A Knowledge Based System Approach to Predict

Defect prediction, using the knowledge based system approach, is being developed for the case of longitudinal surface cracks. Knowledge about longitudinal cracking is in the form of databases, mathematical models and qualitative information. Much of this comes from the technical literature on related topics, but is supplemented by plant experience, as well. Knowledge based systems provide the facility for handling such types of knowledge, with control particulars left up to the system builder The current implementation is forward chaining and uses the OPS5 inference engine. The system involves the use of all three types of knowledge, but more mathematical models are being sought. \:\,'i ' ' Introduction Hot charging has made defect awareness and detection in the continuous casting of steel a more important problem than in the past. If slabs coming off the caster are inspected, thermal units will be lost, and if slabs are passed, unchecked, the risk is higher that finished product quality will be unacceptable to the customer and higher costs will be incurred. It would seem clear that the best approach to the situation is an accurate predictive tool that provides for the confidence that only high quality slabs are passed to the hot rolling mill. The current practice consists of 'setting the controls', according to established practice for the particular grade (or application) being cast, i.e. a table lookup of recommended operating regions for different grades. This has a number of drawbacks. First of all, the acceptance of established procedure may be too conservative under certain conditions. For example, there may be heats which can be cast faster, or sprayed differently than the recommended amounts, in order to hold more heat for the rolling operation, or to increase productivity. Secondly, there are a number of 'response' variables which can significantly affect various defect formations, which are not immediately selectable by the operator, such as variations in casting speed or mold level. Thirdly, table lookups only account for one variable at a time. In addition, tables provide no guidelines or predictions for dispositioning of slabs. This paper will present some of the basic types of information relating to longitudinal cracks that may be used in a knowledge based system. This modelling approach will be defined and contrasted with the more traditional algorithmic approaches. In addition, the framework developed for predicting longitudinal surface cracking will be outlined. Information on Longitudinal Surface Cracking The difficulty is that, while certain problems in casting, such as shell thickness determination (mathematically), are relatively straightforward and dependable, reliable defect prediction methods are not readily available. This is a consequence of the uncertainty with which defects form. In the gase of longitudinal surface cracking, it is obvious from the literature* that the problem is related to irregularity in the mold cooling, both transverse and longitudinal. Irregular heat fluxes will give rise to irregularities in the solid shell thickness and, hence, in the solid shell strength, as well. In this situation, it is easy to envision a thicker portion of the shell pulling away from the wall, while a neighboring thinner area is pushed back against the wall by ferrostatic pressure, leading to a crack, or worse, a breakout. The key issues, then, are thermal and mechanical ones. How large is the difference in neighboring shell thicknesses? How much transverse stress is the shell experiencing, and how does a particular steel respond to these stresses? Methods are described in the literature which attempt to make predictions about cracking, based on finite element stress/strain analysis'. These have primarily been with reference to internal cracks, i.e. further down the caster, but have even been attempted for the case of longitudinal cracking. Such approaches cannot account for the large amount of uncertainty involved, or for various operating features, such as mold level fluctuations. In addition, important qualitative considerations cannot be integrated. The primary sources of knowledge for defect formation, in general, are databases, mathematical models and qualitative considerations. Database information is of two types, relating to properties or phenomenology. Property databases simply consist of tables or plots of various mold, strand and powder properties showing, for example, dependence on temperature. Phenomenological databases cover all of the information in the literature (and elsewhere) which speaks to either the defect, itself, or some important aspect of the defect. For example, peritectic steels (carbon in the range 0.08 to 0.16 %) are expected to exhibit longitudinal surface cracking greater than other grades," as shown in Figure 1. This is related to the shrinking attendent to the peritectic transformation-, as well as microsegregation effects. Similar results can be found for a host of variables : • Chemistry • Operating variables (speed, pouring temperature, spray practice, etc.) • Mold powder characteristics (viscosity, softening temperature, etc.) • strand dimensions • 'Response' variables (variations in speed, mold level, etc.) Other chemistry effects include the Mn/S ratio, which is directly related to ductility considerations. In this case, 'database' refers to the knowledge which has been compiled on the high temperature strength and ductility of steels, as a function of composition*. Casting speed (v), and mold powder viscosity (TJ) have been shown to have the same kind of effects* (see Figure 2), since both will be involved in the even flow of molten flux down the side of the strand. It has even been suggested that a parameter of interest would be TI*V. strand dimension effects have mainly been in the form of width to thickness ratios for slabs, higher ratios leading to more cracking. 'Response variable' effects have been shown to be strong, particularly with mold level variations-. These variables are obviously important, but what about quantities such as the steepest temperature gradient at the midface (at various positions in the mold and just below), the temperature rebound below the mold, etc.? These quantities are of considerable importance, and should be integrated as well. This will lead to greater specificity of the independent variable space. For example, the (average) casting speed will greatly affect the temperature distribution, but will have separate effects, as well. These kinds of information are afforded by the vast work done on modelling in continuous casting. Heat flow modelling has been done on the strand", in the mold, and in the flux layer, as well. This is usually done in the following form (in the strand) : dx V dx ) dy v By ' V p dt J In addition, modelling work has been done in the area of stress and strain determination--. This is a difficult undertaking in the mold, where mesh sizes must be extremely fine for finite element strain analysis (on solid shells only 1 or 2 cm thick), but has been done with reasonable results-. The final source of knowledge about cracking is in the form of qualitative considerations. These are often in the form of heuristics, or rules of thumb, which don't necessarily have any basis in theory, but are found to be true, nonetheless. For example, tube changes have been suggested to have a negative effect on cracking. Also, it has been suggested that longitudinal cracks often happen in bunches. Knowledge Based Systems