VISION RESEARCH
The MicroCert Surface Volume Imaging System For The Inspection
Of Mechanical Properties In Rolled Ferritic Steel Strip

 

EXECUTIVE   SUMMARY

A smart sensory system was developed that takes on-line electromagnetic sensor data and transforms the data stream via an intelligent perceptual network into inspections of mechanical properties.  The goal of this project is to demonstrate that a measurement system can be created that is sensitive to process relevant variables important to the production of high quality steel strip; and, that in the course of sensing and discriminating the effects of causal process parameters on microstructural development,  the system will create a “minimal model” relating sensed electromagnetic patterns in the strip product to extremely accurate final mechanical properties inspections of yield strength, tensile strength and hardness. A “minimal model” constructed by the system would be based on four independent variables, product type, grade, gauge, and sensed patterns of macrostructural organization.

Measurements And System Development - Project A
The initial on-line system will take approximately 12” wide measurements for microstructural mapping down the center line of pickled strip product.  The yield strength, tensile strength, and hardness (the three dependent variables) will then be analyzed with respect to the “minimal four” independent variables enumerated in the previous paragraph. The results of this analysis would be used to optimize the spatial resolution and operating frequencies of the sensory system for maximum sensitivity and selectivity for all products produced on that line. An initial database of extracted macrostructural sensory features will also be constructed. This portion of Project A would take one year.

Information and measurements from phase one will be used in the second phase of the project to seed a heuristic algorithm used in an intelligent perceptual network that would construct a provisional properties inspection model to fine tune the recognition and  extraction of relevant sensory features.  This system will become a surface volume imaging system, capable of conducting in real time, continuously improving, empirically derived inspections of mechanical properties for individual types, grades, and gauges of strip product.  This portion of Project A would take two years.

Integration With Other Technology - Project B
Information from the surface volume imaging smart sensor system will be integrated with other empirically derived data, e.g. data from ultrasonic sensors, at this point.  This information would be used to re-seed the heuristic algorithm of the intelligent perceptual network used to make future smart system inspections of mechanical properties more accurate under a greater variety of production conditions.    This project would take one year.

THE SURFACE VOLUME IMAGING SYSTEM

Introduction
Vision Research developed the MicroCert Surface Volume Imaging System for enhancement of process control over ferritic microstructure during rapid metal forming processes. In addition to being very sensitive to melt chemistry, current prototypes of the MicroCert System are capable of sensing and mapping microstructural features such as grain size, dislocation density, crystallographic texture, and phase distribution with a spatial resolution of centimeters to a depth of several hundred microns into the strip surface. It is further proposed that an integrated intelligent perceptual network would perform mechanical properties inspections on the hot rolled strip based on sensor information and mill data.

How It Works
The sensory portion of the system consists of a superconducting, high Q, radio frequency LC resonator inductively coupled to ferritic strip moving past it. The quality factor (Q) of this resonator is determined to obtain a measure of energy dissipated in the steel alloys coupled to it.  Dissipative energy in the sample results from two sources, hysteresis and eddy currents.  Eddy current loss is relatively microstructure insensitive and is mainly related to chemical composition and temperature while hysteresis loss is determined predominately by microstructural factors.  As an example, if microstructural variables other than grain size, and the factors determining eddy current loss are held constant, grain size will become the dominant factor in differential energy loss with grain size being inversely related to hysteresis loss.  Under these conditions grain size is inversely related to total energy dissipation and thus the measured Q of the resonant circuit [1].

The prototypes of the MicroCert Surface Volume Imaging System use two or more different r.f. frequencies, and hence imaging at different depths in the surface microstructure, to produce a volume image of the strip surface.

Objective
The main objective of this development project is to demonstrate that a measurement system can be created that is sensitive to a wide range, and potentially large number, of relevant process variables involved in the commercial production of hot rolled ferritic steel strip.  Further, in sensing and discriminating the effects of process relevant parameters on microstructural development, the system will, of necessity, create a “minimal model” of the production process that would eventually afford extremely accurate mechanical properties inspections of the coiled strip (e.g. 0.9 ksi yield strength, 0.6 ksi tensile strength, and 2.4 units Vickers hardness resolution, both through-coil and coil to coil).

The goal will be accomplished by an intelligent perceptual network that renders the sensory system capable of discriminating between the effects of the complex non-linear interactions of the multitudinous process variables that are nondestructively imaged in a macrostructural view of the ferritic microstructure as it exists in the coil before shipping to the customer.  This would obviously constitute a quantum leap in process control technology and total quality management.
 

Macrostructure Defined
All samples spatially mapped with various MicroCert  prototype systems to date have shown macrostructural features of microstructural organization on the order of centimeters in size both across the strip surface and along the rolling direction [2].  In some cases there are well defined bands of microstructural organization along the rolling direction for hundreds of meters [3].  The uniformity in microstructural composition that comprises these macrostructural features can be very complex (combinations of different grain sizes, crystallographic textures, dislocation densities, and phase distributions), and they vary through the thickness of the strip as well as across the surface.  This feature of microstructural organization is defined as surface volume macrostructure.

The Importance Of Macrostructural Mapping
Models of chemical reaction-diffusions in membranes, filament and cellular networks, crystalline solutions and in the volumes of relatively thin 3D solution planes [4,5] have repeatedly shown that very small changes in relevant initial conditions, interacting with equally small changes in critical process variables, iterated over a developmental sequence, can produce significantly different macrostructural patterns within a thin planar 3D volume. These self-organizing, spatio-temporal patterns are in some cases [5] similar to the macrostructural features observed in the microstructural mapping of our ferritic steel strip samples.  The models and experimental systems discussed in the above references have such generality that they are usually presented in journals of broad scientific interest like Nature or Science.  This general property of reaction-diffusion models may be described as follows:

The reaction-diffusion model produces patterns that depend uniquely on the initial conditions, the geometry and the scale.  An important aspect of the mechanism is that for a given geometry and scale, the patterns generated for a variety of random initial conditions are qualitatively similar [5].
In terms of steel strip what this implies is that since the width of the strip is very large compared to the thickness of the strip, geometry is a constant.  The thickness of the strip, the gauge, or in the terminology of a chemical reaction-diffusion model, the scale, is a crucial process variable interacting with initial conditions over the developmental process of hot rolling.  Crucial process variables that interact with initial conditions and each other can be identified and discriminated from those with no significant effect by noting how much they modulate the formation of “qualitatively similar” macrostructural patterns. The size and shape of fuzzy, qualitatively similar categories constructed by the intelligent perceptual network is determined by correlation with mechanical properties measurements derived from destructive testing.

Is All This Really Relevant
Hot rolled strip product most certainly fits in the category of a chemical reaction-diffusion in a suitable planar volume, the steel strip; and the iterative nature of the hot rolling reduction process (passing the hot strip from reduction stand to reduction stand) is an excellent example of a developmental sequence with respect to the evolving microstructure and final macrostructure.  The importance of this macroscopic view of steel strip is clearly stated in the J. Majta et. al. paper, Prediction of Mechanical Properties of Steel Strips After Hot Rolling [6]:

In the literature, data for the prediction of mechanical properties of the final product have normally been presented as a function of the solid solution and the grain size components.  However, very often the mechanical response of a material is a consequence of its more complex internal microstructure.  The structure may be described using progressively increasing dimensional scales, beginning at the atomic level and proceeding to a macroscopic, geometric description representable by continuum formulation.  In the case of metals, these two extremes span intermediate features of material organization such as the lattice structure, dislocations, subgrains, and grains [7].  The design of hot forming processes for microalloyed steels requires an understanding of how these features are affected by the thermomechanical history.
The MicroCert system attempts to understand the thermomechanical evolution of macroscopic features by the direct nondestructive imaging of ferritic macrostructural organization at different depths through the cross section of the strip, thereby enabling accurate mechanical properties inspections of over a wide range of production possibilities.

Surely You Must Be Joking
Although at the present time there is only a relatively small amount of empirical evidence obtained from production steel samples of various product types, grades, and gauges, as well as systematic measurements from specifically prepared laboratory samples with known microstructural and/or mechanical properties; it is believed that macroscopic imaging  technology is both sensitive and selective enough to identify the complex non-linear contributions of the most important process relevant variables, enabling final mechanical property inspections with the following resolution: yield strength- 0.9 ksi, tensile strength- 0.6 ksi, Vickers hardness- 2.4 units. This would be accomplished in the following manner:

Vision Research and the Inspection of Mechanical Properties
In its simplest form an ultrasonic mechanical properties sensor usually estimates mean grain size based on propagation and/or attenuation of an ultrasonic pulse through the total thickness of the steel strip.  This mean grain size measurement is then plugged into a model to predict mechanical properties for a particular product type. To be at all useful in a commercial environment the “minimal model” that predicts mechanical properties from raw sensor data should include some estimate of dislocation density in addition to the mean grain size estimate. It is known that a skin pass can differentially affect the mechanical properties of various product types, e.g. ULC vs. HSLA, and may often have a differential effect on various grades of the same product type.  While the sensor and “minimal model” might be very good a making inspections along a single microstructural dimension (e.g. mean grain size averaged through the cross section) of the same product type, e.g. ULC steels,  how about different gauges of the same product type?  Might the sensor and “minimal model” that can resolve yield strength to about 0.9 ksi in same gauge product, obtain an accuracy of only 4.5 ksi when it “looks at” different gauges of the same product? It is easy to imagine how this situation might occur, because it is known that microstructural (and macrostructural) development during the hot rolling process does not scale in a linear manner with changes in product thickness.

The point is that neither a typical ultrasonic sensor nor the Vision Research MicroCert prototype sensors directly measure the mechanical properties of steel strip.  There is a “minimal model”, however primitive, that takes other information (product type, grade, and possibly even some process relevant knowledge) and then makes some prediction about mechanical properties based on sensor measurements transformed according to the assumptions of the model. The steel industry obviously does not need even a “minimal model” to turn electrical signals from a thermocouple into temperature readings, because the transfer function is so very simple. However, one is never developing just a sensor when the subject is as complex as the non-destructive measurement of mechanical properties in hot rolled strip. One is, in reality, developing a sensor that measures some correlate or correlates of the complex microstructural and macrostructural organization in the strip, and a model of how that particular measurement or group of measurements is related to mechanical properties, as determined by good old- fashioned destructive testing of the product.  In any visual system down to that found in the simplest slime worm, the eye has to be attached to brain, a model maker that “makes sense” of the patterns of electromagnetic radiation impinging on the detector within the context of that organism’s industry and occupational environment.

In final analysis the big question is not the resolution of the sensor along a single, well controlled continuum; but how general, robust, and inclusive the predictive power of the model is that interprets the raw sensory data in a complex, non-linear, real world environment. An ultrasound-based sensor is essentially an ear that listens to the echoes of ultrasonic pressure waves created in steel strip.  The Vision Research MicroCert System is more like an eye that images radio frequency electromagnetic radiation absorption patterns in inductively coupled ferritic steel strip.  To get very personal for just a moment, you, as the reader of this document presumably have both sight and hearing to help you in your work, whatever that might be.  Would you, for the sake of economy, be willing to give up your sense of sight?  Is sight merely a redundant luxury to a person who already has hearing?  And how about those brains that are attached to the eyes and ears.  What makes one person a better listener or observer than another; better eyes and ears?  Probably not.  Some people have learned to pay attention to subtle cues in the vast deluge of sensory information impinging on their receptors, and this selective attention gives them a substantial advantage in “making sense” of the message.  Show me a computer of any size and speed that can do speech recognition in a noisy environment equal to the ability of a three-year-old human child.  Do computers just need better microphones and A/D converters in order to do minimally competent speech recognition?  Probably not.

Let’s Get Technical
An ultrasonic sensor and the Vision Research MicroCert sensor can complement each other in the following ways: let’s assume that both the ultrasound-based sensor and MicroCert sensor have roughly the same level of resolution for microstructural properties, e.g. grain size, when the samples employed in testing vary along only one dimension, as opposed to the situation (like a real steel mill) where samples in a test run might vary simultaneously in grain size, percent skin pass reduction, gauge, grade, and product type.  The MicroCert sensor has the ability to measure microstructural properties at different depths through the thickness of the strip, and has been shown to be extremely sensitive to dislocation density at the strip surface.  The ultrasound sensor has the ability to present an integrated view of microstructural properties through the entire cross section of the strip. By combining these two different views of strip microstructure it should be possible to construct a view of strip macrostructure that differentiates between mean grain size refinement and dislocation density due to amount of skin pass reduction.  This macrostructural view should greatly enhance the ability of the “minimal model” to make more accurate mechanical properties determinations.

How These Measurements Will Be Accomplished
The output of the system will be constructed by an asexually reproductive, process driven, genetic algorithm which encodes melt chemistry (the initial conditions, or seed for the process) and line set-up (process parameters which will be iterated over the developmental sequence of the hot rolling process) as input elements of strings for a classifier system. The output strings of the classifier system control the construction of one or more pseudo-color contour plots to display a representation of the macro/microstructural features detected for the entire developmental sequence of the metal forming process.  These color contour plots are for monitoring purposes only.  Unlike neural network AI algorithms containing so called “hidden layers”, the evolution of the entire MicroCert intelligent perceptual network learning process can be closely watched by human observers. The output of the genetic algorithm classifier system also sends control signals back to an extracted feature database for context relevant tuning or “training” of the perceptual mechanism.  This recursive loop can be conceptualized as a process of selective attention executed by the intelligent network. Its effect on system behavior is to greatly enhance the signal-to-noise ratio and the extraction speed for any feature of interest, no matter how complex, which can be discriminated and successfully categorized by the network. (refer back to the last section of Vision Research and the Inspection of Mechanical Properties for clarification)

Once the system has been calibrated using traditional destructive measures of mechanical properties, it can be tuned to discriminate along mechanical properties dimensions, such as hardness, tensile strength and yield strength.  The adaptive nature of the genetic algorithm permits the continually evolving network to discover new and potentially more optimal ways of “looking at” and classifying the complex r.f. electromagnetic fields sensed, and the many process parameters employed in the hot rolling of ferritic steels.

The Theoretical Importance Of Surface Volume Imaging
In many recent computational models of the hot rolling process, microstructural development (e.g. grain refinement) is differentiated  with respect to surface vs. center of the cross section through the thickness of the strip [8 -13]. The strain, strain rate, and temperature are not uniformly distributed through the thickness of the strip during hot rolling. This produces grain size distributions after hot rolling that are non-uniform, with more grain refinement found at the surface, and coarser grains found in towards the center of the strip. The ability to image differential macrostructural development at various depths in the strip should provide important process control information, for the hot rolling process in general; and, more specifically, for fine tuning the line set-up for different gauges of same-grade rolling.

A Telling Example
Recent measurements of the effects of different skin pass reductions under simulated temper roll conditions suggested that a hypothesized increased dislocation density with increasing skin pass reduction starts at the surface and spreads progressively towards the center of the strip.  The Vision Research Q measurements showed a pronounced sensitivity to even the slightest temper mill reduction (fractions of a percent), while measurements made through the thickness of the strip showed a differential sensitivity to higher percentages of temper mill reduction [14]. Since the Vision Research Q measure is known to be a relatively shallow, surface measurement, compared to the through-strip measurements, which can reflect microstructural conditions much deeper in the strip; the differential response function obtained for the two different measurement techniques is expected under the hypothesis of a spreading dislocation density towards the center of the strip with increased skin pass reduction.

The Big Picture
Measurements are currently being conducted with prototypes of the surface volume imaging system to determine whether specific combinations of sensor frequencies and spatial resolutions can further delineate the microstructural effects of skin pass reductions.  However, at this point it would seem that the two very different measurement techniques of ultrasound and Vision Research Q already complement each other in providing a richer view of microstructural development by making possible a total strip volume imaging system.  The same type of intelligent perceptual network that greatly enhances the signal-to-noise performance and feature extraction speed of the surface volume sensory system would, of course, be capable of detecting and classifying total strip volume features as well, making a full volume imaging system much more sensitive and selective to a larger number of process relevant measurements.

Due to the subtle interactions of variations in melt chemistry (product type and grade) and process parameters (mill scheduling and set-up) costly on-line experiments are being conducted all over the globe everyday with each coil of hot rolled strip produced [15-21]. Unfortunately, the results of these valuable experiments are unavailable to the individual producers and the industry as a whole because, at this point in time, some obviously important dependent variables in these experiments are simply unobserved, unrecorded, or not connected in functional relationships to their causal antecedents. An intelligent network, with inputs from the melt chemistry to inspection of the final product before shipping, must be integrated by a “brain”, a model maker of the entire process, that can store and retrieve in real time the accumulated wisdom of countless hours of real-world, steel making experience [2,16,18].

We’ve Got The Sensors; We’ve Got The Brains
A few years ago this proposal would have be absurd; the necessary specialized sensors and fast execution AI software to deal with the complexities involved had not yet been created.  Winfree frames the problem as follows:

To find out whether any particular analytic model actually does these things, it would be necessary to numerically iterate the equations for reactions and diffusion on a three-dimensional grid of globs, starting from suitably arranged initial concentrations.  Moreover, a particularly fine grid would be required because we are looking for an instability and must be sure that it is not an artifact of our dividing the reagent into discrete cells.  Such a numerical simulation is prohibitively time-consuming even on the fastest computers.  Let us instead turn back to the chemical reagent as an analogue computer. [4]
In the case of steel making, the steel strip itself is our analogue computer.  But it must be possible to look into the computer after it has "crunched" all these complex simultaneous interactions and “read” the result.

Put a little differently, Too, et al. state:

...published literature [15]  has indicated that the hot rolling problem is extremely complex,  equations derived on this basis are of high non-linearity in nature.  The solutions to these equations are necessarily computing intensive, and therefore have limited applications to on-line process control where spontaneous response is essential.  In addition, factors such as oxide scaling [and] interfacial boundary friction cannot be treated effectively.  Process control is thus still a challenge to researchers. [16]
The integrated intelligent perceptual network, the combined sensory system and “brain”, is a perpetual model maker that does continually evolving pattern perception across the full developmental sequence of the steel production process [2,18]. The effectiveness of this technology is limited only by the quantity and quality of process relevant data the system is provided with. The First Law of Information Processing still applies: Garbage In, Garbage Out. If the system is seeded with high quality information, ideally, the customer can have custom products, delivered exactly when needed, produced by technologies that are more environmentally “friendly”, at globally a competitive price [17].

In Conclusion
“Smart” mechanical properties inspection systems for strip steel products should include the following design guidelines for optimal properties measurement resolution:

_____

References

1. K. L. Schafer   A sensor and method for the in-situ monitoring and control of microstructure during rapid metal forming processes. U.S. patent #5,420,518, May 30, 1995.

2. K. L. Schafer   “Genetic algorithms for optimizing and automating process control in critical-cycle continuous annealing”, 37th Mechanical Working and Steel Processing Conference and Int’l Symposium on Recovery and Recrystallization in Steel Proceedings, ISS, Warrendale, 1996, 33, 611-613.

3. K. L. Schafer   Presentation of MicroCert Research and Development Program to the American Iron and Steel Institute, Advanced Process Control Program, Project D, Online Mechanical Properties, I.M.I. Boucherville, Quebec, Oct. 10, l996.

4. A.T. Winfree   “Rotating Chemical Reactions”, Sci. Amer., 1974, 230, (6), 82-95.

5. J. D. Murray  “How the Leopard Gets Its Spots”, Sci. Amer., 1988, 258, (3), 80-87.

6. J. Majta, M. Pietrzyk, J.G. Lenard, and J. Janzen   “Prediction of Mechanical Properties of Steel Strips After Hot Rolling”,  37th  Mechanical Working and Steel Processing Conference and Int’l Symposium on Recovery and Recrystallization in Steel Proceedings, ISS, Warrendale, 1996, 33, 89-99.

7. U. F. Kocks   “Reliable Modeling of Complex Behavior”, in Modeling the Deformation of Crystalline Solids, (ed. by T. C. Lowe, A. D. Rollet, et al.) TMS, Warrendale, 1991, 175-188.

8. M. Pietrzyk, J. G. Lenard   Thermal Mechanical Modeling of the Hot Rolling Process. Springer-Verlag, Berlin, 1991.

9. J. Majta, J. G. Lenard, M. Pietrzyk   “On Modeling the Development of the Microstructure and Mechanical Properties of Microalloyed Steels”, Metallurgy and Foundry Engineering, 21 (1), 1995, 9-38.

10. J. Majta, J. G. Lenard, M. Pietrzyk   “Physical and Mathematical Simulation of the Evolution of the Microstructure During Hot Compression of a Nb Steel”, 7th ICM, The Hague, 1995.

11. T. Sakai et al.   “Deformation and Recrystallization Behavior of Low Carbon Steel in High Speed Hot Rolling”,  Trans. Iron Steel Inst. Japan, 1988, 28, 1028-1035.

12. H. Yoshida et al.   “An Integrated Mathematical Simulation of Temperatures, Rolling Loads and Metallurgical Properties in Hot Strip Mills”, Trans. Iron Steel Inst. Japan, 1991, 31, 571-576.

13. S. R. Wang, A. A. Tseng   “Macro and Micro Modeling of Hot Rolling of Steel Coupled by a Micro Constitutive Relationship”, IS&M, Warrendale, Sept., 1996, 49-61.

14. Confidential proprietary information deleted

15. J. J. M. Too, A. A. Tseng   Jour. Of Mat. Processing and Manufacturing Sci, 1992, 1, (2), 121-156.

16. J. J. M. Too et al.   “An Integrated Approach for the Control of Hot Rolling Processes”, 37th Mechanical Working and Steel Processing Conference and Int’l Symposium on Recovery and Recrystallization in Steel Proceedings, ISS, Warrendale, 1996, 33, 555-561.

17. L. W.  Kavanagh   AISI Testimony to the Subcommittee on Interior and Related Agencies of the Committee on Appropriations, U. S. House of Representatives, March 7, 1996.

18. J. H.  Holland   “Genetic Algorithms”, Sci. Amer., 1992,  267, (1), 66-72.

19. S. H. Anderson, M. Buffenoir, R. Oesterreich   “Possible Effects of Secondary Combustion for DRI Melting”, IS&M, Warrendale, April 1997, 27-31.

20. P. E. Nilles   “Recycling and Virgin Materials in the Changing European Steel Industry”, IS&M, Warrendale, April 1997, 33-40.

21. V. Pantea et al.   “The Application of Thermoelectric Power Measurements to the Study of the Recovery and Recrystallization in Cold-Rolled High Strength Low Alloy (HSLA) Steel Sheet” 37th Mechanical Working and Steel Processing Conference and Int’l Symposium on Recovery and Recrystallization in Steel Proceedings, ISS, Warrendale, 1996, 33, 987-995.
 

Home---- Top of Page


Copyright 1997 B. B. Tech Enterprises / Vision Research   63 Greene Street, Suite 204   New York, NY 10012   USA
PHONE:  212-966-0185   FAX:  212-334-4833