25-09-2014, 02:36 PM
ABSTRACT OF THE PROJECT DONE During the pickling process, the under pickling and over pickling defects are usually the direct results of lack of control over process variables including acid concentrations, solution temperature and contact time deriving from pickling line speed. Up to date most pickling lines are run by technical operators, which view the surface of the steel strip time to time and adjust process parameters, by mainly decreasing the line speed; this operation does not always achieves the desired results. The proposed model, composed by two different units, is capable to predict the under pickling occurrence in a pickling process, from coil and process parameters, by also suggesting a defined value for v or its admissible range. By means of decision tree unit, a binary tree has been built allowing a classification of valid process window, from which it is possible to suggest the speed range that should be set in the considered process conditions in order to avoid under pickling defects. In addition, a neuro-fuzzy model called RecBFN has been used in order to
6 MONTHS INDUSTRIAL TRAINING PROJECT REPORT
B.Tech.
Name of Company Bhushan power and steel ltd.
For the period from 1 june 2013 to 30 nov 2013
Department Production
Training Manager Name Mr. B.K Dixit
& Designation AGM, Production
PROJECT REPORT TO REMOVE SURFACE
DEFECT ON STEEL STRIP
SURFACE DURING PICKLING
TRAINING MANAGER SUBMITTED BY Name: Mr. B.K Dixit Name: NIKHIL KOUNDAL
Designation: AGM, Production Branch: Mechanical
Department: Production Univ.Roll No. 100761133068
8. To reduce pickling defects on steel strip surface.
An extremely important part of the finishing line is the pickling process, in which oxides formed during the cold rolling stage are removed from the surface of the steel sheets. The efficiency of the pickling process is mainly dependent on the nature of the oxide present at the surface of the steel, but, also, on process parameters. When acid concentration, solution temperatures and line speed are not properly balanced, then sheet defects like under pickling or over pickling may happen and their occurrence does have a very serious effect on strength and surface appearance of the finished product. Furthermore, product damage from handling or improper equipment adjustment can render the steel unsuitable for further processing. This is the reason why it is important that process significant parameters are controlled and maintained as accurately as possible in order to avoid these undesired phenomena.
A new control algorithm, composed by two different modules, i.e. decision tree and rectangular Basis Function Network, has been implemented to aim of predicting pickling defects and suggesting the optimal speed or the admissible speed range of the steel strip in the process line. In this way the most suitable line speed value can be set in an automatic way or by the technical personnel.
1. Introduction: To attain a complete elimination of the oxide from steel strips surface a mandatory production step. One of the main problems during this process is the difficult evaluation of the pickling state. If on the one hand it is evident that high strip quality requires a complete elimination of the oxide on the strip surface, thus under-pickling cannot be tolerated, on the other hand a too long permanence of the strip in the pickling bath can lead to an erosion of the steel strip surface itself (over-pickling). Both these problems lead to more expenses or losses: in facts under-pickled strips need to be re-processed, while the quality of over-pickled ones must be downgraded.
Removal of oxide is carried out within mechanical methods including polishing method and shot blast method, particularly used for removing dense and adherent oxides scales from stainless steel and chemical methods like electrochemical pickling or chemical pickling methods.
This work is focused on steel chemical pickling bath method (namely pickling), which is the part of the finishing process in the production of steel products in which oxide and scale are chemically removed from the surface of strip steel in order to obtain a clean surface. During the pickling process, steel strip is pulled through three different tanks that contain baths of aqueous acid solution with different concentration: 5, 10 and 15 wt% of hydrochloric acid (HCl) and subsequently steel strip passes through water rinsing tank, with a cascade flow of pickling solution counter current to the direction of the strip movement. Fig.1 shows the pickling process and the involved variables.
In order to prevent and reduce to minimum the iron dissolution from the steel strip, allowing a preferential attack on the scale, in all the baths organic compounds, namely corrosion inhibitors, are added. The goal of the inhibitors addition is to protect the strip from the attack by the hydrochloric acid solution, by means of protective film that sticks on the surface of the steel strip but not on the oxide.
The objective is to identify the conditions when the under pickling phenomena take place, in order to firstly detect and predict their formation. In this paper a model composed by two different modules, a decision tree and a Rectangular Basis Function Network (recBFN), is described. Its aim is to suggest the optimal speed or admissible speed range for the process line, given steel strip parameters (e.g. strip width and thickness, steel grade) and acid bath conditions, with the objective to reduce at minimum the occurrence of under pickling defects. The implemented overall model achieved noticeable results.
2. Problem definition: Over pickling results from the line delays which permit sections of the steel to remain in the acid too long. The presence of an inhibitor reduces iron loss, but when an inhibitor is not used, iron loss during a short delay period appreciably reduces thickness of the steel. This kind of defects must be prevented by adding inhibitor to the acid solution and also by increasing the line speed, if possible. On the other hand, under pickling occurs when the acid solution does not completely dissolve the oxide, due to a low concentration of the acid solution and/or an excessive line speed, which shorten the time of stay of each portion of the strip in the acid solution. The present work is particularly focused on the avoidance of the under pickling phenomenon.
The goodness of pickling process is affected by many variables such as the steel composition, acid solution, scale structure, concentration of corrosion inhibitor, baths temperature and line speed. Usually, under pickling is the direct results of a lack of control over the above-cited process variables, thus there is need to control of pickling process that use neural network model or a hybrid grey box modeling , in order to model the steel pickling process. The use of neural networks and grey box models is appropriate.
The selected variables and their nomenclature are listed below:
Physical coil parameters coil weight, strip thickness o, strip width . Pickling process parameters T1 – temperature of tank 1 , T2 – temperature of tank 2 , T3 – temperature of tank 3 , Trines – temperature of rinse ,under v speed of pickling line , HCl 1 – acid concentration of tank 1 ,HCl 2 – acid concentration of tank 2 , HCl 3 – acid concentration of tank 3 ,Fe2 concentration of tank 1 , Fe2 concentration of tank 2 , Fe2 concentration of tank 3 .Under – under pickling defects
In order to minimize the under pickling phenomena on the surface of the steel strip, the fastest way to govern the effectiveness of the considered process is to set a right value of the speed of steel strip, which affects the permanence time of the steel strip in the pickling bath and the effect of both the hydrochloric acid solution and the corrosion inhibitor on the steel strip surface.
3. Model implementation The model proposed in this work is composed by the union of two different modules, a decision tree-based model and a RecBFN-based model. The variables under P and v are the targets (i.e. they must be predicted), while some of the others are fed as input in the implementation of the developed model.
The first model aims at defining a valid process window, i.e. a range of process parameters ensuring a good behavior of the pickling line. A decision tree model is used, by classifying the input parameters with respect to the associated value of the line speed, by considering only the data contained in the available dataset, that are related to products not affected by pickling defects. The final aim of such system is to suggest to the operator a valid range of the line speed that allows to avoid the under pickling and over pickling defects on the basis of some coil and process parameters.
The second module is a particular kind of neural network and its presence is due to the fact that, when the control is applied on the real line and working at full stretch, there is a decreasing number of under pickling defects on the strip steel surface, because the capability of the control system predict and adjust the speed line will improve through time. Considering this declining frequency of the under pickling defects, a classical learning method which is usually applied to balanced datasets (i.e. datasets where the number of the data belonging to the different classes are balanced) could not be able to detect the anomalous conditions which give raise to under pickling. Therefore a model based on soft computing techniques, which is specialized on imbalanced datasets and is composed by a neuron-fuzzy model called RecBFN (rectangular Basis Function Network), has been implemented and subsequently integrated in the decision tree module.
The following two subsections are devoted to a more detailed description of the two above-mentioned modules, while the third subsection describes how the two modules have been integrated and how the global model works in practice.
a. Decision Tree module A decision tree is a type of tree-diagram that is used for determining the optimum course of action, in situations that are characterized by several possible alternatives with uncertain outcomes. The resulting tree displays the structure of a particular decision and the interrelationships and interplay between different alternatives, i.e. decisions, and possible outcomes. They are widely adopted in the field of operations research and decision analysis, in order to point out the strategy or the procedure that are most suitable to reach a predefined objective. In detail, at each node of the tree, the algorithm chooses one attribute of the data that most effectively splits the set of samples into subsets enriched in one class or the other. Its criterion is the normalized information gain that results from choosing an attribute for splitting the data. The attribute with the highest normalized information gain is chosen to make the decision.
In the present case, a classification algorithm based on a decision tree has been applied to the aim of identifying a valid process window i.e. an optimal range for the line speed value which avoids the occurrence of both under pickling and over pickling defects. The following variables have been selected in the implementation phase of this model, which are related to both product and process: its, W, T1, T2, T3, Trines, HCl_1, Fe2_1, HCl_2, Fe2_2, HCl_3, and Fe2_3.
By considering that the line speed is the only available degree of freedom for technical operators working on the pickling line, the variable v parameter is assumed as target variable in the classification phase. In order to classify the input variables with respect to the target, all the line speed values that range from 104 m /s to more than 385 m /s have been subdivided into the following three different classes of same size:
Class A: 104 < v H 244 class B: 245 < v H 384 class C: v I 385
Moreover a class U of defective coils has been added, in order to take into account those coils where the under pickling phenomenon is previewed to occur independently on the value set for the line speed (for instance because of the low concentrations of hydrochloric acid in the tanks and/or unsuitable temperature values). A binary tree has been built on the defined training set, by means of a decision tree algorithm; where each node represents an input variable and the leaf node represent one of the three previously defined class i.e. the predicted class of the target variable v. Following each branch of the constructed tree, the model is capable to identify a valid process window, showing the line speed range that should be set in the considered process condition in order to avoid under pickling and over pickling defects.
b. RecBFN module In order to correctly detect and classify defective coils also when a high unbalance level will be reached among defective and non-defective coils due to the implementation of the speed line control at full stretch, a neuron fuzzy model called Rectangular Basis Functions Network has been implemented. This kind of neural network, which is depicted in Fig. 2, consists of hidden units, each covering a rectangular area in the input space, using a trapezoidal activation function.
In detail, in the hidden layer of the network, each neuron represents a fuzzy point and it is associated with the two classes, i.e. under pickling defects and no defects. Each neuron takes as input a vector containing the selected parameters computing the degree of membership of the input pattern to the fuzzy point. Lastly the output layer assigns a class to the input pattern processing the input information. The underlying training algorithm allows easy and fast construction of these types of networks and no parameters need to be adjusted; only normalization of the input-data is necessary.
This kind of neural network has been chosen because it works well on unbalanced datasets. Among the pickling process variables, the parameters defined in the testing phase, which are used as inputs of the model, are: T3, HCl_1, Fe2_1, HCl_2, Fe2_2, HCl_3, and Fe2_3. In order to get a prediction of maximum line speed in the case of under pickling defect, v has been used as an additional free input: the whole range of possible line speeds within a predefined range is scanned starting from the lowest bound as far as the inputs configuration is recognized as corresponding to good working conditions. As an under pickling condition is met, a maximum line speed that avoid under pickling is suggested. In this way, by suggesting the first helpful velocity that would avoid under pickling defects, the occurrence of the opposite phenomena (i.e. over pickling), which is due to an excessive time spent in the pickling tank, is prevented.
Integration of the two modules the two different units composed by RecBFN model and decision tree model have been integrated as depicted in Fig. 3.
By using the input variables listed in Par. 3.2, the system firstly runs the RecBFN model in order to evaluate the input parameter against all possible line speed value in a predefined range; if there is a line speed value that, together with other parameters, could cause the under pickling phenomenon, a maximum admissible line speed is suggested. Otherwise the decision tree unit is run, all the input parameters are used to classify the coil with respect to the variable v and a suitable range of line speed values is suggested, in order to prevent the under pickling defects.
4. Results During the implementation phase the dataset has been divided into a training set, which contains the 75% of the total records, and a validation set, that is composed by the remaining 25% of the records. A particular attention has been paid to keep the same proportion between the two existing classes, i.e. defective and non defective coils.
Each of the two modules has been separately designed and trained, as discussed in par. 3.1 and par. 3.2. It must be underlined that the decision tree must be trained only with data referring to standard conditions, while the training of the neural network-based model needs data associated to the situations of no-pickling and under pickling. Finally a global validation of the model has been executed on the remaining 25% of the dataset, which has not been used in the training phase of both modules that compose the model.
The test of such tree that is performed on the validation dataset provided a rate of correct classifications of 75.7%, which means that the decision tree successfully classifies all the inputs with respect to the variable v. For the RecRBF-based module, very low values of the misclassification rate have been achieved on both training and validation sets, namely 0.23% and 0.30% respectively.
In order to evaluate the performance of the global model the following classical and widely adopted indexes in the classification tasks have been exploited.
Considering the confusion matrix, in Table 1, where the columns represent the correctly classified instances, while the rows represent the real instances of the particular class, the performance indexes are defined as follows:
Precision – TP / TP + FP This is a measure of the proportion of the classified matches that are true matches;
Recall – TP / TP + FN This is a measure of the proportion of actual matches that have been correctly classified;
F-Measure – 2*Precision*Recall / Precision + Recall this is a measure of the compromise between precision and recall;
Table 2 shows the obtained results for the four classes, i.e. classes A, B and C that group the decision tree classification, and class U that refers to the RecBFN output for the defective coils.
Table 2. Global model performance
The performance of the global model is noticeably good, because all the values of the adopted indexes are close to one. Thus by means of this model it is actually possible to predict, with a high degree of precision of about 96%, the case in which probably at the end of the pickling process some under pickling defects are formed on the coil surface, by allowing the operator to set a optimal value for v according to the suggestion provided by the model. This is possible thanks to the successful classification of coil and process parameters. The results are computed by taking into account that, when under pickling actually occurs, the fact that predicted line speed is lower than the real one demonstrates the good behavior of the model, as a decrease of the speed line value can lead to an optimal working by avoiding the under pickling phenomenon.
In addition, the very low rate of false alarms (i.e. coils without pickling that are classified as defective), which is equal to 0.034%, shows that the model has a very high probability to predict the under pickled defects when only a few working presents this undesirable phenomena, by also suggesting a maximum value of the line speed, that allows to avoid the occurrence of under pickling defects without conditioning the over pickling rate, because the suggested velocity value is slightly lower than the real value. A training phase should be periodically performed on new data, in order to guarantee the global reliability of the overall system over the time.
5. Conclusions During the pickling process, the under pickling and over pickling defects are usually the direct results of lack of control over process variables including acid concentrations, solution temperature and contact time deriving from pickling line speed. Up to date most pickling lines are run by technical operators, which view the surface of the steel strip time to time and adjust process parameters, by mainly decreasing the line speed; this operation does not always achieves the desired results.
The proposed model, composed by two different units, is capable to predict the under pickling occurrence in a pickling process, from coil and process parameters, by also suggesting a defined value for v or its admissible range. By means of decision tree unit, a binary tree has been built allowing a classification of valid process window, from which it is possible to suggest the speed range that should be set in the considered process conditions in order to avoid under pickling defects.
In addition, a neuro-fuzzy model called RecBFN has been used in order to predict the under pickling defects starting from unbalanced dataset, a situation that is likely to occur when the implemented control works at full stretch and a non optimal working decreases due to the self-learning capability of decision tree. By means of the union of the two different units and starting from coil and process parameters, the global model is able to predict the maximum line speed or the speed range that allows to avoid the under pickling defects.
The only degree of freedom at disposal of the operator during the pickling process is the line speed, which determines the time spent in the tanks by the steel strip and, consequently, how long the acid solution attacks the surface of the steel, with the possibility of formation of under pickling or over pickling defects. The operator will be helped by the suggestion of the maximum line speed or the admissible speed range, that will lead to good working, without the presence of over and under pickling defects, as it is possible to react, by adjusting the speed line so as to carry out the pickling process under optimal conditions, by avoiding the occurrence of under pickling phenomena and obtaining a quality improvement of the final product with a cost reduction.
This model can also be transferred on different line typologies, by simply executing a new training phase of both modules.
As future work, the adaptation of the same global model structure will be done in order to detect over pickling defects, which more rarely happens with respect to the under pickling defects, not allowing at the moment the application of the techniques proposed in this work. The model needs improvements in order to deal with even more imbalanced classes than those present in the under pickling problem. In such way will be possible the optimize the pickling process, allowing at the same time the reduction of both over and under pickling defect.