EP2721453A1

EP2721453A1 - Method for verifying process parameters of a manufacturing process

Info

Publication number: EP2721453A1
Application number: EP11837306.7A
Authority: EP
Inventors: Mikhail Aleksandrovich KALINKIN; Bernhard Lang; Alexander Vladimirovich LOGINOV; Alexey Sergeyevich MININ; Denis Yurievich SARAEV
Original assignee: Siemens AG; Siemens Corp
Current assignee: Primetals Technologies Germany GmbH
Priority date: 2011-09-29
Filing date: 2011-09-29
Publication date: 2014-04-23
Also published as: CN103842920B; WO2013048274A1; CN103842920A

Abstract

The invention describes a method for verifying process parameters of a manufacturing process for prediction of rejections, in particular a hot rolling process of a steel strip for prediction of coil breaks. A database is provided in which a plurality of sets of known process parameters is stored wherein each set comprises an information about the occurrence of rejections and a new set of process parameters which shall be used for manufacturing. The new set of process parameters is processed to find a similar set of process parameters from the stored sets of process parameters in the database such that the process parameters of the new set and the process parameters of the similar set meet one or more predetermined requirements. The information about the occurrence of rejections is evaluated from the similar set of parameters. The information is processed to make a prediction of rejections for the new set of process parameters.

Description

METHOD FOR VERIFYING PROCESS PARAMETERS OF A

MANUFACTURING PROCESS

The invention relates to a method for verifying process parameters of a manufacturing process for prediction of rejections. More specifically, the invention relates to a method for verifying process parameters of a hot rolling process of a steel strip for prediction of coil breaks.

In the following description reference is made to a manufacturing process of hot rolling and pickling of a steel strip. It is to be understood, that the hot rolling process of a steel strip is only an example of a manufacturing process. The problems described and the solution to said problems can be transferred to any manufacturing process in which rejections is a problem.

In steel industry rolling is one of the key processes which aims to reduce the thickness of slabs by compressive forces applied through a set of rolls. In a hot rolling process steel is being heated to about 1.250 °C in a furnace. After being heated in the furnace, slabs are rolled in roughing and finishing mills. The obtained steel strip is cooled by water. Thereafter, the cooled steel strip is coiled and stored in a coil storage for final cooling. Next, the coiled steel strip is uncoiled again, etched in acid and coiled again. During the rolling process (coiling and uncoiling) often various surface defects occur. One type of defects is so-called coil breaks that occur during uncoiling. Steel strips having coil breaks often cannot be reprocessed and therefore are rejections.

There are typically more than 100 process parameters that govern rolling process. All of them may potentially influence coil break occurrence. Therefore, exact reasons of coil breaks are not known. Moreover, process parameters of two strips can be very similar. However, in one case there is a coil break and in another there is no coil break.

Hence, there is a need to predict coil breaks before they occur based on current process parameters of a production system. Furthermore, there is another need to determine safe range of process parameters where the risk of coil break occurrence is minimal. In known solutions there are mechanical means to apply a special supporting and/or uniform pressure during the uncoiling process. However, such mechanical means needs maintenance and makes the production system more complicated.

It is therefore an object of the present invention to provide a method for verify- ing process parameters of a manufacturing process for the prediction of rejections. More specifically, it is an object of the present invention to provide a method for verifying process parameters of a hot rolling process of a steel strip for the prediction of coil breaks.

These objects are solved by a method according to the features of claim 1. Preferred embodiments are set out in the dependent claims.

The invention suggests a method for verifying process parameters of a manufacturing process for prediction of rejections, in particular a hot rolling process of a steel strip for prediction of coil breaks. A database is provided in which a plurality of sets of known process parameters is stored wherein each set comprises an information about the occurrence of rejections, in particular of coil breaks. A new set of process parameters is provided which shall be used for manufacturing of a product, in particular a steel strip. The new set of process parameters is processed to find a similar set of process parameters from the stored sets of process parameters in the database such that the proc- ess parameters of the new set and the process parameters of the similar set meet one or more predetermined requirements. The information about the occurrence of rejections, in particular coil breaks, is evaluated from the similar set of parameters. The information to make a prediction of rejections, in particular of coil breaks, is processed for the new set of process parameters.

The method according to the invention is based on a machine learning method for the prediction of rejections. In particular, the machine learning method is able to predict surface defects in a hot rolling process of a steel strip. The method according to the invention uses a database (which is referred to as a case base) in which all known sets of parameters with the information about rejections are stored. Based on this knowledge a prediction can be made automatically by referring to similar cases, i.e. sets of process parameters, in the past.

The method may in particular be applied to the process of hot rolling of steel strips. However, it is to be understood, that the method according to the invention may be used in any manufacturing process in which lots of different process parameters have to be controlled. The method assists in reducing rejections.

According to a preferred embodiment the step of processing the new set of parameters comprises application of k-Nearest Neighbors (k-NN) algorithm, wherein the predetermined requirement is a matrix. The k-NN algorithm is well known from the prior art and may be used to find a similar set of process parameters with regard to the new set of process parameters which shall be used for manufacturing process. The k-NN algorithm is robust to noise in data, i.e. the sets of known process parameters in the database, and has good self-updatability capabilities. The k-NN algorithm searches for similar process parameters in the database and, based on received results, decides if the new set of process parameters might cause a defect.

It is further preferred that processing the k-NN algorithm comprises application of a genetically optimized similarity function. The similarity function used by the k-NN algorithm may be optimized by the genetic algorithm that determines weight of parameters, i.e. their importance for similarity function. According to that, for application of the genetically optimized similarity function at least some of the process parameters of the set of process parameters are rated with a factor according to a predetermined relevance information. Applying the genetic optimization procedure helps to reduce the amount of process parameters within the sets of process parameters. By applying the genetic optimization procedure important process parameters for defect information can be found while process parameters which are not that important for rejections may be omitted. Therefore, important process parameters will receive bigger weights in the similarity function in order to make the retrieval procedure more efficient. The output from the k-NN algorithm which is the prediction of rejections, in particular a coil break in hot rolling process, for the current process parameters supports plant personnel in decision making on suitability of the current, new set of process parameters.

After manufacturing of the product it is determined whether the manufactured product is rejections or not. In case of manufacturing a steel strip after having rolled the steel strip it is determined whether the steel strip comprises coil breaks or not. It is to be understood, that the manufacturing of the product, in particular the hot rolling process, is only started if the prediction of rejections is positive in that sense that no rejections will be produced with the new set of process parameters.

According to a further preferred embodiment, the new set of process parameters is added to the database together with an information about the occurrence of rejections, in particular coil breaks. It is preferred if the new set of process parameters is added to the database just in case of determined rejections, in particular coil breaks. The information in the database consists of a set of process parameters together with the corresponding information about rejections, i.e. the absence or the presence of a production problem. The database allows determining a safe range of process parameters where the risk of rejections, especially the occurrence of coil breaks, is minimal. Based on the stored sets of process parameters a decision tree, in particular a Classification and Regression Tree (CART), is created, each node of the decision tree predicting a safe range of each of the process parameters. Since each node of the decision tree comprises the information about rejections or not a safe range of each of the process parameters can be derived. The decision tree may be created based on the information of the database in an offline regime. The decision tree may be created after the new set of process parameters has been added to the database. Alternatively the decision tree may be triggered manually.

According to a further preferred embodiment, the set of new process parameters is selected from a total number of process parameters wherein the selection comprises the steps of discarding a first amount of process parameters from the total number of process parameters using the main knowledge resulting in a reduced number of process parameters. A Genetic Algorithm is applied to the reduced number of process parameters to determine a further reduced number of process parameters representing the set of new process parameters. This set comprises the most important process parameters whose parameter values influence rejections. Since only those process parameters are processed which have a significant probability with regards to rejections the processing of the prediction can be enhanced.

Other aspects and advantages of the invention will be apparent from the following description and the appended claims.

Fig. 1 shows a schematic diagram illustrating the method according to the invention,

Fig. 2 shows a schematic diagram illustrating the process of determining those process parameters which have the biggest influence on rejections.

Fig. 1 shows a schematic diagram illustrating the principle of the method for verifying process parameters of a manufacturing process for prediction of rejections. The following embodiment is made with reference to the verifying of process parameters of a hot rolling process of a hot steel strip for prediction of coil breaks. It is to be understood that this is only an example and the method according to the invention may be used for any manufacturing process in which verifying of process parameters may be useful.

The present invention makes use of machine learning methods for both, an online prediction of surface defects and offline determination of a safe range of process parameters that lead to a minimal risk of surface defects in hot rolled steel. The present method makes use of a database CB which is a case base. The database CB provides a plurality of sets of known process parameters wherein each set comprises an information about the occurrence of coil breaks. The set of process parameters may, for exam- pie, consist of continuous casting parameters, hot rolling parameters for a roughing stand, hot rolling parameters for a finishing stand, hot rolling parameters for strip cooling, final strip parameters, pickling parameters, and strip cooling at storehouse parameters.

Continuous casting parameters may consist of the mass fraction of the material processed, such as carbon, manganese, silicon, phosphorus, sulfur and so on, the thickness, width, length, weight and temperature of the slabs. The hot rolling parameters for the roughing stand may consist of the desired final strip thickness and width a strip speed in the first stand, a torque of the first stand, a power of further stands, temperature of water used to cool working rolls, and so on. The hot rolling parameters of the finish- ing stand may consist of the strip speeds of a number of stands, the average rolling strength of the respective stands, the average current of main drive of the stands and the strip temperature after finishing stands. The hot rolling parameters of strip cooling may consist of the number of first water unit enabled, the minimum and maximum numbers of water units enabled and the temperature of water used to cool the strip. The final strip parameters may consist of the strip length, the strip thickness, the strip width, the strip coiling temperature, the strip lateral polythickness, the strip weight and the strip flatness after hot the rolling section. The pickling parameters may consist of the elongation in strengthening machine, the strip speed in pickling line, the average temperatures of the solution in a number of baths, and a number of slices cut from the head of strip. The strip cooling at storehouse parameters may consist of the duration of cooling at storehouse and the average air temperature at the storehouse. It is to be understood that the parameters mentioned above are only examples. The parameters considered may be chosen according to the respective manufacturing process.

Before a steel strip is rolled the process parameters have to be planned for the manufacturing process. In step S 1 a new set of process parameters which are to be used for the hot rolling process is provided. However, at this moment it is not known whether the new set of process parameters will lead to coil break or not. Therefore, in a second step S2 the new set of process parameters is processed to find a similar set of process parameters from the stored sets of process parameters in the database CB such that the process parameters of the new set and the process parameters of the similar set are similar.

Processing the new set of process parameters is done in the so-called online prediction loop. This processing consists of applying the well-known k-Nearest Neighbors (k-NN) algorithm with a genetically optimized similarity function. The k-NN algorithm searches for similar process parameters in the database CB and, based on retrieved results, decides if new process parameters might cause a defect. An advantage of the k- NN algorithm is its robustness to noise in data and the good self-updatability capabilities. The similarity function used by the k-NN algorithm is optimized by a Genetic Algorithm that determines a weight of parameters, i.e. their importance for similarity functions. The genetic optimization procedure is used to distinguish those process parameters which are important for defect formation from others that are not that important. Important process parameters therefore receive bigger weights in the similarity function in order to make a retrieval procedure efficient. This is achieved via the genetic optimization. The output from the k-NN algorithm supports in the decision making of the suitability of the parameters of the new set of process parameters. In particular, the output of the k-NN algorithm gives a prediction of a coil break for the new process parameters.

The procedure of identifying the most relevant process parameters from a total amount of process parameters consists of two main steps and is illustrated in Fig. 2. In step S20 all process parameters which might be used for hot roling process are provided. In the first main step (S21) parameters are discarded by using so-called domain knowledge (which is known as knowledge-driven approach). Furthermore, in the second main step (S23), a selection of the relevant process parameters using the mentioned Genetic Algorithm (known as data-driven approach) is done, resulting in a reduced number of final parameters (S24). More specifically, a list of all available parameters is created (S20). For example, there are totally more than 70 parameters. These parameters may be divided by groups corresponding to three main operations: continuous casting, hot rolling and pickling. It is to be understood, that the number of total parameters may vary from the ex- amples given. Furthermore, the total number of parameters may be divided into a different number of groups.

During the first step (S21) the domain knowledge is used to discard those parameters that most probably do not influence a defect formation. After this step a reduced number of parameters remains among which there are important parameters as well as irrelevant parameters (S22).

During the second step (S23) the Genetic Algorithm will be used. The Genetic Algorithm is a well-known optimization technique, which is known to the person skilled in the art and can be easily found in literature. It is assumed that there is a data set containing records of various process parameters together with corresponding indication of presence or absence of a surface defect. The Genetic Algorithm is used to automatically select a final set of important process parameters (S24).

For a so-called Genetic Algorithm fitness function, classification accuracy of the k-NN algorithm run on this data set with a leave-one-out validation technique may be used. The distance matrix used in the k-NN algorithm is Euclidian. The result produced by the genetic algorithm is a vector of weights where each weight corresponds to a certain process parameter and scales from 0 to 1. A value of "0" means that a corresponding process parameter does not have any importance. A value of "1" indicates the greatest importance. Intermediate values (between 0 and 1) indicate some intermediate importance. Thereby, by sorting the resulting weights by decreasing order and truncating smallest weights, the final set of most important process parameters can be determined.

The final set of process parameters will be used for the step of processing the new set of process parameters and the sets of known process parameters stored in the database. Using the most important process parameters leads to a fast and reliable result in finding the most similar set of known process parameters stored in the database CB.

Referring to Fig. 1 again, in case that the prediction about a coil break (which is taken from the information about a coil break from the known set of process parameters which is the most similar set of process parameters) of the new set of process parame- ters is positive, the hot rolling process is executed (step S3). In step S4 the prediction is verified. That means, the prediction of the k- N algorithm is compared with the actual condition of the steel stip. If the prediction appears to be correct nothing has to be changed in the case base. However, it is to be understood, that a new set of process pa- rameters together with an information about the coil break may be stored in the case base; i.e. the database CB. In case that the output from the k-NN algorithm is wrong, i.e. in fact a coil break or surface defect can be detected although the prediction said that no surface defect will occur, the set of current process parameters together with the information about the surface defect will be added to the database CB together with a correct state of the steel strip. This is done in step S5 in which the database is updated.

The described actions are performed online. The procedure helps plant engineers to prevent coil breaks in real time. However, for personnel it is also important to know the safe range of process parameters in which the risk of a coil break occurrence is minimal. This task is achieved by a so-called offline routine.

The information in the database CB consists of a set of process parameter re-

- cords together with a corresponding diagnose, i.e. absence or presence of coil break, as described above. In step S10 a decision tree, in particular a Classification and Regression Tree (CART), is created based on the database information in an offline regime. The decision tree creation may be triggered manually or when a new case has been added to the database CB. From the result in decision tree a safe range of process parameters can be easily derived to provide a prediction for a single process parameter (step Sl l). The information received from step Sl l may be used to determine single parameters of the new set of process parameters of step S I of the method described. Hence, it might be useful to carry out steps S10 and SI 1 before providing the new set of process parameters in step 1 for the online routine.

The main advantage of the method described is the ability of self learning. The database information is maintained automatically by adding new cases. A further advantage is the ability to explain a prediction by referring to similar cases in the past which are stored as sets of known process parameters in the database. A decision tree may be constructed by an offline routine which can be transformed into a set of rules. An analysis of these rules can assist plant engineers in discovering roots of coil breaks and other steel defects.

Claims

1. A method for verifying process parameters of a manufacturing process for prediction of rejections, in particular a hot rolling process of a steel strip for prediction of coil breaks, comprising the steps of

- providing a database in which a plurality of sets of known process parameters is stored wherein each set comprises an information about the occurrence of rejections;

- providing a new set of process parameters which shall be used for manufacturing;

- processing the new set of process parameters to find a similar set of process pa- rameters from the stored sets of process parameters in the database such that the process parameters of the new set and the process parameters of the similar set meet one or more predetermined requirements;

- evaluating the information about the occurrence of rejections from the similar set of parameters;

- processing the information to make a prediction of rejection for the new set of process parameters.

2. The method according to claim 1, wherein the step of processing the new set of parameters comprises application of k-Nearest Neighbors (k-NN) algorithm wherein the predetermined requirement is a metric.

3. The method according to claim 2, wherein processing the k-NN algorithm comprises application of genetically optimized similarity function.

4. The method according to claim 3, wherein for application of the genetically optimized similarity function at least some of the process parameters of the set of process parameters are weighted with a factor according to a predetermined relevance in- formation.

5. The method according to one of the preceding claims, wherein after manufacturing of the product, it is determined whether the product is rejections or not.

6. The method according to one of the preceding claims, wherein the new set of process parameters is added to the database together with an information about the oc- currence of rejections.

7. The method according to claim 6, wherein the new set of process parameters is added to the database in case of determined rejections.

8. The method according to one of the preceding claims, wherein based on the stored sets of process data a decision tree, in particular a Classification and Regression Tree (CART), is created, each node of the decision tree predicting a safe range of each of the process parameters.

9. The method according to claim 8, wherein the decision tree is created after the new set of process parameters has been added to the database.

10. The method according to one of the preceding claims, wherein the set of new process parameters is selected from a total number of process parameters wherein the selection comprises the steps of

- discarding a first amount of process parameters from the total number of process parameters using domain knowledge resulting in a reduced number of process parameters; and

- applying a Genetic Algorithm to the reduced number of process parameters to determined a further reduced number of process parameters representing the set of new process parameters.