decay is set to 1e-5, the batch size is set to 32, and the training rounds are set to 200.

[0106] Step 3: Use the K-means algorithm to cluster the dynamic feature data set (T, Y) obtained in S2. First, determine the K value, and the SSE curve with respect to the K value is as follows: Figure 4 ​As shown, K=4 is determined, and clustering is performed to obtain the clustering result (T, Y, C).

[0107] Step 4: The different working condition data sets are further divided into training subsets and validation subsets, and are also divided according to the ratio of 8:2. Then, the prediction model of different working conditions is fine-tuned through different working condition data sets. Specifically, the error of the pre-trained model in each working condition is evaluated using the validation subset, and then the model parameters are adjusted based on the error analysis result, including learning rate and weight decay adjustment, and considering increasing or decreasing the number of LSTM network layers. Then, the training subset of each working condition is used as input for retraining, and the adjusted learning rate, weight decay coefficient and network structure are used, and the retraining round is set to 100, and the early stopping strategy is set to prevent overfitting.

[0108] Step 5: A multi-working condition prediction model is constructed, and when predicting, the distance of the sample point to be predicted to each cluster center point is calculated to determine the working condition to which the sample belongs, and then the corresponding model is called for accurate prediction.

[0109] Experimental verification:

[0110] The present application carries out the prediction experiment of the overflow concentration and fineness in the grinding and classification process: the data used are collected from the data collected by the distributed control system (DCS) of a certain concentrator in Jiangxi, including 14 parameters such as feed quantity, front water quantity, pump pool liquid level, pulp concentration, ball mill power, feed end oil pressure, pulp flow, cyclone pressure, overflow flow, overflow concentration, overflow fineness +80 mesh, overflow fineness-200 mesh, slurry pump frequency and make-up water quantity, a total of 34000 data sample points, and the sampling interval is 5s.

[0111] As shown in Table 2, the experimental results before and after dynamic characteristic mining for time lag analysis are compared, Figure 5 The prediction result curve and the slope comparison result graph are shown in Table 3, and the prediction results of the overflow fineness +80 mesh in the single prediction model and the multi-working condition prediction model are shown. The multi-working condition prediction model construction strategy based on time lag analysis and clustering fine-tuning of the grinding and classification process proposed in the present application is significantly improved for the prediction of the overflow concentration and fineness in the time lag analysis and clustering fine-tuning working condition model.

[0112] Table 2 Comparison of prediction results before and after TDMI

[0113]

[0114]

[0115] Table 3 Prediction results of the multi-working condition prediction model (overflow fineness +80 mesh)

[0116] Overflow fineness +80 mesh Single-condition model Muti-condition model RMSE 0.0983 0.0817 R2 0.9917 0.9936 HR(1%) 53.06％ 90.43％ HR(0.5%) 23.72％ 79.97％

[0117] In summary, the grinding and classification process multi-working condition prediction model construction strategy based on time delay analysis and clustering fine-tuning proposed in the present application significantly improves the accuracy of overflow concentration fineness prediction in the grinding and classification process, and is especially suitable for strong time delay or variable working condition beneficiation scenarios.

[0118] It should be noted that in the present application, relationship terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations.

[0119] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can make equivalent replacements or changes within the technical range disclosed in the present application according to the technical solution and inventive concept of the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for constructing a multi-condition prediction model for a grinding and classification process, characterized in that, Includes the following steps: S1. Construction of Dynamic Feature Dataset: Based on the historical operating parameter database of the grinding process, a time-delay correlation analysis algorithm is used to compensate for feature time delays, constructing a dynamic feature dataset for overflow concentration regression prediction to address the problem of deteriorating performance due to time delays; specifically including the following: Collect relevant parameters of the grinding and classification process, preprocess the parameters, and construct the original dataset of the mineral processing process. X , Y ), where the input parameter set X ={ x i }, i =1,2,…, m ,Include m A set of parameters that affect the overflow concentration; output parameter set Y ={ y j }, j =1,2,…, n ,Include n One target value to be predicted; N This represents the total number of sample points; Subsequently, a time-delay correlation analysis algorithm was used to accurately determine the time delay relationship between various variables, explore the dynamic characteristics of the grinding process, and construct a dynamic feature dataset for overflow concentration regression prediction by expanding and serializing the dynamic features. T , Y ),in T This represents the input features after dynamic feature mining; The construction process of the dynamic feature dataset is as follows: For the first step of the grinding and classification process i The input time series of the feature pair j The time series correlation coefficients of the output variables are represented as follows: In the formula, x i ( t ) is the input time series X The i The first feature t time; y j ( t + τ (The output time series to be predicted) Y The j The first feature t + τ time; and The first i The first feature input time series and the second feature input time series and the third ... j The mean of the time series of each output variable; N This represents the total number of sample points; N 0 represents the number of observation points in the time series. If its value is greater than the maximum number of delay points, then the time lag between the two time series is the number of time delays corresponding to the moment of maximum correlation, as shown below: For the i For each input feature, take the maximum delay among all corresponding output sequences, and represent it as follows: Using time delay information, the dataset ( X , Y Dynamic feature expansion is performed on the input variable. i The first feature t The serialization process of a sample is represented as follows: Based on the above process, the dataset combines the variable value of each variable at each time point with the variable values ​​of several preceding time points according to the time lag relationship, forming a feature vector containing dynamic information; S2. Model Building and Pre-training: Select a pre-trained model suitable for sequence data processing, and use the dynamic feature dataset obtained in S1 to pre-train the model to obtain an initial prediction model. S3. Construction of multi-condition datasets: The initial dynamic feature dataset is divided using an unsupervised clustering method to construct different condition datasets in order to identify multiple condition features in the grinding and classification process. S4. Model Training: A pre-training-clustering fine-tuning method for training multi-working-condition models is designed. The initial prediction model pre-trained in S2 is retrained using the multi-working-condition dataset obtained in S3 to obtain a multi-working-condition prediction model that adapts to the characteristics of various jobs. The clustering fine-tuning involves making targeted adjustments to the pre-trained model based on the clustering results. Specifically, for each working condition represented by a cluster, the performance of the pre-trained model on the data for that working condition is evaluated, and the error pattern between the model's prediction results and the actual values ​​is analyzed. Based on the obtained error analysis results, the model adaptively adjusts the connection weights and biases between neurons in the neural network model using the backpropagation algorithm. S5. Prediction Execution: Based on the multi-condition prediction model obtained in S4, a condition identification method based on distance calculation is designed. By calculating the distance from the predicted sample point to each cluster center point, the condition to which the sample belongs is determined, and then the corresponding model is called for accurate prediction.

2. The method for constructing a multi-condition prediction model for a grinding and classification process according to claim 1, characterized in that, The pre-trained model mentioned in S2 is a neural network model, a decision tree model, or a combination thereof, which uses a dynamic feature dataset ( T , Y As a training dataset for pre-trained models, it enables models to learn common features and patterns in large-scale data and master the relationships between various parameters in the grinding process.

3. The method for constructing a multi-condition prediction model for a grinding and classification process according to claim 2, characterized in that, The unsupervised clustering method described in S3 is the K-means clustering algorithm, density-based spatial clustering algorithm, or Gaussian mixture model clustering algorithm; the dynamic feature dataset ( T , Y This is used as the input feature of the clustering algorithm to adapt the clustering to dynamic characteristics, and finally obtains the clustering result of each sample. T , Y,C );in C This represents the clustering results.

4. The method for constructing a multi-condition prediction model for a grinding and classification process according to claim 3, characterized in that, S5 specifically includes the following: Determine the operating condition of the sample point to be predicted, and calculate the distance from the sample point to each cluster center. The distance metric should be consistent with the clustering algorithm, using Euclidean distance, expressed as: In the formula, μ k Indicates the first k Cluster centers of each cluster; D new This represents the processed sample points to be predicted. Based on the principle of closest proximity, the sample point is determined to be the working condition represented by the nearest cluster center, as shown below: In the formula, c k Representing the k A cluster; After determining the operating condition, accurate predictions are made using the corresponding predictive model.

5. A multi-condition prediction model for grinding and classification processes, characterized in that, The model is constructed using any one of the methods described in claims 1-4.

6. A computer device, characterized in that, The computer device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set, or instruction set, and the instruction, program, code set, or instruction set is loaded and executed by the processor to implement the method for constructing a multi-condition prediction model for the grinding and classification process as described in any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, at least one program, code set, or instruction set, which is loaded and executed by a processor to implement the multi-condition prediction model construction method for grinding and classification processes as described in any one of claims 1-4.

Images