Multi-working-condition prediction model in ore grinding classification process and construction method thereof

Abstract

The invention discloses an ore grinding classification process-oriented multi-working-condition prediction model and a construction method thereof, and belongs to the technical field of ore dressing. The method comprises the following steps of: firstly, extracting dynamic characteristics by adopting a time-delay correlation analysis algorithm so as to construct a dynamic characteristic data set for predicting overflow concentration and fineness; thirdly, initial pre-training is carried out on the model through total data, and an initial data set is divided into different working condition data sets through an unsupervised clustering method; then, the working condition data sets are utilized to carry out fine tuning and retraining on the pre-trained model, so that the model can adapt to various working condition characteristics. In the prediction stage, the working condition to which a sample belongs is determined by calculating the distance from a to-be-predicted sample point to each clustering center point, and a corresponding model is called for accurate prediction. Compared with the prior art, the method has the advantages that the accuracy of ore grinding grading prediction is improved, and intelligence and stable operation of ore dressing production are promoted.

Description

Technical Field

[0001] The present invention relates to the technical field of ore dressing, and particularly to a multi-condition prediction model for grinding and classification processes based on time-delay analysis and clustering fine-tuning and a construction method thereof. Background Art

[0002] The grinding and classification process is a key link in the ore dressing process, which has significant dynamic characteristics and time lags. In the grinding process, factors such as the amount of material in the mill, the motion state of the grinding medium, and the feed particle size and hardness are constantly changing dynamically. For example, when the feed rate suddenly increases, the amount of material in the mill increases, the load of the ball mill will increase accordingly, and its motor current will also rise rapidly, which reflects the rapid response of the grinding process to changes in input conditions, that is, the dynamic characteristics. However, after the grinding product enters the classification equipment, the classification effect will not appear immediately. There is a certain time delay from when the material enters the hydrocyclone classifier to when the qualified product and unqualified product after classification are accurately separated and output, which is the time lag. This lag is mainly due to the fact that the material needs to undergo a series of complex physical processes such as sedimentation and separation in the classifier, resulting in a time difference between the change of the operating variable and the obvious change of the output variable. Moreover, various interference factors in the grinding and classification system, such as fluctuations in ore properties and the gradual increase in equipment wear, will also continuously affect the dynamic characteristics and time lag characteristics of the system, making the prediction effect of the overflow concentration and fineness of grinding and classification worse.

[0003] Traditional prediction models usually assume simple linearity and static operating conditions, making it difficult to overcome technical bottlenecks when dealing with the complex characteristics of the grinding and classification process. For example, methods based on simple mathematical models set fixed proportional coefficients to relate production parameters, and cannot capture in real time the influence of complex and variable factors such as the amount of material in the mill, the motion state of the grinding medium, the feed particle size and hardness on the grinding and classification effect. When the mill load changes suddenly due to the change of the feed rate, it is easy to cause prediction deviation of the grinding product quality; methods based on time series analysis rely on the time series characteristics of historical data, but because they are based on the assumption of data stationarity, while the actual grinding and classification are affected by factors such as fluctuations in ore properties and equipment wear, and the data shows non-stationary characteristics, it is difficult to effectively adjust the prediction in the face of dynamic changes, affecting the accuracy; static neural network models only consider the static relationship between input and output data, and do not fully explore the time series characteristics and dynamic change laws of the grinding and classification process. When the key parameters in the mill change rapidly, they cannot effectively capture the dynamic information, resulting in large prediction deviations. In addition, traditional methods are also seriously insufficient in dealing with time lags. Methods based on mathematical models and experience often ignore the time delay from when the material enters to when it is output in the classifier based on instant feedback. Methods based on time series analysis and static neural network models also lack effective processing mechanisms, resulting in the prediction result lagging behind the actual operating conditions, affecting production efficiency, causing unstable product quality, and increasing production costs.

[0004] In response to the defects of the above-mentioned traditional prediction models, this patent proposes a multi-condition prediction model for the grinding and classification process based on time-lag analysis and clustering fine-tuning and its construction method. Summary of the Invention

[0005] The purpose of the present invention is to provide a multi-condition prediction model for the grinding and classification process and a construction method thereof to solve the problems mentioned in the background technology; the present invention effectively improves the model's adaptability and prediction accuracy to the time-lag dynamics and multi-conditions of the grinding and classification process, and can ensure the efficient and stable operation of mineral processing production.

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

[0007] A method for constructing a multi-condition prediction model for a grinding and classification process comprises the following steps:

[0008] S1. Dynamic feature data set construction: Based on the historical operating parameter database of the grinding process, a time-lag correlation analysis algorithm is used to compensate for feature time lag and construct a dynamic feature data set for overflow concentration regression prediction to address the problem of poor performance caused by time lag.

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

[0010] S3. Construction of multi-operating condition data sets: The initial dynamic feature data set is divided using an unsupervised clustering method to construct different operating condition data sets to identify various operating condition characteristics in the grinding and classification process;

[0011] S4. Model training: Design a pre-training-clustering fine-tuning multi-operating-condition model training method. Use the multi-operating-condition dataset obtained in S3 to retrain the initial prediction model pre-trained in S2 to obtain a multi-operating-condition prediction model and adapt the model to the characteristics of various tasks.

[0012] 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 the center point of each cluster, the condition to which the sample belongs is determined, and then the corresponding model is called for accurate prediction.

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

[0014] Collect the relevant parameters of the grinding and classification process, pre-process the parameters, and construct the original data set (X, Y) of the mineral processing process, where the input parameter set X = {x i}, where \(i = 1, 2, \ldots, m\), including \(m\) parameters affecting the overflow concentration and fineness; the output parameter set \(Y=\{y\) j}, where \(j = 1, 2, \ldots, n\), including \(n\) target values to be predicted; \(N\) is the total number of sample points;

[0015] Subsequently, through the time-delay correlation analysis algorithm, the time-delay relationship between variables is accurately determined, and the dynamic characteristics of the grinding process are mined. Through the expansion of dynamic characteristics and serialization, a dynamic characteristic data set \((T, Y)\) for overflow concentration and fineness regression prediction is constructed, where \(T\) represents the input features after dynamic characteristic mining;

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

[0017] For the \(i\)-th feature input time series of the grinding and classification process, the correlation coefficient with the \(j\)-th output variable time series is expressed as follows:

[0018]

[0019] In the formula, \(x\) i (t) is the \(i\)-th feature of the input time series \(X\) at the \(t\)-th moment; \(y\) j (t + τ) is the \(j\)-th feature of the output time series \(Y\) to be predicted at the \(t + τ\)-th moment; and are the means of the \(i\)-th feature input time series and the \(j\)-th output variable time series respectively; \(N\) is the total number of sample points; \(N_0\) is the number of observation points of the time series, and its value is greater than the number of the maximum delay points. Then the time delay of the two time series is the number of time delays corresponding to the moment with the maximum correlation, which is expressed as follows:

[0020]

[0021] For the \(i\)-th input feature, take the maximum delay among all its corresponding output sequences, which is expressed as:

[0022]

[0023] Based on the above time-delay information, the data set \((X, Y)\) is dynamically expanded. For the \(t\)-th sample of the \(i\)-th feature of the input variable, its serialization process is expressed as:

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

[0025] Based on the above process, each variable in the dataset combines the variable values at each time point with the variable values at several previous time points 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 combined model thereof. The dynamic feature dataset (T, Y) is used as the training dataset of the pre-trained model, so that the model can learn general features and patterns from large-scale data and master the relationships between various parameters in the grinding process.

[0027] Preferably, the unsupervised clustering method in S3 is the K-means clustering algorithm, the density-based spatial clustering algorithm, or the 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 belonging result (T, Y, C) of each sample is obtained; where C represents the clustering result.

[0028] Preferably, the clustering fine-tuning in S4 is to specifically adjust the pre-trained model according to the clustering result, specifically: for each working condition represented by a cluster, evaluate the performance of the pre-trained model on the data of this working condition respectively, and analyze the error pattern between the model prediction result and the actual value; according to the obtained error analysis result, with the help of the backpropagation algorithm, make the model adaptively adjust the connection weights and biases between the neurons of the neural network model.

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

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

[0031]

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

[0033] According to the nearest distance principle, determine this sample point as the working condition represented by the nearest clustering center point, as shown below:

[0034]

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

[0036] After determining the belonging working condition, perform accurate prediction through the prediction model corresponding to the working condition.

[0037] The present invention further protects a multi-condition prediction model for the grinding and classification process, which is constructed using the above method.

[0038] The present invention further protects a computer device, which 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 above-mentioned method for constructing a multi-condition prediction model for the grinding and classification process.

[0039] The present invention further protects 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, and the instruction, program, code set or instruction set is loaded and executed by a processor to implement the above-mentioned method for constructing a multi-condition prediction model of the grinding and classification process.

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

[0041] (1) Compared with the existing technologies that mostly ignore the time lag characteristics in the grinding and classification process, only consider the static data at the current moment and cannot reflect the dynamic changes of the process, the present invention uses the time lag correlation analysis algorithm to explore the time delay relationship between each variable, thereby effectively dealing with the time lag dynamics in the grinding and classification process. Based on the time delay relationship between each variable, the present invention constructs a dynamic feature data set for overflow concentration regression prediction. By incorporating rich dynamic information, the data set can fully reflect the change law of the grinding and classification process over time, thereby enabling the model to fully learn and adapt to the time lag dynamics of the grinding and classification process, thereby significantly improving the timeliness and accuracy of the prediction.

[0042] (2) The present invention uses a large amount of historical data of the grinding process to perform initial pre-training on the model, so that the model can learn common features and patterns in large-scale data and have a certain knowledge reserve. Then, by fine-tuning the model under different working conditions, a prediction model for multiple working conditions is constructed. The training process of existing prediction models is mostly to directly train and model the entire amount of data, and only perform a training process once. The pre-training-fine-tuning mode of the present invention can first pre-train on a large amount of data, avoiding the blindness of the model in the initial learning, accelerating the learning progress, and at the same time reducing the model's dependence on some specific working condition data.

[0043] (3) Most of the existing prediction methods construct models based on simple linear assumptions and single working conditions, making it difficult to adapt to the complex and variable working conditions in the grinding and classification process. The present invention uses an unsupervised clustering method, takes a dynamic feature dataset as input, divides the grinding and classification process into different working conditions, and makes targeted adjustments to the pre-trained model according to different working condition datasets. When making predictions, first perform working condition identification, and then use the corresponding working condition model for prediction. In this way, the model can better capture the unique features of each working condition, make up for the deficiencies of the pre-trained single model in dealing with complex working conditions, and significantly improve the prediction accuracy and generalization ability of the model under different working conditions. Description of the Drawings

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings involved in the embodiments will now be briefly introduced. Obviously, the drawings in the following description are only schematic illustrations of some embodiments of the present invention, and those skilled in the art can also construct other forms of drawings based on these drawings without creative labor.

[0045] Figure 1 It is a flowchart of a method for constructing a multi-working condition prediction model for the grinding and classification process proposed by the present invention;

[0046] Figure 2 It is a flowchart for constructing a dynamic feature dataset for overflow density and fineness regression prediction proposed in the embodiments of the present invention;

[0047] Figure 3 It is a schematic diagram for dynamic feature extension and serialization in constructing a dynamic feature dataset proposed in the embodiments of the present invention;

[0048] Figure 4 It is an elbow method curve graph proposed in the embodiments of the present invention;

[0049] Figure 5 It is a comparison graph of overflow density and fineness prediction before and after TDMI proposed in the embodiments of the present invention. Detailed Embodiments

[0050] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments.

[0051] The present invention proposes a multi-condition prediction model for the grinding and grading process and a construction method thereof, focusing on the prediction field, and through precise time-lag analysis technology, deeply analyzing the time lag relationship in the grinding and grading process, effectively overcoming the shortcomings of traditional methods in dealing with time lag. At the same time, using clustering fine-tuning technology, the model is targeted and optimized according to different working conditions, successfully solving the problem that traditional methods are difficult to cope with dynamic characteristics, and can improve the accuracy and stability of predictions, better adapt to the complex and changeable actual production needs of the grinding and grading process, and provide strong technical support for the efficient and stable operation of the mineral processing industry. The multi-condition prediction model for the grinding and grading process and the construction method thereof proposed by the present invention are explained below in conjunction with relevant drawings and specific examples. The specific content is as follows.

[0052] Example 1:

[0053] like Figure 1 As shown, the present invention proposes a method for constructing a multi-condition prediction model for a grinding and classification process, comprising:

[0054] S1. Utilizing massive historical parameter data of the grinding process and advanced time-lag correlation analysis algorithms, a dynamic feature dataset for overflow concentration and fineness regression prediction is constructed. Specifically, the following contents are included:

[0055] S1.1. Data collection and preprocessing: Comprehensively collect various parameters in the on-site distributed control system of the grinding and grading process of the ore dressing plant, including but not limited to feed rate, mill power, classifier overflow concentration, particle size and other data. Clean the collected data, remove outliers and noise data, and ensure the accuracy and reliability of the data. At the same time, normalize the data to make data with different parameters comparable. Based on this, construct the original data set (X, Y) of the overflow concentration prediction of the grinding and grading process with a dimension of (m+n)×N. Among them, the input parameter set X={x i}, i = 1, 2, ..., m, including m parameters that affect the overflow concentration, 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: Apply advanced time-lag correlation analysis algorithms to (X, Y) constructed in S11, such as the time-lag analysis method based on correlation coefficient, to conduct a comprehensive analysis of various variables in the grinding and classification process. By calculating the correlation coefficients between different variables, the time delay relationship between the variables can be accurately determined. In this way, the dynamic characteristics of the grinding process are mined, and the dynamic feature data set (T, Y) for overflow concentration regression prediction is constructed by expanding and serializing the dynamic features, where T represents the input feature matrix after dynamic feature mining. The specific process is as follows: Figure 2 shown.

[0057] The time lag analysis algorithm is as follows: By measuring the correlation between two time series under different time delay states and accurately calculating the correlation coefficients at different time lags, the time lag that maximizes the correlation coefficient is found, and this specific time lag is the optimal time delay between the two time series. The correlation coefficient between the $i$-th characteristic input time series and the $j$-th output variable time series of the grinding and classification process is expressed as follows:

[0058]

[0059] In the formula, $x$ i (t) is the $i$-th characteristic of the input time series $X$ at the $t$-th moment, and $y$ j (t + τ) is the $j$-th characteristic of the output time series $Y$ to be predicted at the $t + τ$-th moment. and are the means of the $i$-th characteristic input time series and the $j$-th output variable time series respectively. $N_0$ is the number of observation points of the time series, and its value should be greater than the number of maximum delay points. Then the time lag between the two time series is the number of time delays corresponding to the moment of maximum correlation, which is expressed as follows:

[0060]

[0061] where represents the number of delays of the $i$-th input characteristic to the $j$-th output prediction variable. To ensure that each input sequence retains the time lag dynamic information for all output sequences, finally, the maximum delay of each input sequence to all output sequences is taken, which is expressed as:

[0062]

[0063] Subsequently, based on the time lag information mined through dynamic features, the dataset $(X, Y)$ is dynamically feature-expanded, so that each data sample not only contains the variable value at the current moment but also incorporates the variable values at historical moments that are strongly correlated with it. By adding the variable values at historical moments with strong correlation, the model can better learn and adapt to the time lag dynamics of the grinding and classification process during training. Specifically, for each input variable, according to the calculated maximum delay $\tau$ i , the variable values at the previous $\tau$ i time points are added as new feature dimensions to the original data sample. For example, if the maximum delay of the feed rate to the overflow concentration is 3 time points, then when constructing the dynamic feature dataset, in addition to the feed rate value at the current moment, the feed rate values at the previous 3 time points will also be included in the data sample. Finally, the dynamic feature dataset $(T, Y)$ for overflow fineness regression prediction is constructed through serialization. For the sample corresponding to the $t$-th moment of the $i$-th characteristic of the input variable, the process is expressed as:

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

[0065] This data set combines the variable values of each variable at each time point with the variable values of several previous time points according to the time lag relationship, forming a feature vector containing dynamic information. This unique modeling method 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 make up for the defect that traditional methods only consider the static data at the current moment and cannot reflect the dynamic changes of the process. The specific implementation process is as Figure 3 shown. For the convenience of drawing, x i (t) and y(t) are respectively defined as and y t .

[0066] S2. Use the full historical data of the grinding process to initially pre-train the model to obtain a pre-trained model. The models used for initial pre-training include but are not limited to neural network models or their combined models. The dynamic feature data set obtained from S1 is used as the training data set for the pre-trained model. Pre-training enables the model to learn general features and patterns from large-scale data, thus having a certain knowledge reserve and laying a foundation for subsequent fine-tuning for different working conditions. Through pre-training, the model initially masters the relationships between various parameters in the grinding process, such as the potential relationships between ore properties, mill speed, feed rate and other parameters and the grinding effect, which helps to improve the efficiency and accuracy of the subsequent training of the model and enables it to adapt to the data characteristics under different working conditions more quickly. The specific contents are as follows:

[0067] Use the Long Short-Term Memory network (LSTM), a neural network model suitable for processing sequence data, as the pre-trained model. This model has significant advantages in learning the complex patterns and long-term trends of sequence data. Its unique structure design can effectively capture the long-term dependencies in the data and process the inflow, outflow and forgetting of information through a gating mechanism, overcoming the limitations of traditional neural networks in processing data with long time spans. In the pre-training stage, first further clean and organize the full historical data of the grinding process to ensure the quality and consistency of the data. Divide the dynamic feature data set (T, Y) constructed by S1 into a training set and a validation set according to a certain ratio, and then conduct initial pre-training on the selected model. The loss function selected in its pre-training process is expressed as follows:

[0068]

[0069] wherein represents the predicted output value of the model, and y i (t) represents the actual value in the dataset. In each round of training, the model predicts the indicators related to the grinding effect, such as the overflow concentration and fineness, based on the input dynamic feature data. By calculating the loss function between the predicted value and the true value, the parameters of the model are updated using the backpropagation algorithm. After multiple rounds of iterative training, the model gradually learns the general features and patterns among various parameters in the grinding process, and initially masters the potential relationships between parameters such as ore properties, mill speed, and feed rate and the 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 using the validation set. 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 the pre-trained model.

[0070] S3. Apply unsupervised clustering methods, such as the K-means clustering algorithm, density-based spatial clustering algorithm, or Gaussian mixture model clustering algorithm, etc., to divide the initial dataset. Use the dynamic feature dataset as the input feature of the clustering algorithm to make the clustering better adapt to the dynamic characteristics of the grinding and classification process. Through clustering analysis, the grinding and classification process is divided into different working conditions, and different working condition datasets are obtained. Each dataset represents a specific working condition. The specific contents are as follows:

[0071] S3.1 Implementation of the clustering algorithm based on the dynamic feature dataset is as follows:

[0072] In view of the characteristics of the data in the grinding and classification process and the clustering requirements, the K-means unsupervised clustering method is used to cluster the above dynamic feature dataset. This algorithm can converge quickly and assigns data by calculating the distance from the data points to the centers of each cluster. Determine the number of clusters K of the clustering as follows. By calculating the sum of squared errors within classes (SSE) under different K values, the expression is as follows:

[0073]

[0074] In the formula, K represents the number of clusters of the clustering, and C k represents the k-th 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 i-th sample point in D that belongs to the k-th cluster, and N d is the number of sample points of the dynamic dataset constructed after time-delay analysis, and μ k represents the clustering center of the k-th cluster, which can be obtained by the following formula:

[0075]

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

[0077] S3.2 performs unsupervised clustering on the grinding and classification process under different working conditions as follows:

[0078] After determining the clustering algorithm and the K value, obtain the dynamic feature dataset (T, Y) from S2 as the input features of the clustering algorithm. These dynamic feature datasets contain the time delay information and dynamic change characteristics in the grinding and classification process. Using this as the input can make the clustering process better adapt to the dynamic characteristics of the grinding and classification process.

[0079] By minimizing the sum of the squares of the distances from all data points to the centers of their respective clusters, the objective function J is defined as the sum of squared errors within the classes, that is:

[0080]

[0081] Finally, obtain the clustering membership results (D, C) for each sample, where C k is the k-th cluster in the clustering result C.

[0082] S4. Retrain the pre-trained model using the described different working condition datasets to make the model adapt to the characteristics of various working conditions; specifically, it includes 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 datasets into training subsets and validation subsets, for example, divide them according to a 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. Use the mean squared error to measure the performance of the pre-trained model on each working condition dataset. For each working condition dataset, calculate the MSE of the pre-trained model respectively, and analyze the error pattern between the model prediction results and the actual values. For example, observe whether there is a systematic bias in the error, what is the fluctuation range of the error, etc. Through the analysis of the error pattern, the deficiencies of the model under different working conditions can be deeply understood, 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 under a certain working condition are generally high, the connection weights of the relevant neurons can be appropriately reduced; when the predicted values fluctuate greatly, the training process of the model can be stabilized by adjusting the learning rate and increasing the regularization term (such as L1 or L2 regularization). In addition, the structure of the network can also be considered for adjustment, 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 are as follows:

[0088] Training process: Use the training subsets of each working condition to retrain the model with adjusted parameters. During the training process, adopt a suitable optimization algorithm and set appropriate number of training epochs and batch size to update the model parameters to ensure that the model can fully learn the characteristics of the working condition datasets.

[0089] Monitoring and adjustment: During the retraining process, use the validation subset to monitor the performance of the model in real time. If it is found that the performance of the model on the validation subset shows a downward trend, such as the validation error starts to rise, it indicates that the model may have overfitting phenomenon. At this time, the training strategy needs to be adjusted in time, such as stopping training early, increasing the regularization strength, etc.

[0090] S5. When making predictions, determine the working condition to which the sample to be predicted belongs by calculating the distance from the sample point to be predicted to the center points of each cluster, and then call the corresponding model for accurate prediction, which specifically includes the following:

[0091] S5.1. Distance calculation and working condition determination are as follows:

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

[0093]

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

[0095]

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

[0097] S5.2. Call the corresponding model for prediction, which is as follows:

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

[0099] Example 2:

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

[0101] Step 1: Collect relevant parameters of the on-site grinding and classification process in the ore dressing plant, including 14 variables such as feed rate, water feed rate, and overflow concentration, with a total of 34,000 sample data in full, as shown in Table 1 for details. Perform preprocessing on the collected data, including data cleaning (removing outliers, missing values, etc.), data normalization (unifying the data to the same scale range), etc., to construct the original data set (X, Y) of the ore dressing process. In this example, X = [x1, x2, …, x 11 , Y = [y1, y2, y3]. Through the time-delay analysis algorithm based on the correlation coefficient, the time-delay characteristics of the grinding and classification process are mined, as shown in Table 1 specifically.

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

[0103] variable variable name unit Number of delay points (τ) Delay time (min) <![CDATA[x1]]> Feed rate t / d 0 0 <![CDATA[x2]]> Water supply t / d 0 0 <![CDATA[x3]]> Pump sump level cm 60 5 <![CDATA[x4]]> Sizing concentration ％ 96 8 <![CDATA[x5]]> Ball mill power kwh 96 8 <![CDATA[x6]]> Feed end oil pressure kpa 18 1.5 <![CDATA[x7]]> Sizing flow <![CDATA[m 3 / h]]> 204 17 <![CDATA[x8]]> Cyclone pressure mPa 180 15 <![CDATA[x9]]> Overflow flow <![CDATA[m 3 / h]]> 288 24 <![CDATA[x 10 > Slurry pump frequency hz 252 21 <![CDATA[x 11 > Add water <![CDATA[m 3 / h]]> 180 15 <![CDATA[y1]]> Overflow concentration ％ 288 24 <![CDATA[y2]]> Overflow fineness + 80 mesh ％ 300 25 <![CDATA[y3]]> Overflow fineness-200 mesh ％ 312 26

[0104] After that, according to Figure 3 Dynamically expand the features of the original data set and serialize it to construct a dynamic feature data set (T, Y) for overflow concentration and fineness regression prediction.

[0105] Step 2: Use the LSTM model as the initial pre-trained model, and its model structure is designed as follows: The number of neurons in the input layer corresponds to 14 feature dimensions of the dynamic feature data set. The middle LSTM layer is set with 128 hidden units. The output end is connected to a fully connected layer and an output layer, and the output layer directly outputs the prediction results of overflow concentration and fineness, with 3 feature dimensions set. Divide the dynamic feature data set (T, Y) constructed in Step S1 into a training set and a validation set according to 8:2, and then perform initial pre-training on the selected model. The loss function uses the mean squared error, the learning rate of the optimizer is set to 1e-5, the weight decay is 1e-5, the batch size is set to 32, and the number of training epochs is 200.

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

[0107] Step 4: Further divide the datasets under different working conditions into training subsets and validation subsets, also divided according to the ratio of 8:2. Then, fine-tune the prediction models for different working conditions through different working condition datasets. The specific operations are as follows: First, use the validation subset to evaluate the errors of the pre-trained model under each working condition. Subsequently, adjust the model parameters based on the results of error analysis, including adjusting the learning rate and weight decay, and considering increasing or decreasing the number of LSTM network layers. Then, use the training subsets of each working condition as inputs for retraining, using the adjusted learning rate, weight decay coefficient, and network structure. Set the number of retraining rounds to 100, and at the same time set the early stopping strategy to prevent overfitting.

[0108] Step 5: Construct a multi-working condition prediction model. When making predictions, determine the working condition to which the sample belongs by calculating the distance from the sample point to be predicted to the center points of each cluster, and then call the corresponding model for accurate prediction.

[0109] Experimental verification:

[0110] The present invention conducts a prediction experiment on the overflow concentration and fineness in the grinding and classification process: The data used comes from the data collected by the Distributed Control System (DCS) at a concentrator site in Jiangxi, including 14 parameters such as feed rate, front water supply, pump sump level, pulp concentration, ball mill power, oil pressure at the feed end, pulp flow rate, cyclone pressure, overflow flow rate, overflow concentration, overflow fineness +80 mesh, overflow fineness -200 mesh, slurry pump frequency, and makeup water volume, with a total of 34,000 data sample points and a sampling interval of 5 s.

[0111] Table 2 shows the comparison of the experimental results before and after the dynamic feature mining for time delay analysis. Figure 5 It is a comparison result graph of the prediction result curve and the oblique line. Table 3 shows the prediction results of the overflow fineness +80 mesh in the single prediction model and the multi-working condition prediction model. The strategy for constructing a multi-working condition prediction model for the grinding and classification process based on time delay analysis and clustering fine-tuning has achieved significant improvements in the prediction of overflow concentration and fineness for the time delay analysis and clustering fine-tuning working condition models.

[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+80mesh 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 construction strategy of the multi-condition prediction model for the grinding and classification process based on time-delay analysis and clustering fine-tuning proposed by the present invention significantly improves the accuracy of predicting the overflow concentration and fineness in the grinding and classification process, and is particularly suitable for beneficiation scenarios with strong time-delay or variable working conditions.

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

[0119] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A method for constructing a multi-condition prediction model for the grinding and classification process, characterized in that, It includes the following steps: S1. Dynamic feature dataset construction: Based on the historical operation parameter database of the grinding process, a time-delay correlation analysis algorithm is used for feature time-delay compensation to construct a dynamic feature dataset for overflow concentration and fineness regression prediction, so as to solve the problem of deteriorated effect caused by time delay; S2. Model construction 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. Multi-condition dataset construction: Use an unsupervised clustering method to divide the initial dynamic feature dataset to construct different condition datasets to identify various condition characteristics in the grinding and classification process; S4. Model training: Design a multi-condition model training method of pre-training - clustering fine-tuning, and use the multi-condition dataset obtained in S3 to retrain the initial prediction model pre-trained in S2 to obtain a multi-condition prediction model and make the model adapt to the characteristics of various operations; S5. Prediction execution: Based on the multi-condition prediction model obtained in S4, design a condition identification method based on distance calculation. By calculating the distance from the sample point to be predicted to the center points of each cluster, determine the condition to which the sample belongs, and then call the corresponding model for accurate prediction.

2. The method for constructing a multi-condition prediction model for the grinding and classification process according to claim 1, wherein The specific content of S1 is as follows: Collect relevant parameters of the grinding and classification process. After preprocessing the parameters, construct the original data set (X, Y) of the ore dressing process. Among them, the input parameter set X = {x i}, i = 1, 2, …, m, includes m parameters affecting the overflow concentration and fineness; the output parameter set Y = {y j}, j = 1, 2, …, n, includes n target values to be predicted; N is the total number of sample points; Subsequently, through the time-delay correlation analysis algorithm, accurately determine the time-delay relationship between variables, mine the dynamic characteristics of the grinding process, and construct a dynamic feature dataset (T, Y) for overflow concentration and fineness regression prediction through dynamic feature expansion and serialization, where T represents the input feature after dynamic feature mining; The construction process of the dynamic feature dataset is specifically as follows: The correlation coefficient between the i-th feature input time series and the j-th output variable time series of the grinding and classification process is expressed as follows: where x i (t) is the i-th feature of the input time series X at the t-th moment; y j (t + τ) is the j-th feature of the time series Y to be predicted at the t + τ-th moment; and are the means of the i-th feature input time series and the j-th output variable time series respectively; N is the total number of samples; N0 is the number of observation points of the time series, and its value is greater than the number of the maximum delay points. Then the time lag of the two time series is the number of time delays corresponding to the moment with the maximum correlation, which is expressed as follows: For the i-th input feature, take the maximum delay in all its corresponding output sequences, which is expressed as: Through the above time-delay information, perform dynamic feature expansion on the dataset (X, Y). For the t-th sample of the i-th feature of the input variable, its serialization process is expressed as: x i (t) → [x i (t), x i (t - 1),..., x i (t - τ i )] Based on the above process, combine the variable values at each time point of each variable with the variable values at several previous time points according to the time-delay relationship to form a feature vector containing dynamic information.

3. A method for constructing a multi-condition prediction model for the grinding and classification process according to claim 2, characterized in that The pre-trained model in S2 is a neural network model, a decision tree model or a combined model thereof. The dynamic feature dataset (T, Y) is used as the training dataset of the pre-trained model, so that the model can learn general features and patterns from large-scale data and master the relationships between various parameters in the grinding process.

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

5. The method for constructing a multi-condition prediction model for the grinding and classification process according to claim 4, characterized in that, The clustering fine-tuning described in S4 is to make targeted adjustments to the pre-trained model according to the clustering results, specifically referring to: for each working condition represented by a cluster, respectively evaluate the performance of the pre-trained model on the data of this working condition, and analyze the error pattern between the model prediction result and the actual value; according to the obtained error analysis result, with the help of the backpropagation algorithm, make the model adaptively adjust the connection weights and biases between the neurons of the neural network model.

6. The method for constructing a multi-condition prediction model for the grinding and classification process according to claim 5, wherein The specific content of S5 is as follows: Determine the working condition to which the sample point to be predicted belongs, and calculate the distance from this sample point to the center points of each cluster; the distance metric is consistent with the clustering algorithm, and the Euclidean distance metric is used, and the expression is: where μ k represents the clustering center of the k-th cluster; D new represents the sample point to be predicted after processing; According to the nearest distance principle, determine that this sample point belongs to the working condition represented by the nearest cluster center point, which is expressed as: where c k represents the k-th cluster; After determining the working condition to which it belongs, perform accurate prediction through the prediction model corresponding to this working condition.

7. A multi-condition prediction model for the grinding and classification process, characterized in that, The model is constructed by the method described in any one of claims 1-6.

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

9. A computer-readable storage medium, characterized in that, At least one instruction, at least one program, a code set or an instruction set is stored in the computer-readable storage medium. The instruction, program, code set or instruction set is loaded and executed by the processor to implement the method for constructing a multi-working condition prediction model for the grinding classification process described in any one of claims 1-7.

Images