Reverse Engineering – Support for an Empirical Process
By John M. Switlik, Boeing Define and Bill Macy, Boeing Phantom Works

This article looks at reverse engineering (RE) as a necessary part of any attempt to model and to build parts or products using CAD/CAM/CAE (CoT) and knowledge based engineering (KBE). The intent is to look at where we are as an industry. With the growing use of advanced methods (CoT augmented with KBE), RE will play an increasing mode.

• For CoT, RE will be central to answering questions involved with determining if the as-built reality meets the as-designed specification, in the sense that quality control will continue to demand an increasing facility with continual measurement.
• For KBE, a complete knowledge loop may not be possible without the empirical feedback that RE can help to provide. Knowledge related to design and build must rely on a larger spectrum of physical and performance attributes than we see currently. The general mechanism for verifying computational states will either be a superset of RE or RE will morph to cover a broader context.

RE has the particular problem that it must deal with data that originates from physical entities and that it must create useful models from this data. Like all forms of modeling, RE face trades related to representation (structure, parameter), methods (in-house/COTS, procedures, algorithms/heuristics, physics), and quality (measurement, fit, continuity, footprint, completeness).

The article provides a general overview of the common issues related to obtaining the data (acquisition) and modeling the data (identification, fitting, smoothing). In the conclusion, the article lists open issues; subsequent articles will cover these two important steps in more detail and will look at future options.

Background
As one looks at the growing computer usage in a modern day setting, ‘ubiquity’ comes to mind, as the possible uses of systems are immense. Applications that deal with designing and building products (CoT and KBE) are one type of computational example that is becoming more involved and complicated with time.

As these types of use expand, it becomes necessary to consider quality concerns, where quality subsumes the methods that support continual verification and validation. Questions that can arise in this regard are several, such as: how well does a design meet a requirement, how well does a production process follow a design, how stable is the decision landscape to support automated control, and so forth.

There are many properties that need to be measured for a quality assessment, with some being more problematic than others. One central property can be modeled via geometry; RE comes about as a method to deal with this property type. In this case, geometry can involve more than representation, as the functional aspects may need to be an inherent part of the definition.

This geometry-oriented RE can be thought of as having two major steps that are each fairly involved.

• Acquisition – depending upon the particular item being measured, there are many options that exist and parameters that need to be controlled. For instance, the item may be in a fixture or may be freestanding. In either case, the captured data will be limited to the perspective and the point of view of the approach used. When there are multiple views to be joined, a mechanism must exist to relate the views reliably. These are simple examples.
• Modeling – the data obtained via Acquisition may be in one of several states after the capturing process. In general, there are strict requirements for data to be handled successfully by CoT. For instance, in the case that the data relates to an existing design, comparative analysis will depend upon how well the data is modeled and how well the data can be mapped back to the design. Too, for RE to influence KBE there needs to be strong mappings between the data and the model.

For each of these two RE steps, there will be trades related to representation, methods, and quality. Since this article covers a fairly large subject, it mostly looks at the general issues. Here and there will be brief reference to subsequent articles that will look at acquisition, modeling, and future options. There is a potential for RE to be an enabler of more powerful KBE, this will be the subject of a future article.

Representation
The concerns of representation deal, in part, with structure and parameter, which will differ between acquisition and modeling. In the acquisition context, there are setup choices related to how the data will be generated. There are many types of acquisition methods with different requirements for data handling. In addition, most entities cannot be captured from one framework; that brings up issues of how different (and potentially disparate) frames of reference will be mapped together. The necessary means for relating views can be more problematic in the freestanding mode than when there is a fixed environment; however this is never an easy issue to be taken lightly.

In the acquisition step, there are issues related to the types of entities that are within the defined view. The naïve notion that highly complicated relationships can be captured with one attempt is often dashed against the reality that identifying entities and their boundaries, let alone their relations, is very difficult.

Representation, in the acquisition mode, deals with how the data is stored. This storage can be in a large range of modes, which ought to support the requirements for the modeling step. One factor might be the amount of data. An open issue is the trade between heavy-density sample data versus less density-sample data with the addition of properties, such as normals, measured diameters, knowledge values (tube bend radii).

In the modeling step, representation must support the acceptance of acquisition data and allow the necessary operations to generate formats that are agreeable to CoT. A lot of the separation activity may be the onus of the acquisition step. In the more complicated situations, tying together acquisition and modeling more closely may be the necessary approach. That is, some types of features may require that RE provide a tight loop between the steps and that there be a means to separate out attributes, such as boundaries. Discussion of this required coupling is a common theme throughout these articles. Such a representation would be a multi-faceted and would require robust transformations between states.

Representation will influence both methods and quality. Given that RE can generate a lot of data, there are computational concerns dealing with time and space constraints. A more fluid flow through the states that are described in the next section will be very important to RE efficacy. As well, there is a continuum of fidelity levels that could be provided by RE for both acquisition and modeling (covered more fully in the Quality section).

Methods
Methods for acquisition and modeling have flavors that are similar to representation. Just as was mentioned in representation, close ties between acquisition and modeling for the more complicated cases is necessary to minimize efforts and maximize quality.

In general, the acquisition step is dependent on a predefined plan. This plan must consider what the product will be and how the data will be processed. Early on in this step, there would be a plan for how frames (points of view) are defined and their extent. An important factor would be the type of data that will be provided (CMM, laser tracker, white light). Also, other factors, such as ambient light, field-of-view, depth-of-field, access, and feature intensity will drive acquisition requirements.

Acquisition can include pre-process of the data and preparation of the data for exchange to modeling. The modeling step will support many higher-order requirements. As a total process, there are four major transition states that can be identified as common to the problem of going from a physical item to a CoT-worthy model of the item or to a KBE-fluent set of attributes.

These states are the following.

1. Obtain data – for geometry, the data may be points or point/normals where the point sets may be dense or not. A sub-process may establish relationships in the data. In terms of geometry, this might mean that a triangulated mesh is made from the points. At one time, this step was principally done in modeling; however some of this relating work can now be done early in the acquisition step. In terms of other data, relationships would facilitate separation of features.
2. Reduce data – this form of data improvement might include point decimation where the optimal point set that captures features is left after some reduction step. One strategy may be to replace multiple points on a primitive entity, such as a plane, with the entity definition. Again, this was formerly done in modeling but can very well be handled in acquisition. At this point, the data will be clean enough to support a reasonable approximation to the physical object via a polygonal surface.
3. Extract – this is where features and other separate entities are handled. Given the richness of the mesh, identifying characteristics can be used to separate out entities by functional role. For instance, determining a hydraulic line (tube) in a data field is an important requirement. Other entities may occur multiply, such fasteners. Currently, software in combination with geometry libraries is an effective means to rapidly populate known geometry in place of the point data. In the future, KBE will play a major role in extracting features or substituting geometry from these data sets based on rules and additional attributes that can be brought to bear.
4. Model – at this point, the data has been fairly well improved and ought to be capable of identifying objects and their attributes such as to support analysis. In terms of CoT requirements, a sub-process may generate surface representation. For instance, in order to make the transformation the NURBS, the data needs to be separated such that there are 4-sided pieces that jointly cover the data and such that the surfaces conform to CoT standards.

These steps can be thought of as independent and in sequence. However, they are linked via the state transitions of the data as it moves toward the knowledge model.

The power of the acquisition and the modeling steps may come from an intelligent coupling. Acquisition does have the presence of the data source that could allow data re-samples thereby providing a more coherent basis to modeling than just a bunch of data. Modeling has access to more advanced knowledge and could help guide acquisition fill in missing pieces. The post-process demarcation (line between acquisition and modeling) will need to be adjustable by context.

The extraction of features results in a diminishing of the cardinality of the sampled points, which is a form of partitioning potentially smarter than separation by applying methods, such as convolution. Too, a method that is sensitive to functional knowledge can help reduce the amount of noisy points (assuming that all the acquired data is not carrying information).

Quality
Quality concerns apply to both the acquisition and modeling steps. As the industry has matured, there has been growth of experience in acquiring data reliably and accurately. Yet, there can be data that is extraneous or that carries misinformation.

In the long list of acquisition hardware, experiential data allows descriptions of performance, such as in the tolerances that can be expected and in the rate of acquisition. One issue is certification in terms of verifying that the vendor specifications are reasonably stated. This can be problematic since the acquisition process can be subject to many possible sources for noise and other interference. A concern may be the time frame that calibration can be expected to hold or that the hardware provides some evaluation of its current state of calibration.

One thrust of RE would look to provide the necessary fidelity required by the application that will use the data. There can be fidelity ratings established along all the states described in Methods that would measure fit and footprint attributes in terms of expected values. The presence of RE successes shows that these issues can be resolved for certain types of situations. An open issue is how to set the fidelity requirement by context.

In terms of supporting as-built analysis, RE’s focus on geometry can lead to a close mapping for fit and form measurement. As well, RE can provide data for design as exemplified by success demonstrated in custom-fit clothes and shoes provided by some manufacturers.

Support for function analysis will be brought forth by the growing use of KBE in that algorithms and heuristics are sensitive to the representation of which fidelity is an important aspect. That is, a method will be directly dependent upon the data configuration. A classic example can be found in the many types of meshes that are possible and in the preference of one over the other given the problem. For instance, one mesh type may be better for flow analysis versus statics analysis, such as stress. RE will need to address these types of concerns when supporting KBE verification.

Conclusion
This article looks at the current state of reverse engineering (RE) in terms of motivation and need. The two steps associated with RE can be identified as acquisition and modeling. That these two need to be fairly closely managed is a prime concern; quality measures need to be applied along both steps. Given that we can separate the two steps, we can look in more detail at each. The next article will address the issues of data acquisition; following that, the modeling issues will be reviewed. A follow-on article will look at the ways that RE can morph to support the growing horizons of advanced computational modeling.

For further discussion of the content of this article, please e-mail the authors.

About the authors
John M. Switlik is an Associate Technical Fellow (Boeing, Commercial Airplane, Wichita Define Systems, KBS, john.m.switlik@boeing.com). His interests cover a broad spectrum: computational support for applied science, mathematics, and management; lifecycle and veracity control for system, application, and domain engineering; integration, use, and quality of computational knowledge.

Bill Macy is a Senior Project Engineer (Boeing, Phantom Works, St. Louis, william.d.macy@boeing.com). His interests focus on the need to leverage reverse engineering metrology and software to provide affordable and better quality engineering, fabrication, and support of products.

Definitions:
CoT - CAD/CAM/CAE/etc. plus "Top Gun" talents. This element does not subsume CyB. KBE – Knowledge Based Engineering NURBS – Non-uniform, rational, b-spline. A standard representation for splines (curve, surface, solid), that is parsimonious, powerfully adaptive, and capable of high-order smoothness.
RE - Reverse Engineering covers all activities related to capturing geometry data about a physical part and to creating a computer model that can be used by CoT. Plus, RE serves as a prototype for a fuller link between computational models and associated physical states.