A method and apparatus for fuzzy optimization control of a grinding classification process

Abstract

The application discloses a kind of fuzzy optimization control method and equipment of grinding classification process, method is: obtaining the historical operation data of grinding classification process, and standard data set is obtained by preprocessing;Wherein operating data includes the time series of several controlled and control variables;According to standard data set, grinding classification process is divided into working condition and determines optimal working condition;Select part controlled variable and calculate its trend characteristics, and according to trend characteristics and optimal working condition, sample is preferably selected and optimization database is generated;Fuzzy rule under different working conditions of process is extracted using weighted optimization WM algorithm, and offline fuzzy rule base is generated;Real-time operation data are obtained and working condition is identified on-line, and the working condition rule base is activated, and control output is obtained according to real-time operation data rule reasoning, simultaneously, on-line control rule is extracted to update fuzzy rule base on-line.The application can reduce the fluctuation of grinding classification process and make it stable in optimal working condition, improve the self-adaptive ability of the process optimization control system.

Description

Technical Field

[0001] This invention belongs to the field of grinding and control technology, specifically relating to a fuzzy optimization control method and equipment for grinding and classifying processes. Background Technology

[0002] Grinding and classification is an intermediate step in ore crushing and mineral flotation. This process uses the grinding action of ball mills and the classifying action of hydrocyclones to reduce ore particles from large to small, outputting a slurry with a suitable particle size. The internal mechanism of grinding and classification is complex, and its internal state parameters cannot be directly detected; this "black box" characteristic increases the difficulty of control.

[0003] Currently, most mineral processing plants in China have a low level of automation, relying heavily on real-time monitoring and adjustments by operators. However, operators lack consistent standards for classifying operating conditions, often making it difficult to maintain equipment at its optimal performance. Therefore, clearly defining operating conditions is crucial for process monitoring. Furthermore, manual operation is subjective, limited in energy, and prone to delays; the accumulation and transmission of experience and knowledge are also challenging. To address this issue, scholars both domestically and internationally have proposed several optimization control methods based on expert systems, fuzzy control, and case-based reasoning, achieving some research results. However, the control rules in these methods are derived from the experience of domain experts. While easy to learn and understand, knowledge acquisition is difficult, updates are cumbersome, and they are not suitable for grinding and classification processes with complex ore sources and variable ore properties.

[0004] With the application of big data technology, data-driven knowledge discovery methods provide an effective way to achieve process optimization control with self-learning and self-updating capabilities. The WM algorithm is an intuitive data-driven machine learning method that can directly extract fuzzy rules from sample data without any prior knowledge and is easy to understand, making it a classic method for fuzzy rule extraction. Furthermore, due to the low grade, uneven particle size distribution, and unstable composition of ore in concentrators, the grinding and classification process exhibits significant time-varying characteristics, making it difficult for offline-extracted fuzzy rules to adapt to the grinding and classification process in the long term. Therefore, researching online update strategies for fuzzy rules is of practical significance for production process control. Summary of the Invention

[0005] To address the existing technical deficiencies in the grinding and classification process, this invention provides a fuzzy optimization control method and equipment for the grinding and classification process, which reduces fluctuations in the grinding and classification process and stabilizes it under optimal operating conditions, thereby improving the adaptive capability of the grinding and classification process optimization control system.

[0006] To achieve the above technical objectives, the present invention adopts the following technical solution:

[0007] A fuzzy optimization control method for a grinding and classification process includes:

[0008] S1: Obtain historical operational data of the grinding and classification process, perform time registration and standardization on the acquired data to obtain a standard dataset; the operational data of the grinding and classification process includes the time series of several controlled variables and control variables of the grinding and classification process.

[0009] S2: Divide the working conditions of the grinding and classification process according to the standard dataset and determine the optimal working conditions;

[0010] S3: Select some controlled variables and calculate the corresponding trend feature data based on their time series data in the standard dataset; select sample data from the standard dataset based on the trend feature data and the optimal working conditions, and combine the selected sample data with the trend feature data to generate an optimized database;

[0011] S4: Using the controlled variable and the control variable as the input and output of the fuzzy rule respectively, and based on the optimized database and the weighted optimization WM algorithm, extract the fuzzy rules under different working conditions of the grinding and classification process, and generate an offline fuzzy rule library;

[0012] S5: Acquire real-time operating data of the grinding and classification process, identify its online operating conditions, activate the fuzzy rule base corresponding to the operating condition, and perform rule reasoning based on the real-time operating data to obtain the output. Apply the output to the grinding and classification process, and at the same time extract online fuzzy control rules and update the fuzzy rule base online based on a sliding window.

[0013] Furthermore, the grinding and classification process includes a grinding process and a classification process; the variables in the grinding process include feed flow rate, grinding noise, and mill power, wherein the feed flow rate is a controlled variable, and the grinding noise and mill power are both controlled variables; the variables in the classification process include makeup water flow rate, sand pump frequency, pump pool level, cyclone concentration, and cyclone pressure, wherein the makeup water flow rate and sand pump frequency are both controlled variables, and the pump pool level, cyclone concentration, and cyclone pressure are all controlled variables.

[0014] Furthermore, the method for time registration of the acquired data is as follows:

[0015] 1) Denoise the historical data to form the original dataset. ,in, This is the variable set for the grinding and classification process, corresponding to the following process variables in order: grinding noise, mill power, pump tank level, cyclone concentration, cyclone pressure, feed rate, sand pump frequency, and makeup water flow rate. The number of process variables; Process variables Time series data; This represents the total number of samples, i.e., the length of the time series data for the process variable.

[0016] 2) Using the feed flow rate as a baseline, calculate the time delay of all other variables relative to the feed flow rate; first calculate... Correlation coefficient estimation ,in for The time delay between them is calculated using the formula shown in equation (1):

[0017] (1)

[0018] in, ; and They are respectively The mean; and They are respectively The standard deviation; calculated according to formula (1). and ,from arrive latency for:

[0019] (2)

[0020] 3) Based on the time delay between variables, analyze the original dataset. Perform time registration to obtain the reconstructed dataset. ,in, Time series data of each process variable The data length is .

[0021] The Z-Score method was used to reconstruct the dataset. Standardization is performed to obtain a standard dataset. :

[0022] (3)

[0023] in, ; Let i be the data of the i-th process variable at time t. for The standardized value; The standard dataset obtained through standardization.

[0024] Furthermore, the standard dataset is divided into grinding process datasets based on the process to which the variables belong. and hierarchical process dataset The grinding and classification processes were both divided into working conditions using a kernel fuzzy C-means clustering algorithm based on state transition. The specific steps are as follows:

[0025] 1) The fuzzy C-means algorithm was used to analyze the grinding process dataset. Initial cluster analysis was performed, with the number of clusters set to c, to obtain the initial cluster center matrix. and the initial membership matrix ,in For the i-th cluster center, For data Belonging to Membership degree;

[0026] 2) Optimize the initial cluster center matrix using a state transition algorithm. The optimal solution for the cluster center matrix is ​​obtained. ;

[0027] 3) The optimal solution obtained by the state transition algorithm is used as the initial value of the cluster center matrix of the kernel fuzzy C-means clustering algorithm, and the kernel fuzzy C-means clustering algorithm is used to process the grinding process dataset. Cluster analysis is performed to obtain the optimal cluster center matrix and the corresponding membership matrix;

[0028] 4) Based on the clustering results of the dataset, the grinding process is divided into c types of working conditions. Each cluster center corresponds to one type of working condition. The optimal working condition is determined based on the production indicators, and the working conditions are divided according to the category to which the data belongs.

[0029] 5) For the hierarchical process and the hierarchical process dataset Following the same method used in steps 1) to 4) above to divide the working conditions of the grinding process, the working conditions of the classification process are divided.

[0030] Furthermore, in step S3, some controlled variables for calculating trend characteristic data include: selecting grinding noise, pump pool level, swirl concentration, and pressure; the principle for selecting sample data in step S3 is: if a sample data in the standard dataset belongs to the optimal operating condition and the corresponding variable trend characteristic is to maintain the optimal operating condition, or its corresponding trend characteristic data is close to the optimal operating condition, then the sample data is selected; the grinding noise trend is selected as a new control variable and added to the selected sample data to form optimized sample data.

[0031] Furthermore, the method for calculating the trend characteristic data of the variables is as follows:

[0032] 1) Set the initial width of the sliding window to W and move the time series of the variables into the window;

[0033] 2) Perform linear regression fitting using the least squares method on the data within the sliding window. If the fitting result satisfies the F-test, proceed to 3); if the fitting result does not satisfy the F-test, proceed to 4).

[0034] 3) Gradually increase the width of the sliding window by 1 and fit the data within the sliding window until the fitting result no longer meets the F-test requirements, then proceed to step 5).

[0035] 4) Calculate the variance and set a threshold to determine whether the data within the sliding window is in a steady state. If it is steady-state data, the sliding window width can be increased until it exceeds the threshold, at which point the trend of the data within the current sliding window can be extracted. Conversely, the sliding window width can be gradually reduced until the F-test is met or the data is below the threshold, at which point the trend of that part of the data can be extracted.

[0036] 5) After extracting a segment trend, the data in the sliding window needs to be cleared, and the sliding window needs to be moved to the next time segment W data points in the time series;

[0037] 6) Repeat steps 2) to 5) until all data in the time series of the variable are fitted. Take the first derivative of the fitted curve to obtain the quantitative characteristic value of the trend, and obtain the trend characteristic data of the variable.

[0038] Furthermore, the method for extracting fuzzy rules under different working conditions in step S4 is as follows:

[0039] 1) Treat all controlled variables of the grinding or classification process as the input variable set of the corresponding process. Each control variable is used as an output variable. ; The number of input variables;

[0040] 2) Determine each input variable based on the distribution of data in the optimized database. and output variables The ranges of the threshold values ​​are respectively and And determine the number of fuzzy subsets respectively and , ;use and Represents a fuzzy set. Input variables of A fuzzy subset For output variables of A fuzzy subset; all of them The fuzzy subsets are combined to obtain A vague area;

[0041] 3) Assign corresponding membership functions to each fuzzy subset of input and output variables;

[0042] 4) For each input / output data pair in the database that conforms to the current fuzzy rules for extraction, optimize the database. The membership degree of each input variable and output variable is calculated according to the assigned membership function under the corresponding fuzzy interval of different fuzzy subsets. The fuzzy subset to which the input data and output data belong is determined based on the principle of maximum membership degree.

[0043] 5) For those by Given a defined fuzzy region, calculate the support for the rules using all data within that fuzzy region:

[0044] (4)

[0045] in, Input variables Fuzzy subset index, ; For the fuzzy subset index of the output variable y; For indexing fuzzy regions; For index The number of input-output pairs within the fuzzy region; Indicates input In fuzzy subsets Membership degree; Indicates output In fuzzy subsets Membership degree;

[0046] 6) Obtain the support matrix :

[0047] (5)

[0048] 7) Select the rule with higher support, i.e., based on the given threshold. Select blur region The inner satisfaction of equation (6) :

[0049] (6)

[0050] in, Indicates the blurred area Maximum support within;

[0051] 8) Select the option from step 7). The corresponding support is used as the weight to calculate the center value av of the fuzzy subset:

[0052] (7)

[0053] in, Let be the number of rules that satisfy equation (6) within the fuzzy region; Output the point corresponding to the maximum value of the membership function for the i-th rule; based on the principle of maximum membership degree, we obtain... The corresponding output fuzzy subset In order to determine the fuzzy rule corresponding to the fuzzy region a;

[0054] 9) Define the confidence level (Conf) to evaluate the reliability of fuzzy rules:

[0055] (8)

[0056] (9)

[0057] in, For weighted variance;

[0058] 10) The finalized fuzzy rules are represented in the following form:

[0059] (10).

[0060] Furthermore, step S5 specifically includes:

[0061] 1) Design a dual sliding window model, wherein: the outer sliding window is of size out_W and contains multiple basic windows, and the outer sliding window slides one basic window at a time; the basic window is of size base_W and contains multiple inner sliding windows; the inner sliding window is of size in_W and slides in_W data items at a time.

[0062] 2) When the real-time running data in_t of the inner sliding window is less than the size of the inner sliding window in_W, online identification is performed based on the real-time running data, the fuzzy rule base of the corresponding working condition is activated, and the control output is obtained through rule reasoning to guide production. At the same time, the data is time-registered, and it is judged whether the data is optimized based on the principle of preferred sample data. If so, the weighted optimization WM algorithm is used to extract online fuzzy control rules.

[0063] 3) When the real-time running data in_t of the inner sliding window reaches the size in_W of the inner sliding window, the fuzzy rule base is updated online according to the fuzzy control rules extracted online in step 2), and in_t is cleared to zero and slid to the next inner sliding window.

[0064] The online update of the fuzzy rule base includes replacing old rules and adding new rules: if the antecedent of a rule is the same but the consequent parameter is different, then the rule is replaced; if there is no fuzzy rule extracted online in the fuzzy rule base, then the fuzzy rule extracted online is added to the activated working condition fuzzy rule base.

[0065] 4) When the real-time running data volume base_t of the basic window reaches the basic window size base_W, the outer sliding window is slid by one basic window size. A kernel fuzzy C-means clustering algorithm based on state transition is used to perform cluster analysis on the data within the slidable outer sliding window, and the similarity between the cluster centers and the data within the data within the outer sliding window before the slide is compared to assess the change in operating conditions. If the similarity between the cluster centers and the cluster centers of the outer sliding window before the slide is less than the set threshold, it indicates a new operating condition. During the sliding of the inner sliding window, a weighted optimization WM algorithm is applied to extract the fuzzy rules for this operating condition and add them to the control rule base. If the cluster centers of the slidable outer sliding window can be matched one-to-one with the cluster centers of the outer sliding window before the slide through similarity, it indicates that no new operating condition has occurred in the current process. The operating condition information is then updated based on the cluster centers of the slidable outer sliding window.

[0066] A fuzzy optimization control device for a grinding and classification process includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor enables the processor to implement the fuzzy optimization control method for the grinding and classification process described in any of the above technical solutions.

[0067] Beneficial effects

[0068] Compared with the prior art, the advantages of the present invention are as follows:

[0069] This invention proposes a process variable time delay analysis method combining nonlocal mean denoising and cross-correlation functions, solving the problems of denoising and time registration in process industry production data. By integrating the global search capability of the state transition algorithm and the nonlinear characteristics of the kernel function, it overcomes the shortcomings of traditional fuzzy C-means clustering algorithms, which are prone to getting trapped in local optima and sensitive to outliers. It also combines production indicators to determine the optimal operating conditions. By combining operating conditions and trend characteristics, it selects superior manual operation data to generate an optimization database, and then uses a weighted optimization WM algorithm to extract fuzzy rules under different operating conditions. Finally, a sliding window model is designed for online extraction and updating of fuzzy rules. These fuzzy rules are applied to the fuzzy optimization control of the grinding and classification process, improving the adaptability of the fuzzy rule base. Based on this invention, the online fuzzy rule optimization control can adjust the control rules in a timely manner according to process fluctuations, thereby reducing process fluctuations and stabilizing at the optimal operating conditions, improving the system's adaptability. Attached Figure Description

[0070] Figure 1 This is a general flowchart of an embodiment of this application;

[0071] Figure 2 Visualization of the clustering results of the grinding process in the embodiments of this application;

[0072] Figure 3The above are the operating characteristic curves of the ball mill in the embodiments of this application;

[0073] Figure 4 Visualize the clustering results of the hierarchical process in the embodiments of this application;

[0074] Figure 5 This application presents a dual sliding window model as an embodiment.

[0075] Figure 6 This is a flowchart illustrating the online extraction and updating of fuzzy rules based on a sliding window, as described in an embodiment of this application.

[0076] Figure 7 Comparison diagram of the grinding sound control effect of the embodiments of this application;

[0077] Figure 8 Comparison diagram of pump pool liquid level control effect in embodiments of this application;

[0078] Figure 9 Comparison diagram of the swirl concentration control effect in the embodiments of this application. Detailed Implementation

[0079] The embodiments of the present invention will be described in detail below. These embodiments are based on the technical solutions of the present invention and provide detailed implementation methods and specific operation processes to further explain the technical solutions of the present invention.

[0080] Example 1

[0081] like Figure 1 As shown, this embodiment provides a fuzzy optimization control method for a grinding and classification process. First, the raw data of the grinding and classification process is preprocessed. The time delay information between process variables is estimated using a cross-correlation function method, and dimensionless processing is performed using the Z-Score standardization method. Then, a kernel fuzzy C-means clustering algorithm based on state transition is used to divide the operating conditions, and the optimal operating condition is determined based on production indicators. Next, an optimization database is generated by combining operating conditions and trend features. Then, fuzzy rules under different operating conditions are extracted based on a weighted optimization WM algorithm, and an offline rule base is generated. Finally, samples are iterated based on a sliding window, and online clustering analysis, online operating condition identification, and online rule extraction are performed on the updated samples, thereby updating the fuzzy rules online. The fuzzy rule base is used to perform fuzzy optimization control on the grinding and classification process. Specifically, the following steps are included:

[0082] S1: Obtain historical operational data of the grinding and classification process, perform time registration and standardization on the acquired data to obtain a standard dataset; wherein, the operational data of the grinding and classification process includes the time series of several controlled variables and control variables of the grinding and classification process.

[0083] Specifically, the grinding and classification process includes a grinding process and a classification process; the variables in the grinding process include feed flow rate, grinding noise, and mill power, wherein the feed flow rate is a controlled variable, and the grinding noise and mill power are both controlled variables; the variables in the classification process include makeup water flow rate, sand pump frequency, pump pool level, cyclone concentration, and cyclone pressure, wherein the makeup water flow rate and sand pump frequency are both controlled variables, and the pump pool level, cyclone concentration, and cyclone pressure are all controlled variables.

[0084] First, the historical data is denoised to form the original dataset. ,in, This is the variable set for the grinding and classification process, corresponding to the following process variables in order: grinding noise, mill power, pump tank level, cyclone concentration, cyclone pressure, feed rate, sand pump frequency, and makeup water flow rate. The number of process variables; Process variables Time series data; This represents the total number of samples, i.e., the length of the time series data for the process variables. In this embodiment, the entire grinding and classification process includes the aforementioned eight variables, therefore... In addition, in this embodiment, the sampling period for the variables is set to 1 minute, and the time series length for each variable is [missing information]. This includes 1442 operational data samples.

[0085] Then, using the feed flow rate as a benchmark, calculate the time delay of the remaining variables relative to the feed flow rate; first calculate... Correlation coefficient estimation ,in for The time delay between them is calculated using the formula shown in equation (1):

[0086] (1)

[0087] in, ; and They are respectively The mean; and They are respectively The standard deviation; calculated according to formula (1). and ,from arrive latency for:

[0088] (2)

[0089] In this embodiment, the time delay information between variables is calculated based on the cross-correlation function method. According to the experience of an operator in a mineral processing plant, the maximum delay time is set to 50 minutes. Based on the feed rate, the time delay of each variable in the grinding and classification process relative to the feed rate is calculated as shown in Table 1.

[0090] Table 1. Time Delay of Various Variables in the Grinding and Classification Process Relative to Feed Rate

[0091]

[0092] Next, the original dataset is analyzed based on the time delay between variables. Perform time registration to obtain the reconstructed dataset. ,in, Time series data of each process variable The data length is .

[0093] Finally, the Z-Score method was used to reconstruct the dataset. Standardization is performed to obtain a standard dataset. :

[0094] (3)

[0095] in, ; Let i be the data of the i-th process variable at time t. for The standardized value; This is a standardized dataset. Based on the process to which the variables belong, the standardized dataset is divided into grinding process datasets. and hierarchical process dataset .

[0096] S2: Divide the working conditions of the grinding and classification process according to the standard dataset, and determine the optimal working conditions based on the production indicators.

[0097] In step S2, to better monitor the operating status of the grinding and classification process, the preprocessed data is divided into operating conditions using a state transition-based kernel fuzzy C-means clustering algorithm, and the optimal operating condition is determined based on production indicators. Considering that in actual production process control, given the instability of the grinding and classification process, it is impossible to simultaneously adjust control parameters while considering the operating conditions of the ball mill, pump tank, and hydrocyclone, the grinding and classification process is divided into a grinding process and a classification process to achieve refined control. The operating status of the ball mill, pump tank, and hydrocyclone is monitored separately, and the optimal operating condition is determined based on production indicators. The working conditions of the grinding process are identified and classified according to... The specific steps for identifying and classifying the working conditions of the grading process are as follows:

[0098] 1) The fuzzy C-means algorithm was used to analyze the grinding process dataset. Initial cluster analysis was performed, with the number of clusters set to c, to obtain the initial cluster center matrix. and the initial membership matrix ,in For the i-th cluster center, For data Belonging to Membership degree;

[0099] 2) Optimize the initial cluster center matrix using a state transition algorithm. The optimal solution for the cluster center matrix is ​​obtained. ;

[0100] 3) The optimal solution obtained by the state transition algorithm is used as the initial value of the cluster center matrix of the kernel fuzzy C-means clustering algorithm, and the kernel fuzzy C-means clustering algorithm is used to process the grinding process dataset. Cluster analysis is performed to obtain the optimal cluster center matrix and the corresponding membership matrix;

[0101] 4) Based on the clustering results of the dataset, the grinding process is divided into c types of working conditions. Each cluster center corresponds to one type of working condition. The optimal working condition is determined based on the production indicators, and the working conditions are divided according to the category to which the data belongs.

[0102] 5) For the hierarchical process and the hierarchical process dataset Following the same method used in steps 1) to 4) above to divide the working conditions of the grinding process, the working conditions of the classification process are divided.

[0103] In this embodiment, based on the operational characteristics of the actual grinding process in production and the experience of the operators, the number of clusters for the grinding process is set to 5. The clustering results are visualized as follows: Figure 2 As shown in Table 2, where “Δ” represents the cluster center, the actual values ​​of the cluster centers after destandardization and the results of the working condition identification are shown in Table 2.

[0104] Table 2. Grinding process condition classification results and corresponding cluster centers

[0105]

[0106] Figure 3 The ball mill's operating characteristic curve shows that when the power is near its maximum value, the ball mill is in its optimal load state, which is the optimal operating condition. Figure 2It can be seen that the ball mill has the highest power under normal operating conditions, which is the optimal operating condition for the ball mill. Based on the operational characteristics of the grading process in actual production and the experience of the operators, the number of clusters in the grading process is set to 3. To better observe the clustering results and the status of each operating condition, standardized data are used to visualize the clustering results in two dimensions: pump pool level-cyclone pressure and pump pool level-cyclone concentration. Figure 4 As shown in Table 3, the actual values ​​of the cluster centers after denormalization and the results of the operating condition identification are as follows. Maintaining stable swirling pressure and swirling concentration is beneficial to ensuring the working efficiency of the hydrocyclone; therefore, operating condition one is the optimal operating condition for the classification process.

[0107] Table 3. Results of the hierarchical process condition classification and corresponding cluster centers

[0108]

[0109] S3: Select some controlled variables, calculate their corresponding trend feature data based on their time series data in the standard dataset, and select the trend feature data of some variables together with the standard dataset to form an optimized database.

[0110] In this embodiment, the controlled variables selected for calculating trend characteristic data include: grinding noise, pump pool level, swirl concentration, and pressure. The method for calculating the trend characteristic data for each selected controlled variable is as follows:

[0111] 1) Set the initial width of the sliding window to W and move the time series of the variables into the window;

[0112] 2) Perform linear regression fitting using the least squares method on the data within the sliding window. If the fitting result satisfies the F-test, proceed to 3); if the fitting result does not satisfy the F-test, proceed to 4).

[0113] 3) Gradually increase the width of the sliding window by 1 and fit the data within the sliding window until the fitting result no longer meets the F-test requirements, then proceed to step 5).

[0114] 4) Calculate the variance and set a threshold to determine whether the data within the sliding window is in a steady state. If it is steady-state data, the sliding window width can be increased until it exceeds the threshold, at which point the trend of the data within the current sliding window can be extracted. Conversely, the sliding window width can be gradually reduced until the F-test is met or the data is below the threshold, at which point the trend of that part of the data can be extracted.

[0115] 5) After extracting a segment trend, the data in the sliding window needs to be cleared, and the sliding window needs to be moved to the next time segment W data points in the time series;

[0116] 6) Repeat steps 2) to 5) until all data in the time series of the variable are fitted. Take the first derivative of the fitted curve to obtain the quantitative characteristic value of the trend, and obtain the trend characteristic data of the variable.

[0117] Let the trend feature dataset of the four selected controlled variables be denoted as . .

[0118] Based on the trend characteristics of some variables from the standard dataset The principle for selecting the optimal sample data is as follows: if a sample data in the standard dataset belongs to the optimal operating condition and the trend characteristic of the corresponding variable is to maintain the optimal operating condition, or its corresponding trend characteristic data is to move closer to the optimal operating condition, then that sample data is selected. Then, the grinding noise trend is selected as a new control variable and added to the selected sample data to form the optimized sample data. All optimized sample data constitute the optimization database.

[0119] Therefore, an optimized database is generated by combining operating conditions, latency information, and trend characteristics. , The set of variables corresponds to the following process variables: grinding sound, grinding sound trend, mill power, pump pool level, cyclone concentration, cyclone pressure, feed rate, sand pump frequency, and makeup water flow rate. Variables Sequence data.

[0120] S4: Based on the optimized database, the weighted optimization WM algorithm is used to extract fuzzy rules under different working conditions in the grinding and classification process, and generate an offline fuzzy rule library.

[0121] For the grinding process, grinding noise, grinding noise trend value, and mill power are selected as input variables, and feed rate is selected as output variable. For the classification process, pump pool level, cyclone concentration, and cyclone pressure are selected as input variables, and makeup water flow rate and sand pump frequency are selected as output variables, respectively. Fuzzy rules are extracted for different operating conditions in the grinding and classification processes, and the specific steps are as follows:

[0122] 1) Treat all controlled variables of the grinding or classification process as the input variable set of the corresponding process. Each control variable is used as an output variable. ; The number of input variables;

[0123] 2) Determine each input variable based on the distribution of data in the optimized database. and output variables The ranges of the threshold values ​​are respectively and And determine the number of fuzzy subsets respectively and , ;use and Represents a fuzzy set. Input variables of A fuzzy subset For output variables of A fuzzy subset; all of them The fuzzy subsets are combined to obtain A vague area;

[0124] 3) Assign corresponding membership functions to each fuzzy subset of input and output variables;

[0125] 4) For each input / output data pair in the database that conforms to the current fuzzy rules for extraction, optimize the database. The membership degree of each input variable and output variable is calculated according to the assigned membership function under the corresponding fuzzy interval of different fuzzy subsets. The fuzzy subset to which the input data and output data belong is determined based on the principle of maximum membership degree.

[0126] 5) For those by Given a defined fuzzy region, calculate the support for the rules using all data within that fuzzy region:

[0127] (4)

[0128] in, Input variables Fuzzy subset index, ; For the fuzzy subset index of the output variable y; For indexing fuzzy regions; For index The number of input-output pairs within the fuzzy region; Indicates input In fuzzy subsets Membership degree; Indicates output In fuzzy subsets Membership degree;

[0129] 6) Obtain the support matrix :

[0130] (5)

[0131] 7) Select the rule with higher support, i.e., based on the given threshold. Select blur region The inner satisfaction of equation (6) :

[0132] (6)

[0133] in, Indicates the blurred area Maximum support within;

[0134] 8) Select the option from step 7). The corresponding support is used as the weight to calculate the center value av of the fuzzy subset:

[0135] (7)

[0136] in, Let be the number of rules that satisfy equation (6) within the fuzzy region; Output the point corresponding to the maximum value of the membership function for the i-th rule; based on the principle of maximum membership degree, we obtain... The corresponding output fuzzy subset In order to determine the fuzzy rule corresponding to the fuzzy region a;

[0137] 9) Define the confidence level (Conf) to evaluate the reliability of fuzzy rules:

[0138] (8)

[0139] (9)

[0140] in, For weighted variance;

[0141] 10) The finalized fuzzy rules are represented in the following form:

[0142] (10).

[0143] This embodiment takes the grinding process as an example. Table 4 shows the threshold interval settings under different working conditions, and Table 5 shows the number of fuzzy sets for the corresponding intervals under different working conditions. All variables involved adopt uniformly distributed Gaussian membership functions. The number of fuzzy rules extracted under different working conditions is shown in Table 6. Taking the partial full grinding working condition as an example, some of its fuzzy rules are shown in Table 7.

[0144] Table 4. Range of values ​​for each variable under different operating conditions during the grinding process.

[0145]

[0146] Table 5. Number of fuzzy sets defined on each variable interval under different operating conditions of the grinding process.

[0147]

[0148] Table 6 Number of fuzzy rules under different working conditions in the grinding process

[0149]

[0150] Table 7. Partial Fuzzy Rules for Partially Saturated Grinding Conditions in the Grinding Process

[0151]

[0152] S5: Acquire real-time operating data of the grinding and classification process, perform online operating condition identification, activate the fuzzy rule base for the identified operating condition, and obtain control output through rule reasoning based on the real-time operating data. Apply the output to the grinding and classification process, and simultaneously extract online fuzzy control rules and update the fuzzy rule base online based on a sliding window. Specific steps include:

[0153] 1) Design a dual sliding window model, wherein: the outer sliding window is of size out_W and contains multiple basic windows, and the outer sliding window slides one basic window at a time; the basic window is of size base_W and contains multiple inner sliding windows; the inner sliding window is of size in_W and slides in_W data items at a time.

[0154] 2) When the real-time running data in_t of the inner sliding window is less than the size of the inner sliding window in_W, online identification is performed based on the real-time running data, the fuzzy rule base of the corresponding working condition is activated, and the control output is obtained through rule reasoning to guide production. At the same time, the data is time-registered, and it is determined whether the data is optimized based on the principle of preferred sample data. If so, the weighted optimization WM algorithm is used to extract online fuzzy control rules.

[0155] 3) When the real-time running data in_t of the inner sliding window reaches the size in_W of the inner sliding window, the fuzzy rule base is updated online according to the fuzzy control rules extracted online in step 2), and in_t is cleared to zero and slid to the next inner sliding window.

[0156] The online update of the fuzzy rule base includes replacing old rules and adding new rules: if the antecedent of a rule is the same but the consequent parameter is different, then the rule is replaced; if there is no fuzzy rule extracted online in the fuzzy rule base, then the fuzzy rule extracted online is added to the activated working condition fuzzy rule base.

[0157] 4) When the real-time running data volume base_t of the basic window reaches the basic window size base_W, the outer sliding window is slid by one basic window size. A kernel fuzzy C-means clustering algorithm based on state transition is used to perform cluster analysis on the data within the slidable outer sliding window, and the similarity between the cluster centers and the data within the data within the outer sliding window before the slide is compared to assess the change in operating conditions. If the similarity between the cluster centers and the cluster centers of the outer sliding window before the slide is less than the set threshold, it indicates a new operating condition. During the sliding of the inner sliding window, a weighted optimization WM algorithm is applied to extract the fuzzy rules for this operating condition and add them to the control rule base. If the cluster centers of the slidable outer sliding window can be matched one-to-one with the cluster centers of the outer sliding window before the slide through similarity, it indicates that no new operating condition has occurred in the current process. The operating condition information is then updated based on the cluster centers of the slidable outer sliding window.

[0158] The dual sliding window designed in this embodiment is as follows: Figure 5 As shown, the online extraction and update strategy for fuzzy rules based on a sliding window is as follows: Figure 6 As shown, the outer sliding window out_W is set to 24 hours, the base window base_W is set to 6 hours, the inner sliding window in_W is set to 2 hours, in_t is the data volume counter for the inner sliding window, and base_t is the data volume counter for the base window.

[0159] The method of the present invention is verified by simulation through the establishment of a nonlinear relationship model based on LSTM of grinding noise, pump pool liquid level and swirl concentration.

[0160] Online analysis of 12 consecutive hours of grinding process production data was conducted, and the grinding sound status under manual control, online fuzzy rule control based on this invention, and offline fuzzy rule control was compared. The control effects are as follows: Figure 7 As shown in the figure, the grinding sound under manual control fluctuates significantly, indicating that the ball mill's operating state is unstable. The grinding sound under both offline and online fuzzy rule control can be basically stabilized at the optimal operating condition. Since online fuzzy rule control can adjust the fuzzy rules in a timely manner according to the fluctuations in the process, it improves the adaptive capability of the control rules. Therefore, the fluctuation of the grinding sound under online fuzzy rule control is smaller.

[0161] Online analysis of 12 consecutive hours of production data from the grading process was conducted, and the pump pool liquid level and swirl concentration states under manual control, online fuzzy rule control based on this invention, and offline fuzzy rule control were compared. The control effects are as follows: Figure 8 and Figure 9As shown in the figure, the grading process is relatively stable for the first 400 minutes. During this period, the fluctuations in the pump pool level and cyclone concentration are significant under manual control, while the online fuzzy rule control results in the most stable states for both. After 400 minutes, anomalies occur in the grading process. At this point, the online fuzzy rule control can adjust in a timely manner according to the process fluctuations, thereby mitigating the large fluctuations in the pump pool level and stabilizing it. In contrast, although the offline fuzzy rule control can ultimately stabilize the pump pool level, it exhibits significant fluctuations during the control process. At this point, the changes in cyclone concentration are not significantly different under the three control methods, but the overall fluctuation amplitude of the cyclone concentration is relatively smaller under the online fuzzy rule control. Therefore, the online updating of the fuzzy control rules in this invention can reduce process fluctuations and improve the system's adaptive capability to a certain extent.

[0162] The above embodiments are preferred embodiments of this application. Those skilled in the art can make various changes or improvements based on them. Without departing from the overall concept of this application, such changes or improvements should fall within the scope of protection claimed in this application.

Claims

1. A fuzzy optimization control method for a grinding and classification process, characterized in that, include: S1: Obtain historical operational data of the grinding and classification process, perform time registration and standardization on the acquired data to obtain a standard dataset; the operational data of the grinding and classification process includes the time series of several controlled variables and control variables of the grinding and classification process. S2: Divide the working conditions of the grinding and classification process according to the standard dataset and determine the optimal working conditions; S3: Select some controlled variables and calculate the corresponding trend feature data based on their time series data in the standard dataset; select sample data from the standard dataset based on the trend feature data and the optimal working conditions, and combine the selected sample data with the trend feature data to generate an optimized database; S4: Using the controlled variable and the control variable as the input and output of the fuzzy rule respectively, and based on the optimized database and the weighted optimization WM algorithm, extract the fuzzy rules under different working conditions of the grinding and classification process, and generate an offline fuzzy rule library. The method for extracting fuzzy rules under different working conditions in step S4 is as follows: 1) Treat all controlled variables of the grinding or classification process as the input variable set of the corresponding process. Each control variable is used as an output variable. ; The number of input variables; 2) Determine each input variable based on the distribution of data in the optimized database. and output variables The ranges of the threshold values ​​are respectively and And determine the number of fuzzy subsets respectively and , ;use and Represents a fuzzy set. Input variables of A fuzzy subset For output variables of A fuzzy subset; all of them The fuzzy subsets are combined to obtain A vague area; 3) Assign corresponding membership functions to each fuzzy subset of input and output variables; 4) For each input / output data pair in the database that conforms to the current fuzzy rules for extraction, optimize the database. The membership degree of each input variable and output variable is calculated according to the assigned membership function under the corresponding fuzzy interval of different fuzzy subsets. The fuzzy subset to which the input data and output data belong is determined based on the principle of maximum membership degree. 5) For those by Given a defined fuzzy region, calculate the support for the rules using all data within that fuzzy region: (4) in, Input variables Fuzzy subset index, ; For the fuzzy subset index of the output variable y; For indexing fuzzy regions; For index The number of input-output pairs within the fuzzy region; Indicates input In fuzzy subsets Membership degree; Indicates output In fuzzy subsets Membership degree; 6) Obtain the support matrix : (5) 7) Select the rule with higher support, i.e., based on the given threshold. Select blur region The inner satisfaction of equation (6) : (6) in, Indicates the blurred area Maximum support within; 8) Select the option from step 7). The corresponding support is used as the weight to calculate the center value av of the fuzzy subset: (7) in, Let be the number of rules that satisfy equation (6) within the fuzzy region; Output the point corresponding to the maximum value of the membership function for the i-th rule; based on the principle of maximum membership degree, we obtain... The corresponding output fuzzy subset In order to determine the fuzzy rule corresponding to the fuzzy region a; 9) Define the confidence level (Conf) to evaluate the reliability of fuzzy rules: (8) (9) in, For weighted variance; 10) The finalized fuzzy rules are represented in the following form: (10)； S5: Acquire real-time operating data of the grinding and classification process, identify its online operating conditions, activate the fuzzy rule base corresponding to the operating condition, and perform rule reasoning based on the real-time operating data to obtain the output. Apply the output to the grinding and classification process, and at the same time extract online fuzzy control rules and update the fuzzy rule base online based on a sliding window.

2. The fuzzy optimization control method for a grinding and classification process according to claim 1, characterized in that, The grinding and classification process includes a grinding process and a classification process. The variables in the grinding process include feed flow rate, grinding noise, and mill power, wherein the feed flow rate is a controlled variable, and the grinding noise and mill power are both controlled variables. The variables in the classification process include makeup water flow rate, sand pump frequency, pump pool level, cyclone concentration, and cyclone pressure, wherein the makeup water flow rate and sand pump frequency are both controlled variables, and the pump pool level, cyclone concentration, and cyclone pressure are all controlled variables.

3. The fuzzy optimization control method for a grinding and classification process according to claim 2, characterized in that, The method for time registration of the acquired data is as follows: 1) Denoise the historical data to form the original dataset. ,in, This is the variable set for the grinding and classification process, corresponding to the following process variables in order: grinding noise, mill power, pump tank level, cyclone concentration, cyclone pressure, feed rate, sand pump frequency, and makeup water flow rate. The number of process variables; Process variables Time series data; This represents the total number of samples, i.e., the length of the time series data for the process variable. 2) Using the feed flow rate as a baseline, calculate the time delay of all other variables relative to the feed flow rate; first calculate... Correlation coefficient estimation ,in for Time delay, The calculation formula is shown in equation (1): (1) in, ; and They are respectively The mean; and They are respectively The standard deviation; calculated according to formula (1). and ,from arrive latency for: (2) 3) Based on the time delay between variables, analyze the original dataset. Perform time registration to obtain the reconstructed dataset. ,in, Time series data of each process variable The data length is .

4. The fuzzy optimization control method for a grinding and classification process according to claim 3, characterized in that, The Z-Score method was used to reconstruct the dataset. Standardization is performed to obtain a standard dataset. : (3) in, ; Let i be the data of the i-th process variable at time t. for The standardized value; The standard dataset obtained through standardization.

5. The fuzzy optimization control method for a grinding and classification process according to claim 2, characterized in that, The standard dataset is divided into grinding process datasets based on the process to which the variables belong. and hierarchical process dataset The grinding and classification processes were both divided into working conditions using a kernel fuzzy C-means clustering algorithm based on state transition. The specific steps are as follows: 1) The fuzzy C-means algorithm was used to analyze the grinding process dataset. Initial cluster analysis was performed, with the number of clusters set to c, to obtain the initial cluster center matrix. and the initial membership matrix ,in For the i-th cluster center, For data Belonging to membership degree ; 2) Optimize the initial cluster center matrix using a state transition algorithm. The optimal solution for the cluster center matrix is ​​obtained. ; 3) The optimal solution obtained by the state transition algorithm is used as the initial value of the cluster center matrix of the kernel fuzzy C-means clustering algorithm, and the kernel fuzzy C-means clustering algorithm is used to process the grinding process dataset. Cluster analysis is performed to obtain the optimal cluster center matrix and the corresponding membership matrix; 4) Based on the clustering results of the dataset, the grinding process is divided into c types of working conditions. Each cluster center corresponds to one type of working condition. The optimal working condition is determined based on the production indicators, and the working conditions are divided according to the category to which the data belongs. 5) For the hierarchical process and the hierarchical process dataset Following the same method used in steps 1) to 4) above to divide the working conditions of the grinding process, the working conditions of the classification process are divided.

6. The fuzzy optimization control method for a grinding and classification process according to claim 2, characterized in that, In step S3, the controlled variables for calculating the trend characteristic data include: selecting grinding noise, pump pool level, swirl concentration, and pressure; the principle for selecting the optimal sample data in step S3 is: if a sample data in the standard dataset belongs to the optimal operating condition and the corresponding variable trend characteristic is to maintain the optimal operating condition, or its corresponding trend characteristic data is close to the optimal operating condition, then the sample data is selected; the grinding noise trend is selected as a new control variable and added to the selected sample data to form the optimized sample data.

7. The fuzzy optimization control method for a grinding and classification process according to claim 2, characterized in that, The method for calculating the trend characteristic data of variables is as follows: 1) Set the initial width of the sliding window to W and move the time series of the variables into the window; 2) Perform linear regression fitting using the least squares method on the data within the sliding window. If the fitting result satisfies the F-test, proceed to 3); if the fitting result does not satisfy the F-test, proceed to 4). 3) Gradually increase the width of the sliding window by 1 and fit the data within the sliding window until the fitting result no longer meets the F-test requirements, then proceed to step 5). 4) Calculate the variance and set a threshold to determine whether the data within the sliding window is in a steady state. If it is steady-state data, the sliding window width can be increased until it exceeds the threshold, at which point the trend of the data within the current sliding window can be extracted. Conversely, the sliding window width can be gradually reduced until the F-test is met or the data is below the threshold, at which point the trend of that part of the data can be extracted. 5) After extracting a segment trend, the data in the sliding window needs to be cleared, and the sliding window needs to be moved to the next time segment W data points in the time series; 6) Repeat steps 2) to 5) until all data in the time series of the variable are fitted. Take the first derivative of the fitted curve to obtain the quantitative characteristic value of the trend, and obtain the trend characteristic data of the variable.

8. The fuzzy optimization control method for a grinding and classification process according to claim 2, characterized in that, Step S5 specifically includes: 1) Design a dual sliding window model, wherein: the outer sliding window is of size out_W and contains multiple basic windows, and the outer sliding window slides one basic window at a time; the basic window is of size base_W and contains multiple inner sliding windows; the inner sliding window is of size in_W and slides in_W data items at a time. 2) When the real-time running data in_t of the inner sliding window is less than the size of the inner sliding window in_W, online identification is performed based on the real-time running data, the fuzzy rule base of the corresponding working condition is activated, and the control output is obtained through rule reasoning to guide production. At the same time, the data is time-registered, and it is judged whether the data is optimized based on the principle of preferred sample data. If so, the weighted optimization WM algorithm is used to extract online fuzzy control rules. 3) When the real-time running data in_t of the inner sliding window reaches the size in_W of the inner sliding window, the fuzzy rule base is updated online according to the fuzzy control rules extracted online in step 2), and in_t is cleared to zero and slid to the next inner sliding window. The online update of the fuzzy rule base includes replacing old rules and adding new rules: if the antecedent of a rule is the same but the consequent parameter is different, then the rule is replaced; if there is no fuzzy rule extracted online in the fuzzy rule base, then the fuzzy rule extracted online is added to the activated working condition fuzzy rule base. 4) When the real-time running data volume base_t of the basic window reaches the basic window size base_W, the outer sliding window is slid by one basic window size. A kernel fuzzy C-means clustering algorithm based on state transition is used to perform cluster analysis on the data within the slidable outer sliding window, and the similarity between the cluster centers and the data within the data within the outer sliding window before the slide is compared to assess the change in operating conditions. If the similarity between the cluster centers and the cluster centers of the outer sliding window before the slide is less than the set threshold, it indicates a new operating condition. During the sliding of the inner sliding window, a weighted optimization WM algorithm is applied to extract the fuzzy rules for this operating condition and add them to the control rule base. If the cluster centers of the slidable outer sliding window can be matched one-to-one with the cluster centers of the outer sliding window before the slide through similarity, it indicates that no new operating condition has occurred in the current process. The operating condition information is then updated based on the cluster centers of the slidable outer sliding window.

9. A fuzzy optimization control device for a grinding and classification process, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the computer program is executed by the processor, the processor causes the processor to implement the method as described in any one of claims 1 to 8.

Images