A method for optimizing control parameters of a powder separator based on a multiple linear regression algorithm

Abstract

The present application relates to the technical field of operating control parameter optimization of a powder classifier, and discloses a method for optimizing operating control parameters of a powder classifier based on a multiple linear regression algorithm, which comprises the following steps: step one, setting of sample parameters; step two, establishment of a multiple linear regression equation model; step three, iterative training of the multiple linear regression equation model; step four, calculation of a bias term w0; step five, calculation of multiple linear regression equation weight values w1, w2, w3, w4, w5, w6 and w7; and step six, acquisition of the operating control parameters of the powder classifier.The present application has the beneficial effect of achieving optimized fitting of the operating control parameters of the powder classifier without causing a fault or stopping the operation of the powder classifier, and obtaining scientific and reasonable historical operating control parameters under the ideal working condition of the powder classifier efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of optimization of operating control parameters of a powder separator, and particularly relates to a method for optimizing operating control parameters of a powder separator based on a multiple linear regression algorithm. BACKGROUND

[0002] A powder separator is a key core commonly used in a cement enterprise. At present, the most widely used and most advanced powder separator is an O-Sepa powder separator, which is known as a high-efficiency powder separator. The eddy classification principle of the O-Sepa powder separator is considered to be advanced and scientific and reasonable. The specific advantages of the O-Sepa powder separator are as follows: low energy consumption, simple operation, convenient fineness adjustment, small wear, simple maintenance, large amount of powder processing, high powder separation efficiency, and the like. In a cement enterprise, the O-Sepa powder separator is usually used in cooperation with a raw mill, a coal mill or a cement mill, and is a key equipment for classifying raw materials, coal powder and cement fineness. The working principle of the O-Sepa powder separator is as follows: the material to be selected is fed into the powder separator from two feeding ports at the upper part, is fully dispersed by a scattering disc and a buffer plate, and falls into a powder selection area. Most of the powder selection air flow comes from a mill, enters the powder selection area through a tangential primary air inlet and a secondary air inlet, and enters the powder selection area horizontally through a guide vane. A cage composed of vertical vanes and horizontal vanes in the powder separator maintains a certain pressure difference between the inside and the outside when rotating, so that the air flow is stable and uniform, thereby creating good conditions for accurate powder selection. The material falls from top to bottom, and each particle is provided with multiple opportunities for repeated selection. Moreover, each selection is carried out under the balanced conditions of accurate centrifugal force and horizontal air flow. Fine powder overcomes the adverse effects of the side wall effect from the outside to the inside. Due to the extension of the powder selection distance, the particles falling into the cone part are selected again by three times of air flow, and the powder separation efficiency is greatly improved.

[0003] In actual production, the main parameters affecting the powder separation efficiency of the O-Sepa powder separator (not considering the improvement of the structure of the powder separator and the local atmospheric pressure) are the rotor speed, the primary air flow, the secondary air flow, the tertiary air flow, the primary air pressure, the secondary air pressure, and the tertiary air pressure. In the prior art, the optimization of the control parameters for improving the efficiency of the powder separator is mainly achieved through long-term production practice and through the trial and error adjustment of the control parameters of the rotor speed, the primary air flow, the secondary air flow, the tertiary air flow, the primary air pressure, the secondary air pressure, and the tertiary air pressure. The main technical problem of the above-mentioned optimization method of the control parameters of the O-Sepa powder separator in the prior art is that: at present, the powder separation efficiency of the O-Sepa powder separator can reach about 85% to 90%, since improving the efficiency of the powder separator needs to balance the energy consumption and the wear and tear of the equipment, and each order of magnitude increase in the efficiency of the powder separator requires too many control parameters to be adjusted, and the uncertainty factors are too many when the running control parameters of the powder separator are adjusted to improve the efficiency of the powder separator, which may even cause equipment failure or shutdown and affect normal production; the efficiency of the powder separator cannot be improved to more than 95%. Based on the above technical defects in the prior art, the inventors have developed a method for optimizing the running control parameters of the powder separator based on a multiple linear regression algorithm, which can well solve the above technical problems in the prior art. SUMMARY

[0004] In order to solve the above-mentioned problems in the prior art, the present application provides a method for optimizing the running control parameters of the powder separator based on a multiple linear regression algorithm, which uses the historical running control data of the rotor speed, the primary air flow, the secondary air flow, the tertiary air flow, the primary air pressure, the secondary air pressure, and the tertiary air pressure of the powder separator as the feature vector sample, simultaneously uses the efficiency of the powder separator corresponding to the historical running data of different rotor speeds, primary air flows, secondary air flows, tertiary air flows, primary air pressures, secondary air pressures, and tertiary air pressures as the target vector to form a multiple linear regression equation, uses the small batch gradient descent algorithm to repeatedly iterate and train the multiple linear regression equation, and forms a multiple linear regression equation model with a fitting goodness close to one; the weight value and the bias term affecting the feature vector and the target vector are calculated by the multiple linear regression equation model with a fitting goodness close to one, under the condition that the weight value and the bias term are known, and under the condition that the ideal efficiency of the powder separator is set, the running control parameter values of the powder separator that are well fitted and matched with the ideal efficiency value of the powder separator are calculated by the multiple linear regression equation formula with a fitting goodness close to one.

[0005] The present application provides a method for optimizing the running control parameters of the powder separator based on a multiple linear regression algorithm:

[0006] Step one, sample parameter setting: set the O-Sepa powder machine efficiency as the target vector Y, the rotor speed as X1, unit: r / min; the primary air flow as X2, unit: m 3 / min; the secondary air flow as X3, unit: m 3 / min; the tertiary air flow as X4, unit: m 3 / min; the primary air pressure as X5, unit: Mpa; the secondary air pressure as X6, unit: Mpa; the tertiary air pressure as X7, unit: Mpa; the above rotor speed, primary air flow, secondary air flow, tertiary air flow, primary air pressure, secondary air pressure, tertiary air pressure as the feature vector; at the same time, select the powder machine efficiency Y1, Y2, Y3, Y4, Y5, Y6, Y7 data corresponding to the historical operation control parameter sample data of X1, X2, X3, X4, X5, X6, X7, and the powder machine efficiency Y1, Y2, Y3, Y4, Y5, Y6, Y7 data as the target vector;

[0007] Step two, establishment of multiple linear regression equation model: the seven historical control data of rotor speed X1, primary air flow X2, secondary air flow X3, tertiary air flow X4, primary air pressure X5, secondary air pressure X6, tertiary air pressure X7 in step one above as the feature vector sample; for the above seven historical control data, the feature vector sample of each historical operation data corresponds to a powder machine historical efficiency sample; according to the linear regression equation y=wx+b, the feature vector sample of seven historical operation data is expressed by the fitting function formula as: ; the above formula is expressed in vector form as: , in which y represents the target vector of the powder machine efficiency; x represents the feature vector sample, w1, w2, w3, w4, w5, w6, w7 are the feature vector sample weight values corresponding to the rotor speed, primary air flow, secondary air flow, tertiary air flow, primary air pressure, secondary air pressure, tertiary air pressure; T represents the transpose of the vector; b represents the bias term; the above X1, X2, X3, X4, X5, X6, X7 historical operation data sample data and the corresponding powder machine efficiency Y1, Y2, Y3, Y4, Y5, Y6, Y7 sample data are fitted in the coordinates to form a fitting plane, which is expressed by the formula as ; in order to facilitate the calculation of the above formula, a supplementary item X0 is added and X0=1, X0 is substituted into the above formula to integrate into a new fitting plane formula (i.e. multiple linear regression equation) as:

[0008] ;

[0009] The integrated new fitting plane formula is converted into the calculation method of matrix ;hw(x) represents the fitting plane, i.e. the multiple linear regression equation;

[0010] Step three, iterative training of the multiple linear regression equation model: the X1, X2, X3, X4, X5, X6, X7 historical operation control data sample data and the corresponding Y1, Y2, Y3, Y4, Y5, Y6, Y7 powder concentrator efficiency sample data in step one are input into the multiple linear regression equation 1 model established in step two in the maximum batch, and the loss function is used to update the X1, X2, X3, X4, X5, X6, X7 historical operation data sample data and the Y1, Y2, Y3, Y4, Y5, Y6, Y7 efficiency sample data along the tangent direction of the iterative training curve descending with an infinitely small learning rate by using the small batch gradient descent algorithm, until the lowest point of the iterative training curve is reached by repeated iterative training, and the multiple linear regression equation model with a fitting goodness tending to one is obtained;

[0011] Step four, calculation of the bias term w0:

[0012] In the multiple linear regression equation model, for each sample, there will be an error i between the true value 3y T and the predicted value 4w (i) x 2 , which is represented by the formula , where i represents the sample matrix, and T represents the vector transpose; the error obeys the normal distribution formula ; substituting the formula into the formula yields the formula ; the formula is expressed as a loss function ; since there is a multiplication calculation in the loss function formula, the multiplication calculation in the above formula is converted to an accumulation calculation expression ; the loss function is obtained by simplifying the formula ; and the weight value solving formula is derived by taking the partial derivative of the loss function and setting the partial derivative to zero; in the above weight value solving formula, w0 represents the weight value, T represents the vector transpose, x represents the historical operation parameter sample matrix including X1, X2, X3, X4, X5, X6, X7, and y represents the powder concentrator efficiency sample matrix including Y1, Y2, Y3, Y4, Y5, Y6, Y7; the weight value solving formula is used to update the X1, X2, X3, X4, X5, X6, X7 historical operation data sample data and the Y1, Y2, Y3, Y4, Y5, Y6, Y7 efficiency sample data along the tangent direction of the iterative training curve descending with an infinitely small learning rate by using the small batch gradient descent algorithm, until the lowest point of the iterative training curve is reached by repeated iterative training, and the multiple linear regression equation model with a fitting goodness tending to one is obtained;The X1, X2, X3, X4, X5, X6, X7 historical operation parameter sample matrix and Y1, Y2, Y3, Y4, Y5, Y6, Y7 powder selecting machine efficiency sample matrix are substituted into the formula, and the w0 bias term value is calculated;

[0013] Step five, the calculation of the weight values w1, w2, w3, w4, w5, w6, w7 of the multiple linear regression equation: the w0 bias term value calculated in step four is brought into the fitting plane formula , and the weight values w1, w2, w3, w4, w5, w6, w7 are calculated respectively;

[0014] Step six, the acquisition of the powder selecting machine operation control parameters: set the ideal efficiency value of the powder selecting machine (the specific value is, for example, 96%), and substitute the bias term w0 and the weight values w1, w2, w3, w4, w5, w6, w7 calculated in step four into the fitting plane formula , and the powder selecting machine operation control parameters X1, X2, X3, X4, X5, X6, X7 that are well fitted with the ideal efficiency value of the powder selecting machine are calculated respectively, and the scientific and reasonable historical operation control parameters under the ideal powder selecting machine efficiency working condition are obtained.

[0015] In step one, the historical operation control parameter sample data of the powder selecting machine X1, X2, X3, X4, X5, X6, X7 and the corresponding powder selecting machine efficiency Y1, Y2, Y3, Y4, Y5, Y6, Y7 data are under the condition that the powder selecting machine has the same feeding amount and model and the powder selecting fineness is controlled at 300 mesh.

[0016] In step one, the historical operation control parameter sample data of the powder selecting machine X1, X2, X3, X4, X5, X6, X7 and the corresponding powder selecting machine efficiency Y1, Y2, Y3, Y4, Y5, Y6, Y7 data are stored in the control system of the powder selecting machine, and each set of historical operation control parameter data corresponds to a set of powder selecting machine efficiency; and the abnormal historical operation control parameter data is removed.

[0017] In step three, for each target amount sample and each feature vector sample, the projection in the coordinate is a black dot, which is the true value, and a multiple linear regression equation is formed on the fitting plane; for each sample in the coordinate, the change of the w0 bias term will adjust the fitting plane upward or downward, and in the process of repeated iteration and training, the fitting plane is fine-tuned by changing the w0 bias term to make it more accurate, and the core influence is the weight values w1, w2, w3, w4, w5, w6, w7, and at this time the bias term b is equal to the bias term w0;

[0018] In step three, the predicted value is the projection point of the true value sample on the fitting multiple linear regression equation; the error is the bias term (distance) between the true value and the predicted value.

[0019] wherein in step three, the error between each sample true value and predicted value is independently having the same distribution, and subject to normal distribution (Gaussian distribution) with mean 0 and variance w 2 .

[0020] The beneficial effects of the present application: by taking the historical operation control parameters of the powder classifier as the feature vector sample and the corresponding powder classifier efficiency as the target vector sample, a multiple linear regression equation model with goodness of fit close to one is formed, the bias term and weight value of the multiple linear regression equation model are calculated by using the above model, under the condition of formulating the ideal powder classifier efficiency, the calculated bias term and weight value are substituted into the multiple linear regression equation model, and the historical operation control parameters of the powder classifier are calculated respectively, thereby realizing the optimized fitting of the operation control parameters of the powder classifier without causing failure or shutdown of the powder classifier, and obtaining scientific and reasonable historical operation control parameters under the condition of formulating the ideal powder classifier efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 is the coordinate graph of the multiple linear regression equation model of the present application;

[0022] Figure 2 is the normal distribution graph of the error in the multiple linear regression equation model of the present application;

[0023] Figure 3 is the coordinate graph of the small batch gradient descent algorithm of the present application;

[0024] Figure 4 is the comparison graph of the powder classifier operation control data and efficiency in the prior art and the powder classifier operation control data and efficiency after fitting and matching of the present application;

[0025] Markings in the figure: 1, multiple linear regression equation, 2, fitting plane, 3, true value, 4, predicted value, 5, intercept. DETAILED DESCRIPTION

[0026] The present application will be further described below in combination with specific embodiments, which are only used to illustrate the present application and not to limit the scope of the present application.

[0027] As shown in Figures 1-4 , the present application provides a method for optimizing the operation control parameters of a powder classifier based on a multiple linear regression algorithm:

[0028] Step one, sample parameter setting: set the O-Sepa powder machine efficiency as the target vector Y, the rotor speed as X1, the unit as r / min; the primary air flow as X2, the unit as m 3 / min; the secondary air flow as X3, the unit as m 3 / min; the tertiary air flow as X4, the unit as m 3 / min; the primary air pressure as X5, the unit as Mpa; the secondary air pressure as X6, the unit as Mpa; the tertiary air pressure as X7, the unit as Mpa; the above rotor speed, primary air flow, secondary air flow, tertiary air flow, primary air pressure, secondary air pressure, tertiary air pressure are the feature vectors; meanwhile, select the powder machine efficiency Y1, Y2, Y3, Y4, Y5, Y6, Y7 data corresponding to the historical operation control parameter sample data of X1, X2, X3, X4, X5, X6, X7, the powder machine efficiency Y1, Y2, Y3, Y4, Y5, Y6, Y7 data are the target vector; the precondition of the historical operation control parameter sample data of the powder machine X1, X2, X3, X4, X5, X6, X7 and the corresponding powder machine efficiency Y1, Y2, Y3, Y4, Y5, Y6, Y7 data is that the powder machine is under the condition of the same type and the same feeding amount, and the powder fineness is controlled at 300 mesh; the historical operation control parameter sample data of the powder machine X1, X2, X3, X4, X5, X6, X7 and the corresponding powder machine efficiency Y1, Y2, Y3, Y4, Y5, Y6, Y7 data are stored in the control system of the powder machine, each set of historical operation control parameter data corresponds to a set of powder machine efficiency; and the abnormal data of the historical operation control parameter is removed.

[0029] Step two, establishment of multiple linear regression equation model: the seven historical control data of the rotor speed X1, the primary air flow X2, the secondary air flow X3, the tertiary air flow X4, the primary air pressure X5, the secondary air pressure X6, the tertiary air pressure X7 in step one are taken as the feature vector sample; for the above seven historical control data, each historical operation data feature vector sample corresponds to a powder machine historical efficiency sample; according to the linear regression equation y=wx+b, the feature vector sample of the seven historical operation data is expressed by the fitting function formula as: ; the above formula is expressed in vector form as: , in the above formula, y represents the target vector powder machine efficiency; x represents the feature vector sample, w1, w2, w3, w4, w5, w6, w7 are the feature vector sample weight values corresponding to the rotor speed, the primary air flow, the secondary air flow, the tertiary air flow, the primary air pressure, the secondary air pressure, the tertiary air pressure; T represents the transpose of the vector; b represents the bias term;

[0030] The historical operation data sample data of X1, X2, X3, X4, X5, X6, X7 and the corresponding powder concentrator efficiency Y1, Y2, Y3, Y4, Y5, Y6, Y7 sample data are fitted in the coordinates to form a fitting plane 2, which is expressed by the formula In order to facilitate the calculation of the above formula, a supplementary bit X0 is added and X0=1, and X0 is substituted into the above formula to integrate into a new fitting plane 2 formula (i.e. multiple linear regression equation 1):

[0031]

[0032] The integrated new fitting plane 2 formula is converted into a matrix calculation method In the above formula, hw (x) represents the fitting plane 2;

[0033] Step three, iterative training of the multiple linear regression equation model: the historical operation control data sample data of X1, X2, X3, X4, X5, X6, X7 and the corresponding Y1, Y2, Y3, Y4, Y5, Y6, Y7 powder concentrator efficiency sample data in step one are input into the multiple linear regression equation 1 model established in step two in the maximum batch, and the loss function is used to adopt the small batch gradient descent algorithm (as shown in Figure 3 ), and the learning rate tends to be infinitely small, and the tangent direction of the iterative training curve is descended, and the historical operation data sample data of X1, X2, X3, X4, X5, X6, X7 and the efficiency sample data of Y1, Y2, Y3, Y4, Y5, Y6, Y7 are iteratively updated until the lowest point of the iterative training curve is reached. The multiple linear regression equation model is obtained by repeatedly iteratively training the fitting goodness tends to one; for each target quantity sample and each feature vector sample, the projection in the coordinates is a black dot for the true value 3, and the multiple linear regression equation 1 (black straight line) is formed on the fitting plane 2; for each sample in the coordinates, the change of the w0 bias term will adjust the fitting plane upward or downward, and in the process of repeated iterative training, the fitting plane is fine-tuned by changing the w0 bias term to make it more accurate. The core influence is the weight value of w1, w2, w3, w4, w5, w6, w7, and the bias term b is equal to the bias term w0 at this time; as shown in Figure 1 For each sample, there is an error between the true value 3 y i and the predicted value 4 w T x (i) , which is expressed by the formula ; the predicted value is the projection point of the true value sample on the fitting multiple linear regression equation 1; the error is the bias term b (distance) between the true value and the predicted value, and i represents the sample in the above formula.Figure 2 The error between each sample true value and predicted value is independently distributed with the same distribution, and subject to a normal distribution (Gaussian distribution) with mean 0 and variance w 2 .

[0034] Step four, calculation of bias term w0:

[0035] In the multiple linear regression equation model, for each sample, there is an error i between the true value 3y T and the predicted value 4w (i) x , which can be expressed by the formula , where i represents the sample matrix, and T represents the vector transpose; the error is subject to a normal distribution formula . Substituting the formula into the formula yields the formula , and the formula is expressed as a loss function . Since there is a multiplication calculation in the loss function formula, the multiplication calculation in the above formula is converted to an accumulation calculation expression . By simplifying the formula , the loss function is obtained. By taking the partial derivative of the loss function and setting the partial derivative to zero, the weight value solving formula is derived, where w0 represents the weight value, T represents the vector transpose, x represents the historical running parameter sample matrix including X1, X2, X3, X4, X5, X6, X7, and y represents the selected powder machine efficiency sample matrix including Y1, Y2, Y3, Y4, Y5, Y6, Y7; using the weight value solving formula , the X1, X2, X3, X4, X5, X6, X7 historical running parameter sample matrix and Y1, Y2, Y3, Y4, Y5, Y6, Y7 selected powder machine efficiency sample matrix are substituted into the formula to calculate the w0 bias term value;

[0036] Step five, calculation of multiple linear regression equation weight values w1, w2, w3, w4, w5, w6, w7: Substitute the w0 bias term value calculated in step four into the fitting plane formula to calculate the weight values of w1, w2, w3, w4, w5, w6, w7, respectively;

[0037] Step six, the acquisition of the operating control parameters of the powder classifier: set the ideal efficiency value of the powder classifier (the specific value is for example 96%), and put the bias term w0 and the weight values w1, w2, w3, w4, w5, w6, w7 calculated in step four into the fitting plane formula to calculate the operating control parameters X1, X2, X3, X4, X5, X6, X7 of the powder classifier that are fitted to the ideal efficiency value of the powder classifier, and obtain the scientific and reasonable historical operating control parameters under the ideal efficiency working condition of the powder classifier.

[0038] In the setting of the sample parameters in the step, the powder classifier efficiency Y1, Y2, Y3, Y4, Y5, Y6, Y7 data corresponding to the historical operating control parameter sample data X1, X2, X3, X4, X5, X6, X7 of the powder classifier are selected, and the powder classifier efficiency Y1, Y2, Y3, Y4, Y5, Y6, Y7 data are the target vectors. The purpose of the above is to use the historical operating control parameter sample data of the powder classifier as the characteristic vector, and use the sample data of the powder classifier efficiency as the target vector, so as to form the multiple linear regression equation 1 on the fitting plane 2 of the coordinates. Figure 1

[0039] In the setting of the sample parameters in the step, the historical operating control parameter sample data X1, X2, X3, X4, X5, X6, X7 and the corresponding powder classifier efficiency Y1, Y2, Y3, Y4, Y5, Y6, Y7 data of the powder classifier are under the precondition that the feeding amount and the model of the powder classifier are the same, the powder fineness of the powder classifier is controlled at 300 mesh, and the abnormal data of the historical operating control parameters are removed. The main purpose of the above is that under the condition that the feeding amount of the powder classifier is the same and the model of the powder classifier is unified, on the one hand, the scientific accuracy of the historical operating control parameter sample of the powder classifier is ensured; on the other hand, the necessary condition for forming the multiple linear regression equation model with the goodness of fit tending to one is provided.

[0040] As shown in Figure 1 , in the establishment of the multiple linear regression equation model in step two, the historical running data sample data X1, X2, X3, X4, X5, X6, X7 and the corresponding sample data Y1, Y2, Y3, Y4, Y5, Y6, Y7 of the powder classifier efficiency are fitted in the coordinates to form the fitting plane 2, and the model of the multiple linear regression equation 1, and as much as possible, the true values of the historical running data sample data X1, X2, X3, X4, X5, X6, X7 and the corresponding sample data Y1, Y2, Y3, Y4, Y5, Y6, Y7 of the powder classifier efficiency are fitted near the multiple linear regression equation 1.

[0041] As shown in Figure 1 , 2 ​, 3, in step three, for each sample in the coordinates, the change of the w0 bias term will make the fitting plane adjust upward or downward, in the process of repeated iterative training, the fine-tuning of the fitting plane by changing the w0 bias term to make it more accurate, the core influence is the weight value of w1, w2, w3, w4, w5, w6, w7, at this time the bias term b is equal to the bias term w0; as shown in Figure 1 , for each sample, there will be an error between the true value 3y i and the predicted value 4w T x (i) , which can be expressed by the formula ; the predicted value is the projection point of the true value sample on the fitting multiple linear regression equation 1; the error is the bias term b (distance) between the true value and the predicted value, where i represents the sample in the above formula; as shown in Figure 2 , the error between the true value and the predicted value of each sample is independent and has the same distribution, and is subject to a normal distribution (Gaussian distribution) with a mean of 0 and a variance of w 2 . The main purpose of this is: using the loss function, using the small batch gradient descent algorithm, with a learning rate that tends to zero, along the tangent direction of the iterative training curve, the historical running data sample data of X1, X2, X3, X4, X5, X6, X7 and the efficiency sample data of Y1, Y2, Y3, Y4, Y5, Y6, Y7 are iteratively trained, so that the error between the true value and the predicted value of the sample is smaller and smaller, and finally a multiple linear regression equation model with a fitting goodness of one is formed.

[0042] As shown in Figure 2 , in step three, the error between the true value and the predicted value of each sample is independent and has the same distribution, and is subject to a normal distribution (Gaussian distribution) with a mean of 0 and a variance of w 2 . According to the normal distribution curve of the error in Figure 2 , it can be known that in the interval where the normal distribution curve approaches 0, the sample data occupies the majority, and the sample data on the left and right sides of the normal distribution curve occupies the minority.

[0043] In step four of the application, the calculation of the bias term w0 is based on the error between the true value and the predicted value of the feature vector sample data in the multiple linear regression equation 1 model, and according to the normal distribution characteristics that the error is independent and has the same distribution, and is subject to a normal distribution with a mean of 0 and a variance of w 2 , the formula is substituted into the normal distribution formula of the error The loss function formula is obtained by simplifying, then the loss function containing multiplication is converted into an accumulated formula expression, then the partial derivative of the loss function is derived to obtain the calculation formula of w0, finally, the historical running data sample data of X1, X2, X3, X4, X5, X6 and X7 and the known data of Y1, Y2, Y3, Y4, Y5, Y6 and Y7 efficiency sample are substituted into the calculation formula of w0, so that the numerical value of the w0 bias term is calculated.

[0044] In step five of the present application, the calculation of the weight values w1, w2, w3, w4, w5, w6 and w7 of the multiple linear regression equation. The main purpose of this is to substitute the numerical value of w0 calculated in step four into the multiple linear regression equation 1 trained by iteration, and calculate the numerical values of the weight values w1, w2, w3, w4, w5, w6 and w7, respectively.

[0045] In step six of the present application, the acquisition of the operating control parameters of the powder separator. The main purpose of this is to substitute the known bias term w0 value calculated in step four and the known weight values w1, w2, w3, w4, w5, w6 and w7 calculated in step five into the multiple linear regression equation 1 trained by iteration, and calculate the operating control parameters X1, X2, X3, X4, X5, X6 and X7 of the powder separator that are well fitted to the ideal efficiency of the powder separator, respectively, so that the optimization and fitting of the operating control parameters of the powder separator are realized without causing failure or downtime of the powder separator, and the scientific and reasonable historical operating control parameters under the ideal efficiency of the powder separator are obtained.

[0046] The steps one to six are mutually interactive and influenced, and the skilled person in the art should regard the steps one to six as a whole. The historical operating control parameters of the powder separator are taken as the feature vector sample and the corresponding efficiency of the powder separator is taken as the target vector sample to form a multiple linear regression equation model with the fitting degree close to one. The bias term and the weight value of the multiple linear regression equation model are calculated by using the above model. The calculated bias term and weight value are substituted into the multiple linear regression equation model under the condition of the ideal efficiency of the powder separator, and the historical operating control parameters of the powder separator are calculated, respectively, so that the optimization and fitting of the operating control parameters of the powder separator are realized without causing failure or downtime of the powder separator, and the scientific and reasonable historical operating control parameters under the ideal efficiency of the powder separator are obtained.

[0047] As Figure 4In the prior art, the adjustment of the operating control parameters of the powder concentrator is achieved by adjusting the rotor speed, the primary air flow, the secondary air flow, the tertiary air flow, the primary air pressure, the secondary air pressure and the tertiary air pressure of the powder concentrator in the production practice. The above adjustment method has a large adjustment range of the operating control parameters of the powder concentrator, which may cause abnormal operation of the powder concentrator, even cause the powder concentrator to malfunction or stop, and affect the normal production operation of the powder concentrator.

[0048] The values of the operating control parameters X1, X2, X3, X4, X5, X6 and X7 of the powder concentrator calculated in step six are input into the virtual central control system software of the powder concentrator, and the central operating parameter control program of the central control system is used for virtual operation, and the operating efficiency of the powder concentrator is simulated by the powder concentrator efficiency calculation module in the control program. Thus, the scientific verification of the values of the operating control parameters X1, X2, X3, X4, X5, X6 and X7 of the powder concentrator calculated in step six can be achieved without causing the powder concentrator to malfunction or stop. When the values of the operating control parameters X1, X2, X3, X4, X5, X6 and X7 of the powder concentrator calculated in step six are verified to be correct, the values of the operating control parameters X1, X2, X3, X4, X5, X6 and X7 of the powder concentrator calculated in step six are input into the central control program of the powder concentrator, and the production is operated normally. From the above comparison, it can be known that the values of the operating control parameters X1, X2, X3, X4, X5, X6 and X7 of the powder concentrator calculated in step six are verified by the production practice, and the production efficiency of the powder concentrator is indeed improved, and unexpected technical effects are achieved. Figure 1

[0049] Many modifications to the above-described embodiments can be apparent to those of ordinary skill in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the application. Accordingly, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.​

Claims

1. A method for optimizing the operating control parameters of an air classifier based on a multiple linear regression algorithm, characterized in that: Step 1: Setting Sample Parameters: Set the O-Sepa classifier efficiency as the target vector Y, the rotor speed as X1 (r / min), and the primary air volume as X2 (m³). 3 / min; the secondary air volume is set to X3, in m³ / min. 3 / min; the tertiary air volume is set to X4, in m³ / min. 3 / min; Primary air pressure is set to X5, in MPa; Secondary air pressure is set to X6, in MPa; Tertiary air pressure is set to X7, in MPa; The rotor speed, primary air volume, secondary air volume, tertiary air volume, primary air pressure, secondary air pressure, and tertiary air pressure mentioned above are the feature vectors; At the same time, the air classifier efficiencies Y1, Y2, Y3, Y4, Y5, Y6, and Y7 corresponding to the historical operating control parameter sample data of X1, X2, X3, X4, X5, X6, and X7 are selected, and the air classifier efficiencies Y1, Y2, Y3, Y4, Y5, Y6, and Y7 are the target vectors; Step Two: Establishing the Multiple Linear Regression Equation Model: The seven historical operating data points from Step One—rotor speed X1, primary air volume X2, secondary air volume X3, tertiary air volume X4, primary air pressure X5, secondary air pressure X6, and tertiary air pressure X7—are used as feature vector samples. For each of these seven historical operating data points, the feature vector sample corresponds to a historical efficiency sample of the air classifier. Based on the linear regression equation y=wx+b, the feature vector samples of the seven historical operating data points are expressed using the fitting function formula as follows: The above formula can be expressed in vector form as follows: In the above formula, y represents the target vector classifier efficiency; x represents the feature vector sample; w1, w2, w3, w4, w5, w6, and w7 are the feature vector sample weights corresponding to rotor speed, primary air volume, secondary air volume, tertiary air volume, primary air pressure, secondary air pressure, and tertiary air pressure; T represents the transpose of the vector; b represents the bias term; the historical operating data samples of X1, X2, X3, X4, X5, X6, and X7 and the corresponding classifier efficiency samples Y1, Y2, Y3, Y4, Y5, Y6, and Y7 are fitted onto the coordinate system to form a fitting plane, which is expressed by the formula: To facilitate the calculation of the above formula, it is proposed to add a supplementary term X0 and set X0=1. Substituting X0 into the above formula, we can integrate it into a new formula for fitting the plane: ； The integrated formula for fitting the plane is converted into a matrix calculation method. In the above formula, hw(x) represents the fitted plane, i.e., the multiple linear regression equation. Step 3: Iterative Training of the Multiple Linear Regression Equation Model: The historical operation control data samples X1, X2, X3, X4, X5, X6, and X7 from Step 1, and the corresponding air classifier efficiency sample data Y1, Y2, Y3, Y4, Y5, Y6, and Y7, are input into the multiple linear regression equation model established in Step 2 in the maximum batch size. The loss function is then used to train the model. The mini-batch gradient descent algorithm is used to iteratively update the historical running data samples of X1, X2, X3, X4, X5, X6, and X7 and the efficiency sample data of Y1, Y2, Y3, Y4, Y5, Y6, and Y7 along the tangent direction of the descent of the iterative training curve with an infinitely small learning rate. This process is repeated until the lowest point of the iterative training curve is reached, resulting in a multiple linear regression equation model with a goodness of fit approaching one. Step 4: Calculation of the bias term w0: In the multiple linear regression equation model, for each sample, the true value y i and predicted value w T x (i) There will be an error between them. Using formula In the above formula, i represents the sample matrix, and T represents the vector transpose; error Formula for following a normal distribution , the formula Substitute into the formula The formula is obtained from , the formula Expressed as a loss function Since the loss function formula involves cumulative multiplication, the cumulative multiplication in the above formula is converted into a cumulative summation expression as follows: By formula Simplify to obtain the loss function Then, by taking the partial derivative of the loss function and assuming the partial derivative is zero, the formula for calculating the weights is derived. In the formula for calculating the weight values ​​above, w0 represents the weight value, T represents the vector transpose, x represents the sample matrix of historical operating parameters including X1, X2, X3, X4, X5, X6, and X7, and y represents the sample matrix of classifier efficiency including Y1, Y2, Y3, Y4, Y5, Y6, and Y7; using the formula for calculating the weight values... Substitute the historical operating parameter sample matrices of X1, X2, X3, X4, X5, X6, and X7 and the classifier efficiency sample matrices of Y1, Y2, Y3, Y4, Y5, Y6, and Y7 into the formula to calculate the value of the w0 bias term. Step 5: Calculation of weights w1, w2, w3, w4, w5, w6, and w7 in the multiple linear regression equation: Substitute the w0 bias term value calculated in Step 4 into the fitting plane formula. In the middle, the weight values ​​of w1, w2, w3, w4, w5, w6, and w7 are calculated respectively; Step Six: Obtaining the Air Classifier's Operating Control Parameters: Set the ideal efficiency value for the air classifier, and simultaneously substitute the bias term w0 and weight values ​​w1, w2, w3, w4, w5, w6, and w7 calculated in Step Four into the fitting plane formula. The values ​​of the classifier's operating control parameters X1, X2, X3, X4, X5, X6, and X7, which are well-matched to the ideal efficiency value of the classifier, were calculated respectively, and the scientific and reasonable historical operating control parameters under the proposed ideal classifier efficiency were obtained.

2. The method for optimizing the operating control parameters of an air classifier based on a multiple linear regression algorithm according to claim 1, characterized in that: In step one, the prerequisite for using the historical operating control parameter sample data of classifiers X1, X2, X3, X4, X5, X6, and X7, and the corresponding classifier efficiency data Y1, Y2, Y3, Y4, Y5, Y6, and Y7, is that the feed amount and model of the classifier are the same, and the fineness of the classifier is controlled at 300 mesh.

3. The method for optimizing the operating control parameters of an air classifier based on a multiple linear regression algorithm according to claim 2, characterized in that: In step one, the historical operating control parameter sample data of classifiers X1, X2, X3, X4, X5, X6, and X7, and the corresponding classifier efficiencies Y1, Y2, Y3, Y4, Y5, Y6, and Y7 are stored in the classifier's control system. Each set of historical operating control parameter data corresponds to the efficiency of a classifier; and abnormal historical operating control parameter data are removed.

4. The method for optimizing the operating control parameters of an air classifier based on a multiple linear regression algorithm according to claim 1, characterized in that: In step three, the predicted value is the projection point of the true value sample onto the fitted multiple linear regression equation; error It is the bias term b between the actual value and the predicted value.

5. The method for optimizing the operating control parameters of an air classifier based on a multiple linear regression algorithm according to claim 4, characterized in that: In step three, the error between the true value and the predicted value for each sample... They are independent entities with the same distribution, and follow a mean of 0 and a variance of w. 2 It follows a normal distribution.

Images