Hot galvanizing coating thickness prediction method based on random forest algorithm

Abstract

The invention discloses a hot galvanizing coating thickness prediction method based on a random forest algorithm, and relates to the technical field of hot galvanizing steel coil production, and the prediction method comprises the steps: producing a hot galvanizing steel coil through a hot galvanizing technology, and obtaining the production historical data of the hot galvanizing technology; preprocessing the historical data and eliminating the time lag of the thickness gauge; performing feature screening based on a mechanism and a correlation coefficient; establishing a random forest network coating thickness prediction model and training; verifying and judging the accuracy and applicability of the random forest network coating thickness prediction model; if the model prediction precision does not meet the requirement, optimizing prediction model parameters; if the requirements are met, putting the prediction model into use; and outputting a zinc layer thickness prediction value. According to the method, the problem of detection time lag in galvanized sheet zinc layer thickness control is solved, and technological parameter adjustment and thickness target control are optimized.

Description

Technical Field

[0001] The invention relates to the technical field of hot-dip galvanized steel coil production, and in particular to a hot-dip galvanized coating thickness prediction method based on a random forest algorithm. Background Art

[0002] Galvanized sheet refers to a steel sheet with a layer of zinc coated on the surface. It is widely used in automobile manufacturing, refrigeration industry, construction industry, heating facilities and other fields. Important technical indicators for measuring the quality of galvanized sheet include the thickness and uniformity of the zinc layer.

[0003] This type of hot-dip galvanized steel coil mainly uses cold-rolled steel coil raw materials, which are unrolled, heated in an annealing furnace, and dipped into zinc liquid in a zinc pot. The excess zinc liquid on the surface is scraped off by an air knife to control the thickness of the zinc layer, and then cooled. In the continuous hot-dip galvanizing production process, the online measurement of the coating thickness value is completed by a thickness gauge, but due to the requirements of the galvanizing production process, the thickness gauge is installed at a distance from the air knife, resulting in a large measurement lag in the system; at the same time, in the actual galvanizing production process, the running speed of the production line changes in real time, resulting in the system measurement lag time also changing in real time. Directly using the information of the thickness gauge for the coating thickness feedback closed-loop control will produce a large overshoot and a long adjustment time, which further increases the difficulty of galvanizing production thickness control and causes large fluctuations in coating thickness. The production process of galvanized sheets is affected by multiple complex physical factors such as aerodynamics and boundary layer flow. The final zinc layer thickness is the result of the combined action of multiple variables such as air knife pressure, air knife distance, and production line running speed. The influence of these variables on the coating thickness is complex and has many disturbance factors, and has strong nonlinearity. At the same time, there are many external disturbances in the actual galvanizing industrial production process, such as: zinc liquid composition and temperature, strip steel specifications, strip steel shape and temperature, etc. These external disturbances will affect the coating thickness and cause large fluctuations in coating quality.

[0004] In summary, the production of galvanized sheet is a typical production process with large time-varying lag, nonlinearity, and strong disturbance, so it is very important to establish a prediction model for the zinc layer thickness of galvanized sheet. However, it is often difficult to establish an accurate system model using traditional mechanism modeling, and the coating thickness cannot be accurately predicted, resulting in the inability to effectively control the coating thickness of the galvanized sheet in a timely manner, causing product quality problems. Summary of the invention

[0005] In order to solve the problems of the prior art, the present invention provides a hot-dip galvanizing coating thickness prediction method based on random forest algorithm; the method can effectively improve the problem of thickness gauge detection lag after air knife process parameters are adjusted in the hot-dip galvanizing process.

[0006] In order to achieve the purpose of the present invention, the scheme adopted by the present invention is as follows:

[0007] A hot-dip galvanizing coating thickness prediction method based on random forest algorithm comprises the following steps: step 1: producing hot-dip galvanized steel coils through hot-dip galvanizing process and obtaining hot-dip galvanizing process production history data; step 2: pre-processing the history data to eliminate the time lag of the thickness gauge; step 3: for the pre-processed data, feature screening is performed in combination with mechanism analysis and correlation coefficient; step 4: establishing a random forest network coating thickness prediction model, inputting historical production data into the random forest network coating thickness prediction model, and training the random forest network coating thickness prediction model; step 5: verifying and judging the accuracy and applicability of the random forest network coating thickness prediction model. Determine; Step 6: If the prediction accuracy of the random forest network coating thickness prediction model does not meet the requirements, optimize the parameters of the random forest network coating thickness prediction model, and further train the random forest network coating thickness prediction model; If the prediction accuracy of the random forest network coating thickness prediction model meets the requirements, the random forest network coating thickness prediction model is put into use; Step 7: Input the characteristic parameters screened out in step 3 to the random forest network coating thickness prediction model, and output the predicted value of the zinc layer thickness; Among them, the characteristic parameters screened out in step 3 include air knife pressure, air knife distance, air knife angle, air knife height, strip speed, strip thickness and zinc liquid temperature.

[0008] Based on the above technical solution, further, in step 1, the historical data at least includes: product specifications and control target data: steel type, coating type, strip thickness, strip width, target coating thickness; parameter data controlled by process personnel: air knife pressure, air knife distance, air knife height; various working conditions, environmental parameters and related parameter data of the previous process flow before the steel plate reaches the air knife: zinc liquid temperature, strip speed, nitrogen pressure, nitrogen flow, actual value of the correction roller, strip tension, air knife angle.

[0009] Based on the above technical solution, further, in step 2, the preprocessing includes processing of abnormal values, and the processing process is: when the operator is calibrating, the data whose thickness gauge value is 0 is removed; the data whose coating thickness setting value does not match the actual thickness is deleted; when the coating specification is about to be changed, the operator adjusts the air knife parameters in advance and removes the data whose tail zinc layer thickness of the current steel coil is the same as that of the next steel coil.

[0010] Based on the above technical solution, further, in step 2, the preprocessing includes the response time lag processing after the air knife parameters are set. The processing process is as follows: During the process of adjusting the air knife parameters, it is difficult for the data measured by the thickness gauge to accurately match the corresponding process parameters. For each record in the dataset of the steel coil, first, it is judged whether the air knife pressure and air knife distance parameters of the steel coil are adjusted, and the deviation between the set value and the actual value is calculated. First, for the air knife pressure adjustment: For each record in the dataset, subtract the actual air knife pressure value from the set air knife pressure value to obtain ΔP. By statistically calculating the data in the stable process, the standard deviation of the actual air knife pressure is obtained as;

[0011] ΔP = P Set - P Act ; In the formula, P Set is the set air knife pressure value, and P Act is the actual air knife pressure value;

[0012] If |ΔP| < σ mbar, it is determined as a record of stable process parameters and the data is retained; if |ΔP| > σ mbar, it is determined as a record of unstable process parameters and the data is excluded;

[0013] Second, when it is detected that the air knife distance parameter in the record entry is adjusted, no corresponding time lag processing is performed.

[0014] Based on the above technical solution, further, in step 2, the preprocessing includes the detection time lag processing of the thickness gauge. The processing process is as follows:

[0015] The thickness gauge measures the thickness of the zinc coating at a distance of d meters in front of the air knife; the thickness of the zinc coating corresponding to the air knife process parameters needs to be retrieved through the corresponding space-time conversion algorithm;

[0016] First, set the speed of the i-th record as v i , define the relative displacement distance as S, with the unit of m. Then, when it comes to the (i + j)-th record, the advancing distance of the steel coil position corresponding to the i-th record is: In the formula, v k is the speed of the k-th record, where k takes values in [i, i + j]; T is the time, with the unit of s;

[0017] Second, set the actual displacement of the thickness gauge lagging behind the air knife as d meters. Then, when the (i + j)-th record satisfies d < S < d + 5, the thickness gauge value of the (i + j)-th record is used as the thickness of the zinc coating corresponding to the air knife process parameters of the i-th record. At this time, d + 5 meters is used as the maximum allowable relative displacement value; if the (i + j)-th record that satisfies the condition of d < S < d + 5 cannot be retrieved through cycling, the i-th data record is excluded.

[0018] Based on the above technical solution, further, in step 3, the mechanism analysis process is: fluid analysis is performed on the process of air knife blowing strip steel, and the formula for coating thickness can be obtained according to the Stokes equation:

[0019]

[0020] Where W is the coating thickness; ρ is the density of the zinc liquid; μ is the viscosity of the zinc liquid, which is related to the temperature of the zinc liquid; V is the strip speed, τ is the shear stress on the surface of the zinc liquid; p is the air knife pressure; and g is the acceleration of gravity.

[0021] Based on the above technical solution, further, in step 3, the process of using correlation coefficient for feature screening is: the calculation method of the Spearman correlation coefficient ρ1 between each data is:

[0022] In the formula, x i is the sample value of a variable in the preprocessed data set in step 2, is the sample mean of the variable; similarly, Y i is the sample value of another variable of the preprocessed data set in step 2, is the sample mean of another variable.

[0023] Based on the above technical solution, further, in step 4, the training process is: data preparation: divide the data into training set and test set, with a ratio of 7:3; input variables: air knife pressure, air knife distance, air knife angle, air knife height, strip speed, strip thickness, zinc liquid temperature; output variable: front thickness gauge value; parameter setting: the number of trees is 278, the maximum depth is set to 38, the minimum sample split is 2, the minimum number of sample leaf nodes is 1, and the maximum number of features is 80%; a random forest network coating thickness prediction model is established based on the network model parameters, and the parameters preprocessed and feature screened in step 3 are input to train the random forest network coating thickness prediction model.

[0024] Based on the above technical solution, further, in step 5, the verification and determination process is:

[0025]

[0026] In the formula, is the predicted value of the front thickness gauge, y i is the actual value of the front thickness gauge, is the average value of the front thickness gauge; n is the total number of data samples; MSE is the mean square error, MAE is the mean absolute error, R 2 The determination coefficient reflects the prediction ability of the random forest network coating thickness prediction model; among them, R 2The larger the MSE and MAE, the better the prediction effect.

[0027] Based on the above technical solution, further, in step 6, the optimization process is: using the grid search algorithm, an exhaustive search is performed within a predefined parameter range to find the optimal parameter combination of the random forest network coating thickness prediction model; wherein the search range is as follows: number of trees: 50-500; maximum depth: 5-50; minimum number of sample splits: 2-20; minimum number of sample nodes: 1-10; and finally the best performing parameter combination is selected.

[0028] Compared with the prior art, the beneficial effects of the present invention are specifically embodied in:

[0029] The present invention obtains historical production data; pre-processes data and eliminates thickness gauge time lag; performs feature screening based on mechanism and algorithm; establishes a random forest network coating thickness prediction model and inputs historical production data to train the network; verifies the accuracy and applicability of the random forest network coating thickness prediction model; inputs process parameters and other influencing characteristic parameters in the production process into the random forest network coating thickness prediction model to obtain a zinc layer thickness prediction value, thereby effectively improving the time lag of detection in the zinc layer thickness control of the galvanized sheet, and optimizing the process parameter adjustment and thickness target control. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A simplified flow chart of the prediction method of the present invention;

[0031] Figure 2 It is a schematic diagram of the hot-dip galvanizing process of the present invention;

[0032] Figure 3 The variation and fluctuation of the zinc layer thickness during the actual production process of the present invention;

[0033] Figure 4 This is a processing effect diagram of step 2 of the present invention;

[0034] Figure 5 This is a random forest network coating thickness prediction model diagram of the present invention;

[0035] Figure 6 The predicted value of the zinc layer thickness model and the measured thickness of the zinc layer;

[0036] Figure 7 is the Spearman correlation coefficient calculation result of the present invention;

[0037] Figure 8 is the prediction accuracy of the model of the present invention.

[0038] Reference numerals:

[0039] 1. Annealing furnace; 2. Zinc pot; 3. Air knife; 4. Thickness gauge. DETAILED DESCRIPTION

[0040] In order to make the purpose and technical solution of the present invention more clear, the technical solution of the present invention will be clearly and completely described below in conjunction with embodiments.

[0041] Example

[0042] Reference Figure 1 and Figure 2 This embodiment provides a method for predicting the thickness of a hot-dip galvanized coating based on a random forest algorithm. Preferably, the hot-dip galvanized coating is preferably the zinc layer of a hot-dip galvanized steel coil. The hot-dip galvanized steel coil is produced by a hot-dip galvanizing process, and its production process is as follows: the cold-rolled steel coil raw material is first unrolled by an unwinder, then annealed in an annealing furnace 1 and immersed in a zinc pot 2, the surface of the steel coil is dipped in zinc liquid in the zinc pot 2, and then scraped by an air knife 3 to control the thickness of the surface zinc layer, and then cooled for a distance to reach a thickness gauge 4 to detect the thickness of the hot-dip galvanized layer, and finally rolled into a hot-dip galvanized steel coil. The prediction method includes the following steps:

[0043] Step 1: Produce hot-dip galvanized steel coils through hot-dip galvanizing process, and obtain the historical production data of hot-dip galvanizing process; among which, the historical data at least includes: production specifications and control target data: steel type, coating type, strip thickness, strip width, target coating thickness, etc.; main parameter data controlled by process personnel: air knife pressure, air knife distance, air knife height, etc.; various working conditions, environmental parameters and related parameter data of the previous process flow before the steel plate reaches the air knife: zinc liquid temperature, strip speed, nitrogen pressure, nitrogen flow, actual value of straightening roller, strip tension, air knife angle, etc.

[0044] Step 2: Preprocess the historical data to eliminate the thickness gauge time lag; the data preprocessing includes abnormal value processing, response time lag processing after the air knife parameters are set, and thickness gauge detection time lag processing.

[0045] In this embodiment, the abnormal value processing process is as follows: when the operator is calibrating, the data with a thickness gauge value of 0 is removed; the data that does not match the set value of the coating thickness with the actual thickness is deleted; specifically, the following situations are included: the set value of the coating thickness of a very few steel coils does not match the actual thickness, and the measured value has abnormal data, and the data of these steel coils is deleted; at some welds, the air knife distance will be adjusted to a relatively large value, and the measured thickness will far exceed the set value at this time, and the data needs to be eliminated; for example, the thickness gauge fails, resulting in a sudden difference between the measured value and the set value, or the parameters are greatly adjusted due to manual intervention in the coil changing process, and this part of the data is not conducive to model training. When the coating specifications are about to be changed, the operator removes the data of the current steel coil with the same tail zinc layer thickness as the next steel coil when adjusting the air knife parameters in advance.

[0046] In this embodiment, the process for dealing with the response time lag after setting the air knife parameters is as follows: During the process of adjusting the air knife parameters, it is difficult for the data measured by the thickness gauge to accurately match the corresponding process parameters. For each record in the dataset of the steel coil, first, it is necessary to determine whether the air knife pressure and air knife distance parameters of the steel coil have been adjusted, and calculate the deviation between the set value and the actual value. 1. For the air knife pressure adjustment: It takes 5 - 10 s for the air knife pressure to adjust to the corresponding set value, and the process data is unstable, which is not suitable as a model training sample. Therefore, for each record in the dataset, subtract the actual air knife pressure value from the set air knife pressure value to obtain ΔP. By statistically calculating the stable process data, the standard deviation of the actual air knife pressure is obtained as: ΔP = P Set -P Act ; In the formula, P Set is the set air knife pressure value, and P Act is the actual air knife pressure value; if |ΔP| < σ mbar, it is determined as a stable process parameter record and the data is retained; if |ΔP| > σ mbar, it is determined as an unstable process parameter record and the data is excluded. 2. Since the air knife distance is adjusted by a high-speed motor for micro displacement and the response speed is extremely fast, when it is detected that the air knife distance parameter in the record entry has been adjusted, no corresponding time lag processing is performed.

[0047] In this embodiment, the process for dealing with the detection time lag of the thickness gauge is as follows: The thickness gauge measures the thickness of the zinc coating at a distance of d meters in front of the air knife. It is necessary to retrieve the thickness of the zinc coating controlled by the corresponding air knife process parameters through the corresponding space-time conversion algorithm. First, taking the i-th data record as an example, set the speed of the i-th record as v i , define the relative displacement distance as S, with the unit of m. Then, when it comes to the (i + j)-th record, the advancing distance of the steel coil position corresponding to the i-th record is: In the formula, v k is the speed of the k-th record, where k ranges from [i, i + j]; T is the time, with the unit of s. Second, set the actual displacement of the thickness gauge lagging behind the air knife as d meters. Then, when the (i + j)-th record satisfies d < S < d + 5, the thickness gauge value of the (i + j)-th record is used as the thickness of the zinc coating controlled by the air knife process parameters corresponding to the i-th record. At this time, d + 5 meters is used as the maximum allowable relative displacement value; if no (i + j)-th record satisfying the condition of d < S < d + 5 can be retrieved through cyclic search, the i-th data record is excluded. Thus, the outliers in the dataset, the response time lag after setting the air knife parameters, and the detection time lag of the thickness gauge are eliminated, and the matching between the process parameters and the thickness gauge thickness value is completed; through the combined processing of the above methods, the best dataset for model training is obtained.

[0048] Step 3: For the preprocessed data, feature screening is performed based on the mechanism and correlation coefficient; specifically, feature screening is performed from the two dimensions of algorithm and mechanism for the preprocessed data in the previous step. The mechanism analysis process is: fluid analysis is performed on the process of air knife blowing strip steel, and the formula for coating thickness is obtained according to the Stokes equation: Where W is the coating thickness; ρ is the density of the zinc liquid; μ is the viscosity of the zinc liquid; V is the strip speed, τ is the shear stress on the surface of the zinc liquid; p is the air knife pressure; and g is the acceleration of gravity.

[0049] The process of feature screening using correlation coefficient is as follows: The calculation method of the Spearman correlation coefficient ρ1 between each data is:

[0050] In the formula, x i is the sample value of a variable in the preprocessed data set in step 2, is the sample mean of the variable; similarly, Y i is the sample value of another variable of the preprocessed data set in step 2, is the sample mean of another variable.

[0051] According to the analysis of the galvanizing process mechanism, air knife pressure, air knife distance, air knife height, zinc liquid temperature, strip speed and strip thickness are all parameters that have a great influence on the coating thickness.

[0052] Step 4: Establish a random forest network coating thickness prediction model, input historical production data into the random forest network coating thickness prediction model, and train the random forest network coating thickness prediction model; in this embodiment, first determine the network model parameters: the number of trees is 278, the maximum depth is set to 38, the minimum sample split is 2, the minimum number of sample leaf nodes is 1, and the maximum number of features is 80%; establish a random forest network coating thickness prediction model based on the network model parameters, and input the parameters preprocessed and feature screened in step 3 to train the random forest network coating thickness prediction model.

[0053] Step 5: Verify and determine the accuracy and applicability of the random forest network coating thickness prediction model; in this embodiment, the verification and determination process is:

[0054]

[0055] In the formula, is the predicted value of the front thickness gauge, y i is the actual value of the front thickness gauge, is the average value of the front thickness gauge; n is the total number of data samples; MSE is the mean square error, MAE is the mean absolute error, R 2The determination coefficient reflects the prediction ability of the random forest network coating thickness prediction model; among them, R 2 The larger the number, the better the ability.

[0056] Step 6: If the prediction accuracy of the model does not meet the requirements, the parameters of the random forest network coating thickness prediction model are optimized, and the random forest network coating thickness prediction model is further trained; if the prediction accuracy of the random forest network coating thickness prediction model meets the requirements, the random forest network coating thickness prediction model is put into use;

[0057] Step 7: Input the process parameters and other influencing characteristic parameters in the production process into the random forest network coating thickness prediction model, and output the predicted value of the zinc layer thickness.

[0058] So far, the prediction of hot-dip galvanizing coating thickness based on random forest algorithm has been completed. The present invention can effectively improve the problem of lag in thickness gauge detection after the air knife process parameters are adjusted in the hot-dip galvanizing process. The deviation between the zinc layer thickness of the galvanized sheet and the actual value of the air knife pressure can be controlled within 3g / ㎡ in a short time, thereby reducing fluctuations in the thickness control process, optimizing the effectiveness of the process parameter adjustment setting, controlling the uniformity and pass rate of the coating thickness, reducing the production cost of the enterprise, and reducing the economic losses of the enterprise caused by the problem of substandard coating thickness.

[0059] The method is further described below in conjunction with specific implementation methods:

[0060] The coating thickness of a roll produced in a hot-dip galvanizing production line of a steel plant is 80g / m 2 The thickness gauge data of the steel coil is attached. Figure 3 When the head of the steel coil passes the air knife, the thickness gauge still measures the thickness data of the previous steel coil.

[0061] Step 1: Obtain the production history data of hot-dip galvanizing process;

[0062] Obtain continuous coating thickness data and air knife parameter data from the thickness gauge on the hot-dip galvanizing production line of the steel plant. Record coating thickness, production speed, temperature, air knife pressure and other related parameters. The data format is shown in Table 1:

[0063] Table 1

[0064]

[0065] Step 2: Preprocess the historical data to eliminate the thickness gauge time lag;

[0066] In this embodiment, outliers in the data are detected and removed, and missing data points are filled by interpolation to ensure data continuity and integrity.

[0067] In this embodiment, the integration method is used to eliminate the hysteresis of the thickness gauge in the data: at time, the speed of the production line is tracked and its change over time is integrated. The actual value of the integrated air knife pressure is set to 200. When the integral value reaches this target, the thickness gauge value recorded at this time will be considered as the coating thickness matching the time. This step is iterated for all thickness gauge values ​​to eliminate the hysteresis of the thickness gauge in the data. The processing effect is as follows: Figure 4 shown.

[0068] Step 3: For the preprocessed data, feature screening is performed based on mechanism and correlation coefficient.

[0069] In this example, the Spearman correlation coefficient was used for correlation analysis, which was chosen based on the low sensitivity of this method to data distribution and its ability to effectively quantify monotonic relationships in non-normally distributed data and situations containing outliers. The correlation between these rankings was calculated by converting the variables in the data set into ranks, and the Spearman coefficient varied from -1 to +1, indicating a completely negative correlation to a completely positive correlation, respectively. This analysis method revealed the variables with the strongest correlation with the zinc layer thickness, providing key information for further development and optimization of the model. Combined with the correlation between the various parameters and the coating thickness shown in Table 2, the correlation cloud diagram of each influencing factor and the zinc layer thickness is shown in the attached figure. Figure 7 shown.

[0070] Table 2

[0071]

[0072] Based on comprehensive consideration of the actual production process and the above-mentioned careful evaluation of the correlation analysis results, the input features finally determined are: air knife pressure, air knife distance, air knife height, air knife angle, strip speed, strip thickness, and zinc liquid temperature.

[0073] Step 4: Establish a random forest network coating thickness prediction model, such as Figure 5 As shown, historical production data is input into the random forest network coating thickness prediction model to train the random forest network coating thickness prediction model; wherein, the training process is: 1. Data preparation: the data is divided into a training set and a test set, with a ratio of 7:3; 2. Input variables: air knife pressure, air knife distance, air knife angle, air knife height, strip speed, strip thickness, zinc liquid temperature; 3. Output variable: front thickness gauge value; 4. Parameter setting: the number of trees is 278, the maximum depth is set to 38, the minimum sample split is 2, the minimum number of sample leaf nodes is 1, and the maximum number of features is 80%. Based on the network model parameters, a random forest network coating thickness prediction model is established, and the data set preprocessed in step 3 is input to train the model.

[0074] Step 5: Verify and determine the accuracy and applicability of the random forest network coating thickness prediction model;

[0075] (1) Calculate the predicted value of the front thickness gauge With the actual value y i Mean square error MSE:

[0076]

[0077] (2) Calculate the predicted value of the front thickness gauge With the actual value y i Mean absolute error MAE:

[0078]

[0079] (3) Calculation of the coefficient of determination R 2 :

[0080] (4) Model evaluation results:

[0081]

[0082] Attached Figure 6 The comparison between the predicted values ​​and actual values ​​of all data points in the test set of the random forest network zinc layer thickness prediction model trained in step 4 is shown. In this scatter plot, the black dots represent the predicted values ​​of the model, and the closeness of the distribution of these points relative to the central diagonal line (y=x) reflects the accuracy of the prediction. MSE<1, MAE<1, R 2 >0.95, and the error between the predicted value and the actual value of the random forest network coating thickness prediction model is within ±2g / m 2 The proportion within the range is 95.30%, and the error is within ±5g / m 2 The proportion within the range reached 99.66%, achieving the actual production control target. At this point, the establishment of the random forest network coating thickness prediction model was completed.

[0083] Step 6: If the prediction accuracy of the model does not meet the requirements, the parameters of the random forest network coating thickness prediction model are optimized, and the random forest network coating thickness prediction model is further trained; if the prediction accuracy of the random forest network coating thickness prediction model meets the requirements, the random forest network coating thickness prediction model is put into use;

[0084] In this embodiment, the optimization process is: using a grid search algorithm, an exhaustive search is performed within a predefined parameter range to find the best hyperparameter combination of the model. The search range is as follows: number of trees: 50-500; maximum depth: 5-50; minimum number of sample splits: 2-20; minimum number of sample nodes: 1-10; and finally the best performing parameter combination is selected, thereby significantly improving the prediction performance and robustness of the random forest prediction model.

[0085] Step 7: Input the feature parameters preprocessed and screened in step 3 to the random forest network coating thickness prediction model, and output the predicted value of zinc layer thickness. Figure 8 shown.

[0086] The above is only an embodiment of the present invention, and its description is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A method for predicting the thickness of hot-dip galvanized coating based on random forest algorithm, characterized in that: The following steps are involved: Step 1: Produce hot-dip galvanized steel coils through hot-dip galvanizing process and obtain hot-dip galvanizing process production history data; Step 2: Preprocess the historical data to eliminate the thickness gauge time lag; Step 3: For the preprocessed data, feature screening is performed by combining mechanism analysis and correlation coefficient; Step 4: Establish a random forest network coating thickness prediction model, input historical production data into the random forest network coating thickness prediction model, and train the random forest network coating thickness prediction model; Step 5: Verify and determine the accuracy and applicability of the random forest network coating thickness prediction model; Step 6: If the prediction accuracy of the random forest network coating thickness prediction model does not meet the requirements, the parameters of the random forest network coating thickness prediction model are optimized, and the random forest network coating thickness prediction model is further trained; if the prediction accuracy of the random forest network coating thickness prediction model meets the requirements, the random forest network coating thickness prediction model is put into use; Step 7: Input the characteristic parameters selected in step 3 into the random forest network coating thickness prediction model, and output the predicted value of the zinc layer thickness.

2. The method for predicting the thickness of hot-dip galvanized coating based on random forest algorithm according to claim 1 is characterized in that: In step 1, the historical data includes at least: Product specifications and control target data: steel type, coating type, strip thickness, strip width, target coating thickness; Parameter data controlled by process personnel: air knife pressure, air knife distance, air knife height; Various working conditions, environmental parameters and relevant parameter data of the previous process before the steel plate reaches the air knife: zinc liquid temperature, strip speed, nitrogen pressure, nitrogen flow, actual value of the correction roller, strip tension, and air knife angle.

3. The method for predicting the thickness of hot-dip galvanized coating based on random forest algorithm according to claim 1 is characterized in that: In step 2, preprocessing includes processing outliers, and the processing process is as follows: When the operator is calibrating, the data with a thickness gauge value of 0 will be removed; Delete the data that does not match the set value of coating thickness with the actual thickness; When the coating specifications are to be changed and the operator adjusts the air knife parameters in advance, the data of the zinc layer thickness at the tail of the current steel coil that is the same as that of the next steel coil will be removed.

4. The method for predicting the thickness of hot-dip galvanized coating based on random forest algorithm according to claim 1 is characterized in that: In step 2, the preprocessing includes the response time lag processing after the air knife parameters are set, and the processing process is as follows: During the process of setting and adjusting the air knife parameters, for each record of the steel coil data set, first determine whether the air knife pressure and air knife distance parameters of the steel coil are adjusted, and calculate the deviation between the set value and the actual value; 1. For air knife pressure regulation: For each record in the data set, subtract the air knife pressure setting value from the actual air knife pressure value to obtain ΔP. By performing statistical calculations on the stable process data, the standard deviation of the actual air knife pressure is obtained as follows; ΔP=P Set -P Act Where, P Set is the air knife pressure setting value, P Act is the actual value of the air knife pressure; If |ΔP|<σmbar, it is determined to be a stable process parameter record and the data is retained; If |ΔP|>σmbar, it is determined to be an unstable process parameter record and the data is discarded; Second, when it is detected that the air knife distance parameter in the record entry is adjusted, no corresponding time lag processing is performed.

5. The method for predicting the thickness of hot-dip galvanized coating based on random forest algorithm according to claim 1 is characterized in that: In step 2, the preprocessing includes the detection time lag processing of the thickness gauge, and the processing process is as follows: the thickness of the zinc layer at the set distance of d meters before the air knife is measured by the thickness gauge; the corresponding time-space conversion algorithm is required to retrieve the zinc layer thickness corresponding to the air knife process parameters; First, set the speed of the i-th record to v i , define the relative displacement distance as S, the unit is m, then when the i+jth record is recorded, the distance the steel coil position corresponding to the i-th record advances is: In the formula, v k is the speed of the kth record, where k is [i,i+j]; T is time in seconds; Second, set the displacement of the actual thickness gauge lag air knife to d meters. When the (i + j)-th record satisfies d < S < d + 5, the thickness gauge value of the (i + j)-th record is used as the zinc layer thickness corresponding to the control of the air knife process parameters of the i-th record. At this time, d + 5 meters is used as the maximum allowable relative displacement. If the (i + j)-th record that satisfies the condition d < S < d + 5 cannot be retrieved through cyclic search, the i-th data record is excluded.

6. The method for predicting the thickness of hot-dip galvanized coating based on random forest algorithm according to claim 1, characterized in that: In step 3, the mechanism analysis process is as follows: perform a fluid analysis on the process of the air knife blowing the strip, and obtain the formula for the coating thickness according to Stokes' equation: In the formula, W is the coating thickness; ρ is the density of the zinc liquid; μ is the viscosity of the zinc liquid; V is the strip speed, τ is the shear stress on the surface of the zinc liquid; p is the air knife pressure; g is the acceleration due to gravity.

7. The method for predicting the thickness of hot-dip galvanized coating based on random forest algorithm according to claim 1, characterized in that: In step 3, the process of feature selection using the correlation coefficient is as follows: The calculation method of the Spearman correlation coefficient ρ1 between each data is as follows: In the formula, x i is the sample value of a variable in the preprocessed data set in step 2, is the sample mean of the variable; Y i is the sample value of another variable of the preprocessed data set in step 2, is the sample mean of another variable.

8. The method for predicting the thickness of hot-dip galvanized coating based on random forest algorithm according to claim 1, characterized in that: In step 4, the training process is as follows: Data preparation: Divide the data into a training set and a test set, with a ratio of 7:3; Input variables: air knife pressure, air knife distance, air knife angle, air knife height, strip speed, strip thickness, zinc liquid temperature; Output variable: front thickness gauge value; Parameter settings: The number of trees is 278, the maximum depth is set to 38, the minimum sample split is 2, the minimum sample leaf node number is 1, and the maximum number of features is 80%; Based on the network model parameters, establish a random forest network coating thickness prediction model, and input the parameters preprocessed and feature-selected in step 3 to train the random forest network coating thickness prediction model.

9. The method for predicting the thickness of hot-dip galvanized coating based on random forest algorithm according to claim 1, characterized in that: In step 5, the verification and determination process is as follows: In the formula, is the predicted value of the front thickness gauge, y i is the actual value of the front thickness gauge, is the average value of the front thickness gauge; n is the total number of data samples; MSE is the mean square error, MAE is the mean absolute error, R 2 The determination coefficient reflects the prediction ability of the random forest network coating thickness prediction model; among them, R 2 The larger the MSE and MAE, the better the prediction effect.

10. The method for predicting the thickness of hot-dip galvanized coating based on random forest algorithm according to claim 1, characterized in that: In step 6, the optimization process is as follows: Use the grid search algorithm to perform an exhaustive search within the predefined parameter range to find the best parameter combination of the random forest network coating thickness prediction model; among them, the search range is as follows: Number of trees: 50 - 500; Maximum depth: 5 - 50; Minimum sample split number: 2 - 20; Minimum sample node number: 1 - 10; Finally, select the parameter combination with the best performance.

Images