Prediction method for mold breakout based on feature vectors and hierarchical clustering

Abstract

A prediction method for mold breakout based on feature vectors and hierarchical clustering is disclosed, which comprises: respectively extracting temperature feature vectors of historical data under sticking breakout and normal conditions and on-line actually measured data to establish a feature vector sample set; performing normalization and hierarchical clustering on the sample set; and checking and judging whether the feature vectors extracted on line belong to a breakout cluster, and then identifying and predicting mold breakout. The method avoids the steps of tedious adjustment and modification of alarm threshold and other parameters, overcomes the artificial dependence of the previous breakout prediction method, has good robustness and mobility; and through temperature feature extraction, achieves accurate identification of sticking breakout temperature patterns, avoids missing alarms and significantly reduces the number of times of false alarms, and greatly reduces the data calculation amount and calculation time, guaranteeing the timeliness of on-line prediction.

Claims

1. A prediction method for mold breakout based on feature vectors and hierarchical clustering comprising the following parts: respectively extracting temperature feature vectors of historical data of sticking breakout and normal conditions and on-line actually measured data, and establishing a feature vector sample set; performing normalization and hierarchical clustering on the sample set; and checking and judging whether the feature vectors extracted on line belong to a breakout cluster, and then identifying and predicting mold breakout comprising the following steps: first step of extracting sticking breakout feature vectors comprising: (1) acquiring historical temperature data of sticking breakout by marking the time when the temperature of a first row of thermocouples in a thermocouple column where a sticking position is located is the maximum, and selecting the temperature data within a first M seconds and a last N−1 seconds, M+N seconds in total; and (2) extracting and constructing temperature feature vectors of the first row and a second row of thermocouples; second step of extracting normal condition feature vectors comprising: (1) acquiring historical temperature data under normal conditions by randomly intercepting temperature data at continuous M+N seconds; and (2) extracting and constructing temperature feature vectors of the first and second rows of thermocouples; third step of extracting on-line real-time temperature feature vectors comprising: (1) collecting and acquiring temperature data of thermocouples of each row and each column on loosed wide face, fixed wide face, left narrow face, and right narrow face copperplates of a mold within the current time and previous M+N−1 seconds, M+N seconds in total in real time; and (2) extracting and constructing temperature feature vectors of the first and second rows of thermocouples; fourth step of establishing a feature vector library comprising: (1) establishing a feature vector sample library D based on the sticking breakout, normal condition and on-line actually measured temperature feature extracted in the first step, the second step and the third step; and (2) performing normalization on the feature vector sample library D, to obtain a feature vector sets, and recording a normalized on-line actually measured temperature feature as S.sub.new the normalization method for the feature vector is as follows: ${\mathrm{x\_nor}}_{\mathrm{ij}}=\frac{x_{\mathrm{ij}}-x_{\mathrm{jmin}}}{x_{\mathrm{jmax}}-x_{\mathrm{jmin}}},i=1,2,\mathrm{.Math.},.Math.S.Math.,j=1,2,\mathrm{.Math.},5$ where x.sub.ij, represents the value of the j-dimensional feature of the ith feature vector in the feature vector set s, x.sub.jmax and x.sub.jmin respectively represent a maximum and a minimum of the j-dimensional feature of all feature vectors, and S represents the total number of vectors in the feature vector set s; fifth step of performing hierarchical clustering on feature vectors comprising: (1) performing hierarchical clustering on the feature vector sets obtained in the fourth step, the specific process including: 1.1) taking each vector sin the feature vector set S as an initial cluster C.sub.i={S.sub.i}, and establishing a cluster set C={C.sub.1, C.sub.2, . . . , C.sub.k}, where s.sub.i, represents the ith vector in S, C.sub.i, represents the ith cluster, i=1,2, . . . , k, k represents the total number of vectors in the feature vector set s 1.2) calculating and determining a distance between any two clusters C.sub.p and C.sub.q in the cluster set C:
d(C.sub.p,C.sub.q)=min(dist(C.sub.pi,C.sub.qj)) where C, represents the ith feature vector in the cluster C.sub.p, C.sub.qj represents the jth feature vector in the cluster Cq, and dist(C.sub.pi, C.sub.qj) represents the Euclidean distance between the feature vectors C.sub.pi and C.sub.qj; and calculating the distance between any two vectors in the clusters C.sub.p and C.sub.q and taking the minimum distance min as the distance between the clusters C.sub.p and C.sub.q; 1.3) marking two clusters C.sub.m and C.sub.n between which the distance is the minimum calculated in the step 1.2, merging C.sub.m and C.sub.n into a new cluster C.sub.{m,n} and adding same to the set C, and deleting the original clusters C.sub.m and C.sub.n, so after cluster addition and deletion, the total number of clusters in the set C is reduced by one at this time; 1.4) performing steps 1.2-1.3 in a loop, when there are only two clusters in the cluster set C, ending the loop, and completing the clustering process; (2) checking whether the clustering result meets the following judgement condition, that is: more than 90% of all sticking breakout feature vectors belong to the same cluster, and the percentage of the normal condition feature vectors in the cluster is less than 20%; if this condition is met, recording this cluster as a breakout cluster C.sub.breakout and recording the other cluster as a normal condition cluster C.sub.normal; otherwise, performing the fifth step again until the clustering result of the feature vector set composed of breakout, normal condition, and actually measured temperature feature meets the above judgement condition; sixth step of identifying breakout and issuing alarm comprising: judging whether the new feature vector s.sub.new, belongs to the cluster C.sub.breakout, if so, issuing a breakout alarm; otherwise, continuing to perform the third through sixth steps; the temperature feature extraction methods involved in the first step (2), the second step (2) and the third step (2) being identical, extracting features of change in the temperature of the same column of thermocouples in a casting direction under different working conditions by using each column of thermocouples as a unit, specifically including: 1st_Rising_Amplitude: first row temperature rising amplitude; 1st_Rising_V_Max: first row temperature rising velocity maximum; 1st_Falling_V_Ave: first row temperature falling velocity average; 2nd_Rising_V_Max: second row temperature rising velocity maximum; 1st_2nd_Time_Lag: temperature rising time lag, that is, time interval between the time when the temperature of the second row of thermocouple starts to rise and the time when the temperature of the first row of thermocouple starts to rise; and thus constructing a feature vector: s=[1st_Rising_Amplitude, 1st_Rising_V_Max, 1st_Falling_V_Ave, 2nd_Rising_V_Max, 1st_2nd_Time_Lag].

2. The prediction method for mold breakout based on feature vectors and hierarchical clustering according to claim 1, wherein the prediction method for breakout is suitable for on-line prediction of breakout of slabs, billets, round billets and beam blanks during continuous casting.

Description

DESCRIPTION OF DRAWINGS

(1) FIG. 1 is a schematic diagram showing distribution of four mold copperplates and thermocouples;

(2) FIG. 2 is a schematic diagram showing extraction of feature vectors of temperature and change rate thereof during sticking breakout;

(3) FIG. 3 is a schematic diagram showing extraction of feature vectors of temperature and change rate thereof under normal conditions;

(4) FIG. 4 is a flow chart showing hierarchical clustering of a feature vector set and breakout identification and warning;

(5) FIG. 5 shows on-line actually measured temperature 1;

(6) FIG. 6 shows a hierarchical clustering and breakout prediction result containing an on-line actually measured temperature feature vector 1;

(7) FIG. 7 shows on-line actually measured temperature 2;

(8) FIG. 8 shows a hierarchical clustering and breakout prediction result containing an on-line actually measured temperature feature vector 2.

DETAILED DESCRIPTION

(9) The present invention will be further illustrated below through specific embodiments in combination with drawings.

(10) The present invention mainly comprises six parts: extracting sticking breakout feature vectors, extracting normal condition feature vectors, extracting on-line real-time temperature feature vectors, establishing feature vector library, performing hierarchical clustering on feature vectors, and identifying breakout and issuing alarm.

(11) FIG. 1 is a schematic diagram showing distribution of four mold copperplates and thermocouples. A slab continuous casting mold is composed of four copperplates, including a fixed wide face copperplate, a left narrow face copperplate, a loosed wide face copperplate and a right narrow face copperplate respectively having a length L of 900 mm, two rows of measuring points are arranged on the horizontal cross-section of the four copperplates L1 (210 mm) and L2 (325 mm) away from the upper opening of the mold, 19 columns of thermocouples are arranged in each row on the fixed wide face copperplate and the loosed wide face copperplate, with the distance L3 between two thermocouples is 150 mm, and each of the two wide face copperplates is provided with 38 thermocouples; each of the left narrow face copperplate and the right narrow face copperplate is provided with a column of thermocouples at the centerline, and each of the narrow face copperplates is provided with 2 thermocouples. The four copperplates are provided with 80 thermocouples in total, and the distances from all thermocouples to the mold copperplate hot face are equal.

(12) First Step: Extracting Sticking Breakout Feature Vectors

(13) (1) acquiring historical temperature data of sticking breakout: marking the time when the temperature of the first row of thermocouples in the thermocouple column where the sticking position is located is the maximum, and selecting the temperature data within the first 15 seconds and the last 9 seconds, 25 seconds in total; and

(14) for the historical temperature of sticking breakout, selecting 30 temperature samples;

(15) (2) extracting and constructing temperature feature vectors of the first and second rows of thermocouples.

(16) FIG. 2 is a schematic diagram showing extraction of feature vectors of sticking breakout temperature and temperature change rate. As shown in FIG. 2, in feature vector extraction, the features of change in the temperature of the same column of thermocouples in the casting direction during sticking breakout are extracted respectively by taking each column of thermocouples as a unit, the specific features and extraction method therefor being as follows: 2.1) Calculating the change rate of temperature data within 5 seconds at the same measuring point, that is:

(17) $v_{\left(r\right)i}=\frac{T_{\left(r\right)i+5}-T_{\left(r\right)i}}{5},i\in \left[1,2,\mathrm{.Math.},20\right]$ where r ∈ [1,2] respectively represents the first and second rows of thermocouples, T.sub.(r)i represents a temperature value of the r.sup.th row of thermocouples at the i.sup.th time, and v.sub.(r)i represents the value of the temperature change rate of the r.sup.th row of thermocouples at the i.sup.th time. 2.2) Determining the starting temperature and time corresponding thereto when the temperature rises:

(18) firstly, acquiring the maximum temperature T.sub.max within 25 seconds in total including the current time and the previous time, and time t.sub.max corresponding thereto, that is:
T.sub.max=max(T.sub.i),i=1,2, . . . ,25,

(19) secondly, traversing forward from T.sub.max, acquiring the previous minimum temperature T.sub.min and time t.sub.min thereof, that is,
T.sub.min=min(T.sub.i),i=1,2, . . . ,t.sub.max
taking T.sub.min, as the starting temperature when the temperature rises, and recording the time corresponding thereto as t.sub.min. 2.3) Extracting corresponding features to construct a feature vector:

(20) 1st_Rising_Amplitude: first row temperature rising amplitude, that is, respectively marking the rising temperature of the first row of thermocouples and the maximum temperature, calculating the difference between the two temperatures, and obtaining a temperature rising amplitude, in □;
1st_Rising_Amplitude=T.sub.(1)max−T.sub.(1)min

(21) 1st_Rising_V_Max: first row temperature rising velocity maximum, that is, extracting the maximum temperature change velocity of the first row of thermocouples obtained in step 2.1), in □/s;
1st_Rising_V_Max=max(v.sub.(l))

(22) 1st Falling_V_Ave: first row temperature falling velocity average, that is, calculating the average temperature velocity of the first row of thermocouples whose temperature change velocity is less than 0 obtained in the step 2.1), in □/s;

(23) $1\mathrm{st\_Falling}\mathrm{\_V}\mathrm{\_Ave}=\underset{v_{\left(1\right)i}<0}{\mathrm{Average}}\left(v_{\left(1\right)}\right),i\in \left[1,2,\mathrm{.Math.},20\right]$

(24) 2nd_Rising_V_Max: second row temperature rising velocity maximum, that is, extracting the maximum temperature change velocity of the second row of thermocouples obtained in the step 2.1), in □/s;
2nd_Rising_V_Max=max(v.sub.(2))

(25) 1st_2nd_Time_Lag: temperature rising time lag, that is, respectively marking the times corresponding to the time when the temperature of the second row of thermocouples starts to rise and the time when the temperature of the first row of thermocouples starts to rise, and taking the time interval between the two as the temperature rising time lag, in s;
1st_2nd_Time_Lag=t.sub.(2)min−t.sub.(1)min

(26) and thus constructing a feature vector: s=[1st_Rising_Amplitude, 1st_Rising_V_Max, 1st_Falling_V_Ave, 2nd_Rising_V_Max, 1st_2nd_Time_Lag]

(27) Second Step: Extracting Normal Condition Feature Vectors

(28) (1) acquiring historical temperature data under normal conditions: randomly intercepting temperature data within continuous 25 seconds;

(29) (2) extracting and constructing temperature feature vectors of the first and second rows of thermocouples.

(30) FIG. 3 is a schematic diagram showing extraction of feature vectors of temperature and temperature change rate under normal conditions. As shown in FIG. 3, in feature vector extraction, the features of change in the temperature of the same column of thermocouples in the casting direction under normal conditions are extracted respectively by taking each column of thermocouples as a unit, the features and extraction method therefor being the same as that in the first step (2).

(31) For the normal condition temperature, 30 temperature samples are selected.

(32) FIG. 4 is a flow chart showing hierarchical clustering of a feature vector set and breakout identification and determination. As shown in the figure, hierarchical clustering of a feature vector set and breakout identification and determination mainly include the following steps:

(33) Third Step: Extracting On-Line Real-Time Temperature Feature Vectors

(34) (1) collecting and acquiring temperature data of thermocouples of each row and each column on loosed wide face, fixed wide face, left narrow face, and right narrow face copperplates of the mold within the current time and previous 24 seconds, 25 seconds in total in real time;

(35) (2) extracting and constructing temperature feature vectors of the first and second rows of thermocouples:

(36) in feature vectors extraction, the features of change in the temperature of the same column of thermocouples in the casting direction during on-line actual measurement are extracted respectively by taking each column of thermocouples as a unit, the features and extraction method therefor being the same as that in the first step (2).

(37) Fourth Step: Establishing Feature Vector Library

(38) (1) establishing a feature vector sample library D based on the sticking breakout, normal condition and on-line actually measured temperature feature extracted in the first step, the second step and the third step, 61 in total;

(39) (2) performing normalization on the feature vector sample library D, to obtain a feature vector set S, and recording a normalized on-line actually measured temperature feature as s.sub.new. The specific normalization method is as follows:

(40) ${\mathrm{x\_nor}}_{\mathrm{ij}}=\frac{x_{\mathrm{ij}}-x_{\mathrm{jmin}}}{x_{\mathrm{jmax}}-x_{\mathrm{jmin}}},i=1,2,\mathrm{.Math.},61,j=1,2,\mathrm{.Math.},5$
where x.sub.ij represents the value of the j-dimensional feature of the i.sup.th feature vector in the feature vector set S, and x.sub.jmax and x.sub.jmin respectively represent a maximum and a minimum of the j-dimensional feature of all the 61 feature vectors.

(41) Fifth Step: Performing Hierarchical Clustering on Feature Vectors

(42) (1) performing hierarchical clustering on the feature vector set S obtained in the fourth step;

(43) 1.1) taking each vector s in the feature vector set S as an initial cluster C.sub.i={s.sub.i}, and establishing a cluster set C={(C.sub.1, C.sub.2, . . . , C.sub.k}, s.sub.i represents the i.sup.th vector in S, and C.sub.i represents the i.sup.th cluster, i=1,2, . . . , 61;

(44) 1.2) calculating and determining the distance between any two clusters C.sub.p and C.sub.q in the cluster set C:
d(C.sub.p,C.sub.q)=min(dist(C.sub.pi,C.sub.qj))

(45) where C.sub.pi represents the i.sup.th feature vector in the cluster C.sub.p, C.sub.qj represents the j.sup.th feature vector in the cluster C.sub.q, and dist(C.sub.pi, C.sub.qj) represents the Euclidean distance between the feature vectors C.sub.pi and C.sub.qj; and calculating the distance between any two vectors in the clusters C.sub.p and C.sub.q, and taking the minimum distance min as the distance between the clusters C.sub.p and C.sub.q;

(46) 1.3) marking the two clusters C.sub.m and C.sub.n between which the distance is the minimum calculated in the step 1.2), merging C.sub.m and C.sub.n into a new cluster C.sub.{m,n} and adding same to the set C, and deleting the original clusters C.sub.m and C.sub.n, so after cluster addition and deletion, the total number of clusters in the set C is reduced by one at this time;

(47) 1.4) performing steps 1.2)-1.3) in a loop, when there are only two clusters in the cluster set C, ending the loop, and completing the clustering process;

(48) (2) checking whether the clustering result meets the following judgement condition, that is:

(49) more than 90% of all sticking breakout feature vectors belong to the same cluster, and the percentage of the normal condition feature vectors in the cluster is less than 20%;

(50) if this condition is met, recording this cluster as a breakout C.sub.breakout, and recording the other cluster as a normal condition cluster C.sub.normal; otherwise, performing steps (1) and (2) again until the clustering result of the feature vector set composed of breakout, normal condition, and actually measured temperature feature meets the above judgement condition;

(51) Sixth Step: Identifying Breakout and Issuing Alarm

(52) judging whether the new feature vector s.sub.new belongs to the cluster C.sub.breakout, if so, issuing a breakout alarm; otherwise, continuing to perform the third steps, the fourth step, the fifth step and the sixth step.

(53) FIG. 5 is a diagram showing temperature of the first and second rows of thermocouples that represents the on-line actually measured temperature 1. The right vertical line in the figure represents temperature data within 25 seconds in total including the current time when on-line detection is performed, and the 24 times before that time. The feature vector obtained after performing feature extraction on the on-line temperature 1 is:
s.sub.new=[0.4,0.18,−1.18,1.72,0].

(54) FIG. 6 is a diagram showing a hierarchical clustering and breakout prediction result of a feature vector set containing an on-line actually measured temperature feature vector 1, that is, containing s.sub.new. As shown in the figure, after normalization and hierarchical clustering, the feature vector set is clustered into two clusters: the left cluster includes all the sticking breakout feature vectors labeled 1 and five normal condition samples labeled 2, the percentage of sticking breakout feature vectors in this cluster is greater than 90% of the total number of sticking breakout samples, and the percentage of normal condition feature vectors is less than 20% of the total number of normal condition samples, that is, six feature vectors, the cluster judgement condition is met, the hierarchical clustering of feature vectors is successful, so the cluster is recorded as a breakout cluster C.sub.breakout; and then the other cluster including more normal condition samples labeled 2 is recorded as a normal condition cluster C.sub.normal. As shown in FIG. 6, the feature vector s.sub.new obtained by performing feature extraction on the on-line actually measured temperature 1, that is, the sample labeled “N”, belongs to the normal condition cluster C.sub.normal after clustering, but does not belong to the sticking breakout cluster C.sub.breakout. Therefore, it is judged as normal condition, the temperature sequence is updated continuously, and the third steps, the fourth step, the fifth step and the sixth step are performed.

(55) FIG. 7 is a diagram showing the temperature of the first and second rows of thermocouples that represents the on-line actually measured temperature 2. The right vertical line in the figure represents temperature data within 25 seconds in total including the current time when on-line detection is performed, and the 24 times before that time. The feature vector obtained after performing feature extraction on the on-line temperature 2 is:
s.sub.new=[7.4,1.06,−0.29,1.36,12].

(56) FIG. 8 is a diagram showing a hierarchical clustering and breakout prediction result of a feature vector set containing an on-line actually measured temperature feature vector 2, that is, containing s.sub.new. As shown in the figure, after normalization and hierarchical clustering, the feature vector set is clustered into two clusters: the left cluster includes all the sticking breakout samples labeled 1 and five normal condition samples labeled 2, the percentage of sticking breakout feature vectors in this cluster is greater than 90% of the total number of sticking breakout samples, and the percentage of normal condition feature vectors is less than 20% of the total number of normal condition samples, that is, six feature vectors, the cluster judgement condition is met, the hierarchical clustering of feature vectors is successful, so the cluster is recorded as a breakout cluster C.sub.breakout; and then the other cluster including more normal condition samples labeled 2 is recorded as a normal condition cluster C.sub.normal. As shown in FIG. 8, the feature vector s.sub.new obtained by performing feature extraction on the on-line actually measured temperature 2, that is, the sample labeled “B”, belongs to the sticking breakout cluster C.sub.breakout after clustering. Therefore, it is judged as breakout, a breakout alarm is issued.

(57) The above embodiments only express the implementation of the present invention, and shall not be interpreted as a limitation to the scope of the patent for the present invention. It should be noted that, for those skilled in the art, several variations and improvements can also be made without departing from the concept of the present invention, all of which belong to the protection scope of the present invention.