WO2015122010A1 - 圧延プロセスの学習制御装置 - Google Patents
圧延プロセスの学習制御装置 Download PDFInfo
- Publication number
- WO2015122010A1 WO2015122010A1 PCT/JP2014/053622 JP2014053622W WO2015122010A1 WO 2015122010 A1 WO2015122010 A1 WO 2015122010A1 JP 2014053622 W JP2014053622 W JP 2014053622W WO 2015122010 A1 WO2015122010 A1 WO 2015122010A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- value
- learning
- learning coefficient
- cell
- rolling
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Ceased
Links
Images
Classifications
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B21—MECHANICAL METAL-WORKING WITHOUT ESSENTIALLY REMOVING MATERIAL; PUNCHING METAL
- B21B—ROLLING OF METAL
- B21B37/00—Control devices or methods specially adapted for metal-rolling mills or the work produced thereby
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
- G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
- G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
- G05B13/0265—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric the criterion being a learning criterion
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Definitions
- the control of the rolling process is to perform rolling so that the rolled material after the manufacture is completed has a desired size and temperature.
- the rolling process control generally includes setting control and dynamic control.
- the rolling phenomenon is predicted by a model formula, and the setting values of the equipment of the rolling equipment, such as the rolling speed, the amount of cooling water, and the roll gap of the rolling mill, so that the rolled material has a desired size and temperature.
- the model formula cannot completely represent the physical phenomenon that occurs in the rolling process.
- the calculation formula representing the model has been simplified for reasons such as reduction of calculation load and convenience of adjustment. For this reason, a deviation occurs between the actual value measured by the sensor and the predicted value calculated by the model formula.
- learning control is performed in which a learning coefficient is provided in the model formula and the learning coefficient is automatically adjusted based on the actual data in order to improve and stabilize the prediction accuracy of the rolling phenomenon.
- Actual data includes finishing temperature and rolling load, and these are collected for the purpose of setting calculation.
- Learning control consists of a plurality of processes, one of which is actual recalculation.
- a model predicted value based on the actual data is calculated using the model formula. This is generally called the actual recalculation value.
- the actual recalculated value is compared with the actual value included in the actual data, and an error of the actual recalculated value with respect to the actual value, that is, a model error is calculated.
- the actual value of the rolling load measured by the load cell is compared with the actual recalculated value of the rolling load calculated from the actual data using the model formula, and rolling A load model error is calculated.
- a learning coefficient is calculated based on the model error.
- the learning coefficient calculated at this time is called an instantaneous value of the learning coefficient.
- the instantaneous value of the learning coefficient calculated from the model error cannot be applied to the next rolled material as it is. Therefore, the instantaneous value of the learning coefficient is passed through a smoothing filter. A value obtained by smoothing the instantaneous value of the learning coefficient is used as an updated value of the learning coefficient.
- the following equation is a specific example of a smoothing filter equation that calculates an update value from an instantaneous value of a learning coefficient.
- the update value of the learning coefficient is calculated by multiplying the deviation between the instantaneous value of the learning coefficient and the previous value of the learning coefficient (previous value of the updated value) by the update gain and adding the previous value of the learning coefficient thereto.
- Znew Zuse * (1- ⁇ ) + Zcur * ⁇
- Znew learning coefficient update value
- Zcur learning coefficient (instantaneous value)
- Zuse Learning coefficient (previous value)
- ⁇ Learning coefficient update gain (filter time constant)
- the update value of the calculated learning coefficient is generally recorded in a stratified table.
- the stratification is a concept for classifying rolling conditions such as thickness, width, strain, strain rate, and temperature of the rolled material. For example, when the thickness is divided into m and the width is divided into n, the stratification table is composed of m ⁇ n cells. Each time the rolling of the material is finished, an updated value of the learning coefficient is calculated and recorded in a cell that matches the rolling conditions of the material.
- a stratified table for recording learning coefficients different learning coefficients for each rolling condition can be appropriately managed, and the rolling phenomenon prediction system is improved. That is, the learning control using the stratified table is an important function for ensuring the prediction accuracy of the model-type rolling phenomenon and ensuring the product quality accuracy and the rolling stability.
- a method of simultaneously updating the learning coefficient of an adjacent cell when updating the learning coefficient of one cell can be considered.
- a learning coefficient is calculated for the cell (i, j) corresponding to the rolling condition by the same method as described above. Note that (i, j) indicates the coordinates of the cell in the stratification table.
- Znew (i, j) Zuse (i, j) * (1- ⁇ ) + Zcur (i, j) * ⁇
- ⁇ Learning coefficient update gain (filter time constant) for the cell
- the learning coefficient is calculated by the following formula. Note that (p, q) indicates the coordinates of adjacent cells in the stratification table, which includes (i-1, j), (i, j-1), (i + 1, j), and (i + 1,1). j) is included.
- This learning method can saturate the learning coefficients of adjacent cells with as few rolling opportunities as possible.
- the learning coefficient of a cell slightly away from the cell cannot be updated. That is, with this learning method, only a limited effect can be obtained with respect to reducing rolling opportunities.
- the learning coefficient of the cell is unstable and greatly fluctuates every time it is updated, the learning coefficient of the adjacent cell is also affected.
- learning control using a stratified table has another problem.
- the problem is that it is difficult to follow changes with time in the rolling process because the cells are subdivided. If there is no rolling for a while for the rolling conditions applicable to a certain cell, the rolling process may change during that time.
- the change in the rolling process mentioned here includes both an active change such as a change in temperature level in hot rolling and a passive change such as deterioration of equipment. If the rolling process changes, the true learning factor will also change. For this reason, when the update value of the learning coefficient recorded in the stratification table remains old, the value may not be appropriate.
- an error included in the model prediction value increases, and the accuracy of the set value of the device decreases.
- the learning term of the unskilled lot is optimized using the learning term of the adjacent mastered lot. can do.
- the learning item of the adjacent lot is not necessarily more familiar than the learning item of the next lot. If the learning item of the next lot is unfamiliar and the learning item of the adjacent lot is also unfamiliar, setting calculation for the next lot must be performed based on the unfamiliar learning item.
- the learning term of an adjacent lot is unstable, that is, if the value fluctuates greatly with each update, the learning term of the next lot that is corrected using it will also become unstable. End up.
- the accuracy of the learning term of the adjacent lot has deteriorated over time due to a change in the rolling process without rolling for a while, the accuracy of the learning term of the next lot to be corrected by using it is also lowered. End up.
- the present invention has been made in view of the above problems, and in the learning control device that learns and manages the learning coefficient of the model formula used for the setting calculation for each rolling condition, the saturation, stability, and the latest degree in each rolling condition. It is an object of the present invention to obtain a learning coefficient satisfying any of the above.
- a learning control apparatus is a learning control apparatus that has a database composed of a plurality of cells that classify rolling conditions, and that manages learning coefficients of model equations used for setting calculation of a rolling process in the database.
- the present invention is configured as follows.
- the learning control apparatus includes an instantaneous value calculation recording means and an update value calculation recording means as means for calculating a learning coefficient and recording it in a database.
- the instantaneous value calculation recording means calculates the instantaneous value of the learning coefficient based on the difference between the actual value measured in the rolling process and the actual recalculated value calculated using the model formula, and the calculated instantaneous value is It is comprised so that it may record with learning date with respect to the cell applicable to the said rolling conditions. With such a configuration of the instantaneous value calculation recording means, instantaneous value history information is recorded for each cell in the database.
- the update value calculation recording means calculates the update value of the learning coefficient based on the instantaneous value of the learning coefficient calculated by the instantaneous value calculation means and the previous value of the learning coefficient under the rolling condition, and the calculated update value is It is comprised so that it may record with learning date with respect to the cell applicable to the said rolling conditions. With this configuration of the update value calculation recording means, update value history information is recorded in the database for each cell.
- each history information of the instantaneous value and the updated value may be managed by a stratified table. That is, the database is configured to include a first stratification table and a second stratification table that share cells, and the instantaneous value recording by the instantaneous value calculation recording means is performed on the first stratification table. The update value recording by the update value calculation recording means may be performed on the second stratified table.
- a learning control apparatus includes: a latest degree evaluation unit that evaluates the latest degree of a learning coefficient; a saturation degree evaluation unit that evaluates a saturation degree of a learning coefficient; and a stability evaluation means that evaluates the stability of a learning coefficient.
- the latest degree evaluation means is configured to evaluate the latest degree of the learning coefficient for each of the plurality of cells based on the update value history information recorded in the database.
- the saturation evaluation means is configured to evaluate the saturation of the learning coefficient for each of the plurality of cells based on the history information of the update values recorded in the database.
- the stability evaluation means is configured to evaluate the stability of the learning coefficient for each of the plurality of cells based on the instantaneous value history information recorded in the database.
- each evaluation result of the latest degree, the saturation degree, and the stability can be managed by the stratified table.
- the database may be configured to further include third to fifth stratification tables that share cells with the first and second stratification tables.
- the latest evaluation result by the latest evaluation means is recorded on the third stratification table, and the saturation evaluation result by the saturation evaluation means is recorded on the fourth stratification table.
- the recording of the stability evaluation result by the stability evaluation means may be performed on the fifth stratified table.
- the learning control apparatus includes a neighboring cell selection unit, an estimated value calculation unit, and a use value determination unit.
- the neighboring cell selection means is a space with the target cell corresponding to the next rolling condition in a space having the rolling condition as a coordinate axis from among the cells whose evaluation results of the latest degree, the saturation degree, and the stability satisfy the criteria. A predetermined number of neighboring cells having a short distance are selected.
- the estimated value calculation means respectively determines the representative value of the learning coefficient in the selected predetermined number of neighboring cells, and based on the coordinates of the target cell and the selected coordinates of the predetermined number of neighboring cells and the representative value, by polynomial interpolation, It is configured to calculate an estimated value of the learning coefficient in the target cell.
- the use value determining means is configured to correct the latest update value of the learning coefficient in the target cell with the estimated value and to determine the use value of the learning coefficient under the next rolling condition.
- the usage value determining means is configured to calculate the usage value by a weighted average of the latest updated value and the estimated value.
- the weighted average according to the evaluation result is such that the weight of the latest update value increases as the evaluation result of the latest degree, saturation degree, and stability for the target cell increases, and the weight of the estimated value increases as the evaluation result decreases. It is more preferable to change the weight coefficient.
- the learning control apparatus configured as described above, it is possible to apply the latest learning coefficient with a high degree of saturation and a high degree of stability with respect to a cell corresponding to the rolling condition of the next material. Thereby, the accuracy of the model predicted value in the setting control can be improved, and consequently the accuracy of product quality and the stability of operation can be improved.
- FIG. 1 shows the configuration and processing flow of a learning control device for a rolling process according to the present embodiment.
- the learning control device 1 includes a learning device 10, a storage device 20, and a setting calculation device 30.
- the learning device 10 is a device that learns the learning coefficient of the model formula used for the setting calculation based on the performance data of the rolling process.
- the storage device 20 is a device that stores the learning coefficient of the model formula and information related thereto in a database.
- This database is composed of a plurality of cells dividing the rolling conditions. More specifically, it is composed of five stratified tables 21, 22, 23, 24, and 25 that share cells.
- the setting calculation device 30 is a device that calculates a model predicted value using a learning coefficient stored in a database, and determines a setting value of a rolling mill device based on the model predicted value.
- the processing unit 11 recalculates the model predicted value based on the actual data of the rolling process.
- the processing unit 12 calculates an instantaneous value of the learning coefficient based on a model error between the actual recalculated value calculated by the processing unit 11 and the actual value included in the actual data.
- the processing unit 13 calculates an updated value of the learning coefficient by passing the instantaneous value of the learning coefficient calculated by the processing unit 12 through a smoothing filter.
- the calculation method of the instantaneous value and the update value of the learning coefficient is the same as that in the conventional learning control described in “Background Art”.
- the learning control device records the instantaneous value and the updated value of the learning coefficient in the cell corresponding to the rolling condition of the rolled material after the rolling of the rolled material is completed. Specifically, the instantaneous value of the learning coefficient is recorded in the first stratification table 21.
- the processing unit 12 selects a cell corresponding to the rolling condition from a plurality of cells constituting the first stratification table 21. Then, the instantaneous value history information is stored in the first stratified table 21 by recording the calculated instantaneous value in time series for the selected cell. More specifically, in the first stratification table 21, the instantaneous value of the learning coefficient is stored as history information together with the learning date and time for the past N rolling materials for each cell. The stored information is numbered from the newest.
- the newly stored information is the first, and every time information is added, a number i + 1 is newly added to the i-th information. Further, if N pieces of information are already stored when information is added, the oldest Nth information is discarded, and an Nth number is newly added to the (N-1) th information.
- the updated value of the learning coefficient is recorded in the second stratification table 22.
- the processing unit 13 selects a cell corresponding to the rolling condition from a plurality of cells constituting the second stratification table 22. Then, the update value history information is stored in the second stratification table 22 by recording the calculated update value in time series for the selected cell. More specifically, in the second stratification table 22, the updated value of the learning coefficient is stored as history information together with the learning date and time for the past M rolled materials for each cell. The stored information is numbered from the newest, as is the history information regarding the instantaneous value of the learning coefficient.
- the newly stored information is the first and the i-th information is newly numbered i + 1, and if M pieces of information have already been stored, the oldest M-th information is discarded. An M number is newly added to the (M-1) th information.
- the learning control device updates the learning coefficient update value and the instantaneous value history information, and then performs three evaluations on the learning coefficient.
- the first evaluation is an evaluation of the latest degree of the learning coefficient, and this is performed for all cells.
- the second evaluation is an evaluation of the degree of saturation of the learning coefficient, which is performed for the cell corresponding to the rolling condition of the material, that is, the cell for which history information has been updated in the current process.
- the third evaluation is an evaluation of the stability of the learning coefficient, and this is also performed for the cell.
- the processing unit 14 evaluates the latest degree of the learning coefficient.
- the processing unit 14 confirms that the update value of the learning coefficient of each cell in the second stratification table 22 is new based on the history information of the update value of the learning coefficient stored in the second stratification table 22. evaluate. If the update of each cell is updated after a certain date and time, the update value of the learning coefficient can be determined to be new. The following formula is used for this determination. ⁇ (k, l) ⁇ time Where (k, l): cell coordinates (where k, l is a combination of all rolling conditions) ⁇ (k, l): Time difference between the update date and time of the learning coefficient for the material and the update date and time of the cell (k, l)
- the processing unit 14 determines that the updated value of the learning coefficient stored in the cell (k, l) is the latest, and sets FCZtime (k, l) to 1. If the above equation is not satisfied, the processing unit 14 determines that the update value of the learning coefficient stored in the cell (k, l) is not the latest, and sets FCZtime (k, l) to 0. Note that FCZtime (k, l) is a numerical value indicating the latest evaluation result of the update value of the learning coefficient. If it is 1, it means that it is the latest, and 0 means that it is not the latest. The processing unit 14 makes the determination based on the above formula for all the cells.
- the latest evaluation value FCZtime (k, l) determined for each cell by the processing unit 14 is recorded in the corresponding cell of the third stratification table 23.
- the processing unit 15 evaluates the saturation of the learning coefficient.
- the processing unit 15 evaluates that the update value of the learning coefficient is saturated for the cell based on the history information of the update value of the learning coefficient stored in the second stratification table 22.
- ⁇ mdf Average value of the change value of the saturation criterion reference value update value
- CZmdf (r, s) is calculated by the following equation.
- the processing unit 15 determines that the update value of the learning coefficient stored in the cell (r, s) is saturated and sets FCZmdf (r, s) to 1. If the above equation is not satisfied, the processing unit 15 determines that the update value of the learning coefficient stored in the cell (r, s) is not saturated and sets FCZmdf (r, s) to 0. If M ⁇ m, the processing unit 15 sets FCZmdf (r, s) to 0. Note that FCZmdf (r, s) is a numerical value indicating the evaluation result of the degree of saturation of the updated value of the learning coefficient. If it is 1, it means that it is saturated, and if it is 0, it means that it is not saturated.
- the saturation evaluation value FCZmdf (r, s) determined by the processing unit 15 is recorded in the cell (r, s) of the fourth stratification table 24.
- Evaluation of the stability of the learning coefficient is performed by the processing unit 16.
- CZcur (r, s) ⁇ cur
- (r, s) Coordinate of the cell
- CZcur (r, s) Standard deviation of change in instantaneous value of learning coefficient of the cell
- ⁇ cur Standard deviation of change in instantaneous value of stability criterion value
- CZcur (r, s) is calculated by the following equation.
- the processing unit 16 determines that the instantaneous value of the learning coefficient stored in the cell (r, s) is stable, and sets FCZcur (r, s) to 1. If the above equation is not satisfied, the processing unit 16 determines that the instantaneous value of the learning coefficient stored in the cell (r, s) is not stable, and sets FCZcur (r, s) to 0. If N ⁇ n, the processing unit 16 sets FCZcur (r, s) to 0. Note that FCZcur (r, s) is a numerical value indicating the evaluation result of the stability of the instantaneous value of the learning coefficient. If it is 1, it means that it is stable, and if it is 0, it means that it is not stable.
- the instantaneous value and the updated value of the learning coefficient are stored for the cell corresponding to the rolling condition of the material. Furthermore, the latest degree of the learning coefficient is evaluated for all the cells, the saturation degree and the stability degree of the learning coefficient are evaluated for the cell, and the evaluation values are also stored in the corresponding cells.
- the setting calculation device 30 obtains a learning coefficient by two methods for a cell to which the rolling condition of the next material corresponds, that is, a target cell that is a model prediction target using a model formula, and applies the learning coefficient from these to the rolling of the next material.
- Use value of learning coefficient is calculated.
- One method is a method of indexing and reading the latest update value of the cell corresponding to the rolling condition of the next material from the history information of the update value of the learning coefficient stored in the second stratification table 22. Processing according to this method is performed by the processing unit 33.
- the coordinate of the cell of the next material is (t, u)
- the update value of the learning coefficient read by the processing unit 33 is Zmod (t, u).
- the other method is a method of calculating an estimated value of a learning coefficient in a cell of the next material using a learning coefficient of a cell in the vicinity of the cell of the next material. Processing according to this method is performed by the processing units 31 and 32. Specifically, first, the processing unit 31 selects a cell that satisfies a predetermined condition from cells in the vicinity of the next material cell.
- the conditions are the following three conditions: 1) Learning is updated with relatively new data. 2) The updated value of the learning coefficient is saturated. 3) The instantaneous value of the learning coefficient is stable.
- the processing unit 31 selects a cell whose evaluation result is determined to be 1 in all of the third stratified table 23, the fourth stratified table 24, and the fifth stratified table 25. Then, three cells having a spatial distance close to the cell (t, u) corresponding to the next material are extracted in order from the selected cells. Here, it is assumed that the selected three cells are cell A (ta, ua), cell B (tb, ub), and cell C (tc, uc).
- the processing unit 32 stores the instantaneous values stored in the first stratification table 21 for the cells A (ta, ua), B (tb, ub), and C (tc, uc) selected by the processing unit 31. Based on the history information, the representative value representing the latest K instantaneous values (for example, the latest K average value or median value, or K-2 averages obtained by removing the maximum value and the minimum value from the latest K values) Value or median).
- the representative value representing the latest K instantaneous values for example, the latest K average value or median value, or K-2 averages obtained by removing the maximum value and the minimum value from the latest K values
- the representative value of cell A (ta, ua) is Zestm (ta, ua)
- the representative value of cell B (tb, ub) is Zestm (tb, ub)
- the representative value of cell C (tc, uc) is Let Zestm (tc, uc).
- the processing unit 32 uses the coordinates of the selected cell A (ta, ua), cell B (tb, ub), cell C (tc, uc) and their respective coordinates.
- a first-order polynomial that holds between the representative values is specified. That is, as shown in FIG. 2, three coordinate points (ta, ua, Zestm (ta, ua)), (tb, The plane passing through (ub, Zestm (tb, ub)), (tc, uc, Zestm (tc, uc)) is obtained. This plane can be determined as satisfying the following equation.
- Zestm (t, u) a * t + b * u + c
- the coefficients a, b, and c of the plane equation are calculated as satisfying the following conditions.
- Zestm (ta, ua) a * ta + b * ua + c
- Zestm (tb, ub) a * tb + b * ub + c
- Zestm (tc, uc) a * tc + b * uc + c
- the processing unit 32 calculates the learning coefficient by the polynomial interpolation using the above formula for the cell (t, u) to which the next material corresponds, so that the learning coefficient satisfying the latest degree, the saturation degree, and the stability degree.
- the estimated value of is calculated.
- the estimated value of the learning coefficient calculated by the processing unit 34 is Zestm (t, u).
- Zuse (t, u) ⁇ * Zmod (t, u) + (1- ⁇ ) * Zestm (t, u) Where Zuse (t, u) is the value used for the learning coefficient.
- the processing unit 35 increases the learning coefficient update value Zmod (t, t, by increasing the value of the weighting coefficient ⁇ if the update value Zmod (t, u) of the learning coefficient is satisfactory in the latest degree, saturation degree, and stability. Give u) a weight. Conversely, if the latest value, saturation, and stability of the estimated value Zestm (t, u) of the learning coefficient are good, the estimated value Zestm (t, u) of the learning coefficient is reduced by decreasing the value of the weighting coefficient ⁇ . Can be given a weight.
- An example of a method for determining the weighting factor ⁇ is shown below.
- the processing unit 36 corrects the model formula with the use value of the learning coefficient calculated by the processing unit 35. Thereby, the precision improvement of the prediction calculation by a model formula is achieved. And setting calculation is performed using the corrected model formula, and the setting value of the apparatus of a rolling mill is determined. By scanning each device in accordance with the set value determined in this way, accuracy of product quality and operational stability can be improved.
- the processing unit 12 corresponds to the instantaneous value calculation recording means
- the processing unit 13 corresponds to the update value calculation recording means
- the processing unit 14 corresponds to the latest degree evaluation means
- the processing unit 15 Corresponds to the saturation evaluation means
- the processing unit 16 corresponds to the stability evaluation means.
- the processing unit 31 corresponds to a neighboring cell selection unit
- the processing unit 32 corresponds to an estimated value calculation unit
- the processing units 34 and 35 correspond to usage value determination unit.
- a routine corresponding to each processing unit is stored as one program in the memory of the computer constituting the learning control apparatus 1. When the program is executed by the processor of the computer, the computer functions as each processing unit.
- learning control device 10 learning device 20 storage device (database) 21, 22, 23, 24, 25 Stratified table 30 setting calculation device
Landscapes
- Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Automation & Control Theory (AREA)
- Mechanical Engineering (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Control Of Metal Rolling (AREA)
- Feedback Control In General (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
Abstract
Description
Znew = Zuse*(1-α) + Zcur*α
ここで、Znew: 学習係数更新値
Zcur: 学習係数(瞬時値)
Zuse: 学習係数(前回値)
α : 学習係数更新ゲイン(フィルターの時定数)
Znew(i,j) = Zuse(i,j)*(1-α) + Zcur(i,j)*α
ここで、Znew(i,j): 当該セルの学習係数更新値
Zcur(i,j): 当該セルの学習係数(瞬時値)
Zuse(i,j): 当該セルの学習係数(前回値)
α : 当該セルの学習係数更新ゲイン(フィルターの時定数)
Znew(p,q) = Zuse(p,q)*(1-α’) + Zcur(i,j)*α’
ここで、Znew(p,q): 隣接するセルの学習係数更新値
Zcur(i,j): 当該材セルの学習係数(瞬時値)
Zuse(p,q): 隣接するセルの学習係数(前回値)
α’ : 隣接するセルの学習係数更新ゲイン(フィルターの時定数)
ε(k,l) < εtime
ここで、(k,l): セルの座標(ただし、k,lは全ての圧延条件の組合せ)
ε(k,l): 当該材の学習係数の更新日時とセル(k,l)の更新日時の時間差
εtime: 最新度の判定基準値
CZmdf(r,s) < εmdf
ここで、(r,s): 当該セルの座標
CZmdf(r,s): 当該セルの学習係数の更新値の変化量の平均値
εmdf: 飽和度の判定基準値
更新値の変化量の平均値CZmdf(r,s)は、次の式によって算出される。
CZcur(r,s) < εcur
ここで、(r,s): 当該セルの座標
CZcur(r,s): 当該セルの学習係数の瞬時値の変化量の標準偏差
εcur: 安定度の判定基準値
瞬時値の変化量の標準偏差CZcur(r,s)は、次の式によって算出される。
1)比較的新しいデータで学習更新がなされている。
2)学習係数の更新値が飽和している。
3)学習係数の瞬時値が安定している。
Zestm(t,u)= a*t+b*u+c
ここで、上記平面の式の係数a,b,cは、以下の条件を満たすものとして計算される。
Zestm(ta,ua)= a*ta+b*ua+c
Zestm(tb,ub)= a*tb+b*ub+c
Zestm(tc,uc)= a*tc+b*uc+c
Zuse(t,u)=β*Zmod(t,u) + (1-β)*Zestm(t,u)
ここで、Zuse(t,u): 学習係数の使用値。
β: 重み係数(0 =< β =< 1)
β=1.0 (更新値Zmod(t,u)が最新度、飽和度、安定度とも満たす場合)
β=0.6 (更新値Zmod(t,u)が最新度、飽和度、安定度の2つを満たす場合)
β=0.4 (更新値Zmod(t,u)が最新度、飽和度、安定度の1つを満たす場合)
β=0.0 (更新値Zmod(t,u)が最新度、飽和度、安定度の1つも満たさない場合)
10 学習装置
20 記憶装置(データベース)
21,22,23,24,25 層別テーブル
30 設定計算装置
Claims (4)
- 圧延条件を区分する複数のセルから構成されるデータベースを有し、圧延プロセスの設定計算に用いるモデル式の学習係数を前記データベースにて管理する学習制御装置において、
前記圧延プロセスで計測された実績値と前記モデル式を用いて算出された実績再計算値とのズレに基づいて前記学習係数の瞬時値を算出し、当該圧延条件に該当するセルに対して前記瞬時値を学習日時とともに記録する瞬時値算出記録手段と、
前記瞬時値と前記学習係数の当該圧延条件での前回値とに基づいて前記学習係数の更新値を算出し、当該圧延条件に該当するセルに対して前記更新値を学習日時とともに記録する更新値算出記録手段と、
前記データベースに記憶された前記更新値の履歴情報に基づき、前記複数のセルのそれぞれについて前記学習係数の最新度を評価する最新度評価手段と、
前記データベースに記憶された前記更新値の履歴情報に基づき、前記複数のセルのそれぞれについて前記学習係数の飽和度を評価する飽和度評価手段と、
前記データベースに記憶された前記瞬時値の履歴情報に基づき、前記複数のセルのそれぞれについて前記学習係数の安定度を評価する安定度評価手段と、
前記最新度、飽和度、及び、安定度の各評価結果が基準を満たすセルの中から、圧延条件を座標軸とする空間において、次圧延条件に該当する対象セルとの空間距離が近い所定個数の近傍セルを選定する近傍セル選定手段と、
選定した前記所定個数の近傍セルにおける前記学習係数の代表値をそれぞれ決定し、前記対象セルの座標と選定した前記所定個数の近傍セルの座標及び前記代表値とに基づいて、多項式補間により前記対象セルにおける前記学習係数の推定値を算出する推定値算出手段と、
前記対象セルにおける前記学習係数の最新更新値を前記推定値で補正し、次圧延条件での前記学習係数の使用値として決定する使用値決定手段と、
備えることを特徴とする圧延プロセスの学習制御装置。 - 前記使用値決定手段は、
前記最新更新値と前記推定値との加重平均によって前記使用値を算出する手段と、
前記対象セルに対する前記最新度、飽和度、及び安定度の評価結果が高いほど前記最新更新値の重みが大きくなり、前記評価結果が低いほど前記推定値の重みが大きくなるように、前記評価結果に応じて前記加重平均の重み係数を変化させる手段と、
を含むことを特徴とする請求項1に記載の圧延プロセスの学習制御装置。 - 前記データベースは、セルを共通にする第1の層別テーブルと第2の層別テーブルとを含み、
前記瞬時値算出記録手段は、前記瞬時値を前記第1の層別テーブルに記録するように構成され、
前記更新値算出記録手段は、前記更新値を前記第2の層別テーブルに記録するように構成されることを特徴とする請求項1又は2に記載の圧延プロセスの学習制御装置。 - 前記データベースは、前記第1の層別テーブル及び第2の層別テーブルとセルを共通にする第3の層別テーブルと第4の層別テーブルと第5の層別テーブルとをさらに含み、
前記最新度評価手段は、前記最新度の評価結果を前記第3の層別テーブルに記録するように構成され、
前記飽和度評価手段は、前記飽和度の評価結果を前記第4の層別テーブルに記録するように構成され、
前記安定度評価手段は、前記安定度の評価結果を前記第5の層別テーブルに記録するように構成されることを特徴とする請求項3に記載の圧延プロセスの学習制御装置。
Priority Applications (5)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201480075773.4A CN106030424B (zh) | 2014-02-17 | 2014-02-17 | 轧制工艺的学习控制装置 |
| PCT/JP2014/053622 WO2015122010A1 (ja) | 2014-02-17 | 2014-02-17 | 圧延プロセスの学習制御装置 |
| US15/119,313 US10124381B2 (en) | 2014-02-17 | 2014-02-17 | Rolling process learning control device |
| JP2015562674A JP6233423B2 (ja) | 2014-02-17 | 2014-02-17 | 圧延プロセスの学習制御装置 |
| KR1020167025045A KR101832653B1 (ko) | 2014-02-17 | 2014-02-17 | 압연 프로세스의 학습 제어 장치 |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| PCT/JP2014/053622 WO2015122010A1 (ja) | 2014-02-17 | 2014-02-17 | 圧延プロセスの学習制御装置 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2015122010A1 true WO2015122010A1 (ja) | 2015-08-20 |
Family
ID=53799772
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/JP2014/053622 Ceased WO2015122010A1 (ja) | 2014-02-17 | 2014-02-17 | 圧延プロセスの学習制御装置 |
Country Status (5)
| Country | Link |
|---|---|
| US (1) | US10124381B2 (ja) |
| JP (1) | JP6233423B2 (ja) |
| KR (1) | KR101832653B1 (ja) |
| CN (1) | CN106030424B (ja) |
| WO (1) | WO2015122010A1 (ja) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020129173A1 (ja) * | 2018-12-19 | 2020-06-25 | 東芝三菱電機産業システム株式会社 | 圧延プロセスの学習制御装置 |
| JP2022106531A (ja) * | 2021-01-07 | 2022-07-20 | 東芝三菱電機産業システム株式会社 | 圧延モデルの学習方法 |
| JPWO2022239157A1 (ja) * | 2021-05-12 | 2022-11-17 | ||
| JPWO2023203691A1 (ja) * | 2022-04-20 | 2023-10-26 |
Families Citing this family (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106777594B (zh) * | 2016-12-01 | 2020-01-17 | 武汉科技大学 | 一种基于连续曲面的轧制过程数学模型的自学习方法 |
| JP6813416B2 (ja) * | 2017-04-10 | 2021-01-13 | 株式会社日立製作所 | プラント制御装置およびその制御方法、圧延機制御装置およびその制御方法並びにプログラム |
| JP7442055B2 (ja) * | 2018-09-03 | 2024-03-04 | パナソニックIpマネジメント株式会社 | 電子密度推定方法、電子密度推定装置、及び電子密度推定プログラム |
| CN110918651B (zh) * | 2019-12-06 | 2020-12-01 | 东北大学 | 基于多分类支持向量机的钢铁冷轧精整生产原料分配方法 |
| US11919060B2 (en) * | 2021-08-16 | 2024-03-05 | The Bradbury Co., Inc. | Methods and apparatus to control roll-forming processes |
| CN114486656B (zh) * | 2021-12-31 | 2022-12-02 | 扬州江净空调制造有限公司 | 用于医疗洁净室的动态环境监测系统 |
| CN116140374B (zh) * | 2023-04-14 | 2023-07-14 | 太原科技大学 | 一种板带轧制过程质量综合预测与工艺调控方法 |
| WO2025253454A1 (ja) * | 2024-06-03 | 2025-12-11 | 株式会社Tmeic | 予測モデルの学習システムおよび学習方法 |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH07200005A (ja) * | 1993-12-28 | 1995-08-04 | Mitsubishi Electric Corp | 学習制御方法 |
| JPH1031505A (ja) * | 1996-07-16 | 1998-02-03 | Mitsubishi Electric Corp | プロセスラインの学習制御方法 |
| JP2839746B2 (ja) * | 1991-06-17 | 1998-12-16 | 株式会社神戸製鋼所 | プロセスラインにおける学習制御方法 |
| JP2000263110A (ja) * | 1999-03-10 | 2000-09-26 | Toshiba Corp | 圧延制御モデル学習装置 |
Family Cites Families (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP3508245B2 (ja) * | 1994-11-04 | 2004-03-22 | Jfeスチール株式会社 | 圧延条件推定モデルによる圧延条件の推定方法 |
| JP4179709B2 (ja) * | 1999-07-15 | 2008-11-12 | 東芝三菱電機産業システム株式会社 | リバース式圧延機の学習制御装置 |
| WO2008018805A2 (en) * | 2006-08-09 | 2008-02-14 | Auckland Uniservices Limited | Process control of an industrial plant |
| JP2009116759A (ja) | 2007-11-09 | 2009-05-28 | Jfe Steel Corp | プロセスラインにおける制御モデル学習方法および装置、ならびに鋼板の製造方法 |
| CN101648208A (zh) * | 2009-08-26 | 2010-02-17 | 苏州有色金属研究院有限公司 | 铝单机架冷轧机的轧制规程制定方法 |
| CN102371279B (zh) * | 2010-08-26 | 2013-06-19 | 宝山钢铁股份有限公司 | 利用辊缝自适应提高精轧带钢厚度精度的控制方法 |
| JP5759206B2 (ja) * | 2011-03-01 | 2015-08-05 | 東芝三菱電機産業システム株式会社 | 学習係数制御装置 |
| CN102728624B (zh) * | 2011-04-13 | 2014-10-01 | 宝山钢铁股份有限公司 | 一种精轧带钢负荷分配设定方法 |
| CN102513351B (zh) * | 2011-12-24 | 2014-01-15 | 东北大学 | 一种带钢冷连轧轧制方法及装置 |
| WO2013171862A1 (ja) * | 2012-05-16 | 2013-11-21 | 東芝三菱電機産業システム株式会社 | 設定計算システムの学習装置及び学習方法 |
| CN104271277B (zh) * | 2012-07-02 | 2016-01-13 | 东芝三菱电机产业系统株式会社 | 温度控制装置 |
| CN103018604A (zh) * | 2012-12-07 | 2013-04-03 | 浙江省电力公司电力科学研究院 | 一种电网设施雷击风险的评估方法及其装置 |
| JP5939175B2 (ja) | 2013-02-19 | 2016-06-22 | 東芝三菱電機産業システム株式会社 | 圧延プロセスの学習制御装置 |
-
2014
- 2014-02-17 JP JP2015562674A patent/JP6233423B2/ja active Active
- 2014-02-17 CN CN201480075773.4A patent/CN106030424B/zh active Active
- 2014-02-17 WO PCT/JP2014/053622 patent/WO2015122010A1/ja not_active Ceased
- 2014-02-17 KR KR1020167025045A patent/KR101832653B1/ko active Active
- 2014-02-17 US US15/119,313 patent/US10124381B2/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2839746B2 (ja) * | 1991-06-17 | 1998-12-16 | 株式会社神戸製鋼所 | プロセスラインにおける学習制御方法 |
| JPH07200005A (ja) * | 1993-12-28 | 1995-08-04 | Mitsubishi Electric Corp | 学習制御方法 |
| JPH1031505A (ja) * | 1996-07-16 | 1998-02-03 | Mitsubishi Electric Corp | プロセスラインの学習制御方法 |
| JP2000263110A (ja) * | 1999-03-10 | 2000-09-26 | Toshiba Corp | 圧延制御モデル学習装置 |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020129173A1 (ja) * | 2018-12-19 | 2020-06-25 | 東芝三菱電機産業システム株式会社 | 圧延プロセスの学習制御装置 |
| JPWO2020129173A1 (ja) * | 2018-12-19 | 2021-02-15 | 東芝三菱電機産業システム株式会社 | 圧延プロセスの学習制御装置 |
| JP2022106531A (ja) * | 2021-01-07 | 2022-07-20 | 東芝三菱電機産業システム株式会社 | 圧延モデルの学習方法 |
| JP7452443B2 (ja) | 2021-01-07 | 2024-03-19 | 東芝三菱電機産業システム株式会社 | 圧延モデルの学習方法 |
| JPWO2022239157A1 (ja) * | 2021-05-12 | 2022-11-17 | ||
| WO2022239157A1 (ja) * | 2021-05-12 | 2022-11-17 | 東芝三菱電機産業システム株式会社 | 圧延プロセスの学習制御装置 |
| JP7323051B2 (ja) | 2021-05-12 | 2023-08-08 | 東芝三菱電機産業システム株式会社 | 圧延プロセスの学習制御装置 |
| JPWO2023203691A1 (ja) * | 2022-04-20 | 2023-10-26 | ||
| WO2023203691A1 (ja) * | 2022-04-20 | 2023-10-26 | 東芝三菱電機産業システム株式会社 | 板クラウン制御装置 |
| CN117320822A (zh) * | 2022-04-20 | 2023-12-29 | 东芝三菱电机产业系统株式会社 | 板凸度控制装置 |
| JP7736177B2 (ja) | 2022-04-20 | 2025-09-09 | 株式会社Tmeic | 板クラウン制御装置 |
Also Published As
| Publication number | Publication date |
|---|---|
| KR20160122208A (ko) | 2016-10-21 |
| KR101832653B1 (ko) | 2018-02-26 |
| CN106030424B (zh) | 2019-03-15 |
| US20170008054A1 (en) | 2017-01-12 |
| JP6233423B2 (ja) | 2017-11-22 |
| JPWO2015122010A1 (ja) | 2017-03-30 |
| US10124381B2 (en) | 2018-11-13 |
| CN106030424A (zh) | 2016-10-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| JP6233423B2 (ja) | 圧延プロセスの学習制御装置 | |
| JP5604945B2 (ja) | 品質予測装置、品質予測方法、コンピュータプログラム、およびコンピュータ読み取り可能な記録媒体 | |
| JP2022132895A (ja) | 合金材料の特性を予測する製造支援システム、予測モデルを生成する方法およびコンピュータプログラム | |
| CN102371279A (zh) | 利用辊缝自适应提高精轧带钢厚度精度的控制方法 | |
| JP2012081518A (ja) | 厚鋼板の冷却制御方法、冷却制御装置及び厚鋼板の製造方法 | |
| CN107045585A (zh) | 一种轧制模型自学习系数确定方法 | |
| JPH11707A (ja) | 圧延機の制御方法及び圧延機の制御装置 | |
| JP7449776B2 (ja) | 熱間圧延ライン制御システムおよび熱間圧延ライン制御方法 | |
| JP6699340B2 (ja) | 厚鋼板の冷却制御方法および製造方法 | |
| JP5817450B2 (ja) | 厚板の板幅制御方法 | |
| JP5939175B2 (ja) | 圧延プロセスの学習制御装置 | |
| JP6900866B2 (ja) | 圧延スケジュール作成装置、圧延スケジュール作成方法、およびプログラム | |
| JP2011220708A (ja) | 鋼材の材質予測装置 | |
| JP2021194690A (ja) | 板厚制御装置、板厚制御方法、プログラム、及び圧延システム | |
| JP6020263B2 (ja) | 圧延荷重の学習制御装置および学習制御方法、並びにこれを用いた金属板製造方法 | |
| JP2007245204A (ja) | 圧延荷重モデルの学習方法及びその装置 | |
| JP2020099982A (ja) | 工作機械の熱変位補正方法、熱変位補正プログラム、熱変位補正装置 | |
| JP7323051B2 (ja) | 圧延プロセスの学習制御装置 | |
| CN119098490B (zh) | 一种基于温度反馈的热轧厚度控制方法、装置及设备 | |
| JP6816629B2 (ja) | 製造プロセス学習制御装置 | |
| JP7252459B2 (ja) | パススケジュール決定装置、パススケジュール決定方法、およびプログラム | |
| JP4155174B2 (ja) | 数式モデルにおける予測誤差補正用テーブル値生成装置及びその方法 | |
| CN114741835A (zh) | 轧制模型的学习方法 | |
| JP6767356B2 (ja) | モデル構築システムおよびモデル構築方法 | |
| CN116955975B (zh) | 一种变形抗力自学习系数诊断方法、装置、设备和介质 |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 14882434 Country of ref document: EP Kind code of ref document: A1 |
|
| ENP | Entry into the national phase |
Ref document number: 2015562674 Country of ref document: JP Kind code of ref document: A |
|
| WWE | Wipo information: entry into national phase |
Ref document number: 15119313 Country of ref document: US |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| ENP | Entry into the national phase |
Ref document number: 20167025045 Country of ref document: KR Kind code of ref document: A |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 14882434 Country of ref document: EP Kind code of ref document: A1 |

