A method for automatically controlling parameters to optimize the overflow particle size range of grinding

Abstract

The present invention discloses a method for automatically controlling parameters to optimize the overflow particle size interval of ore grinding, and belongs to the technical field of ore grinding particle size control. In the production of the present invention, production data is obtained through sensors and sent to the central control computer; in the early stage, a deep forest model is trained through long-term data accumulation and stored in the central control computer; the overflow particle size value obtained by the sensor is judged, and for the case that it is not in the specified interval, the grid method is used to obtain the optimal parameters suitable for each device in the current production state under the deep forest model, and the parameter adjustment results of the sending module are fed back to the production system, and finally the parameter adjustment is completed through the control module. The present invention realizes the automatic adjustment and correction of parameters in the ore grinding process, which greatly improves the stability and production efficiency of the overflow particle size in ore grinding.

Description

Technical Field

[0001] The invention relates to the technical field of grinding particle size control, in particular to a method for automatically controlling parameters to optimize the grinding overflow particle size range. Background Art

[0002] Grinding operation is the most critical link in the entire mineral processing production process, and it plays a connecting role. The distribution of overflow particle size in grinding operation is often a key indicator of mineral processing production. Therefore, how to stably detect the distribution of overflow particle size is the key to improving enterprise production efficiency. With the rapid development of science and technology, although some expensive particle size analyzers can be used to perform online detection of overflow particle size distribution, in the harsh environment of actual production, such particle size analyzers often fail and cannot reflect the operating status of the overall production system in real time.

[0003] At present, there are generally two methods to solve the problem of measuring grinding particle size. One is to develop a new type of particle size analyzer based on the development ideas of traditional detection technology, and realize the direct measurement of grinding particle size in the form of hardware. Although this particle size analyzer can obtain more accurate measurement results, the initial equipment investment is large and the later maintenance cost is high, which makes it impossible for some concentrators to configure such detection equipment in large quantities; secondly, in the harsh environment of actual production, the existing particle size analyzer has poor anti-interference ability, instrument failures occur from time to time, and cannot reflect the operating status of the overall production system in real time, which has certain limitations. Another detection method adopts the idea of ​​indirect measurement, using the detection information that is easy to obtain in the production process, and realizing the estimation of overflow particle size distribution through artificial intelligence technology. In recent years, the soft measurement technology that has emerged in the field of process detection is a concentrated embodiment of this idea.

[0004] The soft measurement research of grinding particle size has always been a popular research direction. In 2017, Zhang Yungang et al. proposed to establish a soft measurement model of grinding particle size using radial basis function neural network, and verified the feasibility of this method in practical applications. In 2020, Zhang Yan et al. proposed a method of establishing a soft measurement model of grinding particle size by improving the gray wolf optimization algorithm and optimizing the least squares support vector machine. The experimental results show that the improved soft measurement model not only has good prediction effect, but also has high optimization accuracy. In 2022, Li Fenfang et al. proposed a method to optimize the TS fuzzy neural network based on particle swarm algorithm to establish a soft measurement model of grinding particle size. The experimental results show that the model of optimizing the TS fuzzy neural network has a better prediction effect.

[0005] However, among the many modeling methods, how to establish a particle size distribution prediction model with short training time, high prediction accuracy and good generalization performance is a difficult problem. Although the machine learning prediction method has obvious advantages, there are still many problems, such as the simple structure of the traditional model, poor interpretability for overflow particle size prediction, and low prediction accuracy. The neural network model has a complex structure, but it has too many hyperparameters, complex training, slow convergence, and is prone to overfitting. In addition, in actual scenarios, how to use the prediction model to perform fast and effective soft measurement and perform certain feedback control on the production process is also a problem that needs to be solved. Summary of the invention

[0006] The purpose of the present invention is to provide a method for automatically controlling parameters to optimize the overflow particle size range of grinding. The method is easy to operate, realizes automatic adjustment and correction of parameters in the grinding process, greatly improves the stability of overflow particle size in grinding and production efficiency, so as to solve the problems raised by the above-mentioned background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A method for automatically controlling parameters to optimize the grinding overflow particle size range comprises the following steps:

[0009] Step 1: Install the PLC terminal unit that needs to collect the corresponding variable data. The main collected data information includes: belt feed volume, belt feeder frequency, feeder start status, ball mill power, grinding sound, current, water supply, circular vibration screen current, pump pool water supplement detection value, pump pool liquid level, hydrocyclone feed pressure, feed flow, feed concentration, underflow pump frequency feedback value, crusher current, magnetic separator current, frequency, high-pressure roller mill current, elutriator discharge valve position value, auxiliary water valve position value, overflow detection value, underflow concentration, water inlet valve position value, and overflow particle size of hydrocyclone and other production information;

[0010] Step 2: During the production process, the installed PLC terminal unit is used to collect and save data, and relevant data preprocessing operations and feature selection are performed on the data to improve data quality;

[0011] Step 3: Select the processed data sets, use the deep forest algorithm to train these data sets and generate corresponding prediction models;

[0012] Step 4: During the production process, the central control computer obtains the real-time monitoring data of each PLC terminal and stores the data in chronological order; at the same time, it is determined whether the overflow particle size is within the normal range. If the value is normal, wait for the next interval judgment; if not, the control parameters need to be adjusted;

[0013] Step 5: When the overflow particle size value is abnormal, the generated deep forest prediction model is used to set the overflow particle size target value to the specified value, and the grid search method is used to obtain the set values ​​of all control parameters in turn;

[0014] Step 6: Send the specific control values ​​to the PLC controller through the sending module, so that the corresponding parameters can be set to automatically control the grinding overflow particle size to the specified range.

[0015] Furthermore, the specific process of data preprocessing for the original data in step 2 is as follows: in order to solve the problem of missing data caused by PLC terminal failure or unexpected data transmission in the actual production environment, linear interpolation is used to supplement the missing values; for the missing point P generated at time t t The missing value Y on (X,Y) can be taken as the value point P near time t. t-a (X 1 ,Y 1 ) and P t+b (X 2 ,Y 2 ), then the corresponding missing values ​​can be calculated and filled, and the calculation formula is as follows:

[0016]

[0017] For outliers in the data set, a detection method based on the Raida criterion is used; first, the overall trend of the data is measured, and then the mean is calculated based on the data after the trend is removed. and standard deviation σ, their formulas are as follows:

[0018]

[0019]

[0020] Then calculate the residual v of each data i , the formula is as follows:

[0021]

[0022] Finally, by passing the residual v i Compared with the range threshold 3σ, if |v i |≥3σ, then the corresponding data x i If it is judged as an outlier, it is corrected by replacing the sample mean;

[0023] For the data after data preprocessing, principal component analysis is used to reduce the dimension so as to select appropriate features for retention. The main process of the algorithm is as follows:

[0024] First, set the data matrix x of n samples and p dimensions as follows:

[0025]

[0026] Among them, x i =(x i1 x i2 ...x ip ) T , i=1,2,...,n

[0027] The integrated k variables can be expressed as:

[0028]

[0029] Where k = 1, 2, 3, ..., n

[0030] The above formula can be expressed in matrix form as G=AX, as follows:

[0031]

[0032] Among them, A is the principal component coefficient matrix, and the model is generally required to meet the following three basic conditions:

[0033] (1) G i and G j Uncorrelated, where i≠j, i, j∈(1, 2, ..., n)

[0034] (2) Var(G 1 )>Var(G 2 )>…Var(G i ), i∈(1, 2, ..., n)

[0035] (3)a m1 2 +a m2 2 +…+a mn 2 =1, m=1, 2, ..., n

[0036] Among them, G 1 is the first principal component, G 2 is the second principal component, …, G n is the nth principal component;

[0037] Generally speaking, when c (c<n) principal components are taken, the cumulative contribution rate of c is required to reach at least 85%; finally, based on the results of principal component analysis, the group of variables with a cumulative contribution rate of more than 85% is taken as the final characteristic variable.

[0038] Furthermore, the specific formation process of the deep forest in step 3 is as follows: first, the data set D is read, and the data set D is divided into training set D according to a certain ratio. 1 and the test set T 1 , D 1 ∈R l×m , where 1 represents the training set D 1 The sample size of m represents the number of feature variables; then initialize the relevant parameters, important parameters such as the sliding window size d i The number of decision trees n in each forest will have a greater impact on model generation; then use a size of d i The sliding window of the training set D 1 Split and generate feature vectors gv of different sizes i , Then the sliding window d i The extracted feature vector gv i They are used to train the random forest and completely random forest models respectively, and then generate two different local feature vectors, which are then concatenated to generate the transformation vector g. i , The vector g i For training the cascade forest model, g i After the nth cascade forest i The level generates an augmented vector Then the augmented vector Concatenate with the original transformation vector to form a feature vector Calculate the eigenvector When the error gradually decreases, the current feature vector is used as the input of the next layer of joint forest; when the error does not decrease for three consecutive layers, the training process is terminated and the cascade forest level is determined; the last layer is used as the evaluation layer to calculate the average of all forest predictions, which is the final deep forest regression prediction result.

[0039] Furthermore, the specific process of adjusting the control parameters using the generated prediction model in step 5 is as follows: first, the number of input variables is set to 1, the importance of all control parameters is sorted in descending order, and the control parameter variable a with the greatest influence is selected. 1, added to the set M; for the variables in the set M, the grid search method is used to explore different values ​​and corresponding overflow particle size changes within the range of the parameters in the set M. At this time, the variables outside the set M (i.e., other unselected control parameters) remain fixed; if the overflow particle size prediction result reaches the normal range during the grid search process, the control parameters are saved and the corresponding machine equipment is adjusted based on feedback; otherwise, the variables in M ​​are replaced with the next parameter to continue the grid search; during the feedback adjustment process, the number of variables in M ​​is guaranteed to be consistent with the number of input variables and all possible selection situations are iterated; if the search cannot be completed under the current number of input variables, the number of input variables is increased by 1 until a suitable parameter is found for feedback adjustment.

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

[0041] The present invention collects relevant variable data in the production process through a PLC terminal unit, and then combines data cleaning and principal component analysis to improve data quality, thereby generating a basic prediction model. In the subsequent production process, the prediction model is used in real time to predict the overflow particle size in the future. The abnormal prediction results are cumulatively grid searched in a certain control parameter order. If the overflow particle size prediction result generated by the prediction model can reach the normal range in this process, the control parameters are recorded and feedback is adjusted; if it cannot reach the normal range, the control parameter variables are replaced in turn, and the number of input variables is increased, so as to perform a new round of search until the effective control parameter adjustment is completed. The present invention can effectively select a variety of variables in grinding production, and can accurately predict the overflow particle size change process in the future after effectively improving the data quality, so that an iterative grid search can be used to quickly find a set of control parameters before the overflow particle size is abnormal, thereby ensuring the normal operation of the operation process. Compared with the traditional direct detection using instruments, the present invention has lower deployment costs and higher stability, greatly saves human resources, and is widely applicable to various grinding scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of the system equipment of the present invention;

[0043] Figure 2 The present invention is a grinding flow chart;

[0044] Figure 3 It is a schematic diagram of the central control structure of the system of the present invention;

[0045] Figure 4 It is a schematic diagram of a model of the present invention;

[0046] Figure 5 This is a comparison chart of the absolute errors predicted by the overflow particle size method of the present invention. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] See also Figure 1-5 In an embodiment of the present invention, a method for automatically controlling parameters to optimize the grinding overflow particle size range is provided, comprising the following steps:

[0049] Step 1: Install the PLC terminal unit that needs to collect the corresponding variable data. The main collected data information includes: belt feed rate, belt feeder frequency, feeder start status, ball mill power, grinding sound, current, water supply, circular vibration screen current, pump pool water supplement detection value, pump pool liquid level, hydrocyclone feed pressure, feed flow, feed concentration, underflow pump frequency feedback value, crusher current, magnetic separator current, frequency, high-pressure roller mill current, elutriator discharge valve position value, auxiliary water valve position value, overflow detection value, underflow concentration, water inlet valve position value, and overflow particle size of hydrocyclone and many other production information. The grinding process is as follows: Figure 2 As shown;

[0050] Step 2: During the production process, the installed PLC terminal unit is used to collect and save data, and relevant data preprocessing operations and feature selection are performed on the data to improve data quality;

[0051] The specific process of data preprocessing for the original data in step 2 is as follows: In order to solve the problem of missing data caused by PLC terminal failure or unexpected data transmission in the actual production environment, linear interpolation is used to supplement the missing values; for the missing point P generated at time t t The missing value Y on (X,Y) can be taken as the value point P near time t. t-a (X 1 ,Y 1 ) and P t+b (X 2 ,Y 2 ), then the corresponding missing values ​​can be calculated and filled, and the calculation formula is as follows:

[0052]

[0053] For outliers in the data set, a detection method based on the Raida criterion is used; first, the overall trend of the data is measured, and then the mean is calculated based on the data after the trend is removed. and standard deviation σ, their formulas are as follows:

[0054]

[0055]

[0056] Then calculate the residual v of each data i , the formula is as follows:

[0057]

[0058] Finally, by passing the residual v i Compared with the range threshold 3σ, if |v i |≥3σ, then the corresponding data x i If it is judged as an outlier, it is corrected by replacing the sample mean;

[0059] For the data after data preprocessing, principal component analysis is used to reduce the dimension so as to select appropriate features for retention. The main process of the algorithm is as follows:

[0060] First, set the data matrix x of n samples and p dimensions as follows:

[0061]

[0062] Among them, x i =(x 11 , x i2 , ..., x ip ) T , i=1,2,...,n

[0063] The integrated k variables can be expressed as:

[0064]

[0065] Where k = 1, 2, 3, ..., n

[0066] The above formula can be expressed in matrix form as G=AX, as follows:

[0067]

[0068] Among them, A is the principal component coefficient matrix, and the model is generally required to meet the following three basic conditions:

[0069] (1) G i and G j Uncorrelated, where i≠j, i, j∈(1, 2, ..., n)

[0070] (2) Var(G 1 )>Var(G 2 )>…Var(G i), i∈(1, 2, ..., n)

[0071] (3)a m1 2 +a m2 2 +…+a mn 2 =1, m=1, 2, ..., n

[0072] Among them, G 1 is the first principal component, G 2 is the second principal component, …, G n is the nth principal component;

[0073] Generally speaking, when c (c < n) principal components are taken, the cumulative contribution rate of c is required to reach at least 85%; finally, according to the results of principal component analysis, the variable group with a cumulative contribution rate of more than 85% is taken as the final characteristic variable;

[0074] Step 3: Select the processed data sets, use the deep forest algorithm to train these data sets and generate corresponding prediction models;

[0075] In step 3, the specific formation process of the deep forest is as follows: first read the data set D and divide the data set D into training sets D according to a certain ratio 1 and the test set T 1 , D 1 ∈R l×m , where l represents the training set D 1 The sample size of m represents the number of feature variables; then initialize the relevant parameters, important parameters such as the sliding window size d i The number of decision trees n in each forest will have a greater impact on model generation; then use a size of d i The sliding window of the training set D 1 Split and generate feature vectors gv of different sizes i , Then the sliding window d i The extracted feature vector gv i They are used to train the random forest and completely random forest models respectively, and then generate two different local feature vectors, which are then concatenated to generate the transformation vector g. i , The vector g i For training the cascade forest model, g i After the nth cascade forest i The level generates an augmented vector Then the augmented vector Concatenate with the original transformation vector to form a feature vector Calculate the eigenvector When the error gradually decreases, the current feature vector is used as the input of the next layer of joint forest; when the error does not decrease for three consecutive layers, the training process is terminated and the cascade forest level is determined; the last layer is used as the evaluation layer to calculate the average of all forest predictions, which is the final deep forest regression prediction result.

[0076] Step 4: During the production process, the main control computer obtains the real-time monitoring data of each PLC terminal and stores the data in chronological order; at the same time, it is determined whether the overflow particle size is within the normal range. If the value is normal, wait for the next interval judgment; if it is abnormal, the control parameters need to be adjusted;

[0077] Step 5: When the overflow particle size value is abnormal, the generated deep forest prediction model is used to set the overflow particle size target value to the specified value, and the grid search method is used to obtain the set values ​​of all control parameters in turn;

[0078] The specific process of adjusting the control parameters using the generated prediction model in step 5 is as follows: First, the number of input variables is set to 1, and the importance of all control parameters is sorted in descending order, and the control parameter variable a with the greatest influence is selected. 1 , added to the set M; for the variables in the set M, the grid search method is used to explore different values ​​and corresponding overflow particle size changes within the range of the parameters in the set M. At this time, the variables outside the set M (i.e., other unselected control parameters) remain fixed; if the overflow particle size prediction result reaches the normal range during the grid search process, the control parameters are saved and the corresponding machine equipment is adjusted based on feedback; otherwise, the variables in M ​​are replaced with the next parameter to continue the grid search; during the feedback adjustment process, the number of variables in M ​​is guaranteed to be consistent with the number of input variables and all possible selection situations are iterated; if the search cannot be completed under the current number of input variables, the number of input variables is increased by 1 until a suitable parameter is found for feedback adjustment. Steps 2-5 are as follows Figure 3 As shown;

[0079] Step 6: Send the specific control values ​​to the PLC controller through the sending module, so that the corresponding parameters can be set to automatically control the grinding overflow particle size to the specified range.

[0080] like Figure 4As shown, the model of a method for automatically controlling parameters to optimize the grinding overflow particle size range in this embodiment mainly includes two parts: a multi-particle size scanning stage and a cascade forest stage. The data is processed by the model as follows: the multi-particle size sliding window scanning process can be expressed as follows: assuming that the input data is a data set with a feature size of n, a sliding window size is set to m, and the step size is 1 by default. The data set is sampled through the sliding window to obtain d = (n-m+1) m-dimensional sub-sample sets, and each sub-sample set is trained through a random forest model and a completely random forest model respectively. After the training is completed, each forest model will generate a probability vector of length c, and then generate a characterization vector of d*c. Finally, all forest results are spliced ​​together to obtain the final output. In the cascade forest stage, each layer of the cascade forest is composed of multiple forests. The design draws on the idea of ​​deep neural networks learning original features layer by layer. Each layer of the cascade structure uses the feature information processed by the previous layer as the input of this layer, and then passes the processed results of this layer and the feature vector obtained in the previous stage to the next layer. When the performance gain of the output result of a certain layer no longer increases, the number of layers of the cascade structure stops increasing, thereby determining the structure of the cascade forest.

[0081] This embodiment also provides an overflow particle size prediction system based on decision tree regression. The system is as follows: the overflow particle size distribution of the hydrocyclone in the grinding process is regressed and predicted by the decision tree regression algorithm. First, the data set D is read and divided into the training set D according to the ratio of 8:2. 1 and the test set T 1 , using the CART regression tree algorithm, where the algorithm for generating the least squares regression tree is as follows: Assume that the training data set D 1 ={(x 1 ,y 1 ), (x 2 ,y 2 )...(x n ,y n )}, in D 1 In the input space, each region is split into two sub-regions by recursive method, and the output value of each sub-region is determined to construct a decision tree. Specifically, the following steps are included:

[0082] Step 1: Traverse the segmentation variables j and scan the segmentation points s, and select a set of optimal segmentation variables j and segmentation points s to minimize the value of formula (8);

[0083]

[0084] Step 2: Use the optimal variables (j, s) obtained by formula (9) to divide different areas, and output the division results:

[0085]

[0086] Among them, x i ∈R m , m=1,2,R 1 (j, s) = {x|x (j) ≤s}, R 2 (j, s) = {x|x (j) >s};

[0087] Step 3: Recursively call steps 1 and 2 for each sub-region until the termination condition is met;

[0088] Step 4: Divide the input space into R 1 , R 2 , …, R n Sub-region, generate the final decision tree g(x):

[0089]

[0090] Where x∈R m ,

[0091] In this example, the data of a mineral processing plant in Maanshan City is used to compare the present invention with the current mainstream decision tree and forest prediction models. The following table compares the prediction results of the overflow particle size method.

[0092]

[0093] like Figure 5 As shown in the figure, the absolute error curve of the deep forest is significantly lower than that of the decision tree model, and is lower than that of the random forest model. In addition, the fluctuation range of the absolute error curve of the deep forest is smaller. Therefore, the prediction effect of the deep forest model established by this method is better.

[0094] In summary, a method for automatically controlling parameters to optimize the overflow particle size interval of grinding of the present invention collects relevant variable data in the production process through a PLC terminal unit, and then combines data cleaning and principal component analysis to improve data quality, thereby generating a basic prediction model, and using the prediction model in real time in the subsequent production process to predict the overflow particle size in the future, and the abnormal prediction results are cumulatively grid searched in a certain control parameter order. If the overflow particle size prediction result generated by the prediction model can reach the normal range in this process, the control parameters are recorded and feedback is adjusted; if the normal range cannot be reached, the control parameter variables are replaced in turn, and the number of input variables is increased, so as to conduct a new round of search until the effective control parameter adjustment is completed. The present invention can effectively select a variety of variables in grinding production, and can accurately predict the overflow particle size change process in the future after effectively improving the data quality, so that an iterative grid search can be used to quickly find a set of control parameters before the overflow particle size is abnormal, thereby ensuring the normal operation of the operation process. Compared with the traditional direct detection using instruments, the present invention has lower deployment costs and higher stability, greatly saves human resources, and is widely applicable to various grinding scenarios.

[0095] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for automatically controlling parameters to optimize the grinding overflow particle size range, characterized in that: The following steps are involved: Step 1: Install the PLC terminal unit that needs to collect the corresponding variable data. The main collected data information includes: belt feed volume, belt feeder frequency, feeder start status, ball mill power, grinding sound, current, water supply, circular vibration screen current, pump pool water supplement detection value, pump pool liquid level, hydrocyclone feed pressure, feed flow, feed concentration, underflow pump frequency feedback value, crusher current, magnetic separator current, frequency, high-pressure roller mill current, elutriator discharge valve position value, auxiliary water valve position value, overflow detection value, underflow concentration, water inlet valve position value, and overflow particle size of hydrocyclone and other production information; Step 2: During the production process, the installed PLC terminal unit is used to collect and save data, and relevant data preprocessing operations and feature selection are performed on the data to improve data quality; The specific process of data preprocessing for the original data in step 2 is as follows: In order to solve the problem of missing data caused by PLC terminal failure or unexpected data transmission in the actual production environment, linear interpolation is used to supplement the missing values; for the missing point P generated at time t, t The missing value Y on (X,Y) can be taken as the value point P near time t. t-a (X1,Y1) and P t+b (X2, Y2), then the corresponding missing values ​​can be calculated and filled, the calculation formula is as follows: For outliers in the data set, a detection method based on the Raida criterion is used; first, the overall trend of the data is measured, and then the mean is calculated based on the data after the trend is removed. and standard deviation σ, their formulas are as follows: Then calculate the residual v of each data i , the formula is as follows: Finally, by passing the residual v i Compared with the range threshold 3σ, if |v i |≥3σ, then the corresponding data x i If it is judged as an outlier, it is corrected by replacing the sample mean; For the data after data preprocessing, principal component analysis is used to reduce the dimension so as to select appropriate features for retention. The main process of the algorithm is as follows: First, set the data matrix x of n samples and p dimensions as follows: Among them, x i =(x i1 x i2 …x ip ) T , i=1,2,...,n The integrated k variables can be expressed as: Where k = 1, 2, 3, ..., n The above formula can be expressed in matrix form as G=AX, as follows: Among them, A is the principal component coefficient matrix, and the model is generally required to meet the following three basic conditions: (1) G i and G j Uncorrelated, where i≠j, i, j∈(1, 2, ..., n) (2)Yes(G1)>Yes(G2)>…Yes(G i ),i∈(1,2,...,n) (3)a m1 2 +a m2 2 +…+a mm 2 =1,m=1,2,...,n Among them, G1 is the first principal component, G2 is the second principal component, ..., G n is the nth principal component; Generally speaking, when c (c < n) principal components are taken, the cumulative contribution rate of c is required to reach at least 85%; finally, according to the results of principal component analysis, the variable group with a cumulative contribution rate of more than 85% is taken as the final characteristic variable; Step 3: Select the processed data sets, use the deep forest algorithm to train these data sets and generate corresponding prediction models; Step 4: During the production process, the central control computer obtains the real-time monitoring data of each PLC terminal and stores the data in chronological order; at the same time, it is determined whether the overflow particle size is within the normal range. If the value is normal, wait for the next interval judgment; if not, the control parameters need to be adjusted; Step 5: When the overflow particle size value is abnormal, the generated deep forest prediction model is used to set the overflow particle size target value to the specified value, and the grid search method is used to obtain the set values ​​of all control parameters in turn; Step 6: Send the specific control values ​​to the PLC controller through the sending module, so that the corresponding parameters can be set to automatically control the grinding overflow particle size to the specified range.

2. The method for optimizing the overflow particle size range of grinding by automatic control parameters according to claim 1, characterized in that: The specific formation process of the deep forest in step 3 is as follows: first, read the data set D, and divide the data set D into a training set D1 and a test set T1 according to a certain ratio, D1∈R l×m , where l represents the number of samples in the training set D1, and m represents the number of feature variables; then initialize relevant parameters, important parameters such as the sliding window size d i The number of decision trees n in each forest will have a greater impact on model generation; then use a size of d i The sliding window is used to split the training set D1 and generate feature vectors gv of different sizes. i , Then the sliding window d i The extracted feature vector gv i They are used to train the random forest and completely random forest models respectively, and then generate two different local feature vectors, which are then concatenated to generate the transformation vector g. i , The vector g i For training the cascade forest model, g i After the nth cascade forest i The level generates an augmented vector Then the augmented vector Concatenate with the original transformation vector to form a feature vector Calculate the eigenvector The training error is , when the error gradually decreases, the current feature vector is used as the input of the next layer of joint forest; When the error does not decrease for three consecutive layers, the training process is terminated and the number of cascade forests is determined; the last layer is used as the evaluation layer, and the average of all forest predictions is calculated, which is the final deep forest regression prediction result.

3. The method for optimizing the overflow particle size range of grinding by automatic control parameters according to claim 1, characterized in that: The specific process of adjusting the control parameters using the generated prediction model in step 5 is as follows: first, the number of input variables is set to 1, the importance of all control parameters is sorted in descending order, and the control parameter variable a1 with the greatest influence is selected and added to the set M; for the variables in the set M, the grid search method is used to explore different values ​​and corresponding overflow particle size changes within the range of the parameters in the set M, and the variables outside the set M (i.e., other unselected control parameters) remain fixed; if the overflow particle size prediction result reaches the normal range during the grid search process, the control parameters are saved and the corresponding machine equipment is adjusted based on feedback; otherwise, the variables in M ​​are replaced with the next parameter to continue the grid search; during the feedback adjustment process, the number of variables in M ​​is guaranteed to be consistent with the number of input variables and all possible selection situations are iterated; if the search cannot be completed under the current number of input variables, the number of input variables is increased by 1 until a suitable parameter is found for feedback adjustment.

Images